Statistical simulation and counterfactual analysis in social sciences

In this paper, we present statistical simulation techniques of interest in substantial interpretation of regression results. Taking stock of recent literature on causality, we argue that such techniques can operate within a counterfactual framework. To illustrate, we report findings using post-electoral data on voter turnout. The analysis of quantitative data, and the estimation of regression models in particular, can now be considered commonplace in the social sciences. There are, of course, notable variations in the ways those analyses are generated (research design, estimation methods, etc.). In the same way, there are discrepancies in terms of standards when it comes to the interpretation of the results and their proper communication. Depending on the nature of the data at hand and the chosen estimation methods, the interpretation phase can be rather equivocal. For instance, displaying the odd ratios, or their natural logarithm, following logistic regressions can be far from intelligible, especially when one is interested in parameters beyond their statistical significance threshold and the direction of their coefficients. Consequently, a greater analytical effort appears to be required to flesh out the proper signification and meaning of parameters, most particularly to express their magnitude. The interpretation of statistical results appears crucial, especially under the lens of knowledge transfer, which involves non-statistical experts (decision-makers, policy analysts, etc.). We contend here that statistical simulation can be put to profit to this end. We also make the argument that this approach is compatible with a counterfactual conception of causality. Although this is not the place to develop a full-fledge