Fixed vs. random effects for browsing data - a simulation

  When you work with trace data — data that emerge when people interact with technology — you will notice that such data often have properties that open up questions about statistical modelling. I currently work with browsing records, obtained at several times from the same users (i.e., a panel data set). A first typical characteristic of such data: Browsing behaviors are skewed. If you’re interested in people reading politically extreme web sites, you will find a few people doing it a lot, and most people doing little to none. A second characteristic relates to the panel nature of the data. If you’re looking at people’s visits to online shops, they do not change their habits much over time — at least these within-person differences are not as great as the differences across people. Panel data lend itself to hierarchical modelling, for example with (1) fixed effects (FE) or (2) random-effects multilevel modelling (RE). For commonalities and differences between these two m