CHAPTER 4 


WINNIPEG HEALTH AND DRINKING SURVEY - WAVE 2

Methodology

This chapter describes the methodology employed throughout this  longitudinal project.  Having already addressed this project's sampling technique and procedure of data collection in the previous  chapter, a brief description of the Wave 2 response rates will be presented in the following section.

Response Rate

Data collection for Wave 2 began during the summer of 1991 and was completed by the fall of 1992.  Of the 1257 persons interviewed in Wave 1, 83 moved away from the city, eight were deemed ineligible (i.e., died or were institutionalized), 57 could not be contacted and 121 refused to be reinterviewed. The number of completed interviews for Wave 2 was 988 and provided an overall response rate of 78.7 percent.
 

Independent Variables and Measures

The following is a list and explanation of the instruments and measures utilized throughout both waves of this project.  A battery of demographic, alcohol abuse, personality and partner abuse items were administered in Wave 1 and then were readministered in Wave 2.
     Items designed to assess changes that might have occurred during the  two year period between Wave 1 and Wave 2 as well as questions that might also elicit further information regarding the circumstances surrounding episodes of partner abuse were added to the follow-up interviews in Wave 2.  These items will be noted below where  applicable.

Demographic variables. Questions were included to measure the following demographic variables: (1) age, (2) gender, (3) marital status, (4) race, (5) religion, (6) employment status and (7) income.
     Items regarding changes with respect to marital status, family size, employment status, as well as income were added to the Wave 2 questionnaire (Appendix B. Part 1).

Alcohol abuse.  Two constructs were selected to assess the extent of alcohol abuse by the respondents (i.e., alcohol consumption and   alcohol dependence).

1) Alcohol consumption:

Alcohol consumption was measured by the Volume Variability Index  (Cahalan & Cisin, 1986; Room, 1982).  This instrument contained nine questions and measured the quantity-frequency (with an added indicator for binge drinking) of wine, beer and liquor consumption. (Appendix B, Part 2).  This measure is particularly relevant because  of the association between alcohol consumption and the perpetration   of violence consistently reported throughout the literature.  The index selected for use in this research was the number of ounces of alcohol consumed per day.

2) Alcohol dependence:

An index was constructed to measure alcohol dependence (Sommer Barnes & Murray, 1990).  The Alcohol Dependence Index was constructed  by assigning "0" and "1" values to scores which respectively fell below and above the scale cutpoints (based on those reported in clinical research on alcohol dependence). The values were then summed to produce a possible scale range of 0 to 3 whereby a score of 0 indicated no indicators of alcoholism while a score of 3 reflected  being alcoholic on all indicators.  It employed the following measures:

A) Michigan Alcoholism Screening Test (SMAST) (Porkorny &  Miller, 1972) is a 13 item test intended to screen individuals in the general population and one which has been widely used in other studies (Appendix B, Part 3). The SMAST was designed to produce a more effective, shorter, self- administered and more easily scored version of the original MAST. In a study comparing  the two versions of the MAST, reliability coefficient alphas computed for the two comparison group scores, yielded coefficients only slightly lower for the SMAST (Seltzer, Vinokur & Rooijen, 1975). However, validity coefficients on the other  hand were found to be slightly higher for the same shortened scale version. In light of these findings, the authors concluded that the SMAST is as effective as the MAST in screening for alcoholism.

B) The Alcohol Dependence Data Schedule (SADD) (Raistrick, Dunbar & Davidson, 1983) is a 15 item instrument which has been administered to both clinical and non-clinical samples and has  been found to strongly distinguish alcoholics from nonalcoholics (Appendix B, Part 4). The split- half reliability obtained by   Raistrick et al. (1983) using the short form (i.e., the 15 item scale used in this study) was .87. Other research by Jorge and Mazur (1985) employing this shortened version obtained a split half reliability of .88 when used in an interview format and .82 when self-administered. Results of the Wave 1 provided an estimate of internal consistency of .82 for females and .68 for males.

3) A subscale of the NIMH Diagnostic Interview Schedule Version III, Revised (DIS) which provides a number of indices of alcoholism that follow the diagnostic criteria of the Diagnostic and Statistical Manual (DSM-III) (American Psychiatric  Association, 1980) was developed by Robins, Helzer, Croughan, Williams and Spitzer (1979). The subscale selected for use in  this research is the "lifetime diagnosis of alcohol dependence/abuse" (Appendix B, Part 5).

Personality.

Several instruments measuring the major dimensions of personality were selected for use in this research. They are as follows:

1) Eysenck Personality Questionnaire (EPQ-R) (1985):

The EPQ-R is a 100 item scale containing three personality subscales:  Neuroticism (EPQN), Introversion-Extraversion (EPQE), and Psychoticism (EPQP). The EPQ-R also contains a validity scale, the Lie Scale (EPQL) (Appendix B, Part 6). The Lie Scale may also be considered a measure of social conformity.  While Neuroticism and  Introversion-Extraversion are well established tests of personality, the Psychoticism dimension is newer and has been a source of  controversy.  Earlier versions of this scale have been criticized for low reliabilities (i.e., Cronbach's Alpha = .74 for males and .68 for females) (Torrubia & Muntaner, 1987) and skewed distributions (Eysenck & Eysenck, 1975).  The most recent version (EPQ- R) (Eysenck, Eysenck & Barrett, 1985) seems to have overcome these handicaps. Wave 1 data provided the following reliability coefficient alphas for the EPQ-R subscales: EPQN (males .85; females .85), EPQE   (males .82; females .80), EPQP (males .60; females .61)  and EPQL (males .83; females .82) (Sommer, 1990; Sommer et al., 1992).

2) MacAndrew Scale (MAC) (MacAndrew, 1965):

The MAC is a 51 item subscale of the Minnesota Multiphasic  Personality Inventory MMPI) (Appendix B, Part 7). It has been successfully cross validated against samples similar to those for which it was originally developed (alcoholics versus non-substance abusing psychiatric patients) (MacAndrew, 1980). MacAndrew (1980) found that 85% of male alcoholics assessed by the MAC scale were classified as "secondary psychopaths" (i.e., neurotic extroverts) according to Eysenck's model of personality.  Whereas results of Wave 1 data indicated that the MAC was a predictor of partner abuse for female respondents (Sommer et al., 1992), tests of reliability performed on these data provided low estimates of internal consistency for both males (.43) (Sommer, 1990) and females (.54) (Sommer et al., 1992).  These were likely due to the multidimensional nature of the scale.  Earlywine, Fine and Martin (1990) considered  the MAC a fallible indicator of underlying constructs. In spite of the MAC's low estimate of reliability, these authors determined that a confirmatory factor analysis demonstrated associations between it and personality measures.  Other research by MacAndrew (1980) showed that the MAC is still considered an appropriate measure because of its widespread use and its demonstrated ability to distinguish between alcoholic and nonalcoholic populations.

3) Neuroticism Index:

In order to avoid the problems associated with multicollinearity, several highly correlated scales were combined to form a Neuroticism  Index (Sommer, Barnes & Murray, 1990).  A conservative approach toward selection of the measures based on the strength of intercorrelations supported this procedure (Sommer et al., 1992).
     Support for combining these measures was also found in the literature on personality (Krisha, 1980; Rosenberg, 1979; Roy, 1977).  The index was constructed by averaging the means of the scales' z scores.
     Transformations were performed to ensure that all composite measures were scored in the same direction.  The following instruments were included in this measure:

A) The Rosenberg Self-Esteem Scale (Rosenberg, 1965) is a 10 item scale which has been found to relate to alcoholism  (Beckman, 1978) as well as with perpetrators of partner abuse (Goldstein & Rosenbaum, 1985; Rouse, 1988; Barnes et al., 1990) (Appendix B, Part 8). This scale provided satisfactory estimates of internal consistency on Wave 1 data for males (Alpha =.83) (Sommer, 1990) and females (Alpha =.86) (Sommer et al., 1992).

B) The Trait Anxiety Scale (Spielberger, 1970), a widely used  measure of anxiety consists of 20 items (Appendix B, Part 9).  Since our earlier research on abuse among male college students  found neuroticism to be a significant predictor of abuse (Barnes, Greenwood & Sommer, 1991), this particular measure was therefore thought to be relevant in this present examination of partner abuse. This measure provided a satisfactory estimate of internal consistency in Wave 1 for males (Alpha =.84) (Sommer, 1990) and females (Alpha =.88) (Sommer et al., 1991).

C) The Baron Ego-Strength Scale is another subscale of the MMPI containing 67 items (Appendix B, Part 10). It measures a general factor of the capacity for personalityintegration (Greene, 1980).  According to Greene (1980), persons who score high on  this measure are thought to have secure self concepts and are able to cope with situationalstress.  Low scorers on the other hand, are more likely to experience chronic, personality problems.  It was suggested that those with less integrated personalities might experience lower levels of impulse control and resort to more primitive methods of coping such as violence (Sommer, 1990). Analyses of Wave 1 data provided somewhat low, yet satisfactory estimates of internal consistency for males (.67) (Sommer, 1990) and females (.70) (Sommer et al., 1992).

D) A fourth measure included in the Neuroticism Index is the EPQN described above.

Family background.

In order to address the issue of violence within the family of origin, two questions were included in the Wave 2 interview schedule (Appendix B, Part 11).

Stress.

Twelve items have been included in the Wave 2 interview schedule to assess the extent of life stress events experienced by abusers and their spouse/partners.  The items selected are similar to some of  those included in the Holmes and Rahe (1967) Life Stress Events Scale (Appendix B, Part 12).

Dependent Variable

Conflict Tactics Scale.

In this project, the dependent variable, partner abuse was measured  by an abridged version of Straus's (1979) Conflict Tactics Scale (CTS) (Form A).  The scale measures the frequency with which abusive actions occur as well as the degree of its severity.  An individual scoring low on the CTS is indicative of someone whose experience with partner abuse is both infrequent and less severe than one scoring high on the same measure.  In this research, both the prevalence (i.e., abuse ever occurring during the course of a relationship) and incidence (i.e., reports of partner abuse incidents during the past year) of perpetrated partner abuse were assessed by this measure.  In its original form, the CTS incorporates the following three modes of conflict resolution: (1) the use of rational discussion, argument and reasoning, (2) the use of verbal and nonverbal acts, and (3) the use of physical force against another person.

For the purposes of this study, six items of the CTS measuring "physical force" tactics were selected for analyses.  Items associated with emotional abuse (i.e., rational discussion, argument and reasoning) were not considered relevant to this investigation of physical abuse and were therefore excluded.  As well, the most severe "physical force" conflict tactics (i.e., beat up, threatened with a knife or gun, used a knife or gun) were also omitted because of low endorsement rates demonstrated in past research (i.e., 0% to 4%) (Brinkerhoff & Lupri, 1988; Malone et al., 1989; Marshall & Rose, 1990; Smith, 1987; Stets & Pirog-Good, 1989).

Furthermore, whereas many studies have employed the full scale in their analyses, portions of the scale have also been used in other research (Brinkerhoff & Lupri, 1988; Roscoe & Benaske, 1985; Smith 1987). In fact, the reliability of an abbreviated version used in Wave 1 of this project (Cronbach's Alpha =.79 for males and Cronbach's Alpha =.94 for females) was found to show greater internal consistency than that of the entire scale used in our earlier research on male perpetrated courtship violence (Cronbach's Alpha =.57) (Barnes, et al., 1991).

Various versions of the scale have been used in face-to-face interviews (Kennedy & Dutton, 1989; Schulman, 1981; Smith, 1987; Sommer, Barnes & Murray, 1990, 1991; Straus et al., 1980), telephone interviews (Straus & Gelles, 1986) and mail surveys (Straus, 1979).
     However as noted previously, the CTS is limited in its ability to elicit information on the following: (1) the circumstances surrounding the violent episode occurred (i.e., who initiated the episode or did it occur in self defence), (2) whether the consumption of alcohol was involved prior to the violent episode, (3) the consequences of violent episodes (i.e., physical injuries) and (4) whether episodes of abuse had been reported to the police.  The  revised version of the CTS employed in this study included items designed to overcome most of the noted limitations of previous research (Appendix C).

Data Analysis

The data analyses conducted in this second phase of the research followed a similar format to that conducted during its first phase using SAS statistical packages (SAS Institute Inc., 1986).  Data from both Waves 1 and 2 of the Winnipeg Health and Drinking Survey (Barnes & Murray, 1989) were used to test the theoretical issues in this research.

In light of the diminished response rate reported earlier for Wave 2, data analysis began with an investigation into the effects of sample attrition. Inspecting the data for possible distortions resulting from missing data is important because nonrandom attrition may pose a threat to internal and external validity (Stacy, Newcomb & Bentler, 1991).  Determinations regarding the proportion of partner abusers from Wave 1 found among those who refused to participate or could not be contacted for Wave 2 data were made.  In addition, possible differences between study dropouts and completers with respect to Wave 1 sample characteristics (i.e., socio-demographic and personality factors, alcohol consumption and CTS scores) were evaluated.

Assessing changes and stability in prevalence rates of partner abuse (i.e., has abuse "ever" occurred) as well as in other independent measures reported in Wave 1 and Wave 2 were achieved by way of frequency and correlational analyses.  As was the case in the analysis of Wave 1 data, Pearson correlations were also performed  within the Wave 2 data to assess the bivariate relationships between partner abuse.

The relationship between alcohol consumption (Wave 2) and partner abuse (Wave 2) were examined for linearity by collapsing the former variable into three levels (i.e., low, medium and high alcohol consumption) and then examining differences in the partner abuse score means.  Linearity is established when the means of the partner abuse scores are found to increase across levels of alcohol consumption.

The final phase of the data analysis tested the major hypotheses of this research using logistic regression analyses.  Given that reported incidence rates for perpetrated partner abuse (10-14%) fall well below the 25 percent cutoff considered to be appropriate for inclusion as a dependent measure in a multiple regression analyses   (Cleary & Angel, 1984), the selection of this approach was considered appropriate.  The application of this procedure in previous spousal violence research (Bland & Orn, 1986; Kalmuss & Seltzer, 1986; Seltzer & Kalmuss, 1988) provided additional support for use in this research.

The logistic regression approach has two distinct features; the dependent variable is dichotomous, and the effects of the independent variables are presented as odds ratios.  In order to conform with the former, this study's dependent measure, "current perpetrated partner abuse" (i.e., the number of times an individual perpetrates partner abuse during the past year) was collapsed into the following two levels: abuse as reported at any level of severity and/or frequency (coded as 1) and no abuse (coded as 2).

According to Halli and Rao (1992), the odds ratio forms the backbone of logistic regression.  In this research, logistic regression analyses made it possible to establish the ratio between perpetrating and not perpetrating current partner abuse based on a number of underlying and situational factors.  The magnitude of each predictor was also assessed through an examination of the parameter estimate in relation to the standard error.  Finally, the fit of the model being tested (i.e., the degree to which the predictors tested were needed to model the dependent measure) was determined by way of chi-square analysis.

In spite of the appropriateness of logistic regression in this research, there are a number of limitations inherent in the use of this approach that are in need of consideration.  For example, a problem associated with the severe skewness of the dependent variable is the potential instability of coefficient estimates using this statistical application (Seltzer & Kalmuss, 1988).  In order to compensate for this problem, the following strategies put forth by  Seltzer and Kalmuss (1988) were adopted in the logistic regression analyses conducted in this research:

1. Rather than including all the main effects and interaction effects in one model, the interactions were estimated in two separate models (i.e., one examining interactions with stress and the other examining interactions with current alcohol consumption).

2. Conservative standards for the interpretation of results were also employed.  Thus, "statistical significance was attributed only when a coefficient was twice the value of the standard error" (p. 481, Seltzer & Kalmuss, 1988).

The second issue relates to a loss of sensitivity in the dependent measure due to the transformations performed.  The quantitative differences once present in the continuous variable, "current perpetrated partner abuse" were eliminated when it was collapsed into two levels.  Thus, instead of assessing the perpetration of current partner abuse as ranging from low to high (with a possible range of 0 to 12), the transformed variable can only be interpreted as indicating either the presence or absence of current perpetrated partner abuse. To the extent that a dichotomous dependent measure is a necessary condition of logistic regression, its resulting lack of sensitivity must be viewed as an unresolved limitation in this approach, to be dealt with theoretically.

Next: Chapter 5 Part 1