[R] Default t test in R is Welch's test, not Student's, and can be very problematic

[Curtis, David]

Hi all.

I'm just writing to draw your attention to this paper, which is Open Access:

Curtis, D. Welch�s t test is more sensitive to real world violations of distributional assumptions than student�s t test but logistic regression is more robust than either. Stat Papers (2024). https://doi.org/10.1007/s00362-024-01531-7
https://link.springer.com/article/10.1007/s00362-024-01531-7?utm_source=rct_congratemailt&utm_medium=email&utm_campaign=oa_20240304&utm_content=10.1007/s00362-024-01531-7

I hadn't really noticed that the default implementation of the t test in R uses Welch's method rather than Student's. I know some people prefer Welch's method but the paper describes how it can be systematically biased and produce extremely low p values in real world scenarios.

If you agree with me, you might want to consider generally using Student's test instead, by using t.test(x, y, var.equal = TRUE).

Code to demonstrate the problem and play around with different scenarios, as described in the paper, is at: https://github.com/davenomiddlenamecurtis/TestTTest

Regards


  *   Dave Curtis

--
Professor David Curtis MD PhD
UCL Genetics Institute
University College London



	[[alternative HTML version deleted]]