Hello everyone.
I am conducting an experiment to evaluate if a compound inhibits a parasite growth quickly or slowly. To do this, I measure the IC50 of this compound when it acts for 24h, 48h or 72h. Up until now, I have been performing the ANOVA test, and Tukey as a post-hoc test. So, for example, if the IC50 of a compound is high (less potent) at 24h but low (more potent) at 72h, and Tukey indicates p < 0.05 for the pair 24h-72h, I'd safely say this compound has slow action.
However, suppose 24h and 72h acts exactly as I described above, but 48h have a high standard deviation. Since Tukey use the pooled variance, it would possibly make the test say that 24h and 72h aren't so different anymore, what is strange, since 24h and 72h alone still tell me the same thing. So, in my case, would it be better to just perform multiple t-tests instead of ANOVA and Tukey?
You say that 48h has a high stadard deviation. But maybe the standard deviations of 24h and 72h are too low?!
If the assumption of homogeneous variances is not justified, you should not use Anova or Tukey's HSD. And if it is justified, then what you see should be considered random variation that should not impact the testing procedure.
Hello Guilherme Souza. First, let's clarify whether you have 3 independent groups or repeated measures, where 3 measurements of IC50 are made on each independent unit of observation at the 3 time points. If it is the latter, we need to restart the conversation. But if it is the former (3 independent groups), you could use a form of one-way ANOVA that does not require homogeneity of variance--e.g., the ONEWAY command in SPSS offers both Welch's F-test and a Brown-Forsythe F-test. There are also several multiple comparison (MC) procedures that do not require homogeneity of variance. A recent simulation study compared some of them, and concluded that Games-Howell is a good choice. Here's the reference:
Sauder, D. C., & DeMars, C. E. (2019). An updated recommendation for multiple comparisons. Advances in Methods and Practices in Psychological Science, 2(1), 26-44. Article An Updated Recommendation for Multiple Comparisons
HTH.
If 48h has a high standard variation, what of 24h and 72h? As professor Jochen Wilhelm , have observed, if assumption of homogeneity of the variances is not met, you shouldn't have used ANOVA (but alternative for unequal variances Welch F test, you should consider running the Games Howell post hoc test. ). NOTE pls: if your sample size is small, and you aren't interested in interaction, the best way is the Multiple T-test so far you can account for the family-wide error. Sir, back to "the case", you have no case because from your statement, "Since Tukey use the pooled variance, #it_would_possibly_make the test say that 24h and 72h aren't so different anymore, what is strange, since 24h and 72h alone still tell me the same thing." You haven't done nor seen the results of any turkey test for the three, you only did for two(pairwise, 24h and 72h)! Do the Turkey test sir, and then complement with a *Dunnett's test (developed especially for multiple-to-one comparisons) and can be used when one fixed “control” group [IC50 for 24h] is compared to all of the other samples/groups, IC50s for 48h and 72h. If it's same conclusion, then compare 48h and 72h. If you don’t have a control group, use Tukey’s Test , known that is better, which is able to account for the familywide error rate, and compare not just for a pair. A good knowledge of your subject area, and research design makes it possible for you to identify that likely problem, and to take the whole experiment, and the post test results into sensible/meaningful decision making given your research question. You can see that better than anyone, and if t-test gives better ground for decision, there is no control, and conditions for it is justfy, it is better here compared to one way ANOVA.
Excuse me, let me say one more thing.
Please watch the video below first, which is about a dose-response study and Estimation IC50. I did this using GraphPad Prism software.
In the X axis, the logarithm and in the Y axis, the answers are repeated three times and in two times, 24h and 48h.
In fact, the important thing is that this type of study is neither a T-test nor an ANOVA. Rather, the best solution is to use a comparison of parameters or curves in the dose response study using graphpad prism.
Question.
Does it make sense to use ANOVA multiple comparison tests to compare two dose-response curves at every dose (or two time course curves at every time point)?
Answer.
My quick answer is No. Longer answer link below...
https://www.graphpad.com/support/faq/does-it-make-sense-to-use-anova-multiple-comparison-tests-to-compare-two-dose-response-curves-at-every-dose-or-two-time-course-curves-at-every-time-point/
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA