Hi, I am using data for the five months between Feb. and June 2026 and for a few earlier years with the same months. For each occupation, for a given year, I divide the total across the 5 months to get an estimate of workers in each occupation for, say, Feb.-June 2026. The problem is that in some occupations, one or more months have 0 observations. I have a question on how what number to divide across the 5 months for a given year. Should I divide by 5 even though some months do not have cases or by the number of months for which a given occupation has cases? I hope this makes sense. Thank you for any guidance.
Rogelio
When estimating frequencies using pooled CPS sample data, you should divide the sampling weight by the number of samples you have pooled. So, when pooling BMS samples from February through June of a single year, you would divide the sampling weight by 5, the number of samples included. In the scenario you described, your estimate for the months that have non-zero frequencies of this occupation would be identical to the estimate for all five months.
Note that your estimates are unlikely to be accurate if the sample sizes are this small. There is no bright-line rule for “too small to study,” but generally single and double digit sample sizes are considered too small to produce reliable estimates. Estimating parameters for single occupations, especially with additional restrictions placed on the analytical sample (e.g., geography or demographics), can often be difficult using CPS data. You may want to consider pooling additional samples together, studying a broader occupation category or multiple occupations together, reducing the restrictions on your analytical sample, or exploring other data sources with larger sample sizes, such as the American Community Survey (data available from IPUMS USA).
Hi Isabel,
Thank you very much for your response and suggestions. This is very useful. Thank you.
Rogelio