0
$\begingroup$

I am running a planned missingness design to pilot some items for a questionnaire I am designing. Specifically, I want to test 80 items and every participant (N = 300+) receives a random 10-item subset of the 80 items. This, however, leads to 20-40 data points per item, thus ~90% of missing values.

As I want to evaluate the unidimensionality with an efa and want to potentially calculate reliabilities, I was wondering whether I could use multiple imputation for the missing data. From my design I know that the data is missing completely at random and I have another variable measuring the same construct which I could use as an auxiliary variable.

Would multiple imputation be a reasonable approach here or shall I just skip the EFA and go with means and SDs for item difficulties?

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ This sounds a little challenging. One issue here is that if questions are designed to / really do measure separate concepts, then what you have available would not contain (much) information that let's you impute the missing values. If they are somewhat (linearly) related, then it will impute especially the part that is related (but miss out on what is not related). Isn't that like that? $\endgroup$
    – Björn
    Commented Jun 21, 2023 at 16:48
  • 1
    $\begingroup$ When you do missing data analyses, there are discrete groups defined by their "missingness patterns". Each unique combination of missing and non-missing variables is a missingness pattern. Having lots of missing data doesn't turn out to be a problem, but having very few people with overlapping missing patterns or nearly complete data does turn out to be a very huge problem. It should be obvious that it will be very hard, nigh on impossible, to find a reliable prediction of response aside from grand averages. $\endgroup$
    – AdamO
    Commented Jun 21, 2023 at 22:08

1 Answer 1

Reset to default
0
$\begingroup$

Given the large number of items relative to the (relatively) small sample and the large amount of planned missing data, an EFA may not give you valid results. You could run a Monte Carlo simulation study with your planned missing design to see whether you can expect unbiased results under the fairly extreme conditions of your study.

Improve this answer
$\endgroup$

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.