What do negative replicate weights mean?

This is a follow-up to a previous question.

I am trying to obtain estimates related to poverty by race from IPUMS-USA based on ACS 2012-2016 5-year sample (e.g., the percentage of African American children below the poverty level within a certain PUMA). Due to the small PUMA geography, the race/ethnicity stratification, and the fact that I am only interested in children <18 living in poverty, my cell sizes are quite small in some cases.

I have tried 2 approaches:

  1. Using strata/cluster/person weight. I recognize that this will give slightly smaller standard errors. However when I did this, all estimates that were equal to 100% had a standard error equal to zero. This seems strange, since the numerator/denominator for all of these observations (the weighted frequency) was less than 200. For such small numbers, it seems suspect that the SE would be 0. Do you have any insight as to why this is happening?

  2. When I tried using replicate weights (see SAS code below), I received an error message in SAS that my replicate weights are either missing or negative. It turns out many of them are negative. What does this mean, and how should I deal with this?

proc surveyfreq data=set1;

table race*poverty;

weight PERWT;

repweights REPWTP1-REPWTP80 / jkcoefs= 0.05 ;


I have read the IPUMS technical documentation for replicate weights and could not find an answer to the meaning of negative replicate weights.

Thank you!

In the first approach, the standard error will be zero because there is no variation in your dependent variable (poverty status). At least from the description above and without looking at your code/data this seems to be a reasonable explanation for this observation. If you’d like, feel free to send your code to ipums@umn.edu and I can take a closer look.

In the second approach, specifically regarding negative weight values, these values help correct for issues related to survey non-response and other practical considerations that arise when collecting ACS data. One solution is to set all negative values equal to zero. More detailed documentation, provided by the Census Bureau, is available here. This document includes SAS code that can be adapted to calculate replicate standard errors manually.