Hi, hopefully I do a decent job of explaining my issue. What I have noticed is that the relative proportion of group insurance policy holders vs dependents to all covered by group insurance (among households with at least one member working at a firm of 1000 or more employees) switches in 2018.
(note: graph uses data, rather than survey, years)
I have been looking into the changes to the health insurance question methodology in 2018 and have applied the VERIFY adjustment on pre-2018 data by changing GRPOWNLY and GRPDEPLY “yes” answers to “no” when in the presence of a “yes” VERIFY answer. I have also applied the adjustment to the 2014 survey due to the 3/8-5/8 sample by removing respondents in the 5/8 sample. Otherwise, I haven’t executed any year-specific code. So, I’m not clear why this switch might have happened and would like to understand what might be the root cause–is there another adjustment that I haven’t applied that I should be aware of, or is it something else?
I noticed that the only survey years where respondents appeared as “not in universe” for both GRPDEPLY and GRPOWNLY are 2019-2023 (i.e., in conjunction with the appearance of the GRPCOVLY variable). I removed those observations, of course. Could it somehow be related?
Update: I see that NIU was only a possible value from 2019-on for GRPDEPLY, but not GRPOWNLY. Do you have any suggestions on how to correct for this?
Bumping this because I haven’t heard anything. Any ideas? TIA!
There are a number of reasons I would not generally expect estimates that use health insurance questions from the 2018 and 2019 ASEC to be similar. Changes to processing of health insurance data were included in the 2019 ASEC. The most significant change that would affect these variables is that a new imputation process was used that grouped people into health insurance units and filled missing data based on the characteristics of that unit. Another relevant change is that the universe of the variables GRPDEPLY and GRPOWNLY changed from being all persons in 1996-2018 to persons covered by employment-based insurance last year in 2019-2023. You can read more about the changes in this user note, this blog post, and this presentation given by Census Bureau staff.
Also, note that the variable VERIFY does not include any “yes” responses in the original data from 2014-onward. We strongly suspect this is due to the 2014 ASEC redesign, although no official Census Bureau documentation confirms this. For more information on the 2014 ASEC redesign, see Census Bureau 2014 ASEC redesign documentation. Using VERIFY to reassign the responses to GRPDEPLY and GRPOWNLY will result in comparability issues between samples before 2014 and samples from 2014-onward.
To try to replicate the change you are seeing, I downloaded an IPUMS CPS extract to look at the share and number of respondents who were the policyholder for employment-based insurance last year (GRPOWNLY) and the share and number of respondents who were a dependent covered by employment-based insurance last year (GRPDEPLY). The universes of both of these variables change between the 2018 and 2019 ASEC (mentioned in the previous paragraph), from all persons to persons covered by employment-based insurance last year. I recoded NIU values as “no” for both variables for the 2019-2023 samples. I did this so that I am comparing the same group of people over time. The people who were NIU in 2019-2023 should have given “no” responses to the two questions had they been asked, since they were not covered by employment-based insurance last year. I limited the analysis to individuals with at least one household member with FIRMSIZE=9.
I see that there is a decline in the number of respondents with “yes” responses to GRPDEPLY in 2019 as compared to 2018. There is small increase in the number of respondents with “yes” responses to GRPOWNLY in 2019 as compared to 2018. These differences are expected, given the many changes between the 2018 and 2019 ASEC that affect measurement of health insurance variables, and given that there are true changes from year to year in people’s insurance coverage.
These tables show the frequencies and shares of “yes” and “no” responses (recall that “no” responses include NIU individuals in 2019-onward) for GRPOWNLY (on top) and GRPDEPLY (on bottom) from the 2017-2020 ASEC.
The Census Bureau provides a Bridge File for 2018, which is the ASEC data from 2018 with the new methods used in 2019 applied to the 2018 data. Using this file as a check can give you an idea of how much of the change may be attributable to the change in methodology in the 2019 ASEC.
Thank you so much for this, @Isabel_Pastoor. I’m combing through it now and will follow up with any questions.
