Composite Means (Dimension) Query

[Jamie Morano]

Hello - hoping that someone could advise on the best coding for determining a "dimension" scoring on patient safety that is comprised of 12 sets of grouped survey questions on a Likert Scale.
In short, we have questions labelled A1, A3, A4, A11 on a 5-point Likert scale (responses 1-5) called "dimension 1). We are following a patient safety model in which each question is scored 1-5 and then a means is created for each question. Thus, I am able to create a means for each question using "mean A1" but am unable to add these means for a first dimension (d1).

Thus far, we've looked at the manuals and online and have tried:

by idnum, sort: egen d1 = mean(A1) ---> this works, but unable to add the other means for A3, A4, A11, and STATA does not output d1.

by idnum, sort: egen d1 = mean(A1)

This also does not work well:
generate d1 = mean A1 + mean A3 + mean A4 + mean A11
generage d1=sum (mean A1 A3 A4 A11)

Each respondent (n=80) has a unique value from 1-5, and we're trying to get a "mean of the mean" for the collective dimension 1.

Thank you for any insights.

[wbuchanan]

Jamie Morano the mean is not mathematically defined in this context. Your data are ordinal in nature (e.g., the distance between the integers is not necessarily equal). Aside from the issues related to taking means of means (e.g., minimizing variance), your sample size is relatively small given what seems to be a highly dimensional survey. How many items are in each of the 12 groups? Do the items load onto unique factors or do some items load on multiple factors (e.g., would item 1 only load on dimension 1 or it is possible for item 1 to load on dimensions 1 and 2)? IRT and CTT methods are typically the way to handle scaling, but given the same size restrictions and dimensionality issues the easiest way to move forward would be summative scores within the dimensions (e.g., add the value of the responses within dimensions). There are a few reasons, but in short it requires fewer assumptions about the relationship between the distribution of observed responses and theta, is extremely easy for practitioners to interpret, and is a sufficient identifying statistic for IRT models that you might use down the road when you have a larger sample (e.g., a Rasch model specifically).

[Jamie Morano]

Thank you for your response and excellent points. Some clarifications and answers:

1) We are attempting to replicate the methodology as outlined in the following publication with a cohort in China (n=40 physicians and n=40 nurses): "Sorra JS, Nieva VF. Hospital Survey on Patient Safety Culture. (Prepared by Westat). AHRQ Publication No 04-0041. Rockville, MD: Agency for Healthcare Research and Quality. Sept 2004." Yes, the sample size is extremely small and certainly a limitation.
2) There are 12 "dimensions" of patient safety, which are the survey questions grouped into 12 thematic groups; each of the 12 dimensions is approximately 4 questions in the Likert scale. The items load into unique factors, i.e., each question belongs only to one thematic dimension group.
3) Agree that summative scores would be the best method. I am attaching the original article with the dimensions mentioned on page 54. Thus far, we have been able to code in Stata the methodology but are stumped in the part of trying to "sum the dimensions".

Any thoughts welcome and much appreciated. AHRQ.Patient.Safety.Survey.pdf

[wbuchanan]

From the documentation in the link (and assuming the variables are named similarly) you would do something like:

qui: egen event_reporting = rowtotal(d1 d2 d3), missing

Beyond that there isn't too much that would be useful. For the size of your total sample (n = 80) there are just too many items (n = 42) to do much. Another issue you may run into is that your sample is fairly different from the original study and you could end up measuring socio-cultural artifacts as much as the actual construct you are trying to measure (e.g., the values on the dimensions could be influenced based on socio-cultural norms in China that are distinctly different from the sample used in the Westat study). Beyond that you might be able to connect with the original authors of the study to see if they could provide item parameters that you could use to estimate the different values of theta and/or see if they've done any additional work testing measurement invariance/parameter drift across cultures.

[Jamie Morano]

Thank you so much for your assistance; we will try this.