He's not wrong, though

It would be interesting to see what the small-sample behaviour is for CIs obtained by inverting a score test instead of the wald or LR approach. iirc, the score test has better small-samples guarantees than either of those, but it's also rarely implemented in standard software.

It has been argued that early textbook publishers found that at 𝑛 ≥ 30 the t distribution’s p-values approximate normal tail probabilities, to the extent that the limited printing space on a textbook page could be spared by referring to z-tables for 𝑛 ≥ 30 (doi.org/10.1037/met0000448)

God my pc is actually crying, this is taking so long.

how even

Studying estimator bias is apparently very woke.

Apart from the philosophical argument not to mix up inference with estimation, I'd also argue that p-values simply aren't appropriate criteria to decide what to visualize. P-values on their own are barely sufficient for statistical inference per se, let alone to decide on effect importance.

Indeed, I agree. My first reply was maybe unclear. I agree that there is little point in taking some effect sizes more seriously than others based purely on a p-value, precisely because inferential errors can occur on either side of the 0.05 cut-off.

...inference can be dubious (e.g., small sample size, violated assumptions, etc.), one risks propagating the inferential procedure's susceptibility to error to the more purely mechanical algorithm of estimation.

I think it is a bad idea, because it conflates estimation and inference. The goal of depicting a regression line is to provide graphical intuition to the interpretation of an effect size, while the goal of plotting CIs is to graphically express inferential uncertainty. Knowing that...

Some would want to be edgy and say they prefer complementary log-log just to sound fancy, but then there are these authentic psychopaths who use linear probability because wreaking havoc is their second nature.

First, cite the appropriate papers to argue the opposite. If the reviewer persists, answer them succinctly as follows:

"Lay thine eyes upon my field of f*cks, and witness how it is empty."

Additional countermeasures may involve wooing the reviewer with sudden declarations of love and affection.

The reason I like the MANOVA here is because it omits the need for aggregating in a way that negates the apparent design of the data: individuals are distinct, measures are repeated within individuals. An average across participants and/or one across days ignores these aspects.

How about a MANOVA with between-subjects binary predictor Treatment, the repeated measures (days) being the dependent variable. No sphericity assumption, while respecting the within-subjects nature of the rep-meas. Excel implementation: real-statistics.com/multivariate...