Multiple Imputation "cannot find an imputed value under the constraints for variable"

[Ri89]

Hi,

I am attempting (and so far failing) to use multiple imputation in SPSS to fill blanks in my data set.

I ran the multiple impuation using the default settings and it worked fine except for the fact I was getting values way below my detection limits, some values were even negative! Following this I ran a custom multiple imputation (fully conditional specification) and set the constraints so that the detection limit was the minimum for imputed values for each variable. However, this resulted in a message stating:

"that after .... draws the imputation algorithm cannot find an imputed value under the constraints for variable .... Check minimum and maximum values specified to determine that they are reasonable and consider raising the number of draws allowed. Execution of this command stops"

So I ran the anaylsis again with values even less than the minimum detection level and increased the number of draws (for both case draws and parameter draws) but still no luck.

Does anyone know where I am going wrong? Any comments would be much appriciated, thanks.

 

[spunky]

Re: Multiple Imputation "cannot find an imputed value under the constraints for varia

did you check the assumptions of whether you have an ignorable or nonignorable non-response pattern? which mechanism follows your missing data more closely, MAR or MCAR?

 

[Ri89]

Re: Multiple Imputation "cannot find an imputed value under the constraints for varia

Thanks for getting back to me.

My missing data is MAR. This is supported by the Little's MCAR test (although it did say in the output that "EM algorithm failed to converge in 25 iterations" - is this bad?)

 

[spunky]

Re: Multiple Imputation "cannot find an imputed value under the constraints for varia

uhm... the EM algorithm did not converge? that's very bad... did you try altering the number of iterations to more? did this happen before or after you changed the limits on the FCS (fully conditional specification) imputation method? thing is you may be choosing values that are either outside of the parameter space of the likelihood function or that are very close to the boundary... but you're not alone on that one. SPSS (at least on it's version 17) has been widely criticised in the statistical literature for it's inefficient estimation under missing data situations...

 

[Ri89]

Re: Multiple Imputation "cannot find an imputed value under the constraints for varia

This message came up before I attempted to change the limits on the FCS method while I was running missing value analysis to look at some descriptive stats on my data.

I just tried increasing the iterations to 100 and still got 'failed to converge', but it seemed fine when I increased the iterations to 1000.

I'm using PASW 18, is this any better than SPSS 17?

Thanks again.

 

[spunky]

Re: Multiple Imputation "cannot find an imputed value under the constraints for varia

i'd assume it's better... i use SPSS 17 and i remember reading some harsh words in the American Statistician when they reviewed their package for multiple imputation so i guess they would have fixed it... besides, PASW 18 was produced after SPSS was sold to IBM so i'm really, really hoping that the engineers in IBM took a closer look at some of the algorithms...

anyways, missing value imputations is a little bit of an art as much as a science. the EM-algorithm used to find the values is constrained by what seems a reasonable domain of the likelihood function... and such domain may or may not be interpretable in the context of real-life data. so if negtive values in your case are non-sensical well... the computer will still give them to you because they are a plausible mathematical solution to your analysis. nevertheless, this doesn't have to be a problem per se... for example, if you're attempting to run something like an OLS multiple regression, what you're reall after are the beta coefficients and the R2. you're not all that interested in the individual points that gave you that and, in such a case, you can still use the data to estimate those parameters. HOWEVER, if you're more interesed in stuff like prediction and you'll need to use some of those non-sensical values, then you do have a problem...

i honestly dont think you're doing anything wrong. it's just the way it is... multiple imputation can be pretty tricky sometimes...

 

[trueblu4you]

Re: Multiple Imputation "cannot find an imputed value under the constraints for varia

I feel totally lost.

I was using SPSS 15.0 and my computer crashed. I ended up not only buying a new computer but also had to buy SPSS 19. In a way I guess that's good because it has a way to do multiple imputation for missed data. Basically, with the SPSS 15 that I had, the best choice I had was the Trend method, but I have been told that the Trend method (even though the best out of the 4 choices I had) was too biased considering my data. Therefore, it was suggested that I use the multiple imputation method. I wasn't able to do that with SPSS 15, but CAN with 19, so I DO want to have accurate (or MORE accurate, less biased) results.

However, I am TOTALLY new to doing such a task and need help rather quickly. I clicked on Analyze --> Multiple Imputation --> Impute Missing Data Values. I went through each of the tabs, scanning the data and made sure the ranges they showed matched the actual ranges that my data needed to have for each variable. I also entered those ranges in the constraints fields. I did 5 imputations, 100 parameter draws, 500 case draws (I have no idea how I'm supposed to pick those last three values). I had SPSS save everything to a new file. I can't see where data were imputed other than I think it's where the fields are now highlighted in a pinkish color. However, there are a LOT of values that did NOT get imputed. So I don't know what I'm doing wrong. I know this is last minute, but I'm needing to get this data cleaned up before the end of the weekend so that I can re-run my analyses... and I've been working on this since I woke up this morning, pretty much non-stop and not making much progress at all.

Thank you SO much in advance for anyone willing to answer my questions. Just be aware though that statistics is NOT my strong suit at ALL, so you may need to phrase your responses in a "simplified" and "guided" way.

Thanks again.

 

[rastiola]

Re: Multiple Imputation "cannot find an imputed value under the constraints for varia

Did you solve this problem? I have a same kind of problem here and i don't know how to deal with this.

 

[Lazar]

Re: Multiple Imputation "cannot find an imputed value under the constraints for varia

Can you tell us a bit more about the problem variable, the exact error, and your missing data model.

In the mean time Amelia II (runs in R but has a GUI point and click interface) by gary king runs on a slightly different algorithm and can read in SPSS data files.