ADP say that they handle payroll for one in six of all companies in America. That is both a large sample size, and probably broadly representative of the economy. There will of course be some business segments that are over or underrepresented but that is different than disregarding the entire report as noise.
ADP reported a meager 29,000 increase in new jobs, but the BLS showed a much larger 139,000 gain. April also showed a similarly wide gap between the two reports.
Through the first five months of 2025, the difference between the two reports has averaged a whopping 63,000 a month.
BLS tracks both numbers, and the private still diverges quite a bit. We saw this again in July. ADP number was -33k private sector jobs, BLS showed growth of 74K for private sector, so massive 100k delta in just the private sector number.
That is quite the delta, but why is it assumed BLS is the more correct one?
From the BLS site on their methodology[0]:
> The Current Employment Statistics (CES) program, also known as the payroll survey or the establishment survey, is a monthly survey of approximately 121,000 businesses and government agencies representing approximately 631,000 worksites throughout the United States.
> All new samples are solicited by computer-assisted telephone interview (CATI), and data are collected for the first 5 months via this mode. After the initiation period, many sample units are transferred to one of several less costly reporting methods that are self-initiated by the respondent.
How is a survey from 121k businesses better than the actual payroll data from a company handling 460k companies[1]?
Isn’t the larger sample generally considered better? Isn’t source data generally considered better than a phone survey of self-reporting?
Looking at the two methodologies, it sounds like BLS data is noise, not ADP.
It doesn't matter how large your sample size is if your sampling method is biased. This could be measuring market share gain/loss in different segments of a steady employment environment.
> disregarding the entire report as noise
Studies with bad / non public sampling methods should be, at a minimum, treated with great skepticism. Why would that not apply here?
A factual number that has less statistical noise or political bias is extremely valuable. Yes, one needs to factor for the biases but that doesn't mean the number should be ignored.
The trend is useful, since one can fairly safely assume that most of the biases haven't radically changed.
ADP data does have a bias, but it is so much data it provides a valuable signal. The importance of understanding the source of your data and how it represents the largest population is something every scientist has been drilled on.
From what I can find they have 1.1m clients and over 900k are small businesses. The vast majority. This makes sense, as there aren’t that many massive companies.
