Using SVYSET for a large survey with final and replicate weights using jackknife

[Julia Merlin]

Hello,

I'm new to posting at Statalist (long time lurker during my master's!), and would appreciate any help that people can offer.

I'm trying to survey set a large, nationally representative sample survey in Stata 16.1 to add a column of data to an existing, published table. The data file contains a set of 200 replicate final interview weights (using jackknife) and one full-sample final interview weight. I used the code below to surveyset my data, and ran a few tabs to see if I could get the same results as the published table before starting any new analyses. I have also added the code I used for tab, in case that's my issue somehow.

I'm getting results that are *slightly* different -- my UNW/obs numbers are the same, but my estimates and SE's are a little off (hundredths place difference, but it matters when rounding), so I'm thinking my issue is with the weights.

Could anyone tell me what I'm doing wrong?? I did 1:1 merge two files to start these analyses, but they both have the same replicate and final weight vars.

svyset [pw=AFNLWGT], jkrw(AREPWT1-AREPWT200) vce(jackknife) mse
svy jackknife: tab A0204, se obs percent

[David Radwin]

Good for you for trying to reproduce published statistics, including standard errors. Too few researchers take this important step, which can lead to undesirable outcomes. Since you found discrepancies, here are some suggestions to try in order:

1. If the data documentation includes instructions for Stata, make sure you are following them. (I suspect you have already done this if possible.)
2. Contact the data provider for assistance. For example, if it is a National Center for Education Statistics (NCES) dataset, you can contact the NCES Help Desk at (800) 677-6987 or [email protected]. Inquiries like this can be helpful to the data provider, too, in case there is an error in the data or publication or as an indication that the instructions for Stata users are not adequate.
3. Look for other publications with estimates and standard errors and try to reproduce those.
4. If all else fails, consider the possibility that whoever made the tables may have done something differently than you. Perhaps they made a mistake or there was a minor revision to the dataset after the tables were published that was not documented. Or they excluded (filtered out) a small number of observations in the estimates, perhaps due to missing values, without noting so. (This outcome could still yield identical numbers of observations in some circumstances.) These things happen, and sometimes they aren't discovered for years. I can speak from experience both making and detecting mistakes. In this event, I suggest briefly describing in a footnote or endnote the discrepancies you found, their magnitude (nearly negligible), and your unsuccessful attempts to resolve it.