Standardizing Effects to the Trial Population

Why Standardize to the Trial Population?

An external control cohort and a clinical trial may differ in baseline characteristics that predict the outcome. A crude comparison of their event rates therefore combines treatment differences with differences in who was observed. Standardization addresses this by predicting outcomes at the covariate values of the target population and averaging those predictions over that population.

The target matters. If the question is the effect among people enrolled in the trial, the relevant covariate distribution is the trial population’s—not the external cohort’s and not an artificial population with equal numbers in each subgroup. Standardized risks and their contrasts then describe the expected outcomes for the trial population under each treatment strategy, subject to the assumptions that connect the observed data to those outcomes.

Definition: Let $A$ denote treatment, $X$ the measured baseline covariates, and $Y(a)$ the potential outcome that would be observed under treatment strategy $a$, for $a=0,1$. Let $S=1$ identify membership in the target trial population. The target average treatment effect (TATE) on the risk-difference scale is $$ \operatorname{TATE}=E\{Y(1)-Y(0)\mid S=1\}. $$ For a binary outcome, the corresponding target risks are $R_a=E\{Y(a)\mid S=1\}$, and $\operatorname{TATE}=R_1-R_0$. A negative risk difference means a lower event risk under strategy 1 than under strategy 0.

The Standardization Formula and Its Assumptions

For each treatment strategy, define $m_a(x)=E\{Y(a)\mid X=x,S=1\}$, the mean potential outcome at covariate value $x$ in the target population. For a binary endpoint, $m_a(x)$ is a risk between 0 and 1. If the fitted model provides these conditional means, the standardized target risk is

R_a=E\{m_a(X)\mid S=1\} \qquad\text{and}\qquad \operatorname{TATE}=E\{m_1(X)-m_0(X)\mid S=1\}.

With $n_T$ trial participants and baseline covariates $X_i$, a direct empirical estimate averages the predictions over trial participants:

\widehat{R}_a=\frac{1}{n_T}\sum_{i:S_i=1}\widehat{m}_a(X_i), \qquad \widehat{\operatorname{TATE}}=\widehat{R}_1-\widehat{R}_0.

This calculation uses each trial participant’s covariates for both treatment predictions. It does not compare the mean prediction for trial participants with the mean prediction for external controls. The latter would average over a different population.

The formula follows from the law of total expectation: partition the target population by its covariates, calculate the mean potential outcome within each covariate pattern, and average those conditional means using the target population’s covariate distribution.

E\{Y(a)\mid S=1\} = E\!\left[E\{Y(a)\mid X,S=1\}\mid S=1\right] = E\{m_a(X)\mid S=1\}.

This identity motivates the calculation, but observed outcomes do not automatically reveal the conditional potential-outcome means. Interpreting fitted predictions as those means requires assumptions. The covariates used for adjustment must be measured before treatment and must capture the relevant confounding and differences between data sources. The outcome, treatment strategies, eligibility criteria, and follow-up period must be sufficiently aligned for the comparison to represent the stated question.

Consistency: a participant’s observed outcome equals the potential outcome under the treatment strategy actually received. Treatment strategies and outcome definitions must be specified clearly enough for this equality to be meaningful.
Conditional exchangeability: within levels of the measured baseline covariates, treatment assignment is not associated with the potential outcomes relevant to the comparison. For an external-control estimate of $Y(0)$, this requires that, conditional on $X$, external controls provide valid information about the control potential outcomes of trial participants.
Conditional transportability: after conditioning on the included covariates, the outcome model learned from its data source applies to the target trial population. For example, control outcomes must not still differ by data source for reasons omitted from $X$.
Support and model adequacy: the covariate patterns in the target population must be sufficiently represented in the data informing each prediction, and the fitted model must estimate the relevant conditional means adequately. Predictions in unsupported regions depend heavily on extrapolation.
No interference: one participant’s treatment does not change another participant’s outcome, unless such interference is explicitly incorporated into the estimand and model.

In a randomized trial, randomization can support exchangeability for trial participants, while the external cohort may provide information about control outcomes. In a nonrandomized comparison, exchangeability must be justified for both treatment strategies. Neither a Bayesian model nor a narrow posterior interval makes these assumptions true; the model represents uncertainty conditional on its structure and assumptions.

Worked Example: Standardizing External-Control Risk by Subgroup

Suppose a hypothetical trial has 100 participants: 60 have a high-risk baseline profile and 40 have a low-risk profile. A fitted model using hypothetical external controls estimates 12-week control risks of 0.20 for the high-risk profile and 0.05 for the low-risk profile. Assume, for this calculation, that these conditional risks apply to trial participants with the same profiles.

The standardized control risk uses the trial’s subgroup proportions:

\widehat{R}_0 = \frac{60}{100}(0.20)+\frac{40}{100}(0.05) = 0.12+0.02 = 0.14.

Thus, the estimated 12-week control risk for a population with the trial’s baseline profile distribution is 14%. Suppose 18 of the 60 participants assigned to treatment experienced the endpoint. Their observed treatment-arm risk is $18/60=0.30$, or 30%. If randomization supports using that risk for $R_1$, the estimated risk difference is

\widehat{R}_1-\widehat{R}_0 = 0.30-0.14 = 0.16.

The estimated risk is 16 percentage points higher under treatment than under control in this hypothetical calculation. The direction and interpretation depend on how the event is defined; here, a higher risk is an unfavorable outcome. The result is causal only if the assumptions connecting the fitted external-control risks to trial participants are credible.

Calculating Individual-Prediction Averages

Subgroup averages are convenient when the model and target population can be summarized by a few covariate patterns. More generally, the model predicts each treatment-specific outcome for every target participant. Averaging those predictions implements the empirical standardization formula, including when the model contains several covariates or nonlinear terms.

Worked Example: Standardizing Predicted Risks for Five Trial Participants

Consider a hypothetical binary endpoint and five trial participants. A fitted outcome model gives the following risks under the two strategies at each participant’s baseline covariates.

Participant	Predicted risk under control, $\widehat{m}_0(X_i)$	Predicted risk under treatment, $\widehat{m}_1(X_i)$
1	0.08	0.04
2	0.10	0.06
3	0.12	0.08
4	0.20	0.14
5	0.30	0.24

The standardized control risk is the average of the five control predictions:

\widehat{R}_0 = \frac{0.08+0.10+0.12+0.20+0.30}{5} = \frac{0.80}{5} = 0.160.

The standardized treatment risk is

\widehat{R}_1 = \frac{0.04+0.06+0.08+0.14+0.24}{5} = \frac{0.56}{5} = 0.112.

Therefore, the standardized risk difference is $0.112-0.160=-0.048$, a 4.8-percentage-point lower predicted risk under treatment for this five-person target population. The risk ratio is $0.112/0.160=0.700$, so the standardized treatment risk is 0.70 times the control risk. These are different effect scales; neither should be substituted for the other without saying so.

Bayesian Standardization and Uncertainty

In a Bayesian analysis, fitted conditional means are uncertain. Standardization should therefore be performed for each posterior draw rather than only once using a single fitted mean. For draw $b$, let $m_a^{(b)}(X_i)$ be the predicted mean under strategy $a$ for target participant $i$. Calculate the standardized risks and their contrast within that draw:

R_a^{(b)} = \frac{1}{n_T}\sum_{i:S_i=1}m_a^{(b)}(X_i), \qquad \Delta^{(b)}=R_1^{(b)}-R_0^{(b)}.

The resulting draws represent posterior uncertainty in the standardized risks and effect under the specified model and assumptions. Their posterior mean or median can summarize the effect; posterior quantiles can provide a credible interval, and the fraction of draws with $\Delta^{(b)}<0$ estimates the posterior probability of a lower risk under strategy 1. These summaries are conditional on the model, prior, observed data, and target covariate distribution.

Worked Example: Standardizing Posterior Draws

Suppose the hypothetical trial target population is 70% low risk and 30% high risk. For one posterior draw, the fitted control risks are 0.12 for low-risk participants and 0.30 for high-risk participants; the treatment risks are 0.08 and 0.20, respectively. Standardize each strategy to the trial proportions:

R_0^{(b)}=0.70(0.12)+0.30(0.30)=0.084+0.090=0.174, $$ $$ R_1^{(b)}=0.70(0.08)+0.30(0.20)=0.056+0.060=0.116, $$ $$ \Delta^{(b)}=0.116-0.174=-0.058.

For this draw, the standardized treatment risk is 11.6%, the control risk is 17.4%, and the risk difference is minus 5.8 percentage points. Repeat the same calculation for every posterior draw, using the same target covariate distribution each time. The distribution of the resulting $\Delta^{(b)}$ values carries model uncertainty into the effect summary; this single draw alone does not provide a credible interval or quantify that uncertainty.

What Standardization Does—and Does Not—Change

Standardization changes the population over which model predictions are averaged. It does not repair a poorly defined endpoint, remove unmeasured confounding, or guarantee that a model learned from external controls applies to trial participants. It also does not make the external controls equivalent to randomized controls by construction. Those are substantive and statistical assumptions, not consequences of the averaging step.

Approach	How the target population enters	Main consideration
Standardization by outcome regression	Predict each strategy’s outcome for target covariates, then average.	Depends on adequate conditional outcome models and predictions relevant to the target.
Weighting	Reweights observed participants so their covariate distribution represents a target.	Depends on the weighting model and can be unstable when weights are extreme.
Matching	Forms matched comparisons among participants with similar covariates.	May change the analyzed population if some target participants cannot be matched.
Crude comparison	Compares observed group averages without covariate standardization.	Generally targets a mixture of group-specific populations and can retain confounding.

These approaches are alternatives for adjusting or defining a comparison, not interchangeable calculations. Standardization is especially direct when the scientific target is the trial population and fitted outcome models provide conditional predictions there. A crude external-control rate is not a substitute: it averages over the external cohort’s covariate distribution. Weighting or matching may be preferable in some settings, but they too require assumptions and careful attention to the population they actually represent.

Diagnostics, Limitations, and a Common Mistake

Before interpreting a standardized estimate, check that the model’s predictions are plausible for the target participants and that the calculation uses the intended trial population and follow-up horizon. For binary outcomes, predicted risks should lie between 0 and 1. Compare predicted risks across important covariate patterns with observed outcomes in the data source used to fit the model, while recognizing that such comparisons cannot establish transportability by themselves. Sparse representation of some target profiles, extreme predictions, or large changes under reasonable model specifications are warnings that the estimate may rely on extrapolation. The estimate can be numerically precise while remaining sensitive to those assumptions.

A further limitation is the choice of target distribution. Averaging over all enrolled trial participants estimates an effect for that trial population as represented in the analysis. If the scientific question instead concerns a subgroup or another population, the averaging distribution must change and the resulting estimand must be stated. Uncertainty in estimated target covariate proportions may also matter when the target is viewed as a broader population rather than the fixed set of trial participants.

Common mistake: “The external-control cohort had a 10% event rate, so the estimated control risk for the trial population is 10%.” This is wrong when the external cohort and trial have different covariate distributions. The external rate averages over the external cohort’s covariates. To estimate trial-population control risk, predict control outcomes at the trial participants’ covariates and average those predictions over the trial population.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A trial population is 80% low risk and 20% high risk. Fitted control risks are 0.10 and 0.25, respectively. What is the standardized control risk?
Using the same target proportions, fitted treatment risks are 0.06 for low-risk and 0.20 for high-risk participants. Calculate the standardized treatment risk and risk difference.
Why is the external cohort’s crude event rate generally not the standardized control risk for the trial population?
In a Bayesian analysis, why should standardization be repeated for each posterior draw?

Solutions. (1) $0.80(0.10)+0.20(0.25)=0.08+0.05=0.130$, or 13.0%. (2) The treatment risk is $0.80(0.06)+0.20(0.20)=0.048+0.040=0.088$, or 8.8%. The risk difference is $0.088-0.130=-0.042$, a 4.2-percentage-point lower risk under treatment. (3) The crude external rate averages outcomes over the external cohort’s covariate distribution, which may differ from the target trial distribution. (4) Each draw has different model parameters and therefore different conditional predictions; repeating the averaging carries posterior uncertainty in those predictions into the standardized risks and their contrast.