0
$\begingroup$

I have ordinal data on three IVs ranging from 1 to 5 as below:

IV1: Not at all Important - Very Important
IV2: Not at all Satisfied - Very Satisfied
IV3: Performs much Worse - Performs much better

As I have some missing data points I ran an Imputation in SPSS. Odd thing is that in my regression almost all my imputation rounds have significant Coefficients except the Pooled one where the coefficients are all highly insignificant.

Any thoughts on how I should interpret this?

$\endgroup$
8
  • 1
    $\begingroup$ What fraction of observations have complete data? Do you get significant variables when you do a (non-recommended) complete case analysis? What imputation method did you use? How was the pooling done? You provided very little information on methods you are using, sample size, collinearities. $\endgroup$ Aug 28, 2014 at 14:17
  • $\begingroup$ Thanks Frank! 1. between 10-20% 2. Yes. 3. Automatic in SPSS 4. Not sure but I had 5 imputations. Pardon my lack of info: Sample Size 252, Collinearities: I had VIF below 10 on all variables. $\endgroup$ Aug 29, 2014 at 12:22
  • $\begingroup$ "Automatic in SPSS" is not a description of the imputation method. $\endgroup$ Aug 29, 2014 at 15:59
  • $\begingroup$ I choose Automatic in SPSS and therefore I cannot (unfortunately) say which one the program ran.. $\endgroup$ Sep 1, 2014 at 6:14
  • $\begingroup$ So you feel comfortable using an undocumented method? Have you at least run simulations to see if the method works? In general I can't say it's good practice to take a completely black box approach to something. $\endgroup$ Sep 1, 2014 at 12:29

1 Answer 1

1
$\begingroup$

You seem to assume that there is a recipe book the statistician consults that has a list of what's right and what's wrong or how to solve this. A few seconds of Googling for SPSS documentation sheds some light on their imputation methods (I'm not familiar with SPSS). And notice you have one of the foremost "experts" responding to you above.

$\endgroup$
3
  • $\begingroup$ Thanks for the link Thomas! This was very helpful! I choose the fully conditional model (as the missing data is arbitrary) and now my pooled model looks better in terms of significance. I'm well aware that there are many factors influencing statistical decisions and that there is no single answer to any statistical question except "it depends". $\endgroup$ Sep 3, 2014 at 6:40
  • $\begingroup$ @GentlemanEddie "looks better in terms of significance". There's no need to play the significance game: mchankins.wordpress.com/2013/04/21/still-not-significant-2 $\endgroup$ Sep 3, 2014 at 14:07
  • $\begingroup$ Hehe, I guess not. Interesting read that link! Personal favorite: "just tottering on the brink of significance at the 0.05 level" But what I meant was that it is more consistent with the rest of my results. For my purpose, this will do just fine. Thanks! $\endgroup$ Sep 4, 2014 at 6:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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