Outcome-Model Standardization to a Target Population

Why Standardize an Outcome Model to a Target Population?

A treatment effect estimated in a study population need not equal the effect in the population where a decision will be made. The populations may differ in characteristics that modify treatment effects. Outcome-model standardization addresses this by estimating outcome means under each treatment for people with the target population’s covariate profiles, then averaging those predictions over the target population.

The source population supplies treatment and outcome data for fitting the outcome model. The target population supplies covariate data that describe whom the effect is intended to represent. The resulting estimate is not automatically causal or transportable: it relies on assumptions connecting treatment assignment, study participation, and potential outcomes.

Definition: Let $S=1$ indicate membership in the source study population and $S=0$ membership in the target population. Let $A$ denote treatment, $X$ the measured covariates used for adjustment, and $Y^a$ the potential outcome that would be observed under treatment level $a$. For a binary treatment $A\in\{0,1\}$, the target-population average treatment effect is $$ \psi_T = E[Y^1-Y^0\mid S=0]. $$ Outcome-model standardization estimates this quantity by predicting the mean outcome under each treatment for target-population covariate values and averaging those predictions over the target covariate distribution.

Identification and the Standardization Formula

The target effect is not observed directly: each person has an observed outcome under only the treatment received. To use source outcomes to estimate it, the following assumptions must be credible for the covariates $X$ included in the analysis.

Consistency: if a person receives treatment $A=a$, then the observed outcome equals that person’s potential outcome under that treatment, $Y=Y^a$. Treatment levels must also be defined clearly enough that this comparison is meaningful.
Conditional treatment exchangeability in the source: $Y^a$ is independent of $A$ given $X$ among people with $S=1$. In a randomized source study this can follow from randomization; in an observational study it requires adequate control of treatment-outcome confounding.
Conditional outcome transportability: for each treatment level $a$, the conditional mean potential outcome is the same in the source and target populations at a given covariate value: $E[Y^a\mid X=x,S=0]=E[Y^a\mid X=x,S=1]$. This says the source outcome model applies to the target after conditioning on $X$.
Positivity: covariate profiles represented in the target must be supported by the source data, and each treatment level must be represented among source individuals at those profiles. Otherwise, the required conditional means cannot be estimated from observed source outcomes without extrapolation.
Correct outcome-model specification: the fitted model must adequately estimate the source conditional mean for each treatment level over the relevant covariate range. This is an estimation requirement, not a consequence of the causal assumptions.

Define the source conditional mean under treatment $a$ as $m_a(x)=E[Y\mid A=a,X=x,S=1]$. Consistency and conditional treatment exchangeability identify this source mean with the source conditional mean potential outcome. Conditional outcome transportability then connects it to the target conditional mean potential outcome. Averaging over the target covariate distribution gives

\psi_T = E[Y^1-Y^0\mid S=0] = E\big[m_1(X)-m_0(X)\mid S=0\big].

The corresponding plug-in estimator fits the outcome model using source data, predicts each target individual’s outcome under both treatment levels, subtracts the predictions, and averages the differences. Equivalently, it averages the predicted outcomes separately by treatment and subtracts the two averages. If the target data contain $n_T$ individuals with covariates $X_1^{(T)},\ldots,X_{n_T}^{(T)}$, the empirical estimator is

\widehat{\psi}_T = \frac{1}{n_T}\sum_{i=1}^{n_T} \left\{\widehat m_1\big(X_i^{(T)}\big)-\widehat m_0\big(X_i^{(T)}\big)\right\}.

This uses individual-level target covariate data, not target outcomes. If the target data are a representative sample, their empirical covariate distribution approximates the target population distribution. If target sampling is not representative, the simple average may not represent that population without an appropriate sampling adjustment.

Worked Examples

Worked Example: Standardizing Risk Differences Across Two Risk Groups

Suppose a hypothetical source study estimates the following conditional risks of an adverse outcome. Assume the outcome is binary and the values below are estimated source conditional means after adjustment for the relevant covariates. The target population is described by 60% high-risk and 40% low-risk individuals.

Risk group	Source risk if untreated	Source risk if treated	Target proportion
Low risk	4%	2%	40%
High risk	20%	14%	60%

Under the assumptions above, use the source conditional risks to predict the target risks under each treatment. The target untreated risk is

\widehat E[Y^0\mid S=0] = (0.40)(0.04)+(0.60)(0.20) =0.016+0.120 =0.136.

Thus the standardized untreated risk is 13.6%. The corresponding treated risk is

\widehat E[Y^1\mid S=0] = (0.40)(0.02)+(0.60)(0.14) =0.008+0.084 =0.092.

The target-population risk difference, expressed as treated minus untreated, is $0.092-0.136=-0.044$, or $-4.4$ percentage points. Under the assumptions, treatment is estimated to reduce the adverse-outcome risk by 4.4 percentage points in this target population. This conclusion is about the target mix of risk groups; it is not the source-population average effect unless the source has the same covariate distribution as the target.

Worked Example: An Age-by-Treatment Interaction

Consider a hypothetical continuous outcome measured in points and age $x$, in years relative to a chosen reference age. A fitted source-population model is

\widehat E[Y\mid A,x,S=1] =10+0.5x-2A+0.4Ax.

The interaction term allows the predicted treatment contrast to vary with age. The target data contain 20 individuals with $x=0$, 30 with $x=1$, and 50 with $x=2$. First calculate the predicted contrast at each age by setting $A=1$ and subtracting the prediction at $A=0$:

\widehat m_1(x)-\widehat m_0(x) =(10+0.5x-2+0.4x)-(10+0.5x) =-2+0.4x.

At $x=0$, the contrast is $-2+0.4(0)=-2$ points. At $x=1$, it is $-2+0.4(1)=-1.6$ points. At $x=2$, it is $-2+0.4(2)=-1.2$ points. Standardizing these individual-level predictions over the target distribution gives

\widehat\psi_T =(0.20)(-2)+(0.30)(-1.6)+(0.50)(-1.2) =-0.40-0.48-0.60 =-1.48\text{ points}.

For a check, the average predicted untreated outcome is $(0.20)(10)+(0.30)(10.5)+(0.50)(11)=10.65$ points. The average predicted treated outcome is $(0.20)(8)+(0.30)(8.9)+(0.50)(9.8)=9.17$ points. Their difference is $9.17-10.65=-1.48$ points, matching the average of the predicted contrasts. Under the assumptions, the estimated target-population mean outcome is 1.48 points lower under treatment. The treatment coefficient $-2$ alone is only the predicted contrast at $x=0$; it is not the target average effect when the model includes an interaction and the target includes other ages.

Worked Example: Standardizing Predicted Risks From a Logistic Model

A hypothetical source study fits a logistic outcome model for a binary event. Suppose its predicted event probabilities for three covariate profiles are as follows, and the target population has 20% of profile 1, 30% of profile 2, and 50% of profile 3.

Target profile	Target proportion	Predicted risk if untreated	Predicted risk if treated
1	20%	0.1192	0.0573
2	30%	0.1824	0.1192
3	50%	0.2689	0.2315

To estimate the target untreated risk, average the predicted probabilities using target proportions:

\widehat R_0 =(0.20)(0.1192)+(0.30)(0.1824)+(0.50)(0.2689) =0.02384+0.05472+0.13445 =0.21301.

The standardized treated risk is

\widehat R_1 =(0.20)(0.0573)+(0.30)(0.1192)+(0.50)(0.2315) =0.01146+0.03576+0.11575 =0.16297.

The standardized risk difference is $0.16297-0.21301=-0.05004$, approximately $-5.0$ percentage points. The interpretation is a predicted target-population risk reduction of about 5.0 percentage points, conditional on the causal and modeling assumptions. The calculations average predicted probabilities, not the logistic model’s coefficients: coefficients describe the model’s conditional log odds, while these weighted predictions produce the specified target risk contrast.

Why the Procedure Works, and What It Does Not Guarantee

The identifying formula follows from iterated expectation. Begin with the mean target potential outcome under treatment $a$ and condition on $X$:

E[Y^a\mid S=0] =E\!\left[E[Y^a\mid X,S=0]\mid S=0\right].

By conditional outcome transportability, the inner conditional mean equals $E[Y^a\mid X,S=1]$. By consistency and conditional treatment exchangeability in the source, this equals $E[Y\mid A=a,X,S=1]=m_a(X)$. Therefore,

E[Y^a\mid S=0] =E[m_a(X)\mid S=0].

Subtracting this expression for $a=0$ from the expression for $a=1$ gives the target-effect formula. In practice, the conditional mean is estimated rather than known, and the target expectation is approximated by the target sample average. Thus the formula’s causal identification assumptions and the outcome model’s estimation accuracy are separate requirements: satisfying one does not ensure the other.

Standardization is useful because it directly targets the population distribution specified in the estimand. It can accommodate effect modification when the outcome model represents it and when predictions are averaged over target covariates. However, a flexible-looking model is not proof that its functional form is adequate. Predictions for target profiles poorly represented in the source may depend heavily on extrapolation, and an omitted effect modifier can make the transported conditional means wrong even if the source model fits its observed outcomes reasonably well.

Common Mistakes and Practical Checks

Using the source covariate distribution for the final average. That gives an average over the source population, not the target. Average predictions using target covariate data when the target population is the estimand.
Using only the treatment coefficient. With treatment-by-covariate interactions, a treatment coefficient is a contrast at a reference covariate value. Predict under both treatment levels for each target profile and average.
Ignoring the model scale. For a binary outcome and a target risk difference, average predicted probabilities under each treatment and subtract. Do not treat a logistic coefficient as a risk difference.
Assuming standardization repairs confounding or selection automatically. It does not. Unmeasured treatment-outcome confounding, unmeasured differences in potential outcomes across populations, or inadequate covariate support can invalidate the result.
Treating a precise estimate as proof of transportability. Precision describes sampling variability under the analysis model; it does not establish that the model or transport assumptions are correct.

Uncertainty should reflect estimation of the source outcome model and, when relevant, sampling of target covariates. A practical approach is to use a variance method appropriate to the source and target sampling designs or a bootstrap that resamples the relevant data according to those designs and repeats the entire model-fitting and standardization procedure. The interval then quantifies sampling uncertainty under the chosen procedure and assumptions; it does not incorporate all uncertainty about whether the causal assumptions hold.

Key takeaway: Outcome-model standardization transports a conditional outcome model to a target population by predicting both treatment outcomes for target covariate profiles and averaging those predictions over the target covariate distribution. A causal interpretation depends on treatment exchangeability, conditional outcome transportability, positivity, consistency, and adequate outcome-model estimation.

Check Your Understanding

Answer each question by referring to the target-population estimand and the standardization steps.

Write the target average treatment effect in potential-outcome notation, and express it using source conditional means $m_1(X)$ and $m_0(X)$.
A source outcome model includes a treatment-by-age interaction. Why is the treatment coefficient alone generally not the target-population average effect?
For a binary outcome, what quantities should be averaged to estimate a target risk difference from a logistic outcome model?
What assumption connects source conditional potential-outcome means to target conditional potential-outcome means?
Why can a target profile with little or no source support make standardized predictions unreliable, even if the fitted model returns a numerical prediction?