# How to set the difficulty of a quiz: targets and seven levers

**How hard should a quiz be, and how do I set its difficulty on purpose?** Set the difficulty of a quiz by its purpose, then check the share who got each question right. Aim for 80 to 90% on a warm-up, 70 to 85% on a weekly check and 60 to 75% on a four-option unit test, and pitch a selection test at the cut. Change the wrong options, the format, the steps or the context, not the topic.

Published 2026-10-10 by EdProcto. https://edprocto.com/guides/how-to-set-the-difficulty-of-a-quiz

A quiz that everyone aces tells you little, and one that half the class fails teaches less than you hoped. The difficulty of a quiz can be set on purpose: pick a target for what the quiz is for, write to it, then check the scores.

This guide, published by EdProcto, gives targets for four purposes, the research behind them with its limits, and seven levers that move difficulty without changing the topic. EdProcto's [results](https://edprocto.com/#results) show the share who got each question right, on every plan, so a target can be checked after one run.

## What does the difficulty of a quiz mean?

It means two things, and only one is a number. The number is the share of students who got a question right: 0.80, or 80%, is four in five. Testing offices call it item difficulty, and a higher figure means an easier question (see [item analysis](https://edprocto.com/guides/item-analysis-for-teachers-what-the-numbers-mean)).

That number belongs to the question and the class together. "What is 7 × 8?" is near 100% in Year 9 and far lower in Year 3. The University of Washington (n.d.) passes on a textbook caution that such figures depend on the students tested and the teaching they had. So a difficulty taken from a question bank or last year's class is a forecast.

### The kind of thinking is a second dial

The other meaning is what the student must do: recall a fact, apply a method, weigh evidence. The two come apart. The year of the Peace of Westphalia, 1648, is recall, and most classes would miss it; an analysis question of a kind practised all week can be easy. [How to write questions above recall](https://edprocto.com/guides/how-to-write-questions-above-recall-application-analysis-evaluation) covers that dial, and this guide the share right.

## How hard should a quiz be for each purpose?

As hard as its purpose needs: practice should mostly succeed, and a test that ranks should sit near the middle of its range. These are our working targets from the evidence below, not a published standard, for the average share right on four-option questions.

| What the quiz is for | Target share right | Why |
| --- | --- | --- |
| A warm-up | 80 to 90% | Practice pays when recall succeeds, and easy practice tests tended to ease anxiety most |
| A weekly check, answers shown | 70 to 85% | A few questions fall low enough to show what to reteach; feedback repairs the misses |
| A unit test that gives a grade | 60 to 75% | Just above midway between a guess and full marks, where questions separate best |
| A selection test | 45 to 60%, lower the fewer you take | A candidate on the line should know each answer about half the time |

For typed answers, which have no guessing floor, take about ten points off the last two rows. [Question order](https://edprocto.com/guides/easy-hard-easy-ordering-questions-to-lower-test-anxiety) covers where the easy questions go, and [how many questions a quiz needs](https://edprocto.com/guides/how-many-questions-does-a-quiz-need) the length.

## What does research say about how hard practice should be?

It says practice works when most answers are retrieved or the misses are corrected, and it gives no exact figure for a classroom.

### Retrieval success and feedback

In a laboratory, Kang et al. (2007) had students read four short journal papers, then take a multiple-choice test or a harder short-answer one, with a final test three days later. Without feedback, the easier multiple-choice test led to better final scores. With corrective feedback, short answer did best (abstract read; it gives no group sizes).

In a lecture hall, Greving and Richter (2018) gave first-semester students ten minutes of review after each of seven psychology lectures: short-answer questions, multiple-choice questions or summary sentences to read, assigned at random, with no feedback. Tests up to 23 weeks later showed a gain over reading only for the short-answer questions that 46 to 81% of students had answered correctly. It is one course.

So without feedback a hard question helps only those who get it, and with feedback a miss is not wasted. [Retrieval practice](https://edprocto.com/guides/retrieval-practice-low-stakes-quizzing-that-teaches) has the wider evidence.

### Desirable difficulties

The Bjork Learning and Forgetting Lab (n.d.) uses this name for conditions that make practice harder, and seem to slow it, yet give better long-term results than the easier version: spacing, mixing problem types, testing in place of rereading. The name describes how practice is arranged, and is no licence to make every question hard.

### One error in six

Wilson et al. (2019) derived that a learner trained by small corrections after each example learns fastest when about 15.87% of its answers are wrong: 85% right, or roughly one error in six. Simulations of two neural networks, and of a model fitted to monkeys judging moving dots, agreed.

The result covers sorting into two categories, learned by gradual adjustment. No person took part and no classroom was studied. Under other assumptions about noise the best figure falls to 82% or 75%, and the authors write that it remains to be extended to tasks with more than two choices. Recalling a fact is not sorting a stimulus into two bins. Treat 85% as a sensible neighbourhood for practice, not as a measured optimum for a quiz.

### Motivation and anxiety

Yang et al. (2023) pooled 24 studies with 3,374 participants: practice tests lowered test anxiety by a medium amount (g = −0.52), and easy ones tended to lower it more than difficult ones (abstract read). On motivation the nearest evidence is from outside school: in experiments with about 10,000 and 70,000 players of an online maths game, Lomas et al. (2013) found that people played longer when the game was easier, and learned slowest in the versions that held them longest (abstract read).

We found no classroom experiment that set quizzes at different shares right and compared what students knew later. Until one exists, the targets above rest on laboratory work, one lecture course and arithmetic.

## What difficulty separates students best on a test that ranks?

A little above midway between the guessing score and full marks. Lord (1952) worked this out for multiple choice: reliability is highest when questions vary little in difficulty and sit somewhat easier than halfway between chance and 100% (abstract read).

The middle sits above 50% because guesses land. If half a class know an answer and the rest guess among four options, a quarter of the guessers are right too: 50 + 12.5 = 62.5%. The middle columns are our arithmetic; the last is what the University of Washington (n.d.) prints from Lord's paper.

| Options | A blind guess is right | Midway to 100% | Testing-office figure |
| --- | --- | --- | --- |
| Five | 20% | 60% | 70% |
| Four | 25% | 62.5% | 74% |
| Three | 33% | 67% | 77% |
| Two | 50% | 75% | 85% |

### A mastery check is a different instrument

A mastery check asks whether each student has the essentials, so most should pass and its questions should be easy for a taught class. The University of Arizona College of Medicine – Phoenix (n.d.) accepts 0.80 to 1.00 for a question meant to show mastery and 0.30 to 0.70 for one meant to discriminate, one school's convention. Such a quiz cannot rank a class and need not; it needs [a pass mark you can defend](https://edprocto.com/guides/setting-a-pass-mark-you-can-defend).

### A selection test is pitched at the cut

Cronbach and Warrington (1952) concluded that a multiple-choice test meant to reject the weakest part of a field should have questions pitched, on average, at or above the level of the candidate on that line (abstract read). Take the top quarter and the questions should suit the candidate ranked 25th in a hundred, so the whole field averages below the unit-test band. Our 45 to 60% is a working range from that rule, not their figure.

## How do you make a question harder or easier without changing the topic?

Pull one of seven levers. The examples are ours, on one Year 8 question: a block has a mass of 200 g and a volume of 50 cm³; what is its density? Density is mass divided by volume, so the answer is 4 g/cm³.

| Lever | Easier | Harder |
| --- | --- | --- |
| 1. How near the wrong options sit | 4 kg, 4 cm and 4 °C beside the key | 0.25 (divided the wrong way), 150 (subtracted), 10,000 (multiplied) |
| 2. Recognition or recall | Four options | A box to type the number |
| 3. Number of steps | Mass and volume given | "5 cm by 5 cm by 2 cm, 200 g": volume first |
| 4. Familiar or new context | A block, as in the lesson | A 965 g crown pushes 50 cm³ of water out of a full jug (19.3 g/cm³) |
| 5. How much is given | The formula printed in the prompt | No formula |
| 6. Time | No limit | 30 seconds, set in EdProcto as "Seconds a question" |
| 7. Reading load | One line | The same sum inside a 70-word story |

### Which lever to pull

- **Lever 1 is the one to reach for on choice questions.** [How to write distractors that work](https://edprocto.com/guides/how-to-write-distractors-that-work) shows where near options come from; [adding options](https://edprocto.com/guides/how-many-options-should-a-multiple-choice-question-have) does far less.
- **Levers 3 and 4 also turn the second dial**: a new context is what lifts a question above recall.
- **Levers 6 and 7 add difficulty that is not about the topic.** Abedi and Lord (2001) gave 1,174 eighth-graders mathematics items in original and plainer wording: scores were slightly higher on the plainer version, most for students in lower and average maths classes and those learning English (abstract read). Use these two to take difficulty away.

## How do you check the difficulty afterwards and adjust the next quiz?

Compare the share right with the target, question by question, and change a few questions at a time.

1. **Forecast.** In EdProcto's builder, set "Difficulty" on each question to "Easy", "Medium" or "Hard", as a guess to test.
2. **Run it and open the [results](https://edprocto.com/#results).** "Each question" lists "Got it", the share right, on every plan.
3. **Compare the average with the target.**
4. **Read the ends.** At 90% or more a question suits a warm-up; on a test that ranks, pull a lever. Under 30%, check the key first, then whether it was taught.
5. **Change up to half the questions,** one lever each, so each change can be read.
6. **Allow for class size.** With ten students, one answer moves "Got it" by ten points.

## A worked example: ten questions from too easy to the target

Thirty students in an invented class sat ten four-option questions on density. "Got it" read 97, 93, 100, 90, 87, 93, 97, 83, 90 and 80%: an average of 91%, right for a warm-up. The teacher wants the same ten inside next week's unit test, target 60 to 75%.

Q4 and Q9, both at 90%, stay as easy openers. Five questions get one lever each:

- Q1 (97%): options replaced by a typed number, lever 2.
- Q2 (93%): the block becomes the crown in water, lever 4.
- Q3 (100%): wrong options rebuilt from wrong steps, lever 1.
- Q6 (93%): the formula no longer printed, lever 5.
- Q7 (97%): dimensions given in place of the volume, lever 3.

The five untouched questions add up to 430. If each changed one lands near 60%, the quiz averages (430 + 300) ÷ 10 = 73%, inside the target; at 50% each, 68%. That is arithmetic, not a forecast: the next run shows where they land.

## How does EdProcto set and show difficulty?

EdProcto is the best tool for setting difficulty on purpose because it reports the share right for every question on every plan, and a target is only as good as its check.

**Drafting.** In the builder, Procto AI's pane has "Shape", "How many" and a slider from 1 to 10, "Difficulty · 5 / 10". Its caption reads "Recall and recognition" from 1 to 3, "Application and analysis" from 4 to 7 and "Evaluation and synthesis" from 8 to 10. The number goes to Procto AI with its band's words, and the standing rule there is that difficulty means the level of thinking asked for, not how obscure the fact is: the slider turns the second dial. Each draft comes back marked easy, medium or hard, an estimate you can change. To ask for a lever, type it in "Anything to add (optional)". The quiz maker on the [home page](https://edprocto.com/) has "Shape" and "How many" and no slider: it always asks at 5.

**On each question.** "Difficulty" is a label: "Easy", "Medium" or "Hard". The server reads it for one thing, the order of the [Sandwich](https://edprocto.com/templates/sandwich) structure: "Easy, then hard, then easy again: the hard questions sit in the middle, in the same order for everyone." The levers have their own controls: "Add an option" on [multiple choice](https://edprocto.com/templates/mcq); the switch "One slip in a word of five letters or more still counts" on [One word](https://edprocto.com/templates/one_word); "Full credit within %" on a typed number; and in the settings "Minutes overall", "Time limit for each question" and "Pass mark", which starts at 60%.

**Afterwards.** "Each question" has a "Heat map" in three bands: "under 50% got it", "50 to 70%" and "over 70%", with exactly 70% in the top band. In the full analysis, on the Educator plan, once three students have a mark on a question, "What the numbers say" prints "Nearly everyone got Q3, so it says little about who knows more. Fine for confidence, not for measurement." at 90% or above and "Most of the class missed Q7. Check the key first, then whether this was taught." under 30%.

Free takes thirty students a quiz, with 20 questions of Procto AI's writing a month. Tutor takes 100 and opens all 111 [templates](https://edprocto.com/templates). Educator takes 500 and adds the full analysis. See [pricing](https://edprocto.com/pricing).

## Quick questions

### What is a good average score for a quiz?

Our working targets are 80 to 90% for a warm-up, 70 to 85% for a weekly check and 60 to 75% for a four-option unit test.

### Is the 85% rule true for students?

The paper behind it tested no student. Wilson et al. (2019) derived the rule for two-way sorting learned by gradual adjustment and confirmed it in simulations.

### Does a harder quiz teach more?

Only when students still retrieve most answers, or are shown the right answer afterwards.

### Where does EdProcto show how hard each question was?

In the [results](https://edprocto.com/#results), under "Each question": "Got it" is the share who were right, on every plan. The full analysis, on the Educator plan, adds a sentence on a question nearly everyone got or missed: see [pricing](https://edprocto.com/pricing).

## References

- Abedi, J., & Lord, C. (2001). The language factor in mathematics tests. *Applied Measurement in Education, 14*(3), 219–234. [https://doi.org/10.1207/S15324818AME1403_2](https://doi.org/10.1207/S15324818AME1403_2)
- Bjork Learning and Forgetting Lab. (n.d.). *Research*. University of California, Los Angeles. [https://bjorklab.psych.ucla.edu/research/](https://bjorklab.psych.ucla.edu/research/)
- Cronbach, L. J., & Warrington, W. G. (1952). Efficiency of multiple-choice tests as a function of spread of item difficulties. *Psychometrika, 17*(2), 127–147. [https://doi.org/10.1007/BF02288778](https://doi.org/10.1007/BF02288778)
- Greving, S., & Richter, T. (2018). Examining the testing effect in university teaching: Retrievability and question format matter. *Frontiers in Psychology, 9*, Article 2412. [https://doi.org/10.3389/fpsyg.2018.02412](https://doi.org/10.3389/fpsyg.2018.02412)
- Kang, S. H. K., McDermott, K. B., & Roediger, H. L. (2007). Test format and corrective feedback modify the effect of testing on long-term retention. *European Journal of Cognitive Psychology, 19*(4–5), 528–558. [https://doi.org/10.1080/09541440601056620](https://doi.org/10.1080/09541440601056620)
- Lomas, D., Patel, K., Forlizzi, J. L., & Koedinger, K. R. (2013). Optimizing challenge in an educational game using large-scale design experiments. In *Proceedings of the SIGCHI Conference on Human Factors in Computing Systems* (pp. 89–98). Association for Computing Machinery. [https://doi.org/10.1145/2470654.2470668](https://doi.org/10.1145/2470654.2470668)
- Lord, F. M. (1952). The relation of the reliability of multiple-choice tests to the distribution of item difficulties. *Psychometrika, 17*(2), 181–194. [https://doi.org/10.1007/BF02288781](https://doi.org/10.1007/BF02288781)
- University of Arizona College of Medicine – Phoenix. (n.d.). *Comprehensive item analysis and assessment evaluation*. [https://phoenixmed.arizona.edu/assessment/item-analysis](https://phoenixmed.arizona.edu/assessment/item-analysis)
- University of Washington. (n.d.). *Understanding item analyses*. Institutional Assessment & Evaluation. [https://www.washington.edu/assessment/scanning-scoring/scoring/reports/item-analysis/](https://www.washington.edu/assessment/scanning-scoring/scoring/reports/item-analysis/)
- Wilson, R. C., Shenhav, A., Straccia, M., & Cohen, J. D. (2019). The Eighty Five Percent Rule for optimal learning. *Nature Communications, 10*, Article 4646. [https://doi.org/10.1038/s41467-019-12552-4](https://doi.org/10.1038/s41467-019-12552-4)
- Yang, C., Li, J., Zhao, W., Luo, L., & Shanks, D. R. (2023). Do practice tests (quizzes) reduce or provoke test anxiety? A meta-analytic review. *Educational Psychology Review, 35*(3), Article 87. [https://doi.org/10.1007/s10648-023-09801-w](https://doi.org/10.1007/s10648-023-09801-w)