Numbers That Held Up

The study defined the labor market as the state of Wisconsin, which sounds neutral and generous. The employer is in Milwaukee County. Nobody commutes to a Milwaukee plant from Vilas County. I pulled the employer's own address data for its last four years of hires. Ninety one percent lived within 35 miles. I then recomputed the study's medians restricted to the counties those hires actually came from, using the same source data the consultant used, changing nothing else. The maintenance technician median moved from 26.40 to 30.15 an hour. Every other benchmark moved the same direction by between four and eleven percent. The strength of this was that I used their numbers and their method. There was nothing to attack except the market definition, and the map was not arguable. Settlement landed near the corrected figure. Total time on it: about nine hours, most of them spent on the hire address data rather than the statistics.

The board came to me because scheduled hours matched forecast demand nicely on an annual basis and yet the store was in constant crisis from mid January to early March and losing good people every spring. The model averaged. It took annual labor hours and distributed them across weeks by a demand index built from sales. What it did not carry was absence. In February, in northern Wisconsin, unplanned absence ran at 14 percent against a 5 percent annual average, from illness, weather and childcare. The model assumed full attendance and the schedule was already at the bone. We rebuilt it in a spreadsheet the general manager could maintain, with a monthly absence factor from their own three years of timesheets and a minimum viable crew constraint by department. Winter scheduled hours went up about 9 percent, summer came down 4, annual cost rose 1.6 percent. Turnover the following year fell from 38 percent to 24. I take no credit for that beyond removing one obvious cause.

A survey of 214 members on a scheduling change. First and second shift appeared to differ substantially in their willingness to accept a rotating pattern. I reported the difference, and I reported it before I had properly checked how the responses had been collected. They had not been collected the same way. First shift filled it in at a meeting, second shift got a link by email a week later, after the change had been publicly discussed and argued about at length. The two groups were not answering the same question in the same conditions, and the difference I found was largely an artifact of when and how they were asked. I caught it four days after I sent the memo. I wrote a correction, drove out and said it to the committee in person, and did not bill for the engagement. The lesson I actually took: ask how the data was collected before you touch it, not after. I now refuse to analyze a survey until somebody tells me the fielding history, and that rule has come from this one afternoon.

One. What is the sample frame, and which employers in your market are not in it? Missing employers are almost never random. Two. What was the response rate, and what was done about nonresponse? Silence here usually means nothing was done. Three. How were benchmark jobs matched? Ask for the survey's own job descriptions, not the titles. Four. Mean or median, and why? Then ask for the other one. If they will not give it, you have learned something. Five. What is the effective date and what aging factor was applied? A study current as of eighteen months ago, aged at a guessed rate, is a forecast wearing a survey's clothes. Six. Where are the confidence intervals? A benchmark based on seven responding employers and one based on ninety should not be presented in the same typeface with the same authority. That is the whole list. It has not changed much since about 1996.