2.2.1. Dependent variables
This study examines two separate dependent variables: frequency of drinking in public settings and in private settings over the past 12 months. Frequency, rather than usual quantity or volume, was used because only frequency and not quantity in different settings was collected in the surveys. These variables are based on the GENACIS Expanded Core questions. The surveys assessed frequency of drinking in various contexts by asking: “Thinking back over the last 12 months, about how often did you drink in the following circumstances? Think of all the times that apply in each situation”. Two situations, or contexts, were chosen to represent Public drinking: “in a bar/pub/disco” and “in a restaurant” and two were chosen to represent Private drinking: “at a party or celebration” and “in your own home”. The eight response categories ranged from “every day or nearly every day” through “once or twice a year” to “never in the last 12 months”. Categories were converted to the metric of days per year using category midpoints. The values for each of the two constituent contexts were summed to indicate the frequency of drinking in each (public and private) setting. Because it is possible to drink in two settings on a given day, the summed frequencies could exceed 365 days. However, exceeding 365 days was extremely rare, so results were not truncated.
Identical or similar questions were asked in each participating country. Sweden only asked these questions in a random third of the full sample; however, the one-third sample was similar in size to those of other countries (Table 1
). Most countries included the two separate questions for frequency of drinking in public settings, i.e.
, in (a) a bar, pub, or disco and (b) restaurant. However, Denmark, Iceland, and Sri Lanka surveys asked about frequency of drinking in a bar, pub, disco, or restaurant/café in a single combined question. Asking multiple questions tends to give higher values than use of a single, combined question. To make responses from surveys more comparable and reduce the methodological ‘penalty’ in the three surveys with the single public setting question, gender specific ratios of frequency of drinking in bars, pubs, and discos versus
in restaurants from similar countries were applied to the gender-specific combined public venue data. For Denmark and Iceland, Swedish ratios were applied. For Sri Lanka, Indian ratios were applied. Restaurant drinking was minimal in India and in Sweden, so this adjustment made very little practical difference.
2.2.2. Independent variables
Country-level variables: Country-level variables to measure gender equality and economic status include both existing indices and indices created specifically for this study.
We included four existing indices and two newly created indices to measure composite gender equality; gender equality in economic participation and opportunity, education, and political participation; reproductive autonomy; and context of violence against women. The existing indices were: the 2008 Gender Empowerment Measure (GEM) and the 2007 Global Gender Gap Index (GGI) Economic Participation and Opportunity, GGI Education, and GGI Political Participation sub-indices [52
]. Indices of women’s reproductive autonomy and context of violence against women were created. In addition to the theoretical reasons for including gender equality in economic participation and opportunity and the context of violence against women described in the introduction, the GEM and other indices were included mainly because of their use in previous research related to gender equality and health [7
The GEM is a composite index that measures gender equality in political participation and decision-making, economic participation and decision-making, and power over economic resources. Higher scores indicate greater gender equality. Sweden has the highest GEM score (0.925), with Denmark and Iceland also highly ranked (0.887 and 0.881 respectively). India, Sri Lanka, and Nigeria have low GEM scores (0.24, 0.371, and 0.198 respectively); while Costa Rica, Argentina, and the United States have moderate scores (0.69, 0.692, and 0.769 respectively). GGI sub-indices estimate relative to men, women’s economic participation and opportunity, educational attainment, and political participation. Higher GGI scores indicate greater gender equality The GGI also has a composite index of country-level gender equality. However, the GGI composite index includes gender differences in life expectancy. Therefore, the GEM was preferred as the composite gender equality indicator.
To our knowledge, there are no existing indices that measure women’s reproductive autonomy across countries. However data about reproductive autonomy and women’s actual control over reproduction are consistently collected and reported in multiple sources. We created a reproductive autonomy index based on the following variables: restrictiveness of abortion laws [56
], contraceptive prevalence [57
], total fertility rate per woman [58
], mean age at marriage for women [57
], and length of maternity leave [57
]. This index reflects a combination of both policy-level reproductive rights and actual reproductive control by women. Adolescent fertility rate and modern contraceptive use were also considered. They were not included because of high correlations with the previously mentioned variables and more missing values. Country-specific indices were created through factor analysis of the five variables: restrictiveness of abortion laws (a five category variable with 1 being most restrictive, 5 least), prevalence of any contraceptive use, total fertility rate per woman, mean age at marriage, and average number of weeks available for maternity leave. A factor analysis revealed a strong single dimensional structure (first eigenvalue of 3.1 comprising 63% of the total variance, all factor loadings larger than 0.7, and with the second eigenvalue less than 1).
To our knowledge, there are also no existing indices that measure the context of violence against women across countries. Recently, many countries have started to collect data on both actual violence against women and countries’ responses to this violence. However, data on actual rates of violence against women are collected inconsistently and are often not comparable across countries. Recent attempts to standardize data collection have moved in a positive direction (see, for example [59
]). However, such data is available from only a subset of countries. Because context of violence against women is a theoretically important factor for this analysis, we created an index based on the best available data. This includes both actual rates of violence against women and country response to such violence. The index is based on the following variables: percent ever sexually assaulted (either attempted or completed) [60
], percent experiencing physical violence by a partner in the past year [59
], percent of population feeling unsafe on the street after dark [66
], homicide rates [61
], attitudes towards wife beating [59
], quality of violence against women legislation [73
], and the number of domains of activity to address violence against women a country engages in as reported to the UN Secretary General [74
]. Even with use of multiple sources, many values of these variables were missing.
Country-specific indices were created through factor analysis. Variables included in the factor analysis were chosen based on completeness of data and findings from pairwise correlations with a wider range of measures that included varying time frames for sexual assault and partner physical violence as well as gender-disaggregated homicide rates. Only total homicide rate was included because there were more missing data for gender disaggregated rates and male and female homicide rates were highly correlated (0.75). The seven variables included: percent of women reporting ever being sexually assaulted, percent of women reporting physical violence against them by a partner in the past year, percent of the population feeling unsafe on the street after dark, rate per 100,000 of mortality caused by homicide, percent of men reporting that violence towards one’s wife was justifiable, quality of legislation within the country punishing violence against women (on a scale of 0–1 with 0 as the highest quality of legislation and 1 as the lowest quality of legislation), and the number of different domains of activity to address violence against women that a country engages in, as reported to the UN Secretary General (on a scale of 0–1, with 0 as having activities in all 7 possible domains). The percent of the population feeling unsafe on the street after dark was initially included in the factor analysis, but produced a very small factor loading and was therefore excluded. Similar to the results for the reproductive autonomy factor, a strong single dimension emerged from the analysis (first eigenvalue of 3.7 comprising 62% of the variance, all factor loadings larger than .6, and with a second eigenvalue of less than 1).
Gross Domestic Product per capita 2006 (GDP) and the Human Development Index (HDI) were both considered as indicators of country-level economic status. The two measures were highly correlated (0.82) and multilevel findings from the HDI were similar to GDP. Therefore, only results for GDP are reported here as it was less correlated with the other country level variables than the HDI (See Table 3
). Differences in findings between HDI and GDP are noted in the text.
Missing values Data for each country, with the exception of Isle of Man, were available for GGI sub-indices and GDP. Missing data was dealt with by substituting values with those from similar countries and by multiple imputation. First, country-level data for the Isle of Man were unavailable. Data from the United Kingdom was deemed to be the most appropriate country based on both current and prior British influence and therefore were directly substituted. Second, for the 2008 GEM data were unavailable for India and Nigeria. GEM scores from 1999 and 1996, respectively, were used.
As standard HLM models require complete country-level data and as the pattern of missingness of country-level data was well-dispersed across countries, the remaining missing values were imputed within each gender equality domain. For the reproductive rights index, Iceland was missing the contraceptive prevalence rate. For the context of violence against women index, values for 24 country-variable pairs (about 22% of observations) were missing. The missing values were for sexual assault (eight countries), physical violence from partner (seven countries), and quality of legislation punishing violence against women (one country). Data were imputed in the NORM program [75
]. This program imputes continuous data assuming a multivariate normal distribution of the data. Missing data were imputed 10 times and the average value was substituted for all missing values. In order to examine the variability in the estimates produced from the multiple imputations, the range of each of the imputed values was estimated. For all data imputed, variation across imputations was very small. The range from the smallest to the largest values across all variables ranged from only −1% to +2%. Given the small amount of variability in imputations, it was decided that it was unnecessary to estimate the model for each of the multiply imputed datasets and then combine resulting model estimates.
Individual Level Variables: Age, gender, and marital status were taken from responses to the GENACIS surveys in each country. Across countries, age was asked as a continuous variable. Marital status, although asked with slightly different possible categories across country, was coded as 1 if the respondent was married or living with a partner and 0 otherwise. Gender was coded as 0 if female and 1 if male.