Approximately 50% of cases are missing data on one of my predictor variables. With the default option selected (listwise treatment of missing data), the models produced are weak. This is probably because the listwise option reduces n substantially.

The alternative (pairwise exclusion), when selected, produces a strong model (the total variance explained is about 50%) with a number of significant predictors (the variable with 50% missing data is a significant predictor in this model).

However, this sounds a bit too optimistic. I've read that when pairwise exclusion is selected, SPSS will base degrees of freedom for significance testing on the number of cases with complete data (in this case, 32) rather than on the total number of cases. From what I understand, this means that the significant effects may be exaggerations.

Am I right to be concerned about the potential for exaggerated effects when pairwise exclusion is selected? Or are the parameter estimates (and the model as a whole) still trustworthy?

  • $\begingroup$ Can you provide a citation for the statement you read about SPSS's degrees of freedom being based on the listwise N? Also, if this is true, it will dampen, not exaggerate, significance as compared with results based on the total N. $\endgroup$ – rolando2 Sep 29 '12 at 11:29
  • $\begingroup$ For the pairwise option in regression, the SPSS manual reads: "Cases with complete data for the pair of variables being correlated are used to compute the correlation coefficient on which the regression analysis is based. Degrees of freedom are based on the minimum pairwise N." $\endgroup$ – Archaeopteryx Sep 29 '12 at 13:32

When you have so much missing data, the first concern is why the data are missing. They can be missing completely at random (MCAR), missing at random (MAR) or not missing at random (NMAR). Searching on missing data here, or on any of those terms in Google, should give you lots of information.

Neither listwise nor pairwise deletion are good options with so much missing. If the data are MCAR or MAR, then it is certainly worthwhile looking at multiple imputation. Even if they are NMAR, multiple imputation may be best.

I don't know about SPSS capacity with regard to multiple imputation (I am not an SPSS user) but both R and SAS have excellent abilities in this regard.

  • $\begingroup$ The data are missing because they were collected by two individuals, and one of the individuals didn't collect any data for that variable. Since data collection was fairly evenly spilt, this means that 50% of cases lack data on that variable (missingness of all other variables is less than 5%). I've been reading on various missing data treatments. Mean substitution does produce a good regression model (in this case). But many seem to advise against it. My worry is that, due to the large proportion of the missing data, generalizability will be severely limited no matter what I do. $\endgroup$ – Archaeopteryx Sep 29 '12 at 13:36
  • $\begingroup$ That sounds like MCAR, then, as long as the two individuals collected data on similar cases. If that is not true, then it's NMAR, possibly severely so. $\endgroup$ – Peter Flom Sep 29 '12 at 13:38

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.