The ADP Mimics the Flawed Methodology of the BLS

Diving Into ADP Methodology

An investigation of ADP Methodology shows it seeks to mimic the BLS jobs report.

Much of the data processing used in preparing the NER mimics the methodology used by the BLS to calculate payroll employment for the CES report. One difference is that the ADP NER includes any active employee within the company, whereas the BLS tracks employees only if they are paid during the applicable month.

ADP then goes through all kinds of contortions to account for 5-week pay cycles and seasonal adjustments, and even weather, something the BLS does not do.

ADP has two dummy variables, one for odd payroll cycles and the other explicitly for August.

It acknowledges that “VAR models create a tradeoff between oversimplification and over-parameterization.”

Over-Parameterization

Instead of three explanatory variables represented by three lags of industry growth rates, our model now uses one explanatory variable represented by a moving average of the first two lags. Moreover, using a moving average of BLS industry growth rates helps reduce the inherent volatility and thus noise in the data.

August Effect

BLS consistently reports lower levels of employment changes in the month of August, likely in part due to low response rates by larger companies as many of their employees take vacations during the month. This effect is captured in our model by including a dummy variable for the month of August.

The BLS underestimates August so ADP put in a dummy variable to duplicate the BLS error.

ADP Huge Sample Set

The ADP has a huge sample set. It processes nearly 20% of the nation’s private payrolls, but not in the same percentages as jobs in the real economy.

That is the one variable it needs to account for. Instead of all the other nonsense like dummy variables actually designed to introduce errors, I suggest the ADP should strive to minimize errors.

An excellent start would be to weed out duplicate social security numbers and all the associated over-counting by the BLS.

Alas, the BLS does not do that, so ADP mimics that error too.

What’s Worse?

• Silly methodology?
• Attempts to mimic silly methodology?

Since ADP frequently and wildly misses the BLS estimates why bother?

Let me suggest the ADP try to produce a real number instead of attempting to game, and often badly, what the BLS will report.

Mish