Standardizing Census occupation categories with microdata, 1970-2024

Thomas M. Menino Convention & Exhibition Center 

Respondents to the U.S. Decennial Census and related surveys are classified into hundreds of detailed occupation categories. The classification systems change periodically, creating breaks in time series. Researchers may want a unified occupation category set to extend for a long time period, to study the effects of technological change, for example. Standard crosswalks and unified category systems from IPUMS and other researchers accomplish this, but they reclassify the data coarsely, based only on the original occupation category. In theory they could make more use of particular "dual-coded" data sets in which specialists have applied multiple category systems based on more of the respondent attributes, such as their industry of work, age, income, and geographic location. Modern machine learning tools can do this on a large scale to impute occupations to millions of observations. Meyer and Asher (2021) applied random forest methods to accomplish this on an experimental scale. We extend the methods of Meyer and Asher (2021) to cover employed persons in the Population Census and Current Population Survey from 1970 to 2024, assigning best-matching 1990 Census occupations on the basis of several variables. We compare the assignments to earlier systems (e.g. IPUMS's occ1990) to see what changes the new method induced, and test the resulting distributions of occupations for how well they match known trends and how much the resulting occupation categories jump in size or attributes when the classification systems changed.