Bootstrap

[Camila Cotrim]

Hello,

thank you very much for sharing the online tool. 

I'm quite new to estimation statistics and have some questions related to the bootstrap resampling.

I have attached a figure, so I hope it can help.

In this figure, I'm trying to analyse the effect of certain molecules on the activity of my protein. Sample 1 is blank and my control, and sample 3 is a positive control - a molecule that increases protein activity. 

When analysing the results of the other molecules using the mean difference, I have assumed that the greater the mean difference, the greater is the potential of the molecule to increase activity. However, looking at the bootstrap resampling (5000 as default) I noticed that samples 2, 3,4,6,7 and 12 don't approach the normal distribution nor the skewed distribution, which I assume is due to the small sample size and high variability. Based on that, my questions are: Is this the best way to analyse the data? Is there any bootstrap correction (similar to BCa correction) that can be applied in this case?

Thank you in advance,

Camila

[Joses Ho]

Hi Camila,

As you pointed out, the small sample sizes and high variability pose a very difficult challenge for data analysis, whether you are using estimation statistics, or null-hypothesis testing.

BCa correction is widely thought to be the most robust against non-normal distributions, such as the ones you have here.

You could try Cliff's delta (which the webapp provides) as an effect size rather than the mean difference, but given than some groups (eg 2, 3, 4) have no numerical overlap with the control group, your CIs will be virtually null-width.

Hope this helps,

Joses