SCIENCEYE BLOG

Contact

Causal Machine Learning: Signposts, Levers and HDL

8 min read

Painterly illustration of a glowing signpost pointing toward a heart shaped beacon a brass lever on a stone plinth and glowing roots beneath the heart

SHARE:

By Dr. Lee Stoner

If you want to see why causal machine learning matters, start with high-density lipoprotein (HDL), the “good cholesterol.” For decades, people with more of it had fewer heart attacks. In one large analysis, each standard-deviation rise in HDL went with 38% lower odds of a heart attack.1 Raising HDL looked like an obvious way to prevent them. That turned out to be wrong.

Our earlier post, Why Machine Learning Can’t Answer ‘What If?’, asked whether a finding shows a cause or only a prediction. Researchers blur the two often: in 180 cohort studies from top medical journals, 26% mixed them up.2 This post follows HDL to show what causal machine learning can, and cannot, do about it.

A strong predictor can still be the wrong target.

Table 1. What this article covers.
Objective Outcome
Why a strong predictor can be the wrong targetTell a signpost from a lever
What causal machine learning tries to doKnow what its answers rest on
Why explaining a model is not explaining a causeRead model explanations with care

The HDL Lesson: A Signpost, Not a Lever

To be fair, HDL is a useful predictor. It feeds QRISK3, a heart-risk calculator built from English general-practice records. QRISK3 separates higher- from lower-risk patients well: its C statistic, a measure of this where higher is better, reached 0.88 in women.3 But predicting risk is not the same as changing it.

A 2007 trial gave 15,067 patients a statin, either alone or with torcetrapib, a drug that raises HDL. HDL rose by 72.1%, yet cardiovascular events rose too (hazard ratio 1.25, 95% confidence interval 1.09 to 1.44).4 The drug had off-target effects. Genetics offers a cleaner test, because genes are handed out at conception, before lifestyle or illness can sway them.1 People whose genes raise their HDL had no lower heart-attack risk (odds ratio 0.93, 95% confidence interval 0.68 to 1.26).1 HDL was a signpost of risk, not a lever. Its link to fewer heart attacks may come from hidden shared causes that raise HDL and protect the heart.

Signpost or lever: the HDL story Sketch of causes. Three boxes on the left: an HDL-raising drug, genes that raise HDL, and hidden shared causes that were not measured. The drug, the genes and the hidden causes each point to HDL level. The drug and the hidden causes also point to heart-attack risk. There is no arrow from HDL level to heart-attack risk, because raising HDL did not lower risk. An orange pulse runs along the drug’s two arrows. Signpost or lever? HDL-raising drug (torcetrapib) Genes that raise HDL Hidden shared causes (unmeasured) Exposure HDL level Outcome Heart-attack risk and events No arrow: raising HDL did not lower risk Sketch, not a validated DAG Signpost or lever: the HDL story Sketch of causes. Three boxes at the top: an HDL-raising drug, genes that raise HDL, and hidden shared causes that were not measured. The drug, the genes and the hidden causes each point to HDL level. The drug and the hidden causes also point to heart-attack risk. There is no arrow from HDL level to heart-attack risk, because raising HDL did not lower risk. An orange pulse runs along the drug’s two arrows. Signpost or lever? Sketch, not a validated DAG HDL-raising drug (torcetrapib) Genes that raise HDL Hidden shared causes (unmeasured) Exposure HDL level Outcome Heart-attack risk and events No arrow: raising HDL did not lower risk
Figure 1. A sketch of the HDL story, not a validated causal diagram. The drug raised HDL and also raised events; genes that raise HDL did not lower risk.1,4 The two arrows from the hidden causes are an untested assumption, and the missing arrow from HDL to risk is itself a claim. It cannot show how effects differ between people.

What Causal Machine Learning Tries to Do

Causal machine learning tries to find the levers before a trial does. Instead of asking who is at risk, it asks what would happen to an outcome if something changed, such as a treatment, and how that effect differs between people.5 That is a counterfactual question: what would the outcome have been under the option a person did not get?

The catch: the data cannot check the method’s key assumption: that every hidden shared cause, like those behind HDL, was measured. More data shrinks random error, but a missed shared cause biases every record alike. A 2026 roadmap for health data asks researchers to justify these assumptions and treat results as hypotheses.6 An ordinary risk model is no shortcut: its parameters and predictions should not be read causally.7

Table 2. Prediction and causal questions ask different things of the same data.6,7
Question What it estimates What it must assume
Who is likely to have a heart attack?Risk, from a clinical prediction modelNothing causal: its inputs are not levers
What if everyone took the drug?The average treatment effectNo unmeasured shared causes; each kind of patient received each option
Who would benefit most?Individualized treatment effectsThe same assumptions, holding in every subgroup

Causal machine learning can find levers, but only where the data hold the hidden causes.

Why Explaining a Model Is Not Explaining a Cause

What about explainable artificial intelligence (AI)? These methods show which inputs a model relied on, and how much, to reach a prediction. They explain the model, not the outcome. If a risk model ranks HDL near the top, HDL helps the model predict, which is exactly what HDL did for decades. A 2021 viewpoint argued that current explainability methods are unlikely to deliver their promises for decisions about individual patients.8 Ordinary models share the trap: reading every number a model reports as an effect leads to mistakes.9

Two questions about HDL Two panels. Left, the question explainable AI answers: did the model use HDL? Answer: yes, HDL helps it predict. Right, the causal question: does changing HDL change risk? Answer: in the trial and genetic studies, raising HDL did not lower risk. Explainable AI asks Did the model use HDL? Yes HDL helps it predict risk (a signpost) A causal question asks Does changing HDL change risk? Not when HDL was raised in a trial or by genes (not a lever) Two questions about HDL Two panels. Top, the question explainable AI answers: did the model use HDL? Answer: yes, HDL helps it predict. Bottom, the causal question: does changing HDL change risk? Answer: in the trial and genetic studies, raising HDL did not lower risk. Explainable AI asks Did the model use HDL? Yes HDL helps it predict risk (a signpost) A causal question asks Does changing HDL change risk? Not when HDL was raised in a trial or by genes (not a lever)
Figure 2. Explainable AI answers the left-hand question; only causal evidence answers the right-hand one.8

Takeaway

A predictor is a signpost, not a lever: it shows where risk is, but only a cause is something you can pull to change it. Next time a headline, app or health report says a number “lowers your risk”, ask whether that was shown by changing the number or only by measuring it. Before chasing a test number, ask your doctor: “Has changing this number been shown to help?”

“Prediction and causal explanation are fundamentally distinct tasks of data analysis.”10

Arnold and colleagues, International Journal of Epidemiology, 2021

References

Each source below was checked in PubMed or arXiv, and every link opens the original. Reference 6 is a preprint that has not yet been peer reviewed.

  1. Voight BF, Peloso GM, Orho-Melander M, Frikke-Schmidt R, Barbalic M, Jensen MK, et al. Plasma HDL cholesterol and risk of myocardial infarction: a mendelian randomisation study. Lancet. 2012;380(9841):572-80. doi:10.1016/S0140-6736(12)60312-2
  2. Ramspek CL, Steyerberg EW, Riley RD, Rosendaal FR, Dekkers OM, Dekker FW, et al. Prediction or causality? A scoping review of their conflation within current observational research. Eur J Epidemiol. 2021;36(9):889-98. doi:10.1007/s10654-021-00794-w
  3. Hippisley-Cox J, Coupland C, Brindle P. Development and validation of QRISK3 risk prediction algorithms to estimate future risk of cardiovascular disease: prospective cohort study. BMJ. 2017;357:j2099. doi:10.1136/bmj.j2099
  4. Barter PJ, Caulfield M, Eriksson M, Grundy SM, Kastelein JJP, Komajda M, et al. Effects of torcetrapib in patients at high risk for coronary events. N Engl J Med. 2007;357(21):2109-22. doi:10.1056/NEJMoa0706628
  5. Feuerriegel S, Frauen D, Melnychuk V, Schweisthal J, Hess K, Curth A, et al. Causal machine learning for predicting treatment outcomes. Nat Med. 2024;30(4):958-68. doi:10.1038/s41591-024-02902-1
  6. Tjandra D, Chang T, Parbhoo S, Ranganath R, Kurepa Waschka A, Mitchell W, et al. Causal machine learning is not a panacea: a roadmap for observational causal inference in health [preprint]. arXiv. 2026:2605.20782. arxiv.org/abs/2605.20782
  7. Lin L, Sperrin M, Jenkins DA, Martin GP, Peek N. A scoping review of causal methods enabling predictions under hypothetical interventions. Diagn Progn Res. 2021;5(1):3. doi:10.1186/s41512-021-00092-9
  8. Ghassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care. Lancet Digit Health. 2021;3(11):e745-50. doi:10.1016/S2589-7500(21)00208-9
  9. Westreich D, Greenland S. The table 2 fallacy: presenting and interpreting confounder and modifier coefficients. Am J Epidemiol. 2013;177(4):292-8. doi:10.1093/aje/kws412
  10. Arnold KF, Davies V, de Kamps M, Tennant PWG, Mbotwa J, Gilthorpe MS. Reflection on modern methods: generalized linear models for prognosis and intervention, theory, practice and implications for machine learning. Int J Epidemiol. 2021;49(6):2074-82. doi:10.1093/ije/dyaa049

Useful Resources

FAQs

What is causal machine learning?

It is machine learning that estimates what would happen to an outcome if something changed, such as a treatment, rather than only predicting the outcome. This Nature Medicine perspective sets out its uses and cautions.

Is HDL cholesterol still worth measuring?

Yes, as a signpost. After all, HDL helps risk calculators such as QRISK3 estimate who is at risk. The trial and genetic studies above found that raising it did not lower that risk. Your doctor can explain your own numbers.

What is a confounder?

A confounder is a common cause of both a supposed cause and its outcome, so it links them without one causing the other. Our post on directed acyclic graphs shows how researchers map them.

Can explainable AI show causes?

No. It shows what a model used to make a prediction, not what changing an input would do. Our post Why Machine Learning Can’t Answer ‘What If?’ explains the difference.

How is causal machine learning different from a trial?

A randomized trial lets chance decide who gets a treatment, as the National Institutes of Health guide to clinical trials explains. Causal machine learning works with data where people were not randomized, so it must assume the hidden causes were measured. Our post comparing study designs goes further.

Leave a Reply

Your email address will not be published. Required fields are marked *