ERROR: Within-imputation COVB matrix has diagonal entries less than 0 for _Imputation_= 868 in

[cdriesse139]

Hello you all,

I have been trying to impute missing data using the delta adjusted pattern imputation to compare it to the control-based pattern imputation. 

Unfortunately I am getting the above error message which I am unable to solve. Could somebody advice?

My syntax is as follows:

proc mi data=milk.missingwide out=milk.FNDSdimp seed=128 nimpute=1000;
class sex single precovid randomization;
FCS discrim(sex/details) discrim(single/details) reg(FNDSbase/details) reg(FNDS1/details) reg(FNDS2/details) reg(FNDS3/details) reg(FNDS4/details);
mnar model(FNDSbase / modelobs= (randomization='0'))
adjust (FNDS1/ shift=1 adjustobs=(randomization='1'))
adjust (FNDS2/ shift=1 adjustobs=(randomization='1'))
adjust (FNDS3/ shift=1 adjustobs= (randomization='1'))
adjust (FNDS4/ shift=1 adjustobs= (randomization='1'));
var sex single precovid FNDSbase FNDS1 FNDS2 FNDS3 FNDS4;
run;

*reverting back to long dataset type;
data milk.FNDSdimplong;
set milk.FNDSdimp;
week=1; FNDS=FNDS1; output;
week=2; FNDS=FNDS2; output;
week=3; FNDS=FNDS3;output;
week=4; FNDS=FNDS4; output;
run;

proc mixed data=milk.FNDSdimplong;
class subj randomization sex single week precovid;
model FNDS = randomization|week FNDSbase sex single precovid /ddfm=kr2 solution covb;
repeated week / subject=subj(randomization) type=AR(1) ;
rANDOM subj(randomization);
by _imputation_;
lsmeans randomization*week /pdiff cl;
ods output SolutionF=milk.mxFNDSdparms covb=milk.mxFNDSdcovb;
run;

proc mianalyze parms(classvar=full)=milk.mxFNDSdparms covb(effectvar=rowcol)=milk.mxFNDSdcovb;
class randomization week sex single precovid;
modeleffects randomization week randomization*week FNDSbase sex single precovid;
run;

[ballardw]

We generally recommend including the code and and all the messages from the LOG when discussing errors so we can see exactly where SAS reported something and possibly other relevant information.

I this case look at your data for _IMPUTATION_=868 in data=milk.FNDSdimplong.

I suspect you will find something odd about the set of model variables for those observations.

[cdriesse139]

Thank you for your advice. Taking a closer look at the log indeed provided an answer to the question. While the covariance matrix that was the best fit for the non-imputed data, this covariance matrix was no longer the best fit when missing data was imputed.
Always good to see it with a fresh pair of eyes.