Every hiring team believes its process is rigorous. Almost none of them can say what their process predicts, because almost none of them have measured it. Personnel psychology has, repeatedly, for about a century, and the answer is uncomfortable for anyone whose process is built on interviews.
Top rankedof the common selection procedures
The largest recent reassessment of selection research found that the structured interview is the single best predictor of job performance among common hiring methods, ahead of cognitive ability tests, work samples and job knowledge tests.
The word doing the work in that sentence is "structured". The same body of research is equally clear that an interview without structure is one of the weaker things you can do with an hour.
| Corrected estimate, 2022 | Older estimate, 1998 | |
|---|---|---|
| Structured interview | .42 | .51 |
| Work sample test | .33 | .54 |
| Cognitive ability test | .31 | .51 |
| Job knowledge test | .26 | .48 |
| Unstructured interview | .19 | .38 |
Validity coefficients: 0 means the method tells you nothing about later performance, 1 means it tells you everything. Nothing in hiring comes close to 1. The 1998 column is the set of figures most articles still quote; the 2022 paper showed the corrections behind them had been systematically overapplied, and revised every one downward.
Two things are worth sitting with. The first is that a structured interview predicts performance more than twice as well as an unstructured one, using the same hour of the same people's time. The second is that the honest numbers are lower than the ones you have read elsewhere. Hiring methods are weaker predictors than the industry likes to claim, including the good ones.
What structure actually means
Structure is not formality, and it is not a rubric written after the fact. It is a small set of decisions made before you meet anyone, and held constant across everyone you meet.
- Every candidate answers the same questions, in the same order.
- The questions are about the work, not about the person: what they would do, or what they have done, in a situation the job actually contains.
- Each answer is scored on a defined scale, against criteria written before the first candidate applied.
- The scoring happens per answer, not as an overall impression at the end.
The review of the structured interview literature finds that these components are separable: you can adopt some and not others, and the gain scales with how many you adopt. Standardising the questions matters; standardising the evaluation matters at least as much.
That last point is where most teams stop early. They agree on a question list, then let each interviewer form a holistic impression and write it up. The questions were the cheap half.
Why the assessment goes before the interview
If structured evaluation predicts better than unstructured conversation, the scheduling follows. An interview is the most expensive thing in the process: it consumes a senior person, synchronously, per candidate. Spending it on people you know nothing about is spending your most limited resource at the point of maximum uncertainty.
A structured assessment sent before the call inverts that. Everyone who applies does the same work, scored the same way, at a cost that does not rise with volume. The interview then does the thing an interview is uniquely good at: probing an answer you have already read.
The order is the intervention
Nothing above requires new tooling. A team that moves its existing structured questions from the call to a form, and scores them before scheduling, has already made the change that matters.
Building an assessment that measures the job
The failure mode of a technical assessment is testing something adjacent to the work: puzzle-solving under time pressure, recall of algorithms nobody implements, or willingness to spend a weekend unpaid. Each of those measures something real. None of them is the job.
- 1Take a task from the last month of the actual role. Strip it to something self-contained.
- 2Give the candidate the context a colleague would have: the codebase conventions, the constraint that matters, the reason the task exists.
- 3Let them use whatever they would use at work, documentation and search included. You are hiring someone who will have those.
- 4Cap the time, and mean it. An assessment that expands to fill a weekend selects for people with free weekends.
- 5Write the scoring criteria before you send it. If you cannot say what a good answer contains, the task is not ready.
Work sample tests, the family this kind of task belongs to, hold up well across the research, though more modestly than the older literature claimed. The updated meta-analysis of work sample validity is the basis for the figure in the table above.
Knowing whether it worked
An assessment you never evaluate is a ritual. Three numbers tell you most of what you need, and all three are yours to measure rather than anyone else's to supply.
- Completion rate: how many people who start it finish. A falling rate usually means the task got longer, not that candidates got worse.
- Pass rate: how many clear your threshold. Very high means the task is not discriminating; very low usually means it is testing something other than the job.
- Interview-to-offer rate among people who passed. This is the one that tells you whether the assessment is predicting anything. If it rises after you introduce the assessment, the assessment is working.
Deliberately, none of those has a target here. Benchmarks published for them come from vendors describing their own customers, and a number from someone else's funnel tells you nothing about yours. Your own baseline, measured before you change anything, is the only comparison that means something.
The bottom line
The research does not say hiring is solved. It says the gap between a structured process and an unstructured one is larger than the gap between most of the things teams argue about, and that the structured version is available to anyone willing to write the questions down first.