How to standardize Likert Scale in Stata

[Tian Xin]

Dear Statalists,

Currently I'm working with a "self-rated health" (SRH) variable from a panel data. However, the answer designed for the SRH question is inconsistent over the years. The question is "How would you rate your health status", while the answer for Wave 1 was "1 healthy 2 fair 3 relatively unhealthy 4 unhealthy and 5 very unhealthy" and for Wave 2 was "1 excellent 2 very good 3 good 4 fair 5 poor". I would like to ask, how would I standardize the variable to a five-points scale, to make it comparable across the years? Many thanks in advance!

Cheers,
Tian

[Nick Cox]

So far as I can see, on this information, the definitions changed. No white magic can reverse or correct for that. I

[daniel klein]

I agree with Nick.

You could try and run CFA models for the different years and see whether you can establish some kind of measurement invariance to support the claim that you are still measuring the same thing, despite different wording. Sensitivity analysis of any kind might also be helpful. However, be prepared for critical comments from reviewers how may or may not buy such an arguments.

Best
Daniel

[John Mullahy]

Tian: Beyond what Nick and Daniel have suggested, I'd note that there's a literature on issues akin to the one you are confronting that might be helpful for you to consult. For example: https://link.springer.com/article/10...654-008-9287-6

[Marcos Almeida]

Building on excellent replies from previous messages, perhaps you should consider item response theory. For example, you could compare answers for the whole data or by year.

[wbuchanan]

In a way this is parroting what Nick has already mentioned, but I don’t think you’ll be able to make a reasonable enough argument about the change in the scales. The big issue is that the response sets exist in different dimensions. The IRT tools in Stata currently are for univariate models only, so they would necessarily be a good solution. You could also consider creating a series of latent indicators for each degree of the response set as a way of modeling the error in the measurement and constructing the rest of the latent scale with those values.

[daniel klein]

The IRT tools in Stata currently are for univariate models only,

I think (or hope) that you can get around this calling gsem directly?

Best
Daniel

[wbuchanan]

daniel klein
I'm honestly not sure. It would probably be really easy to screw up setting up constraints and things like that. The only other possible way I could think about approaching it would be to treat the two time periods like demographic groups and test DIF across the two groups on that item. It still doesn't address the dimensionality issue, but would at least provide an argument that the change in the dimension didn't affect the scale or create any parameter drift issues.

[Tian Xin]

Thank you Nick, Daniel, John, Marcos and Wbuchanan for all the excellent replies. I think the key message is that we can hardly correct or rescale the answer due to the inconsistent measures of SRH. What we can try is to argue that we are measuring the same thing if it is the case.

I'm not familiar with confirmatory factor analysis, item response theory... But I will have a good look at them and try. Also will look at the paper suggested by John. I will come back with some results or, more likely, some questions. Many thanks in advance.

Cheers,
Tian

[wbuchanan]

Tian Xin
any info about the direction you ended up taking in the end?