eduKate Singapore

VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Next useful routeCanonical hubHELP

How AI Works | Doubly Robust Estimation — Outcome Models, Propensities and Cross-Fitting

—

by

eduKate Asia

Did you know? Estimating a treatment effect from observational data often requires two predictions that can each be wrong in a different way. A propensity model predicts who was likely to receive the treatment; an outcome model predicts what outcomes are expected given a person’s characteristics and treatment status. Doubly robust estimation combines them so that, under strong causal assumptions and suitable statistical conditions, an average treatment effect estimator can remain consistent if either one of those two nuisance models is correctly specified. That is a precise advantage, not permission to ignore confounding or data quality.

The useful workflow is to define the causal question before training, verify measured confounders and treatment overlap, fit treatment-assignment and outcome models without leaking each evaluation record into its own nuisance prediction, calculate an augmented inverse-probability-weighted score, then inspect extreme weights, uncertainty and alternative specifications. Cross-fitting is especially helpful when flexible machine learning models are used. If unmeasured common causes or absent treatment overlap defeat identification, no algebraic trick can recover the missing evidence.

This article explains present-day causal machine learning through first principles, an exact four-person calculation, a changed-context exercise and practical diagnostics. The worked observations are deliberately invented, with no claim to measure any real student’s or patient’s benefit. The underlying theory builds on Bang and Robins (2005) and the research on double/debiased machine learning.

Contents

Start with the counterfactual question, not an attractive algorithm

A supervised prediction question asks what outcome is likely for a person given observed information. A causal treatment-effect question asks how an outcome would differ if the same person received one intervention rather than another. We cannot normally observe both potential outcomes for one individual under the same conditions. Observational treatment groups also differ because assignment was not random. A model that predicts outcomes accurately can still fail to answer what the treatment caused.

Define A as a binary treatment indicator, with A=1 for treatment and A=0 for comparison. Let Y be the observed outcome, X a set of pre-treatment covariates and Y(1), Y(0) the potential outcomes under the two treatment choices. A common target is the average treatment effect, E[Y(1)−Y(0)] for a well-defined population. The minus sign represents a causal contrast, not merely a difference between two convenient columns in a spreadsheet.

The treatment itself must be specified carefully. ‘Extra tuition’, ‘a new medicine’ or ‘a training course’ might hide multiple versions, doses and levels of participation. If two people counted as treated received radically different interventions, the intended causal contrast can become ambiguous. The result also depends on who belongs to the target population and when outcomes are measured. These are design questions, not hyperparameters.

Suppose students who seek extra practice already differ in motivation, available time and previous understanding. Their observed grades may differ even without the practice programme. Comparing treatment-group and comparison-group means cannot separate baseline differences from the programme’s effect. A valid analysis needs a defensible account of the assignment process and the relevant measured confounders before using outcome prediction.

This article is narrower than a general introduction to predictive AI. It asks how to estimate one population-level causal effect when two nuisance functions can be estimated from observed data. The existing inverse propensity weighting guide develops one component. Here we study why adding an outcome residual correction can change the estimator’s robustness.

Three identification assumptions that the formula cannot manufacture

Consistency means the observed outcome corresponds to the potential outcome under the treatment actually received, within a well-defined intervention. This assumption becomes strained when the same label covers inconsistent versions of treatment or when one person’s treatment changes another person’s outcome. Interference is especially plausible in classrooms, social networks and policy interventions. The analyst must describe the intervention and relationships rather than dismiss those issues.

Conditional exchangeability means that, after adjusting for the measured covariates X, treatment assignment carries no remaining information about potential outcomes. Informally, the recorded variables must make treated and untreated people sufficiently comparable for the causal contrast. This does not mean the original observations were randomised. It is a substantive, generally untestable claim about measured and unmeasured causes. A beautifully fitted neural network cannot repair a confounder that was never observed.

Positivity, also called overlap, requires that people with relevant covariate profiles in the target population have nonzero chances of being treated and untreated. If everyone with a particular critical characteristic always receives treatment, the untreated outcome for that profile is not observed. The analysis then depends on extrapolation or a changed target population. A doubly robust score with huge inverse weights reveals instability; it does not create the missing comparison cases.

Temporal ordering belongs in the design as well. A feature measured after treatment may be an outcome or mediator rather than a valid baseline confounder. Conditioning on it can distort the causal interpretation, sometimes introducing selection bias. Define a time zero, treatment decision window, baseline covariate window and follow-up outcome window before fitting nuisance models.

The causal-inference literature, including Bang and Robins, makes double robustness a claim within an identification framework. It is not ‘robust against all bad data’. If exchangeability, consistency or positivity fails, even a perfectly optimised estimating equation may target the wrong quantity.

The two nuisance models solve different partial tasks

The propensity score is e(X)=Pr(A=1|X), the estimated probability of receiving the intervention given baseline features. In observational data it describes the treatment assignment mechanism, not the expected outcome. The outcome models are m1(X)=E[Y|A=1,X] and m0(X)=E[Y|A=0,X]. They describe conditional expected outcomes in the observed treatment groups. In a causal interpretation with exchangeability, they help represent potential-outcome means.

Call these nuisance functions because the main target is the treatment effect rather than either prediction surface itself. A propensity model can classify treatment decisions well while still being badly calibrated at important probabilities. An outcome model can predict outcomes well in the overall data yet be inaccurate for rare covariate profiles needed to compare treatment choices. Inspect both under the requirements of the causal estimator.

A plain outcome-regression estimator averages m1(X)−m0(X) across the target sample. A simple inverse-probability-weighted estimator uses observed outcomes weighted by the inverse of their probability of receiving the actual treatment. The first leans on the outcome model; the second leans on the assignment model. An augmented estimator combines a fitted outcome contrast with weighted residual corrections.

The residual for a treated observation is Y−m1(X), the departure of the actual outcome from its treated-outcome prediction. The residual for an untreated observation is Y−m0(X). The augmented score uses the appropriate residual and scales it by the corresponding treatment-assignment probability. In expectation, the correction can compensate for misspecification of one nuisance component when the other is correct, under the assumptions described in the original theory.

Do not train the propensity model on post-treatment outcomes, use features unavailable at assignment time or tune only for its apparent classification accuracy. Treatment probabilities near zero or one have unusual importance because they produce large weights. The nuisance-model evaluation should match the statistical job that it serves, not simply whichever metric looks impressive on a dashboard.

Read the augmented score one term at a time

For one observation i, a common augmented inverse-probability-weighted score for the ATE is ψ_i = m1(X_i) − m0(X_i) + A_i/e(X_i) × [Y_i−m1(X_i)] − (1−A_i)/(1−e(X_i)) × [Y_i−m0(X_i)]. The estimator averages these scores over the target sample. This form is for a binary intervention and a specified population; other estimands and study designs need different scores.

The first term is an outcome-model estimate of the treatment contrast for that person’s baseline profile. The second correction is active only when the person was treated, because A=0 removes it for an untreated observation. The third is active only for untreated observations and is subtracted. The treatment-specific residual is adjusted by the inverse probability of the treatment actually observed.

The formula should be read with its domain restrictions. e(X) must lie between zero and one, away from the extremes enough for stable estimation. If e(X)=0 for an actually treated observation or e(X)=1 for an untreated observation, a denominator becomes zero. Numerical clipping may prevent a divide-by-zero exception, but it does not prove that causal overlap exists. Clipping changes the estimator and must be reported as a modelling choice.

An augmented score can be negative even when the observed outcome is positive, because it includes residual corrections and counterfactual contrasts. It is not an individual treatment effect estimate with a ground-truth label. Summing its parts and dividing by the correct number of independent observations gives a sample estimate of the population quantity under the chosen assumptions.

The apparent symmetry matters: treated residuals receive positive weighting; untreated residuals are subtracted after weighting. A sign error or a mistaken propensity denominator can reverse the estimate. Before fitting a large model, verify the score with small hand-calculated cases like the four-row exercise that follows.

Worked example: four observations and a 2.833-unit estimate

Use four invented observations, with outcomes measured in arbitrary score units. Person 1 is treated: A=1, Y=8, e=0.50, m1=7 and m0=4. The outcome-model contrast is 7−4=3. Their treated residual is 8−7=1, so the weighted correction is 1/0.50 × 1 = 2. The untreated correction is zero. Score ψ1=3+2=5 units.

Person 2 is treated: A=1, Y=6, e=0.75, m1=6 and m0=4. The model contrast is 2; the residual is 6−6=0, so the augmentation vanishes. Score ψ2=2 units. Notice that this person’s propensity, 0.75, is not added or averaged into the outcome directly. It matters through a correction that is zero here.

Person 3 is untreated: A=0, Y=3, e=0.25, m1=6 and m0=4. The model contrast is 2. The untreated residual is 3−4=−1 and the inverse comparison probability is 1/(1−0.25)=1/0.75=4/3. The correction is minus (4/3 × −1), which adds 4/3. Score ψ3=2+4/3=10/3≈3.333 units. The double negative is intentional.

Person 4 is untreated: A=0, Y=5, e=0.50, m1=7 and m0=4. The model contrast is 3. Its untreated residual is 5−4=1 and comparison inverse probability is 1/(1−0.50)=2. Subtracting that correction gives ψ4=3−2=1 unit. A sign error here would incorrectly produce five instead of one.

Average the four scores: [5+2+(10/3)+1]/4 = (34/3)/4 = 17/6 ≈ 2.8333 outcome-score units. The fractions allow readers to check rounding. It is a sample estimating-equation result for artificial inputs, not proof that any actual intervention adds 2.8333 points. Four records are nowhere near sufficient for reliable causal inference, and these invented nuisance predictions do not validate the assumptions.

Why ‘doubly robust’ is a conditional mathematical property

The double-robustness idea can be understood by considering two hypothetical error situations. First, suppose the outcome models m1 and m0 are correctly specified. Their conditional residuals have mean zero under the appropriate groups, so the weighted augmentation has zero expected contribution under suitable regularity conditions even if the propensity model is not perfectly specified. The outcome-regression contrast carries the correct mean information.

Second, suppose the propensity model e is correctly specified but the outcome models are inaccurate. The weighted residual corrections can, in expectation, offset the bias from the misspecified outcome contrast. This is the special structure of the augmented score. It does not require both nuisance models to be perfect simultaneously for consistency in the relevant asymptotic setting.

Now suppose both models are wrong. There is no general double-robustness guarantee. Bias can remain or become severe. Furthermore, the mathematical claim presumes correct causal identification and enough data for the required convergence and variance conditions. Using flexible machine learning models does not automatically satisfy those conditions. A method that is robust to one nuisance-model error can still fail through confounding, measurement error, positivity violations or selection bias.

The name can also mislead people into imagining that a result is twice as accurate as an ordinary estimate. There is no such arithmetic meaning. It provides two routes to consistent estimation, not two independent proofs of causality. Both nuisance models are learned from much of the same data and can share bad features or hidden assumptions. Their failures are often correlated.

Bang and Robins’ 2005 Biometrics article introduced a widely cited modern treatment of doubly robust estimators in missing-data and causal models. The phrase refers to a proven estimating-equation property within specified models, not to a general assurance that combining any two AI predictions yields trustworthy results.

Extreme propensities can overwhelm the correction

Consider a treated observation with propensity e=0.02. Its treated residual multiplier is 1/e=50. If the observed outcome exceeds m1(X) by one unit, the augmentation contributes fifty units to that person’s score. For a comparable treated observation with e=0.50, the same one-unit residual contributes only two units. The first contribution is twenty-five times as large. This is a consequence of the formula, not necessarily a programming bug.

An untreated observation with propensity e=0.98 similarly has comparison probability 0.02 and an inverse factor fifty. When near-deterministic treatment assignment leaves few counterexamples, any observed exception carries enormous weight. Such cases can dominate a small sample estimate, lead to wide uncertainty and make findings highly sensitive to one record or propensity-model error.

A practical overlap audit should inspect the distribution of e(X) separately among treated and comparison observations. Plot or summarise near-zero and near-one tails, report how many people have limited comparable counterparts, and check what happens if the target population is restricted to a supported region. A report that lists only a mean propensity can conceal the exact observations driving instability.

Trimming, clipping or using alternative overlap weights can improve numerical stability in selected settings, but each can alter the estimand or introduce bias. Never silently cut an inconvenient probability from 0.02 to 0.10 and continue calling the result the same population effect. State the rule, motivation, affected cases and new target interpretation. If no defensible support exists, an honest answer may be that the desired causal contrast is not identified from these data.

The neighbouring inverse propensity weighting guide explains the standalone role of these weights. This article’s distinct contribution is the outcome residual augmentation and the conditions under which it changes the consistency argument.

Cross-fitting limits training-on-the-same-record bias

Modern nuisance models can be flexible: boosted trees, regularised regressions, random forests or neural networks. Such models may fit their own training observations unusually well, especially in small or high-dimensional samples. If the residual corrections are computed on the same records used to tune both nuisance functions, the estimates may inherit overfitting and regularisation bias. Cross-fitting separates model-fitting observations from the records used to evaluate scores.

Imagine 200 independent fictional records divided into two folds of 100 each. Fit the propensity and outcome models on the first 100 records, then predict both nuisance functions and compute ψ_i only for the second 100. Separately fit fresh nuisance models on the second fold and score the first fold. Every record then receives out-of-fold nuisance predictions. Combine the 200 scores and divide by 200 for a full-sample estimate.

More folds can use more training data per fitted nuisance model, but raise computational cost and sometimes instability. The number of folds, sample structure and independence units must fit the study. If five records are repeated measurements from one participant, putting some in training and others in held-out scoring can leak participant-specific information. Split by the unit that determines independence and assignment, not merely by arbitrary row number.

Cross-fitting is one of the central ingredients of double/debiased machine learning, along with appropriately orthogonal score construction. The formal claims require conditions on nuisance convergence and statistical dependence. Cross-fitting alone does not solve unmeasured confounding or positivity, and it does not convert a weak outcome model into a correct one by magic.

Preserve fold assignments, timestamps, feature creation rules and the source model used to predict each record. A result is only auditable if another analyst can verify that the nuisance prediction for observation i really came from a model not fitted on i or its leakage-equivalent duplicates. This is the same disciplined separation that honest validation requires in other machine learning problems.

Doubly robust estimation versus outcome-only and propensity-only methods

Outcome regression is attractive when observed covariates predict outcomes well. It models both treatment possibilities and averages their difference. But it can extrapolate badly to covariate combinations rarely observed in one treatment group. A polished predictive score is not proof that both potential-outcome surfaces are correctly specified in the regions needed for a causal comparison.

Inverse-probability weighting attempts to create a pseudo-population in which treatment assignment is balanced on recorded features, assuming a correct propensity model and proper support. It avoids committing to an explicit outcome regression but can have large variance when some treatment probabilities are tiny. The score may be unstable even if average propensity accuracy looks excellent.

The augmented estimator begins with outcome predictions and corrects them using propensity-weighted observed residuals. Under the relevant assumptions it has the familiar double-robust consistency property. This often makes it appealing when both nuisance functions can be estimated credibly. Yet finite-sample performance can still be poor, especially with extreme weights or flexible overfitting models.

Targeted maximum likelihood estimation is another framework combining outcome estimation and information about treatment assignment through a targeted updating step. It is related but not identical to simply averaging a raw augmented score. Different versions, estimands and model choices carry their own theoretical guarantees. One should not call every method involving two regressions ‘AIPW’.

Rather than rank methods by complexity, compare how they behave under the plausible data-generating mechanisms and the intended decision. The most advanced estimator is not always the right one for tiny, badly confounded data. A clear descriptive comparison may be more defensible than a highly precise causal number unsupported by identification.

Build the analysis before looking at the effect estimate

Start with a target-trial style question: who could receive treatment, when assignment occurs, what comparison is meaningful, which outcome is measured and how follow-up is defined. State inclusion and exclusion rules before using outcome information to refine treatment groups. Otherwise the sample can be quietly selected to favour a particular answer.

Draw a causal diagram or at least list plausible common causes of treatment and outcome. Record baseline measurements for those causes without adjusting indiscriminately for variables measured after the intervention. A covariate can be a confounder, mediator, collider or irrelevant predictor depending on the causal structure. Predictive feature importance does not classify these roles correctly on its own.

Estimate and inspect the treatment-assignment mechanism. If one subgroup virtually never receives treatment, acknowledge the limited comparison. Use descriptive balance diagnostics to see whether weighting reduces observed covariate imbalance, and remember that balance on observed variables cannot prove exchangeability with respect to unobserved causes.

Plan the nuisance fitting protocol, including any hyperparameter tuning and cross-fitting, before reading the final effect estimate. Select a main estimator and sensitivity analyses. Record the original data and code versions. Report confidence or uncertainty intervals using methods appropriate to the estimator, clustering and fold structure. A standard deviation of per-row scores is not automatically the right interval if observations are dependent.

Finally examine substantive plausibility without turning it into answer hunting. A surprising effect can indicate a real finding, but it can also signal a sign bug, support violation, post-treatment covariate or incorrectly handled outcome scale. Trace the calculation and assumptions rather than replacing the model repeatedly until the effect matches expectations.

What a confidence interval can and cannot repair

A point estimate such as 2.8333 units seems exact because the arithmetic contains four decimals. Its uncertainty can be enormous, especially with four invented cases. In a real study, standard errors should reflect the sampling design, independent unit, estimating equation and nuisance-estimation procedure. Clustered assignments, repeated outcomes and cross-fitted scores can require additional care.

Orthogonal scores reduce sensitivity to certain first-order nuisance-estimation errors under suitable conditions. This supports inferential methods even when flexible machine learning is used. It does not imply the final interval accounts automatically for model selection, measurement error, interference, hidden confounding or departures from positivity. A frequentist confidence interval quantifies selected sampling uncertainty under a maintained model, not the probability the entire causal story is true.

Bootstrap strategies may help in some configurations but cannot be assumed valid for every complex estimator without checking the underlying theory and dependency structure. Always distinguish the standard error from the spread of the observed Y values and from the dispersion of predictions. These quantities answer different questions. Extreme AIPW scores can lead to unstable estimated variance that requires inspection.

Sensitivity analysis should ask how the result changes when plausible unmeasured confounding, model misspecification or overlap restrictions are considered. Such exercises cannot identify exactly what hidden factors exist, but they can show whether a decision depends on fragile assumptions. Reporting an impressive narrow interval while ignoring an implausible treatment-assignment story is a false sense of certainty.

Use human-readable scale language. A 2.8-point effect on a ten-point quiz has a different meaning from 2.8 units on a continuous laboratory measure. If the outcome has a ceiling or meaningful subgroup variation, a single population average may conceal relevant consequences. No AIPW equation decides what improvement matters to the people affected.

Eight ways a correct formula can still produce a wrong conclusion

Failure one is untreated confounding. A missing factor that affects both treatment and outcome violates the exchangeability story. Repair it through better study design, richer pre-treatment measurement or honest sensitivity analysis; do not claim that two fitted models erase hidden causes. Failure two is lack of overlap. Diagnose it through propensity distributions, rare treatment patterns and support checks; change the estimand if justified.

Failure three is leakage from future or post-treatment variables. Repair it by defining time zero and rebuilding features from information available before assignment. Failure four is a propensity denominator bug. A treated record uses e, while an untreated record uses 1−e. Verify both with hand calculations before scaling up.

Failure five is calculating nuisance predictions in-sample while claiming out-of-fold cross-fitting. Repair it by recording explicit train and score indices for every fold. Failure six is ignoring duplicate participants or households across folds. Repair it by splitting at the appropriate independent unit and modelling interference or clustering where material.

Failure seven is changing the outcome or comparison group after seeing the result. The new analysis may answer a different question, and selective exploration invalidates the original inferential story. Keep an immutable primary analysis plan and clearly label exploratory revisions. Failure eight is silently clipping extreme propensities. Report the altered rule, affected sample and interpretation rather than hiding unstable support.

Notice that only one of these eight is purely an arithmetic issue. Most concern whether the data and experimental design support the causal contrast. The role of reliable machine learning is to estimate the nuisance functions within a valid design, not to replace the design with a more persuasive number.

Changed-context exercise: comparing two approved training options

Imagine an organisation comparing optional training A with a standard programme. Participation is not randomly assigned, and teams differ in experience. An analyst records baseline experience, workload and prior performance, then uses outcome models and a propensity model to estimate the effect of A on a later measured skill score. Before doing any arithmetic, they must decide whether programme versions are comparable, whether team members influence one another and whether all important assignment causes were measured.

Suppose a hypothetical treated employee has e=0.20, observed skill score Y=72, predicted treated outcome m1=70 and predicted untreated outcome m0=65. The outcome-model contrast is 70−65=5 score units. The treated residual is 72−70=2; its inverse-propensity correction is 2/0.20=10. The augmented score is 5+10=15 score units. This is a contribution to the average estimating equation, not an individual ground-truth treatment effect.

Now change one condition: the same case has propensity e=0.02, with every other value held fixed. The treated residual correction becomes 2/0.02=100, and the score becomes 105 units. The outcome’s observed scale has not suddenly changed. The large score is an instability signal that the model is extrapolating from a rare treatment pattern. A report should inspect why the estimated propensity is so small and whether the target comparison is supported.

If the organisation uses one team leader to approve treatment for an entire team, individual employees’ assignments are not independent. The analyst may need team-level folds and cluster-aware uncertainty. Splitting teammates across nuisance-model train and evaluation sets would not necessarily protect against leakage or shared outcomes. The design change matters more than the algebra.

The transfer lesson is that the same formula moves between domains only when the causal question, time zero, observation unit and support conditions are rebuilt. AIPW is not a universal stamp of validity, and rare-event weights should prompt investigation rather than a celebratory claim about a larger treatment benefit.

Limits and ethical use of observational treatment estimates

A doubly robust estimate cannot establish causation when its identification assumptions are unsupported. It cannot prove that a model correctly controlled all confounders, remove bias from wrongly recorded outcomes or guarantee overlap in groups that lack untreated comparisons. It does not estimate an individual patient’s personal benefit with certainty. Group averages can conceal large variation and even important harm to subgroups.

Using automated causal scores in high-stakes decisions can amplify injustice if relevant groups were underrepresented or assignment was shaped by inaccessible past policies. Inspect which individuals have support and which are effectively extrapolated from others. Document the target population, excluded groups, uncertainty and ethical constraints before translating an effect into policy.

Data minimisation still matters. A propensity model may tempt investigators to collect as many personal details as possible, yet not every variable is a valid confounder or appropriate to process. Use a reasoned causal set, consent and lawful governance where applicable, and protections for sensitive records. No technique requires making unnecessary individual-level information public.

Double robustness is a valuable statistical property when its conditions hold; it is not a factual guarantee about the world. If both nuisance models are flawed or key causal assumptions fail, it is entirely possible to produce a precisely calculated but misleading estimate. The honest ending may be that the data cannot support the proposed causal claim.

All worked numbers above are fictional and designed for reproducible arithmetic. The linked primary research describes general theory rather than certifying the performance of an AI system on a live study in this article.

Frequently asked questions

What is doubly robust estimation? It combines treatment-assignment and outcome models in an augmented estimating equation that can remain consistent under suitable causal and statistical assumptions when either nuisance model is correctly specified.

Does that mean both models can be wrong? No. If both are misspecified, the usual double-robustness guarantee does not apply. Unmeasured confounding, no overlap and bad timing can also invalidate the causal interpretation.

What does AIPW stand for? Augmented inverse probability weighting. An outcome-model contrast is augmented by propensity-weighted residuals from observed treated and untreated outcomes.

Why divide the residual by a treatment probability? The correction reweights observations according to the probability of receiving the treatment state actually observed. Rare assignments can receive large weights, which is both the method’s leverage and a source of instability.

Is a negative individual score an error? Not necessarily. AIPW scores include adjusted residuals and can lie outside the raw outcome range. Inspect the formula and denominators before deciding whether it is a mistake.

What is cross-fitting? It scores each record using nuisance predictions from models fitted on other observations or folds, reducing overfitting-related dependencies when appropriately implemented.

Does high propensity classification accuracy prove validity? No. The model must estimate relevant probabilities reliably, especially tails. Causal identification also requires measured confounders and suitable treatment overlap.

Can I use AIPW on data from a randomised trial? It can be appropriate, often with known randomisation probabilities and outcome adjustment, but the correct estimand, design and inference method should be specified. Random assignment does not eliminate every missing-data or interference issue.

Is the effect estimate an individual benefit? No. The average score estimates a population contrast under assumptions. It does not reveal both potential outcomes for any one individual.

Reader pathways, related explanations and primary research

If you are a student, reconstruct the four invented scores and check that the sum is 34/3 and the mean 17/6. Then change the third person’s observed outcome to four, keeping everything else fixed, and compute the new score: the untreated residual becomes zero, so ψ3 becomes two rather than 10/3. The mean becomes [5+2+2+1]/4 = 2.5 units. This arithmetic is a practice problem, not new causal evidence.

If you build a causal-ML pipeline, begin with an explicit treatment, outcome, time zero, covariate diagram and target population. Check measured confounding and overlap before selecting models; use leakage-safe cross-fitting, store nuisance predictions, audit extreme corrections and preserve the complete estimating equation. The earlier inverse propensity weighting guide is a useful companion, while causal forests addresses a different goal of treatment-effect heterogeneity.

Primary research: Bang and Robins (2005), Doubly Robust Estimation in Missing Data and Causal Inference Models; and Chernozhukov and colleagues (2018), Double/debiased machine learning for treatment and structural parameters. These are sources for the estimating-equation properties and cross-fitting argument, not endorsements of an illustrative effect estimate.

Continue with the How AI Works guide and education that transfers into independent evidence checks. No observational causal analysis, model fitting or live-data claim was performed here.

Discover more from eduKate Singapore

Subscribe now to keep reading and get access to the full archive.

Continue reading