Causal ML (Causal Machine Learning)

Causal machine learning uses flexible models to estimate causal effects that vary across units — the effect for this customer or this segment rather than the average. It sits between two fields and inherits a constraint from each, which is the most useful thing to understand about it.

What it relaxes is the functional form. What it does not relax is identification. A flexible model can learn a complicated relationship between covariates and outcome without you specifying it; it cannot learn whether the comparison it is fitting is causal.

Row two is the structural fact, and it is the same one that defines the field: individual effects “cannot be identified, i.e., they cannot be expressed as a function of the observed data,” while “average causal effects can sometimes be identified from data, even if individual causal effects cannot.” So an estimate of a per-unit effect is a modeled quantity, not a measured one — even a perfect model has no per-unit label to be right about.

The assumptions do not get easier

“Causal inference from observational data requires two elements: data and identifiability conditions,” and a more expressive model can estimate functional relationships more flexibly, but does not supply additional observations or establish the identification conditions. Two of the conditions bite harder when the target is a per-segment effect.

• Exchangeability has to hold conditional on “known measured covariates” — so an unmeasured confounder biases the estimate exactly as it would in a regression, and a flexible model will fit the biased relationship confidently and smoothly.
• Positivity requires that “the probability of receiving every value of treatment conditional on \(L\) is greater than zero” — and now it must hold within every region you report an effect for. If nobody in a segment was ever treated, there is no comparison to make there, however confidently the model extrapolates into it.

The second point is the practical failure mode people meet first: a model returns a plausible effect for a segment that contains no treated units, because interpolation does not know it has left the data. Checking overlap per segment is not optional. See causal effect for the conditions and causal inference for the surrounding field.

Which is also why experimental data is worth so much more here than elsewhere: “in ideal randomized experiments the identifiability conditions hold by design,” so causal ML on experimental data is estimating heterogeneity within a valid comparison rather than arguing for one. See A/B testing.

Evaluation is the hard part

With no per-unit truth, you cannot score predictions against labels. What you can do is check whether the model’s ranking is useful in aggregate: group units by predicted effect and compare the actual average outcomes of treated and untreated within each group. If the top group shows a larger realized difference than the bottom group, the ranking carries information; if the groups look alike, it does not, regardless of how well the model fits.

A randomized holdout supports these within-group comparisons through its assignment design, subject to adequate sample sizes and implementation checks. Observational evaluation can also support causal claims under explicit identification assumptions and suitable adjustment; raw treated-versus-untreated group differences do not establish those assumptions. The curve-based versions of this check are covered under uplift curve and Qini coefficient, and the decision-facing use is uplift models.

One boundary worth stating. A predictive model cannot be repurposed for an intervention decision: a model that ranks customers by churn probability ranks them by risk, not by how much a retention offer would help, and those orderings can be almost unrelated. Substituting one for the other is the most expensive mistake in this area and it produces campaigns that target people who were going to stay anyway. See treatment effect and, for a specific estimator family, causal trees.

References

• Miguel A. Hernán & James M. Robins, Causal Inference: What If (Chapman & Hall/CRC, 2020)