# income elasticity + regression + household survey data

I'll try to be concise and up to the point:

Problem: estimate income elasticities for various products from household survey data.

Given: I have to big complex household survey datasets that uses two stage stratified sampling. However, I have the design specified (strata variable) for only one, and replicate weights for the other data set.

The idea: is to estimate some for of a regression expenditure_for_product ~ household's income and then use the coefficients to estimate it's elasticity

What is unclear: To estimate simple statistics as mean expenditure for a product or its S.E. one has to properly account for the survey design (especially for variance estimation). But I am not sure how if at all I need to adjust it in my regression case because the literature survey in fact does not provide a clear answer ( whether the weights (and correspondingly the survey design) should be used in my regression case)

Software: there is in fact R package survey which is capable of incorporating (not sure exactly how, as help files does not specify or I can't locate it) the survey design into regressions but this of course does not mean I actually need to do it in my case.

A small digression: For example, say that log(expenditure_for_product) ~ log(household's income) captures the relationship nicely. In that case, the regression coefficient would be all that I need. However, as there are many zero purchases reported by households for goods, I cannot model it in logs. I had an idea of aggregating the data into percentiles so that there would be 100 points with (average income of x percentile, average expenditure for some product of x percentile). But how can if at all should adjust this idea for survey design to get correct estimates?

In essence: In essence I am looking for a theoretically correct* way to calculate income elasticities given complex household survey data and thus references on this matter or suggestions would be very helpful.

*Note, I cannot incorporate price effects because I do not have price information available, so I realize that it will not be too precise.

Thank you!

it sounds like you need to run svyglm on a complex-sample survey design object that you created with the svydesign or svrepdesign functions?
if you are talking about the consumer expenditure survey, the entire setup has already been done for you all the way to an example regression call in the r language. just walk through that code and change the svyglm call to what you're studying. even if you're not talking about the consumer expenditure survey, it sounds quite analogous to what you need - so the code might be instructive anyway :) thanks
• hey, thanks for your answer Anthony. in fact, I have been using your website to get acquainted with how to use survey package to analyze complex design surveys, but this questions is not how to do, but whether, do I really need to do it? as I've digressed, if I decide to for example log-log model, then I would somehow transform the data (for example - aggregate into percentiles) to get rid of zero expenditures. but then I of course loose the information about survey design. but the key is - maybe I do not even need this information for my purposes? thanks for your thoughts though! – Sarunas Jun 14 '14 at 8:11
• after some thinking I concluded (without proof) that it would not be a good idea to ignore extra information that is available (survey design), so I accepted your answer. kids - use the survey package! – Sarunas Jun 22 '14 at 14:16