Using all available data with missing data in repeated measures
I have data from a pre-post design, but for some subjects I only have one data point, not both. Is there a way to analyze these data in SPSS that can make use of all available data?
Resolving the problem
Yes, you can analyze these data using the MIXED procedure (Linear Mixed Models in the menus). Time is specified as a fixed effect and as the indicator of a REPEATED effect, and either the compound symmetry (CS) or heterogeneous compound symmetry (CSH) covariance structure will be appropriate for paired data. In order for this analysis to be valid, the missing data mechanism must be MAR (missing at random) or MCAR (missing completely at random), and not NMAR (not missing at random). For a discussion of these mechanisms and their implications for modeling of repeated measures data, see Applied Longitudinal Analysis, by Garrett Fitzmaurice, Nan Laird, and James Ware, 2004, Wiley.
Another alternative is the GENLIN procedure with generalized estimating equations (GEE). You would similarly declare time as repeated, typically with exchangeable specified as the structure for the working correlation matrix (exchangeable is the same as compound symmetric). GENLIN offers generalized linear models, so this is an option for binary, ordinal, count, and other non-normally distributed data. However, missing data assumptions for GEE are MCAR, rather than simply MAR.
Beginning with Release 17, you can also use the multiple imputation options in the SPSS/PASW MVA option to create complete data sets for analysis in any appropriate modeling procedure.
Translate this page: