I am analysing a data that has normal distribution confirmed by Shapiro-Wilk Normality test (p=0.56) and Q-Q plot.
I divided my data in to two groups.
Following which, I applied Welsch test two-sided to find out significant features with a threshold of 0.05. I got around 40 significant features. Post this I applied corrections by using Benjamini-Hochberg FDR , now the result was 0 significant feature.
Can anyone throw light on this topic?
Hello Rakeshkumar,
You didn't say how many total features were compared, but given that you found about 40 to be significant without statistical adjustment for aggregate type I error risk, and that adjusted tests yielded none, I'd guess it was a lot.
Any of the adjustment procedures (Scheffe, Tukey, Holm, Bonferroni, B-Hochberg, etc.) accomplish the developer's stated goal of controlling aggregate risk of one or more type I errors by making the "significance" threshold for any single comparison more stringent. So, for your case, the more stringent threshold was such that none of your comparisons using the adjusted method was statistically significant. Perhaps in the future, it would make sense to determine the most important features to be compared (prior to analysis), and restrict the comparisons to just those features. Alternatively, you could increase the sample size to add statistical power to compensate for the more conservative tests that the adjustment procedure invokes.
Good luck with your work!
Hello Rakeshkumar Yadav . I am confused. You said your data are normally distributed, and reported one normality test result. That suggests there is just one dependent variable. But then, you talked about multiple (Welch) t-tests, 40 of which are statistically significant when you make no correction. Please clarify by answering the following questions.
Thanks.
PS- Bear in mind that some people reckon statistical tests of normality as precursors to t-tests are not helpful. Here's a short conference presentation that expresses my view on that topic.
Presentation Silly or Pointless Things People Do When Analyzing Data: 1. ...
Bruce Weaver David Morse Thank you Sir for your responses.
I have formed two groups on the basis of polluted and non-polluted sites (n=13 sites and n=11 sites). I want to know about diferentially abundant feature (Abundance of genes (total=300 genes)).
Since, I checked that my data is normally distributed, i did not apply non-parametric test (that's what I read).
After few readings, I thought instead of t-test, Welch t-test would be the appropriate choice.
Yesterday, I applied the same test on another data which involved the same sites and group. Prior to FDR correction I got 90 genes deferentially abundant, however, post FDR I got 4 genes. (This data too, as per the SW test was normally distributed).
Am I doing something wrong here?
Am I applying wrong statistics here? Please guide.
And surely, I will go through that link.
Kindly help.
Regards,
Rakeshkumar
Bruce Weaver Assuming that prior normality test should not be used as the criteria for t-tests and assuming that my sample size is big enough (I do not know how to prove "BIG SAMPLE SIZE", confused), I applied White's non-parametric test this time on the first data. I got the similar result as the earlier one where post-FDR correction on significant features remain.
Now highly confused!!
Please Lead!!!!
Hello Rakeshkumar Yadav . Re the point about small vs large samples, bear in mind that you can also rely on what is known about the variables from other studies. See the final paragraph of this BMJ Statistics Note, for example:
- https://www.bmj.com/content/338/bmj.a3166
I know almost nothing about genetics, so don't know how "abundance" is computed or what the distribution of abundance values usually looks like. Perhaps others can comment on that.
Re your other question, with 300 t-tests (one per gene), I'm not surprised that only 4 remain significant when you use the FDR approach.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics