You wrote that you "have doubts that I have perfomed the analysis correctly". I wonder if you have performed the correct analysis at all. You have proportion data, which is obviousely not normal distributed, so ANOVA is not a sensible choice in the first place. Why don't you use a model based on the beta distribution, or a quasi-binomial model, or at least use logit-transformed values? The next issue is that the data of adjacent time-points are not independent. Analyzing such a time-course as an ANOVA-design (each time-point being considered a "group") seems strange. It would make more sense to model the kinetics, either functionally or using a flexible basis like splines.

To sum up: it looks like you have choosen a bicycle to climb a tree and now you ask if the color of the saddle is a problem.

This is how I read it. I may be wrong. Happy to learn why your approach was in fact smart and sensible.

Jochen Wilhelm, I am not a GraphPad user, but I see that it is able to estimate repeated measures ANOVA models:

• https://www.graphpad.com/guides/prism/latest/statistics/repeated_measures.htm

I assume that Daria Malysheva treated Time as a repeated measures factor, not as a set of independent groups. (Note that if GraphPad is like SPSS, then the Greenhouse-Geisser correction (for violation of the sphericity assumption) is only available for RM factors.)

Daria Malysheva, you appear to be saying that GraphPad allows you to use something it calls "Dunnett's test" on a repeated measures factor. If it does indeed allow that, I think it is wrong to call it by that name, because Dunnett's test was designed for use with independent groups. As the Wikipedia page says, it uses a pooled error term:

"...the t-statistics are all derived from the same estimate of the error variance which is obtained by pooling the sums of squares for error across all (treatment and control) groups."

Source: https://en.wikipedia.org/wiki/Dunnett%27s_test

But for multiple comparisons involving the levels of a repeated measures factor, the usual advice is NOT to use the overall error term from the ANOVA model (i.e., MSTreatment x Subjects). Instead, one should use an error term based on only those levels involved in the comparison. I don't know the details of how GraphPad calculates (what it calls) Dunnett's test for repeated measures factors, but I doubt it's correct! YMMV.

Thom Baguley, I'm tagging you on this, because I suspect you have discussed some of these issues in your textbook (Serious Stats).

Jochen Wilhelm Thanks for your response! If I were to just analyse the rates of the reaction, that would have been pretty straightforward, it's what I've seen in articles from biochemical journals. However in this particular case I was interested in a pairwise analysis at each point, becuase, from what I've seen, data from similar experiments seem to be presented as simple bar graphs with SD's, rather than a rate plot, at least in more biologically-leaning articles. The use of ANOVA was just me extrapolating what I've seen in other articles (though in some cases it was just one substrate being researched so it was one-way so perhaps the distribution wasn't as skewed?). Thank you for suggesting distributions that could better suit my model. I have found some ANOVA analogues for non-normally distributed data, but they don't seem to be implemented in the software I'm familliar with. If I "unskew" the data (logarithming comes to mind, though I suspect it won't be as simple as that in this case) would a standard ANOVA still be innapropriate?

Bruce Weaver, yes it was an RMANOVA, apologies for not specifying. Thank you for pointing out the correct terminology and use of the Dunnett test. I'll see what options I have for the error term, I'm somewhat limited by what's available in GraphPad, unfortunately my programming skills aren't quite good enough yet for me to easily perform this analysis independently.

Bruce Weaver , thanks for commenting. I mean that the expected values across adjacent time-points are not independent, at least if the sampling rate is faster than mayor changes in the signal. This is different to the independence due to repeated-measures from the same subject or some common-variance structure. Here it is about stability of a substance, so one would expect some kind of exponential decay. At any later time point one can expect a lower signal. This information is valuable and should be used for modelling. Treating time-points as different treatments at which entirely differnet things can happen at any time-point (like ANOVA does) just does not seem something smart to do.

Daria Malysheva , I know that. One biologist starts using some inappropriate or at least suboptimal way or tool for an analysis, others see it and reapeat it, and soon the literature is filled with a quasi-standrad of an analysis which is suboptimal or inapproriate without being questioned anymore. It's usually not a good idea to learn statsitics by reading what other life scientists present in their papers. Better read good books from people having studied the subject.

Regarding "to unskew the data": I mentioned the logit transformation. The log-transformation will be similar when all values are < 0.5.

Bruce Weaver RM ANOVA should use paired t tests and probably a multiplicity correction as this doesn't require sphericity and respects the correlated structure in the data.

The GG correction is conservative and the HF correction liberal so I have seen a suggestion that you could average the two epsilon estimates. However generally I think HF is OK when epsilon is close to 1 and GG better when its close to lower bound.

Thom Baguley Thank you!

Jochen Wilhelm Thank you for elaborating about the logit transformation and for the general advice regarding statistics.