Should we care (more) about data aggregation?

Abstract

Aggregation tools transform multidimensional data into indices. To investigate how the design of an aggregation process affects regression results, we build democracy indices that differ regarding their scale and aggregation function. Using the democracy–growth nexus as a testing ground, we illustrate that the choice of the aggregation procedure significantly affects OLS and 2SLS estimates since different methods produce systematically different index values for observations at the lower and upper end of the autocracy–democracy spectrum. We also illustrate that dichotomous measures produce significantly smaller OLS estimates than continuous measures due to lower discriminating power. Whether continuous and dichotomous indicators create different 2SLS estimates depends on their design. Because of the methodological similarities of democracy indicators and other social science indicators, we expect similar consequences for other empirical analyses.