State-level unemployment rates

CPS
1

I tried to find an answer to this, but I didn’t succeed in finding anything. Apologies if I missed a previous answer about this.

I’m working with the CPS basic monthly microdata to look at state-level employment/unemployment trends for sub groups (e.g., race/ethnicity). Due to the sample size issues, I’m doing this for quarters rather than single months.

As a check to make sure I was using the data correctly, correct weights, etc., I tried to replicate the published BLS numbers for the not seasonally adjusted unemployment rate for the U.S. for May and for June. I was able to do this successfully.

I also was able to replicate the preliminary not seasonally adjusted unemployment rate for Virginia for May that was reported by BLS last month. However, my estimate for June is a good deal different from what was reported by BLS last week (about 0.8 percentage points difference).

So I think my questions are:

  1. Am I likely doing something wrong, or is it generally impossible to reproduce the state-level unemployment rates (and I just happened to match the published number for May)?

and

  1. As part of the BLS state-level employment releases, they initially provide preliminary numbers, and then those get revised the following month. Do you have any knowledge or insights into that process? The CPS basic monthly data don’t get revised, right? So what are the revisions based on, and do they show up in any public microdata?

Thanks.

2

The national unemployment rate numbers are based entirely on CPS. Because the CPS is less reliable for state-level estimates, BLS instead uses a model-based estimate for state-level unemployment rates as part of the Local Area Unemployment Statistics (LAUS) program. Because of this, the numbers produced from CPS microdata will not always match up with the state-level estimates from BLS.

The LAUS estimates use firm-level surveys from the Current Employment Statistics program and data on UI claims, which are revised after their initial preliminary release, so the initial LAUS unemployment rate estimate is also preliminary.

3

Perfect! Thanks. Exactly what I was looking for.

4

I’m doing the same calculation for Illinois: unemployment by race. I’d like to also look at the covid variables.

If I were to produce my own estimates using the microdata for Illinois is the method below acceptable? (using stata). FYI, When I ran the numbers using this method for July 2021, my unemployment rate estimate was lower than BLS’s which makes sense according to your response. I see others produce estimates with a caveat that these are “preliminary” I’m fine doing that with more recent data if I can look at data by race.

Using svyset:

svyset cpdid [pw = wtfinl]

unemployment:

gen unemployment = .
replace unemployment = 0 if labforce == 2 & age >= 16
replace unemployment = 1 if (empstat == 21 | empstat == 22 | empstat)

Race (white, black, latinx):

gen race2 = .
replace race2 = 0 if race == 100 & hispan == 000
replace race2 = 1 if race == 200 & hispan == 000
replace race2 = 3 if hispan > 0 & hispan < 901

svy: tab unemp
svy: tab if race2 == 0 (and so for blacks and latinxs)

or (without svyset)

tabstat unemp [aw = wtfinl],by(race2)

race2 Mean

0 .0499294
1 .1584633
3 .0636967

Total .0665681

It’s a lower estimate

In terms of the covid variables, I’d do basically the same thing as unemployment. Here’s the procedure I’d take for working remotely due to covid:

gen remote = .
replace remote = 0 if covidtelew == 1 | covidtelew == 2 (those in the universe)
replace remote = 1 if covidtelew == 1 (those who did not work remotely)

tabstat remote if unemp == 0 [aw = wtfinl],by(race2)

race2 Mean

0 .8413992
1 .9020904
3 .905842

Total .8603091

Thanks!

5

It is beyond the scope of the IPUMS User Support team to review code or analytical decisions of individual researchers, so I won’t give an opinion of the “acceptability” of your method. I will note that it is common to pool data from multiple months of the CPS survey when presenting statistics for subgroups or geographies lower than national. Although this gives estimates that are less current, they are also less noisy. For example, Minnesota reports unemployment rates for demographic subgroups using a 12-month moving average of data from CPS.