Many socio-economic studies rely on panel data as they also provide detailed
demographic information about consumers. For example, advertisers use TV and web
metering panels to estimate ads effectiveness in selected target demographics.
However, panels often record only a fraction of all events due to non-registered
devices, technical problems, or work usage. Goerg et al. (2015) present a
beta-binomial negative-binomial hurdle (BBNBH) model to impute missing events in
count data with excess zeros. In this work, we study empirical properties of the
MLE for the BBNBH model, extend it to categorical covariates, introduce a penalized
maximum likelihood estimator (MLE) to get accurate estimates by demographic group,
and apply the methodology to a German media panel to learn about demographic
patterns in the YouTube viewership.