Challenges when Generalizing Psychological Measurements across Populations : Applications in Machine Learning and Cross-Cultural Comparisons

Abstract: In order to ascertain the validity and applicability of psychological theories, models, and measurements, it is important to examine their generalizability across different assessment situations. In this thesis, I examine how the application of measures outside of their initial domain may cause complications. This is applied to two fields where such considerations of generalizations may be especially beneficial: machine learning models and cross-cultural comparisons. Paper I explored whether text-based machine learning models of personality with a broad set of predictors, or models based on a set of more constrained but more psychologically meaningful predictors, better predicted personality in one of two text domains. The former models provided equal or superior prediction in the same domain in which it was trained compared to the latter models, but equally poor or poorer prediction in the other domain. Paper II reexamined the results of an article that, like the cross-cultural studies re-examined in Paper III, found that over time and across states in the U.S., higher gender equality was associated with larger gender differentiation, here in names given to children. Re-analyses showed that there was no such systematic association across time, and that the differentiation across states was confounded with a more strongly associated cultural/language predictor. Paper III re-examined multiple studies that have assessed that association across countries. Here, it was shown that cultural differences, as indicated by cultural regions, other measures such as individualism, and data quality indicators, better explained the variation in differences across countries. When controlling for cultural/language regions, the association with gender equality disappeared or, sometimes, reversed. These results indicate the degree to which different cultural factors are interrelated, and suggests the need for complementary methods. In conclusion, this thesis exemplifies the importance of considering how models and measures may interact with and generalize across situations. This is true whether it supports greater generality or situational specificity of different psychological measures.