FINAL REPORT OF TASK FORCE ON GAMBLING ADDICTION IN MARYLAND:
PROFILE OF PATHOLOGICAL GAMBLERS UNDERGOING TREATMENT IN THE STATE OF
MARYLAND
FINAL REPORT OF TASK FORCE ON GAMBLING ADDICTION IN MARYLAND
by
Valerie C. Lorenz, Ph.D., Robert M. Politzer, Sc.D., & Robert A.
Yaffee, Ph.D.
TABLE OF CONTENTS
LETTER TO THE SECRETARY, DEPARTMENT OF HEALTH AND MENTAL
HYGIENE . . . . . . . . . . . . . . . . . . . . . . . . ii
TABLE OF CONTENTS . . . . . . . . . . . . . . . . . . . . . vi
INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . 1
Fact Sheet. . . . . . . . . . . . . . . . . . . . . . . . 2
Selected Comments of Survey Respondents . . . . . . . . . 3
Establishment and Purpose of the Task Force . . . . . . . 4
Membership of the Task Force. . . . . . . . . . . . . . . 5
Acknowledgements. . . . . . . . . . . . . . . . . . . . . 7
Work of the Task Force. . . . . . . . . . . . . . . . . . 9
CONCLUSIONS AND RECOMMENDATIONS - SUMMARY. . . . . . . . . . 12
PATHOLOGICAL GAMBLING. . . . . . . . . . . . . . . . . . . . 19
Types of Gamblers . . . . . . . . . . . . . . . . . . . 21
Clinical Definition . . . . . . . . . . . . . . . . . . 24
The Stages of Pathological Gambling . . . . . . . . . . 25
Criminal Behavior . . . . . . . . . . . . . . . . . . . 28
Treatment and Recovery. . . . . . . . . . . . . . . . . 29
Public Health Impact. . . . . . . . . . . . . . . . . . 30
The Epidemiologic Model . . . . . . . . . . . . . . . . 31
HISTORY OF PATHOLOGICAL GAMBLING TREATMENT IN MARYLAND . . . 35
Legislation . . . . . . . . . . . . . . . . . . . . . . 36
Beginnings. . . . . . . . . . . . . . . . . . . . . . . 37
Johns Hopkins Center for Pathological Gambling. . . . . 38
Washington Center . . . . . . . . . . . . . . . . . . . 43
Taylor Manor Hospital . . . . . . . . . . . . . . . . . 44
Changing Point. . . . . . . . . . . . . . . . . . . . . 45
Epoch House . . . . . . . . . . . . . . . . . . . . . . 45
National Center for Pathological Gambling, Inc. . . . . 46
Maryland Council On Compulsive Gambling . . . . . . . . 47
Hotline . . . . . . . . . . . . . . . . . . . . . . . . 47
Further Developments. . . . . . . . . . . . . . . . . . 49
Current Treatment Options Elsewhere . . . . . . . . . . 51
PREVALENCE OF GAMBLING ADDICTION IN MARYLAND . . . . . . . . 54
ECONOMIC AND SOCIAL IMPACT OF GAMBLING ADDICTION . . . . . . 58
PROFILE OF MARYLAND PATHOLOGICAL GAMBLERS IN PROFESSIONAL
TREATMENT PROGRAMS. . . . . . . . . . . . . . . . . . . 62
The Nature of the Gambling Problem. . . . . . . . . . . 63
A Profile of the Maryland Pathological Gambling Patient:
1983-1989. . . . . . . . . . . . . . . . . . . . . 64
A Statistical Model of the Severity of the Gambling
Problem for Maryland Pathological Gambling
Patients: 1983-1989. . . . . . . . . . . . . . . . 66
Recommendations . . . . . . . . . . . . . . . . . . . . 68
PROFILE OF MARYLAND GAMBLERS ANONYMOUS RESPONDENTS . . . . . 69
PROFILE OF MARYLAND GAM-ANON RESPONDENTS . . . . . . . . . . 72
REPORT OF THE COMPULSIVE GAMBLING HOTLINE. . . . . . . . . . 74
LIABILITY OF THE GAMING INDUSTRY FOR MARYLAND'S PATHOLOGICAL
GAMBLING PROBLEM. . . . . . . . . . . . . . . . . . . . 78
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . 81
APPENDICES
APPENDIX A: A WORD ON ROBERT A. YAFFEE, PH.D.. . . . . . . . 87
APPENDIX B: GAMBLERS ANONYMOUS SURVEY. . . . . . . . . . . . 89
Introduction. . . . . . . . . . . . . . . . . . . . . . 91
Methodology . . . . . . . . . . . . . . . . . . . . . . 92
Results . . . . . . . . . . . . . . . . . . . . . . . . 92
Discussion. . . . . . . . . . . . . . . . . . . . . . . 97
Tabulation of Survey Questions and Responses. . . . . . 100
APPENDIX C: GAM-ANON SURVEY. . . . . . . . . . . . . . . . . 112
Introduction. . . . . . . . . . . . . . . . . . . . . . 114
Methodology . . . . . . . . . . . . . . . . . . . . . . 115
Results . . . . . . . . . . . . . . . . . . . . . . . . 115
GamAnon Respondents' Requests for Help. . . . . . . . . 117
GamAnon Respondents' Messages for the Governor and
Legislators. . . . . . . . . . . . . . . . . . . . 118
Conclusions . . . . . . . . . . . . . . . . . . . . . . 119
Tabulation of Survey Conducted by the Task Force of
Maryland GamAnon Chapters. . . . . . . . . . . . . 120
APPENDIX D: COMPULSIVE GAMBLING HOTLINE -- 1-800-332-0402:
FISCAL YEAR 1990 FINAL REPORT. . . . . . . . . . . . . 124
Background. . . . . . . . . . . . . . . . . . . . . . . 126
Legitimate Calls. . . . . . . . . . . . . . . . . . . . 127
Lottery Calls -- U.S. and Maryland. . . . . . . . . . . 140
Public Relations Efforts. . . . . . . . . . . . . . . . 142
Legislative Activity. . . . . . . . . . . . . . . . . . 143
Summary . . . . . . . . . . . . . . . . . . . . . . . . 144
Representative Calls Received by the Compulsive
Gambling Hotline . . . . . . . . . . . . . . . . . 146
Varied and Different Hotline Calls. . . . . . . . . . . 153
APPENDIX E: PROFILE OF PATHOLOGICAL GAMBLERS UNDERGOING
TREATMENT . . . . . . . . . . . . . . . . . . . . . . . 154
Research Objectives . . . . . . . . . . . . . . . . . . 156
Profile of the Pathological Gambler in Maryland . . . . 162
Personal History of Abuse and Consequences. . . . . . . 166
APPENDIX F: SEVERITY OF COMPULSIVE GAMBLING AND CO-ADDICTION
IN MARYLAND . . . . . . . . . . . . . . . . . . . . . . 192
APPENDIX G: A REVIEW OF PREVALENCE ESTIMATES . . . . . . . . 211
A More Accurate Estimate. . . . . . . . . . . . . . . . 214
Post-Stratification Weights . . . . . . . . . . . . . . 215
Our New Estimate. . . . . . . . . . . . . . . . . . . . 216
Other Statistical Concerns. . . . . . . . . . . . . . . 217
APPENDIX E
PROFILE OF PATHOLOGICAL GAMBLERS
UNDERGOING TREATMENT IN THE STATE OF MARYLAND
Prepared for the Task Force on Gambling Addiction in Maryland
by
Robert A. Yaffee, Ph.D.
Research Consultant, Academic Computing Facility
Courant Institute of Mathematical Sciences
New York University
Submitted February 6, 1990
Research Objectives
There are several objectives of this analysis. The first
goal is to provide a new preliminary profile of the pathological
gambler in the State of Maryland. Of interest are the background
socio-demographic characteristics, financial circumstances, and
psychological attributes of this patient population. Secondly,
this analysis will include a discussion of the preferences for
types of gambling. Thirdly, from the factors significantly
related to the depth of the gambling debt incurred, a statistical
model to explain these phenomena will be postulated, estimated,
and fitted. On these bases, explanations will be advanced to
clarify the nature of the problem of compulsive gambling in
contemporary Maryland.
This analysis has not been attempted before in the State
of Maryland. It provides new and useful information for the
assessment of the nature of the patient population. Such
information should prove useful to researchers and therapists
as seeking to assess the general nature of the problem as it is
confronting them today. It provides policy-makers with an
informational base upon which to assess the nature of the social
problem and its social costs. The information generated by this
study may be useful to scientists seeking to test earlier or
preliminary findings in different locations, preparatory to
tendering nationwide generalizations.
The Data and its Sources
The data for this analysis come from a survey of patients
undergoing therapy for this psychiatric illness. These data are
offered by the three treatment centers in State of Maryland that
have clinical programs dedicated to the treatment of pathological
gambling. These data were provided by The National Center for
Pathological Gambling, the Washington Center, and Taylor Manor
Hospital, in a fashion so that this analyst remains unaware of
the identity of any particular patient.
The time frame during which these data have been collected
dates from the inception of these programs until the Summer of
1989. The Taylor Manor Hospital program was begun in 1983. The
National Center for Pathological Gambling data was started in
1986, while the Washington Center data stems from patients
treated at that facility since 1983. Overall, these data come
from approximately four to six years of treatment of patients in
Maryland during the 1980s.
For this analysis, merged data were taken from at least two
treatment centers. The number of Maryland patients in the
National Center for Pathological Gambling to date was 94. The
number of Maryland patients at Taylor Manor was 93, and the
number in the Washington Center for Pathological Gambling
clinical program was 59. If all three clinical surveys asked the
question, the data from all of them were included as long as the
coding of that question was consistent. If all three treatment
centers asked the question but the coding of the particular
variable from only two treatment centers was consistent, then the
data were culled from those programs with consistent coding. If
only one treatment center asked a question, the question, with
one exception, was not included in this analysis. The total
number of possible observations was therefore 246, but instances
of missing data often limited the number of observations to less
than 246. The larger sample size permitted analyses that
otherwise could not have been performed.
The larger sample size allowed for greater precision and
large sample estimation techniques. The precision of
calculations for all percentage data is rounded to the tenths of
a percent. For logit models, the precision is rounded to the
thousandths. The logit models, requiring multiple cross-
tabulation estimations without many zero cell entries, are made
possible by this collection of data. The resulting logit model
partly explains effects significantly associated with the
severity of the compulsive gambling problem.
Moreover, after coding to preclude duplication of cases
within the treatment subpopulations, traces of respondent
identity have been eliminated. Thus the identity and anonymity
of the respondents have been guaranteed to the fullest extent
possible. In keeping these confidences, this Task Force has
complied with our ethical obligations as researchers and
clinicians to the patients who will find that they can again have
reason to accurately confide in the Task Force so that it can
help the policy makers formulate policy in the public and patient
interest.
Methodology
Statistical Techniques
The statistical techniques utilized to attain the stated
objectives are basically two-fold. For a profile of the
contemporary patient population, a frequencies analyses was
undertaken. The percentages that describe the patient population
provide useful information that indicates what sort of people the
therapists are now treating. When the percentages are presented,
they are presented for the category being discussed. When a
number is given, the number or "n" refers to the total number of
respondents giving valid answers to the question. In order to
calculate the total number responding to that question, the
reader merely multiplies the proportion represented by the
indicated percentage by the n supplied. The product of this
calculation is the number of respondents fitting into the answer
category under discussion. This notation is used because some of
the treatment centers providing data to the merged survey may not
have asked particular questions, so that the sample size varies
from question to question. Also, some respondents provided no
answer to particular questions.
Background data consist of socio-demographic information,
financial circumstances and psychological attributes which is
provided by a percentage analysis. The same sort of analysis was
used to examine the gambling preferences of these patients. But
to examine the severity of the gambling problem, crosstabulations
of the severity of gambling problem ratio (a trichotomized ratio
of the gambling debt to the annual personal income) by other
variables were used to screen for candidates to a logit model.
The accompanying Pearson chi-square and likelihood ratio chi-
square tests of significance were used to assess the level of
significance. When problems with the chi-square preclude its
use, measures of significance of proportions may be used.
Relationships with chi-square significances of p < .05 are then
chosen for inclusion into the broader multifactor categorical
analysis. Because most of the variable were coded as either
ordered or unordered categories, categorical data analysis of a
multifactor approach is used for general model building.
A particular type of categorical data analysis -- namely,
logit analysis -- was selected for this general model building.
Logit analysis is based on a multidimensional crosstabulation.
It can be construed as a kind of analysis of variance model
utilizing the logit transformation of a dependent variable. The
logit transformation is merely the natural log of the odds ratio
and the odds ratio is the probability of an event happening
divided by the probability of its not happening. Probabilities
can be assessed by proportions (or percentages, if you will) of
cell counts in a crosstabulation. Although logit analysis is
more elaborate than a simple crosstabulational analysis, it was
selected for several reasons. Logit analysis is appropriate when
the dependent variable is dichotomous, ordinal, or categorical in
nature. Most of the variables in this study are categorical or
ordinal in nature. Even if the dependent variable were a
continuous ratio, the logit analyses would generally be more
robust than commonly used techniques of discriminant function
analysis or classical regression analysis with such continuous
dependent variables in that validity is dependent on fewer
assumptions. Yet logit analyses differ from the earlier
crosstabulations that were taken from the studies performed by
the different treatment centers.
The data from all of the major treatment centers are used to
bestow this analysis with more sensitivity and power. Unlike the
crosstabulational analyses, which has difficulty handling more
than three variables at a time, logit modeling has the advantage
of being able to accommodate a larger scale multi-variable
approach by which it is possible to assess the significance and
magnitude of the main effects of more components as well as their
respective interactions on the model. This provides a better way
to examine the nature of catalyst or suppressor main effects as
well as interactions. The more elaborate logit model allows
examination of a multitude of effects that may be antecedent to
or confounding of a simple bivariate crosstabulational analysis.
The logit analyses permit, without serious violation of model
assumptions, the study of the nature of relationships among
categorical, nominal, and ordinal variables, when the dependent
variable is binary, ordinal or nominal, as is the case in this
study. The maximum-likelihood estimation of the techniques is
not as susceptible to problems of nonlinearity, misestimation,
heteroskedasctity, and nonnormality as are ordinary least squares
estimation procedures found in the classical regression
analyses. Because the logit model is a more elaborate and
elegant means of handling these data, it is used to assess the
nature of the severity of the gambling problem among Maryland's
patient gamblers today.
If the dependent variable were ordinal or categorical, it
might be objected that one could employ discriminant function
analysis. For that analysis to be valid, the assumptions of
multivariate normality and equality of the error co-variance
matrices would have to hold. In practice, these conditions are
seldom met. In fact, when most predictor variables are not
continuous, the technique usually fails to give accurate results.
It might be objected that one could use classical regression
analysis with a dichotomous dependent variable to obtain similar
results. The classical regression model presupposes a continuous
dependent variable, which means that it is inappropriate for such
analyses. Using ordinary least squares estimation, the classical
regression model tends to overpredict and underpredict the values
of the dependent variable leading to bias. If the dependent
variable were dichotomous (for example, coded either 0 or 1),
ordinary regression would produce predictions higher than the 1
and lower than the 0 value. It would also predict values between
the two values of the dichotomous dependent variable. If the
dependent variable were slightly ordinal, with a few ordered
values, ordinary regression would similarly mispredict the values
of the dependent variable. If the dependent variable were
unordered and polytomous, classical regression would be inap-
propriate. The F tests for significance testing would not work
and there would be no clear way to determine whether the pre-
dictors should be included in the model. The coefficient of
determination, the R square, would not work either so that there
would be no reliable way to assess the strength of the overall
model. Therefore, when the dependent variable is an ordered
typology or ordinal polytomy, as is the level of the gambling
problem, the use of a logit analysis with cumulative logits is
indicated. The cumulative logit makes use of the ordering effect
in the dependent variable and is tantamount to a series of logits
with cut points between the odds ratios that proceed up the level
of the ordinal dependent variable. For this purpose, a part of
the Statistical Analysis System (SAS) User's Group International
Supplementary Library, the SAS logist procedure, was selected.
The procedure uses maximum likelihood estimation and cumulative
logits for ordinal dependent variables.
When the dependent variable was binary coded, rather than
ordered polytomous, logistic regression was performed. The
transformation of the dependent variable here is a simple
logistic one, where cumulative logits need not be used to take
advantage of the ordering. The regression is performed on the
natural log of the odds ratio of the moderate or heavy gambling
severity to the lower level of gambling severity. Logistic
regression, with its maximum likelihood estimation, is also
applied here. To provide for criterion validation of the results
of the analysis, when logistic regression was used, the
programming was done using the categorical modeling procedure in
SAS, the logistic regression procedure in SPSS/pc and the logit
procedure in LIMDEP. These programs shall be discussed later
in the report. These binary logistic regression programs
produced the same regression coefficients and significance test
results.
Secondary Analysis
This analysis is a secondary analysis. It combines the data
from several different surveys in retrospect. While many of the
same questions were used in the surveys conducted by the treat-
ment centers, the coding of the responses to the questions
frequently varied. One of the treatment centers dichotomized
approximately 17 of the 41 variables used in its survey. Many
of its questions pertained to whether the patients had or are
abusing substances, while another treatment center coded its
questions according to whether the abuse was past, present, or
giving rise to dependence or past dependence. Another treatment
center only used 14 of the common variables. One treatment
center decided to use a broader question than another treatment
center: asking about all addictions or problems of substance
abuse of the individual parents while another asked about
maternal drug abuse, maternal alcohol abuse, etc., separately.
The task of combining these responses into a universal coding
scheme for an omnibus program for analysis often necessitated
using the broadest categorizations of answer categories at the
loss and cost of information. Where there was insufficient
agreement on the questions asked, the data had to be drawn from
two rather than three treatment centers. If the question was
asked by only one treatment center, it was generally rejected for
the purpose of this aggregate data analysis. Therefore, the
sample size varied with the number of sources of the data.
Profile of the Pathological Gambler in Maryland
A profile of the patient pathological gambler in Maryland
includes an analysis of the nature of the compulsive gambling
problems of the patient population. When the patient began the
gambling, his gambling preferences, the depth of his gambling
debt, the severity of it compared to the size of his annual
income, the extent of his insurance to cover treatment for his
impulse control disorder, and his pending legal problems are
examined as the first part of the nature of his gambling problem.
The analysis of the profile of the compulsive gambler in Maryland
also includes a basic description of demographic, socioeconomic,
parental and psychological characteristics of the compulsive
gambling patients in Maryland. In the treatment of the
demographic attributes of the gambler, the age, race, sex, number
of siblings, and marital status of the gambler are considered.
The profile of the socioeconomic characteristics of these
patients will include the employment status, education and annual
personal income. In the inquiry about influential parental
roles, questions deriving from alcoholic consumption, gambling,
and early demise are addressed. The patients' excesses and
possible consequences are then considered. The histories of
physical or sexual, drugs, alcohol, food abuse are examined.
Whether the patient ever attempted suicide, was an inpatient or
outpatient, was ever incarcerated or has pending legal problems
are examined. In this way, the patient compulsive gambler
profile is developed.
The Nature and Severity of the Gambling Problem
The nature and severity of the gambling problem are partly
addressed by focusing on the gambling preferences of patients as
well as on the effects associated with enhanced severity of the
gambling problem. Most of the contemporary Maryland patients
began the gambling sometime between 10 and 30 years of age. Only
7.8 percent (n=180) commenced gambling before the age of 10, and
only 7.7 percent started wagering after the age of 30. Eighty-
four percent inaugurated their gambling activities during their
teens or prime of life. The nature of the gambling problem
involves the types of gambling that the bettor favors.
The most favored type of gambling of these patients
indicates much about the demand for opportunity to gamble among
those addicted. A plurality (30.6%, n=144) in Maryland indicated
that the racetrack was their most favored betting pastime.
Slightly more than a fifth of these respondents maintained that
poker machines were their favorite form. The casino was favored
by sixteen percent. Many gamblers had more than one favorite
type of gambling.
Pre-eminent among the second most favored type of gambling
among the patients who gambled in more than one way was the
lottery. Almost 36 percent of the patients maintained that this
was their second most favored gambling activity. The casino was
ranked second among the second most favored types while cards
were ranked third. The precise tallies for these types of
gambling preferences may be found in Tables 1 and 2.
The depth of the gambling debt provides an indication of the
severity of the problem. Approximately forty-eight (47.8
percent, n=136) of these individuals owe less than $10,000.
About 25 percent of them owe between $10,000 and $50,000.
Another quarter of them owe more than $50,000. About twelve
(12.2) percent owe more than $100,000 in gambling debt. But the
depth of their gambling problem may better be understood as a
ratio between their gambling debt (numerator) and their annual
personal income (denominator). Persons with a ratio of less than
one do not owe as much as they earn in a year. Persons with a
ratio of more than one owe more than they earn each year. Fifty-
seven percent (n=100) of the gamblers owe one-half or less than
one half of their annual income. Seventy-three percent owe up to
their annual personal income. Another twelve percent owe between
one time and two times their annual personal income. Fifteen
percent of these persons owe in gambling debt more than twice
what they earn in a year. The severity of the gambling problem
ratio is a key construct that will be used in a more complex
mathematical model later in this paper.
Other indications of the severity of the gambling
predicament involve the number of patients with pending legal
problems and lack of insurance coverage for treatment. Questions
have been put to respondents concerning pending legal problems at
two of the treatment centers, while only one treatment center
inquired about the existence of insurance coverage. If an
assessment of the need for support of treatment is to be made,
the extent to which it is already financed must be considered.
The disposition of the court cases may also be affected by
whether the patient had begun treatment as well.
Whether the patient had begun treatment may have depended on
his being able to finance it. For these reasons, the proportion
with and without insurance coverage is considered. It was found
that almost one-fifth (20.3 percent, n=153) of the respondents
indicated that they had pending legal problems, whereas more than
a quarter of those surveyed (28.7%, n=94) reported having no
insurance coverage. Of those that have insurance, one percent
indicated that they had good inpatient insurance but poor
outpatient coverage, 4.3 percent reported VA or Champus, 27.7
percent mentioned HMO or other similar coverage, while 38.3
percent noted having BC/BS or other similar insurance. In all,
almost two-thirds of the patients lacked solid insurance coverage
and one-fifth of these persons were beset with legal problems.
The nature of this problem is somewhat peculiar to Maryland
at the present time. Maryland has its own laws and its own
sources of gambling. The opportunities and constraints differ
from those of, say, New Jersey. In New Jersey, JFK Medical
Center data reveal that, unlike the current situation in Mary-
land, the most favored kind of gambling is horse betting
(52 percent), casino betting (46 percent), sports (41 percent),
non-casino cards and craps (24 percent) and the legal lottery
(13 percent). The nature of the gambling problem within each
state may depend on the opportunities and regulations within the
state and the easy access and proximity of that state to other
opportunities in neighboring states.
Demographic Characteristics of the Gambling Patient
The compulsive gambling patient at the time of this analysis
is for the most part a married, middle-aged, male Caucasian from
a family with two brothers or sisters. But without further
analysis of the age, sex, race, marital status, and number of
siblings, this description would be guilty of propagating an
oversimplified stereotype. Thirty-nine and nine-tenths percent
(n=226) of the respondents said that they were between 30 and
39 years of age. Another 27 percent reported being between
40 and 49 years old. Approximately 16 percent (16.4%) described
themselves as being between 20 and 29 years of age. Only
two percent claimed that they were under 21. Over 50 and younger
than 60 were 11.9 percent. Only 2.2 percent admitted being
senior citizens (over 65 years of age). Most of these patients,
then, seem to be middle-aged.
When gender and race were considered, it was found that 85
percent (n=246) of the compulsive gamblers are male and 15
percent are female. As to race, 86.2 percent (n=217) are white,
12 percent are black, 0.9 percent are Asian, and 0.5 percent are
Hispanic. "Others" amounted to 0.5 percent. The patients are
typically white males, although women and blacks also constituted
significant proportions of this patient population.
Most of these patients are married (58.9%, n=241). Seven-
teen percent are divorced and 12 percent are single. Widows or
widowers constitute 7.1 percent. Three and three-tenths percent
are separated, while 1.7 percent live with a partner.
These patients do not generally come from very large
families. The largest percentage (23.9%, n=159) come from
families with two siblings. Seventeen and sixth-tenths percent
come from families with three brothers or sisters. Thirty-two
and one-tenth percent come from families with one or no siblings.
Thus, almost three-fourths of these persons come from families
with three or less siblings.
Socio-economic Characteristics of the Compulsive Gamblers
In general, these patients are fully-employed as clerical or
sales persons, have at least a high-school education and earn
less than $30,000 a year. Of the 239 persons on whom data was
available, 83.3 percent were fully employed, 2.9 percent were
part-time employed, and 13 percent were unemployed. Less than
one percent of these patients were retired or students.
A plurality of patients were clerical or sales persons, with
smaller portions of this population having come from management
or executive positions or the professions. Of 229 patients,
43.7 percent had clerical or sales occupations. Almost 38 per-
cent (37.6%) held executive or management positions. Almost 14
(13.5) percent were in the professions. Slightly more than four
percent were in business. Less than one percent were housewives
or students.
The educational level of these patients was mostly at the
high school graduate level. Almost a quarter of them had dropped
out of high school and more than half had completed 12 years of
schooling. If one can assume that the respondents had not
repeated years of school and had not skipped grades, some
inferences can be made about the level of educational attainment.
A little more than one-fourth (26 percent, n=246) had dropped out
of high school. More than half (52.4 percent) had graduated high
school. Another 4.1 percent had some college, while another
11.8 percent had two through four years of college. Approxi-
mately six (5.7) percent had gone to graduate school.
The income level of these patients is distributed fairly
evenly across the income spectrum, trailing off a little at the
upper end. Approximately 27 percent (27.2 percent, n=239) earn
$10,000 per year or less. Another 21.8 percent earn between
$11,000 and $20,000 per annum. Nearly 27 percent (26.8 percent)
earned between $20,001 and $30,000. About 13 percent (12.6
percent) earned between $30,001 and $40,000. Less than twelve
percent (11.7 percent) made more than $40,000 per year.
History of Parental Abuse and Loss
These patients come from families with a substantial amount
of parental abuse of alcohol, gambling, and early demise. In a
substantial portion of cases, the father had a problem with
alcohol or gambling. In a surprisingly large portion of cases,
the mother was reported to have passed away before the patient
had turned 18 years of age. Of the 169 patients answering the
question of whether one of the parents had an alcohol problem,
37.9 percent stated that their parents were plagued by such a
problem. While 10.5 percent of the mothers (n=76) had an alcohol
problem, some 37.3 percent of the fathers (n=75) were said to
have had such a problem. Some 37.9 percent (n=169) of the
parents had a gambling problem. Eight and two-tenths percent
(n=170) of the mothers and 23.7 percent (n=170) of the fathers
were reported to have had a gambling problem.
A large portion of the patients lost a parent during
childhood. More than 50 percent (50.3 percent, n=169) of the
mothers and 14.2 percent of the fathers (n=169) had died before
the patient was 18 years old. This history may be indicative of
inadequate resistance to indulgence and early loss of parental
support and guidance.
Personal History of Abuse and Consequences
Substantial proportions of the patient population have been
abused and overindulge. This has led to large portions of them
undergoing treatment and incarceration. Almost four-tenths
(41.1 percent, n=168) have been subjected to physical or sexual
abuse in earlier years. More than one fourth (26.7 percent,
n=187) have had or do have a drug problem, while more than half
(50.8 percent, n=187) have had or do have an alcohol problem.
Fifteen percent (n=187) were reported as overeating in some
manner (whether indulging in quantities of sweets, salts, or
other foods). More than one quarter of the patients have
attempted suicide (25.7 percent, n=167) and many (48 percent,
n=173) have been outpatients before. More than one fifth (22.1
percent, n=172) have been inpatients before and more than one-
fifth had pending legal problems (20.3 percent, n=153).
Approximately 13 percent (13.2 percent, n=174) have found
themselves in jail or prison. From these tallies, it is clear
that this patient population has substantially experienced other
abuses and their consequences.
What are the gambling preferences of these patients? This
variable was one with which there was some difficulty. The
preferences were coded differently by different treatment
centers. When I tried to use the lowest common denominator of
codings that were mutually exclusive and collectively exhaustive
for purposes of statistical inference, the category containing
the largest percentage of gamblers was the miscellaneous or other
category. This classification of the values of this variable,
while statistically correct, was conceptually poor. I decided to
throw out the classification used by that one treatment center,
drop its cases, and employ that classification which the other
two treatment centers used. The problem with this categorization
was that although it was conceptually rich, it was not
statistically amenable to tests of significance because the
values of the variable were not mutually exclusive.
----------------------------------------------------------------
Table 1
FIRST GAMBLING PREFERENCE
FORM OF CUMULATIVE CUMULATIVE
GAMBLING FREQUENCY PERCENT FREQUENCY PERCENT
-----------------------------------------------------------
BINGO 1 0.7 1 0.7
CARDS 14 9.7 15 10.4
CASINO 23 16.0 38 26.4
DICE,BAR BOOT 1 0.7 39 27.1
HORSE/DOGS 44 30.6 83 57.6
LOTTERY/NUMBS 8 5.6 91 63.2
POKER MACHINES 30 20.8 121 84.0
POOL 1 0.7 122 84.7
SPORTS 19 13.2 141 97.9
STOCKS/OPTIONS 1 0.7 142 98.6
ANYTHING 2 1.4 144 100.0
Missing 9
----------------------------------------------------------------
Table 2
SECOND GAMBLING PREFERENCE
FORM OF CUMULATIVE CUMULATIVE
GAMBLING FREQUENCY PERCENT FREQUENCY PERCENT
-----------------------------------------------------------
Missing 9
Missing or none 44
BINGO 1 0.9 1 0.9
BUSINESS 1 0.9 2 1.8
CARDS 14 12.8 16 14.7
CASINO 17 15.6 33 30.3
DICE,BAR BOOT 2 1.8 35 32.1
HORSE/DOGS 10 9.2 45 41.3
LOTTERY/NUMBS 39 35.8 84 77.1
POKER MACHINES 10 9.2 94 86.2
POOL 1 0.9 95 87.2
SLOT MACHINES 1 0.9 96 88.1
SPORTS 10 9.2 106 97.2
STOCKS/OPTIONS 2 1.8 108 99.1
ANYTHING 1 0.9 109 100.0
----------------------------------------------------------------
The patients may have had some difficulty deciding which
category properly characterized their preference. They in some
cases could have legitimately chosen casino or cards, casino or
slots, poker machines or casinos, etc. This means that there is
probably measurement error built into the responses to this
variable. Instead of using this variable for statistical
significance testing, which is precluded by the nature of the
answer categories, the mere frequencies will be presented. From
these runs performed to date, the favorite gambling preference is
the horses or dogs, second is the poker machines, third is the
casinos, and fourth is sports. Among the second most favorite
kind of gambling, the lottery is pre-eminent, then comes casino
gambling followed by cards. Fourth place for the second prefer-
ence is tied by poker machines, horses/dogs, and sports. These
preferences are presented in more detail in Tables 1 and 2 above.
Whether those other matters are significantly associated
with the gambling and the depth of the gambling debt is the
important question to be addressed. That this population has
experienced other abuses does not necessarily mean that those
abuses are related to the gambling problem. They may be wholly
dissociated or partly associated with the problem. This question
is addressed through crosstabulational analysis. The dependent
variable for this analysis is a measure of the seriousness of the
gambling problem, constructed by taking the gambling debt,
measured in thousands of dollars, and dividing it by the annual
personal income of the gambler, also measured in thousands of
dollars. This ratio of gambling debt to income then becomes the
criterion variable in the remainder of the analysis concerning
the severity of the gambling problem.
Crosstabulations of Variables Related to the Severity Ratio
Several variables were found to be possibly significantly
related to the severity of the gambling problem ratio, according
to chi-square or ordinal correlation significance tests -- among
them, the educational experience. Other dummy variables included
whether the father had a gambling problem or whether the mother
died before the patient was 18 years of age. Whether the patient
had been physically or sexually abused, had a drug, alcohol or
overeating problem were also addressed. Whether the patient had
been incarcerated before, and whether he had legal problems
pending, were other dummy variables whose relationship with the
severity of the gambling problem were tested. But not all of
these relationships were significant across the board of tests,
as can be seen in Table 3.
Some relationships were of dubious validity due to the large
portion of cells in the crosstabulation with expected frequencies
less than five. Sparse data artificially inflates the chi-
square, tending to produce spurious indications of significance.
Both the relationship between the severity of gambling problem
and overeating, and the severity of the gambling problem and
whether the patient had ever been jailed, exhibited chi-square
invalidity following from sparse data. Unless the tests of
significance were further corroborated by those applied, on the
one hand, to the Somers' D or, on the other hand, to the
Stuart's Tau-C, these relationships were generally discarded.
Most of the remaining significant relationships, although
significant, are of weak magnitude. That is, such relationships
exhibit correlations of less than or equal to .15. Relationships
between the severity of the gambling problem and each of the
following variables were found to be weak: education, whether
either parent had a gambling problem, and whether the patient was
physically or sexually abused. The relationships between the
severity of the gambling problem and whether the patient was ever
incarcerated, on the one hand, and whether the patient overeats
on the other hand, were found to have weak Tau-Cs and moderate
Somers' Ds. Most of these relationships dropped out of
significance in the later model.
Those relationships whose bivariate crosstabulations were
found to be significant at a level of .10 or less were the
patient's education, income, whether the mother died before the
patient was 18 years of age, whether the patient had been
physically or sexually abused, and whether the patient had a drug
problem. These variables became candidates for the multifactor
logit/logistic models discussed later, and therein were found to
be significant at the .05 level. Other variables apparently
almost significantly related to the severity of the gambling
problem may have manifestations of intervening or antecedent
relationships.
----------------------------------------------------------------
Table 3
Bivariate Tests of Significance and Correlation
between Severity of the Gambling Problem Ratio
and Other Variables
| | Likelihood | |
|Pearson | Ratio | |Somers'
Variable |Chi-sq df p| Chi-sq p |Tau-C ASE | D ASE
| | | |
Education |15.06 8 .06| 18.76 .02* | .06 .04 |.07 .04
Annual | | | |
Income |50.14 10 .00| 64.47 .00**| .29 .04**| .23 .06**
| | | |
Either Parent | | |
Gambled| 9.08 2 .01| 10.12 .00**|-.06 .06 |-.08 .08
Mother | | | |
Gambled| 3.32 2 .19| 5.74 .06 |-.05 .03 |-.17 .10
Father | | | |
Gambled| 7.58 2 .02| 8.71 .01* |-.07 .06 |-.07 .08
| | | |
Mother Died | | |
Early |40.02 2 .00| 43.03 .00**| .45 .07**| .45 .07**
Physical/Sexual | | |
Abuse | 9.10 2 .01| 9.84 .01**| .17 .08**| .18 .08**
Abused | | | |
Drugs |15.57 2 .00| 20.15 .00**|-.24 .05**|-.30 .06**
| | | |
Abused | | | |
Alcohol|12.20 2 .00| 12.42 .00**|-.23 .07**|-.23 .07**
Overeats |14.43 2 .00| 22.46 .00**|-.18 .03**|-.36 .04**
Ever | | | |
Jailed |13.76 2 .00| 11.60 .00**| .17 .06**| .37 .12**
Pending Legal | | |
Problem|13.71 2 .00| 13.90 .00**| .24 .07**|-.37 .10**
Significance (when n -> large): * p < 05 ** p < .01
----------------------------------------------------------------
The educational level was found to be significantly
related to the severity of the gambling problem. The direction
of the relationship appears to be a positive one. The largest
proportion of those with lower levels of gambling problem had the
lowest level of education. The largest proportion of those with
a moderate level of gambling problem had more than 12 but less
than 15 years of education. Meanwhile, the largest proportion of
patients with a high severity of gambling problem had at least 14
years of education. The magnitude of the relationship, whether
indicated by the Stuart's Tau-C or the Somers' D of .16 (signifi-
cance = .053), does not appear to be strong at all.
----------------------------------------------------------------
Table 4
Gambling Problem Severity Ratio
Crosstabulated with Patient Education
Educational Level in Years of Schooling
Gambling (Assuming no grades skipped or repeated)
Problem
Severity HS HS Some College Some
Ratio Dropout Grad College Graduate Grd Schl Total
-------- ------------------------------------------- ------
Low 54 86 7 18 14 179
84.4% 66.7% 70.0% 62.1% 100.0% 72.8%
Medium 5 22 1 4 0 32
7.8 17.5 10.0 13.8 0.0 13.0%
High 5 21 2 7 0 35
7.9 16.3 20.0 24.1 0.0 14.2%
------------------------------------------- ------
Total 64 129 10 29 14 246
Row Percent 26.0% 52.4% 4.1% 11.8% 5.7% 100.0%
Note: Each cell contains both count and column percentage.
----------------------------------------------------------------
Similarly, patient income level was also found to be
significantly related to the severity of the gambling problem.
Although this relationship appeared to be significant, patient
income was not included as an independent variable in the
multivariate model, as it had already been defined as part of the
dependent variable. Concern about the distortion of linkage
between the dependent gambling problem severity ratio and the
linear combination of independent variables, as a result of
correlated errors, provided the basis for this decision.
There appears to be a significant relationship between the
extent of the gambling severity ratio and the death of the mother
before the patient was 18 years of age. The joint distribution
of data, presented in Table 5 below, between the early death of
the mother and the depth of the gambling problem, is charac-
terized by a Pearson chi-square of 40.02 and a likelihood ratio
chi-square of 43.30 with 2 degrees of freedom. The significance
level of both of these coefficients is p < 0.00.
----------------------------------------------------------------
Table 5
Gambling Problem Severity Ratio Crosstabulated
by Mother's Demise before Patient was 18 years old
Gambling Problem Mother Did Mother Did
Severity Ratio Not Die Die Row
--------------------------------- Total
Low 75 37 112
89.3% 43.5% 66.3%
Medium 3 24 27
3.6 28.2 16%
High 6 24 30
7.1 28.2 17.8%
----------------------------------
Total 84 85 169
49.7% 50.3% 100%
Note: Each cell contains both count and column percentage.
----------------------------------------------------------------
In Table 5, these chi-square coefficients are not plagued
with 20 percent or more of its cells having expected frequencies
less than 5, and therefore this significance test appears to be
valid. The magnitude of the relationship appears to be fairly
strong with Stuart's Tau-C and Somers' D both of .45, but with
significance levels of .065. In short, those patients with low
severity ratios tend not to have had their mother die before they
were 18, whereas those with middle or high severity ratios tend
to have had their mothers pass away before they were 18 years of
age.
Another interesting significant relationship emerged. The
linkage between the severity of the gambling problem ratio and
the patient's having been physically or sexually abused appears
according to the chi-square coefficients to be significant. The
Pearson chi-square is 9.95 and the likelihood ratio chi-square is
9.84 with 2 degrees of freedom. For both of these coefficients,
the significance level is p < .00. In Table 6, these data are
presented.
----------------------------------------------------------------
Table 6
Gambling Problem Severity Ratio Crosstabulated
by Patient's Physical or Sexual Abuse
Gambling Problem Not Was
Severity Ratio Abused Abused Row
--------------------------------- Total
Low 71 40 111
71.7% 58.0% 66.1%
Medium 18 9 27
18.2 13.0 16.1%
High 10 20 30
10.1 29.0 17.9%
----------------------------------
Total 99 69 168
58.9% 41.1% 100%
Note: Each cell contains both count and column percentage.
----------------------------------------------------------------
Although this relationship appears to be significant, it is
not a strong relationship. The Stuart's Tau-C is only .17 and
the Somers' D is only .17. Both of these coefficients had less
significant levels of .08 and .07, respectively. Notwithstanding
the weakly moderate relationship, there appears to be some
evidence of past physical or sexual abuse.
The question of cross-addiction often arises. The evidence
from this patient population is that it does not exist. The
existing evidence indicates a lack of cross-addiction. From
Table 7, the data are presented showing the joint distribution
between patient's past or present drug problem and the severity
of the gambling problem. With the Pearson chi-square being 16.57
and the likelihood ratio chi-square being 20.15 with 2 degrees of
freedom, the significance level of both of these coefficients is
p < .00. The significance is further supported by a lack of
cells with expected frequencies less than 5. This points to a
statistically significant relationship. Yet the direction of the
relationship is a negative one. The distribution of those with
drug problems shows a greater concentration among those with low
gambling severity than does the distribution of patients without
drug problems. Among those without drug problems, there are
larger proportions of patients with middle or high gambling
severity ratios. The magnitude of this relationship is indicated
by the Stuart's Tau-C being -0.235 with a significance level of
.05 and a Somers' D of -0.299 with a significance level of .059.
In sum, the more likely a person is to have a drug problem, the
less likely he is to have a more severe gambling problem.
----------------------------------------------------------------
Table 7
Gambling Problem Severity Ratio Crosstabulated
by Patient Past or Present Drug Problem
Gambling Problem Not a Has a
Severity Ratio Problem Problem Row
--------------------------------- Total
Low 84 46 130
61.3% 92.0% 69.5%
Medium 26 1 27
19.0 2.0 14.4%
High 27 3 30
19.7 6.0 16.0%
----------------------------------
Total 137 50 187
73.3% 26.7% 100%
Note: Each cell contains both count and column percentage.
----------------------------------------------------------------
A Model Explaining Seriousness of the Gambling Problem
In order to develop the model of the seriousness of the
gambling problem, the gambling problem severity ratio was con-
structed and used as the basis for the logit analysis. Candidate
predictors were selected from collapsed versions of variables
found to have crosstabular relationships with significance levels
of .10 or less. To explain the model, we first consider con-
struction of the dependent variables, then the selection of the
predictor variables. Afterward, the explanation of the model,
its indications of fit, and the interpretation of its coeffi-
cients are discussed. This ratio was trichotomized into low,
medium and high levels. The low level extended from 0 to 0.234;
the medium level, from 0.233 to 0.8667; and the high level
spanned the region above 0.868 to the maximum of 58.8. Because
this gambling problem severity ratio is an ordered typology, a
logit analysis utilizing cumulative logits was selected.
The logit used here is the natural log of the odds ratio of
the gambling problem variable. The natural log of a number is
the power to which 2.718 is taken to generate that number, with
two exceptions. The natural log of 1 is defined as 0 and the
natural log of 0 is undefined. For example, the natural log of
10 is 2.303. That is, when 2.718 is taken to the 2.303 power,
one obtains the number 10. The odds ratio is the probability of
being characterized by one of the categories (levels) of the
gambling problem variable divided by the probability of not being
characterized by that level. The probability of being in a group
may be empirically obtained by the proportion of total cases in
that group, as long as the observations are independent of one
another. The categories should be mutually exclusive and
collectively exhaustive. In other words, the odds ratio -- given
a variable of two levels, high and low -- is the probability of
being in the upper level of the variable divided by the
probability of being in the lower level. If the gambling
severity problem had only two levels, the logit could be
represented as follows:
Because the gambling problem severity has been constructed
as an ordered trichotomy, cumulative logist, instead of regular
logist, are used. Thus, two equations formulate our model. The
first formula utilizes the natural log of the odds ratio of the
top two levels compared to the bottom level. The second formula
utilizes the natural log of the odds ratio of the top level
compared to the two lower levels. These transformations of the
severity of gambling problem ratio become the dependent variable
in this analysis. Cumulative logits provide us with two
dependent variables:
The single formula provided is a summary formula explaining
the factors that contribute to the explanation of logit1 and
logit2. Alpha1 may be construed as a dummy variable used when we
are examining the logit1 as a dependent variable and alpha2 as a
dummy variable used when we are examining the logit2 dependent
variable. The generic summary formula of the cumulative logit
provided is as follows:
Cum logit = Alpha1 + Alpha2 + B1X1 + B2X2 + ... + BnXn
The B's here are regression coefficients in the formulae.
They are the coefficients expressing the change in the logit that
accompanies a unit change in the independent variable under
consideration. For a total change in the logit score that
accompanies a unit change in particular variables, the whole
formula, including the alpha (intercept value), must be used for
the calculation. The X's are the individual variables.
The estimation process is accomplished by maximum
likelihood. The observed values of the dependent variable are
compared to the fitted values of the model. Generally, the model
estimates the parameters for the alphas and Bs which maximize the
probability of obtaining the observed set of data. This may be
done by calculating the estimated likelihood function, taking its
natural logarithm, and then, computing the partial derivatives
with respect to the variable or alpha. The resulting formula may
be set to zero for computation of the maximum. The values of the
coefficients which maximize this likelihood function are then
used as the coefficients in the formula. More specifically, the
Gauss-Newton maximum likelihood algorithm utilizing step halving
with the Gauss increment was utilized. This method is robust
to the violation of several ordinary least squares regression
assumptions.
The fitting strategy employed involved selection of
candidate predictors from significant crosstabulations with the
trichotomized dependent variable, testing the null model, the
main effects model, a model with two-way interactions, and models
with higher order interactions. The testing a model with higher
order interactions was obviated by the lack of improvement in the
fit when the first order-interactions were all found to be non-
significant. Higher order interactions (three-variable) were
tested but encountered an excessive number of empty cells in the
crosstabulations formed. Hence, these interaction models were
abandoned. Then the model was fine-tuned by collapsing
categories of the predictor variables to improve the fit or the
correlation between the observed computation of the dependent
variable from the formulated equation (observed logit scores)
with the predicted logit scores.
This final recoding of scores yielded the same variables
coded as follows: Education was coded as having had high school
dropout or not. Whether the mother died before the patient was
18 years old, whether the patient has or has had physical or
sexual abuse, and whether the patient has or has had a problem
with drug abuse are dummy variables. Together these variables
with the maximum likelihood estimation, yielded the following
formulae.
Logit1 = -3.021 + 1.104*Education + 2.144*MotherDiedEarly
+ 1.132*PhysicalOrSexualAbuse - 1.598*DrugProblem
Logit2 = -4.178 + 1.104*Education + 2.144*MotherDiedEarly
+ 1.132*PhysicalOrSexualAbuse - 1.598*DrugProblem
For a more elaborate assessment of these formulae the reader
may consult Table 8, where the B (the regression coefficient in
the ordinal logit analysis), its standard error, the chi-square
or Wald statistic (the square of the ratio of the regression
coefficient to its standard error), and the significance level of
the Wald statistic (which is distributed as a chi-square).
Table 8
Cumulative Logit Analysis
| Variable | B | Standard Error | Chi-square (Wald) Statistics | p |
|---|
| Alpha1 | -3.021 | .600 | 25.29 | 0.000** |
|---|
| Alpha2 | -4.178 | .647 | 41.70 | 0.000** |
|---|
| Education | 1.104 | .463 | 5.68 | 0.017* |
|---|
| Mother Died Early | 2.144 | .432 | 24.68 | 0.000*** |
|---|
| Physical or Sexual Abuse | 1.132 | .380 | 8.86 | 0.003** |
|---|
| Drug Problem | -1.598 | .608 | 6.90 | 0.009** |
|---|
Significance Levels: * p < .05; ** p < .01; *** p < .001
-2 log Likelihood = 292.42 , p = 0.00;
Main Effect Model Likelihood Ratio Chi-Square = 59.33, 4 df, p = 0.00;
-2 log Likelihood of Model Chi-Square with Variables Included = 233.09
Somers' Dxy = .608 Gamma = .674
If we wished to know the influence on the odds of the
gambling severity problem on the part of a unit increase in the
particular variable, controlling for the influence of all other
variables, we could examine the change in odds ratios in Table 9.
The change in the odds of being either of the top two levels
(over being in the lower category) of the gambling severity
problem as a result of a unit change in the respective variable
is provided under the Logit 1 listing of Table 9. The change in
the odds of being in the high category as a result of a unit
change in the respective independent variable is given in the
Logit 2 listing of Table 9. The most powerful influence on the
change of the severity of the gambling problem is the death of
the mother before the patient was 18 years of age. The second
most powerful influence on this odds ratio is the physical or
sexual abuse of the patient. The education of the patient has
approximately the same magnitude of an effect on the odds ratio,
while the existence of a drug problem is the only other addiction
which has a negative influence on the severity of the gambling
problem. That is to say, the existence of such a problem is
inversely related to the increased severity of the gambling
problem.
Table 9
Regression Coefficients (Bs) and Odds Changes
for Each Variable of Two Cumulative Logit Models
| Variable | Regression
Coefficients
for Logit 1
-- Moderate
or Very
Severe
Compared to
Low Severity | Regression
Coefficients
for Logit 2
-- Very
Severe
Compared to
Moderate or
Low Severity | Partial Odds
or exp(B)
(where B
equals the
Regression
Coefficient) |
|---|
| Intercept | -3.021 | -4.178 | |
|---|
| Education | 1.104 | 1.104 | .332 |
|---|
| Mother Died Early | 2.144 | 2.144 | 8.534 |
|---|
| Physical and Sexual Abuse | 1.132 | 1.132 | 3.102 |
|---|
| Drug Problem | -1.598 | -1.598 | .202 |
|---|
The strength of the model is indicated by the Somers' D of
.608 and a Gamma of .674. These coefficients represent the
correlation between the observed a predicted values, corrected
for ties and the correlation not corrected for ties,
respectively. These correlations, falling as they do on a scale
from 0 to 1, represent a reasonably good fit.
The relative strength of the variables in increasing the
odds of being in either of upper levels of the severity of the
gambling problem are, in decreasing order, the demise of the
mother before the age of 18, whether the patient was a high
school graduate, the experience of past physical or sexual abuse,
and lastly whether or not the patient has or had a drug problem.
The past or present existence of a drug problem was negatively
related to the increased severity of the gambling problem. Even
so, the odds change effected by a unit increase in drug problem
was small.
A Logistic Regression
Many persons find the cumulative logit analysis a little
arcane. They prefer to deal with a binary dependent variable,
coded in two values, such as a comparison of moderate through
heavy gambling severity to light gambling severity. The natural
log of the odds of having a moderate or heavy gambling problem to
a light gambling problem they might consider a simpler or more
elegant dependent variable. Since the proportion of cases in the
high severity level is not large, collapsing the high and medium
levels might seem reasonable. Such an analysis was run in three
different statistical packages with the following results.
When the gambling compulsiveness ratio is collapsed so that there
is the low level and the upper two levels, the predictive
capability of the model is improved, yielding a better goodness
of fit.
Thus, it is possible to formulate the relationship between
the logit and the significant variables so that the model is
consistent with the data. The aggregate size of the difference
between the predicted and the observed proportions in the
multiway classification system undergirding the model is not
statistically significant. That is to say, a very good fit
between the model and the data is obtained with the logistic
regression. The formulated model is:
Logit = -3.133 + 1.188*Education
+ 1.014*PhysicalOrSexualAbuse - 1.759*DrugProblem
+ 2.371*MotherDiedEarly
Table 10
Logostic Regression of a Binary
Gambling Severity Ratio
| Variable | B | Standard Error | Chi-
square (Wald) Statistic | p |
|---|
| Alpha1 | -3.133 | .649 | 23.28 | 0.000*** |
|---|
| Education | 1.188 | .503 | 5.58 | 0.018* |
|---|
| Mother Died
Early | 2.371 | .450 | 27.73 | 0.000*** |
|---|
| Physical or Sexual
Abuse | 1.014 | .431 | 5.80 | 0.016* |
|---|
| Drug
Problem | -1.789 | .626 | 8.01 | 0.005** |
|---|
Significance Levels: * p < .05; ** p < .01; *** p < .001
-2 log Likelihood = 213.56 , p = 0.000***;
Model Likelihood Ratio Chi-Square = 61.219, 4 df, p = 0.000***;
Goodness of Fit = 146.317, 161 df, p = .793
Somers' Dxy = .673 Gamma = .735
Table 11
Regression Coefficients (B's) and Odds Ratios Changes (Increases) for Each Variable of the Binary Dependent Variables Model in the Maryland Patient Model
| Variable | B | Partial Odds = exp(B) |
|---|
| (Intercept) | -3.133 | 4.359 |
|---|
| Education | 1.188 | 3.281 |
|---|
| Mother Died Early | 2.371 | 10.708 |
|---|
| Physical or Sexual Abuse | 1.014 | 2.757 |
|---|
| Drug Problem | -1.789 | .172 |
|---|
The relative effect of each of the above variables on the
odds of a patient having a moderate to severe gambling problem
compared to having a light problem may be found in Table 11.
Clearly, the early death of the mother is the most powerful among
these. Next is the existence of physical or sexual abuse in the
past or current life of the patient. The least powerful, among
these significant effects, is that of the past or present
existence of a drug problem.
How well does this estimated model fit the data? A good
model is one that results in a high likelihood, which means a
small value for the log of the likelihood. If an estimated model
fits the data perfectly, the likelihood is 1 and -2 times the
natural log is 0. If the degrees of freedom are computed with
N - p, where N = the number of observations and p = the number of
predictor variables, then the significance level of the model is
greater than .05 and the model would appear to be consistent with
the data. The -2 log likelihood listed below compares this model
with the perfect model and finds no significant difference. If
one uses the goodness of fit statistic, which is equal to the sum
of the residuals divided by P(1-P) where P is the predicted value
of each observation, and the same formula for the degrees of
freedom, the significance level remains greater than .05 and the
model still appears to fit well.
Table 12
Goodness fo Fit Statistics for the Binary Model
| Chi-square | df | Significance |
|---|
| -2 * log (Likelihood) | 152.339 | 161 | 0.675 |
|---|
| Goodness of Fit | 146.137 | 161 | 0.793 |
|---|
| Model Chi-square | 61.219 | 4 | 0.000 |
|---|
| Improvement | 61.219 | 4 | 0.000 |
|---|
There is a significant improvement in fit between the null
model and this binary logistic model as indicated by the model
and improvement chi-square in the above table. With this simpler
model, we obtain a classification table that indicated that the
false positive rate of prediction for this model was 34.4 percent
while the false negative rate was 14.7 percent. The sensitivity
of the model was 73.7 percent, the specificity of the model was
79.8 percent. The total correct predictions was 77.7 percent.
These statistics indicate a substantial proportion of explanation
of the gambling severity on the part of the patient. The Somers'
D and the Gamma coefficients are fairly large as well. With a
nonsignificant goodness of fit, our model predicts without an
aggregate significant difference to the observed severity of the
gambling problem on the part of these patients.
Table 13
Classification Table for Logistic Regression Model
| | Predicted Severity |
| | Lower | Higher | Total |
Observed
Severity | Lower | 87 | 22 | 109 |
| Higher | 15 | 42 | 57 |
| Total | 103 | 64 | 166 |
Correct Prediction Rate 77.7%
Sensitivity: 73.7% False Positive Rate: 34.4%
Sensitivity: 79.8% False Negative Rate: 14.7%
Discussion
The problem severity ratio has been constructed to indicate
the extent of the problem. If the ratio is very low, then there
is no great problem. When this ratio is low enough, the gambler
can recoup his losses without too much ado. He merely has to
stall or to arrange for installment payments to pay his debts.
When this ratio becomes high enough, it becomes increasingly
difficult for the gambler to pay his debts on time. This
encumbers him in a number of ways. It deprives him of reserves
with which to take care of unforeseen problems and expenditures
incurred as a result. It overloads his reserves with pressing
demands for payment, jeopardizes if not destroys his credit, and
deprives him of needed resources with which to pay for the
necessities of life. The higher this ratio the more
insurmountable this problem becomes. All of this seems common-
sensical enough.
Mark Nicolich has argued that one should use a strictly
defined measure of gambling behavior, such as the number of bets
that one places within a period of time. Nicolich maintains that
the size of the gambling debt is not a measure of gambling
behavior per se. He demurs that the problem severity ratio is a
measure of how poor the bettor gambles. But the seriousness of
the gambling problem is not captured by the number of bets that a
gambler has made in a year. Up to a point, the number of bets
made indicates social and/or legitimate recreation. Such recre-
ation may not be problematic at all. Recreational wagering may
be different from problematic compulsive gambling. The serious-
ness of the gambling problem is reflected by the inability of the
gambler to stop in face of mounting losses. Given the variables
included in this analysis, the seriousness of the problem is best
indicated by the extent of the gambling debt relative to access
to resources with which to discharge that debt.
The meaning of the equation characterizing the severity of
the gambling problem indicates that existence of past physical
abuse of the patient and the early death of the mother may
provide a harsh and non-supportive emotional environment for the
patient. Education may lead to greater gambling given this past.
But there is no evidence of alcoholic co-addiction. Alcoholism
was not found to be significantly related to the severity of the
gambling problem. There may be some cross or co-addiction with
other substance abuse. Drug abuse was found to be inversely
related to the severity of the problem. Yet this equation was
found to be far from complete. Somers' D was .668 so that the
R square analog would be approximately .454. This means that
other relevant variables were not included in this analysis.
The model fitting strategy began with the null model, then
the main effects model, and the two-way interactions. Because
the two-way interactions were non-significant, higher order
interactions were not attempted. The resulting model was a main-
effects model noted above.
External Validity
These data represent the patient population, not a mere
sample, in the State of Maryland during the past few years. The
respondents were those individuals receiving treatment in the
clinical programs specifically addressing the problems of
compulsive gamblers in Maryland. The fact that a random sample
was not taken is not a problem with population data. Inferences
are not being made to a larger population but rather to the
relationships that are posited in the study. The descriptive
character of the problem will be useful in providing policy
makers with a better understanding of the problem with which they
must deal. The generalizing of these findings to larger
populations must be made with a tentative orientation that is
dependent upon replication of this study in other states. This
study may make a contribution to this evolution of our under-
standing of the problem. Nor should one believe that the
findings of this study will remain immutable as the circumstances
and cultural milieu evolves. There will be need for repeated
analyses to discover which characteristics change and at what
rates. It is hoped that this study will contribute toward
attainment of these objectives.
Co-addiction, Cross-addiction and General Therapy
From these data, there appear to be some evidence of co-
addiction or cross-addiction among compulsive gamblers. When
almost one half of the gamblers have or have had an alcohol
problem, one might have reason to believe that this possibility
should be investigated further. Without clear distinction as to
what addiction ran concurrent to what other addiction, it is
difficult to tease out the prevalence or incidence of co-ad-
diction. This evidence suggests serial rather than co-addiction.
Without a control group and random assignment we cannot know
that there is a significant difference between the mass public
and the compulsive gambler's alcohol or drug addiction rate. To
be sure, there is no evidence in this sample that the existence
of the alcohol problem is significantly related to the severity
of the gambling problem. While the bivariate relationships
indicated the possibility of a significant negative relationship
between alcohol problems and the gambling problem severity ratio,
the alcohol problem variable turned out, in the multivariate
model, to be a byproduct of other relationships, and not to be
directly significantly related to the gambling severity ratio.
While one-fourth of the compulsive gambling patients have had or
do have a drug problem, the existence of this problem is signifi-
cantly negatively related to the gambling severity ratio. The
prevailing evidence has been explained as an indication of a
quest for clarity and control of mind over situation or events.
Therefore, we do not have enough solid evidence to be sure that
there is a substantial co-addiction problem. Nor can we say that
there is no cross-addiction problem with complete assurance. If
we were to look at those in treatment for alcoholism and those in
treatment for drug abuse, there may within those samples be
insignificant and insubstantial cross-addiction. Until those
possibilities are explored, it would appear as though compulsive
gamblers need special counseling that may not be properly
provided by counselors who specialize in treating other addictive
behaviors.
Substantive Recommendations
There are substantive recommendations and methodological
recommendations. First, the substantive recommendations will be
made and second, the methodological recommendations will be made.
It is interesting that the favorite form of gambling is the
track, whether horse or dogs. From the review of the socio-
economic status of the gambling patient, it is clear that the
background of these individuals is modest in income and educa-
tional attainments. A fifth of them have pending legal problems.
More than a fourth of them have attempted suicide. Almost one-
half of them have been outpatients before, while more than a
fifth have been inpatients. Yet more than one-fourth of those
queried on this subject do not have insurance. These individuals
appear to be seriously in need of treatment.
The model of the severity of the gambling problem is a
secondary analysis utilizing variables that not had been designed
for this purpose. Although the models fit the data, lack of a
more perfect fit and better correct prediction rate indicates
that the model may be improved. That is, not all of the relevant
variables that explain the severity of the gambling problem have
been included in the analysis. This subject should be investi-
gated more thoroughly with a view toward explaining and predict-
ing more accurately the severity of the gambling problem. Better
predictability here should point the way toward better therapy.
That poker machines have become so popular as to rank
second, next to the racetrack, among the most favored kinds of
gambling, is a little surprising. Such popularity shows the
potential of new computerized machinery to entice the compulsive
gambler. These machines are in bars, stores, and restaurants.
Among the second most favored kind of gambling, the
lottery/numbers is found to be the most salient. Because very
little numbers activity has been reported, this type of gambling
is inferred to be primarily lottery participation. This is
indicative of a new kind of addict that is being bred by state
government activity. Some might say that it is merely another
form of opportunity that is being seized by the would-be
entrepreneur and gambler. More than likely, it is a combination
of the personality and the situational opportunity that is taking
hold here.
The state may derive from the lottery a considerable amount
of income. States have conducted lotteries for this purpose
since colonial times. Historically, lotteries have brought in
substantial sums of funds. Although it might be politically
unpopular to oppose the lottery altogether, there may be means of
dealing with the problem created by this form of legalized
gambling.
If the state creates this form of opportunity, is the state
morally or ethically obligated to provide for the treatment of
the addicts that are created, nourished, and ruined amidst the
flourishing of these circumstances? Should a part of the pro-
ceeds be given to the treatment programs or to insurance compa-
nies as a subsidy to cover treatment costs for such individuals?
If the state were to levy a tax on the wagering, it could use the
proceeds from that tax for funding the treatment of those
certified as addicted. The tax could be levied on the type of
betting in accordance with the proportions spent by those in
treatment for pathological gambling on those forms of gambling.
Perhaps the way to tax the betting would be in proportion to the
amount spent on different types of gambling. It is possible to
conduct a poll of the patient compulsive gamblers to see how they
have spent their money wagering each year. Based on the relative
percentages, the state could tax the forms of gambling in order
to provide funds to cover treatment of the certified victims of
this pastime.
Contrary to popular belief, these data provide very little
evidence of co-addiction. This does not mean that there is no
co-addiction. The model of the gambling severity problem indi-
cates that alcoholism is not significantly related to the severi-
ty of the gambling problem at all. It furthermore suggests that
drug abuse may be or have been negatively related to the gravity
of this problem. Even if there were substantial co-addiction,
for these tendencies to be operating in different directions
points to the possibility of different mechanisms at work in the
personalities of those who may be or may have been serially
addicted or co-addicted. We need to investigate the differences
between present or past abuse more carefully and to explore the
subject of co versus serial addiction more carefully. In this
survey, the questions were worded so that it was not possible to
completely distinguish past from present abuse. In future
surveys, this distinction should be made possible and the
question of co- versus serial addiction should be delved into
more deeply. Times of onset and times of cessation should be
included so that the full nature of serial and co-addiction may
be accurately examined. In that way, these findings may point
toward whether there should be a general or a specialized
treatment program for persons so afflicted.
Methodological Recommendations
With the lack of the control group, the percentages by
themselves only serve to characterize the patient gambler. They
do not serve as the basis for a comprehensive analysis. For this
reason, whenever there appears to be a conflict between the
characterization of the gambler by the percentages or a charac-
terization of the severity of his problem by a correlation
analysis, the latter, with its tests of significance, should be
given priority. But the bivariate analysis, when in conflict
with the multivariate analysis, should yield to the latter. The
multivariate approach controls for the antecedent and intervening
variables more than does the bivariate approach. To more fully
develop this approach, a logistic path analysis of these
variables may be undertaken later.
Bert Holland argued that the restriction of range of the
sample to compulsive gambling patients may reduce the signifi-
cance and magnitude of the coefficients. If only part of a
real correlation is measured because of constriction of the scope
of the sample, then only a smaller part of a more substantial
correlation might be detected. If this were the case, a negative
relationship between drug abuse and the severity of the gambling
problem might be understated. That is, a more powerful negative
relationship might indeed hold in reality. The same restriction
of range of the sample might cause an eclipsing of a minor
relationship between alcoholism and severity of the gambling
problem. These possibilities should be examined in further
research. For the time being, what stands out is the lack of a
significant relationship between alcoholism, on the one hand, and
the gravity of the gambling problem on the other. Also salient
in these findings is the negative relationship between drug abuse
and seriousness of the gambling problem.
In the future, it would be of service to the State of
Maryland and its citizens were the treatment programs to commis-
sion envoys to join together in a task force with a view toward
deciding upon the basic questions and answer categories to be
included on their patient surveys. Questions should be designed
so that it is clear to the respondent that the answer categories
are distinct yet comprehensive. If the gambling preferences are
asked, casino activities should be clearly distinct from similar
non-casino activities. These categories must be mutually exclu-
sive and collectively exhaustive, if statistical tests of signif-
icance are to be properly performed. It goes without saying that
this would not preclude the treatment programs from asking their
own questions in the way that they wish. But the collective
decision of how to get what information in order to answer which
questions and the consistent compliance with that decision would
make the job of writing the program to do the analysis as well as
the job of analyzing the findings easier and more likely to
succeed.
Underlying concepts may be analyzed if multiple indicator
methods are applied. If multiple observed indicators of the same
concept should be used in later surveys, the underlying concepts
linked to these indicators may be analyzed with greater reliabil-
ity. Scales with improved reliability may be used to tap these
underlying constructs. More in-depth analyses may be conducted.
Efforts should be undertaken to obtain large sample sizes to
maintain the power of the statistical tests and to allow for the
use of many advanced large sample statistical tests. Cooperation
among the treatment centers would facilitate this effort.
Future research should include a control group to provide
baseline estimates. Studies of others in treatment and self-
help groups should be undertaken to see whether there is evidence
of compulsive gambling among those persons. If the patients
interviewed could be interviewed again at a later date as part of
a panel analysis, more fruitful analyses would be possible. A
more longitudinal perspective could shed more light on the
etiology of this problem and the efficacy of their different
treatment modalities.
Cooperation among the treatment centers should be encour-
aged. Once the collective decision to use a particular coding is
made, the centers should seriously try to adhere to the consis-
tent coding and question usage. In these ways, a collective
research effort could yield more and better information on this
important subject.
ACKNOWLEDGEMENTS
While I take full responsibility for the assumptions,
inferences, judgments and errors in the foregoing analysis of
socio-demographic characteristics, financial circumstances and
psychological attributes of the patient population of the State
of Maryland under treatment for pathological gambling, I am
grateful to many persons who assisted me in this work.
I express my appreciation to Edi Franceschini, Deputy
Director of the Academic Computing Facility ("ACF") of the
Courant Institute of Mathematical Sciences of New York
University, for providing permission for this analysis to be
performed on the IBM 4341 mainframe.
To Drs. Valerie Lorenz, Robert Politzer, Mark Nicolich, and
Winthrop Munro, I express my grateful thanks for helpful comments
and criticism.
I also thank Bert Holland, a colleague at the Academic
Computing Facility of the Courant Institute, for very
constructive suggestions.
To Clifford C. Clogg, current editor the Journal of the
American Statistical Association and Professor of Sociology at
Pennsylvania State University, I am very grateful for very
helpful suggestions. But I must give thanks to the works of Leo
Goodman for blazing the intellectual path I had to take.
To Yolanda Ramirez, also of ACF at Courant, I express my
appreciation for her entering the Taylor Manor Data into machine-
readable form on the IBM OS/MVS system.
I am grateful to James Gray for letting me use the "out of
town computer equipment" with which I wrote much of this
analysis.
Previous part of this document.
Next part of this document.
This page has been accessed 
times. Counter courtesy of Web-Counter.