I am comparing the difference in shell area (mm^2) between two species of clams under ambient environmental conditions over a one year period using monthly measurement intervals. Unfortunately, part-way into the experiment the clams of both species started experiencing heavy mortality (i.e. “Species 1” – 38% mortality at month 4, and 94% mortality at month 8; “Species 2” – 7% mortality at month 4, and 50% at month 8).
As a consequence, I have many missing values in my data set due to this mortality.
I know that SAS usually drops subjects with missing variables in PROC MIXED and GLM.
Does anyone know a way to do a Repeated Measures analysis on growth measurements without loosing the data from the dead organisms?
Thanks, but I'm positive SAS drops subjects with missing data for PROC ANOVA, GLM, and MIXED...I've tried doing my analysis with all three. If you know of any other procedure that might be helpful though, I'd like to hear it.
I wonder if I'm misunderstanding the issue because as far as I understand things, you should be getting different results with PROC GLM and PROC MIXED for repeated measures. In the Class Levels section of the GLM output, are you getting a note that observations have been removed due to missing data? In the Dimensions section of the PROC MIXED output, how many observations are used and not used? Does the number used match the total number of non-missing values or just for the subjects with complete data?
I found this in the SAS Course Notes for Mixed Model Analyses:
"In the presence of missing repeated measures for a subject, the MIXED procedure does not exclude this subject from the analysis; instead it uses all the available data. This method (likelihood-based ignorable analysis) leads to a valid analysis when the missing data can be assumed missing at random (MAR)."
Unfortunately, it doesn't sound like your data is missing at random, so there might be better methods than PROC MIXED anyway.
There is an example in the third link that shows how GLM drops 4 of the 8 subjects because of missing data, but MIXED uses the available data from all 8 subjects. The two methods produce different p-values because of this.
Thanks for the help!
Actually, It was I who was misunderstanding the issue. I started to finally figure that out, and then your response confirmed it for me. Thanks for the links as well, I really appreciate it!