Principal components analysis for complex survey data

[Prest]

Hello here.

I am analysis national survey data with a complex study design therefore I would like to incorporate sampling weights

and account for clusters and stratum.

I am using food frequency questionnaire data. I derived 24 food groups from 112 food items. I want to use Principal

components analysis to derive Dietary patterns. Please which procedure can I use to  to run principal components

analysis for this kind of study design? 

I know the proc pls procedure and the proc factor , but these are seem to be for data obtained by simple random sampling.

Any assistance would be accepted.

[PaigeMiller · Accepted solution]

You would need to compute the survey weights yourself (or perhaps you already have them), and save the weights in a variable in the data set. Then you could include this variable in the WEIGHT statement within either PROC PRINCOMP or PROC FACTOR. There is no WEIGHT statement in PROC PLS.

[PaigeMiller]

Adding

The usual reason you want to perform principal components on the data is so that you can analyze the data in fewer dimensions, such as by plotting in fewer dimensions to detect outliers or clusters, or to use the fewer dimensions in some subsequent analysis.

The usual reason you do complex surveys is because you have clusters or strata that you can't (or haven't) sampled using the same proportion in each cluster or strata, and the survey weights allow you to estimate some population parameter taking into account the amount of sampling in each cluster or strata.

However, putting the two together, it sounds strange to me to use the survey weights to estimate "population principal components", as I don't see the purpose of surveys and weighting compatible with the idea of principal components. That doesn't mean you can't do it, but it does mean I'm not comfortable doing this (you may be comfortable doing this).

[Prest]

Thank you so much.

You seem to imply that including the complex survey information may not be necessary during derivation of Principal components. That the procedure stays the same as in simple random sample data analysis.

Did I get you right?