Yes, you can use more robust tests for highly skewed data. However, to define highly skewed is somewhat arbitrary. In the video below, I show that a non-parametric test has a higher statistical power for highly skewed data compared to a parametric test: ua-cam.com/video/mOzVwv9ob9Q/v-deo.html
Thank you for your video. In the case of n>30 and highly skewed data, would you prefer a non-parametric test over the option of bootstrap? e.g. in a scenario where you analyze group differences, would you use a Mann-Whitney U Test or an unpaired t-test with 10k bootstrap samples?
Hard to say because there are many types of skewed distributions. Anyway, in this video: ua-cam.com/video/mOzVwv9ob9Q/v-deo.html I show that the MWU test has higher statistical power than the t-test for a log-normal distribution. I also tried permutation tests, such as the one shown in this video: ua-cam.com/video/v7u8lHgoWig/v-deo.html and bootstrap confidence intervals (not shown in the video though) and they had a power between the MWU and the t-test. Thus for a log-normal distribution, MWU performs best. However, for other types of skewed distributions, you might get different results.
Im waiting already to your next video. Btw, cant we use more robust tests when the data is highly skewed? And how much is highly skewed?
Yes, you can use more robust tests for highly skewed data. However, to define highly skewed is somewhat arbitrary. In the video below, I show that a non-parametric test has a higher statistical power for highly skewed data compared to a parametric test:
ua-cam.com/video/mOzVwv9ob9Q/v-deo.html
Thank you for your video. In the case of n>30 and highly skewed data, would you prefer a non-parametric test over the option of bootstrap? e.g. in a scenario where you analyze group differences, would you use a Mann-Whitney U Test or an unpaired t-test with 10k bootstrap samples?
Hard to say because there are many types of skewed distributions. Anyway, in this video:
ua-cam.com/video/mOzVwv9ob9Q/v-deo.html
I show that the MWU test has higher statistical power than the t-test for a log-normal distribution. I also tried permutation tests, such as the one shown in this video:
ua-cam.com/video/v7u8lHgoWig/v-deo.html
and bootstrap confidence intervals (not shown in the video though) and they had a power between the MWU and the t-test. Thus for a log-normal distribution, MWU performs best. However, for other types of skewed distributions, you might get different results.
@@tilestats thank you very much for your fast reply!
Thanks for the video
Many thanks to him indeed
His videos come as a help in difficult times
great