We will use this space each week in the news blast for Justin to share new ideas he thinks may be of interest for our behavioral & experimental community. If you ever have ideas for topics, please share your ideas with
This Week's BRITE Idea:
For this week's BRITE ideas we want to highlight a methodological issue that often comes up in studies. The issue is what to do if you observe an outcome at two points in time for the same person with a treatment in between -- should you a) use the change in outcomes (i.e., the difference) as your measure or should you b) run your regression with the second outcome and control for the initial outcome level? That is, should you difference or control for prior outcomes?
I saw a nice
discussing this issue and illustrating the fact that this involves bias-variance trade-offs. For observational studies, there is no obvious right answer. However, for most experimental studies where you have randomized treatments, the right answer is actually to control for prior outcomes rather than use differences. The reason is that using the difference is better for dealing with bias, but with randomization the bias should be small. In these cases, it is better to control because that improves the variance. I think this is a reasonably well known issue, but thought it was nice to have this blog post summarizing the points.