# Matching BLS Job Tenure Estimates

Greetings, I’m trying to match the BLS estimates for job tenure and am failing to match or even see consistent time trends. I don’t expect to match exactly, but these estimates seem pretty off. Wondering if others have insight on this.

I’m replacing JTYEARS to missing if >=99.96 and then producing estimates in Stata using iquantile. I limit to wage and salary workers age 16 and over based on the BLS universe.

`iquantile jtyears if age>=16 & !inlist(classwkr,13,14,26, 29) [aw=jtsuppwt], by(year) allobs`

I’m 4-2% lower than BLS in 2016-2020; 7% higher than BLS in 2020/2014; and 8% lower than BLS in 2010.

Would love it other have insight!

I do not have an answer for you currently, but I am looking into this and I hope to have more information for you next week.

I wasn’t able to find anything that would explain this discrepancy. I have some thoughts:

1. The issue could be with the method used for computing the weighted median. iquantile uses a particular linear interpolation method. I found different results using -epctile-. BLS may use yet another method.

2. BLS may have more detailed internal data on job tenure that they use for their calculations. Although there are observations with non-integer value for JTYEARS, they are relatively few. I wonder if the internal data may have more detail, since all of the BLS’s medians are in the middle of 4 and 5. However I’m a bit confused because according to this technical note, more detail was asked about months of tenure only for those who answered that they had worked at their current job for one or two years. So I wouldn’t expect medians that are fractions between 4 and 5, even with weights. That suggests to me that BLS may be using some other interpolation method, such as kernel density estimation.

I suggest reaching out to the BLS directly with this question.

Thanks, really appreciate your efforts on this! I ran the medians a few different ways in both Stata and R and consistently find these slight differences. I’m not overly concerned about the magnitude of the difference but was curious if others had guidance. My estimates of the percent of employees with 10+ years of tenure are spot on, so I’m similarly assuming this has to do with weighted median methods.

1 Like