1
$\begingroup$

I am using a certain 7 item Likert scale in my research and have collected my data - however I've now realised that I left one of the items (Item 7) off my initial survey. I have since collected all the data and am now looking to analyse it so there is no possibility of collecting this information retrospectively.

What can I do about this "Missing" data?

I have looked into imputing the data for this item based on mean substitution (or multiple imputation?) - however I'm not sure that this is really 'missing' data in that the participant didn't skip the question, it was just never included..

Should I be looking to impute this data for item 7 so that the total score for this scale is then comparable to past research? or should I just use the data collected for the 6-items and note this as a limitation?

$\endgroup$
4
  • $\begingroup$ What do you mean you left one of the items off? $\endgroup$ Commented Aug 28, 2019 at 12:04
  • $\begingroup$ It's a 7 item scale and when I was entering it into the online survey platform I somehow managed to leave off the last item. So I only asked 6 out of the 7 items on the scale. And usually, the 7-items would be summed to create a total score but I only have scores for 6 items. Hope that makes sense. $\endgroup$
    – LSSS
    Commented Aug 28, 2019 at 13:14
  • $\begingroup$ Do you have some individuals that answered item 7 at all? I'm thinking you might be able to use imputation to predict the missing values based on the people that have item 7 complete. If you don't have that though, there is no way to predict the missing data if it's missing for everyone. $\endgroup$
    – Emma Jean
    Commented Oct 2, 2019 at 0:50
  • $\begingroup$ I was wondering have you found a solution for this issue? I am having the same problem, I forgot to include 1 item from 7-items scale and I can't find any solution, unfortunately. $\endgroup$
    – lena
    Commented Jan 27, 2021 at 22:13

1 Answer 1

1
$\begingroup$

This is missing data by definition.

In missing data, people normally discuss three main cases. Missing at random means that the data can be considered to be missing completely at random, as if some TA wrote a program to delete 5% of the responses at random. This essentially never happens. Mean imputation would be proceeding as if the data are missing completely at random.

There is missing conditionally at random (MCAR). This means you are willing to assume that the data are missing, conditional on other variables in the regression model. For example, before the missing data, men might have had a lower mean score than women. Multiple imputation would account for this.

Not missing at random (NMAR) means the data are missing, and the probability of missingness depends on variables you can't measure. In real life, the data are probably a mix of MCAR and NMAR.

One complicating factor is that you left an entire question out. I think you can't consider the data to be missing totally at random. In multiple imputation, you would be imputing a few sets of values for that question conditional on how other people responded and conditional on the covariates you were putting into the regression model.

First, if half the people missed Q7, the imputation model bases its imputations based on how the other half responded. If Q7 is missing in its entirety, I think this is a problem on its face.

Second, another way to consider it is what proportion of the data are missing. One guideline in my field is that 5-10% missingness is something you can handle through multiple imputation. You have just under 15% of the data missing.

Third, each scale measures a latent construct. For example, in a widely-used depression survey, we have affect, cognitive, and somatic symptoms of depression. The last question is suicide ideation. If you totally deleted one question, you are slightly changing the nature of the latent construct you're measuring in a qualitative sense. Or, if I deleted the suicide ideation question, I've now lost the ability to measure very severe cases of depression - you could consider that to be a quantitative but not a qualitative change to the latent construct.

I know this is a rather late response, but I would vote for treating the scale as a 6-item scale and reporting the error. If this means a resulting paper is not publishable, then I would go back and re-sample people with the full questionnaire.

$\endgroup$

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