(a) The concept of averaging. There is nothing wrong with averages. People here like maximizing expected utility, which is an average. “Effects” are typically expressed as averages, but we can also look at distribution shapes, for instance. However, it’s important not to average garbage.
(b) The fact that population effects and subpopulation effects can differ. This is true, and not surprising. If we are careful about what effects we are talking about, Simpson’s paradox stops being a paradox.
(c) The fact that we should worry about confounders. Full agreement here! Confounders are a problem.
I think one big problem is just the lack of basic awareness of causal issues on the part of the general population (bad), scientific journalists (worse!), and sometimes even folks who do data analysis (extremely super double-plus awful!). Thus much garbage advice gets generated, and much of this garbage advice gets followed, or becomes conventional wisdom somehow.
That depends. Mostly they are used as single-point summaries of distributions and in this role they can be fine but can also be misleading or downright ridiculous. The problem is that unless you have some idea of the distribution shape, you don’t know whether the mean you’re looking at is fine or ridiculous. And, of course, the mean is expressly NOT a robust measure.
There are three separate issues:
(a) The concept of averaging. There is nothing wrong with averages. People here like maximizing expected utility, which is an average. “Effects” are typically expressed as averages, but we can also look at distribution shapes, for instance. However, it’s important not to average garbage.
(b) The fact that population effects and subpopulation effects can differ. This is true, and not surprising. If we are careful about what effects we are talking about, Simpson’s paradox stops being a paradox.
(c) The fact that we should worry about confounders. Full agreement here! Confounders are a problem.
I think one big problem is just the lack of basic awareness of causal issues on the part of the general population (bad), scientific journalists (worse!), and sometimes even folks who do data analysis (extremely super double-plus awful!). Thus much garbage advice gets generated, and much of this garbage advice gets followed, or becomes conventional wisdom somehow.
That depends. Mostly they are used as single-point summaries of distributions and in this role they can be fine but can also be misleading or downright ridiculous. The problem is that unless you have some idea of the distribution shape, you don’t know whether the mean you’re looking at is fine or ridiculous. And, of course, the mean is expressly NOT a robust measure.