An attrition rate needs a base

Doubling a half-year rate assumes the second half repeats the first. For Apex, 22.3 percent on the average base becomes 44.6 percent annualised. The word "annualised" is doing real work in that sentence: it says the figure is a projection, built from six months of data, under an assumption that the next six months look the same. Drop the word, and 44.6 becomes a fact about a year that has not happened.

The assumption fails in both directions. If the Sales exodus was a one-off, the second half will be quieter and 44.6 overstates. If it was the start of something, the second half will be worse and 44.6 understates. The half-year figure does not know which, and neither does the doubling. What the number can honestly say is "22.3 percent in H1, which would be 44.6 for the year if H2 repeats it", and that is the sentence to put on the slide, unshortened.

The same trap runs the other way. A quarter's rate multiplied by four, a month's rate multiplied by twelve: every one of those multiplications is a forecast, and the shorter the period it was built from, the noisier the forecast. It also breaks a comparison: an annualised half-year against last year's measured full year is one projection beside one measurement, and the gap between them is partly the assumption.

"Is 22.3 percent bad" has no answer until three things are fixed: bad compared with whom, over what period, and counting which leavers. The comparison most people reach for is a national figure, and the best open one is the US Bureau of Labor Statistics quits rate from its JOLTS survey, which counts voluntary quits during a month as a percent of employment. It is monthly. Apex's figure is a half-year. The scale conversion is the self-study of the build-along below, and getting it wrong by a factor of six or twelve is the single most common way a benchmark misleads.

The second comparison is with last year, and there the question is whether last year's figure used the same base, the same period, and the same definition of leaver. Usually nobody knows, because it was never written down. The rate card in the build-along is how to make sure yours is.

The group rate hides the split, and the split is the finding. Apex's eight departments on the start base: Sales 50.0 percent (54 of 108), Customer Service 37.6 (35 of 93), Operations 25.2 (55 of 218), Marketing 19.7 (15 of 76), Engineering 11.3 (24 of 213), HR 10.7 (8 of 75), Finance 8.3 (12 of 145), R&D 7.4 (7 of 94). The four entities tell a flatter story: Digital 25.0, Logistics 23.5, Manufacturing 19.3, Retail 17.7, all on the start base, all constructed.

Two rules for any split. First, the base goes under every rate, as "54 of 108", not as a footnote. A rate with its base can be re-added: 54 plus 35 plus 55 and so on reaches 210, and the bases reach 1,022, so the split reconciles to the bridge from Foundations session 1. A rate without its base cannot be checked against anything. Second, the same base rule everywhere. Sales on the start base beside R&D on the average base is a comparison of two conventions, not of two departments.

Transfers matter here too: a person who moved from Sales to Marketing in March is a Sales leaver and a Marketing joiner on the department split, and neither on the group bridge. Apex has 291 rows with a transfer flag, so the rule has to be written before the split is made.

HR's rate is 10.7 percent, from 8 leavers of 75. One more resignation and it is 12.0; one fewer and it is 9.3. On a base of 75, each person is 1.3 points, which is arithmetic (100 over the base) and not a property of HR. Marketing, on 76, has the same problem. Operations, on 218, moves 0.5 of a point per person and can carry a trend; HR cannot, and a dashboard that colours HR red at 12.0 and green at 9.3 is colouring one person's decision.

The rule of thumb: read the base before the rate, and if the base is under about a hundred, read the count instead. "8 people left HR" is a sentence a manager can act on. "HR attrition is up 2.7 points" is the same fact dressed up as a trend. Small groups also make the survivor problem from Foundations session 1 worse: with 75 people, the handful who left before the snapshot would have moved the number visibly, and they are not in the table.

The four entity rates (25.0, 23.5, 19.3, 17.7) and the eight department rates both sum to 210 leavers over 1,022 starters, because they are two cuts of the same bridge. A department that spans entities, like Operations, will show one rate on the department view and be spread across four on the entity view; both are correct, and a reader who compares Operations' 25.2 with Digital's 25.0 is comparing a column with a row. The habit is to say which cut a rate comes from, in the same breath as its base. Deep-dive session 3 rebuilds both cuts from the table and shows the reconciliation as a check that runs every month.

Apex records eleven exit reasons among its 210 leavers: Better Opportunity 33, Career Growth 33, Compensation 29, Management/Culture 23, Relocation 19, Work-Life Balance 19, Unknown 17, Personal/Health 15, Involuntary/Performance 10, Retirement 9, Further Study 3. All constructed. "Regretted" is a rule that sorts those eleven into two piles. The rule used here counts Better Opportunity, Career Growth, Compensation, Management/Culture and Work-Life Balance: 137 people, 65.2 percent of the 210.

Every line in that rule is arguable. Is a relocation regretted? The person would have stayed if the company had a site in the new city, so maybe. Are the 17 Unknowns regretted? Nobody asked, or nobody answered, and putting them either side moves the figure by eight points. The honest way to report a regretted rate is with the categories beside it, so that the reader can see which pile each one landed in and argue. A regretted figure reported alone, with the rule in somebody's head, is a number that cannot be disagreed with because it cannot be seen.

The same discipline applies to "voluntary", "controllable", "avoidable" and every other adjective a dashboard puts in front of "attrition". Each is a line through the categories. Show the line.

The sentence that usually follows a regretted rate is "and replacing each of them costs 1.5 to 2 times salary". That range circulates widely and has no measured source: it is attributed to an analyst's estimate rather than to a study that counted the costs. The professional body most often cited for it, SHRM, publishes its own range as 50 to 200 percent of salary, which is wide enough to contain almost any conclusion and is itself a range of estimates. A cost of attrition figure is not a fact about a company until that company has counted its own recruiting, onboarding and vacancy costs, and most have not.

What can be said without a cost figure is often stronger: 137 people whose reasons the company could in principle have addressed, 54 of them from one department of 108, most of them under two years' tenure. Those are counts, they reconcile, and a board can act on them without multiplying anything by an unmeasured constant.