Book a meeting
All articles
India Talent Market17 November 2026·6 min read

Negative Marking, Explained: A Wrong-Answer Penalty Measures Confidence Along With Skill

A wrong-answer penalty buys real measurement precision, but it also scores how certain a candidate needs to be before answering. What the evidence says, and a five-step check before you use it in a hiring test.

By AssessAll Editorial

Negative marking is a scoring rule that subtracts a fraction of a mark for every wrong answer, so a skipped question costs nothing and a wrong one costs something. It sharpens measurement by suppressing blind guessing. It also adds a second trait to every score: the candidate's willingness to answer when they are not certain.

Most Indian candidates have spent years inside that rule. India's large competitive examinations — civil services, staffing commission, banking, medical and engineering entrance — commonly apply a fractional penalty for wrong answers, and coaching for them drills attempt strategy as a discipline in its own right. Candidates bring that strategy into corporate hiring assessments. Whether your test should reward it is a design decision, and most teams make it by copying whatever examination they happen to remember.

What the penalty is actually for

On a four-option item scored number-right, a candidate who knows nothing earns 25% by filling the sheet at random. That noise sits on top of the real signal, and a wrong-answer penalty is the oldest fix: price the guess so random answering returns zero.

The effect is larger than most practitioners assume. Akyol, Key and Krishna analysed 9,273 first-time takers of the Turkish university entrance examination, where each 45-question section used five options, +1 for correct and −0.25 for wrong. Raising the penalty from 0.25 to 1 point improved precision about as much as lengthening each section from 45 to 70 questions. Run the trade the other way and the same penalty would let 27 questions carry the validity of the original 45 (NBER Working Paper 22401).

Espinosa and Gardeazabal reached a compatible conclusion by modelling each item as a lottery: the optimal penalty is relatively high, and stays high even when candidates do not behave fully rationally. They acknowledged that it disadvantages risk-averse candidates, and judged that bias small next to the measurement error it removes (*Journal of Mathematical Psychology*, 2010).

The case for negative marking is not folklore. It is a real precision gain, and a cheaper one than adding items.

What the penalty also measures

The counter-case is not that the penalty fails to remove guessing. It is that it puts something else in.

The Standards for Educational and Psychological Testing call this construct-irrelevant variance: "variance in test-taker scores that is attributable to extraneous factors that distort the meaning of the scores," which reduces the validity of any interpretation you place on them (NCME glossary; the Standards).

Baldiga tested that with 406 subjects answering 20 history items under either a −0.25 penalty or none. Under the penalty, women were more than 50% more likely than men to skip a given question. Controlling for separately measured risk tolerance and stated confidence narrowed the gap to roughly one-third more likely but did not remove it, and there was no gender difference in knowledge, measured by a forced-response section of the same test. Each skipped item cost close to a quarter of a point, and women with the same knowledge scored about 0.44 points lower, roughly a twelfth of a standard deviation. Under no-penalty scoring the pattern vanished: all but one subject answered every question (*Management Science*, 2014).

That is a 20-item laboratory test, and the author says so. The mechanism, though, is not in doubt. A penalty converts a disposition — how certain you need to be before committing — into score variance, and disposition is not what you are hiring for.

Why this lands hardest at the top of your funnel

Hiring selects from the upper tail, which is precisely where the effect concentrates.

Espinosa and Gardeazabal's field experiment found that under a correction-for-guessing rule, women omitted more items, got fewer correct and received lower grades — and that the average gap was concentrated in the upper tail. The gap also depended strongly on how the rule was framed to candidates. They titled the paper for personnel selection and college admission rather than education generally (*B.E. Journal of Economic Analysis & Policy*, 2020).

Akyol and colleagues found the same direction from the other side: the certainty threshold a candidate needs before answering rises with ability. Stronger candidates were more cautious, not less.

Two counterweights:

  • The aggregate effect in that exam was small. In the Turkish counterfactual, forcing everyone to attempt every question raised the female share of the top 5% only from 39.9% to 40.2%. Risk aversion explained about 10% of the observed gender gap at the top.
  • But measurement error does not disappear because the funnel looks unchanged. The same analysis reports that a standard Rasch model ignoring skipping can estimate equally able men as roughly 5% higher in ability. A model blind to omissions mis-ranks individuals even when group totals come out close.

For a decision made one candidate at a time, the second point is the operative one.

What it costs in reliability

The direct empirical evidence on reliability runs against the penalty. Cecilio-Fernandes and colleagues gave 298 medical students from four Dutch schools 200 previously used progress-test items in a two-by-two crossover, scoring the same responses both ways. Person reliability came out at 0.80 and 0.82 under number-right scoring against 0.73 and 0.77 under formula scoring. Item reliability was effectively identical either way, at 0.94 to 0.96, and formula scoring produced about three times as many dysfunctional items. Their conclusion: number-right tests had the better psychometric properties (*BMC Medical Education*, 2017).

Both bodies of evidence can hold at once. A penalty removes guessing noise and adds risk-attitude noise. Which dominates depends on how many options your items carry, how good your bank is, and who sits the test. That is an empirical question about your assessment, not a settled one.

How to make the decision on your own test

Conditions that favour a penalty

  • Four or five options per item, where random guessing is otherwise worth 20–25% of the score.
  • A short test you cannot lengthen, since the penalty buys precision that extra items would otherwise have to buy.
  • A pass/fail certification decision near the middle of the distribution rather than top-of-tail selection.

Conditions that argue against one

  • Two- or three-option items, where the penalty is hard to calibrate and the guessing baseline is high anyway.
  • Top-of-tail selection, where the omission effect concentrates.
  • Any process carrying adverse-impact obligations you would have to defend, including the US Uniform Guidelines for multinational funnels.

A five-step check before you switch it on

  1. Score your existing response data both ways. You already hold the raw responses. Compute reliability under number-right and under your penalty rule on the same data, as the Dutch crossover did, and let the numbers decide.
  2. Compute omission rates by subgroup. Count unanswered items by gender, by language of schooling, by region. If omission rates differ and knowledge-matched scores differ with them, the penalty is doing the damage.
  3. State the rule plainly in the instructions. Framing moved the gap in both the 2014 lab study and the 2020 field experiment. A candidate who misreads the rule is being scored on inference, not skill.
  4. Never pool penalty and non-penalty sections into one composite without rescaling. They are on different metrics.
  5. Price the alternative. Scoring number-right and instructing candidates to attempt everything removes the omission trait outright. More items, adaptive delivery or constructed-response scoring remove the guessing problem at its source rather than correcting for it afterwards.

One structural point is worth naming: the guessing-correction debate exists only because of the multiple-choice format. Scenario items graded on a written response, of the kind AssessAll scores with AI, have no 25% floor to correct for, so the trade-off never arises. Adaptive delivery narrows it from the other side, reaching a given precision in fewer items.

A caveat on the evidence: these studies are Turkish, Dutch and American, and no published Indian estimate could be verified for this piece. Measure omission rates on your own candidates rather than assume the figures transfer — especially in a market whose candidates have been trained on penalty scoring far more intensively than these samples were.

The takeaway

Negative marking is a defensible choice that buys real precision, and an indefensible default when nobody has checked what it costs. Score your own response data both ways, look at who is leaving items blank, and decide — rather than inherit the rule from an examination you are not running.

#negative marking#psychometrics#test design#scoring#hiring assessment

Measure it, don't guess it.

Start free with 100 credits — or write to solutions@bodhih.com.

Start free