Hello! This feels like a bit of a silly question, but after reviewing previous recommendations in the help section for creating 5-year ACS samples by adding the 1-year ACS samples and dividing by 5 (which I read produces slightly different results than IPUMS published 5-year samples, but more or less consistent), is it safe to assume that new / average replicate weights can be calculated using the same approach? It feels too easy…
Asking with the intent of creating 2013-2017 5YR data in advance of IPUMS mid-January release.
Thanks much!
OMG I DEFINITELY wouldn’t try that at home.
Roughly speaking, the replicate weights work by faking the sampling process and taking some sample units in while keeping others out of sample, in a way that resembles the actual sampling design. Since I don’t know exactly how the Census Bureau pulls the sample, I simply cannot recreate the patterns of units that would need to be taken or omitted together in a meaningful way. (And my understanding is that the samples from the different years are not independent – at the very least, if you were sampled in 2015, you probably won’t be sampled again for a few years; so even if I had some idea about the units taht could go together in 2015, it would help me much how these units could interact with other units from 2013, 2014, 2016 and 2017 in attempting to form my own version of the 5-year product.)
Just taking the 1-year replicate weights and dividing them by five is basically telling the software that a certain group of units is going together in the joint 2013-through-2017 sample. So if in a given replicate, units 123 and 345 from year 2013 are included together with units 678 and 789 from year 2017, you are making a statement that this is how Census would have sampled them. But in reality you don’t really know that.