An official website of the United States government
This document describes the methods used to produce the new metric for occupational separations. The method involves two completely separate calculations, first for calculating workers who leave an occupation and find employment in a different occupation (occupational transfers), and second for calculating workers who leave the labor force entirely (labor force exits). The two methods are described in detail below; some aspects of the methods are the same, but are repeated so that the descriptions are comprehensive for both methods. Both methods use CPS microdata files, and the narratives include the microdata variable names used in performing the calculations; CPS microdata files and data dictionaries are available from the Census Bureau website at https://thedataweb.rm.census.gov/ftp/cps_ftp.html.
To estimate workers who leave an occupation to enter a different occupation, we first estimate historical transfers using data from the CPS Annual Social and Economic Supplement (sometimes known as the March supplement). Question 46 in the supplement asks respondents whether the longest job they held in the prior calendar year is the same as their current job. If it is not, question 47 asks them what that job was, including information about the occupation, industry, and class of worker. This data is coded into supplement variables that define the longest job in the previous year. From these variables, we use the SOC major occupational group (WEMOCG), the NAICS major industry group (WEMIND), and the class of worker (LJCW). Supplement question 41 asks respondents how many hours they usually worked per week in the prior calendar year. In the supplement variable HRCHECK, the interviewer codes the worker as part time (1-34 hours per week) or full time (35+ hours per week) based on the response to question 41.
We also use demographic data on workers from the supplement, specifically age (A_AGE), sex (A_SEX), race (PRDTRACE), ethnicity (PEHSPNON), citizenship (PRCITSHP), educational attainment (A_HGA), and current occupation (PEIOOCC for detailed occupation and A_DTOCC for major group). Finally, we note the year the data were collected.
We do not directly look at historical occupational transfers by occupation, because many occupations would have small and unreliable sample sizes. Instead, we use a regression to determine how different factors affect the likelihood to transfer.
The regression uses a probit model, with the specification as follows:
Prob(transfer) = ƒ(age, sex, occupation, education, occupation*education, race, ethnicity, citizenship, full time status, class of worker, industry, year)
The dependent variable is defined as a respondent employed in a different major occupational group than in the prior calendar year (WEMOCG not equal to A_DTOCC).
For the independent variables:
The universe is all respondents in the ASEC supplement who meet all of the following criteria:
Data from 10 supplement years are used to improve sample sizes and mitigate cyclical issues.
The output of the regression is a series of coefficients for each of the independent variables.
The coefficients of the historical data regression provide information on the probability that a worker with those characteristics will leave their current occupation for another occupation. To project the number of workers who are expected to transfer, we apply these coefficients to the current demographic structure of occupations. Current data on occupations comes from the monthly CPS data. All months from the current base year are used, along with all months from the previous year in order to boost sample sizes. The independent variables are taken directly from each monthly data respondent, for all respondents who are employed (ALFSR = 1 or 2). Note that variable names in the monthly data do not always match the variable names from the supplement; equivalent variables are identified and used.
The parameters for each respondent, plus the regression coefficients, generate a z-score for the probability that that worker will leave their current occupation for another occupation. The z-score is converted into a numeric probability for each respondent. That probability is multiplied by the respondent weight (PWCMPWGT) and then summed by occupation to generate a numeric value for the number of workers in that occupation projected to transfer. This is divided by total employment in the occupation to generate a rate of transfers for each occupation.
The model generates the percent of workers projected to transfer occupations over a nine-month period of time. This nine-month rate is divided by nine to get a monthly rate and then multiplied by 12 to get the annual rate provided in the data. In cases where multiple SOC occupations are aggregated into one CPS occupation, the rate from the CPS occupation is assigned to all of the component SOC occupations. As a result, some detailed SOC occupations have identical rates.
Annual occupational transfer rates are applied to the average of base and projected employment for an occupation to develop a projection for the average annual number of occupational transfers over the projection period. For summary SOC occupations, the number of occupational transfers from component detailed occupations is summed and used to calculate an occupational transfer rate for the summary occupation.
To estimate workers who leave the labor force entirely, we first estimate historical exits using data from the monthly CPS. Monthly CPS data includes respondents who are in the sample for consecutive months, on an in-for-4, out-for-8, in-for-4 month pattern. Individual respondents are matched by records with equivalent household ids (HRHHID and HRHHID2), person line number (PULINENO), sex (PESEX), race (PTDTRACE), and age (PEAGE). Once the universe of matched records is created, respondents are identified as either labor force leavers, if they were in the labor force (PEMLR 1-4) for each of the first four months of their rotation but out of the labor force (PEMLR 5-7) for each of the second four months, or labor force stayers if they were in the labor force for the entirety of both four-month periods. Respondents with all other combinations of labor force status are excluded from the data.
We do not look directly at historical labor force exits by occupation, because many occupations would have small and unreliable sample sizes. Instead, we use a regression to determine how different factors affect the likelihood to exit.
The regression uses a probit model, with the specification as follows:
Prob(exit) = ƒ(age, sex, age*sex, occupation, education, occupation*education, race, ethnicity, citizenship, full time status, class of worker, industry, year)
According to the rules stated above, the dependent variable is defined 1 if they left the labor force or 0 if they remained in the labor force.
For the independent variables:
The universe is all respondents in the monthly CPS data who meet all of the following criteria:
Ten years of year-to-year matched data are used to improve sample sizes and mitigate cyclical factors.
The output of the regression is a series of coefficients for each of the independent variables.
The coefficients of the historical data regression provide information on the probability that a worker with those characteristics will leave the labor force. To project the number of workers who are expected to leave, we apply these coefficients to the current demographic structure of occupations. Current data on occupations comes from the monthly CPS data. All months from the current base year are used, along with all months from the previous year in order to boost sample sizes. The independent variables are taken directly from the each monthly data respondent, for all respondents who are employed (PEMLR = 1 or 2).
The parameters for each respondent, plus the regression coefficients, generate a z-score for the probability that that worker will leave the labor force. The z-score is converted into a numeric probability for each respondent. That probability is multiplied by the respondent weight (PWCMPWGT) and then summed by occupation to generate a numeric value for the number of workers in that occupation projected to leave. This is divided by total employment in the occupation to generate a rate of leaving for each occupation.
The model generates the percent of workers projected to exit the labor force over a nine-month period of time. This nine-month rate is divided by nine to get a monthly rate and then multiplied by 12 to get the annual rate provided in the data. In cases where multiple SOC occupations are aggregated into one CPS occupation, the rate from the CPS occupation is assigned to all of the component SOC occupations. As a result, some detailed SOC occupations have identical rates.
Annual labor force exit rates are applied to the average of base and projected employment for an occupation to develop a projection for the average annual number of labor force exits over the projection period. For summary SOC occupations, the number of labor force exits from component detailed occupations is summed and used to calculate a labor force exit rate for the summary occupation.
Last Modified Date: September 4, 2019