A comparison of recent estimates of monthly job creation by the Bureau of Labor Statistics (BLS) and estimates generated by Automatic Data Processing (ADP) reveals wide disparities. For example, in February, BLS showed 536,000 jobs created versus only 179,643 for ADP. The following chart details some of the extreme differences between the estimates.
The chart inevitably raises questions about where such dramatic differences come from and exactly how many jobs are, in fact, being created. The answer to the first question has three aspects – differences in data, differences in methodology, and estimation error.
ADP provides a detailed presentation of its data sources and methodology (https://adpemploymentreport.com/common/docs/ADP-NER-Methodology-Full-Detail.pdf). ADP uses its payroll database, which it claims covers about 20% of US private payroll employment, as its main source. It then works through a process involving many steps to massage the data, delete outliers, match firms into more than 460,000 pairs that report employment in two consecutive months, and perform seasonal adjustments. ADP notes one difference between its methodology and that used by BLS, which is that ADP counts active employees, whereas BLS counts only employees who are actually paid during the month. The ADP data are then aggregated into several asset-size classes and assigned to the 13 industry segments. Sophisticated statistical methods are then employed to clean the data, to make seasonal adjustments, and then to make forecasts, which are then aggregated. Of course, each step in the process can be the source of estimation errors. Further, and importantly, ADP notes that its data doesn’t match the size distribution and industry composition of the general population, so it attempts to make compensating adjustments that can be an additional source of estimation error.
BLS produces the monthly Current Employment Statistics job estimates. The sample is larger than that employed by ADP, covering about 1/3 of non-farm employees and a separate sample of government employment. It represents all 50 states and Washington, DC (https://www.bls.gov/web/empsit/cestn.htm#section1). The industry data are drawn from the universe of firms covered in the tax system. BLS assigns each firm to one of 13 industry categories and sorts firms into 8 size classes based upon the number of employees. Detail is provided on the percentage of firms represented in each size class as well as the proportion of total employment represented. The sample is constant throughout the year, unlike that of ADP. As is the case with the ADP approach, pairs of firms are employed, and the data are screened to remove records deemed to be out-of-scope. BLS also provides estimates of the measurement error both in terms of the number of jobs and relative standard errors. Interestingly, one standard error for total nonfarm employment is 68,265 jobs, and the monthly report also provides confidence intervals around the estimates. These intervals are almost always ignored in the reporting. But it is important to recognize their importance when we assess the monthly jobs report. For example, the reported 559,000 jobs for the month of May, which is an estimate made with 90% confidence, actually means that the number of jobs created in May, could in fact, range anywhere from a low of 446,000 to a high of 671,300. Note that even this highest possible number falls far below the number estimated by ADP. Put another way, a jobs number that falls between plus or minus 112,000 could actually mean that zero jobs were created. We have not been able to find published confidence intervals for the ADP forecasts.
In addition to standard forecast errors, there is another reason to be cautious about putting too much emphasis on a particular jobs number. We need to remember that shocks to the system, such as the financial crisis and/or the COVID-19 pandemic, have had economic consequences that are not likely to be factored into the statistical procedures, since the estimation methods rely upon past data that may not include such events. A shock can add to errors, and we are uncertain as to how large or important those errors may turn out to be. So, the bottom line is that investors must be cautious about giving too much weight to a single data release in the case of the employment data, and especially to releases put out by ADP.
Robert Eisenbeis, Ph.D.
Vice Chairman & Chief Monetary Economist
Email | Bio
Links to other websites or electronic media controlled or offered by Third-Parties (non-affiliates of Cumberland Advisors) are provided only as a reference and courtesy to our users. Cumberland Advisors has no control over such websites, does not recommend or endorse any opinions, ideas, products, information, or content of such sites, and makes no warranties as to the accuracy, completeness, reliability or suitability of their content. Cumberland Advisors hereby disclaims liability for any information, materials, products or services posted or offered at any of the Third-Party websites. The Third-Party may have a privacy and/or security policy different from that of Cumberland Advisors. Therefore, please refer to the specific privacy and security policies of the Third-Party when accessing their websites.
Sign up for our FREE Cumberland Market Commentaries
Cumberland Advisors Market Commentaries offer insights and analysis on upcoming, important economic issues that potentially impact global financial markets. Our team shares their thinking on global economic developments, market news and other factors that often influence investment opportunities and strategies.
