FAQ/Poisson292019-11-11 12:23:04PeterWatson282013-03-08 10:17:58localhostconverted to 1.6 markup272008-07-24 11:42:20PeterWatson262008-07-24 11:42:01PeterWatson252008-07-24 11:40:07PeterWatson242008-07-24 11:39:43PeterWatson232008-07-24 11:29:58PeterWatson222008-07-24 11:29:32PeterWatson212008-07-24 11:26:12PeterWatson202008-07-24 11:25:03PeterWatson192008-07-24 11:13:18PeterWatson182008-07-24 09:26:05PeterWatson172008-07-24 09:25:06PeterWatson162008-07-24 09:23:43PeterWatson152008-07-23 14:37:03PeterWatson142008-07-23 14:36:39PeterWatson132008-07-23 13:17:26PeterWatson122008-07-23 13:15:00PeterWatson112008-07-23 13:13:53PeterWatson102008-07-23 13:13:12PeterWatson92008-07-23 13:11:41PeterWatson82008-07-23 13:11:07PeterWatson72008-07-23 13:10:21PeterWatson62008-07-23 13:08:04PeterWatson52008-07-23 13:06:48PeterWatson42008-07-23 11:58:16PeterWatson32008-07-23 11:57:46PeterWatson22008-07-23 11:55:06PeterWatson12008-07-23 11:54:09PeterWatsonHow do I compare observed numbers correct to those expected by chance in a multi-choice task?Suppose a questionnaire has a set of k possible responses to N questions of which one of the k choices is the correct response. Assuming by chance each of the k responses to each question is equally likely the number expected by chance equals N/k. Suppose we observe x $$\leq$$ N responses for a patient. We wish to see how likely x responses are given the patient responds at random to k choices on each of N questions. To do this we can assume the number of correct responses, x, follows a Poisson distribution, Po(m) of general form: $$P(X=x|\mbox{random responses}) = [mx ]/[x! ]e-m where $$m$$ is the expected total number of correctly answered questions from the N questions. Since this equals N/k we can rewrite the above as: $$P(X=x|\mbox{random responses}) = [(N/k)x ] / [x!] e-(N/k) one-tailed p-value = P(X \leq x) = (sum from 1 to x) P(X=x|\mbox{random responses}) = 0.5(two-tailed p-value) For large N the Poisson distribution is approximately Normally distributed with mean and variance both equal to N/k so we can analogously obtain a one-tailed p-value as: P(X $$\leq$$ x) = Probit( [x-(N/k)] divided by [Square Root{N/k}] ). Example Suppose we have 14 questions each with 3 possible responses of which only one is correct and a patient gets a total score of 3 correct responses. We wish to determine how likely it is we would observe 3 or fewer responses given the patient has responded at random (1 sided p-value). The expected number of correct responses assuming the patient is responding at random is 14/3=4.67. Using the Poisson distribution P(X $$\leq$$ 3) = P(X = 0) + P(X = 1) + P(X = 2) + P(X = 3) = (1 + 4.67 + [4.672 ]/[2] + [4.673 ]/[6])e-4.67 = 0.31. We can think of the p-value as the sum of numbers correct which are less likely assuming random responses than the observed one. It turns out that P(X $$\geq$$ 6) also accounts for less likely occurrences than the observed 3 correctly answered questions, given random responses. So the exact two-sided p-value is the sum of the Poisson probabilities P(X $$\leq$$ 3) and P(X $$\geq$$ 6) which equals 0.64. P(X $$\leq$$ 3) is approximately equal to P(X $$\geq$$ 6) because the Poisson distribution is symmetric about its expected value, N/k, the two-tailed p-value can also be computed as approximately equal to twice the one-tailed p-value = 0.31 x 2= 0.62. This is close to the p-value of 0.64 using the sum of the poisson probabilities above. The approximate or exact p-values from a Poisson distribution both conclude that there is no evidence to suggest a score of 3 on a three-choice task of 14 questions differs from chance responses. We can also evaluate probabilities of observing 3 correct responses due to chance using the Normal approximation to the Poisson distribution: $$P(X \leq 3) = Probit (\frac{3-4.67}{\sqrt{4.67}}) = Probit(-0.772) = 0.22$$ with a two-sided p-value of 0.44. The two-tailed probability equals $$P(X \leq 3) + P(X \geq 6.33)$$ since the Normal distribution probabilities are, like those of the Poisson distribution, symmetric about the mean of 4.67. Of course we can't observed 6.33 correct responses but this is a continuous approximation to the discrete Poisson distribution - like joining lines between frequency bars on a histogram of the number of correct responses! As with the Poisson distribution we conclude there is no evidence to suggest getting 3 questions correct on a three-choice task of 14 questions differs from chance. The exact Poisson two-sided p-value and its Normal approximation may be computed using a spreadsheet. In practice for over 30 questions (N) the Poisson and Normal approximations should closely agree. For less than 30 questions the Poisson is preferable as it is a discrete distribution assuming, as in this example, only integer values can occur (ie that numbers of correctly answered questions are whole numbers).