Good day forum, first I want to say thank you for such engaging content. I love the deep topics brought forward in this forum and I have learned a great deal about SPSS and statistical methods. I appreciate the work that everyone puts into this forum.
On to my topic, I have an entry-level question about interpreting binary logistic regression output, and I want to know if I am on the right track. I have done a bit of research outside of this forum, but I value the opinions here. I will describe what I am hoping to learn, provide output, and share my interpretation; I would appreciate your review of my interpretation.
In full disclosure, this is related to work on my master's thesis.
Using the Variations of Democracy dataset, I am attempting to understand if 1) a model for predicting the occurrence of genocide is possible, and 2) if religion [broadly] enhances or detracts from that model. Using the V-Dem dataset as the source of independent variables (IV) for the 2022 year, I recode one of my 6 IVs as a dichotomous variable to reflect the absence or presence of a situation and use the other variables in their native format. My dependent variable (DV) is coded by me as binary and generated from another publicly available source, the GenocideWatch website.
When running the regression analysis one IV is designated as a categorical variable (absence/presence) and the Hosmer and Lemeshow goodness of fit test is turned on as the only option. The IVs are entered as a group of 5 for block 1 and 1 for block 2. The output for this test is attached.
I will interpret this for our discussion, not as I would for publication. Obviously Block 0 is our baseline and the important points here are that we have ample cases with 178 of our 179 providing data to the analysis, our Classification Table demonstrates that the baseline demonstrates a 74.7% predictive capability when predicting the absence of genocide, and the model is statistically significant as shown in Variables in the Equation for Block 0.
Shifting to Block 1 where we enter the non-religious IVs, the Omnibus Tests of Model Coefficients indicates model significance with a p < .001, which is corroborated by the Hosmer and Lemeshow Test showing a significance of .347. The Model Summary shows that between 26.3 and 38.8 percent of the variance in the DV can be explained by these 5 DV, however, Variables in the Equation demonstrates that only 2 of the DVs show statistical significance, and only one demonstrates that for every unit increase in domestic autonomy we would expect a 1.099 increase in the log-odds of reported genocide. The Classification Table reports a 5.9% increase in predictive capability over baseline with a Percentage Accuracy in Classification of 80.3%. This is comprised of a positive predictive value of 66.7% and a negative predictive value of 83.1%. The model shows statistical significance, but lacks practical significance based on the p-values reported in Variables in the Equation and only 2 of the 5 IV showing statistical significance (performance legitimation .016, freedom from political killings .005).
Considering Block 2 where the religious IV is entered, the Omnibus Tests of Model Coefficients indicates continued model significance above Block 1 with a p < .001, which is again corroborated by the Hosmer and Lemeshow Test with a value of .153, although this is less significant than Block 1. The Model Summary shows that between 26.6 and 39.3 percent of the variance in the DV can be explained by the 6 DVs, and again Variables in the Equation table shows that while 2 variables add significantly to the model, only a one-unit increase in domestic autonomy would demonstrate an expected increase of 1.087 increase in the log-odds of reported genocide. In this block the Classification Table shows an increase of 8.4 percentage points over baseline and 2.8 percentage points over the predictive capability of Block 1, with a percentage accuracy in classification of 83.1%. This is comprised of a positive predictive value of 75.9% and a negative predictive value of 84.6%, both increases over Block 1 measures. In this case, while the model shows statistical significance it continues to lack practical significance based on only 2 of the 6 IVs demonstrating statistical significance in the equation (performance legitimation .013, freedom from political killings .007).
I am asking if this assessment aligns with the data, or am I mischaracterizing or possibly misinterpreting anything? This analysis answers my questions, although not as I would have expected. Thank you in advance and I look forward to your insights.
------------------------------
Sjon Woodlyn
------------------------------