An effect is estimated by subtracting a constructed baseline representing what would have obtained absent the intervention, attributing the residual to the intervention. The move reduces a causal question to an observational baseline plus trivial arithmetic — so the inference's strength is the baseline's credibility, never the subtraction.
How would you explain it like I'm…
The Twin-Plant Trick
Suppose you water one plant and leave a twin plant alone, then see how much taller the watered one grew. You subtract the un-watered twin's height to find out what the water did. The twin shows what would have happened without the water, so the extra growth is the water's effect.
Subtract the Would-Have-Been
Counterfactual subtraction is a way to measure the effect of something by subtracting a baseline that stands for what would have happened without it. You take the outcome you actually saw, subtract the would-have-been outcome, and call the leftover the effect of the change. The subtraction itself is easy arithmetic; the hard and important part is building a believable baseline, because the no-change world never actually happened, so you cannot just look it up. If your baseline guess is good, your answer is good; if your baseline is wrong, the whole estimate is wrong even though the subtraction was correct. That is why all the care goes into how you build the baseline, not into the subtracting.
Observed Minus Baseline
Counterfactual subtraction is the pattern of estimating a quantity of interest, an effect or impact, by subtracting a baseline that represents what would have obtained absent the intervention. The estimate is the observed outcome minus the counterfactual baseline, and the residual is attributed to the intervention. Crucially, the whole validity rests on the baseline being a credible stand-in for the unobserved counterfactual, not on the arithmetic, which is trivial. Structurally it is a reduction: a causal question (what did the intervention do?) is turned into an observational question (what would have been observed without it?) plus a cheap subtraction. Different methods, like difference-in-differences, randomised control, synthetic control, or a gene knockout, are just different constructions of the same baseline, with the subtraction identical across all of them. It is powerful because it imports ordinary estimation machinery onto a causal question, but fragile because the baseline construction, unlike the subtraction, is not estimable from the data and rests on identifying assumptions that live outside the arithmetic.
Counterfactual subtraction is the structural pattern of estimating a quantity of interest, an effect, attribution, impact, or signal, by subtracting an observed or modelled baseline that represents what would have obtained absent the intervention or event of interest. The estimate is the difference between the observed outcome and the counterfactual baseline, and the residual after subtraction is attributed to the intervention; the move's validity rests entirely on the credibility of the baseline, not on the arithmetic, which is trivial. Four commitments: the baseline is not observed under the intervention but constructed to represent its absence; the subtraction nets out anything common to the observed and counterfactual conditions, so only the difference survives; the residual is attributed to the intervention; and the inference's strength is the baseline's credibility, not the subtraction step. The structural logic is a reduction: a causal question (what did the intervention do?) is reduced to an observational question (what would have been observed absent it?) plus an arithmetic operation. The arithmetic is cheap and estimable; the observational question, having no direct answer, is decided by the design of the baseline. Substrate-specific methods, difference-in-differences, randomised control, synthetic control, climate counterfactual simulation, gene knockout, BATNA comparison, are different constructions of the same counterfactual baseline, the subtraction invariant across all. It is powerful because it transports substrate-neutral estimation machinery (point estimate, standard error, robustness checks) onto an otherwise philosophical causal question, and fragile because the baseline construction, unlike the subtraction, is not estimable from the data; identifying assumptions carry the inference and live outside the arithmetic.
It exposes a move invisible in substrate vocabulary but identical across substrates: "the drug reduced mortality by X%" reads as "drug-arm minus placebo-arm was X% — and the placebo arm's credibility as a counterfactual is what the inference is really about."
It reduces a sprawling cross-substrate vocabulary to one operation with one shared failure mode and a four-move catalogue keyed to the baseline: audit, quantify, stress-test, disclose — replacing a separate critique apparatus for trials, evaluations, and attribution studies.
It separates the estimable part (the subtraction, with its standard error) from the assumed part (the baseline's fidelity), making the orthogonal point that tightening the standard error does nothing to repair a non-credible baseline.
Causal inference to climate: an econometrician recognises climate attribution as difference-in-differences with a no-forcing simulation in the role of the parallel-control unit.
Genetics and pharmacology: knockout and placebo trials are the same subtraction with different baseline-construction techniques.
The portable lesson: an effect estimate's credibility is bounded by its least-defensible baseline, not by the precision of its arithmetic.
A phase-III trial reports a 4-point mortality reduction as treatment-arm minus placebo-arm mortality; randomisation makes the arms exchangeable, so the residual is the drug's effect — but larger arms only tighten the standard error and do nothing to repair the baseline if blinding fails or dropout is differential.
Parents (1) — more general patterns this builds on
Counterfactual SubtractionpresupposesCounterfactuals — The ESTIMATOR that operationalises a counterfactual into a number: observed minus a constructed baseline standing in for the unobserved counterfactual. The file: 'one operationalisation of the broader concept', sitting 'one level below causality'. Presupposes counterfactuals (the metaphysical target).
Counterfactual Subtraction is not Counterfactuals because counterfactuals name the metaphysical unobserved alternative world, whereas this is the estimator that builds a baseline as a stand-in and nets it out.
Counterfactual Subtraction is not Selection Bias because selection bias is one failure mechanism of the baseline, whereas this is the whole subtraction structure within which that mechanism does its damage.
Counterfactual Subtraction is not Effect Size because effect size is the magnitude the subtraction yields, whereas this is the structural move that produces any magnitude and locates its credibility in the baseline.