Creating summative measures while identifying/imputed missing data

[John Brent]

Hello All –

I am trying to create a number of summative scales (egen newvar = rowtotal (var1 var2…). When doing so, however, missing data/cases are retained once the scale is constructed (because it has some value from another var). With that, how would I compute the summative measure only for those respondents that have no missing values, then multiply impute the values for the rest of the sample?

Thank you,
John

[Rich Goldstein]

I don't understand - it sounds as though some of your observations have missing data on some items that make up the scale and some have no missing data on these items; is that correct? if yes, then since you have decided on passive imputation (rather than "just another variable" imputation), you should run -mi impute- using all the data; then when you sum, those with no missing data will have exactly the same result as if you had broken the data into 2 pieces while those with imputed values can now be summed; note that using -mi impute- on just some of the data may be inefficient or even biasing but you don't supply enough detail to be sure; if I have misunderstood, please clarify

[John Brent]

Hello -

Yes, some observations have missing data on items making up the scale and others have no missing data on these items. Note: I have not imputed anything at this point.

With that, when I construct the final summative measures (added together from single items) to use in models, descriptives report as though the cases within missing data are not missing anything at all (because they have values added together from other items making the scale).

So, do I create the summative measures from all single items then mi impute? Doing this leads to the issue above where final scales have nothing to impute because cases with missing data still have values from other single items they have data for.

Does this help to clarify?

[Rich Goldstein]

you don't say how you summed but my guess is that you used -egen- which, in many cases, will just ignore that some of your variables have missing values; you, however, need to decide whether you want to impute the missing values for the individual items (generally called passive in the lit) and then sum or whether you want to sum, making sure that the total is missing if any (or at least some minimum number of) items have missing values and then impute the total (often referred to as "just another variable" in the imputation lit - note that neither of these strategies is perfect but you have to choose which you want - there is a lot of lit on this issue; one place to start, as it discusses this issue, is van Buuren, S (2018), Flexible imputation of missing data, second edition, CRC Press

if your issue is that you want the sum to be missing if any (or if a certain number of) items are missing, you can use other -egen- function (e.g., rowmiss or rownonmiss) to get a count of the number (non)missing and then use that

[John Brent]

Thank you again for the insight and follow-up. The initial post identifies that egen was used to construct the summative measures. While I understand the passive approach (imputing at the item level and then constructing summative measures, I do not understand the JAV method. It seems to me that once items are summed, cases with missing data will have some value for the final summed variable (see below for example):

id	item 1	item 2	item 3	item 4	item 5	summed var
1	1	. (missing)	1	0	. (missing)	2 (out of 5)

Imputing for the summed measure, using the JAV approach, would then impute nothing for the summed measure seeing that there is a value – though missing data is present within its composition?

I am 100% likely getting this wrong – both at the conceptual and practical level. So, any guidance and STATA how-tos would be appreciated more than you know!

Thank you once more!