Tomás Varnet Pérez

A suggestive question they can ask is whether it would make sense to write the discussion and conclusion reversing the roles of exposure and outcome (since associations are symmetric).

On Chile's national college common app platform, smart personalized info/assistance helped cut mistakes and boosting placements into higher-ranked programs by 20 percent, from Tomás Larroucau, Ignacio A. Rios, Anaïs Fabre, and Christopher Neilson https://www.nber.org/papers/w34164

This is one of those probability facts that drives my usual advice to people seeking intution for probability theory: Stop seeking intuition! It's not intuitive, and that's why it is so useful.

You can learn examples and reform your intution in time. But better to just trust the axioms and compute.

To avoid an LLM in the loop, perhaps some success could be had by translating and de-translating the text into one (or more) other languages.

Regarding 1., just off the top of my head, you could use LLM's to paraphrase or re-express the same content, in a way that gets rid of any idiosyncratic style that may be identifiable. Additionally, can request to (probabilistically) replace activities, places, family relationships.

In the same flavor of one of my favourite vignettes, @rmcelreath.bsky.social 's Statistical Rethinking on Akaike and the conception of the AIC:

Excerpt from an Interview with Jamie Robins (2022): "While waiting to cross a street in 1988, I suddenly realized how to construct a model that satisfied all my requirements - the structural nested failure time model (SNFTM)"

Another exhibit to the underappreciated importance of 'idle' activities for insights and connections.

Scikit-learn’s Defaults are Wrong This recent Tweet erupted a discussion about how logistic regression in Scikit-learn uses L2 penalization with a lambda of 1 as default options. If you don’t care about data science, this sou…

Is it any of these? ryxcommar.com/2019/08/30/s...

datascience.stackexchange.com/questions/10...

github.com/scikit-learn...

If you're trying to *predict* the likelihood of being an axe murderer, association is enough. That is true and uncontroversial. But not sure this is the point in discussion here.

Unsure what is the interpretation here. Is it 1) 'being an axe murderer' is seen as a latent class one belongs to even before commiting a murder (akin to a principal stratum), so that it makes them buy an axe even before its used as a weapon or 2) The scenario is a premeditated axe murder

However, you just introduced partial knowledge of that causal process, by saying that 'purchasing an axe' is a cause and that 'bodies in the car' are an effect.

Abstract from an article which says: Abstract The complex mechanisms of heredity are little appreciated by non-specialists, in some measure, because of misunderstandings that are perpetuated when words used for technical terms have other, more widely understood, folk meanings. When a word has both technical and folk meanings, it is the responsibility of the specialist to avoid promoting confusion by either using extremely cautious and precise language when using the term or, in cases when confusion is inevitable, abandoning the term in favor of one without a widely understood folk meaning. The study of heredity is beset by such confusion, and the term heritability appears to be at the heart of some of the confusion. In this article, I discuss both the technical and folk meanings of heritability and examine the bridge between them. By continuing to use the term heritability, we risk promulgating serious misunderstanding about the workings of heredity, therefore I suggest selectability as an alternative term to avoid such pitfalls.

As Stoltenberg (1997) was commenting a couple decades ago about 'heritability':

From the Guardian: To be clear, no one is saying a weak grip will kill you directly, in the way that heart failure will. But it usually indicates problems far beyond your hands and wrists. We don’t yet have the scientific studies to back this up, but it seems reasonable to assume that anything that improves it will also improve your overall health. It will certainly improve your quality of life.

"To be clear, no one is saying correlation implies causation ... but it seems reasonable to assume that it does."

Thanks, Guardian.

"According to the CDC, in 10 percent of those drownings, the adult will actually watch the child do it, having no idea it is happening."

Text from Philip Stark article 'Pay no attention to the model behind the curtain' (2022): Yep, the ‘facts.’ They proceed to estimate the effect that climate change will have on mortality, crop yields, energy use, the labor force, and crime, at the level of individual counties in the United States through the year 2099. They claim to be using an ‘evidence-based approach.’ ^35 Footnote 35: To my eye, their approach is ‘evidence-based’ in the same sense that alien abduction movies are ‘based on a true story.’.

Are you looking for an applied paper where they use a DAG to recognize they have confounding by indication or rather a theoretical treatment of the situation?

The status quo is unpaid work, but I haven't seen any proposal to transform the publishing system that wishes to keep it that way.

Continuation of the above: "Several test procedures develop critical values n⁎, such that micronumerosity is a problem only if n is smaller than n⁎. But those procedures are questionable. Remedies for micronumerosity If micronumerosity proves serious in the sense that the estimate of μ has an unsatisfactorily low degree of precision, we are in the statistical position of not being able to make bricks without straw. The remedy lies essentially in the acquisition, if possible, of larger samples from the same population. But more data are no remedy for micronumerosity if the additional data are simply “more of the same.” So obtaining lots of small samples from the same population will not help. If we return from this fantasy to reality, several lessons may be drawn. • Multicollinearity is no more (or less) serious than micronumerosity. Exact multicollinearity (R²ⱼ = 1) is a close analogue of exact micronumerosity (n = 0). When a research article complains about multicollinearity, readers ought to see whether the complaints would be convincing if “micronumerosity” were substituted for “multicollinearity.” • For example, if a test for exact multicollinearity is reported, the null hypothesis being R²ⱼ = 1, readers ought to consider whether they would test the null hypothesis n = 0. Or if a test for orthogonality is reported, the null hypothesis being R²ⱼ = 0, readers ought to consider whether they would test the null hypothesis that n is large. It is quite sensible to measure n, but would one want to undertake a statistical test on the true value of n? • For another example, if a rule is proposed to decide whether the collinearity is severe (how large R²ⱼ has to be before one says that there is a multicollinearity problem), readers ought to consider whether it is plausible to develop a rule that decides how small n has to be before one says that there is a small-sample-size problem."

What follows is from Chapter 23.3. of Goldberger (1991): "Econometrics texts devote many pages to the problem of multicollinearity in multiple regression, but they say little about the closely analogous problem of small sample size in estimation a univariate mean. Perhaps that imbalance is attributable to the lack of an exotic polysyllabic name for 'small sample size'. If so, we can remove that impediment by introducing the term micronumerosity. Suppose an econometrician set out to write a chapter about small sample size in sampling from a univariate population. Judging from what is now written about multicollinearity, the chapter might look like this: 1. Micronumerosity The extreme case, 'exact micronumerosity', arises when n = 0; in which case the sample estimate of μ is not unique. (Technically, there is a violation of the rank condition n > 0: the matrix 0 is singular.) The extreme case is easy enough to recognize. 'Near micronumerosity' is more subtle, and yet very serious. It arises when the rank condition n > 0 is barely satisfied. Near micronumerosity is very prevalent in empirical economics. 2. Consequences of micronumerosity The consequences of micronumerosity are serious. Precision of estimation is reduced. There are two aspects of this reduction: estimates of μ may have large errors, and not only that, but [the variance of the sample mean; DG] will be large. Investigators will sometimes be led to accept the hypothesis μ = 0 because [the ratio of the sample mean to its standard error; DG] is small, even though the true situation may be not that μ = 0 but simply that the sample data have not enabled us to pick μ up. The estimate of μ will be very sensitive to sample data, and the addition of a few more observations can sometimes produce drastic shifts in the sample mean. The true μ may be sufficiently large for the null hypothesis μ= 0 to be rejected, even though [the variance of the sample mean; DG] = σ2/n is large because of micronumerosity. But …

I always think of this excerpt from 'A Course in Econometrics' (Goldberger, 1991), where the tongue-in-cheek concept of micronumerosity (small sample size) is introduced as a parallel:

Micronumerosity leads to loss of precision, drastic changes with additional data, wrong hypothesis testing, etc.

Article title: "P-Curve Won’t Do Your Laundry, But It Will Distinguish Replicable from Non-Replicable Findings in Observational Research: Comment on Bruns & Ioannidis (2016)"

Makes me think of the title of this reply by Simonsohn and co.

The analog'd be smth like "Causal DAGs won't give you the lowest mean squared error estimator for your data and parametric context but it will distinguish between nonparametrically identifiable and nonidentifiable estimands", but catchy

Had also escaped my Greenland papers radar for quite some time somehow

Limitations of individual causal models, causal graphs, and ignorability assumptions, as illustrated by random confounding and design unfaithfulness - European Journal of Epidemiology We describe how ordinary interpretations of causal models and causal graphs fail to capture important distinctions among ignorable allocation mechanisms for subject selection or allocation. We illustr...

Not an exhaustive article, and also partially unorthodox (cf. 'random' or 'epidemiological' confounding), but Greenland and Mansournia's (2015) paper touches on some limitations.

Informally, I'd say a causal DAG only shows you what causes what (plus some neat identification implications from this)

But that is leaving quite some heavy work as an exercise to the reader. Maybe more critically, it also does not give the reader a way to concretely contest some of these assumptions or claims behind the conclusion, as they are not presented in any way, much less an explicit and unambiguous one. 3/2

I get what you are hinting at though. There is some sort of implicit assumption on the sparsity of variables involved and a vague restriction on which causal directions would make sense in that system, in which maybe the previous questions can be quantified as affirmative. 2/2

In which specific way would it be evidence? Inductively, is it more likely that it is causative given that they are associated than they are not (whatever that statistical model would look like)? From an error perspective, could estimating an association falsify the claim that there is a cause? 1/2

Which is an aspect that is often overlooked, if not outright conflated with identification. But I'd argue the issue there is with the over-selling of what a causal DAG can tell you rather than with causal DAGs not fulfilling their promise.

I'd say 2) is perhaps too ambiguous as to evaluate its wrongness. Causal DAGs are tools to establish non-parametric identification. I'd assume that 'understanding OS', as you suggest, goes beyond mere identification, and actually cares about estimation with finite data/blocks.

$119 Springer cancer treatments book: ‘As an AI language model …’ Peter Purgathofer looked over “Advanced Nanovaccines for Cancer Immunotherapy” by Nanasaheb Thorat, published in March by Springer Nature. On page 25, he found why we pay Springer the big bucks: [h…

We're so not ready for the AI slop deluge.

$119 Springer cancer treatments book: ‘As an AI language model …’ – Pivot to AI

That professional chess players burn thousands of calories through a match just by sheer cognitive effort.

Maybe not the main point of the tweet, but this is only in the (causal) risk difference scale, no? Average causal effects are not averages of contrasts, but contrasts of averages. Linearity of expectation makes this the same in the difference scale, but else an ATE is not an average of ICE's.