Which selection method really works? 85 years of research summarized

Why this research matters

In 1998, Frank Schmidt and John Hunter published a meta-analysis in Psychological Bulletin that summarized 85 years of research into personnel selection. The article, "The Validity and Utility of Selection Methods in Personnel Psychology", has since become one of the most cited papers in HR psychology.

The central question was simple. Which selection methods actually predict how well someone will perform in a job? And which methods that many companies use do not actually work?

The answer was based on thousands of studies and millions of employees. Not on opinions, not on gut feeling, but on data. And the conclusions are surprising to many HR professionals.

What does "validity" mean in plain language?

Validity sounds like a statistical concept, but the idea is simple. It is the answer to the question: if I use this method to select someone, how well does it predict how that person will perform?

A validity of .00 means the method predicts nothing: you might as well have tossed a coin. A validity of 1.00 would mean the method predicts perfectly. That does not exist in practice.

A validity of .51 (structured interview) means this method uses 51% of the theoretically maximum predictive power. That may sound modest, but in the context of human behavior it is exceptionally high.

• By comparison: the relationship between smoking and lung cancer has a correlation of about .40. Nobody doubts that smoking causes lung cancer. A selection method with a validity of .51 has a stronger statistical relationship than that.

The practical consequence is this. If you use a method with validity .51 instead of hiring at random, you hire significantly better employees on average. Over hundreds of hires, that translates into millions of euros in productivity difference.

The reality: how most organizations select

Let's be honest. Most organizations select like this. A recruiter screens CVs on work experience and education, invites a shortlist for an interview, and decides based on the click. That feels logical. But what does science say about this approach?

• CV screening on work experience (validity .18): after 5 years in the same job, extra experience predicts almost nothing about future performance. Yet most vacancies ask for "at least 7-10 years of experience". Schmidt and Hunter found that the relationship between experience and performance levels off completely after 5 years.
• CV screening on education (validity .10): within a group of applicants for the same job, the difference in years of education predicts almost nothing. Someone with a master's degree does not by definition perform better than someone with a bachelor's degree.
• Unstructured job interview (validity .38): no fixed questions, no fixed scoring. The interviewer forms an impression based on a free conversation. Two interviewers who speak with the same candidate regularly reach opposite conclusions.

Do you combine these three methods? Then you use only a fraction of the available predictive power. You leave most of your selection accuracy on the table. And that costs money. Every misfit you hire costs your organization tens of thousands of euros in lost productivity, onboarding and possibly turnover.

The ranking: what really predicts job performance?

Schmidt and Hunter ranked 19 selection methods by predictive validity. Below are the most important ones, grouped by effectiveness:

--- High predictive value ---

• Work sample tests (validity .54): the candidate performs part of the actual work. The highest validity of all individual methods, but only usable if the candidate already knows the work.
• Structured interviews (validity .51): every candidate gets the same questions, scored on the same criteria. Just as predictive as cognitive tests, and a world of difference from the unstructured interview. Read more about the STAR method and BARS scoring.
• Cognitive tests / GMA (validity .51): measure general mental ability. The best-researched method, with thousands of studies. Ability tests are often administered as part of an e-assessment or a live assessment center.
• Integrity tests (validity .41): measure reliability and conscientiousness. They do not correlate with intelligence, which is why they add a lot in combination with cognitive tests.

--- Moderate predictive value ---

• Unstructured interviews (validity .38): the "ordinary job interview" without a fixed structure. Significantly less predictive than structured interviews. Yet this is the most widely used selection method in the world.
• Traditional assessment centers (validity .37): role plays, group assignments and in-tray exercises. Lower than structured interviews or cognitive tests. A modern assessment center that combines these methods scores considerably higher.
• References (validity .26): limited value, partly due to legal restrictions on what previous employers may share.

--- Low predictive value ---

• Work experience in years (validity .18): the most commonly used filter in vacancy texts. But after 5 years in the same job, it predicts virtually nothing.
• Education in years (validity .10): determines what level of job someone can get, but within a job barely predicts who performs better.
• Interests (validity .10): people choose jobs that match their interests, but once in the job, interests do not determine how well they perform.
• Age (validity -.01): no relationship at all with job performance.

The conclusion is uncomfortable but clear. The two methods most organizations rely on, CV screening (.10-.18) and the unstructured interview (.38), are in the bottom half of the ranking.

The power of combining

The most important finding of Schmidt and Hunter is not which method works best, but which combinations work best. A single method is never enough.

The three strongest combinations for predicting job performance:

• Cognitive test + integrity test: combined validity of .65. The highest of all combinations studied. This is because integrity tests (which mainly measure conscientiousness) do not correlate with intelligence. They measure something fundamentally different, which is why they add a lot. Cognitive tests are often administered through an e-assessment.
• Cognitive test + structured interview: combined validity of .63. The interview partly measures different things than the cognitive test: work experience, personality, communication. Together they reach 63% of the maximum predictive power. Read more about the structured interview.
• Cognitive test + work sample: combined validity of .63. Very strong, but only usable for experienced candidates.
• Compare this with the standard approach: CV screening on experience (.18) and education (.10), followed by an unstructured interview (.38). That combination uses less than half of the predictive power that is available. The difference between .38 and .65 sounds small, but translates into tens of thousands of euros per hire in productivity difference.

The switch does not have to be radical. Simply by structuring your existing interviews (from .38 to .51) and adding an assessment, you jump from the bottom half to the top 3. Read more about how to set up an effective recruitment and selection process.

The economic impact of better selection

Schmidt and Hunter also calculated what better selection yields in euros. The core finding is that the difference in productivity between an average and an above-average employee is at least 40% of the annual salary.

With a salary of EUR 60,000, this means that an employee at the 84th percentile produces at least EUR 24,000 more per year than an average employee. The difference between a below-average and an above-average employee is double that: EUR 48,000 per year.

For management jobs and professionals, the difference is even larger: 48% of the average salary. With a salary of EUR 80,000, that is a difference of EUR 38,400 per year.

If you hire 10 people per year and switch from unstructured interviews (.38) to a combination of structured interview + assessment (.63), you improve your selection by 66%. Over those 10 hires, that can mean hundreds of thousands of euros in productivity difference. Per year.

This is not theory. It is mathematics, based on 85 years of data.

What does this mean for your organization?

Most organizations select based on CV screening (education + experience), an unstructured interview, and gut feeling. According to Schmidt and Hunter, that is a combination with low predictive value.

The step to better selection does not have to be big:

• Step 1: Structure your interviews. Use fixed questions per competency, with follow-up suggestions and assessment criteria. This alone raises validity from .38 to .51. That is an improvement of 34%. Read how to set this up in our article on structured interviewing.
• Step 2: Add assessments. A cognitive test or personality assessment measures things the interview does not measure. The combination is stronger than either one separately. An integrated assessment center that combines cognitive tests, personality questionnaires and a structured interview reaches the highest predictive value. Read more about assessments in selection.
• Step 3: Base everything on the job profile. The interview questions, the assessment criteria and the scoring criteria must all be derived from the same competencies. That makes them comparable and combinable. Read how to set this up from the job architecture.

The result is a selection decision based on multiple data sources, all linked to the same competencies. That is what Schmidt and Hunter call "the optimal combination".

Common mistakes in selection

Based on the research by Schmidt and Hunter, these are the most common mistakes:

Too much weight on work experience. After 5 years in the same job, extra experience predicts almost nothing. Yet most vacancies ask for "at least 7-10 years of experience".

Unstructured interviews as the only method. A free conversation feels informative, but its predictive value is limited. Two interviewers who interview the same candidate in an unstructured way often reach different conclusions.

Not using assessments. Many organizations see assessments as "too heavy" or "too expensive". But the data is clear. Assessments are among the strongest predictors of job performance.

Selecting on education. Within a group of applicants for the same job, the difference in years of education predicts almost nothing. Someone with a master's degree does not by definition perform better than someone with a bachelor's degree in the same job.

Trusting gut feeling. The human brain is poor at predicting future behavior based on a 45-minute conversation. Structured methods compensate for the biases we all have.