I have binary data on exam questions (correct/incorrect). Some individuals might have had prior access to a subset of questions and their correct answers. I don’t know who, how many, or which. If there were no cheating, suppose I would model the probability of a correct response for item $i$ as $logit((p_i = 1 | z)) = \beta_i + z$, where $\beta_i$ represents question difficulty and $z$ is the individual’s latent ability. This is a very simple item response model that can be estimated with functions like ltm’s rasch() in R. In addition to the estimates $\hat{z}_j$ (where $j$ indexes individuals) of the latent variable, I have access to separate estimates $\hat{q}_j$ of the same latent variable which were derived from another dataset in which cheating was not possible.

The goal is to identify individuals who likely cheated and the items they cheated on. What are some approaches you might take? In addition to the raw data, $\hat{\beta}_i$, $\hat{z}_j$, and $\hat{q}_j$ are all available, although the first two will have some bias due to cheating. Ideally, the solution would come in the form of probabilistic clustering/classification, although this is not necessary. Practical ideas are highly welcomed as are formal approaches.

So far, I have compared the correlation of question scores for pairs of individuals with higher vs. lower $\hat{q}_j -\hat{z}_j $ scores (where $\hat{q}_j - \hat{z}_j $ is a rough index of the probability that they cheated). For example, I sorted individuals by $\hat{q}_j - \hat{z}_j $ and then plotted the correlation of successive pairs of individuals’ question scores. I also tried plotting the mean correlation of scores for individuals whose $\hat{q}_j - \hat{z}_j $ values were greater than the $n^{th}$ quantile of $\hat{q}_j - \hat{z}_j $, as a function of $n$. No obvious patterns for either approach.


I ended up combining ideas from @SheldonCooper and the helpful Freakonomics paper that @whuber pointed me toward. Other ideas/comments/criticisms welcome.

Let $X_{ij}$ be person $j$’s binary score on question $i$. Estimate the item response model $$logit(Pr(X_{ij} = 1 | z_j) = \beta_i + z_j,$$ where $\beta_i$ is the item’s easiness parameter and $z_j$ is a latent ability variable. (A more complicated model can be substituted; I’m using a 2PL in my application). As I mentioned in my original post, I have estimates $\hat{q_j } $ of the ability variable from a separate dataset $\{y_{ij}\}$ (different items, same persons) on which cheating was not possible. Specifically, $\hat{q_j} $ are empirical Bayes estimates from the same item response model as above.

The probability of the observed score $x_{ij}$, conditional on item easiness and person ability, can be written $$p_{ij} = Pr(X_{ij} = x_{ij} | \hat{\beta_i }, \hat{q_j }) = P_{ij}(\hat{\beta_i }, \hat{q_j })^{x_{ij}} (1 - P_{ij}(\hat{\beta_i }, \hat{q_j }))^{1-x_{ij}},$$ where $P_{ij}(\hat{\beta_i }, \hat{q_j }) = ilogit(\hat{\beta_i} + \hat{q_j})$ is the predicted probability of a correct response, and $ilogit$ is the inverse logit. Then, conditional on item and person characteristics, the joint probability that person $j$ has the observations $x_j$ is $$p_j = \prod_i p_{ij},$$ and similarly, the joint probability that item $i$ has the observations $x_i$ is $$p_i = \prod_j p_{ij}.$$ Persons with the lowest $p_j$ values are those whose observed scores are conditionally least likely -- they are possibly cheaters. Items with the lowest $p_j$ values are those which are conditionally least likely -- they are the possible leaked/shared items. This approach relies on the assumptions that the models are correct and that person $j$’s scores are uncorrelated conditional on person and item characteristics. A violation of the second assumption isn’t problematic though, as long as the degree of correlation does not vary across persons, and the model for $p_{ij}$ could easily be improved (e.g., by adding additional person or item characteristics).

An additional step I tried is to take r% of the least likely persons (i.e. persons with the lowest r% of sorted p_j values), compute the mean distance between their observed scores x_j (which should be correlated for persons with low r, who are possible cheaters), and plot it for r = 0.001, 0.002, ..., 1.000. The mean distance increases for r = 0.001 to r = 0.025, reaches a maximum, and then declines slowly to a minimum at r = 1. Not exactly what I was hoping for.

  • 5
    $\begingroup$ This is a tough problem because you have very little information about the nature of the cheating. How do you differentiate a cheater from a student who studied extra hard? Without more information, you can't. One possibility is if students can cheat by copying off each other, or if subsets of students had access to the same answers. If this is the case, you could create a distance function between students (lower distance means they did well on the same questions) and look for patterns here. This would be more conclusive IMO. $\endgroup$
    – rm999
    Mar 5 '11 at 0:28
  • 3
    $\begingroup$ Levitt and Dubner describe their approach in Freakonomics (freakonomicsmedia.com ). $\endgroup$
    – whuber
    Mar 5 '11 at 1:56
  • 1
    $\begingroup$ @whuber Thanks, I will look up the paper (assuming it's published). I listened to the audiobook, but can't remember the details of how they identified cheaters (who were teachers who were fudging students' answers, I believe). $\endgroup$
    – lockedoff
    Mar 5 '11 at 7:58
  • 1
    $\begingroup$ If I recall the Freakonomics case, it involved spotting children in the same school/class who had (a) large jumps in attainment compared with a year earlier, (b) different answers for the earlier easier questions, and (c) identical sequences of answers for later harder questions, so being suggestive of a teacher filling in answers which the children had left blank. $\endgroup$
    – Henry
    Mar 7 '11 at 0:24
  • 1
    $\begingroup$ @Henry Here's the paper: pricetheory.uchicago.edu/levitt/Papers/JacobLevitt2003.pdf $\endgroup$
    – lockedoff
    Mar 7 '11 at 16:00

Ad hoc approach

I'd assume that $\beta_i$ is reasonably reliable because it was estimated on many students, most of who did not cheat on question $i$. For each student $j$, sort the questions in order of increasing difficulty, compute $\beta_i + q_j$ (note that $q_j$ is just a constant offset) and threshold it at some reasonable place (e.g. p(correct) < 0.6). This gives a set of questions which the student is unlikely to answer correctly. You can now use hypothesis testing to see whether this is violated, in which case the student probably cheated (assuming of course your model is correct). One caveat is that if there are few such questions, you might not have enough data for the test to be reliable. Also, I don't think it's possible to determine which question he cheated on, because he always has a 50% chance of guessing. But if you assume in addition that many students got access to (and cheated on) the same set of questions, you can compare these across students and see which questions got answered more often than chance.

You can do a similar trick with questions. I.e. for each question, sort students by $q_j$, add $\beta_i$ (this is now a constant offset) and threshold at probability 0.6. This gives you a list of students who shouldn't be able to answer this question correctly. So they have a 60% chance to guess. Again, do hypothesis testing and see whether this is violated. This only works if most students cheated on the same set of questions (e.g. if a subset of questions 'leaked' before the exam).

Principled approach

For each student, there is a binary variable $c_j$ with a Bernoulli prior with some suitable probability, indicating whether the student is a cheater. For each question there is a binary variable $l_i$, again with some suitable Bernoulli prior, indicating whether the question was leaked. Then there is a set of binary variables $a_{ij}$, indicating whether student $j$ answered question $i$ correctly. If $c_j = 1$ and $l_i = 1$, then the distribution of $a_{ij}$ is Bernoulli with probability 0.99. Otherwise the distribution is $logit(\beta_i + q_j)$. These $a_{ij}$ are the observed variables. $c_j$ and $l_i$ are hidden and must be inferred. You probably can do it by Gibbs sampling. But other approaches might also be feasible, maybe something related to biclustering.

  • $\begingroup$ I read the first part of your answer and think it's promising. Two quick notes -- this was multiple choice so probabilities of guessing correctly are 25% or 20%. You are correct in that we may assume a subset of questions was leaked before the exam. Will return to this on Sunday or Monday. $\endgroup$
    – lockedoff
    Mar 5 '11 at 8:12

If you want to get into some more complex approaches, you might look at item response theory models. You could then model the difficulty of each question. Students who got difficult items correct while missing easier ones would, I think, be more likely to be cheating than those who did the reverse.

It's been more than a decade since I did this sort of thing, but I think it could be promising. For more detail, check out psychometrics books

  • $\begingroup$ Usually, cheating or guessing might be incorporated directly into an IRM. This is in essence what a 3-PL model intend to do, as it includes a parameter for difficulty, discrimination, and guessing which acts as a lower asymptote for the probability of endorsing an item. However, it has been proven to be unrealistic in most situations, and other dedicated person-fit statistics have been developed alongside (either in educational testing or psychological assessment). Meijer, Person-Fit research: An introduction. APM (1996), 9: 3-8 has a nice review on aberrant response patterns. $\endgroup$
    – chl
    Mar 6 '11 at 9:45
  • $\begingroup$ @chl Thanks! I studied this stuff in grad school, but that was long ago - my last class was in 1996 or so. $\endgroup$
    – Peter Flom
    Mar 6 '11 at 12:27
  • $\begingroup$ @chl Thanks for your suggestions. The model in my question is in fact an item response model (a Rasch or 1PL model with fixed discrimination parameter). I think the suggestion to look at individuals with aberrant performance is a good start, but I am looking for an approach that takes advantage of the additional information provided by the correlation in cheaters' responses for items on which there was cheating. You can imagine that if we used your procedure to identify cheaters, for example, they would perform well on similar difficult items. $\endgroup$
    – lockedoff
    Mar 7 '11 at 15:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.