As I understand it, the introduction of EARNWEEK2 in April 2023 involved two changes:
- new rounding criteria, applied retroactively to months prior to April 2023
- new top-coding, applied in April 2023 onward
I’m observing earnings increases starting in April 2023, but I’m not sure if it’s a true increase or just a product of the top-code changes. The median values (which theoretically should be unaffected by increasing a top-code) are also going up, so it could be a combination of the two.
My r code is below. I am starting with an IPUMS extract that includes the following relevant columns:
# filter to earnings eligibility flag = yes
filter(ELIGORG == 1) %>%
# make null some NIU rows that remain in the data
mutate(EARNWEEK = if_else(EARNWEEK == 9999.99, NA_real_, EARNWEEK),
EARNWEEK2 = if_else(EARNWEEK2 == 9999.99, NA_real_, EARNWEEK2)) %>%
# summary stats
group_by(YEAR, MONTH) %>%
summarise(earnweek_avg = weighted.mean(EARNWEEK, w = EARNWT, na.rm = TRUE),
earnweek_min = min(EARNWEEK, na.rm = TRUE),
earnweek_median = weightedMedian(EARNWEEK, w = EARNWT, na.rm = TRUE),
earnweek_max = max(EARNWEEK, na.rm = TRUE),
earnweek2_avg = weighted.mean(EARNWEEK2, w = EARNWT, na.rm = TRUE),
earnweek2_min = min(EARNWEEK2, na.rm = TRUE),
earnweek2_median = weightedMedian(EARNWEEK2, w = EARNWT, na.rm = TRUE),
earnweek2_max = max(EARNWEEK2, na.rm = TRUE))
Beginning in April 2023, the Census Bureau instituted a “dynamic” topcode for weekly earnings in which the top three percent of cases are topcoded and assigned a value that is a weighted average of the weekly earnings values of all cases that fall within that top three percent. As a result, topcode values provided by the Census Bureau change monthly for the MISH 4 group from April 2023 - March 2023 and for all respondents in the outgoing rotation groups from April 2024 onward. Prior to April 2023, the topcode for weekly earnings represented four to five percent of the values while the dynamic topcoded samples have only 3% of values topcoded. IPUMS does not round original topcoded values when applying new rounding rules to originally un-rounded months or rotation groups. The Census Bureau recommends using the lowest topcode for the months of interest. For further information about topcoding, see the Census Bureau’s user note or IPUMS CPS documentation on topcoding of usual hourly earnings.
Any change in the way income is measured or recorded likely affects mean income estimates. However, top-coding changes should not impact estimates of median income. The median may be affected by rounding, but in your graphs, the median income appears to increase two months following the change. I do not think it is unreasonable to conclude that an actual increase in median earnings is occurring. To determine the extent to which the new top-coding methods are affecting your post-April 2023 estimates of mean income, I would recommend performing your analysis with and without respondents with top-coded income values and comparing your results. I would also encourage you to review how your field handles topcoded values in data; for example, you may be interested in this paper on using the pareto distribution approach for topcoded values in public use CPS data.