How many questions should a quiz have? As many as its score has to carry. A quiz that only tells you what to reteach tomorrow can be three to five questions. A score that grades students, or passes some and fails others, needs enough questions that luck stops deciding it: about 20 to 40 for a grade, 40 or more for a pass.
Luck comes in two kinds: the guess that lands, and the accident of which questions you happened to ask. Both shrink as a quiz grows, at a rate this guide works out and sets against the minutes a class will bear.
EdProcto publishes this guide, and its results report how consistently a quiz measured and how long each question took, so the length can be checked on your own class.
How many questions should a quiz have for each purpose?
Set three to five questions when you will read every answer, 20 to 40 when the total becomes a grade, and 40 or more when it passes or fails someone. The table is our recommendation from the arithmetic below, not a published standard, and it assumes multiple choice with three or four options.
| What the score is for | Questions | Why |
|---|---|---|
| Practice, or finding one misconception | 3 to 5 | You act on each answer. EdProcto's quiz maker on the home page writes a sample of 3 or 5 |
| A weekly check, added to others | 8 to 10 each | No single one decides anything |
| A grade on a unit | 20 to 40 | A standard error of 10 to 7 points |
| A pass or fail decision | 40 to 80 | 7 to 5 points, with blind guessing far below the pass mark |
| Certification or selection | 90 or more | The strictest reliability, from an exam blueprint |
True or false needs more at every level, as the next section shows, and no length makes a pass mark right: that takes its own method.
Why are five questions too few to decide anything?
Five questions are too few because chance alone moves a five-question score by about 20 points either way. Burton (2001) names two sources of that chance: a test holds only a small proportion of the questions that could be asked, so a student may be lucky or unlucky in the selection, and random guessing lowers reliability further unless it is discouraged. He concludes that many multiple-choice tests, as usually run, are too unreliable for their purpose. We could read only his abstract, so the figures below are ours.
Luck from guessing
A student who guesses every answer averages 50% on true or false, 33% with three options and 25% with four, but the average hides the spread. This table gives the highest score that at least one blind guesser in twenty reaches, by exact binomial arithmetic.
| Questions | True or false | Three options | Four options |
|---|---|---|---|
| 5 | 80% | 60% | 60% |
| 10 | 80% | 60% | 50% |
| 20 | 70% | 50% | 40% |
| 40 | 62.5% | 45% | 37.5% |
On five questions a lucky guesser looks like a student who knows the work. At forty, with three or four options, the luckiest sit well under any pass mark, and true or false is still not safe. The true or false guide gives the chance of passing by guessing and the marking that removes it.
Luck in which questions were asked
The second kind reaches students who never guess. Take a student who truly knows 70% of everything you could have asked. Each question you pick is one this student gets right with a chance of 0.7, so a short score wanders.
| Questions | Standard error, in points | At least nine scores in ten fall within | Chance of scoring under 60% |
|---|---|---|---|
| 5 | 20 | 40–100% | 16% |
| 10 | 14 | 50–90% | 15% |
| 20 | 10 | 55–85% | 11% |
| 40 | 7 | 57.5–80% | 6% |
| 80 | 5 | 61–78% | 2% |
In plain words: on a ten-question quiz this student's score will often land anywhere between about 50% and 90%, and about one time in seven it falls under 60%.
The standard error is the binomial one, the square root of 0.7 × 0.3 ÷ the number of questions, so halving it takes four times the questions.
How does reliability grow as a quiz gets longer?
Reliability climbs steeply over the first twenty or thirty questions and slowly after that. Reliability is the share of the differences between students' scores that is real, from 0 to 1; Cronbach's alpha, explained in the item analysis guide, is its usual estimate. The Spearman-Brown formula gives the growth: make a quiz with reliability r into one n times as long, with questions of the same quality, and the new reliability is n × r ÷ (1 + (n − 1) × r). Ellis and Sijtsma (2024) trace the formula to papers by Spearman and by Brown in 1910 and prove that it still describes the growth when questions are drawn at random from one pool, not only when they are perfectly matched. For a short quiz that mixes unrelated topics, their simulations show it can promise too much.
The table starts from a ten-question quiz with a reliability of 0.50, an assumption for illustration and not a finding.
| Questions | Reliability | Added by the last doubling |
|---|---|---|
| 10 | 0.50 | start |
| 20 | 0.67 | 0.17 |
| 40 | 0.80 | 0.13 |
| 80 | 0.89 | 0.09 |
| 160 | 0.94 | 0.05 |
Each doubling doubles the ratio of signal to noise. At 0.50 that ratio is 1 to 1, and 2 to 1 is 0.67. At 0.80 it is 4 to 1, and 8 to 1 is only 0.89: most of the noise had already gone. So doubling a short quiz buys far more than doubling a long one.
How high is high enough? Jogerst et al. (2021), citing Downing (2004) and an older textbook, give 0.9 for high-stakes examinations such as national licensing, 0.8 for moderate stakes and 0.7 for lower stakes such as a medical school's own quizzes. Tavakol and Dennick (2011) found acceptable values reported from 0.70 to 0.95. On these figures the three bands take 24, 40 and 90 questions. Downing (2004) says what a low figure means: scores can be expected to vary widely on retesting.
How many questions fit in the time you have?
Divide the minutes by the time one question takes, which runs from about half a minute for a true or false statement to two and a half for a question on a case. The allowances are our working figures, set above the measured averages because an average hides the slowest students.
| Question shape | Measured | Our allowance | Fit in 10 minutes |
|---|---|---|---|
| True or false | Three attempted for every two multiple choice (Frisbie, 1973) | 30 seconds | 20 |
| Multiple choice, recall | 36 to 41 seconds on average (Schneid et al., 2014) | 45 seconds | 13 |
| Multiple choice on a scenario | 88 minutes on average for 50, untimed (Jensen et al., 2013) | 110 seconds | 5 |
| Extended matching on a case | 100 within four hours (Beullens et al., 2002) | 150 seconds | 4 |
| Short answer, a sentence or two | Nothing we read | 120 seconds | 5 |
Options cost time: Schneid et al. (2014) calculated that three options in place of four or five made room for 16% more questions in an hour. A quiz that does not fit its time starts to measure speed.
How do you work out the length of your own quiz?
Start from the objectives and the minutes, not from a round number.
- Name the decision. Take the range from the first table.
- Count the objectives. Give each at least three questions if you will report on it separately, so eight objectives make 24. That floor is our working rule; Jensen et al. (2013) wrote their 50-question examinations with at least two for every learning outcome.
- Check the minutes. If 24 do not fit, test five objectives properly and hold three for next week. Do not thin every objective to one question.
- Run it and read the result. EdProcto's results give Cronbach's alpha and the pace, which say whether to add questions or minutes next time.
The floor comes from the standard error: for our 70% student a topic score built on two questions has an error of 32 points, and on three, 26. Where areas differ in weight, the exam blueprint guide turns weights into counts.
When do more questions make a quiz worse?
More questions make a quiz worse when the extra ones are weaker than the rest, or when they do not fit the time. Tiredness is the usual objection, and the studies we read do not support it. Ackerman and Kanfer (2009) gave 239 first-year university students test sessions of three and a half to five and a half hours: reported fatigue rose with the hours, yet mean performance was higher in the longer sessions. Jensen et al. (2013) gave one section of a biology course 100-question examinations in place of 50, and those students did no worse on the second fifty than on the first. Beullens et al. (2002) saw no measurable effect of fatigue on scores across 100 questions in four hours. All three studied university students. We found nothing comparable for young children.
Padding is the real cost. The Spearman-Brown gain assumes the added questions are as good as the first. A question that stronger students miss more often than weaker ones pulls reliability down: Beullens et al. (2002) reached an alpha of 0.80 by deleting the questions that correlated negatively with the total score. Ellis and Sijtsma (2024) warn that a high reliability says nothing about the quality of the questions and may only show that there are many, and Tavakol and Dennick (2011) note that an alpha above 0.90 may point to redundant ones. The guide to writing multiple-choice questions says what a good one looks like.
Is one long quiz better than several short ones?
Several short quizzes added together do the work of one long one, usually better. The arithmetic does not care whether forty questions were answered at once or as four tens: the total has the 7-point error of the forty-question row, not the 14 of a single ten. Twelve weekly quizzes of eight questions make 96, and an error under 5 points.
A bad day costs a twelfth of the total, and the quizzing itself helps students remember, as the guides on retrieval practice and weekly quizzes in a lecture course set out. Two cautions: students learn between quizzes, so the total describes the term and not one day, and no single short quiz should carry a decision that only the total can bear.
How does EdProcto help you set and check the length?
EdProcto is the best tool for settling the length of a quiz, because after the first run it tells you whether that length was enough: it reports the reliability this guide turns on.
The quiz maker on the home page takes your material and a shape, such as multiple choice or true or false, and Procto AI writes a sample of 3 or 5 questions from it. In your account the builder writes more, and the plan caps what Procto AI writes in a month: 20 questions on Free, 50 on Tutor, 100 on Learner, the self-study plan, and no monthly cap on Educator, within a daily ceiling.
Time is a setting: a limit in minutes for the whole sitting, a limit in seconds for each question, or neither. The Practice preset sets no limit; Standard quiz and Exam set an overall one, and Exam also hands the quiz in when time runs out.
After the quiz, the results show each question's difficulty and discrimination, Cronbach's alpha for the quiz, and pace: the median number of seconds the class spent on each question and the median time to finish. With at least eight submissions, an alpha under 0.5 brings a finding that the quiz measured inconsistently, through too few questions or several faulty ones. A question that took more than three times as long as a typical one is flagged too.
Results question by question come with Free, as do ten quizzes a month, enough for a weekly check. The full analysis, which adds discrimination, distractors and reliability, is on Educator. The pricing page lists every plan, and the templates show all 111 shapes.
Quick questions
How many questions should a 10-minute quiz have?
About thirteen recall multiple-choice questions, twenty true or false statements or five scenario questions, on our working allowances of 45, 30 and 110 seconds each.
Is a 10-question quiz enough for a grade?
Not by itself. A student who truly knows 70% of the topic scores anywhere from 50% to 90% on ten questions, nine times in ten. Add several short quizzes together, or set 20 to 40 questions.
How many questions does a test need to be reliable?
It depends on the questions. As an illustration, a quiz at 0.50 with ten questions reaches 0.70 at 24, 0.80 at 40 and 0.90 at 90 by the Spearman-Brown formula. EdProcto's results report Cronbach's alpha for the quiz you ran, so you can check.
Do true or false quizzes need more questions than multiple choice?
Yes. On twenty questions one blind guesser in twenty still reaches 70% with two options, against 40% with four. Frisbie (1973) found that students attempt three true or false statements for every two multiple-choice questions, so the extra statements cost little time.
References
- Ackerman, P. L., & Kanfer, R. (2009). Test length and cognitive fatigue: An empirical examination of effects on performance and test-taker reactions. Journal of Experimental Psychology: Applied, 15(2), 163–181. https://doi.org/10.1037/a0015719
- Beullens, J., Van Damme, B., Jaspaert, H., & Janssen, P. J. (2002). Are extended-matching multiple-choice items appropriate for a final test in medical education? Medical Teacher, 24(4), 390–395. https://doi.org/10.1080/0142159021000000843
- Burton, R. F. (2001). Quantifying the effects of chance in multiple choice and true/false tests: Question selection and guessing of answers. Assessment & Evaluation in Higher Education, 26(1), 41–50. https://doi.org/10.1080/02602930020022273
- Downing, S. M. (2004). Reliability: On the reproducibility of assessment data. Medical Education, 38(9), 1006–1012. https://doi.org/10.1111/j.1365-2929.2004.01932.x
- Ellis, J. L., & Sijtsma, K. (2024). Proof of reliability convergence to 1 at rate of Spearman–Brown formula for random test forms and irrespective of item pool dimensionality. Psychometrika, 89(3), 774–795. https://doi.org/10.1007/s11336-024-09956-7
- Frisbie, D. A. (1973). Multiple choice versus true-false: A comparison of reliabilities and concurrent validities. Journal of Educational Measurement, 10(4), 297–304. https://doi.org/10.1111/j.1745-3984.1973.tb00807.x
- Jensen, J. L., Berry, D. A., & Kummer, T. A. (2013). Investigating the effects of exam length on performance and cognitive fatigue. PLoS ONE, 8(8), Article e70270. https://doi.org/10.1371/journal.pone.0070270
- Jogerst, K., Park, Y. S., & Petrusa, E. (2021, March 1). Are your assessment scores and feedback reliable? A statistical review for the surgical educator. RISE: Resources in Surgical Education. American College of Surgeons. https://www.facs.org/for-medical-professionals/news-publications/journals/rise/articles/assessment-scores/
- Schneid, S. D., Armour, C., Park, Y. S., Yudkowsky, R., & Bordage, G. (2014). Reducing the number of options on multiple-choice questions: Response time, psychometrics and standard setting. Medical Education, 48(10), 1020–1027. https://doi.org/10.1111/medu.12525
- Tavakol, M., & Dennick, R. (2011). Making sense of Cronbach's alpha. International Journal of Medical Education, 2, 53–55. https://doi.org/10.5116/ijme.4dfb.8dfd