Using weights: when and how?

[Giorgio Piccitto]

Dear all, I would like to know your opinion about the use of weights (an issue totally new to me).

I am running some analyses (both descriptive and OLS regression) on two appended dataset. Each of them contains a variable which is its own probability weight: how should I use them in a single regression (where I use the combined data of both datasets)? And also, should I use the weights also in the descriptive analyses?

Thanks in advance for your support.

Best, G.P.

[Dirk Enzmann]

This is (at least in the first place) not a matter of opinion. Although you may hope for an easy and effortless way to understand the issue, unfortunately there is no Nuremberg Funnel magically sparing you the effort to read and learn by yourself. As a start you should read the Stata manual [U] 20.24 Weighted Estimation, but if the topic is totally new to you, you should read some introductory material, such as

• Vallian, R. & Dever, J. A. (2018). Survey Weights: A Step-by-Step Guide to Calculation. College Station, TX: Stata Press.
• Lohr, S. L. (1999). Sampling: Design and Analysis. Pacific Grove, CA: Duxbury Press.

[David Radwin]

Another suggestion:

Bollen, K. A., Biemer, P. P., Karr, A. F., Tueller, S., & Berzofsky, M. E. (2016). Are survey weights needed? A review of diagnostic tests in regression analysis. Annual Review of Statistics and Its Application, 3, 375-392. https://www.annualreviews.org/doi/ab...-011516-012958

[Marcos Almeida]

The previous replies IMHO are clarifying enough. Your question seems to deal with two different issues. One, the theoretical background for using - or not - weights in regression analysis. Two, the command to apply weights in Stata. Last but not least, since you remarked you are dealing with probability weights, you are probably (sorry for the pun) dealing with survey-structured data. Being this so, probability weights are part and parcel of survey data analysis.

[Giorgio Piccitto]

Thanks all for your suggestion, I will go through this literature trying to build my own knowledge on the issue.

Sorry if I was not so clear in expressing my doubt: I was wondering indeed how should I apply in an appended dataset two different weights, referring one to one survey and the other to the second (the two parts of the combined dataset).

Could you suggest me something?

Thanks and best, G.

[Steve Samuels]

Create a new variable equal to the weight variable in each survey and use that.

[Giorgio Piccitto]

Dear Steve,
how can I combine two weights in one variable?

Thanks for the suggestion, and sorry for the question.

[Steve Samuels]

gen   new_weight= survey1weight if survey==1
replace new_weight = survey2weight if survey==2

This calculation assumes that both weights were not scaled (e.g. transformed to sum to sample size or a fixed constant) or were scaled in the same way. To find out, look at the results of

total survey1weight
total survey2weight

[Giorgio Piccitto]

Dear Steve,
thanks a lot, I will try your solution.

Best, G.

[Bryan Sayer]

If by appended, you mean one joined after the other (like two DHS countries) then there really is just one weight variable and you just need to rename one of them or do what Steve says. However, there is discussion of whether you need to rescale the weight when combining surveys. I think it depends on what population you are making inferences about. But that is a discussion of significant involvement.

However, if you mean two surveys of overlapping populations that are merged together (like men and women in a DHS survey to make a couple) then you probably want the weight corresponding to the smaller population. But again, this is a significant discussion.