The hiring funnel

Screen-in rate: 120 of 400, 30.0 percent. The base is everyone who applied. This is the ratio the rest of tonight cares most about, because the screen is the stage with the most people and the least time per person, and it is where the two studies in Part 3 found the trouble.

Interview rate: 40 of 120, 33.3 percent. The base is the screened, not the applicants. Quote it as "a third" and someone will hear "a third of applicants get an interview", which would be 133 people, not 40.

Offer rate: 12 of 40, 30.0 percent. The base is the interviewed. It reads the same as the screen-in rate and describes a stage with a tenth of the people in it. Acceptance rate: 10 of 12, 83.3 percent. The base is offers made, and it is the one ratio a company reads as a verdict on itself: two people said no.

All four are constructed for Apex, and the arithmetic is the point, not the values. The rule is the same as for attrition: the base goes beside the ratio, every time, and the base of a funnel ratio is the stage before it unless somebody says otherwise. The overall yield, 10 of 400, is the one ratio whose base is the top of the funnel, and it is the number a hiring manager should hear first: 400 applications bought 10 people.

Every stage can be read two ways. Against the stage before, it is a conversion: how selective was this step. Against the top of the funnel, it is a yield: how much of the original 400 is still here. At the interview stage, Apex reads 33.3 percent one way and 10.0 percent the other (40 of 400). Neither is wrong, and a report that mixes them, a conversion on one line and a yield on the next, will be read as a funnel that leaks 23 points between two rows that describe the same forty people. The build-along draws both columns side by side so that the difference is visible rather than argued.

Time-to-fill measures the vacancy: from the day the requisition was opened, or approved, to the day an offer was accepted. It is the number a hiring manager feels, because it is how long the seat was empty, and it includes the days before anyone applied. It is driven by approval speed, posting speed, sourcing and every stage of the funnel.

Time-to-hire measures the candidate: from the day the person who was eventually hired entered the funnel to the day they accepted. It is the number the candidate feels, and it is the better measure of how fast the process itself runs once there is somebody in it. It excludes the days the requisition sat waiting.

Both stop at acceptance, not at the first day, because the notice period belongs to the candidate's old employer. Both definitions are standard practice with no single source, and both drift: some teams start fill at posting rather than approval, some start hire at first contact rather than application. The only defence is to write the start and end events beside the number, exactly as the base goes beside a rate. A time-to-hire quoted without its start event is a number without a base.

An average time-to-fill is computed over vacancies that were filled. The ones still open at the reporting date are not in it, and they are, by construction, the slow ones. This is the survivor problem from Foundations session 1 wearing a different hat: the average describes the vacancies that closed fast enough to be counted. The median is less distorted than the mean, and the honest report says how many requisitions were still open on the date and how old the oldest was. Deep-dive session 2 measures the same effect on tenure, where it is called censoring, and the fix is the same: count the ones that have not finished yet, and say so.

Take Apex's screening stage and split it by sex with an invented split: 240 men applied and 80 were screened in, 160 women applied and 40 were screened in. Two selection rates, 33.3 and 25.0 percent, each with its base. The impact ratio is the lower rate over the higher: 25.0 over 33.3, which is 0.75. It is one number, it is computed at one stage, and it can be computed at every stage and for every group the company records.

The US Uniform Guidelines on Employee Selection Procedures (29 CFR 1607.4(D), 1978) give the rule of thumb that a selection rate under four-fifths of the highest group's rate is generally regarded as evidence of adverse impact. That is the one sentence of law on the page. What counts as a group, what to do when the numbers are small, how a screening tool shifts the ratio, and what the New York City audit law asks employers to publish are all taught with a working simulator in the AI + HR course, and deep-dive session 4 does the arithmetic in pandas.

The habit for tonight is smaller: a funnel report carries the impact ratio at its widest stage, with both rates and both bases printed, the way an attrition rate carries its base. A ratio of two rates reported as a single decimal is two undeclared bases at once.

Bertrand and Mullainathan 2004, "Are Emily and Greg More Employable Than Lakisha and Jamal?", American Economic Review 94(4) 991-1013, sent otherwise comparable resumes to real job advertisements with names chosen to sound white or African-American and counted the callbacks. Resumes with white-sounding names received about 50 percent more callbacks. The stage they measured is the screen: the moment a person reads a resume for a few seconds and decides who goes forward. It is the widest stage, the fastest decision, and the one where the impact ratio in the figure above is computed. That is why the metric belongs at the top of the funnel and not at the offer.

Hoffman, Kahn and Li 2018, "Discretion in Hiring", Quarterly Journal of Economics 133(2) 765-800, studied firms that introduced a job test to rank applicants for low-skill service jobs, where managers could still hire against the test's recommendation. Test-based hiring raised completed tenure by about 25 percent (0.23 log points), and the hires managers made by overriding the test left sooner than the ones the test would have chosen. The study does not say managers know nothing; it says that in that setting, the exceptions they made were on average worse than the rule they overrode. For a funnel it means the screening stage is measurable, and that "the manager's judgement" is a stage with a selection rate like any other, which can be compared with the alternative it replaced. Deep-dive session 9 returns to the same paper when the screening tool is an algorithm.