Recency Bias in Performance Reviews (and Four More)

Recency bias is the tendency to weight recent events more heavily than earlier ones. In a performance review it means the last six weeks quietly stand in for the whole year.

It is the most-discussed review bias, and it is genuinely worth fixing. It is also not the most expensive one, which is the part most articles leave out.

How recency bias actually shows up

Not as an obvious error. It shows up as a rating that feels right and is built from a non-representative sample:

• Someone who struggled in spring and delivered in autumn is remembered as strong
• Someone who carried the first half and had a quiet Q4 is remembered as coasting
• A single incident in November outweighs eleven quiet months of reliability

The manager is not being careless. They are answering an impossible question — summarise a year — with the only data available, which is what they can currently recall.

The tell: if review outcomes correlate with what happened in the final quarter more than with the year, you have it. That is measurable, and almost nobody measures it.

The four that usually cost more

1. Availability bias

Visible work is remembered; invisible work is not.

The engineer whose systems never broke produced no memorable events. The person who prevented three crises produced no crises. Meanwhile whoever fought a visible fire — possibly one they helped cause — is vivid.

Why it costs more than recency: recency distorts when you look. Availability distorts what counts as work at all, and it systematically penalises prevention, maintenance and enablement — exactly the work that is hardest to replace.

2. Idiosyncratic rater effect

The largest single source of variance in most rating data is not the person being rated. It is the person doing the rating. Some managers are generous, some severe, and the spread between them routinely exceeds the spread between employees.

Practical consequence: a rating tells you about a manager-employee pair, not an employee. Comparing ratings across managers without accounting for that compares raters.

3. Halo and horns

One strong attribute colours everything. A confident communicator is rated higher on technical judgement; someone who missed a deadline in March is rated lower on collaboration in a way that has nothing to do with collaboration.

How it shows up in data: a person's ratings across unrelated dimensions correlate far more than they should. If your competency scores move together almost perfectly, you are measuring one impression five times.

4. Similarity bias

Managers rate people like themselves more highly — similar background, communication style, working pattern. It is the bias with the most serious equity consequences, and the hardest to see from inside, because it feels like recognising quality.

What actually counters them

Ordered by how much they help relative to effort.

Contemporaneous records. The single highest-value change. Notes written within a week of the event, not reconstructed at year end. This directly defeats recency, substantially helps availability, and gives every other countermeasure something to work with. See evidence-based performance reviews.

Structured behavioural evidence. Require a specific, dated example for each rating. Not "strong collaborator" but what they did, when, and what followed. Vague ratings are where halo and similarity live.

Multiple raters. Peer input reduces the idiosyncratic rater effect more than any amount of manager training, because it dilutes a single rater's calibration rather than trying to correct it.

Rate dimensions separately, in separate sittings. Scoring five competencies in one pass guarantees halo. Even a short break between them helps.

Calibration sessions with evidence present. Calibration without evidence is managers arguing from impressions, which mostly surfaces who argues best. With records on the table it becomes a real comparison.

Measure your own distributions. Ratings by manager, by demographic group, by tenure, by quarter of last notable event. You are looking for patterns that reflect the rater rather than the rated. This is the step almost nobody does, and it is the only one that tells you whether any of the others worked.

What doesn't work

Unconscious bias training alone. Awareness does not reliably change ratings. It is worth doing alongside structural changes and is not a substitute for them.

Telling managers to "consider the whole year." The instruction does not create data they no longer have.

Forced distributions. They replace one distortion with a stronger one — now the rating is constrained by quota rather than performance, and managers optimise around it.

More competencies. Twelve dimensions rated from one impression is one impression recorded twelve times, with more halo, not less.

The honest summary

You cannot remove bias from human judgement, and a review process that claims to has usually just hidden where it operates.

What you can do is narrow the gaps where it operates unchallenged — by making the record contemporaneous, the evidence specific, the raters plural, and the resulting distributions something you actually inspect.

Recency is the easiest to fix and the one everyone names. Availability and the idiosyncratic rater effect are larger, quieter, and worth more of your attention.

Related

• Evidence-based performance reviews — the method
• Why annual performance reviews don't work
• Continuous performance management