My study is designed in such a way that I have individuals with varying numbers of observations of the predictors, but only one observation of the outcome. Like a multilevel model, but with the outcome at the 'group' level. Following the conclusions of Foster-Johnson and Kromrey (2018), I plan to conduct an OLS regression analysis of group means, with White's heteroscedasticity adjustment.
My concern lies in the fact that the number of observations of the predictors differs massively between individuals... the range is between 2 and about 9000. I have read that the OLS estimators would be unbiased, and the correction produces robust standard errors. But I can't shake the feeling that perhaps there are additional ways to address this huge difference in group size. Calculating a weighted mean for the OLS maybe? Or is the correction really enough? Would love to hear thoughts or recommendations on how to approach this. Thanks!
Foster-Johnson, L., & Kromrey, J. D. (2018). Predicting group-level outcome variables: An empirical comparison of analysis strategies. Behavior research methods, 50(6), 2461-2479.