I think both versions are used in the literature and both refer to the same issue: what to do if scoring a questionnaire (and occasionally a rating scale) if items are missing. It’s actually a very important issue if allowing respondents to omit items though it may be becoming less important as increasingly measures are delivered through online forms and apps that don’t allow omission of items (so you end up with some people who might have returned a pro-ratable score returning nothing).
This is related to, actually the simplest approach to, the statistical issues involved when “imputing” missing data in analyses of aggregated data.
Details #
Ideally, all measures that are going to allow some omission of items should have stated pro-rating rules and all reports analysing data using any scale should be clear that the scoring followed the designers’ pro-rating rules (and it the report word count allows, should say what that was and perhaps how many responses were pro-rated). That very rarely happens.
So what is a pro-rating rule? It should always have two components:
- how many of the items in the score may be missing for pro-rating to be allowed (or what proportion of the items in the score, which comes to the same thing!)
- what to do with the completed items to get the pro-rated score.
For example, for all the CORE system measures our rule is the same and has these components.
- If more than 10% of the items that contribute to the score are missing you can’t pro-rate and the score is treated as missing.
- If fewer than 10% of the items items are missing take the mean of the remaining item scores (and multiply by 10 if using the so-called “clinical scoring”).
So for the ten item YP-CORE and CORE-10 measures scores can be pro-rated if only one item is missing.
I have occasionally seen a prorating rule in which missing items are treated as scoring zero (note: this is actually very different from the more typical rescaling). I have also seen some measures which allow to my mind very high proportions of items to be pro-rated. Another issue is that some measures, e.g. the 25 item RCADS (Revised Child Anxiety and Depression Scale) where there is pro-rating rule for the anxiety score and for the depression score but for the total score the pro-rating rule is that it cannot be pro-rated if either of those subscale scores is missing.
There are a lot of quite interesting theoretical issues about the pros and cons of different pro-rating rules, both ones assuming similar logic of responding is being used by most respondents (roughly the same as supposing a strong shared factor structure) and other interesting questions that arise if we thing that there may be subsets of respondents who may respond to the measure rather differently from other, or most, respondents. I have found to my mind remarkably little theory, simulation or analyses of real data exploring these issues.
Try also #
- Imputation
- RCADS (Revised Child Anxiety and Depression Scale)
- Scoring
- Standardising/normalising
- Transforming/transformations
Chapters #
Never explicitly covered in the OMbook.
Online resources #
In my shiny apps these apps apply pro-rating rules to raw data:
Dates #
First created 27.vi.26.
