Countries where Women Have more Positive Interactions with Economic Decisions and Legal Rights Have Lower Homicide Rates: an Exploratory Study

For purposes of this paper, we are focused on homicide rates in general, and not domestic-violence oriented homicides. We return to this later in the manuscript.

Although the original WBL data contained information for 186 cases, we could not locate homicide rates for seven of those countries in the IHME database: (1) West Bank and Gaza, (2) Palau, (3) San Marino, (4) Hong Kong, SAR, China, (5) Korea, Republic, (6) Kosovo, and (7). St. Kitts and Nevis.

Other fit statistics, such as the AIC, have been used with this methodology. With respect to the AIC, it is very similar to the BIC but does not vary with sample size (Nagin 2005, p.74) and oftentimes both BIC and AIC tell a substantively similar story.

This type of finite mixture modeling has been used extensively in criminology to examine individual variability in offending patterns (Piquero 2008), but has also been used in other clinical and social sciences to study a variety of topics (Nagin and Odgers 2010). The methodology has also been applied to countries in previous research (Piquero and Piquero 2006).

It is important to note that by increasing the number of groups identified, we observed both a smaller number of cases in the ‘new’ group as well as a worse BIC score. As Nagin (2005) and others have noted, it is also the case that simply having one more group does not always provide more insight, especially if the group being added is simply the creation of one more group that is spliced off from another group thereby leaving fewer and fewer cases in the additional groupings (Piquero 2008). For this portion of the analysis, the two-group BIC was −781.67, which was better than the other model permutations such as the three-group model (BIC = −786.89) and the four-group model (BIC = −790.87). The appropriateness of the model was examined with the posterior probabilities as recommended by Nagin (2005) and these were higher than the models with additional groups. Moreover, when running the three and four group model, the sample sizes in each group became too small and led to concerns associated with power to detect associations. In the three-group model, 17.34% of the sample was in group one, 82.18% of the sample was in group two, and only 0.43% of the sample was in the third group. When we estimated the four-group model, we found similar issues– where less than 1 % of the sample was in group four. Given that these additional subgroups were so small (<1% of the sample) and did not indicate latent classes that were unique in shape from the two-group solution, we selected the two-group model for this portion of our analysis.

As noted above, we estimated up to 9 models and the corresponding BIC values were: 2-GROUP (BIC = −608.92), 3-GROUP (BIC = −610.70), 4-GROUP (BIC = −614.09), 5-GROUP (BIC = −618.83), 6-GROUP (BIC = −623.50), 7-GROUP (BIC = −626.20), 8-GROUP (BIC = −630.01), 9-GROUP (BIC = −634.94).

At first glance, the homicide rate for Group 3 is much higher than it should be, relative to its other group neighbors. The issue here is that one country in particular, El Salvador, had a very high homicide rate (56.69) but they were grouped into this category because their women’s index score approximated the other countries in the model.

GDP information was unavailable on Worldometer for 7 countries. However, GDP for these countries was collected from a different source. The countries are listed as follows, with their respective GDP source: (1) Cabo Verde; Statista, (2) Cote D’Ivoire; World Bank, (3) Dijbouti; World Bank, (4) Eritrea; Statista, (5) Kosovo; Trading Economics, (6) Puerto Rico; Statista, (7) Tawain; Trading Economics.

Population density information was unavailable on Our World in Data for 3 countries. However, population density for these countries was collected from a different source. The countries are listed as follows, with their respective population density source: (1) Eritrea; Macro Trends, (2) Korea, Rep.; Macro Trends, (3) Tawain; Statista.

An anonymous reviewed suggested that we estimate a negative binomial regression model (using the originally, skewed homicide rate variable) in lieu of the OLS model that used the log of the homicide rate. The results of the negative binomial regression were substantively the same with respect to coefficient size and significance to those reported using the skewness-corrected homicide rate.