Social Science Statistics Blog

Unless I am missing the point -- and sadly that occurs too often -- rounding is the same thing as adding a uniformly distributed "mismeasurement" term to the variable. Further because with rounding the exact distribution of the mismeasurement term is known, one can correct for the bias in the estimates. For example, in an OLS regression the bias is a reduction in the absolute value of the coefficients (other than the constant, whose variable values are presumably not mismeasured).

I quit reading JASA over 20 years ago so I am not going to try to get the referenced article and see what is being offered. I find it hard to believe that (again for instance) in regression treating the rounding as a missing data problem will improve on treating it as the actual variable plus the mismeasurement term whose distribution is known. But if that is the case, I would be interested in how and why -- maybe enough to travel to an academic library :-).

If you encounter this sort of problem regularly, it might be worth the trip to the library. I won't try to go through the paper on the blog, but sometimes you need heavier artillery than a mismeasurement term. Suppose, for example, that different people have different coarsening (rounding) patterns. Some people, for example, might round to the nearest half-year, while others might round to the nearest year. If you think this might be going on, you'll need to model the kind of person who's giving you the information before applying a rounding adjustment.