Can Supervised Machine Learning Be Used to Identify Family Firms Using a Sophisticated Definition?

The ability to identify family firms is a critical element of any empirical family businesses study, however obtaining representative time series data that applies complex family firm definitions can be challenging and resource consuming. In this method development paper, I propose that supervised machine learning may be used as a scalable imputation approach after a ‘learner’ has been trained with a combination of survey and registry data. I demonstrate this technique using survey data from 7,153 firms that applies the ‘European definition’ of family firms, which captures elements of voting rights and governance. I combine this data with population registry data containing person- level family relationships who are affiliated with the firms through ownership, board membership, or several other ways. After training several random forest, boosting tree and neural network machine learners, I find that random forest and boosting tree learners obtained the best performance by being able to correctly classify roughly 9 out of 10 firms in holdout samples. Furthermore, several avenues of further improving the accuracy of prediction exist, but their applicability depends on the intended use of the predicted data. Future directions and implications to empirical family business research are discussed.