Why don't the results of the Wilcoxon matched-pairs test from Prism or InStat match the results of another program or book?

Ambiguity in the test itself

When performing the Wilcoxon matched-pairs test, there are two ways to do the calculations in some cases. The ambiguity only comes up only when some of the matched subjects have exactly the same value in each measurement. For example, if you are comparing measurements before an intervention with those after the intervention, there are two ways to compute the Wilcoxon test if some of the subjects have the same value before and after the intervention.

Prism and InStat use the method suggested by Wilcoxon and described in S Siegel and N Castellan, Nonparametric Statistics for the Behavioral Sciences and in WW Daniel, Applied Nonparametric Statistics (and many others). The subjects that show no change are simply eliminated from the analysis, reducing N. Other books show a different method that still accounts for those subjects, and this alternative method gives a different P value.

Continuity correction bug in  Prism 3 and 4

The approximate calculation of the P value includes a correction for continuity:

Z = (|T - Mt| - 0.5) / SigmaT

The value 0.5 is so called "continuity correction". Prism 3 and 4 used 0.25  instead of 0.5. Prism computes an exact P value when the sample size is tiny. With larger samples, the continuity correction matters little. So this bug has only a tiny effect on P values. It was fixed in Prism 5.00 and 5.0a.

Bug in Prism 3 and 4 reporting approximate vs exact p value

if the number of pairs is 15 or greater, Prism computes an approximate P value. With fewer than 15 pairs, with no ties, Prism computes an exact P value. The problem is when there are fewer than 15 pairs but there are ties. Prism 3 and 4 report that it computes the P value using an approximation, but in fact uses the 'exact' algorithm, which is not exact when ties are present. Fixed in 5.00 and 5.0a.  

.