Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Residential housing code violation prediction: a study in Victorville, CA using geographically weighted logistic regression
(USC Thesis Other)
Residential housing code violation prediction: a study in Victorville, CA using geographically weighted logistic regression
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Residential Housing Code Violation Prediction:
A Study in Victorville, CA Using Geographically Weighted Logistic Regression
by
Matthew Dean Pugh
A Thesis Presented to the
Faculty of the USC Graduate School
University of Southern California
In Partial Fulfillment of the
Requirements for the Degree
Master of Science
(Geographic Information Science and Technology)
May 2016
Copyright © 2016 by Matthew Dean Pugh
I dedicate this paper to my loving parents, my brother, and to all of my family and friends who
supported me throughout this entire process. Without you, none of this would have been
possible.
iv
Table of Contents
List of Figures .....................................................................................................................................vii
List of Tables .................................................................................................................................... viii
Acknowledgements ............................................................................................................................. ix
List of Abbreviations ............................................................................................................................ x
Abstract ................................................................................................................................................ xi
Chapter 1 Introduction.......................................................................................................................... 1
1.1 Motivation................................................................................................................................. 3
1.2 Research Questions .................................................................................................................. 6
1.3 Overview of Research Design ................................................................................................. 6
Chapter 2 Background and Related Work ........................................................................................ 10
2.1 Property Values and Code Enforcement ............................................................................... 11
2.1.1. What is Code Enforcement?........................................................................................ 11
2.1.2. High Quality Landscaping Increases Property Value................................................ 12
2.1.3. Managing Residential Properties through Code Enforcement .................................. 14
2.1.4. The Effect of Code Enforcement on a City ................................................................ 16
2.2 Crime Prediction .................................................................................................................... 18
2.2.1. Traditional GIS Based Approaches ............................................................................ 19
2.2.2. Regression Modeling Techniques ............................................................................... 21
2.3 Logistic Regression Techniques............................................................................................ 21
2.4 Geographically Weighted Regression Techniques .............................................................. 23
Chapter 3 Data and Methods .............................................................................................................. 27
3.1 Research Design ..................................................................................................................... 28
3.1.1. Non-spatial Logistic Regression Technique .............................................................. 28
3.1.2. Geographically Weighted Logistic Regression Technique ....................................... 30
3.2 Data Requirements & Data Sources...................................................................................... 32
3.2.1. Dependent Variable ..................................................................................................... 33
v
3.2.2. Independent Variables ................................................................................................. 33
3.2.3. Variable Justification ................................................................................................... 37
3.3 Procedures & Analysis ........................................................................................................... 38
3.3.1. Study Areas .................................................................................................................. 39
3.3.2. Individual Analysis of the Three Study Areas ........................................................... 40
3.3.3. Combined Area Analysis............................................................................................. 41
3.3.4. Geographically Weighted Logistic Regression Analysis .......................................... 42
3.3.5. Model Validation and Making Predictions ................................................................ 44
Chapter 4 Results and Predicted Violations ...................................................................................... 46
4.1 Binary Logistic Regression Results ...................................................................................... 46
4.1.1. Interpretation of Results .............................................................................................. 47
4.1.2. Logistic Regression Model Results ............................................................................ 54
4.1.3. Logistic Regression Coefficients ................................................................................ 56
4.2 GWR 4 Results ....................................................................................................................... 60
4.2.1. Interpretation of GWR4 Results ................................................................................. 61
4.2.2. GWR4 Model Results.................................................................................................. 64
4.3 Selection of the Prediction Model ......................................................................................... 67
4.4 Predictions .............................................................................................................................. 69
4.4.1. Prediction of Violations............................................................................................... 69
4.4.2. Model Validation ......................................................................................................... 73
Chapter 5 Discussion and Conclusions ............................................................................................. 75
5.1 Findings .................................................................................................................................. 75
5.1.1. Non-spatial Findings.................................................................................................... 75
5.1.2. Spatial Findings ........................................................................................................... 80
5.1.3. Predictions .................................................................................................................... 82
5.2 Relation to Previous Work..................................................................................................... 84
5.3 Limitations and Future Work ................................................................................................ 86
5.3.1. Major Limitations ........................................................................................................ 87
5.3.2. Future Work ................................................................................................................. 89
5.3.3. Conclusions .................................................................................................................. 91
Appendix A: Area 2 and Area 3 Maps .............................................................................................. 97
vi
Appendix B Example SPSS Output................................................................................................... 99
Appendix C Logistic Regression Results Full Summary Table .................................................... 102
Appendix D Example GWR4 Output .............................................................................................. 110
Appendix E GWR4 Coefficient Maps............................................................................................. 122
vii
List of Figures
Figure 1 The City of Victorville .......................................................................................................... 2
Figure 2 Study Areas 1, 2, and 3.......................................................................................................... 8
Figure 3 A simple logistic regression curve...................................................................................... 29
Figure 4 Area 1 neighborhood and observed violations ................................................................. 40
Figure 5 Global regression result for Area 1 in GWR4 ................................................................... 63
Figure 6 Area 1 bandwidth selection output ..................................................................................... 63
Figure 7 Local model output for Area 1 ............................................................................................ 64
Figure 8 Floor area variable coefficient values of the GWLR analysis .......................................... 67
Figure 9 Predicted housing code violations based on logistic regression ....................................... 70
Figure 10 Neighborhood comparison of predictions ........................................................................ 72
Figure 11 Model validation neighborhood ........................................................................................ 74
Figure 12 GWR4 map of coefficients for days to previous violation variable ............................... 81
viii
List of Tables
Table 1 Independent variable list ....................................................................................................... 34
Table 2 Dummy variables for GWR4 ............................................................................................... 43
Table 3 Area 1 multicollinearity test ................................................................................................. 48
Table 4 Area 1 variable selection with non-significant variables.................................................... 49
Table 5 Area 1 remaining variables once non-significant variables were removed ....................... 49
Table 6 The null model significance test .......................................................................................... 49
Table 7 The null model predictions ................................................................................................... 50
Table 8 Area 1 model significance test ............................................................................................. 51
Table 9 Area 1 pseudo R squared values .......................................................................................... 51
Table 10 The coefficient and odds ratio output table ....................................................................... 53
Table 11 The overall number of cases predicted correctly for Area 1 ............................................ 54
Table 12 The results of the model iterations and key statistics ....................................................... 54
Table 13 Results of the coefficient calculations, the odds ratio, and significance
for the logistic regression models ............................................................................................. 56
Table 14 GWR4 model iteration summary table containing important statistics ........................... 65
Table 15 Observed predicted accuracy of the model ....................................................................... 73
ix
Acknowledgements
I would like to thank my advisor, Dr. Karen Kemp, for her guidance and encouragement in
completing this project. I would also like to thank my committee members, Dr. Robert Vos and
Dr. Su Jin Lee, for their help in getting me to this point. I thank my parents, Allan and Donna,
and my brother Steven for their constant support throughout my entire student career. Also, my
friends Randy, Audrey, Bobby, Monique, Thomas and his wife Pilar, and Robbi for their support
and understanding, and for reminding me to have a little fun from time to time. I also thank
Robert Chang, PsyD for his sound advice and encouragement. I would also like to acknowledge
my employers, the City of Victorville and Mimi Song Company for allowing me to use their
resources to complete this project.
x
List of Abbreviations
ABC Alcoholic Beverage Control
AIC Akaike information criterion
BLR Binary logistic regression
CD Compact disk
GIS Geographic information systems
GWR Geographically weighted regression
GWLR Geographically weighted logistic regression
IDRE Institute for Digital Research and Education
NAD North American Datum
OLS Ordinary least squares
SFR Single-family residence
SPSS Statistical Package for Social Sciences
UCLA University of California Los Angeles
USC University of Southern California
VIF Variance inflation factor
xi
Abstract
Cities throughout the country are constantly striving to improve their perceived image. Whether
it is requiring lush landscaping in commercial developments, or simply making sure that the trim
on a house is properly painted, cities are constantly struggling to get citizens to comply with
municipal codes. Such is the case in the City of Victorville, CA, where economic recovery has
been slow following the 2008 housing market crash, leaving poorly maintained properties in its
wake. Presently, Victorville’s code enforcement staff is doing a proactive enforcement survey of
all single-family homes in the city in an effort to “clean up” these properties. However, the
survey is inefficient and is taking up a good amount of officer time, leaving commercial and
industrial areas of the city neglected. This project was able to predict which houses in
Victorville are likely to have a code enforcement violation that requires action from staff in order
to better allocate resources to areas that require more attention and pull resources from areas that
do not require attention. The primary question here is what property attributes can be used to
predict the occurrence of a code enforcement violation? Several have been selected, including
property value, length of ownership, and presence of a previous violation. A binary logistic
regression analysis was run on three areas of the city containing approximately 2,200 homes that
have already been surveyed in order to train a model for predicting the remaining 29,000 homes.
Geographically weighted logistic regression was then employed to factor in spatial variation in
the relationships between the response variable and the explanatory variables. The success of
this model will make Victorville’s code enforcement more efficient, and it is a model that any
city can employ to make its own code enforcement departments more effective.
1
Chapter 1 Introduction
In the City of Victorville, which is situated in the Mojave Desert region of Southern California
(Figure 1), there has been a complete shift in the quality of residential housing due to the
construction boom of 2004 to 2008. Developers from all over came to the City and built low
quality homes at a high volume in order to make the largest profit possible. Compounding this
issue was a lack of planning and urban design that is essential in creating sustainable
neighborhoods. Once the market crashed at the end of 2008, developers left town to seek their
riches elsewhere. Left in their wake were architecturally basic, cheaply constructed single-
family homes that are already showing signs of decay and blight. In the years that followed,
property values remained low and a new wave of investors flooded the City to take advantage of
homes that had been lost to bank foreclosures, short sales, and utter abandonment. The result
was a high number of renter occupied homes that went unmaintained because occupants had no
sense of place and owners of rental properties went unaccountable because they did not live in
the area. This left front yards devoid of landscaping with weeds, deteriorating fences, inoperable
vehicles, trash and debris, and an overall look of decay.
For these reasons, the City has adopted a new code enforcement policy of proactive
enforcement of municipal laws that govern the aforementioned property maintenance issues in an
effort to clean up the City, add some stability to residential districts, and allow neighborhoods to
both retain and increase their property values. This has resulted in the hiring of three new code
enforcement officers and an influx of new code enforcement action cases that take hours of
manpower to rectify. In addition, as the seven total officers concentrate on residential properties,
commercial and industrial properties get neglected. This creates inefficiency that the code
enforcement manager and City Council need to address. This research project attempted to find
2
a better way of allocating those code enforcement resources by using statistical analysis
techniques and a set of explanatory variables to predict which areas in the City are likely to have
the greatest need for enforcement action. By doing this, City officials should be able to better
manage their personnel, budget, and infrastructure in addition to providing a better level of
service to the City’s residents and businesses.
Figure 1: The City of Victorville
3
1.1 Motivation
The City’s code enforcement division is currently conducting a citywide proactive
investigation on single-family residential neighborhoods, logging which properties have housing
code violations and which properties do not. Full neighborhood investigations are conducted and
formal written notices of violation are being given to property owners or the occupant of the
residence. The result is a set of notices with a property address that can be mapped easily at the
parcel level; properties that do not have a violation are not given a notification. The code
enforcement division has investigated roughly 2,300 single-family properties out of
approximately 29,000 total single-family residences as of July 11, 2015, which is the date that
the data used in this project was obtained from the City. The senior code enforcement officer
estimates that this project will take another 18 to 24 months to complete, if not longer. This is
putting a strain on officer time as they must also continue to address the other properties in the
City where code enforcement action is necessary, such as in commercial or industrial areas.
Making this process more efficient would greatly benefit the City as a whole and free up officer
resources to fulfill all responsibilities for city code enforcement.
To do so, this project made predictions on which single-family residences are likely to
have housing code violations and which are not likely to have housing code violations using the
binary logistic regression technique. This technique is able to model binary outcomes such as
yes/no, pass/fail, dead/alive, or 0/1. The model used various measures that either pertain to the
property itself or to the neighborhood it is situated in and gave each property a value between 0
and 1 for where it fell on the logistic regression curve, or line of best fit. The result was a yes or
no value depicting whether a code violation was likely or not. The resulting model can be used
by the City’s code enforcement staff to identify neighborhoods that should be given extra
4
attention or neighborhoods that will require very little attention. This should give staff a clear
picture of the current state of the City and how they should proceed with their proactive
inspection program to most efficiently manage the situation.
The hope is that this research project will extend beyond the scope of this analysis and be
a great benefit to the City of Victorville. As mentioned before, the City is having great difficulty
in keeping its residential housing stock in proper repair because of various economic and social
issues. A successful regression model should be able to reliably predict areas where property
maintenance is likely to be an issue, and the local government will be able to efficiently allocate
its code enforcement resources to areas that will require the most attention. This will also give
city officials a glimpse into the main contributing factors that cause residential properties to fall
into decay, thus allowing officials to act accordingly. In addition, this project could help to
alleviate the strain on property values from poor maintenance and hopefully raise residents’ pride
in their neighborhood and community. Furthermore, it could be used by other cities that are also
struggling with their residential zones because the variables used in the analysis are not specific
to Victorville. Instead, they are variables that are related to the characteristics of the homes
themselves and could be obtained in any county or city that maintains property characteristic
data and assessed value data.
To the best of my knowledge, binary logistic regression modeling has not been used to
predict the occurrence of residential properties that require code enforcement action. Therefore,
this project adds to previously conducted studies for prediction. Spatial predictive analysis has
been conducted on crime (Antolos et al. 2013; Liu & Brown 2003), property abandonment
(Morckel 2014), fire hazard potential (Rodrigues, de la Riva, & Fotheringham 2014; Martinez-
Fernandez, Chuvieco, & Koutsias 2013) groundwater spring potential (Ozdemir 2011), riverbank
5
erosion potential (Atkinson et al. 2003), and landslide hazard potential (Kundu et al. 2013). This
project gives another aspect to what the binary logistic regression technique can be used for in
spatial analysis and prediction. Furthermore, this project will give researchers another look into
how spatial regression techniques can be implemented in manners that effect public policy and
decision-making for housing regulation.
In addition, there has been little research into how a binary logistic regression model can
be applied to geographic data, and more specifically, how the occurrence of a dependent variable
is affected by the location of another dependent variable in the dataset; in this case, a housing
code violation. Therefore, this project adds to work done using geographically weighted
regression with a binary outcome. Research conducted previously includes that of Atkinson et
al. (2003) where they employed a geographically weighted logistic regression (GWLR) model
for erosion susceptibility in order to study the effects of local phenomena that varied with
location, such as distance upstream. In addition, Luo and Kanala (2008) employed GWLR to
model urban growth and land use change over time. Also, the work of Martinez-Fernandez,
Chuvievo and Koutsias (2013) accounted for local variations in wildfire occurrence variables
using GWLR. This project will add to this previous research and will give an additional field of
study for the GWLR technique to be employed beyond physical geography or sociology.
Moreover, the ability to predict other municipality related phenomena such as which
properties are likely to become rental homes or which properties are likely to have crime could
be profoundly useful to a local government. The techniques employed here could ultimately lead
to other models that may be used by governing authorities to make their internal processes more
efficient. This could lead to more man hours available for other projects and more budgetary
funds to apply to other programs or departments. Even though this project is covering a very
6
small aspect of what local governments do, it potentially has the ability to provide lawmakers
and officials with the ability to simplify their processes and provide a more effective form of
government.
1.2 Research Questions
The goal of this project is to predict whether or not a single-family home in the City of
Victorville is likely to have a code enforcement violation. This project will answer the following
questions:
1. Can certain property attributes predict the occurrence of code enforcement
violations?
2. Are there relationships between a home with a code enforcement violation and its
neighboring properties?
3. Which single-family residential properties in Victorville are likely to have a
violation?
Questions 1 and 2 will be answered as a result of the logistic regression and
geographically weighted regression techniques. These questions are necessary because they
essentially determine if this project has any validity in the field of geography. Question 1 will
determine if the subsequent model is of any value scientifically and Question 2 will determine if
spatial location does in fact play a role in how the physical characteristics of a neighborhood can
be affected if there are a few properties that are substandard or run down. Question 3 is of
course the goal of this research project.
1.3 Overview of Research Design
Since this model was predicting a binary outcome and some of the explanatory variables
are non-continuous, techniques such as ordinary least squares and traditional geographically
7
weighted regression cannot be used. Following the examples of Lee and Sambath (2005), Wu
and Zhang (2013), Martinez-Fernandez, Chuvieco, and Koutsias (2013), and Rodrigues, de la
Riva, and Fotheringham (2014), a binary logistic and a geographically weighted logistic
regression were created using various explanatory variables related to individual properties.
Variables used were lot size, assessed land value, ownership type, length of ownership, year of
construction, or whether or not a property is tax defaulted. The study focused on three specific
areas within the City that have already been fully surveyed by code enforcement staff (see Figure
2). The binary logistic regression analysis was conducted on each of these areas independently.
They were then combined into a single dataset where the analysis was conducted on a set of
random samples from the dataset as well as on the data as a whole. The model was then
analyzed for whether or not it violated any of the key assumptions of logistic regression, which
are outlined later in this document.
The process was then repeated using only geographically weighted logistic regression
within the GWR4 software developed by Tomoki Nakaya and distributed by Arizona State
University (Nakaya 2009). The intent was to improve upon the logistic regression model by
introducing spatial variation into the equation. This sought to illustrate that single-family
residences can be affected by neighboring residences in regards to the need for code enforcement
action. The GWR4 software produced variable coefficient values and a constant value that were
then used in the geographically weighted logistic equation to make predictions for the remaining
single-family homes in the City.
8
Figure 2: Study Areas 1, 2, and 3
The model was validated by evaluating the important assumptions of binary logistic
regression and whether or not the model violated any of them. Next, the model was tested on an
9
area of the City that was proactively surveyed after the date of data collection of this project to
determine how well the model performed as a predictive tool. The model performed well in
terms of identifying neighborhoods that are likely to have violations, but it was not strong
enough to confidently identify individual properties that are likely to have a violation.
The following chapters explain the details of how these predictions were made and the
steps that were necessary to arrive here. Chapter 2 discusses the role of code enforcement and
how property owners and neighborhoods are affected by code enforcement. It also discusses
how crime incidents are predicted using geographic statistics because crime prediction has many
similarities to code enforcement prediction, and it describes the applications of how logistic
regression and geographically weighted logistic regression can be used as a prediction tool in the
realm of geography. Chapter 3 outlines the methodology employed in creating this model, what
data was needed, and how the predictions were made. Chapter 4 discusses the results of the
analysis and how well the model performed as a predictive tool. Finally, Chapter 5 illustrates the
limitations of this model, how it tied in to the related work that precedes it, and what can be done
in the future to make this a stronger predictive tool.
10
Chapter 2 Background and Related Work
Cities in the United States have been dealing with the phenomena of residential housing decline,
crime, and urban blight for decades. Generally, housing that was once new and attractive to
potential residents slowly falls into a state of decay as homes or apartment buildings fall behind
on general maintenance such as painting, landscape pruning and trimming, window replacement,
roof replacement, or fence replacement. This, in conjunction with higher crime rates, gives the
appearance that residential areas are dark, dangerous and no longer desirable places to live. As a
result, cities must use their code enforcement resources to force residents to perform these
essential property maintenance tasks to improve not only the neighborhood, but also the overall
appearance of the city so that new residents will continue to move there, thus increasing the
city’s tax base and revenue (Anderson and Cordell 1988).
The City of Victorville is currently in this process of neighborhood revitalization through
proactive code enforcement. This research project attempted to predict whether or not a property
within the City would have some kind of code enforcement violation. In addition, an
understanding of how the occurrence of an incident can be predicted through spatial analysis was
essential. Crime prediction using spatial analysis was a good comparison because crime
incidents and code enforcement incidents share similar attributes, such as they must be reported,
they are point based, and they involve violation of a law. Next, various regression modeling
techniques were analyzed to determine how prediction of incidents is possible through both
traditional statistical methods as well as through spatial prediction techniques such as
geographically weighted regression. Finally, the culmination of this review leads to a scientific
and tested method with which the research questions of this research project were answered.
11
2.1 Property Values and Code Enforcement
Observations of most cities in the country reveal that there are areas with high property
values and areas with low property values. Often times, areas with the lower values are the
oldest in age, have the highest crime rate, have high occupancy turnover rates, and are very
poorly maintained. The primary mechanism with which a municipality or other governing
agency can attempt to address this is through enforcement of housing codes designed to maintain
standard living conditions such as adequate lighting and ventilation, fire egress, heating and
cooling, and proper roof maintenance (Meier 1983).
2.1.1. What is Code Enforcement?
Cities across the country employ code enforcement officers to ensure that both building
and construction codes as well as municipal codes are properly met. These codes can be
anything from requiring that a citizen obtain a building permit for their new patio cover to
requiring that a citizen pull all of the weeds from their front yard or move the inoperable vehicle
off the driveway and out of sight from neighbors. Should citizens choose to disobey the code
and the code enforcement officer, the office has the authority to issue a monetary citation against
the property with the violation as an incentive for the citizen to comply. If compliance is still not
achieved, a recorded lien can be placed on the property preventing sale or transfer of ownership
until fines are paid and compliance is met. These violations are reported in a manner similar to
crime insofar as a citizen must report the violation or file a complaint with the city’s code
enforcement department. An officer must then respond to the report or complaint and enforce the
necessary codes just as a police officer would enforce the necessary laws or penal codes. As the
City of Victorville attempts to improve its image by cleaning up the residential properties
through proactive code enforcement, special attention will be given to front yard aesthetics and
12
landscaping.
2.1.2. High Quality Landscaping Increases Property Value
There have been several studies conducted that show a positive correlation between
landscaping and greenspace and its effect on property values. This is intuitive, especially for
most homebuyers as homes that are the most visually appealing on the outside attract the most
attention. Des Rosiers et al. (2002) looked at how tree cover, ground cover, lawn cover, and
landscape structures increased property values for homes that had these amenities over homes
that did not. They found that landscape features such as a hedge or flower arrangement increased
homes values by up to 3.9% in some cases, while decorative structures such as patio covers
increased property values by as much as 12.4%. The architecture of the building was even found
to have a significant impact on the results as bungalow and cottage style homes with quality
landscaping had the highest values over row houses with similar landscaping.
A study by Dombrow, Rodriguez, and Sirmans (2000) conducted in Baton Rouge, LA
attempted to assess the value added by mature trees on a residential property. They used a
multiple regression approach using home sale data from the area and found that regional
assessors typically added around 2% to the total value of the home if it had mature trees on the
property that were aesthetically appealing and provided shade and other benefits to the property.
This coincides with a previous study done by Anderson and Cordell (1988) in Athens, GA that
found that single-family homes with landscaping with trees subsequently saw an increase in sales
prices by 3.5 to 4.5% over properties that did not. Both of these studies used a regression model
and found statistically significant correlations between the sale price and the presence of
landscaping and trees.
A study conducted by Luttik (2000) also looked at how natural landscapes can have
13
substantial effect on the value of a home. While the study of Des Rosiers et al. (2002) only
surveyed 760 single-family homes in Quebec, Canada, Luttik (2000) surveyed almost 3,000
single-family homes over a large region of the Netherlands. Other than the number of homes in
the survey, these studies were very similar. Both used a hedonic price model where they analyze
the effect of the total price by specific property characteristics. In essence, they both determine
the effect of landscaping, location, property size, and architecture style separately and then
combine them into a total property value. While Des Rosiers et al. (2002) looked at property
landscapes, Luttik (2000) looked at location adjacent to greenspace, view of greenspace, view of
water features, and landscape diversity. They found that homes located directly on a lake had the
highest increase in property value at roughly 12%, those that were bordering an area with a park,
forest, or lake had increases of roughly 5% to 8%, and homes not located near greenspace
adjacent to more urban areas with multi-story apartment buildings actually saw a decrease in
property value by up to 7%. Unfortunately, the City of Victorville does not have any scenic
water features; however, there are still greenspaces such as parks and golf courses that can have
an impact on surrounding property values.
In addition to Luttik (2000), the study of Conway et al. (2008) attempted to model how
property values are affected by greenspace in surrounding areas. Conway et al. (2008) expanded
on the hedonic approach of Luttik (2000) to include a spatial lag model to account for the spatial
autocorrelation determined to be present in the hedonic model. They discovered that even with
spatial autocorrelation removed through the spatial lag model, there are positive impacts on
property value from greenspace. However, the effect deteriorates with distance as greenspaces
greater than 400 feet away from a property were shown to have negligible impacts on property
value. This clearly shows that Tobler’s First Law of Geography is in effect here, as objects in
14
close proximity to one another exert greater impact than objects further apart (Tobler 1970).
What these studies show is that there is a significant effect on property values from
greenery and mature landscaping. Trees, water features, scenic views, and overall access to open
space are shown to increase residential property values by roughly 3% to as much as 12% in
some instances. In all of these cases, a regression model was employed in order to see how well
these landscaping and open space characteristics affected the dependent variable of property
value. Each one sought to determine which factors were most influential in order to make
predictions on property values in future transactions. For this project, these studies act as a
justification for the City of Victorville’s current proactive code enforcement policies that are
aimed at increasing property values by ensuring that residents maintain their property appearance
with proper landscape maintenance.
2.1.3. Managing Residential Properties through Code Enforcement
For decades, cities of all sizes have attempted to keep residences within their jurisdictions
in a good state of repair through code enforcement and other forms of criminal and legal
processes. Many studies were done in the 1960s to analyze what needed to be done with regards
to housing conditions. This came during a time of heavy housing reform that ultimately led to
the creation of the Department of Housing and Urban Development (HUD) in 1965, which was
charged with the regulation of housing in the United States. Housing acts such as the Fair
Housing Act of 1968 and the Housing and Community Development Act of 1974 were also
passed in an attempt to increase housing quality in the country. These acts also created Section 8
housing, which is a government program that helps renters offset their rental costs if they have a
low enough income level to qualify (Teater 2011).
Two studies done during this time still have relevance in 2015. One such study published
15
in the Harvard Law Review by Carlton, Landfield, and Loken (1965) examined how municipal
housing codes were enforced at the time and offered a different approach to more effective
results. Another study from the Columbia Law Review by Gribetz and Grad (1966) looked at
how code enforcement evolved over time, what the primary issues were at the time of the article,
and what could be done to alleviate some code enforcement inefficiencies.
Both of these studies pointed out several issues with the overall code enforcement
process, including the idea that full results are almost never achieved and that it relies on a
criminal process that is often slow and inefficient. Carlton, Landfield, and Loken (1965) note for
rental housing that code violations are often given to the property landlord and not to the tenant
who is more often than not the party responsible for these violations. This leads to a conflict
between the landlord and the tenant on who should ultimately take responsibility for the
violation, which in turn disrupts the mitigation process. They also discuss the difficulty of
enforcing housing codes across a municipality in an un-biased manner. This is largely due to the
fact that many housing codes are broad and require much interpretation that gives way to unfair
enforcement and potential corruption. The authors went on to describe the various “remedies” of
code enforcement such as fines, liens, and a judicial process that are still present in today’s code
enforcement process. In addition, they describe how New York City also imposed rent control
and withholding proceedings to enforce housing code violations.
These discussions were reinforced in Gribetz and Grad (1966) as they also discussed
New York’s rent control process, as well as the idea that the code enforcement process is slow
and has many hurdles that must be cleared both physically and administratively. The researchers
point out that the criminal code enforcement process does not work very well on major violations
because there is often a financial barrier preventing an owner or tenant from correcting the issue.
16
This leads to a long and drawn out process of citations and inspections that ultimately increase
financial burden due to fines and various other actions that require some form of monetary
transaction that does not go directly to correction of the violation. They propose a stratified
system of enforcement where infrequent violators are given more leniency versus a “hard-core”
violator that must be taken to court.
Both of these articles give several insights into how housing code violations can be dealt
with by the authority having jurisdiction. The laws and processes that were discussed here are
still in effect today and cities across the nation, including the City of Victorville, still follow the
process of citations, fines, liens, and ultimately criminal sanction to address the violations with
cities. However, as the articles discuss, these tools and processes are often inefficient or slow
which causes violations to linger and passes financial obligations for these processes on to
property owners and other violators. This in turn causes a circular effect resulting in more
required action by both the city and the property owners which leads to resentment and hostility
on both accounts (Carlton 1965; Gribetz and Grad 1966).
2.1.4. The Effect of Code Enforcement on a City
Many have argued that code enforcement has either a positive or a negative impact on a
city. On one hand, there is the idea that active code enforcement delays residences from falling
into decay as they age, and code enforcement is essentially not needed in high income level
neighborhoods. On the other, there is the notion that people living in run down residences do not
have the disposable income to pay for routine maintenance of their homes. The result is a
downward trend that is nearly impossible to escape. This section looks at both arguments.
The positive effects of code enforcement were studied by Ron Meier (1983) where he
analyzed how the City of Pasadena, CA managed code compliance through an inspection
17
program that was required each time a residence changed occupancy. The city was divided using
existing Census tracts and were then categorized by income level. He then analyzed the change
in property value over time and compared it to how active the code compliance inspection
program was in these tracts. The result was that upper-middle class homes were not affected by
the code inspection program which saw a natural market value increase due to neighborhood
affluence. The lower class neighborhoods that had the most active code inspection activity saw
some increase in value from code compliance inspections, but it was not significant compared to
middle class residences. This study found the most significant impact on property value came in
these middle class neighborhoods where code compliance inspections uncovered minor
violations that were easily fixed because either the cost of making these repairs was passed on in
the selling price, or the landlord was able to take care of these issues before new tenants
occupied the residence. These findings indicate that there is a need within Pasadena to allocate
code enforcement resources to less affluent neighborhoods because the effect is more substantial.
The negative aspect of code enforcement was discussed in the work of Miller (1973),
Ross (1996), and Burby et al. (2000). Each of these studies discussed how too much code
enforcement can prove to be detrimental to code enforcement’s goal of cleaner and safer
neighborhoods. Miller (1973) discussed the economics of code enforcement with emphasis on
the idea that some code enforcement is good, but there is a point where a breaking point is
reached and property owners must simply “walk away” from the property, especially if they are
renting. Miller (1973) looked at code enforcement in terms of five levels of enforcement with 1
being the lowest and 5 being the highest. He notes that at level 1, there is no financial burden
placed on the owner but there is no oversight on how that property is maintained. At level 2 and
3, there is some financial burden, but the property is maintained because of pressure from
18
inspectors or officers. At level 4, the property owner begins to take a loss on the property
because the financial burden of code enforcement requirements, property taxes, and other costs
exceeds what they earn in rent, but they continue to hold the property. At level 5, the financial
burden is too great and the owner walks away and abandons the property or sells it. At the
neighborhood scale, a code enforcement level of 5 would cause more problems than it would fix
because there would be a high level of abandoned properties that would be subject to vandalism
and blight.
In addition, the work of Ross (1996) and Burby et al. (2000) further emphasize that too
much code enforcement leads to urban decline. Both of these studies give examples of how too
much government oversight or too much discretion on the part of the inspectors ultimately places
too much of a burden on developers and property owners to make it feasible for them to continue
building or owning property. Furthermore, they both discuss that code enforcement must be
tailored to each neighborhood based on social and economic factors in order for it to succeed in
improving neighborhood quality while at the same time encouraging residents to remain. Ross
(1996) goes on to point out that overbearing code enforcement leads to property abandonment,
just as in Miller (1973), but he also notes that abandonment is contagious and can cause low
income neighborhoods that are otherwise stable to change into slums that are ridden with crime
and drug abuse.
2.2 Crime Prediction
Crime enforcement and code enforcement share several similar characteristics. Both
require an officer to respond to a citizen’s call or complaint, both require a location to respond
to, and it is advantageous to be able to predict where these instances will happen. For this
reason, this section takes a look at the work that has been done in geographic information
19
systems (GIS) to predict the occurrence of crime in order to draw similarities and make
assumptions as to how various techniques can be applied to code enforcement prediction.
2.2.1. Traditional GIS Based Approaches
There are various methods in GIS that can be used to analyze points and visualize
patterns. Murray et al. (2001) outlines several different spatial analysis techniques for
understanding patterns in crime data for Brisbane, Australia. These techniques include
exploratory analysis, cartographic display, and optimization-based clustering. These techniques
are largely visual in nature and require that the analyst has knowledge of the data being
displayed. Furthermore, these techniques make prediction difficult because underlying
relationships may not be displayed in the output maps. Murray et al. (2001) goes on to discuss
the various spatial statistical analysis techniques that have more predictive capabilities. Global
techniques, such as Moran’s I and box maps, depict location based relationships across an entire
study area. This can potentially give the same weight to features far from a subject point as it
does to a nearby feature. Localized techniques such as Moran’s scatter plot and local indicators
of spatial autocorrelation (LISA) place more emphasis on nearby features. These techniques
show spatial clustering and statistical significance among the data.
2.2.1.1. Hot Spot Mapping and Analysis
The most common form of crime analysis is hot spot mapping. This method can be
employed in many ways. Clustering approaches using hierarchical and partitioning techniques
allows analysts to see crime clustering in small geographic areas and group them based on pre-
specified criteria. The major disadvantage to this technique is that it is difficult to distinguish the
number of significant clusters in the data (Grubesic and Murray 2001). Studies have looked to
improve the traditional hot spot analysis.
20
The work of Liu and Brown (2003) used a newly developed density model that analyzes
transitional density as a means of hot spot detection and compared it to standard hot spot
techniques. Findings indicated that the transitional model outperformed the hot spot models in
all but one comparison. They were essentially able to improve upon hot spot detection by
incorporating density estimates over time and space and by isolating features within the data that
had the most explanatory power. This was also done by Xue and Brown (2004) where they
incorporated the assumption of criminal preference in burglary data and found that this
significantly improved on the hot spot detection techniques. These studies clearly show that
simple clustering and hot spot detection do not always depict all spatial relationships in data and
that there are better techniques available.
2.2.1.2. Kernel Density Estimation
One technique that improves upon hot spot mapping is kernel density estimation. In the
writing of Chainey et al. (2008), they used kernel density estimation in conjunction with standard
deviational ellipses, thematic maps and grid analysis to test the accuracy of hot spot analysis as a
spatial predictor of crime. The kernel density estimation out-performed all other techniques
based on predictive accuracy index. Nakaya and Yano (2010) take kernel density estimation to a
new level in crime analysis in Japan. They employed a new technique of 3D visualization in a
space-time cube where the variable of time is added to the typical kernel density estimation.
What this does is show temporary association between two known crime hot spots that may only
exist for a short period of time.
Crime prediction using kernel density estimation can also be employed using recent
spatial data from social media, such as Twitter. This media captures the geographic location of
an individual’s “tweet” along with the text within the tweet itself. A study in this phenomenon
21
was conducted by Gerber (2014) in Chicago, IL to predict the occurrence of crime. Gerber
(2014) created a training model using one month’s worth of “tweets” specifically posted for
crimes using kernel density estimation and found that the standard kernel density estimation was
significantly improved for 19 out of 25 crime types studied. He did caution that it is unclear as
to why the Twitter data enhanced the kernel density estimation, but this study still demonstrates
how kernel density can be used to predict crime using non-traditional data sources.
2.2.2. Regression Modeling Techniques
Crime has also been predicted using various regression methods. Bayesian regression
(Wheeler and Waller 2009), Tobit regression (Osgood, Finken, and McMorris 2002), and logistic
regression (Antolos et al. 2013) have all been used in the past. Tobit regression is intended for
data that is “censored,” meaning that there is an obstacle or a value limitation. Tobit regression
was actually shown to improve standard modeling techniques such as Ordinary Least Squares
(OLS) by eliminating unnecessary components of the model. Bayesian regression is used as a
hierarchical approach to correct increased coefficient variance created when using
geographically weighted regression (GWR). Both of these methods have their merits, but the
logistic regression technique is what was employed in this research project because logistic
regression is used for binary dependent/response variables (Antolos et al. 2013).
2.3 Logistic Regression Techniques
Logistic regression modeling techniques can be employed in any field where the final
outcome is dichotomous or binary. Examples include yes or no, present or not present, dead or
alive, and 0 or 1. This section looks at three studies where logistic regression was utilized to
make a prediction without employing any kind of geographical weighting factors. These studies
employ similar methodology in the fact that they use a stepwise regression model using portions
22
of a known dataset to train their model while the other portion of the known dataset is used to
validate the predictions made by the model.
The first study performed by Perestrello de Vasconcelos et al. (2001) attempted to predict
the probabilities of wildfire ignition in Portugal. They sought to compare the outcomes of a
logistic regression model and a neutral networks model. A stepwise logistic regression model
was created using several environmental variables such as topography and distance to man-made
features to find a probability of ignition value. They found that although the logistic regression
model proved to be a strong predictive tool, the neutral networks method made more significant
predictions and was more robust. This research project loosely followed the logistic regression
model techniques employed here, specifically using the logistic regression model to determine
which variables provide the most significant predictive capabilities. However, because this study
was more of a comparison between two different methodologies, this research project did not
follow it directly because it does not factor in geographical variability in the data.
The work of Ozdemir (2011) employed similar methodology as Perestrello de
Vasconcelos et al. (2001) with regard to using a stepwise logistic regression. However, Ozdemir
(2011) uses the technique as the primary function for predicting groundwater spring potential in
Turkey. Here, he uses variables such as land use, lithology, slope aspect, and elevation in raster
data to create a spring probability raster which showed the areas that are likely to support a
spring and which areas are not. The data training points in this research consisted of a set of
known control pixels within the landscape raster that contained a spring and an equal number of
pixels from the areas in the raster that are not known to contain a spring that were chosen by a
random selection tool within ArcGIS. The logistic regression model created here had an overall
predicted accuracy of approximately 95%, which indicates that it is an extremely strong model.
23
Unfortunately, the methodology used in this study cannot be used in this research project
because it used raster data to predict if a spring could or could not be present. However, as with
Perestrello de Vasconcelos et al. (2001), this research project utilized the model validation
principles of checking predicted outcomes with known control observations.
A similar study conducted by Kundu et al. (2013) also used logistic regression techniques
to predict landslide occurrence using raster data. The primary difference from the two
previously discussed works involving logistic regression was that this study used 31 explanatory
variables in the model. Simplicity becomes the issue here. In the study, the researchers used the
forward stepwise method with their logistic regression model which began by calculating the
model with no explanatory variables, followed by introducing only statistically significant
variables one at a time until all of the significant variables were in the model. Variables that
were not statistically significant at the 0.05 level were thrown out. Kundu et al. (2013) kept only
seventeen of the thirty one initial explanatory variables, which is nearly half. Even though
Kundu et al. (2013) employed the SPSS software program, their methods were not utilized for
the purposes of this research project because they used raster data, there was a very large number
of explanatory variables, and the sampling method was similar to Ozdemir (2011). However,
using the forward stepwise method of logistic regression was considered when determining
which explanatory variables were to be used within the context of this research project.
2.4 Geographically Weighted Regression Techniques
Spatial location plays an important role in regression modeling. One regression
technique that factors spatial location is geographically weighted regression (GWR). Atkinson et
al. (2003) employed GWR to model riverbank erosion in Wales using various geomorphological
variables such as streamflow, material flow, meander, history, and vegetation. They used a
24
univariate method of GWR, and explored a logistic form of GWR to fit their model to each
explanatory variable rather than fitting many variables to one model as in multivariate
regression. They did note that the Gaussian model was not used in this analysis because it only
explains spatial autocorrelation between variables. Their primary goal in this study was to
determine relationships rather than predicting outcomes, therefore, this methodology was not
utilized for the purposes of this research project. However, Atkinson et al. (2003) was useful in
understanding how GWR can be used to determine spatial relationships between explanatory
variables and that GWR can be used for other academic purposes beyond prediction.
The predictive capability of GWR was explored in Erener, Sebnem, and Düzgün (2010)
where they compared logistic regression and neutral network techniques to geographically
weighted regression techniques to show which method was the strongest predictive tool. Their
primary purpose was to improve upon the non-spatial methods of logistic and neutral network
models by incorporating spatial factors. Their study involved prediction of landslide
susceptibility in Norway and included variables such as precipitation, slope, aspect, geology, and
others. Many of these variables were raster data; therefore, the methodology of Erener, Sebnem,
and Düzgün (2010) was not used in this research project. However, their findings are what was
of interest. Even though they found that the logistic regression technique was reliable and
provided strong predictive capabilities, the GWR model proved to be stronger. Their study
yielded pseudo R squared values between 0.14 and 0.19 for the logistic regression model, and
0.54 for the GWR model. This was a profound improvement. Erener, Sebnem, and Düzgün
(2010) also note that the predictive accuracy of the GWR method was greater than that of logistic
regression. This study is a great example of how GWR improves upon traditional non-spatial
techniques, which is the reason that this research project is utilizing geographically weighted
25
logistic regression to predict if a single-family home has or does not have a code enforcement
violation.
Another study that compared logistic regression and GWR was that of Saefuddin,
Setiabudi, and Fitrianto (2012). Here, they examined the differences between a global logistic
regression model and a local geographically weighted logistic regression model in predicting the
poverty level of regions in Indonesia based on the Human Development Index. The global
logistic regression method was performed first, followed by the local geographically weighted
logistic model. For the global logistic model, the researchers note that the odds ratio is a better
way to interpret the results of the model because it is easier to interpret than the model
coefficients. This was considered in this research project when interpreting the results of the
logistic regression output for the areas that have already been proactively surveyed by code
enforcement. For the local geographically weighted logistic regression model, Saefuddin,
Setiabudi, and Fitrianto (2012) note that the weighting function and bandwidth distance
selections are highly important in the GWR model because these selections determine how the
spatial location of other instances in the data affect the instance being analyzed. They note that
the Gaussian weighing function was problematic because of the difficulty of assigning weights to
all data in the study area. They recommended using the bi-square function instead because of its
ability to remove points in the dataset where the distance between the subject point and its
weighting point is greater than the selected bandwidth distance. Essentially, the bi-square
method calculates the regression using only data points that are in the specified bandwidth
distance. This was considered for the purposes of this research project when preparing the
geographically weighting logistic regression model in the GWR4 software. Finally, Saefuddin,
Setiabudi, and Fitrianto (2012) found that GWR was a superior model to logistic regression
26
because it was able to fit data points more accurately and was the best method for their analysis
of poverty.
Sections 2.3 and 2.4 are good examples of how logistic regression and geographically
weighted regression can be combined to make a stronger predictive model. This research project
followed suit with the aforementioned studies and utilized multiple aspects of their
methodologies. Chapter 3 discusses four additional works where logistic regression and GWR
were used to make predictions for spatially occurring phenomena and how the methodology of
these studies was utilized to compose the methodology of this research project. It also discusses
the data needs and variable justification for the model that was created to predict code
enforcement violations.
27
Chapter 3 Data and Methods
This study was undertaken to determine if it is possible predict if a single-family residence (SFR)
in the City of Victorville (herein referred to as the City) is likely or not likely to have a housing
code violation in an effort to make the code enforcement process more efficient. Similar to
crime prediction, this study looks at both spatial and non-spatial factors that have the potential to
predict this outcome. Since the prediction is a binary outcome (yes or no), a non-spatial logistic
regression model was used to determine the key explanatory variables that have the greatest
influence on this outcome. Spatial factors were subsequently considered by employing
geographically weighted logistic regression using the key explanatory variables from the non-
spatial logistic model. The result was a model that was used to determine which of the 29,000
plus single-family residences in the City are likely to have a housing code violation.
This chapter explores the methodology employed to reach a prediction of either yes, a
housing code violation is likely, or no, a violation is not likely for every SFR in the City. The
methodology was based in part on studies done by Lee and Sambath (2005), Wu and Zhang
(2013), Martinez-Fernandez, Chuvieco, and Koutsias (2013), and Rodrigues, de la Riva, and
Fotheringham (2014) where both non-spatial and spatial logistic regression techniques were
utilized to determine the binary outcome of environmental factors. In addition, it explains how
the independent variables were determined, the process of the non-spatial logistic regression
technique and how its primary assumptions were met, and the process of determining the spatial
variability in this study.
Chapter 3 is composed of three sections. The first section describes the non-spatial and
spatial logistic regression techniques and how the methodology was selected for this study. The
second section discusses the data that was needed for this study and how it was acquired, and it
28
discusses justification for each of the independent variables. Finally, the third section discusses
the procedures that were followed to reach a predicted outcome for each SFR in the study area.
3.1 Research Design
An important aspect of this research project was determining which regression technique
to use in order to predict if a SFR has a housing code violation. Many of the variables that were
selected were categorical and non-continuous, eliminating the ability to do the more widely
utilized ordinary least squares (OLS) and geographically weighted regression (GWR) techniques
because with those techniques all variables must be continuous and they must be on interval or
ratio scales. Furthermore, the traditional linear regression technique of statistical prediction
could not be employed because the outcome is binary. Therefore, the method chosen for this
study was binary logistic regression because it capably handles the binary outcome, nominal and
non-continuous variables, and binary variables.
3.1.1. Non-spatial Logistic Regression Technique
The binary logistic regression (BLR) technique attempts to fit a set of data along a line of
best fit, similar to a traditional linear regression. However, this line of best fit does not follow a
straight line, but instead follows a logarithmic “S” shaped curve because the binary outcome of
yes/no or 1/0 causes points to be concentrated around two locations on the curve (see Figure 3).
Furthermore, the BLR model assumes the independent variables are non-linear functions of each
other, meaning that multicollinearity should be checked prior to beginning the regression
analysis, which is similar to the OLS model. BLR also assumes the actual dependent variable
and the independent variables do not share a linear relationship, but rather a linear relationship
between the logit of the dependent and independent variables. Other assumptions include the
fact that the dependent variable does not have to be normally distributed, the independent
29
variables do not have to be interval or ratio scale, and there needs to be a large sample size
(Anderson, 1982). The prediction equation of the BLR model uses the odds ratio and a constant
to compute the probability that a dependent variable value will fall between either 1 or 0, or yes
or no. The cut off value is typically set at 0.5, meaning that if the probability is greater than 0.5,
the value is assumed to be 1/yes and if it falls below 0.5 it is assumed to be 0/no.
Figure 3: A simple logistic regression curve
The binary logistic regression prediction equation is given in Equation 1 below, where
log(p
i
/ 1- p
i
) is the prediction percentage, α is the constant, β is the variable coefficient, and X is
the variable value.
(1)
Lee and Sambath (2005) used a logistic regression model to study landslide
susceptibility. Here they used the logistic regression modeling tools available in IBM’s SPSS
statistics software to train a model using known landslide data and multiple independent
variables to then predict which regions in their greater study area were most susceptible to
30
landslide. The initial portion of this study largely followed the methodology of Lee and Sambath
(2005) where the SPSS program was used to train a logistic regression model to be used in
predicting the likelihood of a housing code violation for 29,000 homes using a sample of 2,300
homes within the city.
3.1.2. Geographically Weighted Logistic Regression Technique
The geographically weighted logistic regression (GWLR) technique introduces a spatial
aspect to the binary logistic regression model by including the X and Y coordinate points to the
logistic regression equation. Essentially, a logistic regression is created for each instance in a
spatial dataset based on a selection of surrounding instances. A distance band, or kernel, must be
specified to determine how much influence each occurrence exerts on the others. This kernel
determines the number of surrounding data points that get factored into the regression equation
of the data point being regressed. This measure is crucial in the GWLR technique and must be
calculated carefully because it determines the degree of spatial influence in the GWLR equation.
The GWLR equation (2) is given below, where log(p
i
/ 1- p
i
) is the prediction percentage,
α is the constant, β is the variable coefficient, (u
i
,v
i
) are the coordinates of the variable, and X is
the variable value. The surrounding points within the kernel distance are weighted against the
data point being calculated so that points close to the subject point exert more influence than
points that are further away. The (u
i
,v
i
) value in the equation is essentially a second coefficient
value that expresses the geogaphic influence exerted on the variable value once the degree of
influence from the surrounding data points within the kernel is determined. The value of (u
i
,v
i
)
is the result of the weighting function applied to the location of the data point being regressed.
31
(2)
The studies of Wu and Zhang (2013), Martinez-Fernandez, Chuvieco, and Koutsias
(2013), and Rodrigues, de la Riva, and Fotheringham (2014) all utilize the GWLR technique to
model various environmental phenomena. Wu and Zhang (2013) began by using the non-spatial
logistic regression method first to train their model, then compared that to the results of the
GWLR method and found that GWLR created a more effective prediction model because it
accounts for more of the spatial heterogeneity of the occurrences. Martinez-Fernandez,
Chuvieco, and Koutsias (2013) followed similar methodology to model the occurrence of
wildfire. They first modeled the fire occurrence using linear and logistic regression, as well as
ordinary least squares before utilizing geographically weighted regression to measure the effect
of spatial relationships in their data. In this case, traditional GWR was used instead of GWLR
because their dependent variable was not binary. However, they did use GWR3 software to
compute the spatial aspect of this model because it was more robust than the tools available in
the ArcGIS software available at the time. This was a contributing factor in the selection of the
GWR4 software used for this research project because it allowed the analysis to go beyond the
ArcGIS platform where tools for more defined GWR analyses, such as geographically weighted
logistic regression, are not yet available. Finally, Rodrigues, de la Riva, and Fotheringham
(2014) followed the same methodology as the former, but instead used a random sample of
wildfire occurrences to train both their logistic regression and GWLR models.
This research project followed the techniques of these four independent studies. A non-
spatial logistic regression model was created using a sample of approximately 2,300 single-
family homes from three separate neighborhoods within the City. This model was then used to
32
determine which independent variables were statistically significant, which is a key assumption
within logistic regression. The analysis then moved to GWLR within the GWR4 software
obtained from Arizona State University to measure the spatial variability within this model and
to make housing code violation predictions for the entire City’s single-family residences. This
process is outlined in more detail in Section 3.3: Procedures and Analysis.
3.2 Data Requirements & Data Sources
Data from several sources were needed for this research project. The San Bernardino
County Assessor’s Office and the City of Victorville were the primary sources of the required
data, as well as the California Department of Alcoholic Beverage Control. Most of this data was
free to download over the Internet; however, some data from the Assessor had to be purchased
and data from the City of Victorville had to be manually input from paper reports. The cut-off
date for data collection was July 11, 2015, which is the day both the City’s code enforcement
data was collected and the date that Assessor data was downloaded. Data beyond this point in
time was not considered even though analysis of the data took place two months later.
The San Bernardino County Assessor’s Office provided much of the data in the form of
an Assessor’s parcel feature class that is available to download for free on the County’s website.
This data contained an Assessor’s parcel number (APN) that acted as the primary database key to
match a property’s attributes from other sources, and the situ address that was used to join
properties from the proactive survey. The Assessor’s Office also provided street centerlines
shapefile for free, which was used to compute two independent variables. Other Assessor data
was collected through Mimi Song Company of Ontario, California using their access to First
American Title Company’s MetroScan program which links to the Assessor’s live database.
The City of Victorville provided the proactive code enforcement data, previous code
33
enforcement data, as well as rental property business license data used to determine the type of
occupancy of the property. California’s Department of Alcoholic Beverage Control provided
location of liquor stores that was used in conjunction with street centerlines to compute distance
based variables.
3.2.1. Dependent Variable
The dependent variable in this study was the City of Victorville’s proactive code
enforcement survey. Through the City’s current proactive survey, code enforcement officers
must visit each property within a target neighborhood and note the presence of any violation
seen. These violations can range from unmaintained landscaping, presence of in-operable
vehicles, illegally parked vehicles, dilapidated building conditions, or unpermitted structures.
Officers then give the resident a written notice of the violation(s), retaining a copy of this notice
for the administrative records. At the time of data collection for this study, code enforcement
staff had surveyed three distinct neighborhoods outlined in Section 3.3.1 later.
In order to convert these reports into spatial data, each address of the cited properties had
to be logged into a spreadsheet, parsed, and recombined into a format of address number and
street name to successfully join it to the parcel feature class. This was done for all three
neighborhoods in the same manner. Properties within these neighborhoods were selected based
on the streets noted in the survey area. Selected properties in the proactive survey were given an
attribute value of “YES” while properties that did not appear in the proactive survey were given
an attribute value of “NO,” thus forming the binary dependent variable.
3.2.2. Independent Variables
The independent variables for the model were created using several different processes.
Some were obtained directly from the data source, others were obtained through simple
34
calculations using other variables, and others were obtained using network analysis. Table 1 lists
these variables.
Table 1: Independent variable list
Independent Variable
Property Size (Square Feet)
Total Property Value (Dollars)
Value Per Building Square Foot (Dollars)
Floor Area Ratio (Building Size \ Property Size)
Corporation Ownership (Yes/No)
Occupancy Type (Renter/Owner)
Number of Building Stories (1/2)
Structure Age (Years)
Length of Ownership (Days)
Tax Default Status (Yes/No)
Previous Code Case Present (Yes/No)
Number of Cases (2005 to Present)
Days Since Previous Violation (Days)
Number of Neighbor Cases
Distance to Liquor Store (Cartesian) (Feet)
Distance to Liquor Store (Network) (Feet)
3.2.2.1. County Assessor Data
Variables collected directly from the San Bernardino County Assessor’s office were the
size of the structure, the corporation ownership (yes/no), the tax default status, and the number
of building stories. Structure size, ownership type, and building stories were obtained by
downloading the data through the MetroScan program for all Victorville parcel numbers. The tax
status variable was obtained by purchasing the San Bernardino County Tax Collector’s TR345
report, which is only available on CD. The parcels contained in this report were joined to the
parcels feature class based on APN and exported to create a separate feature class.
Other variables were derived by computing County data with other County, City, or
geometric data. The total property value variable was computed by adding the land value and
35
structure value attributes provided in the parcels feature class using field calculator. The value
per building square foot was computed in Field Calculator by dividing the total property value
by the square footage of the structure. It does not include the square footage of the garage. The
property size was calculated using the Calculate Geometry option in ArcMap using the State
Plane Coordinate System CA Zone V NAD 1983 Datum. The floor area ratio was computed by
adding the structure square footage and the garage square footage together, then dividing by the
property square footage. Structure age was calculated by subtracting the “year built” field from
the Assessor data by the current year (2015), leaving the age of the structure in years.
3.2.2.2. City Data
City data are all available through a Request for Public Records document through the
City’s Records Department at no cost. Data was extracted through the City’s permit and
licensing software called “Tidemark.” To do this, Tidemark had to be queried using the MS
Query function in Microsoft Excel. Data for rental homes was extracted by using the business
license category code for rental units and legacy data for previous code enforcement cases were
extracted for the entire database, which dated back to early 2005.
Rental licenses were joined to the county parcels feature class based on APN and
exported as a separate feature class before being spatially joined to the feature class containing
the independent variables in this analysis. Parcels that were successfully spatially joined were
given a value of “renter” in the new occupancy field, and the remaining parcels were given a
value of “owner.” It is worth noting here that the rental licenses used in this variable were only
the rental homes known to the City at the time of this analysis. Unfortunately, this database does
not account for all of the rental homes in the City, but it does account for most. The actual
number of rentals in the City has not been determined.
36
The code enforcement data was used to create the previous code case present, days since
previous violation, code case count, and number of neighbor cases variables. Previous code case
present was simply a yes or no binary variable indicating that the property had at least one
housing code violation in the last ten years. If a case was successfully joined to a parcel, it was
given a value of “yes,” and a value of “no” was given to the remaining parcels. Days since
previous violation was calculated in Excel by sorting date values chronologically and removing
duplicate APN rows for older cases on the same parcel. This left the most recent case, and the
number of days between the case date and July 11, 2015 was calculated before being joined to
the rest of the variables. Code case count was also created in Excel by using the Consolidate
tool, which essentially counted the number of cases associated with each APN. Number of
neighbor cases was created by using a buffer selection of parcels containing a code case and
appending that count as a new field in the feature class containing the study variables.
3.2.2.3. Other Data Sources
The California Department of Alcoholic Beverage Control (ABC) provided the location
of commercial establishments that sell alcohol for off-site consumption. All locations within
Victorville were selected, as well as those in the surrounding cities of Hesperia, Apple Valley,
and Adelanto in order to ensure that the closest liquor store was chosen for each SFR in the city.
A simple street network was constructed using Network Analyst and the street centerlines feature
class provided by the County of San Bernardino. The network was then solved to calculate the
travel distance in feet from each SFR to the nearest liquor store location. In addition, the Near
tool in ArcGIS was used to calculate the Cartesian distance from each SFR to the nearest liquor
store location. This created the network distance and the near distance variables that were then
appended to the variables feature class.
37
3.2.3. Variable Justification
Variables in this analysis were chosen for a number of reasons. Some pertained to the
building itself, some pertained to the occupants, and others to the neighborhood of the SFR.
Structure age and total property value were natural choices because in most cities, buildings
constructed under less thorough building code standards are most likely to become dilapidated
and lose their value. The property size was selected to determine if neighborhoods that did not
have a lot of space between buildings had a higher likelihood of housing code violations while
neighborhoods with ample space between buildings were less likely. The value per square foot
and floor area ratio variables were created to normalize the total value variable and the property
size variable, respectively.
With regards to the SFR occupants, the property ownership type, occupancy type, length
of ownership, and tax default status variables were selected. Property ownership type was
chosen to provide insight into whether or not the entity who owned the building played a role in
how well it was maintained. The thought here was that if a building was owned by married
persons or single persons who lived at the residence, it would be better maintained than a
residence owned by a corporation from a different region. The occupancy type variable was
selected to determine if there was a difference in the level of building care between renters, who
often have high turnover rates, and the owners of the SFR, who often stay for years and have a
certain level of pride in their home, generally. Length of ownership sought to expand upon this
idea that ownership stability led to a high level of building care. Tax default status was chosen
because intuitively if a resident does not have money to pay their property taxes, they likely do
not have money to pay for maintenance of their residence. Code enforcement related variables
were also added to account for occupants who have a tendency of property neglect or multiple
offenses.
38
Neighborhood variables were used to determine if there was any effect on a property
based on the quality of the neighborhood it is situated in. The number of nearby code
enforcement cases was used to account for Tobler’s First Law that if a property that was
surrounded by code enforcement prone properties, it would be likely to have a code violation as
well. This was obtained by using the Average Nearest Neighbor tool in ArcGIS to determine the
average distance between SFRs’ parcel centroids in each study area, which was then used as the
selection distance for each residence to select the number of cases within that distance. This
effectively accounted for adjacent SFRs to a subject SFR, even if they were separated by a street
or alley on any side.
The distance measures to the nearest liquor store were created to introduce a crime
element as it is noted by many studies that liquor stores create crime hot spots that can have a
spill-over effect to nearby properties (Block and Block, 1995; Speer et al. 1998; Britt et al. 2005;
and Toomey et al. 2012). This creates a strain on property values, which is what Gibbons (2004)
and Linden and Rockoff (2008) show in their analysis of crime rates and property values. As
discussed in Chapter 2 of this document, lower property values increase the likelihood of
housing code violations.
3.3 Procedures & Analysis
The following discussion outlines the steps that made up this research project. Three
individual neighborhoods were analyzed first, followed by five random samples of all three, and
concluded with an analysis of the entire dataset, creating nine distinct analysis areas. A non-
spatial logistic regression was conducted for all nine, followed by a geographically weighted
logistic regression analysis. Areas 1, 2, and 3 were analyzed individually because they were
three distinct datasets collected by code enforcement staff at different times. They also consisted
39
of different sample sizes, 581 homes, 316 homes, and 1,301 homes respectively. Sample size is
an important factor in logistic regression, which was tested here. Five random samples of Areas
1, 2, and 3 were analyzed to determine if single-family homes were consistent across the three
areas or if their neighborhoods played an important role in their variability. Finally, all homes in
the three study areas were analyzed to compare how the model performed to the other, smaller,
samples of data.
3.3.1. Study Areas
The study area for this research project was the City of Victorville in the Mojave Desert
of Southern California (Figure 1). Within the City, three distinct neighborhoods were analyzed.
Area 1 (shown in Figure 4) on the north end of the City consisted of the Cypress Point housing
tract which was approximately 60 percent built out at the time of this study. The neighborhood
consisted of moderate grade construction homes roughly two to eight years in age on parcels
ranging in size from 6,000 to 10,000 square feet on average. The neighborhood has a
homeowners association, so residents have a fair amount of resident oversight when it comes to
the aesthetics of the neighborhood. Area 2 (shown in Appendix A) on the east end of the City
consisted of two separate neighborhoods that were in close proximity to each other. One
neighborhood had low-grade construction homes that were approximately ten to fifteen years
old, while the other neighborhood had moderate to upper end construction homes that ranged in
age from brand new to fifteen years old. Area 3 (shown in Appendix A) was on the west end of
the City and consisted of low to moderate construction grade homes approximately ten to fifteen
years old. Each area had unique characteristics and residents of varying economic status.
40
Figure 4: Area 1 neighborhood and observed violations
3.3.2. Individual Analysis of the Three Study Areas
Variables were first examined for multicollinearity in order to not violate a key
assumption of logistic regression that collinear variables are not present in the model. This was
done by performing a linear regression using only the independent variables. In IBM’s SPSS
22.0 software, the linear regression tool was used for this purpose. The first independent
41
variable was placed in the dependent variable dialog box with the remainder of the independent
variables placed in the independent variable dialog box. The test was iterated by removing the
most recent independent variable from the dependent variable dialog box and swapping it with
one from the independent variable dialog box until all independent variables were tested. The
test produced variance inflation factor, or VIF, values explaining how each of the variables was
related to the others. A VIF above 3.0 indicated that there was some degree of multicollinearity,
and a value above 5.0 indicated that multicollinearity is present. For this research project,
variables with a VIF above 3.0 were removed from the analysis.
A non-spatial logistic regression was then performed on each study area individually.
Area 3 was done first because it had the greatest number of homes surveyed, and since one of the
assumptions of logistic regressions is a large sample size, this was a logical starting point. The
dependent variable and independent variables were uploaded to the SPSS software program and
the logistic regression was performed using the “enter” method, which utilizes all of the
independent variables in the model. After the initial iteration of the model, variables that were
not statistically significant predictors at the 0.05 level according to SPSS were removed before
running the model a second time. Statistical significance is one of the key assumptions of the
logistic regression model and is used to determine the “goodness-of-fit” of the model. Chapter 4
explains these outputs and what they mean for the model.
Area 1 and Area 2 were then analyzed in the same manner as Area 3 where
multicollinearity was first tested using simple linear regression, followed by the “enter” method
for logistic regression.
3.3.3. Combined Area Analysis
The three areas were then combined into a single dataset and five random samples of 500
42
homes were taken using the random sample generation tool within SPSS. For each random
sample, the same procedure was followed in the individual area analysis; non-significant
variables were removed and multicollinearity was addressed.
A logistic regression was then performed on the entire dataset to compare how it
performed to the individual analyses of Areas 1, 2, and 3, as well as on the five random samples.
Again, this procedure followed the same steps as the previous analyses where multicollinearity
was addressed using a linear regression of the variables, followed by the logistic regression
analysis.
3.3.4. Geographically Weighted Logistic Regression Analysis
The role of geography was tested in this analysis by using the GWR4 software program
to perform a geographically weighted logistic regression (GWLR) analysis. This was performed
in the same manner as the non-spatial logistic regression analysis, where Areas 1, 2, and 3 were
analyzed separately, followed by the five random samples of the data and concluding with the
dataset as a whole. However, since GWR4 cannot handle nominal scale data, dummy variables
had to be created for the dependent variable as well as for each categorical independent variable
that passed the multicollinearity tests. Table 2 gives a breakdown of these dummy variables.
43
Table 2: Dummy variables for GWR4
Within GWR4, the model was set to Logistic (binary), the latitude and longitude
coordinates for each single-family home’s parcel centroid were input and the spherical option
was selected. Next, the independent variables that remained at the end of the non-spatial logistic
regression model for each area were input as local variables. The kernel type was set to Adaptive
Gaussian using the nearest neighbor method and the bandwidth selection method was set to the
golden selection method where the GWR4 software determines the optimum bandwidth measure
for each dataset. This was done because the incremental spatial autocorrelation and Moran’s I
tools in ArcGIS did not produce any peak values for any of the datasets, meaning there is no
distance where spatial relationships are most pronounced, according to these ArcGIS tools.
Finally, the selection criteria was set to the Akaike Information Criterion (AIC) because the
GWR4 manual states that this is the most suitable method when using a Gaussian kernel type.
These parameters were set for each iteration of the GWLR analysis over the study areas.
SPSS Variable
Value
GWR4 Variable
Value
Owner 1
Renter 0
Yes 1
No 0
Yes 1
No 0
Yes 1
No 0
Yes 1
No 0
Occupancy Type
Default Status
Previous Code Case Present
Corporation Ownership
Proactive (Dependent)
44
Once each area was run through the GWR4 software, the output table containing variable
coefficients and predictions was analyzed for accuracy to compare against the non-spatial
logistic regression model. This was done by comparing the y column in the output table, which
contained the actual observed dependent variable value for each case, with the yhat column
which contained the predicted probability value for each case as calculated by GWR4. Yhat
values below 0.5 were re-valued at 0 and values above 0.5 were re-valued at 1. If y and yhat
matched, a new value was calculated at 0 meaning that the model correctly predicted the
dependent variable. If the y and yhat did not match, a new value of 1 was calculated indicating
that the model did not correctly predict the dependent variable. The total number of incorrectly
predicted cases was subtracted from the total number of observed cases, and the value was then
divided by the total number of cases to arrive at a prediction accuracy percentage number. The
prediction accuracy percentage number was then compared to the same number calculated in the
logistic regression model found in SPSS.
3.3.5. Model Validation and Making Predictions
The ultimate goal of this research project was to predict which houses in the City had a
likelihood of having a housing code violation. To make these predictions, the binary logistic
regression equation was calculated for each house using the constant value and the variable
coefficient values from the SPSS logistic analysis for the best performing model, which was the
combined dataset of Areas 1, 2, and 3. Predicted probabilities were calculated using a logit
transformation process because the logistic regression equation makes predictions using the log
odds scale. First, the log (short for logarithm) odds value for each single-family home was
calculated using the logistic regression equation where the constant was added to the products of
the variable value and its corresponding coefficient (see Equation 1). Next, the log odds values
45
were multiplied by the exponential value (e), which is the inverse of a logarithm, so it
mathematically converts the log odds (which is the logarithm of odds) to simple odds by
cancelling the logarithm. Then, the odds values were converted to probabilities by dividing the
odds value by 1 plus the odds value. This calculation determined the percentage of likelihood
that a house had a violation (Simon 2013). Prediction percentages that were above 0.50 were
given a value of “yes/1” and those below 0.50 were given a value of “no/0.” The results were
then mapped by joining the new dataset back to the parcels feature class for visual analysis.
The model validation process method was straight forward. Code enforcement continued
to conduct their proactive survey in new areas of the City. One such area consisted of 376
homes in the Brentwood neighborhood near the geographic center of the City. The predicted
values of the logistic regression model were cross-referenced with this area’s actual survey data
to determine how well the model actually performed. An overall accuracy percentage was
determined by dividing the total number of correct predictions by the total number of actual
proactive enforcement citations.
46
Chapter 4 Results and Predicted Violations
This chapter analyzes the outputs of both the non-spatial logistic regression models and the
geographically weighted logistic regression (GWLR) models. Each iteration of the analysis over
Areas 1, 2 and 3, as well as the five random samples and the combined areas produced differing
statistics of varying strength and predictive capability. Section 4.1 looks specifically at the
logistic regression model, how multicollinearity was minimized, the predictive capability of the
models, and each models’ statistical significance. Section 4.2 describes how each iteration of the
model was affected by incorporating geographic variability using the GWR4 software and
whether or not the logistic regression model was improved with GWLR. Section 4.3 discusses
the reasons why the final prediction model was chosen, and Section 4.4 discusses the model
validation using observed data from field inspections that were collected after the initial data
collection of this research project. It also provides predictions for the 29,000 single-family
homes in Victorville that were not part of the model training.
4.1 Binary Logistic Regression Results
Results of the logistic regression models were analyzed using UCLA’s Institute for
Digital Research and Education (IDRE) Annotated SPSS Output Logistic Regression resource.
This is a free online resource center where annotated statistical interpretation instructions can be
found for several different kinds of statistical outputs, including logistic regression. The
following section discusses the key statistics in the SPSS output results, including the Wald chi-
squared test used to determine if the constant is statistically significant from zero, the score and
significance test used to determine if an independent variable was a good predictor, the test for
overall model significance given by the chi-square statistics and its p-value, and the pseudo R
squared values used to interpret model goodness-of-fit. The overall percentage value from the
47
Step1 Classification table is of importance because it gives the percentage of cases that were
correctly classified by the logistic regression model. Furthermore, the Wald and significance test
were analyzed to determine if the coefficient values for each variable were statistically
significant from zero, allowing the null hypothesis to be rejected. The output also contains the
coefficient values for each variable and the constant value that were used to create the logistic
regression prediction equation. An example of the SPSS output and which statistics were of
importance for the purposes of this research project can be found in Appendix B.
4.1.1. Interpretation of Results
This section discusses in detail how the results of the logistic regression output of SPSS
in the context of the Area 1 values for reference. Each table shown in this section contains the
portions of the logistic regression output that were of importance as previously explained. The
tables are followed by a discussion on what these outputs mean, how they equate to the strength
of the model and the relationships between the dependent and independent variables.
Interpretation began with the multicollinearity test of the variables. These tests showed
that the code case, total value, and near distance variables had VIF values greater than 3 meaning
that multicollinearity was present. Code case was removed because it provided the least amount
of information compared to the case count, days to previous violation, and nearby cases variables
that it was collinear with. Total value was removed because it was not normalized to the
structure size where value per square foot was normalized. Near distance was removed because
the network distance variable is a more accurate representation of reality. Residents cannot pass
through walls or structures and would likely travel in a vehicle to a liquor store. Once these
variables were removed, the linear regression test of the independent variables showed little
multicollinearity with VIF values all below 3. Table 3 shows the final iteration of the
48
multicollinearity test for Area 1 while the remainder of the multicollinearity tests for the other
iterations of the model can be found in Appendix C.
Table 3: Area 1 multicollinearity test
VARIABLE VIF VALUE
Property Size (LOTSQFT) 1.519
Floor Area Ratio (FLOOR_AREA) 2.547
Number of Building Stories (NOSTORY_DUM) 1.908
Length of Ownership (LENGTH_OWN) 1.245
Structure Age (STRUCT_AGE) 1.019
Value Per Square Foot (VALUE_PSF) 1.306
Days to Previous Violation (DAYS_TO_VI) 2.593
Number of Cases 2005 to Present (CASE_COUNT) 2.664
Tax Default Status (TAX_STATUS_DUM) 1.040
Occupancy Type (OCCUPANCY_DUM) 1.185
Network Distance to Liquor Store (NETWORK_DI) 1.169
Number of Nearby Cases (NEARBY_CAS) 1.306
The variables that did not show multicollinearity were then input into the logistic
regression model for Area 1 which yielded non-significant independent variables in the form of
structure age, value per square foot, tax status, occupancy, and corporation owned. Removal of
these variables did not affect the statistical significance of the predictive power of the other
variables because each one is analyzed separately within SPSS. Once non-significant variables
were removed, the remaining variables were statistically significant at the 0.05 level, upholding
the assumption that the logistic regression model is fit using only significant variables. Table 4
shows the variable significance table before the non-significant variables were removed and
Table 5 shows the variables that remained after the non-significant variables were removed.
49
Table 4: Area 1 variable selection with non-significant variables
VARIABLE SIGNIFICANCE
Property Size (LOTSQFT) .022
Floor Area Ratio (FLOOR_AREA) .000
Number of Building Stories (NOSTORY_DUM) .000
Length of Ownership (LENGTH_OWN) .009
Structure Age (STRUCT_AGE) .208
Value Per Square Foot (VALUE_PSF) .698
Days to Previous Violation (DAYS_TO_VI) .000
Number of Cases 2005 to Present (CASE_COUNT) .000
Tax Default Status (TAX_STATUS_DUM) .208
Occupancy Type (OCCUPANCY_DUM) .934
Network Distance to Liquor Store (NETWORK_DI) .000
Number of Nearby Cases (NEARBY_CAS) .000
Corporation Owned (CORP_OWNED) .504
Table 5: Area 1 remaining variables once non-significant variables were removed
VARIABLE SIGNIFICANCE
Property Size (LOTSQFT) .022
Floor Area Ratio (FLOOR_AREA) .000
Number of Building Stories (NOSTORY_DUM) .000
Length of Ownership (LENGTH_OWN) .009
Days to Previous Violation (DAYS_TO_VI) .000
Number of Cases 2005 to Present (CASE_COUNT) .000
Network Distance to Liquor Store (NETWORK_DI) .000
Number of Nearby Cases (NEARBY_CAS) .000
The null model was then analyzed by looking at the p-value of the constant and the
overall percentage value. As Table 6 shows, the p-value of the constant (or intercept) was .000
meaning that it was statistically significant and the null hypothesis could be rejected.
Table 6: The null model significance test
B Standard Error Wald Deg. Freedom Sig. Exp(B)
Constant -1.119 .096 135.078 1 .000 .326
In SPSS, the null model (which excludes independent variables) predicts all values to be
“no.” The overall percentage value in the output is the number of predicted “no” values that were
50
correct in the dependent variable data. This essentially shows how many “no” values and how
many “yes” values are in the training dataset, which in the case of Area 1 was 75.4% “no”
values. The percentage was calculated by dividing the number of correct “no” values, 438, by
the total cases in the dataset, 581. The remaining 143 cases had a value of “yes” because the null
model predicted them incorrectly. Table 7 below, which is the output from SPSS, shows the
total “no” values, the total “yes” values, and the overall accuracy percentage of the null model.
Essentially, if independent variables were not included in the logistic regression equation, the
model would be able to accurately predict 75.4% of the cases in the dataset. This percentage
value became the benchmark against which the regression model was evaluated once the
variables were input for this particular model.
Table 7: The null model predictions
Predicted
Observed NO YES Percentage Correct
PROACTIVE NO 438 0 100.0%
PROACTIVE YES 143 0 0.0%
Overall Percentage 75.4%
The next output analyzed was the model significance test, which is indicated by the
Omnibus Tests of Model Coefficients table as shown in Table 8. Here, the model entry at the
bottom of the table indicates the statistical significance of the model. The chi-square value of
154.096 and p-value of .000 indicate that this is a significant model because the significance
threshold is the .05 level. The degrees of freedom column was not used to interpret the logistic
regression model using the Enter method, as is only of value in a stepped model in SPSS.
51
Table 8: Area 1 omnibus model significance test
Chi-Square Deg. Freedom Sig.
Step 154.096 8 .000
Block 154.096 8 .000
Model 154.096 8 .000
The pseudo R squared values in the Model Summary Table were then analyzed. These R
squared values cannot be interpreted in the same manner as ordinary least squares (OLS) because
they are not calculated the same. The Cox and Snell R squared value is always calculated to be
on a scale from 0 to 0.75, and the Nagelkerke R squared value calculates an adjustment to the
Cox & Snell R squared to place it on a range from 0 to 1 in order to give it the appearance of an
OLS R squared value, which is easier to interpret. IDRE advises careful interpretation of these
values, but the intent is to show how much of the variation in the dependent variable is explained
by the independent variables. According to Clark and Hosking (1986), a pseudo R squared value
above 0.20 for Cox & Snell indicates a model is acceptable. These values are shown in Table 9.
Table 9: Area 1 pseudo R squared values
Step
-2 Log
likelihood
Cox & Snell
R Square
Nagelkerke R
Square
1 494.347 .233 .346
The Variables in the Equation output table shows the model coefficients (which are on
the log odds scale), the Wald and significance test of the coefficients, and the odds ratio. These
statistics are primarily used for interpreting how each independent variable affects the model in
terms of positive and negative correlations. For example, if the odds ratio of the variable is less
than 1, an increase in the value of that variable will cause a decrease in the odds of the event
occurring, which is an inverse relationship. If the odds ratio is greater than 1, this indicates that
an increase in the value of the variable will cause the odds of the dependent variable occurring to
52
increase, which is a direct relationship. If the odds ratio is exactly 1, there is no relationship
between the dependent variable and the corresponding dependent variable. In SPSS, the software
automatically rounds the odds ratio to three significant figures. The option to increase the
number of significant figures is available. In this research project, the number of significant
figures was increased to six. This caused odds ratios that were exactly 1.000 at three significant
figures to become either slightly less than 1 or slightly more than 1, indicating that there is a very
slight relationship between the dependent and independent variables in all iterations of the
model.
For the floor area variable, the odds ratio is .019 meaning that if the floor area is
increased, the odds of having a code enforcement violation decrease in Area 1. Since the Wald
test is significant at the 0.05 level (the p-value is .025), the null hypothesis that the coefficient is
equal to zero can be rejected. For the lot size, length of ownership, days to violation, and travel
distance variables, the odds ratio was 1.000 when rounded to three decimal places. The output
table was modified to show odds ratios to six decimal places, revealing a very small value for
each of these variables. This indicated that there was a very slight inverse or direct relationship
to the dependent variable, depending on the value of the odds ratio. The odds ratios and
significance tests for the remaining variables are shown in Table 10.
It is worth noting here that although interpreting the odds ratios is interesting and
provides much information about the model, it is not a vital factor in making predictions, which
is the purpose of this research project. The primary focus for the Variables in the Equation
output table was on the variable coefficient values and the value of the constant, both shown in
the column labeled “B.” These values are used in forming the logistic regression equation to
make predictions. Interpretation of a .000 coefficient value is given in Chapter 5 as it relates to
53
the scale of the units of measurement of the variable.
Table 10: The coefficient and odds ratio output table with variables in the equation
Variable B
Std.
Error
Wald d.f. Sig. Exp(B)
Lot Size (LOTSQFT) .000 .000 .931 1 .335 1.000
Floor Area (FLOOR_AREA) -3.978 1.772 5.040 1 .025 .019
Number of Stories
(NOSTORY)
-.140 .326 .185 1 .667 .869
Length of Ownership
(LENGTH_OWN)
.000 .000 4.281 1 .039 1.000
Days to Previous Violation
(DAYS_TO_VI)
.000 .000 22.152 1 .000 1.000
Number of Previous Cases
(CASE_COUNT)
.147 .115 1.645 1 .200 1.158
Travel Distance to Liquor Store
(NETWORK_DIST)
.000 .000 20.211 1 .000 1.000
Number of Nearby Cases
(NEARBY_CASE)
.059 .033 3.304 1 .069 1.061
Constant 2.396 1.242 3.718 1 .054 10.976
The final output table is the Block 1 Classification Table shown in Table 11. This table
provides the most important piece of information in the output. This table shows how well the
model performed with the variables included. Again, for Area 1, the null model with no
variables included was able to correctly predict 75.4% of the cases. With the variables included,
the model was able to correctly predict 84.3% of the cases. This is an increase of nearly 10%,
meaning that this is a good predictive model and performed well against the null model. The cut
value of .500 simply indicates that if the predicted probability of a case was below that cut value,
it was given a value of “no” and if the predictive probability was above that cut value, it was
given a value of “yes.” If more certainty was to be given to the predicted probabilities, this value
could be increased to .600, which would in turn cause borderline predictions around a value of
.500 to be given a value of “no” in the model and only the stronger predicted probabilities (those
54
above .600) would be given a value of “yes.”
Table 11: The overall number of cases predicted correctly for Area 1
Predicted
Observed NO YES Percentage Correct
PROACTIVE NO 410 28 93.6%
PROACTIVE YES 63 80 55.9%
Overall Percentage 84.3%
4.1.2. Logistic Regression Model Results
This section contains the results of the logistic regression models for Areas 1, 2, and 3,
the five random samples, and the combined areas. The summary tables contained in this section
show the most interesting results that pertain to the interpretation of the logistic regression
models. Each table is followed by a discussion on the key differences between the various
iterations of the model. Chapter 5 discusses what these results mean and what findings can be
made from them about the data. To begin, Table 12 shows the key statistics of the iterations of
the logistic regression model used to compare and make conclusions about the strength of each
model.
Table 12: The results of the model iterations and key statistics
Model
Null
Model
Constant
Sig.
Null
Model
Accuracy
Chi-
square
Value
Sig.
Cox &
Snell R
Squared
Nagelkerke
R Squared
Prediction
Model
Accuracy
Accuracy
Difference
Model
Assessment
Area 1 -1.119 .000 75.4 154.096 .000 0.233 0.346 84.3 8.9 Good
Area 2 -0.371 .001 59.2 124.989 .000 0.327 0.441 76.9 17.7 Very Good
Area 3 -0.261 .000 56.5 274.345 .000 0.190 0.255 71.1 14.6 Poor
Sample 1 -0.464 .000 61.4 99.047 .000 0.180 0.244 72.6 11.2 Poor
Sample 2 -0.456 .000 61.2 126.419 .000 0.223 0.303 76.2 15.0 Good
Sample 3 -0.456 .000 61.2 111.749 .000 0.200 0.272 73.2 12.0 Poor
Sample 4 -0.405 .000 60.0 119.855 .000 0.213 0.288 75.2 15.2 Good
Sample 5 -0.498 .000 62.2 133.114 .000 0.234 0.318 77.4 15.2 Good
Combined -0.484 .000 61.9 506.853 .000 0.206 0.280 74.9 13.0 Good
55
In all of the models, the null model constant value was negative and they each had a
significance value of 0.000, so the null hypothesis that the model’s constant was equal to zero
could be rejected at the 0.01 level. Area 1 had the highest null model accuracy percentage and
was approximately 15% greater than the other models. In every model except the combined
areas, the chi-squared value was fairly low and every model’s chi-squared significance value was
0.000 indicating that these were all statistically significant models. The pseudo R squared values
varied with Random Sample 1 having the lowest value and Area 2 having the highest value.
The model assessment column in Table 12 indicates if the resulting model for each area
was good, very good, or poor. This assessment was based on the combination of the model’s
pseudo R squared values, the significance value of the omnibus test of model coefficients which
is indicated by the chi-squared value and corresponding significance value, and the change in the
predicted accuracy from the null model to the prediction model. To be classified as a good
model, the omnibus test had to yield a low chi-squared value with a significance value less than
0.05, the pseudo R squared values had to be greater than 0.20, and there had to be an increase in
the prediction accuracy from the null model to the prediction model.
The accuracy of the prediction model was the most important piece of information to take
from this table. Area 1 produced the highest overall value but had the lowest accuracy difference
over its corresponding null model. However, this was a good model because the pseudo R
squared values were above 0.20 and the overall accuracy percentage was high. The best model
came in Area 2 because it had the highest pseudo R squared values and the greatest change in
accuracy over the null model. Area 3 along with Random Samples 2 and 3 were the weakest
models because the pseudo R squared values were at or below the 0.20 threshold and the change
56
in accuracy percentages were low. The remaining models all had pseudo R squared values above
0.20 and the change in accuracy was 15% or higher, meaning that these were acceptable models.
4.1.3. Logistic Regression Coefficients
This section contains a brief discussion of dependent and independent variable
relationships and what it means for the corresponding regression equation coefficients. Table 13
contains the results of the coefficient calculations in SPSS, which are on the log odds scale, the
converted odds ratio used for interpretation, and the significance value of the odds ratio. The
null hypothesis here is that the coefficient value is equal to zero and the statistical significance
level to reject this null hypothesis was the 0.05 level. As previously explained, the odds ratio
determines the likelihood of the dependent variable occurring with a change in the independent
variable in either the positive or negative direction.
Table 13: Results of the coefficient calculations, the odds ratio, and significance for the logistic
regression models
Variable Name and (Alias)
Coefficients & Relationships
Variable
Coefficient
Variable Odds Ratio
- Exp(B)
Odds Ratio
Significance
AREA 1
Lot Size (LOTSQFT)
0.000 1.000 0.335
Floor Area (FLOOR_AREA) -3.978 0.019 0.025
Number of Stories (NOSTORY) -0.140 0.869 0.667
Length of Ownership
(LENGTH_OWN)
0.000 0.039 1.000
Value per Sq Ft (VALUE_PSF)
0.001 1.001 0.928
Days to Previous Violation
(DAYS_TO_VI)
0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
0.147 0.200 1.158
Travel Distance to Liquor Store
(NETWORK_DIST)
0.000 1.000 0.000
Number of Nearby Cases
(NEARBY_CASE)
0.059 1.061 0.069
Constant
2.396 10.976 0.054
57
Variable Name and (Alias)
Coefficients & Relationships
Variable
Coefficient
Variable Odds Ratio
- Exp(B)
Odds Ratio
Significance
AREA 2
Value per Sq Ft (VALUE_PSF) 0.001 1.001 0.928
Days to Previous Violation
(DAYS_TO_VI)
-0.001 0.999 0.000
Number of Previous Cases
(CASE_COUNT)
0.067 1.069 0.695
Number of Nearby Cases
(NEARBY_CASE)
0.001 1.001 0.973
Constant
1.261 3.527 0.154
AREA 3
Length of Ownership
(LENGTH_OWN)
0.000 1.000 0.109
Days to Previous Violation
(DAYS_TO_VI)
-0.001 0.999 0.000
Number of Previous Cases
(CASE_COUNT)
-0.250 0.779 0.033
Occupancy (OCCUPANCY)
0.071 1.074 0.668
Cartesian Distance to Liquor Store
(NEAR_DIST)
0.000 1.000 0.201
Travel Distance to Liquor Store
(NETWORK_DIST)
0.000 1.000 0.015
Number of Nearby Cases
(NEARBY_CASE)
0.034 1.034 0.140
Constant 2.114 8.284 0.000
RANDOM SAMPLE 1
Days to Previous Violation
(DAYS_TO_VI)
0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
-0.016 0.984 0.919
Number of Nearby Cases
(NEARBY_CASE)
0.060 1.062 0.035
Constant 0.351 1.421 0.369
RANDOM SAMPLE 2
Number of Stories (NOSTORY)
0.839 2.315 0.000
Days to Previous Violation
(DAYS_TO_VI)
-0.001 0.999 0.000
Number of Previous Cases
(CASE_COUNT)
-0.168 0.846 0.210
Cartesian Distance to Liquor Store
(NEAR_DIST)
0.000 1.000 0.926
58
Variable Name and (Alias)
Coefficients & Relationships
Variable
Coefficient
Variable Odds Ratio
- Exp(B)
Odds Ratio
Significance
Travel Distance to Liquor Store
(NETWORK_DIST)
0.000 1.000 0.058
Number of Nearby Cases
(NEARBY_CASE)
0.016 1.016 0.637
Constant 1.225 3.405 0.034
RANDOM SAMPLE 3
Floor Area (FLOOR_AREA)
-2.158 0.116 0.094
Number of Stories (NOSTORY) 0.399 1.491 0.112
Structure Age (STRUCT_AGE)
0.007 1.007 0.359
Days to Previous Violation
(DAYS_TO_VI)
0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
-0.091 0.913 0.446
Cartesian Distance to Liquor Store
(NEAR_DIST)
0.000 1.000 0.685
Travel Distance to Liquor Store
(NETWORK_DIST)
0.000 1.000 0.009
Number of Nearby Cases
(NEARBY_CASE)
0.019 1.019 0.554
Constant
2.119 8.324 0.011
RANDOM SAMPLE 4
Number of Stories (NOSTORY) 0.598 1.818 0.008
Value per Sq Ft (VALUE_PSF)
-0.017 0.983 0.005
Days to Previous Violation
(DAYS_TO_VI)
0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
0.012 1.012 0.938
Occupancy (OCCUPANCY)
-0.168 0.845 0.510
Travel Distance to Liquor Store
(NETWORK_DIST)
0.000 1.000 0.013
Number of Nearby Cases
(NEARBY_CASE)
0.061 1.063 0.039
Constant
2.111 8.254 0.004
RANDOM SAMPLE 5
Floor Area (FLOOR_AREA) -0.479 0.620 0.727
Number of Stories (NOSTORY)
0.602 1.826 0.025
Days to Previous Violation
(DAYS_TO_VI)
-0.001 0.999 0.000
Number of Previous Cases
(CASE_COUNT)
-0.158 0.854 0.320
Occupancy (OCCUPANCY)
-0.170 0.844 0.518
59
Variable Name and (Alias)
Coefficients & Relationships
Variable
Coefficient
Variable Odds Ratio
- Exp(B)
Odds Ratio
Significance
Cartesian Distance to Liquor Store
(NEAR_DIST)
0.000 1.000 0.864
Travel Distance to Liquor Store
(NETWORK_DIST)
0.000 1.000 0.101
Number of Nearby Cases
(NEARBY_CASE)
0.055 1.056 0.086
Constant
1.414 4.111 0.110
COMBINED AREAS
Floor Area (FLOOR_AREA)
-0.623 0.536 0.309
Number of Stories (NOSTORY) 0.437 1.548 0.000
Days to Previous Violation
(DAYS_TO_VI)
0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
-0.088 0.916 0.203
Cartesian Distance to Liquor Store
(NEAR_DIST)
0.000 1.000 0.659
Travel Distance to Liquor Store
(NETWORK_DIST)
0.000 1.000 0.000
Number of Nearby Cases
(NEARBY_CASE)
0.022 1.022 0.143
Constant 1.602 4.961 0.000
The most important thing to note here is a variable with a coefficient of 0.000 does not
impact the outcome of the logistic regression equation. It causes the associated variable to drop
out of the equation because the product of the coefficient value and the variable value are zero.
Since the logistic regression equation is the addition of the constant value with the products of
the coefficient and variable values, adding a zero value does not change the outcome because any
number when added to zero is still the same. That being said, many of the variables in these
regression models could not reject the null hypothesis of the coefficient equaling zero. This
detracts from the overall strength of the model, which is discussed later in Chapter 5.
In all iterations of the model, the days to previous violation variable was a significant
predictor and had a statistically significant odds ratio at or very close to 1.000. This means that
60
there was an issue with this variable. The number of significant digits of the output tables was
increased from 3 to 6, which revealed that the coefficients of this variable did in fact have a
value, albeit extremely small. So here, the null hypothesis that the variable’s coefficient is zero
can be rejected even though the odds ratio shows no positive or negative relationship. In
addition, the null hypothesis for the constant in each model could only be rejected for the
combined areas, Area 3, and Random Samples 2, 3 and 4. Also note that in four of the six
instances where the number of building stories was a significant predictor, it also showed a
significant relationship between the independent and dependent variables.
One will also notice the absence of the corporation owned yes/no and tax status variables
in all iterations of the model, meaning that these variables contributed nothing to the logistic
regression model training and showed no relationships between the dependent and independent
variables. Occupancy only appeared twice and structure age appeared only once in the model
iterations, which is discussed further in Chapter 5.
4.2 GWR 4 Results
The following section discusses the results of performing the geographically weighted
logistic regression analysis using the GWR4 software. Areas 1, 2, and 3, the five random
samples, and the combined areas were input into the GWR4 software using only the remaining
variables from the logistic regression models from SPSS to determine if there was any effect of
location on the data. For all but one of the model iterations, the local geographical weighting
model improved upon the global model for goodness of fit to some extent. Furthermore, the
geographical weighting was able to increase the model prediction percentage, although, many of
these increases were small at approximately one percentage point.
61
The next section describes how the results of the GWR4 software were interpreted in the
context of Area 1’s data for a reference. Interpretations were based off of the GWR4 user
manual, which explains how to perform the model, what the different tool selections mean,
suggestions on which kernel types to use based on the model type, and what each of the output
statistics means. This manual is attached to the GWR4 download file from Arizona State
University.
4.2.1. Interpretation of GWR4 Results
The GWR4 software first conducts a “global” regression analysis where local variation in
the data is not accounted for. This is similar to, but not the same as the logistic regression
calculations of SPSS; the SPSS outputs give more information and were used to removed non-
significant predictors from the model. The global regression of GWR4 is not important for
making predictions, but is still worth noting because the percent of deviance explained value was
used to compare the global and local regression models of GWR4. Also, GWR4 produces an
Akaike Information Criterion (AIC) value that is used to compare the fit with different models.
Essentially, the smaller the AIC, the better the model fit. For Area 1, the percent deviance
explained value was 0.23 and the AIC was 503.116. Percent of deviance explained can be
interpreted in the same manner as the pseudo R squared values of SPSS. The next output was
the optimal bandwidth distance as computed by the golden bandwidth selection search method
of GWR4. Chapter 3 explained why the golden bandwidth selection method was chosen to
determine the optimal bandwidth measure, but it essentially determines where the spatial
variability is most pronounced in the data. This bandwidth is given as the number of nearest
neighbors and not in a distance. This is because the kernel type is adaptive, meaning the actual
distance between the analyzed cases to the furthest point in the kernel will vary. The bandwidth
62
value is the number of cases that fall within the adaptive kernel that produces the optimal amount
of influence on the case being analyzed, according to GWR4. For Area 1, this was the 488
nearest neighbors out of a total of 581 different cases.
The next output of GWR4 was the local regression model where spatial relationships are
considered. The percent of deviance explained in Area 1 when spatial variation was accounted
for was 0.27, which is an increase over the global model. It also produced coefficient values
based on several different aggregation methods, which are not important for this research project.
The y and yhat comparison for Area 1 produced an overall correct prediction percentage of
84.9%, which is a very slight increase from the logistic regression percentage of 84.3%. Here,
the local model out-performed the global model by increasing the percent of deviance explained
value, but the overall accuracy did not increase significantly. Figure 5 shows the results of the
GWR4 global model calculated by the software, Figure 6 shows the optimum kernel bandwidth,
and Figure 7 shows the results of the local model. A full GWR4 output and an example of the
listwise table showing the y and yhat values can be found in Appendix D.
63
Figure 5: Global regression result for Area 1 in GWR4
Figure 6: Area 1 bandwidth selection
64
Figure 7: Local model output for Area 1
4.2.2. GWR4 Model Results
This section contains the results of the GWR4 model iterations. Table 14 summarizes
these results beginning with the global regression model’s AIC value and deviance explained
value. That is then followed by the bandwidth distances selected by the golden bandwidth
selection search of GWR4 and the corresponding total number of cases in each model. It also
shows the outputs of the local regression model, including the AIC value and percentage of
deviance explained when spatial relationships are considered. The remaining portion of the table
shows the accuracy comparisons between the logistic regression models of SPSS and those of the
GWR4 software.
65
Table 14: GWR4 model iteration summary table containing important statistics
Each iteration of the GWR4 model decreased the AIC value of the global model in the
local model with the exception of Area 2 where it was increased by 1. This indicates that the
GWLR model fit better than the non-spatial logistic regression model everywhere except for
Area 2. Note too that the percent of deviance explained was much lower for the GWR4 software
than in the SPSS pseudo R squared values. This emphasizes the importance of using the SPSS
software for performing the non-spatial regression calculations because it proved to be more
accurate in terms of the percentage of variation explained in the model.
The optimum bandwidth selections were rather large compared to the total number of
cases in each model. This is a strong indicator of the absence of spatial relationships in the data
because these bandwidths encompass most, if not all of the cases in each study area; for each
random sample, the bandwidth was equal to the number of cases. This means that the optimum
bandwidth calculation, which is designed to pick up the point in the data where the spatial
relationships are most pronounced, failed to find any sort of peak relationship point.
Furthermore, the data sampling method employed for this research project may have been the
cause of the large bandwidth sizes due to a lack of organization in the collection of the observed
violation data. Unfortunately, this could not be controlled because the data was collected by City
AIC
% Deviance
Explained
Total Cases
in Model
Optimum
Bandwidth
AIC
% Deviance
Explained
Logistic
Model
GWLR
Model
Diff.
Area 1 512.3469 0.23 581 488 502.0703 0.28 84.3% 84.9% 0.6%
Area 2 312.3737 0.29 316 216 313.3339 0.33 76.9% 80.1% 3.2%
Area 3 1523.20885 0.15 1301 954 1513.80266 0.17 71.1% 72.2% 1.1%
RS 1 575.878 0.14 500 500 559.1855 0.19 72.6% 73.2% 0.6%
RS 2 555.4258 0.18 500 500 540.761 0.23 76.2% 75.6% -0.6%
RS 3 574.0514 0.17 500 500 564.7624 0.21 73.2% 74.6% 1.4%
RS 4 569.1571 0.18 500 500 550.7589 0.24 76.0% 75.6% -0.4%
RS 5 547.9625 0.20 500 500 535.4414 0.26 77.4% 77.8% 0.4%
Combined 2431.0602 0.17 2198 897 2336.4292 0.23 74.9% 76.3% 1.4%
MODEL
Global Model Bandwidth Local Model Accuracy
66
code enforcement officers before the time of this research project. This is discussed further in
Chapter 5 in the future work discussion.
The percent of deviance explained by the local model was increased over the global
model in every iteration. However, this did not equate to a higher prediction accuracy for every
iteration. For Random Samples 2 and 4, the accuracy of the predictions actually decreased. The
rest of the accuracy percentages were approximately positive 1%, with the highest being in Area
2 at 3.2%. These increases are not very pronounced, and the implications of this are explained in
Chapter 5 as well.
Spatial variation was also observed between the three study areas when analyzing the
coefficient values of each variable in the GWLR output of the Combined Areas. For each study
area, there was very low spatial variability among the data within each area, and each area
differed from the other two. For example, the floor area variable had all data in Area 1 less than
-0.50 standard deviations from the mean. In Area 2, all of the floor area data was in the -0.50 to
0.50 standard deviation range. Area 3 was the only study area to show variability in its data
largely because of its larger sample size. Here, data points in the northwest section were less
than -0.50 standard deviations, data points in the middle of the study area were between 0.50 and
1.7 standard deviations, and data points near the eastern edges of the study area were near the
mean in the -0.50 to 0.50 standard deviation range. The absence of spatial variability in the data
within each study area was likely caused by the large bandwidth size, which is discussed further
in Chapter 5 along with the implications of the lack of variability in the GWLR output. Figure 8
below depicts the floor area coefficient analysis of the GWLR output for the Combined Areas.
The remaining variable coefficient maps from the Combined Areas can be found in Appendix E.
67
Figure 8: Floor area variable coefficient values of the GWLR analysis showing low variability.
4.3 Selection of the Prediction Model
The model iterations from logistic regression and geographically weighted logistic
regression ranged from very good to poor when considering pseudo R squared values, percent of
deviance explained, and overall prediction accuracy. This discussion explains the reasoning for
why the final model used to make predictions for the entire City of Victorville was selected.
Of the non-spatial logistic regression models, Area 1 performed the best in terms of
number of correctly predicted violations while Area 2 performed the best in terms of the amount
of variation explained as shown in the pseudo R squared values of the model. These two models
68
would have been the top candidates had it not been for the neighborhood variation issue, which
is explained in Chapter 5. Area 3 produced too weak of a model for consideration. All of the
random samples were disregarded because the fluctuations in the significant predictor variables
was too great, and the fluctuations in the strength of each sample’s model would make
predictions highly volatile.
The geographically weighted regression models were not considered for similar reasons.
Area 1 and especially Area 2 were able to increase the accuracy of predictions when spatial
relationships were considered, however, each area on its own did not encompass enough
neighborhood variability to make good predictions for the entire city. Area 3 was also able to
increase the accuracy, but its high AIC value and low percentage of deviance explained values
were too low to be considered. The random samples were too inconsistent to be considered,
especially since two of the samples actually reduced the accuracy of the model predictions. The
combined areas in the spatial model produced an extremely high AIC value meaning that the
model was not a good fit when spatial relationships were considered.
Therefore, the combined areas dataset of the non-spatial logistic regression models was
the top candidate to make predictions across the entire city because it encompassed the most
neighborhood variation within the data. It also had acceptable pseudo R squared values and it
had three variables that had statistically significant variable relationships. Furthermore, the
increase in the prediction accuracy over its null model was acceptable, and it contained the
largest sample size. As a reminder, a large sample size is a primary assumption of the logistic
regression technique.
69
4.4 Predictions
This final section of Chapter 4 looks at the predictions made by the combined areas non-
spatial logistic regression model that was selected as explained in the previous section. The first
part of this section discusses the predicted outcomes of the logistic regression equation for the
single-family homes that were not part of the model training. The second section discusses how
the model was validated by calculating the accuracy of the predictions with observed violations
from field inspections that occurred after the data collection date of this research project, which
was July 11, 2015.
4.4.1. Prediction of Violations
Following the logistic regression equation calculation and the log odds transformation
process outlined in Section 3.3.5, predictions were made for all of the single-family homes in the
City. Figure 9 shows a very small section of the City at the intersection of Bear Valley Rd and
Amethyst Rd where there is a high density of homes. The red dots on the map indicate a housing
code violation is more than 50% likely while a blue dot indicates that a violation is less than 50%
likely. According to the model, 7,483 homes are likely to have a code violation. This is
approximately 26% of the 29,000 single-family homes in the City. As a reference, if the number
of predicted violations was divided evenly among the seven code enforcement officers in
Victorville, they would each have to take on over 1,000 individual cases, which is a substantial
workload.
70
Figure 9: Predicted housing code violations based on logistic regression
71
Figure 10 shows an area of the City where there is a clear contrast between the
predictions in four neighborhoods. Neighborhoods 1 and 2 outlined in red show a high density
of probable housing code violations while Neighborhoods 3 and 4 show a very low density of
probable housing code violations. This illustrates how neighborhoods can possibly be selected
for proactive enforcement due to high likelihood of violations versus neighborhoods that can be
overlooked in proactive code enforcement due to a very low likelihood of housing code
violations.
Figure 10 also shows that the predictions can potentially depict differences among
neighborhoods. Neighborhood 4 outlined in blue is a new housing tract that was built in the last
10 years. Neighborhood 3 is a tract of single story homes on very small lots, with a requirement
that only seniors can be the primary resident. Neighborhoods 1 and 2 were both built in the mid-
1990s and are likely beginning to show signs of decay. This appears to come across in the
predictions, and is explained further in Chapter 5.
72
Figure 10: Neighborhood comparison of predictions
73
4.4.2. Model Validation
The neighborhood selected for model validation is shown in Figure 11. Again, it
consisted of 376 single-family homes in the Brentwood neighborhood in the center of the City.
The overall observed accuracy of the model was calculated in exactly the same manner as it was
in the training model by comparing the number of correctly predicted “no” violations with the
number of observed “no” violations, as well as the number of predicted “yes” violations with the
number of observed “yes” violations. These were then totaled and compared against the total
number of homes in the neighborhood. Table 15 shows the outcome of this calculation. The
combined area training model calculated a 74.9% overall accuracy in SPSS. The observed
accuracy from comparing the predicted with the observed indicated a 50.3% overall accuracy.
This was calculated by adding the total number of correct “no” predictions and the total correct
“yes” predictions and dividing that by the total number of cases, so (125 + 64) / 376. The
implications of this outcome are discussed in detail in Chapter 5.
Table 15: Observed predicted accuracy of the model
TOTAL NO YES % Correct
NO 231 125 106 54.1%
YES 145 81 64 44.1%
TOTAL 376 50.3%
Predicted Observed
74
Figure 11: Model validation neighborhood
75
Chapter 5 Discussion and Conclusions
The results of the analysis proved to be quite interesting and many of the outcomes were
unexpected. This chapter outlines the major findings from the non-spatial and spatial regression
models and answers the research questions outlined in Chapter 1. In addition, this chapter
discusses how this research project contributes to the work previously completed using a
combined analysis methodology of logistic regression and geographically weighted regression as
well as using regression modeling as a tool for making predictions for spatially occurring
incidents, such as crime. Furthermore, this chapter discusses what the major limitations were in
this research project and what work can be done in the future to improve upon its findings.
5.1 Findings
This section discusses the major findings of this research project. It illustrates what was
of significance in the variable selection for each model, what relationships were found between
the dependent and independent variables, what relationships could be found spatially using the
GWLR technique, and the implications of the model predictions. It also discusses how each
iteration of the model compared to the others and why the observed percentage of correct
variables did not match up with the training model.
5.1.1. Non-spatial Findings
The multicollinearity tests revealed that the code case yes/no, lot sq ft, and total value
variables showed collinearity in nearly all of the models. This outcome was actually expected
because the code case yes/no variable simply showed if a code violation had been present in the
past which is expressed in the days to previous violation and number of previous cases variables.
Note here that the days to previous violation and case count variables did not show collinearity
76
because they were on different scales of measurement; days to previous violation was a length of
time and case count was a tally, meaning they tell different stories. The total value variable was
normalized by dividing it by the square footage of the home, which was the likely cause of the
collinearity between it and the value per sq ft variable. These were both dollar measurements of
the value, and once total value was removed for the reasons expressed in Chapter 4, the VIF
value of the value per sq ft variable dropped to approximately 1.3 which means there was little to
no collinearity with other variables. Finally, the lot square footage variable expressed
collinearity with the floor area variable because the floor area variable was the building square
footage divided by the lot square footage, meaning that the lot square footage information was
also expressed in the floor area variable. Again, once the lot square footage variable was
removed, the VIF of floor area dropped to below 2.0 indicating little multicollinearity. The
exception to this was in Areas 1 and 2 where the near distance and network distance variables
showed collinearity with each other. This was likely because Area 1 and 2 were the smaller
sample sizes in terms of geographic area covered, so the near and network distances were
similar. Area 3, the combined areas, and the random samples encompassed larger geographic
areas giving these distances more variability. Near distance was removed as explained in
Chapter 4 and the VIF value of network distance dropped below 2.0 in both Areas 1 and 2.
The nine iterations of the non-spatial logistic regression model produced results that were
unique to each sample of data. As Chapter 4 illustrated, each training model had a different set
of significant predictor variables. For example, Area 1 retained lot size, floor area, number of
stories, length of ownership, value per sq ft, days to previous violation, previous case count,
network distance, and number of nearby cases while Area 2 only retained value per sq ft, days to
previous violation, previous case count, and number of nearby cases as significant predictors.
77
This was likely caused by variability in the data between each of the study areas. As explained
in Chapter 3, each study area where proactive enforcement inspections were conducted had
different neighborhood characteristics. Area 1 had a homeowner’s association, Area 2 had a
very low income neighborhood and a high income neighborhood, and Area 3 was somewhere in
the middle of these. The neighborhood characteristics, which were not accounted for in this
research project, likely caused the differences in which predictors were significant for each area.
Furthermore, variables that were initially thought to be good predictors of code violations
were not significant predictors at all. For example, the corporation owned variable was not a
significant predictor in any of the models. The initial thought was that if a property was owned
by a corporation, it was likely to be rented which would create occupancy turn over and lead to a
higher likelihood of a violation. According to the logistic regression models, this was
resoundingly not the case. In addition, structure age only appeared as a significant predictor in
one of the models. The initial thought with this variable is that the structure age and the
likelihood of a violation would be a direct relationship, meaning that as age increases, the chance
of a violation also increases because the older homes must be maintained and if the occupant
does not make these repairs, code violations appear. These models indicate that the structure age
is not a significant determinant of code violations. Another variable that was thought to be a
significant predictor was the tax default variable, which did not appear in any of the regression
models. Here, the reasoning was that if an owner is defaulting on their property taxes, they are
not maintaining their property because they cannot afford it. This was somewhat touched upon in
Chapter 2 in the discussion that if there is a financial burden on an owner, caused by too much
code enforcement or otherwise, the property is likely to not be maintained. This relationship
may be more complicated than a simple binary variable and is worth exploring further.
78
The independent and dependent variable relationships also varied between the different
iterations of the model. The most interesting item came with the days to violation variable where
in most of the models, the coefficient value was near zero, the odds ratio was near 1.000, but the
odds ratio was significant at the 0.01 level. It was discovered that the coefficient did not equal
zero once the number of significant digits was increased from 3 to 6 in the output tables. These
coefficient values were extremely small (0.000487 for example) but they still played a role in
determining the outcome of the dependent variable. This was also the case with the network
distance and length of ownership variables when they were significant predictors in the models.
It was also found that the number of building stories had a positive relationship in the models
where it was a significant predictor except for in Area 1 where it was a negative relationship, and
in Random Sample 3 where the significance value was greater than 0.05 and the null hypothesis
could not be rejected. This means that when the number of stories value increases from 0 (single
story) to 1 (two story), the likelihood of having a code violation increases. Again, it was not a
significant predictor in all the models, but this was still an interesting finding because the number
of stories was not initially thought to have a significant relationship with code violations.
The variation in the strength of the different models was also interesting. In Area 2
where the sample size was smallest, the pseudo R squared values and model accuracy increase
were the highest. Area 2 also had the second fewest number of significant predictors in the
model at four. Random Sample 1 was the weakest model in terms of pseudo R squared values,
but it performed better than Area 1 in terms of increasing the model’s prediction accuracy. The
five random samples showed some consistency in their null model accuracy with all percentages
at approximately 61%. However, this was not reflected in the prediction model accuracy as two
random samples performed lower than the other three. Also note here that the pseudo R squared
79
values vary from 0.18 to 0.234, meaning that there is inconsistency in model strength. This is an
indication that there is possible variability in the neighborhoods where the data was collected;
otherwise the random samples would have been more consistent in terms of model accuracy and
strength because each sample would have had data that was similar.
In addition, neighborhood variability would have been expected to have appeared in the
results of the GWLR variable coefficient maps because that is what the GWLR analysis method
is intended to depict; the differences from one data point to the next. This was not the case in this
research project as the coefficient maps only showed variability in Area 3 of the Combined
Areas dataset used to make the predictions. Areas 1 and 2 showed little to no variability among
the data. This was likely due to the design of this research project in both the sampling of the
observed code violation data and in the bandwidth selection method. A true random sample of
data across the entire city would have likely showed the variability in the data more accurately
because there would have been data points that encompassed more of the distinct neighborhoods
within the city. Furthermore, determining a bandwidth measure that encompassed only the
subject SFR and properties immediately adjacent to it would have likely depicted more of the
variability in the data within each study area. The optimum bandwidth selection tool in the
GWR4 software was not able to find the spatial variation that is likely present in this data.
Again, a true random sample of data points encompassing the entire city in conjunction with a
more precise bandwidth selection method would likely give a large boost to the predictive
capabilities of the GWLR analysis.
Research question #1 of this project was, “can certain property attributes predict the
occurrence of code enforcement violations?” The answer to this is complicated and varies
spatially, but generally the answer is yes. Some variables, such as number of stories or days to
80
previous violation, were always significant predictors. Others variables, such as structure age or
corporation owned yes/no, were rarely found to be significant predictors, if at all. The general
results of this analysis show there are in fact certain property characteristics that can predict the
occurrence of housing code violations to some degree. However, the accuracy of these predictor
variables is not strong enough to initiate code enforcement administrative action on because they
do not account for all variations in the City’s data and should be considered with great caution.
5.1.2. Spatial Findings
The geographically weighted logistic regression (GWLR) models showed that there was
almost no spatial variability among the dependent variable and the independent variables. This
was clear in the optimum bandwidth selection tool in the GWR4 program. This tool selects the
bandwidth measure where the spatial variability is most pronounced in the data, which is not un-
like the incremental spatial autocorrelation tool found in ArcGIS. In all of the random samples
of the data, the optimum bandwidth was equal to the total size of the sample at 500. In Area 1,
the bandwidth was 488 out of 581 homes, 216 out of 316 for Area 2, 954 out of 1,301 for Area 3,
and 897 out of 2,198 for the combined areas. These large bandwidth sizes indicate that there is
very little that can be explained geographically in the model. The initial thought in using GWLR
was that houses that had code violations would have a “spill-over” effect on to nearby homes,
similar to the effects of crime discussed in the writings of Block and Block (1995), Speer et al.
(1998), Britt et al. (2005), and Toomey et al. (2012) as mentioned in Chapter 3. According to the
GWR4 program calculations, this is not the case for code enforcement violations.
Existence of neighborhood variation was confirmed when the coefficient values of the
combined areas, which was used to make the predictions, were mapped. For each variable
coefficient, the results show that each neighborhood varies from the other two, but there is very
81
little variation inside the confines of each neighborhood area. The largest area, Area 3, was the
only one that showed some degree or internal spatial variation in the coefficient values. Figure
12 below shows the days to violation coefficient values mapped across the three study areas.
Notice that all points inside Area 1 fall between 0.5 and 1.2 standard deviations from the mean
while all points in Area 2 are greater than -1.5 standard deviations from the mean. Most of Area
3 falls between -0.5 and 0.5 standard deviations with a few points falling in other ranges, though
still within one standard deviation from the mean. So again, the model could not find any spatial
variation among the single-family homes in the dataset. The remainder of the coefficient maps
showing very similar outcomes can be found in Appendix E.
Figure 12: Map of the Days to Violation variable coefficient values from the GWLR analysis
82
The presence of variation in the neighborhoods is also present in the inability of the five
random samples of the training data to make consistent predictions. As the results of the GWLR
analysis show, two of the random samples of the data actually caused the model to decrease in
prediction accuracy. In addition, the percent of deviance explained value was rather low
compared to Area 1 and Area 2. Also, in the logistic regression model, the pseudo R squared
values were not consistent and two of the samples had values that were at or below the
acceptable threshold value of 0.20. So even though these samples are using the same dataset,
there are differences in the data that are causing skewed results. This is likely because each
random sample potentially has more samples from one neighborhood over another, and because
the three study areas are so different, having a higher percentage of samples from one or the
other is causing the data change. Section 5.3 explains how this phenomena can be reduced by
employing a better data sampling method.
Research question 2 asked if the occurrence of a code enforcement violation is partially
the result of the effect of neighboring properties, and the answer to this is no. The large optimum
bandwidth values combined with the lack of spatial variability in the prediction model
coefficient values shows that a spatial relationship between the dependent variable and the
independent variables does not exist. However, as discussed earlier, using an alternative
bandwidth selection method may reveal spatial relationships among neighboring properties. This
is discussed in Section 5.3.
5.1.3. Predictions
The primary goal of this research project was the prediction of housing code violations
throughout the City of Victorville. This was accomplished; however, the first thing to note here
is the 50.3% accuracy when the model was validated using a proactive inspection area that was
83
collected after the initial data collection of this project. This is of course much lower than the
74.9% accuracy rate of the model training data. This result unfortunately suggests that the model
was not a good predictor of housing code violations at the single-family home level. This is
likely due to the weak relationships between the independent variables and the dependent
variable found in the non-spatial logistic regression analysis of the odds ratios and their
significance values.
However, this model has the potential to be able to predict housing code violations at the
neighborhood level. A careful study of the four neighborhoods within the City of Victorville
showed that homes in brand new housing tracts were largely predicted as having no violations.
This is quite intuitive because these new homes are built to current code and have not had time to
go unmaintained. In addition, neighborhoods that are regarded as “bad” areas of the City were
predicted to have a higher density of possible violations, which is also intuitive and follows suit
with the work of Meier (1983). In Meier (1983), he discussed how lower class neighborhoods
have higher occurrence of code enforcement activity in Pasadena, CA. Areas in Victorville that
are known to be lower class (in terms of quality) were predicted to have high occurrences of
code violations. Though this could not be confirmed in the model validation process, the results
of the predictions could be used in other methods to identify which neighborhoods are likely to
have violations and which are not. This is discussed in Section 5.3.2.
This model was able to predict, to some degree, which houses in the City of Victorville
currently have housing code violations, which answers research question 3. Of course, the
previous discussions warrant extreme caution when acting upon these predictions due to low
percentages of accuracy and weak relationships between the dependent and independent
variables. This is also due to the fact that the accuracy was only tested on a small neighborhood
84
in the City, so neighborhoods with different characteristics could exhibit different degrees of
prediction accuracy.
5.2 Relation to Previous Work
The results of this research project followed suit with many of the studies discussed in
Chapter 2 of this document. The work of Des Rosiers et al. (2002) discussed how high quality
landscaping can increase the value of a property. For new housing tracts in the City,
development code standards require that a high quality landscape be installed prior to the home
being permitted for occupancy by building inspectors. The model of this project predicted that
these new homes are not likely to have code violations, and a contributing factor to this is a
quality landscape, which leads to a high initial property value. Also, Luttik (2000) discussed
how greenspace has a positive effect on surrounding property value. The model predicted that
the homes immediately adjacent to the City’s golf course had a low likelihood of housing code
violations, which could be the result of higher home values according to Luttik (2000) and Meier
(1983). Again Meier (1983) suggested that high property values lead to low code enforcement
activity because residents are able to maintain their properties on their own.
This research project is also a contribution to the GIS based approaches to crime
prediction. The binary dependent variable of code enforcement violation present yes/no could
easily be changed to burglary present yes/no (Antolos et al. 2013) or simply crime present
yes/no. The work of Murray et al. (2001) discussed many examples of how crime can be
predicted using various spatial and statistical techniques such as Moran’s I analysis or LISA
analysis. Now, a logistic regression model could easily be included in this list because it is also
employing methodology using spatial autocorrelation in the form of the GWLR technique.
Crime was also modeled using logistic regression and GWR in the work of Wheeler and Waller
85
(2009), and Antolos et al. (2013). Wheeler and Waller (2009) were showing how GWR
techniques can be improved on by using Bayesian regression and Antolos et al. (2013) used
logistic regression to model the occurrence of burglary. This research project used ideas from
both of these studies where the logistic regression method was initially used and was then
improved upon by using a more refined regression technique, GWR in this case. Though the
results of this research project did not coincide with the work of Antolos et al. (2013), the results
do show that it is possible to model code violations in a manner similar to crime modeling.
This research project also adds another field that can be studied with logistic regression
modeling. Section 2.3 of Chapter 2 outlined several studies where logistic regression was
employed, including wildfire ignition (Perestrello de Vasconcelos et al. 2001), groundwater
spring potential (Ozdemir 2011), and landslide susceptibility (Kundu et al. 2013). In each of
these previous studies, the logistic regression technique was used to train a dataset to make
predictions over a study area; this is exactly what was done in this research project. These
previous studies were conducted for phenomena that occurred in the physical sciences. This
research project adds a social science aspect to the realm of possible subjects that can be studied
using the logistic regression technique.
Section 2.4 of Chapter 2 outlined several studies that employed geographically weighted
regression techniques to make predictions of spatially varying phenomena, and some of them
even went so far as to compare the GWR technique to the logistic regression technique. In
Erener, Sebnem, and Düzgün (2010) and in Saefuddin, Setiabudi, and Fitrianto (2012), the
researchers were able to improve the predictive capability of non-spatial regression models with
the use of GWR. In this research project, it was found that there were very slight increases in the
predictive accuracy of the GWLR model over its corresponding logistic regression model.
86
Future work on code enforcement prediction using the same methodology but stronger predictive
variables could produce results that align with Erener, Sebnem, and Düzgün (2010) and
Saefuddin, Setiabudi, and Fitrianto (2012) in terms of model improvement.
Finally, this research project also contributes to the work of Wu and Zhang (2013),
Martinez-Fernandez, Chuvieco, and Koutsias (2013), and Rodrigues, de la Riva, and
Fotheringham (2014) in that it gives another example where logistic and geographic regression
modeling can be successful in making predictions over space. This research project was able to
combine the methodologies in the aforementioned studies to produce a model that was able to
make predictions with some degree of success. Even though these predictions were only 50%
accurate at the parcel level when compared to a small sampling of observed code violations used
for validation, it appeared to be successful in identifying neighborhood level clusters of
violations that could potentially be addressed by code enforcement staff in the proactive
inspection program. However, these neighborhood level clusters of violations were not field
verified in this research project, and further investigation into the validity of identifying
neighborhoods using parcel level predictions within logistic regression is necessary before this
claim can be fully supported.
5.3 Limitations and Future Work
There were several observed limitations in this study. The results all indicate that there
are likely several key variables that were not accounted for in this regression model. Also, the
results show that there is little to no spatial variation between the dependent variable and its
explanatory variables as indicated in the results of GWR4. This section addresses these
limitations and what work can be done to improve these results.
87
5.3.1. Major Limitations
The most noticeable issue in this research project was the low pseudo R squared values in
the non-spatial models and the low percentage of deviance explained values in the spatial
models. As explained earlier, these values mean that there is much more to the story when it
comes to predicting housing code violations in a city. This was one of the key contributing
factors to the low observed accuracy of the predictions. The variables that were used in this
research project were mostly related to the physical condition of the house, such as floor area,
structure age, and assessed value. These variables did not include information on the
neighborhood that these homes were in. Also, there were not enough variables that explained
demographics or socio-economic condition, which play very important roles in determining the
quality of a neighborhood and whether or not there are probable housing code violations.
Furthermore, variables did not explain enough about the residents of each SFR, which will likely
be a difficult explanatory variable to produce at the parcel level due to laws limiting data sharing
from government agencies, including the US Census Bureau. Interpolations of US Census tract
or block level data to arrive at parcel level data would also introduce the modifiable areal unit
problem into the model. Also, conducting the analysis at the Census tract or block level in order
to capture more demographic and economic data for the analysis would further emphasize the
modifiable areal unit problem.
Another key limitation of this research project was the lack of significant relationships as
explained by the odds ratio of the SPSS logistic regression outputs. As explained earlier, even
though variables were determined to be statistically significant predictors, there were not many
statistically significant relationships between the dependent variable and the explanatory
variables. Many of the coefficient values were very close to zero and their converted odds ratios
were extremely close or at 1, meaning that there was no significant relationship present. This
88
diminishes that overall quality of the model, which is a vital part of making accurate predictions.
Limitations were also noticed in the collection of the data. Code enforcement officers
exercise a great deal of discretion when determining if a home has a housing code violation.
Each officer’s interpretation of the code could be different from the others, meaning that there
could be data discrepancies based on which officers performed the initial proactive inspections.
This actually relates to the work of Ross (1996) and Burby et al. (2000) where they discussed
government discretion and how it varies between government officials, interpretation of codes,
and how this can cause violations to be overlooked. This potentially could create data quality
issues with the model. This was not specifically analyzed in this research project, but it would
be worth studying.
Having three sample areas that were miles apart was also a limitation in this analysis.
Because of this, it was difficult to produce a model that encompassed more diversity among
homes and neighborhoods so that these factors could be accounted for in the predictions. It
would have been more accurate to take a random sample of the City in its entirety rather than
random samples of data within the three study areas. This would have been able to capture a
more accurate depiction of the variability in the City’s SFRs. The City contains many diverse
neighborhoods and having data from only three of these neighborhoods produced low accuracy
when tested against a neighborhood that had different characteristics than the sample areas. The
fact that the City has so much variability in its housing and that codes can be interpreted
differently by different officers or inspectors makes predicting violations a very difficult task,
and the results of this research project are a good indicator of that.
Finally, training the model using clusters of data points instead of across the entire study
area could potentially bias the prediction model. For example, if there are high code violation
89
counts that come from the clusters of training data, the model will have a hard time
differentiating data points where there is a low likelihood of a violation because the model does
not have any examples of this. It would be the same kind of bias if the training data had a low
occurrence of code violations as it would not be able to pin-point SFRs with a high likelihood of
a violation with any ease. The fact that logistic regression uses a binary dependent variable
makes minimizing this model bias difficult because the logistic model cannot determine the
degree or severity of the violation. This bias could potentially be reduced by either having a
good sampling of data points that encompasses neighborhoods with high violations and
neighborhoods with low violations, or by making the dependent variable nominal or ordinal so
the model can have other training information to be used in predictions. Also, changing the
dependent variable to nominal or ordinal would require a different regression technique, meaning
that the binary logistic technique of this research project could not be used.
5.3.2. Future Work
There is a great deal of room for future work with this research project. The first area
would be to find better prediction variables. As mentioned earlier, variables that incorporate
demographic or socio-economic factors such as household income or a normalization of
household income against the assessed property value would be useful. As Meier (1983)
discovered in his study, areas with higher household income levels saw a much lower need for
code enforcement action because property owners were financially able to maintain their
properties as required by city regulations. Another variable that should be incorporated would be
some sort of neighborhood variable. There needs to be something that can describe the state of a
neighborhood to make stronger predictions and account for differences among neighborhoods.
These variables could consider crime rates, dwelling unit density, ethnic groups, age of residents,
90
single parent households, or educational achievement.
Other methods of variable collection could also be employed. Remote sensing could
potentially be used to capture SFRs where the front yard has gone from green grass to brown or
dead grass using infrared sensors and image classification software. This would show SFRs that
violate the live landscaping requirement of the City. Aerial photography could also be used to
show SFRs with the presence of inoperable vehicles. For example, if an aerial photograph shows
a vehicle in the driveway or on the street in a February flight, and that vehicle appears in a July
flight and has not moved, there is a good chance that the vehicle does not run; Code says
vehicles that do not run must not be visible from the street. Aerial photography could also be
used to locate SFRs that have trash and debris that must be abated using basic visual scans of
photographed areas. Using remote sensing would be a time consuming and expensive process,
but it could yield information that is vital to predicting if other SFRs are likely to have the same
type of code violations.
Hot spot analysis of the code violation predictions could be useful in strengthening the
claim that the model created in this research project was potentially able to identify
neighborhoods where there is a higher likelihood of violations. So instead of visually scanning
each neighborhood in the prediction data to determine if there is a high or low occurrence of
violations, the hot spot analysis could identify these areas more quickly. Furthermore, the hot
spot analysis could be used to identify neighborhoods with the highest intensity of violations, as
well as neighborhoods with the lowest intensity. This would help City officials to allocate
resources to the most intense hot spots first, followed by less intense hotspots. This could
produce quick changes in the neighborhood characteristics of these areas, and it would show
elected officials of the City that the proactive enforcement program does have some degree of
91
impact, hopefully in the positive direction.
Future work should also include more sound sampling methods to produce more
consistent datasets. Having three clusters of data in three very different neighborhoods likely
caused issues in the project. If data collection practices were more systematic, say by using only
one officer to do all of the inspections, and if more data was collected from different
neighborhoods, such as the half acre lot neighborhoods which are much lower in dwelling
density, the model could potentially recognize more of the neighborhood variability in the data
and make stronger predictions. Also, adding more areas to verify the accuracy of the model
should increase the 50.3% observed accuracy number.
5.3.3. Conclusions
Thus, it can be seen that even though this project was not able to produce highly accurate
predictions of housing code violations, there were achievements in determining how well
selected variables performed in making predictions and in determining how much of a role basic
geography principles played in this phenomena. Furthermore, the prediction model generated
using the regression techniques could potentially be used to identify neighborhoods that have a
high likelihood of violations and others that have a low likelihood of violations. This is based on
observed knowledge of these neighborhoods and how the predictions seem to correlate to the
neighborhood characteristics. Also, identifying which variables expressed multicollinearity with
other variables will aid in the replication of this research project because there will not be such a
strong effort on data collection.
In replication of this research project by other cities, the advice would be to obtain
stronger predictor variables, collect a random sample data from the entire city instead of three
clusters of SFRs in order to better explain the role of geography, and to use more validation
92
areas.. Stronger predictor variables could make the relationship between the dependent and
independent variables stronger. Also, collecting a random sample of data points across the entire
city area would make the kernel bandwidth in the GWLR analysis more equipped to depict the
spatial relationships in the data. Finally, providing more validation areas could potentially
increase the model’s prediction accuracy to a point where it could be trusted. Doing so could
make this prediction model more accurate in order to one day be able to make code enforcement
departments in any city more effective.
93
REFERENCES
Anderson, J. A. 1982. Logistic regression. Handbook of Statistics. North-Holland, New York:
169-191.
Anderson, Linda M., and H. Ken Cordell. 1988. Influence of trees on residential property values
in Athens, Georgia (USA): a survey based on actual sales prices. Landscape and Urban
Planning 15 (1): 153-164.
Antolos, D., D. Liu, A. Ludu, and D. Vincenzi. 2013. Burglary crime analysis using logistic
regression. Human Interface and the Management of Information. Information and
Interaction for Learning, Culture, Collaboration and Business, Lecture Notes in
Computer Science 8018: 549-558.
Atkinson, P. M., S. E. German, D. A. Sear, and M. J. Clark. 2003. Exploring the relations
between riverbank erosion and geomorphological controls using geographically weighted
logistic regression. Geographical Analysis 35 (1): 58-82.
Block, R. L., and C. R. Block. 1995. Space, place and crime: hot spot areas and hot places of
liquor-related crime. Crime and Place 4 (2): 145-184.
Britt, H. R., B. P. Carlin, T. L. Toomey, and A. C. Wagenaar. 2005. Neighborhood level spatial
analysis of the relationship between alcohol outlet density and criminal
violence. Environmental and Ecological Statistics 12 (4): 411-426.
Burby, R. J., P. J. May, and R. C. Paterson. 1998. Improving compliance with regulations:
choices and outcomes for local government. Journal of the American planning
association 64 (3): 324-334.
Carlton, R. E. 1965. Enforcement of municipal housing codes. Harvard law review 78 (4): 801-
860.
Chainey, S., L. Tompson, and S. Uhlig. 2008. The utility of hot spot mapping for predicting
spatial patterns of crime. Security Journal (19) 1-2: 4-28
Clark, W. A., and P. L. Hosking. 1986. Statistical methods for geographers. Wiley, New York.
Conway, D., C. Q. Li, J. Wolch, C. Kahle, and M. Jerrett. 2008. A spatial autocorrelation
approach for examining the effects of urban greenspace on residential property values.
Journal of Real Estate, Finance, and Economics (2010) 41: 150-169.
Des Rosiers, F., M. Theriault, Y. Kestens, and P. Villenevue. 2002. Landscaping and housing
values: an empirical investigation. The Journal of Real Estate Research 23 (1): 139-161.
Dombrow, J., and M. Rodriguez. 2000. The market value of mature trees in single-family
housing markets. The Appraisal Journal 68 (1): 39-43.
94
Erener, A., H. Sebnem, B. Düzgün. 2010. Improvement of statistical landslide susceptibility
mapping by using spatial and global regression methods in the case of More and Romsdal
(Norway). Landslides 7: 55-68.
Gerber, M. S. 2014. Predicting crime using Twitter and kernel density estimation. Decision
Support Systems (61): 115-125.
Gibbons, S. 2004. The costs of urban property crime. The Economic Journal 114 (499): F441-
F463.
Gribetz, J., and F. P. Grad. 1966. Housing code enforcement: sanctions and remedies. Columbia
law review 66 (7): 1254-1290.
Grubesic, T. H., and A. T. Murray. 2001. Detecting hot spots using cluster analysis and GIS.
Proceedings from the Fifth Annual International Crime Mapping Research Conference
26.
Kundu, S., A. K. Saha, D. C. Sharma, and C. C. Pant. 2013. Remote sensing and GIS based
landslide susceptibility assessment using binary logistic regression model: a case study in
Ganeshganga watershed, Himalayas. Journal of the Indian Society of Remote Sensing 41
(3): 697-709.
Lee, S., and t. Sambath. 2006. Landslide susceptibility mapping in the Damrei Romel area,
Cambodia using frequency ratio and logistic regression models. Environmental Geology
50: 847-855.
Linden, L., and J. E. Rockoff. 2008. Estimates of the impact of crime risk on property values
from Megan's laws. The American Economic Review: 1103-1127.
Liu, H., and D. E. Brown. 2003. Criminal incident prediction using a point-pattern-based density
model. International Journal of Forecasting 19: 603-622.
Luo, J., N. K. Kanala. 2008. Modeling urban growth with geographically weighted multinomial
logistic regression. Geoinformatics 7144: 7144M-1 – M-11.
Luttik, J. 2000. The value of trees, water and open space as reflected by house prices in the
Netherlands. Landscape and Urban Planning 48 (3): 161-167.
Martinez-Fernandez, J., E. Chuvieco, and N. Koutsias. 2013. Modelling long-term fire
occurrence factors in Spain by accounting for local variations with geographically
weighted regression. Natural Hazards Earth System Science 13: 311-327.
Meier, R. B. 1983. Code enforcement and housing quality revisited: the turnover case. Urban
Affairs Quarterly 19 (2): 255-273.
Miller Jr., L. Charles. 1973. The economics of housing code enforcement. Land Economics 49
(1): 92-96.
95
Morckel, V. 2014. Predicting abandoned housing: does the operational definition of
abandonment matter? Community Development 45 (2): 121-133.
Murray, A. T., I. McGuffog, J. S. Western, and P. Mullins. 2001. Exploratory spatial data
analysis techniques for examining urban crime. The British Journal of Criminology (41):
309-329.
Nakaya, T., K. Yano. 2010. Visualizing crime clusters in a space-time cube: an exploratory data-
analysis approach using space-time kernel density estimation and scan statistics.
Transactions in GIS 14 (3): 223-239.
Nakaya, T. Geographically Weighted Regression (version 4.0). Windows. Tempe, AZ: Arizona
State University GeoDa Center for Geospatial Analysis and Computation, 2009.
Osgood, D. W., L. L. Finken, and B. J. McMorris. 2002. Analyzing multiple-item measures of
crime and deviance II: Tobit regression analysis of transformed scores. Journal of
Quantitative Criminology 18 (4): 319-347.
Ozdemir, A. 2011. Using a binary logistic regression method and GIS for evaluating and
mapping the groundwater spring potential in the Sultan Mountains (Aksehir, Turkey).
Journal of Hydrology 405: 123-136.
Perestrello, M. J., S. Silva, M. Tomé, M. Alvim, and J. M. Cardoso Pereira. 2001. Spatial
prediction of fire ignition probabilities: comparing logistic regression and neural
networks. Photogrammetric Engineering & Remote Sensing 67 (1): 73-81.
Rodrigues, M., J. de la Riva, and S. Fotheringham. 2014. Modeling the spatial variation of the
explanatory factors of human-caused wildfires in Spain using geographically weighted
logistic regression. Applied Geography 48: 52-63.
Ross, H. L. 1996. Housing code enforcement and urban decline. Journal of Affordable Housing
& Community Development Law 6 (1): 29-46.
Saefuddin, A., N. A. Setiabudi, and A. Fitrianto. 2012. On comparison between logistic
regression and geographically weighted logistic regression: with application to
Indonesian poverty data. World Applied Sciences Journal 19 (2): 205-210.
Simon, Steve. 2013. Calculating predicted probabilities from a logistic regression model. The
Monthly Mean. July 31. Accessed November 27, 2015.
http://www.pmean.com/13/predicted.html
Speer, P. W., D. M. Gorman, E. W. Labouvie, and M. J. Ontkush. 1998. Violent crime and
alcohol availability: relationships in an urban community. Journal of Public Health
Policy 19 (3): 303-318
Teater, B. A. 2011. A qualitative evaluation of the section 8 housing choice voucher program.
Qualitative Social Work 10 (4): 503-519.
96
Tobler, W. R. 1970. A computer movie simulating urban growth in the Detroit region. Economic
Geography 46: 234-240.
Toomey, T. L., D. J. Erickson, B. P. Carlin, K. M. Lenk, H. S. Quick, A. M. Jones, and E. M.
Harwood. 2012. The association between density of alcohol establishments and violent
crime within urban neighborhoods. Alcoholism: Clinical and Experimental Research 36
(8): 1468-1473.
Wheeler, D.C, and L. A. Waller. 2008. Comparing spatially varying coefficient models: a case
study examining violent crime rates and their relationships to alcohol outlets and illegal
drug arrests. Journal of Geographic Systems (2009) 11 (1): 1-22.
Wu, W., L. Zhang. 2013. Comparison of spatial and non-spatial logistic regression models for
modeling the occurrence of cloud cover in north-eastern Puerto Rico. Applied Geography
37: 52-62.
Xue, Y., and D. E. Brown. 2006. Spatial analysis with preference specification of latent decision
makers for criminal event prediction. Decision support systems 41: 560-573.
97
Appendix A: Area 2 and Area 3 Maps
98
99
Appendix B Example SPSS Output
100
101
102
Appendix C Logistic Regression Results Full Summary Table
Multicollinearity Predictors Coefficients & Relationships
Variable Name and (Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Sig.
Keep In
Model?
Variable
Coefficient
Variable
Odds
Ratio -
Exp(B)
Odds
Ratio
Sig.
AREA 1
Lot Size (LOTSQFT) 4.116 1.520 0.022 YES 0.000 1.000 0.335
Floor Area
(FLOOR_AREA)
1.889 1.718 0.000 YES -3.978 0.019 0.025
Number of Stories
(NOSTORY)
2.842 1.475 0.000 YES -0.140 0.869 0.667
Length of Ownership
(LENGTH_OWN)
1.184 1.088 0.009 YES 0.000 0.039 1.000
Structure Age
(STRUCT_AGE)
1.047 1.020 0.208 NO
Total Assessed Value
(TOTAL_VALU)
7.646 REMOVED
Value per Sq Ft
(VALUE_PSF)
4.186 1.215 0.698 YES 0.001 1.001 0.928
Code Case Yes/No
(CODECASE)
5.955 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
7.812 2.314 0.000 YES 0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
2.553 2.443 0.000 YES 0.147 0.200 1.158
Tax Defaulted Yes/No
(TAX_STATUS)
1.033 1.019 0.208 NO
Occupancy
(OCCUPANCY)
1.152 1.035 0.934 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
5.772 REMOVED
Travel Distance to Liquor
Store (NETWORK_DIST)
1.171 1.106 0.000 YES 0.000 1.000 0.000
Number of Nearby Cases
(NEARBY_CASE)
3.229 1.966 0.000 YES 0.059 1.061 0.069
Corporate Ownership
(CORP_OWNED)
1.172 1.154 0.504 NO
Constant 2.396 10.976 0.054
AREA 2
Lot Size (LOTSQFT) 4.116 REMOVED
Floor Area
(FLOOR_AREA)
1.889 1.718 0.265 NO
Number of Stories
(NOSTORY)
2.842 1.475 0.820 NO
103
Multicollinearity Predictors Coefficients & Relationships
Variable Name and (Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Sig.
Keep In
Model?
Variable
Coefficient
Variable
Odds
Ratio -
Exp(B)
Odds
Ratio
Sig.
Length of Ownership
(LENGTH_OWN)
1.184 1.088 0.811 NO
Structure Age
(STRUCT_AGE)
1.047 1.020 0.640 NO
Total Assessed Value
(TOTAL_VALU)
7.646 REMOVED
Value per Sq Ft
(VALUE_PSF)
4.186 1.215 0.048 YES 0.001 1.001 0.928
Code Case Yes/No
(CODECASE)
5.955 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
7.812 2.314 0.000 YES -0.001 0.999 0.000
Number of Previous Cases
(CASE_COUNT)
2.553 2.443 0.000 YES 0.067 1.069 0.695
Tax Defaulted Yes/No
(TAX_STATUS)
1.033 1.019 0.643 NO
Occupancy
(OCCUPANCY)
1.152 1.035 0.206 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
5.772 REMOVED
Travel Distance to Liquor
Store (NETWORK_DIST)
1.171 1.106 0.845 NO
Number of Nearby Cases
(NEARBY_CASE)
3.229 1.966 0.000 YES 0.001 1.001 0.973
Corporate Ownership
(CORP_OWNED)
1.172 1.154 0.616 NO
Constant 1.261 3.527 0.154
AREA 3
Lot Size (LOTSQFT) 5.881 REMOVED
Floor Area
(FLOOR_AREA)
8.405 1.726 0.247 NO
Number of Stories
(NOSTORY)
2.656 1.752 0.156 NO
Length of Ownership
(LENGTH_OWN)
1.202 1.170 0.026 YES 0.000 1.000 0.109
Structure Age
(STRUCT_AGE)
1.006 1.004 0.775 NO
Total Assessed Value
(TOTAL_VALU)
7.826 REMOVED
Value per Sq Ft
(VALUE_PSF)
6.509 1.345 0.452 NO
104
Multicollinearity Predictors Coefficients & Relationships
Variable Name and (Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Sig.
Keep In
Model?
Variable
Coefficient
Variable
Odds
Ratio -
Exp(B)
Odds
Ratio
Sig.
Code Case Yes/No
(CODECASE)
9.396 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
2.843 2.843 0.000 YES -0.001 0.999 0.000
Number of Previous Cases
(CASE_COUNT)
2.823 1.115 0.000 YES -0.250 0.779 0.033
Tax Defaulted Yes/No
(TAX_STATUS)
1.005 1.004 0.357 NO
Occupancy
(OCCUPANCY)
1.161 1.149 0.003 YES 0.071 1.074 0.668
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.191 1.191 0.015 YES 0.000 1.000 0.201
Travel Distance to Liquor
Store (NETWORK_DIST)
1.184 1.181 0.003 YES 0.000 1.000 0.015
Number of Nearby Cases
(NEARBY_CASE)
1.120 1.082 0.000 YES 0.034 1.034 0.140
Corporate Ownership
(CORP_OWNED)
1.110 1.114 0.278 NO
Constant 2.114 8.284 0.000
RANDOM SAMPLE 1
Lot Size (LOTSQFT) 6.066 REMOVED
Floor Area
(FLOOR_AREA)
8.364 1.541 0.434 NO
Number of Stories
(NOSTORY)
2.178 1.552 0.071 NO
Length of Ownership
(LENGTH_OWN)
1.117 1.111 0.314 NO
Structure Age
(STRUCT_AGE)
1.033 1.031 0.172 NO
Total Assessed Value
(TOTAL_VALU)
7.542 REMOVED
Value per Sq Ft
(VALUE_PSF)
5.998 1.262 0.351 NO
Code Case Yes/No
(CODECASE)
9.091 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
2.736 2.732 0.000 YES 0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
2.820 2.803 0.000 YES -0.016 0.984 0.919
Tax Defaulted Yes/No
(TAX_STATUS)
1.034 1.033 0.927 NO
105
Multicollinearity Predictors Coefficients & Relationships
Variable Name and (Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Sig.
Keep In
Model?
Variable
Coefficient
Variable
Odds
Ratio -
Exp(B)
Odds
Ratio
Sig.
Occupancy
(OCCUPANCY)
1.140 1.130 0.549 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.746 1.738 0.063 NO
Travel Distance to Liquor
Store (NETWORK_DIST)
1.694 1.694 0.095 NO
Number of Nearby Cases
(NEARBY_CASE)
1.338 1.198 0.000 YES 0.060 1.062 0.035
Corporate Ownership
(CORP_OWNED)
1.088 1.088 0.551 NO
Constant 0.351 1.421 0.369
RANDOM SAMPLE 2
Lot Size (LOTSQFT) 7.635 REMOVED
Floor Area
(FLOOR_AREA)
10.123 1.530 0.149 NO
Number of Stories
(NOSTORY)
2.308 1.581 0.001 YES 0.839 2.315 0.000
Length of Ownership
(LENGTH_OWN)
1.136 1.127 0.135 NO
Structure Age
(STRUCT_AGE)
1.024 1.020 0.121 NO
Total Assessed Value
(TOTAL_VALU)
8.783 REMOVED
Value per Sq Ft
(VALUE_PSF)
6.818 1.291 0.564 NO
Code Case Yes/No
(CODECASE)
9.475 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
10.677 2.253 0.000 YES -0.001 0.999 0.000
Number of Previous Cases
(CASE_COUNT)
2.353 2.344 0.000 YES -0.168 0.846 0.210
Tax Defaulted Yes/No
(TAX_STATUS)
1.054 1.046 0.303 NO
Occupancy
(OCCUPANCY)
1.140 1.071 0.892 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.821 1.804 0.048 YES 0.000 1.000 0.926
Travel Distance to Liquor
Store (NETWORK_DIST)
1.659 1.633 0.010 YES 0.000 1.000 0.058
Number of Nearby Cases
(NEARBY_CASE)
1.456 1.302 0.000 YES 0.016 1.016 0.637
106
Multicollinearity Predictors Coefficients & Relationships
Variable Name and (Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Sig.
Keep In
Model?
Variable
Coefficient
Variable
Odds
Ratio -
Exp(B)
Odds
Ratio
Sig.
Corporate Ownership
(CORP_OWNED)
1.112 1.112 0.918 NO
Constant 1.225 3.405 0.034
RANDOM SAMPLE 3
Lot Size (LOTSQFT) 1.082 1.510 0.336 NO
Floor Area
(FLOOR_AREA)
1.807 2.170 0.001 YES -2.158 0.116 0.094
Number of Stories
(NOSTORY)
2.075 1.676 0.004 YES 0.399 1.491 0.112
Length of Ownership
(LENGTH_OWN)
1.172 1.170 0.740 NO
Structure Age
(STRUCT_AGE)
1.044 1.037 0.033 YES 0.007 1.007 0.359
Total Assessed Value
(TOTAL_VALU)
3.248 REMOVED
Value per Sq Ft
(VALUE_PSF)
2.805 1.311 0.796 NO
Code Case Yes/No
(CODECASE)
9.376 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
10.516 2.092 0.000 YES 0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
2.223 2.196 0.000 YES -0.091 0.913 0.446
Tax Defaulted Yes/No
(TAX_STATUS)
1.053 1.047 0.153 NO
Occupancy
(OCCUPANCY)
1.207 1.148 0.314 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.969 1.953 0.022 YES 0.000 1.000 0.685
Travel Distance to Liquor
Store (NETWORK_DIST)
1.804 1.776 0.000 YES 0.000 1.000 0.009
Number of Nearby Cases
(NEARBY_CASE)
1.386 1.293 0.001 YES 0.019 1.019 0.554
Corporate Ownership
(CORP_OWNED)
1.085 1.075 0.136 NO
Constant 2.119 8.324 0.011
RANDOM SAMPLE 4
Lot Size (LOTSQFT) 7.351 REMOVED
Floor Area
(FLOOR_AREA)
8.822 1.557 0.163 NO
107
Multicollinearity Predictors Coefficients & Relationships
Variable Name and (Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Sig.
Keep In
Model?
Variable
Coefficient
Variable
Odds
Ratio -
Exp(B)
Odds
Ratio
Sig.
Number of Stories
(NOSTORY)
2.082 1.573 0.032 YES 0.598 1.818 0.008
Length of Ownership
(LENGTH_OWN)
1.226 1.209 0.504 NO
Structure Age
(STRUCT_AGE)
1.018 1.016 0.304 NO
Total Assessed Value
(TOTAL_VALU)
7.917 REMOVED
Value per Sq Ft
(VALUE_PSF)
5.786 1.395 0.025 YES -0.017 0.983 0.005
Code Case Yes/No
(CODECASE)
9.925 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
11.466 2.510 0.000 YES 0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
2.698 2.682 0.000 YES 0.012 1.012 0.938
Tax Defaulted Yes/No
(TAX_STATUS)
1.052 1.035 0.284 NO
Occupancy
(OCCUPANCY)
1.233 1.164 0.001 YES -0.168 0.845 0.510
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.955 1.952 0.096 NO
Travel Distance to Liquor
Store (NETWORK_DIST)
1.730 1.740 0.006 YES 0.000 1.000 0.013
Number of Nearby Cases
(NEARBY_CASE)
1.498 1.329 0.000 YES 0.061 1.063 0.039
Corporate Ownership
(CORP_OWNED)
1.103 1.105 0.879 NO
Constant 2.111 8.254 0.004
RANDOM SAMPLE 5
Lot Size (LOTSQFT) 6.725 REMOVED
Floor Area
(FLOOR_AREA)
9.282 1.547 0.016 YES -0.479 0.620 0.727
Number of Stories
(NOSTORY)
2.291 1.592 0.000 YES 0.602 1.826 0.025
Length of Ownership
(LENGTH_OWN)
1.175 1.159 0.698 NO
Structure Age
(STRUCT_AGE)
1.023 1.021 0.173 NO
Total Assessed Value
(TOTAL_VALU)
7.295 REMOVED
108
Multicollinearity Predictors Coefficients & Relationships
Variable Name and (Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Sig.
Keep In
Model?
Variable
Coefficient
Variable
Odds
Ratio -
Exp(B)
Odds
Ratio
Sig.
Value per Sq Ft
(VALUE_PSF)
5.160 1.250 0.573 NO
Code Case Yes/No
(CODECASE)
7.867 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
9.722 2.598 0.000 YES -0.001 0.999 0.000
Number of Previous Cases
(CASE_COUNT)
2.613 2.595 0.000 YES -0.158 0.854 0.320
Tax Defaulted Yes/No
(TAX_STATUS)
1.020 1.018 0.423 NO
Occupancy
(OCCUPANCY)
1.167 1.110 0.003 YES -0.170 0.844 0.518
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.888 1.873 0.010 YES 0.000 1.000 0.864
Travel Distance to Liquor
Store (NETWORK_DIST)
1.802 1.771 0.001 YES 0.000 1.000 0.101
Number of Nearby Cases
(NEARBY_CASE)
1.320 1.220 0.000 YES 0.055 1.056 0.086
Corporate Ownership
(CORP_OWNED)
1.115 1.113 0.248 NO
Constant 1.414 4.111 0.110
COMBINED AREAS
Lot Size (LOTSQFT) 2.682 REMOVED
Floor Area
(FLOOR_AREA)
4.073 1.551 0.000 YES -0.623 0.536 0.309
Number of Stories
(NOSTORY)
2.144 1.580 0.000 YES 0.437 1.548 0.000
Length of Ownership
(LENGTH_OWN)
1.157 1.139 0.503 NO
Structure Age
(STRUCT_AGE)
1.002 1.002 0.597 NO
Total Assessed Value
(TOTAL_VALU)
4.048 REMOVED
Value per Sq Ft
(VALUE_PSF)
3.286 1.288 0.444 NO
Code Case Yes/No
(CODECASE)
9.106 REMOVED
Days to Previous Violation
(DAYS_TO_VI)
10.761 2.410 0.000 YES 0.000 1.000 0.000
Number of Previous Cases
(CASE_COUNT)
2.516 2.495 0.000 YES -0.088 0.916 0.203
109
Multicollinearity Predictors Coefficients & Relationships
Variable Name and (Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Sig.
Keep In
Model?
Variable
Coefficient
Variable
Odds
Ratio -
Exp(B)
Odds
Ratio
Sig.
Tax Defaulted Yes/No
(TAX_STATUS)
1.010 1.009 0.341 NO
Occupancy
(OCCUPANCY)
1.173 1.101 0.070 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.833 1.822 0.000 YES 0.000 1.000 0.659
Travel Distance to Liquor
Store (NETWORK_DIST)
1.724 1.713 0.000 YES 0.000 1.000 0.000
Number of Nearby Cases
(NEARBY_CASE)
1.387 1.258 0.000 YES 0.022 1.022 0.143
Corporate Ownership
(CORP_OWNED)
1.105 1.105 0.752 NO
Constant 1.602 4.961 0.000
110
Appendix D Example GWR4 Output
111
112
113
Multicollinearity Predictors Coefficients & Relationships
Variable Name and
(Alias)
Initial
VIF
Value
Final VIF
Value
Predictor
Significance
Keep
In
Model?
Variable
Coefficient
Vaiable
Odds
Ratio -
Exp(B)
Odds Ratio
Significance
AREA 1
Lot Size (LOTSQFT) 4.116 1.520 0.022 YES 0.000 1.000 0.335
Floor Area
(FLOOR_AREA)
1.889 1.718 0.000 YES -3.978 0.019 0.025
Number of Stories
(NOSTORY)
2.842 1.475 0.000 YES -0.140 0.869 0.667
Length of Ownership
(LENGTH_OWN)
1.184 1.088 0.009 YES 0.000 0.039 1.000
Structure Age
(STRUCT_AGE)
1.047 1.020 0.208 NO
114
Total Assessed Value
(TOTAL_VALU)
7.646 REMOVED
Value per Sq Ft
(VALUE_PSF)
4.186 1.215 0.698 YES 0.001 1.001 0.928
Code Case Yes/No
(CODECASE)
5.955 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
7.812 2.314 0.000 YES 0.000 1.000 0.000
Number of Previous
Cases (CASE_COUNT)
2.553 2.443 0.000 YES 0.147 0.200 1.158
Tax Defaulted Yes/No
(TAX_STATUS)
1.033 1.019 0.208 NO
Occupancy
(OCCUPANCY)
1.152 1.035 0.934 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
5.772 REMOVED
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.171 1.106 0.000 YES 0.000 1.000 0.000
Number of Nearby
Cases
(NEARBY_CASE)
3.229 1.966 0.000 YES 0.059 1.061 0.069
Corporate Ownership
(CORP_OWNED)
1.172 1.154 0.504 NO
Constant 2.396 10.976 0.054
AREA 2
Lot Size (LOTSQFT) 4.116 REMOVED
Floor Area
(FLOOR_AREA)
1.889 1.718 0.265 NO
Number of Stories
(NOSTORY)
2.842 1.475 0.820 NO
Length of Ownership
(LENGTH_OWN)
1.184 1.088 0.811 NO
Structure Age
(STRUCT_AGE)
1.047 1.020 0.640 NO
Total Assessed Value
(TOTAL_VALU)
7.646 REMOVED
Value per Sq Ft
(VALUE_PSF)
4.186 1.215 0.048 YES 0.001 1.001 0.928
Code Case Yes/No
(CODECASE)
5.955 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
7.812 2.314 0.000 YES -0.001 0.999 0.000
Number of Previous
Cases (CASE_COUNT)
2.553 2.443 0.000 YES 0.067 1.069 0.695
Tax Defaulted Yes/No
(TAX_STATUS)
1.033 1.019 0.643 NO
115
Occupancy
(OCCUPANCY)
1.152 1.035 0.206 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
5.772 REMOVED
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.171 1.106 0.845 NO
Number of Nearby
Cases
(NEARBY_CASE)
3.229 1.966 0.000 YES 0.001 1.001 0.973
Corporate Ownership
(CORP_OWNED)
1.172 1.154 0.616 NO
Constant 1.261 3.527 0.154
AREA 3
Lot Size (LOTSQFT) 5.881 REMOVED
Floor Area
(FLOOR_AREA)
8.405 1.726 0.247 NO
Number of Stories
(NOSTORY)
2.656 1.752 0.156 NO
Length of Ownership
(LENGTH_OWN)
1.202 1.170 0.026 YES 0.000 1.000 0.109
Structure Age
(STRUCT_AGE)
1.006 1.004 0.775 NO
Total Assessed Value
(TOTAL_VALU)
7.826 REMOVED
Value per Sq Ft
(VALUE_PSF)
6.509 1.345 0.452 NO
Code Case Yes/No
(CODECASE)
9.396 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
2.843 2.843 0.000 YES -0.001 0.999 0.000
Number of Previous
Cases (CASE_COUNT)
2.823 1.115 0.000 YES -0.250 0.779 0.033
Tax Defaulted Yes/No
(TAX_STATUS)
1.005 1.004 0.357 NO
Occupancy
(OCCUPANCY)
1.161 1.149 0.003 YES 0.071 1.074 0.668
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.191 1.191 0.015 YES 0.000 1.000 0.201
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.184 1.181 0.003 YES 0.000 1.000 0.015
Number of Nearby
Cases
(NEARBY_CASE)
1.120 1.082 0.000 YES 0.034 1.034 0.140
Corporate Ownership
(CORP_OWNED)
1.110 1.114 0.278 NO
116
Constant 2.114 8.284 0.000
RANDOM SAMPLE 1
Lot Size (LOTSQFT) 6.066 REMOVED
Floor Area
(FLOOR_AREA)
8.364 1.541 0.434 NO
Number of Stories
(NOSTORY)
2.178 1.552 0.071 NO
Length of Ownership
(LENGTH_OWN)
1.117 1.111 0.314 NO
Structure Age
(STRUCT_AGE)
1.033 1.031 0.172 NO
Total Assessed Value
(TOTAL_VALU)
7.542 REMOVED
Value per Sq Ft
(VALUE_PSF)
5.998 1.262 0.351 NO
Code Case Yes/No
(CODECASE)
9.091 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
2.736 2.732 0.000 YES 0.000 1.000 0.000
Number of Previous
Cases (CASE_COUNT)
2.820 2.803 0.000 YES -0.016 0.984 0.919
Tax Defaulted Yes/No
(TAX_STATUS)
1.034 1.033 0.927 NO
Occupancy
(OCCUPANCY)
1.140 1.130 0.549 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.746 1.738 0.063 NO
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.694 1.694 0.095 NO
Number of Nearby
Cases
(NEARBY_CASE)
1.338 1.198 0.000 YES 0.060 1.062 0.035
Corporate Ownership
(CORP_OWNED)
1.088 1.088 0.551 NO
Constant 0.351 1.421 0.369
RANDOM SAMPLE 2
Lot Size (LOTSQFT) 7.635 REMOVED
Floor Area
(FLOOR_AREA)
10.123 1.530 0.149 NO
Number of Stories
(NOSTORY)
2.308 1.581 0.001 YES 0.839 2.315 0.000
Length of Ownership
(LENGTH_OWN)
1.136 1.127 0.135 NO
Structure Age
(STRUCT_AGE)
1.024 1.020 0.121 NO
117
Total Assessed Value
(TOTAL_VALU)
8.783 REMOVED
Value per Sq Ft
(VALUE_PSF)
6.818 1.291 0.564 NO
Code Case Yes/No
(CODECASE)
9.475 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
10.677 2.253 0.000 YES -0.001 0.999 0.000
Number of Previous
Cases (CASE_COUNT)
2.353 2.344 0.000 YES -0.168 0.846 0.210
Tax Defaulted Yes/No
(TAX_STATUS)
1.054 1.046 0.303 NO
Occupancy
(OCCUPANCY)
1.140 1.071 0.892 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.821 1.804 0.048 YES 0.000 1.000 0.926
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.659 1.633 0.010 YES 0.000 1.000 0.058
Number of Nearby
Cases
(NEARBY_CASE)
1.456 1.302 0.000 YES 0.016 1.016 0.637
Corporate Ownership
(CORP_OWNED)
1.112 1.112 0.918 NO
Constant 1.225 3.405 0.034
RANDOM SAMPLE 3
Lot Size (LOTSQFT) 1.082 1.510 0.336 NO
Floor Area
(FLOOR_AREA)
1.807 2.170 0.001 YES -2.158 0.116 0.094
Number of Stories
(NOSTORY)
2.075 1.676 0.004 YES 0.399 1.491 0.112
Length of Ownership
(LENGTH_OWN)
1.172 1.170 0.740 NO
Structure Age
(STRUCT_AGE)
1.044 1.037 0.033 YES 0.007 1.007 0.359
Total Assessed Value
(TOTAL_VALU)
3.248 REMOVED
Value per Sq Ft
(VALUE_PSF)
2.805 1.311 0.796 NO
Code Case Yes/No
(CODECASE)
9.376 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
10.516 2.092 0.000 YES 0.000 1.000 0.000
Number of Previous
Cases (CASE_COUNT)
2.223 2.196 0.000 YES -0.091 0.913 0.446
Tax Defaulted Yes/No
(TAX_STATUS)
1.053 1.047 0.153 NO
118
Occupancy
(OCCUPANCY)
1.207 1.148 0.314 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.969 1.953 0.022 YES 0.000 1.000 0.685
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.804 1.776 0.000 YES 0.000 1.000 0.009
Number of Nearby
Cases
(NEARBY_CASE)
1.386 1.293 0.001 YES 0.019 1.019 0.554
Corporate Ownership
(CORP_OWNED)
1.085 1.075 0.136 NO
Constant 2.119 8.324 0.011
RANDOM SAMPLE 4
Lot Size (LOTSQFT) 7.351 REMOVED
Floor Area
(FLOOR_AREA)
8.822 1.557 0.163 NO
Number of Stories
(NOSTORY)
2.082 1.573 0.032 YES 0.598 1.818 0.008
Length of Ownership
(LENGTH_OWN)
1.226 1.209 0.504 NO
Structure Age
(STRUCT_AGE)
1.018 1.016 0.304 NO
Total Assessed Value
(TOTAL_VALU)
7.917 REMOVED
Value per Sq Ft
(VALUE_PSF)
5.786 1.395 0.025 YES -0.017 0.983 0.005
Code Case Yes/No
(CODECASE)
9.925 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
11.466 2.510 0.000 YES 0.000 1.000 0.000
Number of Previous
Cases (CASE_COUNT)
2.698 2.682 0.000 YES 0.012 1.012 0.938
Tax Defaulted Yes/No
(TAX_STATUS)
1.052 1.035 0.284 NO
Occupancy
(OCCUPANCY)
1.233 1.164 0.001 YES -0.168 0.845 0.510
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.955 1.952 0.096 NO
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.730 1.740 0.006 YES 0.000 1.000 0.013
Number of Nearby
Cases
(NEARBY_CASE)
1.498 1.329 0.000 YES 0.061 1.063 0.039
Corporate Ownership
(CORP_OWNED)
1.103 1.105 0.879 NO
119
Constant 2.111 8.254 0.004
RANDOM SAMPLE 5
Lot Size (LOTSQFT) 6.725 REMOVED
Floor Area
(FLOOR_AREA)
9.282 1.547 0.016 YES -0.479 0.620 0.727
Number of Stories
(NOSTORY)
2.291 1.592 0.000 YES 0.602 1.826 0.025
Length of Ownership
(LENGTH_OWN)
1.175 1.159 0.698 NO
Structure Age
(STRUCT_AGE)
1.023 1.021 0.173 NO
Total Assessed Value
(TOTAL_VALU)
7.295 REMOVED
Value per Sq Ft
(VALUE_PSF)
5.160 1.250 0.573 NO
Code Case Yes/No
(CODECASE)
7.867 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
9.722 2.598 0.000 YES -0.001 0.999 0.000
Number of Previous
Cases (CASE_COUNT)
2.613 2.595 0.000 YES -0.158 0.854 0.320
Tax Defaulted Yes/No
(TAX_STATUS)
1.020 1.018 0.423 NO
Occupancy
(OCCUPANCY)
1.167 1.110 0.003 YES -0.170 0.844 0.518
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.888 1.873 0.010 YES 0.000 1.000 0.864
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.802 1.771 0.001 YES 0.000 1.000 0.101
Number of Nearby
Cases
(NEARBY_CASE)
1.320 1.220 0.000 YES 0.055 1.056 0.086
Corporate Ownership
(CORP_OWNED)
1.115 1.113 0.248 NO
Constant 1.414 4.111 0.110
COMBINED AREAS
Lot Size (LOTSQFT) 2.682 REMOVED
Floor Area
(FLOOR_AREA)
4.073 1.551 0.000 YES -0.623 0.536 0.309
Number of Stories
(NOSTORY)
2.144 1.580 0.000 YES 0.437 1.548 0.000
Length of Ownership
(LENGTH_OWN)
1.157 1.139 0.503 NO
Structure Age
(STRUCT_AGE)
1.002 1.002 0.597 NO
120
Total Assessed Value
(TOTAL_VALU)
4.048 REMOVED
Value per Sq Ft
(VALUE_PSF)
3.286 1.288 0.444 NO
Code Case Yes/No
(CODECASE)
9.106 REMOVED
Days to Previous
Violation
(DAYS_TO_VI)
10.761 2.410 0.000 YES 0.000 1.000 0.000
Number of Previous
Cases (CASE_COUNT)
2.516 2.495 0.000 YES -0.088 0.916 0.203
Tax Defaulted Yes/No
(TAX_STATUS)
1.010 1.009 0.341 NO
Occupancy
(OCCUPANCY)
1.173 1.101 0.070 NO
Cartesian Distance to
Liquor Store
(NEAR_DIST)
1.833 1.822 0.000 YES 0.000 1.000 0.659
Travel Distance to
Liquor Store
(NETWORK_DIST)
1.724 1.713 0.000 YES 0.000 1.000 0.000
Number of Nearby
Cases
(NEARBY_CASE)
1.387 1.258 0.000 YES 0.022 1.022 0.143
Corporate Ownership
(CORP_OWNED)
1.105 1.105 0.752 NO
Constant 1.602 4.961 0.000
121
122
Appendix E GWR4 Coefficient Maps
123
124
125
126
127
128
129
Abstract (if available)
Abstract
Cities throughout the country are constantly striving to improve their perceived image. Whether it is requiring lush landscaping in commercial developments, or simply making sure that the trim on a house is properly painted, cities are constantly struggling to get citizens to comply with municipal codes. Such is the case in the City of Victorville, CA, where economic recovery has been slow following the 2008 housing market crash, leaving poorly maintained properties in its wake. Presently, Victorville’s code enforcement staff is doing a proactive enforcement survey of all single-family homes in the city in an effort to "clean up" these properties. However, the survey is inefficient and is taking up a good amount of officer time, leaving commercial and industrial areas of the city neglected. This project was able to predict which houses in Victorville are likely to have a code enforcement violation that requires action from staff in order to better allocate resources to areas that require more attention and pull resources from areas that do not require attention. The primary question here is what property attributes can be used to predict the occurrence of a code enforcement violation? Several have been selected, including property value, length of ownership, and presence of a previous violation. A binary logistic regression analysis was run on three areas of the city containing approximately 2,200 homes that have already been surveyed in order to train a model for predicting the remaining 29,000 homes. Geographically weighted logistic regression was then employed to factor in spatial variation in the relationships between the response variable and the explanatory variables. The success of this model will make Victorville’s code enforcement more efficient, and it is a model that any city can employ to make its own code enforcement departments more effective.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Predicting the presence of historic and prehistoric campsites in Virginia’s Chesapeake Bay counties
PDF
A spatial analysis of violent crime cold spots: testing the capable guardian component of the routine activity theory
PDF
A comparison of GLM, GAM, and GWR modeling of fish distribution and abundance in Lake Ontario
PDF
Modeling nitrate contamination of groundwater in Mountain Home, Idaho using the DRASTIC method
PDF
Using Maxent to model the distribution of prehistoric agricultural features in a portion of the Hōkūli‘a subdivision in Kona, Hawai‘i
PDF
Predicting archaeological site locations in northeastern California’s High Desert using the Maxent model
PDF
A model for emergency logistical resource requirements: supporting socially vulnerable populations affected by the (M) 7.8 San Andreas earthquake scenario in Los Angeles County, California
PDF
Selection of bridge location over the Merrimack River in southern New Hampshire: a comparison of site suitability assessments
PDF
Modeling geopolitics in Tikal through least cost paths
PDF
Archaeological least cost path modeling: a behavioral study of Middle Bronze Age merchant travel routes across the Amanus Mountains, Turkey
PDF
Precision agriculture and GIS: evaluating the use of yield maps combined with LiDAR data
PDF
Beach morphodynamic change detection using LiDAR during El Niño periods in Southern California
PDF
That sinking feeling: predicting land subsidence in California’s San Joaquin Valley with a spatial regression model
PDF
Assessing the transferability of a species distribution model for predicting the distribution of invasive cogongrass in Alabama
PDF
Obesity and healthy food accessibility: case study of Minnesota, USA
PDF
Estimating populations at risk in data-poor environments: a geographically disaggregated analysis of Boko Haram terrorism 2009-2014
PDF
Analysis of future land use conflict with volcanic hazard zones: Mount Rainier, Washington
PDF
Applying least cost path analysis to search and rescue data: a case study in Yosemite National Park
PDF
Out-of-school suspensions by home neighborhood: a spatial analysis of student suspensions in the San Bernardino City Unified School District
PDF
Using pattern oriented modeling to design and validate spatial models: a case study in agent-based modeling
Asset Metadata
Creator
Pugh, Matthew Dean
(author)
Core Title
Residential housing code violation prediction: a study in Victorville, CA using geographically weighted logistic regression
School
College of Letters, Arts and Sciences
Degree
Master of Science
Degree Program
Geographic Information Science and Technology
Publication Date
03/01/2016
Defense Date
12/07/2015
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
code,code enforcement,geographically weighted logistic regression,geographically weighted regression,hot spot,Housing,logistic regression,logit,OAI-PMH Harvest,regression,Victorville,violation
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Kemp, Karen K. (
committee chair
), Lee, Su Jin (
committee member
), Vos, Robert O. (
committee member
)
Creator Email
matt.pugh87@gmail.com,matthedp@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c40-216934
Unique identifier
UC11278306
Identifier
etd-PughMatthe-4171.pdf (filename),usctheses-c40-216934 (legacy record id)
Legacy Identifier
etd-PughMatthe-4171.pdf
Dmrecord
216934
Document Type
Thesis
Format
application/pdf (imt)
Rights
Pugh, Matthew Dean
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Tags
code
code enforcement
geographically weighted logistic regression
geographically weighted regression
hot spot
logistic regression
logit
regression
violation