GENLIN fails to produce pooled parameter estimates with multiple-imputed data set
When running the GENLIN (Analyze-Generalized Linear Models) procedure on a multiple-imputed data set, I have been unable to obtained pooled estimates, although the results for the original sample and all of the imputation samples appear to be stable and reasonable. For example, I ran a logistic regression model in GENLIN with a data set that had 5 imputation samples. Estimates were provided for the original data and all 5 imputed samples. However, the Pooled section of the Parameter Estimates table was empty, with the footnote:
'At least one of the vectors or matrices in the model being added is different in size from the corresponding vector or matrix in previously added models. This model cannot be added.'
When I used the LOGISTIC REGRESSION command (Analyze->Regression->Binary Logistic) to run the same model on this imputed data set, that procedure produced the same parameter estimates as GENLIN for all predictors in the original sample and all 5 imputations. The LOGISTIC REGRESSION run also printed pooled parameter estimates.
Why does GENLIN fail to produce pooled estimates for a well-defined model? Would it be appropriate to manually pool the results using Rubin's rules?
Resolving the problem
This problem was filed as a defect and fixed in Release 18.0.1. The error occurred because the pooling of estimates was halted when any of the coefficients had an empty standard error ("Std. Error") cell. The parameter labeled "(Scale)" had no standard error in the affected models, but this condition is legitimate for many of the models available in GENLIN. The pooling was successful with the LOGISTIC REGRESSION procedure results for the same model because that procedure does not print a "(Scale)" Parameter, so all of the nonredundant parameters had nonempty Standard Error cells.
In Release 18.0.1 and later versions, the pooling is not halted when the Scale parameter has an empty standard error cell. When the model is only available through GENLIN, applying Rubin's rules to the nonpooled parameter estimates is a workaround for Releases prior to 18.0.1.
Translate this page: