What is the difference between person and replicate weights in ACS?


I am comparing demographic characteristics for a county across multiple years of the ACS. I want to make sure that my standard errors are correct so I can perform tests of statistical significance, but I am not sure whether I should use person weights or replicate weights. What would the command be to apply person weights in Stata? I’m not sure what the cluster should be, a household?



Please disregard this question. I’ve figured it out. However, I would be interested in seeing model code for applying sample weights in Stata. Currently, I am using the following code. I have recoded the variable educd into a set of indicators.

svyset serial [pweight = perwt]

svy linearized : mean ed_lths ed_hs ed_AA ed_BA, over(year)

The problem is that when I use a recoded dummy variable that uses only one value of the original variable, e.g. ed_BA where ed_BA includes only educd==101, the command does not generate results. I think this has to do with the person weights because I can get a result when they are not being used.

I don’t understand why the weights cause this to occur. I wonder whether the procedure I am following is incorrect. I am calculating these statistics for a single county.



While person weights will produce correct point estimates (e.g., means, proportions), you can use additional statistical techniques that account for the complex sample design to produce more accurate standard errors and statistical tests. Using replicate weights generates the most accurate standard error calculations. You can find more information about replicate weights here, along with a link to sample STATA code. While not as accurate as replicate weights, we also make available STRATA and CLUSTER variables for purposes of calculating standard errors. You can find more information about using these variables here, including sample STATA code. This STATA code may help resolve your follow-up question.

Hope this helps.