Scholars in the social sciences are increasingly relying on machine learning (ML) techniques to construct data from large corpora of text and images.
The ML-generated variables are subsequently utilized in statistical analysis to address substantive questions through regression and hypothesis testing.
However, this approach can introduce substantial bias and lead to incorrect inferences due to prediction errors during the machine learning stage.
In this paper, we present an approach that incorporates ML-generated variables into regression analysis while ensuring consistency and asymptotic normality.
The proposed approach leverages a small-scale human-coded sample to capture the bias in the naive estimator, without the need for strict assumptions about the structure of prediction errors. Furthermore, we have developed diagnostic tools to assess how much additional ML-generated data and/or human coding reduce uncertainties in the main analysis.
We illustrate the effectiveness of our method by revisiting a study on the sources of election fraud with ballot image data and regression analysis.