Comparing the fits of two models, when one is ambiguous. New choice in Prism 5.0b and 5.02.

New to Prism 5 is the diagnosis of an 'ambiguous' fit. This term is unique to Prism, but the idea is not novel. Prism reports 'ambiguous' when the dependency of one or more parameters is nearly 1.00.  This means that the parameters are so intertwined, that there are an infinite number of combination of parameter values that generate curves that fit equally well. The curve would fit well, but the best-fit values of the parameters are almost arbitrary.

If you choose to compare two models on the Compare tab of nonlinear regression and one or both fits is 'ambiguous', Prism 5.00, 5.01 and 5.0a does not continue with the comparison. If the fit of the other model is not ambiguous, Prism chooses that one to graph regardless of its sum-of-squares. If both fits are ambiguous, then Prism plots no curve.

If you are opening a file created by Prism 4, this means that Prism 5 may (in rare cases) plot a different model than Prism 4 did. Prism 4 fits the one with the lower sum-of-squares if the fit is enough better to be statistically significant (F test) or have a lower AICc. If one fit is ambiguous, Prism 5 picks the other model, regardless of sum-of-squares.  In some cases, Prism 5 will plot no curve even though Prism 4 did.

If the goal of the comparison is to find a scientifically plausible model, then it makes sense to avoid models with ambiguous fits, regardless of how well they fit the data. If your goal is just to plot an attractive curve, perhaps to use as a standard curve for interpolation, then plotting a model whose fit is ambiguous might make sense.

We updated Prism 5.02 (Windows) and 5.0b (Mac) to give you a new choice in the Compare tab:

[ ] If one fit is ambiguous, choose the other without formal comparison.

It is checked by default to match previous versions of Prism.