Jeremy Labrecque 🇨🇦

Cohort fertility rates for the United States, by age 40, 45 and 50.

Most graphs of the fertility rate depict the 'period fertility rate', which is based on a single year's data and doesn't necessarily reflect how many children women actually have across their lifetimes.

I've used data from the Human Fertility Database to show the cumulative number instead:

Science is grounded in observation. Measurement is a tool for observation. Measurements should be evaluated for validity and reliability/uncertainty. Scientists who use measurements without understanding their properties are not really scientists at all.

People worry about black-box models.

This is black-box data.

Since the Ezra & Ta-Nehisi discussion is still happening: the main point I think most are missing is that Klein is saying the role of the journalist-intellectual is to do strategic politics, whereas Coates says the role of the journalist-intellectual is to tell the truth

…about the plausibility of the causal assumptions which requires strong knowledge of causal inference and strong substantive knowledge.

Yeah. One thing that I think is often missing when people push for causal language in the research question is that it in NO WAY makes your answer more causal. The only thing that gives a stronger causal interpretation to your estimate is convincing arguments…

One reason I oppose associational or other non-causal language in research questions is it lets people think they don't need causal inference to answer causal questions.

(When the underlying question is causal, of course. Which it almost always is.)

I'm gonna be saying "was silence not an option?" from now on

We argued with the editors but they failed to understand this basic idea. (And even if the student was interested in the causal effect, all the adjustment variables were mediators so you shouldn't adjust for them anyway). The student ended up taking the paper to another journal.

I knew a student who just wanted to report the unadjusted difference in some outcomes by sex. To know whether these were different in men and women. The editors insisted they needed to adjust for a whole host of other variables (as though the student was trying to answer a causal question).

And, in my experience at least, epidemiologists have the opposite problem you describe. They don't over-interpret the unadjusted estimate. They assume that the only things worth knowing must be adjusted for every available covariate. Which is very wrong.

Another reason might be that when your adjusted estimate is far from the real world, that suggests that you're relying on your model for inference which is a good thing to know, at least I like to know how worried I need to be about model specification.

And the unadjusted estimate does tell you something. In the absence of selection bias it is the descriptive difference in the outcome by exposure group. Descriptive statistics tell you the state of the real world which is very useful and important to know. But, ok maybe you're not interested in that

What you're saying is that you think you know the DGM with respect to age and therefore any estimate without adjusting for age is non-informative. I would say that you should never be so sure of your DGM. The data could have been sampled in a weird way. Or your population could be different.

Congrats to Moldova!!! Russia threw everything to destabilize the election and failed spectacularly.

Democracy won 🙌🏼

So exhausting..

A simple example, if I read a paper and adjustment for a variable I believe is a strong confounder results in no or little change in estimate (of a collapsible effect measure), that might make me wonder whether that variable is poorly measured, for example, or that there is some other problem.

If the estimate changed (or didn't change) in an unexpected way and you can't explain it, that's a sign that you don't understand your data-generating mechanism really at all. And I would therefore put much less credence in your adjusted estimate.

I've never really questioned reporting both so I'm willing to change my mind here.

But my way of thinking about this is: to convince me that your adjusted estimate is even close to the causal effect, you have to convince me that you understand the data-generating mechanism pretty well.

And to be fair, many academic journals also ban "causal" words like effect/impact/due to.

Canada Post “lost” 1 billion dollars last year?

How about, “it cost Canadians 1 billion dollars to have a national postal service” which works out to costing about $25 a year per person (population of Canada in 2024 = 40 million). Seems like a pretty reasonable cost to me.

I'd be happy to give a (virtual) talk to your lab or department or whatever if you think that would help. I have a talk specifically for this topic.

This is not because AI is generally useless. It is a tool that has to be carefully tested for possible uses and implemented in ways that create a net benefit. Like all tools. This is not happening.
So it may point to the possibility that a lot of business decision makers are useless.

People are making Rapture jokes like there's no tomorrow

The audience is health scientists so I wouldn’t consider causal effect to be jargon. But even among a general audience I don’t think “genuine relationship“ would be clearer. but maybe I’m wrong

What is a "genuine relationship"?

There's language for what they're talking about. It's the causal effect.

Demystifying Clone‐Censor‐Weighting to Studying Treatment Initiation Windows: An Example Using Publicly Available Synthetic Medicare Claims Data Background Clone-censor-weighting (CCW) can compare treatment regimens that are initially indistinguishable (such as starting treatment within specific time windows) without using landmarks or…

Demystifying Clone‐Censor‐Weighting to Studying Treatment Initiation Windows: An Example Using Publicly Available Synthetic Medicare Claims Data - Webster‐Clark - 2025 - Pharmacoepidemiology and Drug Safety - Wiley Online Library

When I talk about associational language people often accuse me of being the language police. I'm from Quebec so I know first hand that language policing sucks.

But look at that sentence in my previous tweet and tell me language isn't already being policed...

A sentence I just read in a published study:

"By identifying and adjusting for biases such as confounding, selection bias, and information bias, epidemiologists accurately estimate the genuine relationship between exposures and outcomes"

...the genuine relationship...

But also we spend most of our time teaching people how to answer questions and spend very little time on how to ask questions so in another way, it's not very surprising that people operate this way.

journals.lww.com/epidem/fullt...