Dear all,

I am using household panel data across three waves (2011, 2013, and 2015). Although the vast majority of households surveyed in 2011 appeared in year 2013 (with an attrition rate of < 5%), there exists a very large attrition rate (about 61%) between 2013 and 2015. It is most likely that the data is Missing Not At Random (MNAR), because households in some regions of the study country were affected by political unrest in year 2015. A closer look at the data indicated that most of the sample households left in the final round were in locations affected by the political unrest in the country.

Currently, I am focusing on the households that remained in the sample, but still worried whether a potential attrition bias could impact my analyses. So far, I have tried to deal with attrition bias using Inverse Probability Weighting (IPW) approach. Prior to calculating IPW, I tested whether attrition in my panel data model is random using a probit in which the dependent variables takes the value one for households which drop out of the sample after the first wave and zero otherwise. The test result indicated the attrition is nonrandom.

Yet, I am not quite sure if using IPW approach indeed addresses an issue arising from such a large attrition rate (>60%). Does anyone suggest whether using IPW approach addresses this attrition bias? Do you have any suggestion if this is an acceptable rate of attrition or whether there is statistically recommended/reasonable rate of attrition in general?

Many Thanks for your help!
Abebayehu