Demographic Characteristics and Number of Extramarital Affairs

19004 words (76 pages) Essay

18th May 2020 Statistics Reference this

Disclaimer: This work has been submitted by a university student. This is not an example of the work produced by our Essay Writing Service. You can view samples of our professional work here.

Any opinions, findings, conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of


In this report, the relationship between the demographic characteristics of the individual and the number of extramarital affairs a person had in one year is discussed. The classical Poisson, quasi-Poisson, negative binomial regression models for count data, which belong to the family of generalised linear models, are introduced and applied to the “Affairs” dataset. Incorporating the over-dispersion and excess zeros problems, hurdle and zero-inflated regression models is also considered for modelling. Stepwise regression is used as a strategy for selecting variables. Moreover, the results of estimating the coefficients for the model are compared to find the model, which can fit the “Affairs” data best. As the measurement scales discussed in the report, it is better to treat the response variable “affairs” as ordinal data rather than count data. Hence, ordinal logistic regression is applied to the dataset in the advanced chapter.

Keywords: GLM, Poisson model, negative binomial model, hurdle model, zero-inflated model, ordinal logistic regression, stepwise regression.

1. Introduction

Although there are so many articles about extramarital affairs published in the popular press and media, scientific research on this sensitive topic developed slowly (Glass et al.,  2004). In a previous study, evaluating the psychobiological correlates of extramarital affairs, Fisher et al. found that when compared with men without extramarital affairs, a higher androgenisation is present in men with infidelity (Fisher et al., 2009). Moreover, unfaithful subjects showed a higher risk of conflicts within the original family and higher stress at work (Bandini et al., 2010). As extramarital affairs seem to be an alpha-male activity (Fisher, 2012), most of researches paid more attention to males than females. However, marital infidelity is not only about males. So the purpose of this report is to test if there is a relationship between the demographic characteristics (include gender) and the number of having extramarital affairs. And then, choose a model that can fit the data best. There are 601 observations on 9 variables for the study, and the data are available from the magazine surveys leading by Psychology Today and Redbook in 1969 and 1974 (Fair, 1978).

The extent of having extramarital affairs is by no means small. In the two surveys used in this study, 27.27%

men and 22.86% 

women had had at least one extramarital relationship during the year of the study (Fair, 1978). Therefore, some generalised linear models will be fitted with the variable “affairs” as response and the other eight variables as explanatory variables. Furthermore, variable selection and model comparison are implemented to find the best models for this problem. Besides, incorporating the measurement of scales (details in Chapter 2), the outcome variable “affairs” is treated as count data in Chapter 4, and it is treated as ordinal data in Chapter 5. As a result, the purpose of the count data regression in this report is inference, and the purpose of ordinal logistic regression is prediction.

Get Help With Your Essay

If you need assistance with writing your essay, our professional essay writing service is here to help!

Find out more

The structure of the project is as follows. Chapter 2 describe the data and Chapter 3 introduced the methods. Chapter 4 contains a standard analysis of count data regression models: explanatory analysis, model verification, model comparison and model selection. Chapter 5 includes ordinal logistic regression and summary of conclusions, discussion of limitations and the further task is discussed in Chapter 6.

2 The Data


This project uses data from Fair (1978). The data are gathered from two magazine questionnaires on sex in the 1960s and 1970s. Psychology Today (PT) conducted the first survey in 1969, and the second survey was conducted in 1974 by Redbook (RB). The questionnaires included the question about the number of extramarital affairs a person had in one year as well as some questions about other aspects of sexual behaviour. That is, various demographic and economic characteristics of the individual (Fair, 1978).

As the methods of statistical analysis depend on the measurement scales of the response (Dobson, Annette J; Barnett, Adrian G, 2008), the measurement scales of variables need to be discussed first. The four fundamental levels of measurement scales (nominal, ordinal, interval and Ratio) provided by Entry (Entry, 1990)  are used in this report. The definitions are given below:

  1. Nominal scale: a scale used for labelling variables into different categories and does not involve a quantitative value or order.
  2. Ordinal scale: a scale used when there is some natural order or ranking between the categories. For example: diastolic blood pressures groups as 70, 7190, 91110, 111130,  131 mmHg
  3. Interval scale: a scale used when rank-ordered data categories may be placed equidistantly on a scale with an arbitrary zero point. Besides, the difference between any two values is meaningful.
  4. Ratio scale: a scale used when equidistant rank-ordered data categories may be placed on a scale with an absolute zero point.

Nominal and ordinal data are categorical variables as well as interval and ratio scale are continuous variable (Dobson, Annette J; Barnett, Adrian G, 2008). Table 1 shows the data description for the “Affairs” dataset.

From the table, it is clear that all predictor variables in the dataset are categorical variable. Besides, it is better to treat the response variable “affairs” as an ordinal categorical variable rather than a count variable. Since the value recorded for the response variable “affairs” ( y

) is not always the real number of times of having affairs. For example, when y=12

, it means that the observation cheated every month, every week, or even every day. However, the topic of the dissertation belongs to the theme “Count data regression models”, so the response variable “affairs” is treated as count data in Chapter 4 (Standard analysis of the data) and as an ordinal variable in Chapter 5 (Advanced chapter).

Table 1 Data description

Variable Name


Description of Variable

Values of Variable




How often engaged in extramarital affairs during the past year?

0 = none, 1 = once, 2 = twice, 3 = 3 times, 7 = 4-10 times,  12 = monthly/weekly/daily





0 = female, 1 = male





17.5 = under 20, 22 = 20–24, 27 = 25–29, 32 = 30–34, 37 = 35–39, 42 = 40–44, 47 = 45–49, 52 = 50–54, 57 = 55 or over




No. of years married

0.125 = 3 months or less, 0.417 = 4–6 months, 0.75 = 6 months-1 year, 1.5 = 1–2 years, 4 = 3–5 years, 7 = 6–8 years, 10 = 9–11 years, 15 = 12 or more years.




Are there children in the marriage?

0 = no, 1 = yes




How religious

1 = anti, 2 = not at all, 3 = slightly, 4 = somewhat, 5 = very




Level of education

9 = grade school, 12 = high school graduate, 14 = some college, 16 = college graduate, 17 = some graduate work, 18 = master’s degree, 20 = Ph.D.





1– 7, according to Hollingshead classification  (reverse numbering)




How rate marriage

1 = very unhappy, 2 = somewhat unhappy, 3 = average, 4 = happier than average, 5 = very happy


3. Description of the methods

In this section, the theory and its implementation in R for some basic count data regression models and their zero-augmented extensions are described (see Table 2 for an overview). In addition, the other methods used in this report are also showed in this Chapter.

The Poisson, negative binomial models and ordinal regression model (OLR) are described in a generalized linear model (GLM) framework. The Poisson regression model is provided in R by the glm()

function (Chambers JM, Hastie TJ, 1992) in the stats package. The glm.nb()

and polr() 

functions are in theMASS package, respectively (Venables WN, Ripley BD, 2002). Besides, rather than ordinary least squares (OLS), the methods above use maximum likelihood estimation (MLE) to estimate the parameters, which means that they are depended on large-sample approximations.

Poisson regression is often used for modelling count data. However, as count data sets sometimes show over-dispersion (or have a large number of zeros), the classical Poisson regression model for count data does not always fit the data well. To solve the over-dispersion problem, quasi-Poisson model and negative binomial (NB) regression model, which including additional shape parameter can be considered (Achim Zeileis, Christian Kleiber, Simon Jackman, 2008). Another two models, which can also solve these two problems, are hurdle and zero-inflated model. The hurdle and zero-inflated extensions of these models are provided by the functions hurdle()

and zeroinfl()

in package pscl (Jackman, 2008). The function polr()

in package MASS can fit the model when the outcome variable is treated as an ordered factor. The table below summarised the models considered in this report. Further information about the models and relative code in R can be found in the respective references.

Table 2 Methods of statistical analysis considered in the report


Explanatory Variables





Categorical & Continuous

generalized linear model


Maximum Likelihood

Negative Binomial (NB)

Maximum Likelihood


Zero-inflated Poisson (ZIP)

Maximum Likelihood

Zero-inflated NB (ZINB)

Maximum Likelihood


Categorical & Continuous

generalized linear model

Maximum Likelihood

3.1 Generalised linear model (GLM)

Generalised linear regression is helpful in the social sciences and economics (Achim Zeileis, Christian Kleiber, Simon Jackman, 2008). In this section, the framework, assumptions and the main idea of generalised linear regression models (GLM) are shown.

3.1.1 Assumptions of GLM

  1. The observations  Y1,Y2, ...,Yn

     are independently distributed.

  2. The dependent variable  Yi

    is from an exponential family (e.g. binomial, Poisson, multinomial, normal).

  3. A linear relationship between the transformed response (the link function) and the explanatory variables.
  4. The homogeneity of variance does NOT need to be satisfied.
  5. Errors need to be independent but NOT normally distributed.

(Hilbe, 2014) 

3.1.2 Main ideas of generalised regression modelling

The primary motivation for the model is to figure out the relationships between the number of having affairs and other aspects of characteristics. The classical generalised regression modelling process, which is given by the book “An Introduction to Generalized Linear Models” (Dobson, Annette J; Barnett, Adrian G, 2008), involves four steps:

  1. Specifying three components to models:
  1. The linkage – an equation linking the outcome and explanatory variables.
  2. Systematic Component – specifies the explanatory variables ( X1, X2,  Xk

    ) in the model and creates linear predictor; e.g., β0 + β1x1 + β2x2++βkxk


  3. Random Component – the probability distribution of outcome variable ( Y


  1. Estimating parameters in the models.
  2. Checking how well the models fit the data.
  3. Making inferences: for example, testing hypotheses and calculating confidence intervals for parameters.

Since different statistical analysis methods have different linkage, and statistical analysis methods depend on the measurement scales of the response and explanatory variables. The next section summarised the model frame for the GLM models considered in this report.

3.1.3 Model frame

Generalised linear models (GLMs) are generalised from the linear models, which describe the dependence of a scalar variable yi(i=1,,n)

on a vector of  xi

. The conditional distribution yi|xi

of response variable belongs to the exponential family, which can be written as:

                                         fy;θ=expaybθ+cθ+dy,                                            (1)


Where bθ

is the natural parameter and the functions ay, bθ, cθ 

and dy 

are known from which member of the family is used (Achim Zeileis, Christian Kleiber, Simon Jackman, 2008).

For example, for the binomial and Poisson distribution, ay=y

, the conditional mean and variance of yi 

are given below (Dobson, Annette J; Barnett, Adrian G, 2008):

E[y xi= cθbθ

Var[y xi= bθcθcθbθbθ3

The dependence of the conditional mean E[y xi= μi

can be described via

                                                                gμi= xiTβ,                                                                         (2) 


where g()

is link function and β

is the vector of regression coefficients, which are estimated by maximum likelihood (ML) method (Chambers JM, Hastie TJ, 1992). Treat the variable “affairs” as count data
Poisson model

The classical regression model used for modelling count data is the Poisson regression. The probability density function for the response y


fy;θ=exp(θ)θyy!,                                                        (3)

which is of exponential family.  Canonical link for Poisson model is g(θ) = log(θ)

resulting in a log-linear relationship between mean of the response y

and the linear predictors. Since the variance in the Poisson model is identical to the mean, over-dispersion, which means that the variance is bigger than the mean, may appear in practice (Hilbe, 2014). In R, Poisson model can be easily applied in the glm()

call just by setting  family = poisson


Quasi-Poisson model

One way of modelling over-dispersed count data is using quasi-Poisson model, which means using the mean and the variance function from the Poisson GLM, but leaving the dispersion parameter unrestricted (Jay M. Ver Hoef, Peter L. Boveng, 2007). Such that



Where λ>1

. In the equation, λ

is an over-dispersion parameter which can be estimated from the data. In R, the quasi-Poisson model can be fitted with the glm()

function, by simply setting  family = quasipoisson


Negative binomial models

Another formal way of dealing with over-dispersion is to assume a negative binomial (NB) distribution for  yi|xi

, which can be regarded as a gamma mixture of Poisson distributions. The probability density function of NB models can be written as:

fy;μ,θ=Γ(y+θ)Γθy!μyθθμ+θy+θ ,                                       (4)

Where μ

is the mean, θ

is a shape parameter and Γ

is the gamma function. For every  θ

, this fits equation (1) and thus is another particular case of the GLM framework (Jay M. Ver Hoef, Peter L. Boveng, 2007). In R, the coefficient β

can be calculated using the function glm.nb()

from the package  MASS

. Treat the variable “affairs” as ordinal data

If there is an apparent natural order among the response categories, then ordinal logistic regression models can be considered. Suppose Y

is an ordinal outcome with J

categories. Then  PYj

is the cumulative probability of Y

less than or equal to the j

th category  (= 1, 2, K  1)

. It is clear that PYJ=1

. The log odds (or the logit

) of being less than or equal j

th category is defined as


for  j=1, 2, J1  CITATION Bru06 2057  (Bruin, 2006)


Due to the parallel lines assumption in R system (which means that the intercepts are different for each category but the slopes are the same among categories), the ordinal logistic regression models have the following general form:


for  j=1, 2, J1

(McCullagh, 1980).

In R, the coefficient β

can be calculated using the function polr()

, and the ordinal logistic regression model is parameterized as


Where ηi=βi

and p is the number of η



in R is a kind of proportional odds logistic regression models, with proportional odds assumption. And to have the model return the Hessian which is used to get standard errors, Hess=TRUE

is specified.

3.2 zero-augmented model

In addition to over-dispersion, many count data sets have more zero observations than allowed for the Poisson model. Two kinds of models that can solve both of the problems are Hurdle models and zero-inflated models (Lambert D, 1992).

The density of the zero-inflated is a mixture of a point mass at zero I0(y)

and a count distribution  fcount(y;x,β)

. The probability of observing a zero count is π=fzero0;z,γ.

Hence, the probability density function of zero-inflated models can be written as:

fzeroinfly;x,z,β,γ=πI0y+1πfcounty;x,β,                          (5)


where I()

is the indicator function and the corresponding equation for the mean is

μi= πi0+1πiexpxiTβ,                                          (6)

using the canonical log link (Achim Zeileis, Christian Kleiber, Simon Jackman, 2008). In R, zero-inflated count data models can be fitted with the zeroinfl()

function from the pscl


The hurdle model combines two-component models: a hurdle component models zero vs. larger counts and a truncated count component, such as Poisson or negative binomial, which is employed for positive counts (Achim Zeileis, Christian Kleiber, Simon Jackman, 2008).

That is, the hurdle model combines a zero hurdle model fzero(y;z,γ)

(right-censored at y = 1) and a count data model fcount(y;x,β)

(that is left-truncated at y = 1):

fhurdley;x,z,β,γ=fzero0;z,γ                                                                    if y=01fzero0;z,γfcounty;x,β1fcount0;x,β                  if y>0         (7)

The model parameters β, γ

, and potentially one or two additional dispersion parameters θ

(if fcount

or fzero

or both are negative binomial densities) are estimated by maximum likelihood. The corresponding mean regression relationship is given by:

logμi=xiTβ+log1fzero0;zi, γlog1fcount0; xi, β,               (8)

again the canonical log link is used. In R, hurdle count data models can be fitted with the hurdle()

function from the pscl package (Jackman, 2008).

3.3 Stepwise regression

In statistics, there are many different strategies for selecting variables for a regression model. The usual method to choose a model is (backward/ forward) stepwise. “Backward stepwise” is used in this report, and the details are described below (Vincent Calcagno, Claire de Mazancourt, 2010).

  1. Fit the full model first and find out the terms that are not statistically significant, that is, whose removal does not signicantly reduce the fit of the model.
  2. Removes the least-significant one, thus obtaining a new “simplified” model.
  3. Repeat the procedure until all effects in the formula are found significant.

Besides, several methods that can determine whether the removal of a given term at each test is significant (Venables WN, Ripley BD, 1997).

  1. Hypothesis test tools, such as the chi-square test, or likelihood ratio (LR) test, involve specifying a significance threshold for the p values (e.g., 5%).
  2. Information criteria (IC), e.g., the Akaike information criterion (AIC).

Choosing model by AIC in a stepwise algorithm is convenient, and the choice of predictor variables can be carried out by an automatic procedure in R via  step()

CITATION Dob08 l 2057 (Dobson, Annette J; Barnett, Adrian G, 2008).

There are three directions for step()

function to select the model – “both”, “backward”, or “forward”. The only fundamental difference between backward and forward stepwise selection is whether there is the explanatory variables in the model at first (Hilbe, 2014):

  • without predictors (forward);
  • with all the predictors (backward).

As in GLM, the deviance is not simply related to the maximised log-likelihood (Achim Zeileis, Christian Kleiber, Simon Jackman, 2008). There is a potential problem when using the stepwise algorithm in GLM. Nevertheless, the Poisson and Binomial families have fixed scale by default (Hilbe, 2014), so it is not necessary to amend it in this case.

3.4 The variance inflation factor (VIF)

Multicollinearity occurs when there are high correlations among predictor variables, leading to unreliable and unstable estimates of regression coefficients (Allison, 2012). In this report, the variance inflation factor (VIF) is used to diagnose the multicollinearity. The VIF can be calculated by fitting a linear regression of that predictor on all the other predictors, and then the R2

can be obtained from the regression. The VIF is just 1/(1R2)

. It can be accomplished via vif()

in R. Moreover, for VIF in R, if the degree of freedom for any predictors in an unweighted linear model is bigger than 1, then generalised variance-inflation factors (Fox, J; Monette, G, 1992) are calculated.

Though the interpretations for GVIF is different, the rule of GVIF(1/(2*Df)) <2

(Philipp, 2019) is used in some publications, which would equal to  

a VIF of 4 for one-coefficient variables. The same criterion used in this report. That is, if GVIF(1/(2*Df))2

, multicollinearity appears at a high rate.

3.4 Over-dispersion Test


The function dispersiontest()

can be found in the package AER in R, which implements a test for over-dispersion by Cameron and Trivedi (Cameron; Trivedi, 1990).

As in a Poisson model, the mean E(Y)=μ

and the variance Var(Y)=μ

are the same, the over-dispersion test sets  VarY=μ+cf(μ)

, where the constant c<0

means under-dispersion and c>0

 means over-dispersion. The function f()

is the monotone function (linear is the default). The test statistic used is a t

statistic which is asymptotically standard normal under the null.

H0:c=0  v.s.  H1:c0

If the p-value in the output is smaller than 0.05, then the null hypothesis will be rejected. That is, there is enough evidence of over-dispersion.

4 Standard analysis of the data

In this section, the variable “affairs” is treated as a count variable; the relative models mentioned in the previous section are applied to the “Affairs”. Before the models are fitted, fundamental exploratory analysis of the dataset is carried out. At the end of the chapter, fitted models discussed in this chapter are compared. As the mean functions for the models are similar but the fitted likelihoods are different, the models’ performance differently when solving over-dispersion and/or the number of zero counts problems.


4.1  Exploratory data analysis

As there are no missing values in the dataset, the first step of exploratory data analysis is to obtain a first overview of the response variable. A histogram of the frequency of extramarital sex is carried out below. Since the values of the variable “affairs” are not always the real frequency, argument  names.arg

is used to re-label the frequency.

The figure below shows both substantial variation and a rather large number of zeros (see Figure 1). Besides, the second step in the exploratory analysis is always to look at the partial relationships, which can be shown by pairwise bivariate displays of the response variable against each of the predictor variables. Only the plot for the number of having affairs against age group is shown in this report, and the code for analogous displayed can be found in the R code section.


Figure 1 How often engaged in extramarital sexual intercourse during one year

Simple scatterplot is not useful as variables are ordinal categorical variables producing overlap. Boxplot or bar plot can overcome the problem of overlap, but as there are a large number of outliers in the dataset, the bar plot performance better. Besides, the numbers of observations for each group are different, prop.table

is used instead of table

to eliminate the population effects.  Figure 2 shows the bar plot of the number of having affairs against age group.

All displays show that the number of having affairs increases or decreases with the independent variables as expected. For example, the bar plot in Figure 2 shows that both the probability of having affairs and the frequency of cheating are higher when people at their middle age or when they are rather young, which fits the common sense.

Figure 2 The number of having affairs against years of married (percentage)

As multicollinearity is a common problem when estimating linear or generalised linear models, the most widely used diagnostic for multicollinearity – generalised variance-inflation factors (GVIF) – is used in the report (the details are in Section 3.4). Table 3 shows the degree of freedom and the (generalised) variance inflation factor (GVIF) for each predictor.

Table 3 GVIF for the predictors














































From the table, Multicollinearity could be ignored safely in this case.

4.2 Poisson regression

As a first attempt to capture the relationship between the number of having extramarital affairs and all predictor variables, the basic Poisson regression model is used. The backward stepwise regression procedure identified the model which included 8 predictors “gender”, “age”, “yearsmarried”, “children “, “religiousness”, “education”, “occupation” and “rating”, as the one which produced the lowest value of AIC. 

Significant coefficients are expected after the stepwise regression procedure, but it is not true in the case. Most coefficients are highly significant except coefficients for “gender” and “children”. Here are some reasons: the stepwise regression procedure in R selects the model based on Akaike Information Criteria (AIC), not p-values. The purpose is to find the model with the smallest AIC by removing or adding variables. Since each variable in the model is a categorical variable, this leads to likelihood-ratio tests allow higher p-values when selecting the models (Achim Zeileis, Christian Kleiber, Simon Jackman, 2008). Since the coefficients for “gender” and “children” are not significant, they are deleted one by one via deleting the variable with the biggest p-value in Chi-square tests.

Also, the test of over-dispersion mentioned in Chapter 3 is used, and there is evidence of over-dispersion ( c

is estimated to be 5.57

, and the p-value is ) which reject the assumption of equidispersion (i.e. c=0

). Therefore, it is more reasonable to explore other models to solve the over-dispersion (and excess zeros) problem. The values of coefficients are suppressed here and are presented in tabular form in Table 7.

4.3 Quasi-Poisson Model

As AIC is not defined for the quasi-Poisson model, the stepwise regression step()

  in R cannot proceed, so choosing a model by the chi-square test in a stepwise algorithm is considered. As a result, the predictor “occupation”, “gender”, “children”, “education”, “yearsmarried” are deleted in order. As the estimated dispersion of λ=6.639418

, which is larger than 1, confirming that over-dispersion is present in the data. Table 4 shows the deviances and AIC of Quasi-Poisson model.

Table 4

Null deviance: 2925.5  on 600  degrees of freedom

Residual deviance: 2250.4  on 584 degrees of freedom


From the table, the residual deviance for the Poisson model is  D=2250.4 

, which is very large compared with the  X2(584)

, indicating a poor fit if the Quasi-Poisson is the correct model for the response. That is, even if the over-dispersion is showed by the estimated dispersion of λ

, more models could be explored to find the most suitable model for the “Affairs” dataset. The values of coefficients are presented in Table 7.

4.4  negative binomial model

A more formal way to deal with over-dispersion in a count data regression model is to use a negative binomial model. As shown in Table 7, both the values of regression coefficients and significance of coefficients are somewhat similar to the quasi-Poisson. Thus, in terms of predicted means, the two models give very similar results. Table 5 shows the deviances and AIC of the negative binomial model.

Table 5

Null deviance: 408.73  on 600  degrees of freedom

Residual deviance: 340.85  on 584 degrees of freedom

AIC: 1479.2

From Table 5, the residual deviance for the Poisson model is  D=340.85 

, which is relatively small compared with the  X2(584)

distribution, indicating a good fit if the negative binomial is the correct model for the response.

4.5  Zero-inflated regression & Hurdle regression

In this section, the zero-inflated regression model and the Hurdle regression model are applied to the “Affairs” dataset. The backward stepwise regression procedure identified the model which included four predictors “gender”, “age”, “religiousness” and “rating”, as the one which produced the lowest value of AIC. As the hurdle regression model and ZINB regression model are based on the same variables, the better model could be selected via likelihood. Table 6 shows the log-likelihood of the two models. From the table, the log-likelihoods are almost same, and zero-inflated model is a bit better than Hurdle model in this case. See Table 7 for a more concise summary.

Table 6



Hurdle regression

-683.4 on 37 Df

Zero-inflated regression

-683.1 on 37 Df

4.6  Comparison of the models

Table 7 Summary of fitted count regression models for Affairs dataset





Poisson   Quasi-P       NB





     Hurdle-NB                   ZINB



  Count     Zero          Count        Zero





















































































































































































After fitting several count data regression models to the “Affairs” data set, it is interesting to look at what these models have in common and what their differences are.

The estimated regression coefficients in the count data model are compared first. The result (see Table 7) shows that there are some small differences: as for GLM, the value and significance of coefficients for the  Quasi-Poisson regression model are rather similar to the negative binomial model, and are different from the Poisson regression model. As the over-dispersion test, the Hosmer-Lemeshow test and the exploratory analysis discussed in the previous section, the negative binomial model captures the over-dispersion in the “Affairs” dataset better. In summary, the negative binomial regression model fits the “Affairs” dataset best in the three GLM.

Find out how can help you!

Our academic experts are ready and waiting to assist with any writing project you may have. From simple essay plans, through to full dissertations, you can guarantee we have a service perfectly matched to your needs.

View our services

Besides, the hurdle and zero-inflation models lead to similar results on this data set. Their mean functions for the count component are very similar. Though the fitted coefficients for zero component are somewhat different, the signs of the coefficients match, that is, for zero component, the signs for the two zero-augmented models are just inversed. Here are some reason, for the hurdle model, the zero component describes the probability of observing a positive count. But for the ZINB model, the zero-inflation component predicts the probability of observing a zero count from the point mass component.

As a result, if the response variable is treated as count data. The ZINB model is preferred since it has a more helpful interpretation: the zero component determines whether a person had extramarital affairs or not, and the count component shows how many affairs that person had. Besides, the significant variables are always marital happiness, age and degree of religiosity.

5. Advanced analysis of the data

In this section, the response variable “affairs” is treated as ordinal data, and ordinal logistic regression described in Chapter 3 is applied to the “Affairs” dataset. The polr()

command from the MASS

package in R is used to estimate the ordered logistic regression model.

To explore the predicted probabilities of different times of having affairs, the dataset “Affairs” is divided into two dataset – training dataset and test dataset at first. The training dataset is used to fit the parameters, and the test dataset is used to calculate the predicted probabilities.

5.1 Obtaining parameters on the test dataset

The final model selected by backward stepwise regression method includes five predictors: “gender”, “age”, “yearsmarried”, “religiousness” and “rating”. As the output does not include the p-value, the p-value can be calculated by comparing the t-value against the standard normal distribution. Though it is only right with infinite degrees of freedom, it is reasonably approximated by large samples (Chambers JM, Hastie TJ, 1992). After calculating the p-value, “gender” and “yearsmarried” are deleted since they are not always significant.

Confidence intervals can also be calculated for the parameter estimates. Assuming a normal distribution, the confidence interval can be obtained via confint.default()

in R. Note that profiled CIs are not symmetric, if the 95% CI does not cross 0, the estimated parameter is statistically significant.

The coefficients from the model are difficult to interpret because they are scaled in terms of logs. Hence, the coefficients and Cis are converted into the odds ratios (OR) and confidence intervals for odds ratio. The OR and confidence intervals can be gotten via exponentiate the estimates and confidence intervals. A table which includes the value of coefficients, the odds ratios and the confidence intervals is shown below with insignificant coefficients highlighted.

Table 8 The coefficients for the final OLR



Odds Ratio





























































































Here are some interpretations of the model statistics in Table 8.

Coefficients :

Here are some examples of the interpretation of the coefficients:

  • An individual aged 20–24  (z2=22)

    , compared to an individual aged under 20  (z2=17.5)

    , is associated with a lower likelihood of having many affairs in marriage. The 95% CI of the odds ratio doesn’t include 1, which means that the coefficient for people aged 20-24 is statistically significant at the 5% level.

  • An individual who is very religious (z5=5)

    , compared to an individual lack of faith (z5=1)

    , is associated with a lower likelihood of having many affairs in marriage. The 95% CI of the odds ratio doesn’t include 1, which means that the coefficient is statistically significant at the 5% level.

  • An individual who has a somewhat unhappy marriage (z8=2)

    , compared to an individual who has a miserable marriage (z8=1)

    , is associated with a lower likelihood of having many affairs in marriage. The 95% CI of the odds ratio contains 1, which means that the coefficient is not statistically significant at the 5% level.

Other coefficients can be explained similarly.


  • Mathematically, the intercept 0|1

    corresponds to logit[P( 1)]

    . It can be interpreted as the log of odds for the people who never had affairs in that year versus the log of odds for the people who had had affairs.

  • Similarly, the intercept 1|2

      corresponds to logit[P( 2)]

    . It can be interpreted as the log of odds for the people who never had affairs or just cheated once in the past year versus the log of odds for the people who had affairs at least twice.

Other intercepts can be explained similarly.

5.2  Making predictions on the test dataset

By using the intercept and slope values from Table 8, the probability of having affairs in a different frequency can be calculated for the observations in the test dataset.

For example, for the first observation in the test dataset, who is aged 25 -29  (z2=27)

, has somewhat religiousness  (z5=4)

and whose marriage is happier than average (z8=4)

The probability corresponding to never have affairs in the past year will be calculated as:

logit[P 1= 3.92132.3244*1+1.6870*1+1.5343*1

logit[P 1=1.6244

P( 1)= exp(1.6244)/(1+exp(1.6244)) = 0.8354011

In our case,  P( 1) = P(=1) =0.8354011

Similarly, the probability of that observation only had once affairs will be calculated as:

logit[P 2= 3.55552.3244*1+1.6870*1+1.5343*1


P( 2)= exp(1.9902)/(1+exp(1.9902)) = 0.8797643

Hence, P= 2= P 2 P 1

                                   = 0.8797643 0.8354011



Similarly,  P= 3=0.02079924

, P= 4=0.028035073

                   P= 5=0.04136479

, P= 6=0.030019260

The predicted probabilities of having different times of affairs for different observations in test dataset can be calculated in R easily. Figure 3 is a set of plots for all of the predicted probabilities for the various conditions.

Figure 3 plots for all of the predicted probabilities for the different conditions.

From the plot, it is clear that the size of the test dataset is quite small, hence the plots for some conditions are incomplete. And since all predictors are categorical variable, the lines in the plot are broken lines. Though more evidence is needed before any definitive conclusion can be drawn, the benefits and shortcomings of using an ordinal logistic regression model to fit the “Affairs” dataset are discussed below.

The significant variables for the final ordinal logistic regression (OLR) model are the same as the quasi-Poisson and the negative binomial regression model – rating

(marital happiness), age

and religiousness

(degree of religiosity). But the OLR can give the predicted probabilities of having different times of affairs for different observations exactly. Besides, as the response variable “affairs” is an ordinal data, OLR is a more proper method to fit the “Affairs” dataset better. However, there are also some limitations of treating the response variable as ordinal data:

  1. There are rather few statistical methods designed for ordinal variables.
  2. The parallel lines assumption in R system (which means that the intercepts are different for each category but the slopes are the same among categories) for OLR is rather difficult to test.

6 Conclusions and discussion

The primary purpose of this paper has been to develop a model to explain the relationship between the number of having affairs and other variables related to the demographic characteristics. The model frame for both count data regression and ordinal logistic regression, as well as their implementation in the R system for statistical computing, are reviewed. Incorporating the over-dispersion and excess zeros problems, hurdle and zero-inflated regression models is also considered for modelling. Moreover, the profits and limitations of treating the response variables as count data or ordinal data are discussed.

 As a result, though the models applied are different, the coefficients for rating

(marital happiness), age

and religiousness

(degree of religiosity) are always significant.  Among several count data discussed in this report, the zero-inflated regression model can explain the data best. For the ordinal regression model, a set of plots for predicted probabilities for the different conditions are showed. Though more data are needed to obtain a more complete plot for each cell.

For the further task, it is meaningful to study on if the observations having affairs with the same person or not, their marital status after several years (which means that if cheating in the marriage would change the status of marriage), or even the parental influence to individuals. As the extent of having extramarital affairs is by no means small, this topic has reality content to research.

Cite This Work

To export a reference to this article please select a referencing stye below:

Reference Copied to Clipboard.
Reference Copied to Clipboard.
Reference Copied to Clipboard.
Reference Copied to Clipboard.
Reference Copied to Clipboard.
Reference Copied to Clipboard.
Reference Copied to Clipboard.

Related Services

View all

DMCA / Removal Request

If you are the original writer of this essay and no longer wish to have your work published on the website then please: