Which estimation method for missing data using PROC GLIMMIX?

[dsuchoff1]

I have a relatively small agronomic dataset that is normally distributed. We were not able to collect results from some plots in the field and thus have missing data. Should I be using a different estimation technique besides the default method=RSPL? Additionally, is it appropriate to allow for unbounded covariance parameters when analyzing an incomplete dataset?

Thank you,

David

[SteveDenham · Accepted solution]

If the plots are missing at random RSPL is fine as a method.  If the missingness is not at random, multiple imputation is probably the best choice.

As far as negative variance components, see the rather long thread here https://communities.sas.com/t5/Statistical-Procedures/Interpreting-the-random-effect-solution-in-a-m... .  @WillTheKiwi makes some good points about negative estimates of variance components, which I still have some issues with (but only on theoretic grounds),

SteveDenham

[WillTheKiwi]

Why are you using Glimmix, if the data (i.e., the residuals) appear normally distributed? I know you can use Glimmix for such data, but why not Mixed, which probably gives you more control over the structure of the residuals and less convergence and infinite-likelihood problems. And I always use ddfm=Sat in Mixed, which seems to give sensible degrees of freedom for everything. As for negative variance, yeah, I don't have any problem with it. When a true variance is small and the uncertainty is large, sampling uncertainty can easily result in an observed negative variance, and even if the sample variance is positive, the only way you can get sensible confidence limits is to allow them (i.e., the lower  limit) to be negative.