Predicting the outcomes of Gaming Referenda

1
Predicting the outcomes of Gaming Referenda

• Ercan Sirakaya
• Texas AM University, USA
• Dursun Delen
• Oklahoma State University, USA
• Hwan-Suk Choi
• University of Guelph, Canada

APacCHRIE Sixth Biennial Conference on Tourism
in Asia May 27-29, 2004 Phuket, Thailand
2
Outline

• Introduction
• Significance
• Purpose of the Study
• Methodology
• Model
• Results
• Discussion

3
Introduction

• Many small communities throughout the US are left
  in backwash despite extensive infusions of state
  and federal funds
• the development of a community depends on
  diversification of its economic base
• communities with heavy tourism activity may be
  able to overcome the problem of seasonality and
  economic instability by diversifying their
  product and adding new products for tourist
  consumption
• Las Vegas and Atlantic City are classic examples
  of the ability of gambling to support year-round
  tourism

4
Introduction cont

• many communities offer a variety of incentives
• One example can be found in the strong support
  for abolition of laws prohibiting gambling
• Proponents of gaming argue gaming-tourism can
  create significant positive long-term
  sociocultural-benefits
• better living conditions,
• more leisure opportunities,
• stronger cultural identity, etc.) and
• economic-benefits
• improved job opportunities,
• better standard of living,
• increased disposable income,
• increase in tax revenue, and more) for many
  communities (Roehl 1999).

5
Introduction cont

• Opponents argue
• gambling activities defy religious beliefs and
  work ethics,
• invite political corruption,
• swindling, money laundering and organized crime,
• erode traditional family and societal values and
  responsibilities,
• perpetuate activities of illegal and criminal
  nature, and
• instigate irresponsible behavior (DiClemente,
  Story and Murray. 2000 Giacopassi, Nichols and
  Stitt 1999 Gupta and Derevensky 1998 Roehl
  1999 Winters and Anderson 2000).

6
Status of Gambling in the US

• total revenues 68.7 billion, a 158 increase
  over the 1990 figures (American Gaming
  Association 2002a, 2003).
• In 2001, direct employment in the casino
  industry 364,000 with wages totaling 11.5
  billion.
• 3.6 billion in taxes revenue

7
Introduction cont

• There are limits to the public's acceptance of
  gambling, and attempts to pass gambling laws may
  fail and/or create divisions within communities
  unless the factors behind public
  support/opposition of gaming are understood
• There is a need to assess the determinants that
  differentiate a community that passes a gambling
  ballot from a community that strongly opposes
  gaming.

8
Significance

• Ascertaining these determinants would help
  identify communities with high acceptance
  probability before a gaming ballot comes to a
  vote thereby preventing potential conflicts and
  divisions between the industry and communities.

9
Past Studies

• Few studies in tourism have tested models to
  predict community support for commercial gaming.
  
• Silverberg and Ulbrich (1996) used secondary data
  collected in South Carolina (US) to build a model
  for predicting ballot outcomes.
• This effort, though commendable, suffered from a
  lack of conceptual framework and
  multicollinearity problems of variables.

10
Purpose

• The present study was undertaken in order to gain
  an understanding of the factors affecting both
  legalization and prohibition of gaming activities
  by using a computational tool (ANN).

11
Conceptual Model
12
Artificial Neural Network Model

• Artificial Neural Networks has been shown to be
  superior when compared to other forecasting
  techniques in analyzing data with
  multicollinearity problems (Uysal and Roubi
  1999).
• ANN are capable of predicting new observations
  (e.g., ballot outcomes) from other observations
  (on the same or other variables) after executing
  a process of so-called learning-from-
  existing-data (Haykin 1998).

13
Study Method

• Data both primary and secondary data collection
  techniques.
• e-mail survey of all 50 state-election offices
  and related divisions.
• 64 ballots of twenty-two states proposing gaming
  ballots between the years 1919 and 1998 were
  extracted from the information received.
• only the results of gambling related ballots
  proposed during 1990s were used (14 states with
  24 ballot results California, Colorado, Florida,
  Georgia, Louisiana, Nebraska, New Jersey, North
  Dakota, Oklahoma, Ohio, South Dakota, Texas,
  Utah, and Washington)

14
Variables Used

• Ballot Type I (Gambling) Unary encoding (UE)
• Ballot Type II (Wagering) Unary encoding (UE)
• Percent population voted
• Medium family income
• Percent population church members
• Percent population male
• Poverty level
• Unemployment rate
• Percent population minority (non-white)
• Percent population older than 45
• Metropolitan statistical area (MSA) Unary
  encoding (UE)
• Not MSA Unary encoding (UE)

Note dependent variables are percent nay votes
and percent yes votes. The data set includes
1287 records from 1287 counties.
15
Steps in Neural Network Modeling

• Step 1. Observations with missing and/or null
  values were removed from the dataset.
• Step 2. Observations (rows) in the data set were
  randomized among themselves.
• Step 3. The randomized data set (1287 records)
  was split into three separate data files
• training data, cross-validation data, and testing
  data.
• the data set was split into three parts using the
  following percentages 60 for training (773
  observations), 20 for cross validation (257
  observations), and 20 for testing (257
  observations).

16
Steps in Neural Network Modeling

• Step 4. Sensitivity analysis was performed to
  determine the cause and effect relationship
  between the inputs and outputs of a trained
  neural network model.
• Step 5. To establish reliability, data were
  converted to binary classification problem

17
ANN Architecture Specifications

• Multi Layered Perceptron (MLP) with one hidden
  layer was used.
• The input layer included PEs for 12 independent
  variables with 10 PEs in the hidden layer.
• The output layer included only one dependent
  variable nay- or yes-votes in percentages.

18
ANN Architecture Specifications
19
NN Training/Testing Results

• Two neural network models were trained one for
  percent- nay-votes (Model-1) and another for
  percent-yes-votes (Model-2) using data from 1287
  individual counties in the U.S.
• approximately 82 accuracy
• the calculation of the accuracy of the model
  prediction in terms of what the actual vote (for
  versus against) would turn out to be for both
  models

20
NN Training/Testing Results

• Out of 257 test data records, in Model-1
  (percent-nay-votes) the trained neural network
  predicted the 201 counties against vote-turn-out
  accurately, and
• in Model-2 (percent-yes-votes) the trained neural
  network predicted 198 counties for vote-turn-out
  accurately.
• This finding indicates that in each of the two
  models, on average, the neural network models
  predicted the vote outcome accurately for four
  out of every five counties on the data set that
  they have never seen.

21
Sensitivity Analysis on Neural Network Model

• Sensitivity analysis is a method for extracting
  the cause and effect relationship between the
  inputs and outputs of a trained neural network
  model (Davis 1989)

Figure 4. Sensitivity Values for percent Yes
presented in a Column Plot
22
Sensitivity Analysis on Neural Network Model
Figure 5. Sensitivity Values for Percent Nay
resented in a Column Plot
23
Implications

• Theoretical standpoint
• a significant relationship between abolition of
  prohibitionary laws for gaming and
  socio-demographic and geographic variables.
• increased proportion of minority populations
  within a geographical space gave rise to
  legalization of gaming
• proximity to population centers (operationalized
  as MSAs) was an important variable for
  legalization of gaming (confirmed Lanes research
  1993). A lottery adoption study conducted by
  Filer et al (1988) also reported similar results.
  

24
Implications

• Theoretical standpoint
• As with earlier studies (Curriden 1993 Shapiro
  1996), increased church memberships within the
  general public was also found to be a sensitive
  variable to changes in voting behavior toward
  gaming.
• This study supports the findings of Martin and
  Yandle (1990) who estimated the probability of
  the occurrence of lottery in a state. Their
  results suggested that the percentage of a
  states population that is Baptist was negatively
  related to the likelihood of occurrence of a
  lottery.

25
Implications cont

• Practical standpoint
• Be able to assess communities with high
  probability of passing gaming ballots (e.g., with
  the help of GIS).
• resources devoted to promoting gambling can be
  efficiently utilized and potential conflicts that
  may arise between the industry and the
  communities may be prevented before a promotional
  campaign is initiated.

26
Conclusion

• one can speculate on the directionality of the
  sensitivity values. For example, areas where
  there is a proportionally higher church
  membership, gaming proposals, regardless of their
  gaming type, are expected to be rejected.
• Increases in minority populations and gaming
  proposals in metropolitan counties versus rural
  areas can result in the acceptance of wagering
  laws.

27
Limitation

• unlike regression analysis, sensitivity analysis
  does not allow the researcher determine the
  direction of the effects.
• Artificial neural networks are known as a
  black-box approach to solving complex problems.
  Though empirical results are generally favorable
  compared to other techniques, they lack the
  theoretical explanations of independent
  variables.

28
Further Studies

• examine voting outcomes based on various types of
  gaming.
• variables expected to explain the acceptance of
  low impact gaming (e.g., bingo, lottery, horse
  racing) might be very different than the
  variables used in predicting casino gambling.
• Other lines of research should be directed toward
  improving the forecasting ability of the models.
  Including other variables, newly developed
  techniques such as genetic algorithms or rough
  sets and/or collection of primary data can help
  improve predictive power of current models.

29
ANY QUESTIONS?
No to war, and human suffering.