On the confidence tab of the nonlinear regression dialog, you choose how Prism should deal with difficult fits. If you choose the recommended method (new in 8.2), difficult parameters (and their confidence interval) will be marked as "unstable". If you choose the other method, entire fits may be flagged as "ambiguous" and certain parameters will be preceded with a tilde/squiggle (~). The The method used to identify an unstable parameter is very different than the method to identify an ambiguous parameter, so the results will vary based on your choice.

Prism 8 improved some algorithms, so sometimes fits curves faster than Prism 7 did. The results should be the same or similar in almost all cases. Prism 8 fixed a bug so now reports Hougaard's skewness properly even if with unequal weighting. Prism 6 and 7 reported incorrect results for Hougaard's skewness with unequal weighting.

Prism 5 and 6 use the same algorithms, so should always report identical results. Prism 7 introduced slightly improved algorithms, so the results might vary a bit, but the differences are trivial.

Prism 4 used slightly different algorithms, so curve fitting results can differ from results with later Prism versions in these cases:

•If your fit is labeled "Ambiguous" by Prism 5 or later, you know that some of the parameters are not determined precisely. Prism 4 presented a full set of results in this case, but the results are not useful when the fit is ambiguous.

•If you chose no weighting, check the sum-of-squares from the two programs. The goal of regression is to minimize that sum of squares, so see which version of Prism found a fit with the smaller sum-of-squares. Prism 5 and later have a few improvements in the fitting algorithm, so occasionally it can find a better fit than did Prism 4. The differences, if any, are usually trivial.

•If you chose to weight by the Y values (or the Y values squared), Prism 5 and later handle weighting differently than did Prism 4. Prism now weights by the Y value of the curve, while Prism 4 (and earlier releases) weighted by the Y value of the data. Weighting by the Y value of the curve is better, so the results of Prism 5 and later are more correct. Since the weighting is computed differently, you can't directly compare the weighted sum-of-square values reported by the two versions of Prism.

•When you compare two models, Prism now does an extra step. If one of the models is ambiguous, then Prism chooses the other model, without doing the F test or AIC comparison.

•Prism now offers more rules for defining initial parameter values. If your equation uses one of these new rules, Prism 4 might not be able to find a reasonable fit until you tweak those initial values. In particular, Prism now has smarter rules for fitting sigmoidal log(dose) vs. response curves.

•If you entered data as mean, SD (or SEM) and N, then Prism 4 (by default) fits the means and weights by the sample size (N). This is one of the two options on the weighting tab (the other option is to fit means only, ignoring N). Prism is now smarter by default (although you can choose to just fit the means and ignore N and SD values). It accounts not only for sample size N but also for the SD (or SEM) values you enter. With Prism 5 or later (but not Prism 4), you'll get exactly the same results from data entered as mean, SD and N as you would have by entering raw data. Prism 4 only accounts for differences in N, but not SD. The best fit values of the parameters, and thus the appearance of the curve, is the same with Prism 4 and 5. But Prism 5 and later does a smarter job with standard errors, confidence intervals, and comparisons of models.