From Conditional Risks to a Population Contrast
A treatment model may report how event risk varies with treatment and baseline covariates. But a conditional comparison—say, among patients with a particular covariate profile—is not necessarily the population-level contrast of interest. Standardization converts covariate-adjusted predictions into risks for a stated target population by averaging them over that population’s covariate distribution.
In a competing-risks setting, the outcome must be the cumulative incidence of the event of interest, with competing events handled as competing outcomes rather than treated as ordinary censoring. The result is a total effect: it compares the event risk under each treatment strategy, allowing the treatment to affect both the event of interest and competing events.
The target distribution matters. It might be the covariate distribution of the study participants, a defined clinical population, or another population for which the needed covariate data and treatment support are available. A standardized risk is therefore not a universal property of a treatment: it is a risk for a specified population and time horizon.
Identification and Estimation
A standardized contrast is causal only if the observed data identify the counterfactual risks. For the baseline-treatment setting considered here, the core conditions are consistency (the observed outcome under received treatment agrees with that treatment’s counterfactual outcome), conditional exchangeability (no unmeasured confounding of treatment and potential outcomes given \(L\)), and positivity (each treatment being compared has positive probability at covariate profiles represented in the target population). Treatment versions must also be sufficiently well defined for consistency to be meaningful.
Censoring must not selectively hide event outcomes after conditioning on the variables used to handle it. In practice, the method used to estimate conditional cumulative incidence must correctly account for censoring under the censoring assumptions. In addition, the outcome model must estimate the conditional risks adequately, or the estimator must otherwise consistently recover them. These requirements are distinct: a well-fitting outcome model cannot repair unmeasured confounding, and causal assumptions cannot repair a badly misspecified model.
The justification is the law of total probability applied to the counterfactual event indicator. First partition the target population by \(L\); then, under exchangeability and consistency, replace the counterfactual conditional risk in each covariate stratum with the observed risk among people receiving \(a\) in that stratum:
The middle equality uses conditional exchangeability and positivity; the last uses consistency. Positivity is needed so the conditional observed risk is defined and estimable at the target covariate values. This argument identifies the estimand; it does not prescribe a particular model or prove that a chosen model fits.
A practical estimator fits an outcome model that can provide conditional cumulative-incidence predictions at the chosen horizon. For each target-population member \(i\), predict the risk under each treatment, keeping that person’s covariates fixed, and average the predictions:
Here the \(n\) covariate profiles are those of the chosen empirical target population. If the target is external, its covariate distribution must instead be supplied or estimated. Uncertainty intervals should account for estimating the outcome model and the standardized risks—for example, by a suitable bootstrap that refits the model in each resample.
Worked Standardizations
Worked Example: Standardizing Over Two Covariate Strata
Suppose a hypothetical cohort’s target population is 60% lower-risk and 40% higher-risk at baseline. A correctly specified model, under the identification and censoring conditions stated above, predicts the following cumulative incidences of the event of interest by 2 years. Assume complete ascertainment through 2 years for this illustration.
| Baseline stratum | Target proportion | Risk under \(A=0\) | Risk under \(A=1\) |
|---|---|---|---|
| Lower-risk | 0.60 | 0.10 | 0.06 |
| Higher-risk | 0.40 | 0.30 | 0.18 |
Standardize the risk under \(A=0\): $$ \widehat F_{k,\mathrm{std}}^0(2) =(0.60)(0.10)+(0.40)(0.30) =0.06+0.12=0.18. $$
Under \(A=1\): $$ \widehat F_{k,\mathrm{std}}^1(2) =(0.60)(0.06)+(0.40)(0.18) =0.036+0.072=0.108. $$
The standardized risk difference is \(0.108-0.18=-0.072\), or 7.2 percentage points lower under \(A=1\) by 2 years. The risk ratio is \(0.108/0.18=0.600\). These are estimated causal contrasts only if the identification assumptions hold; the calculations alone do not establish them.
Worked Example: Standardizing Predictions for Individual Covariate Profiles
Consider a hypothetical target population of four people with covariate profiles \(L_1,\ldots,L_4\). A fitted model provides conditional 1-year cumulative-incidence predictions for event \(k\). Assume the target population is exactly these four profiles and that the model and causal assumptions are adequate.
| Profile | Predicted risk under \(A=0\) | Predicted risk under \(A=1\) |
|---|---|---|
| \(L_1\) | 0.05 | 0.04 |
| \(L_2\) | 0.10 | 0.08 |
| \(L_3\) | 0.20 | 0.16 |
| \(L_4\) | 0.25 | 0.20 |
Give each profile weight \(1/4\). The standardized risks are $$ \widehat F_{k,\mathrm{std}}^0(1) =\frac{0.05+0.10+0.20+0.25}{4} =\frac{0.60}{4}=0.150, $$ and $$ \widehat F_{k,\mathrm{std}}^1(1) =\frac{0.04+0.08+0.16+0.20}{4} =\frac{0.48}{4}=0.120. $$
Thus, the estimated risk difference is \(0.120-0.150=-0.030\), a 3.0-percentage-point reduction by 1 year in this target population. The risk ratio is \(0.120/0.150=0.800\). These predictions are averaged over the same four profiles under both treatment assignments; that is what makes the comparison standardized to one common population.
Worked Example: Competing Events Remain Part of the Risk
In a hypothetical population, half the people are in each of two baseline strata. A model estimates the event-\(k\) cumulative incidence by 3 years as follows. Each prediction accounts for the possibility that a competing event occurs first. Assume the model is adequate and the causal identification and censoring conditions hold.
| Stratum | Target proportion | Risk under \(A=0\) | Risk under \(A=1\) |
|---|---|---|---|
| Stratum 1 | 0.50 | 0.12 | 0.09 |
| Stratum 2 | 0.50 | 0.24 | 0.16 |
For \(A=0\), the standardized event-\(k\) risk is $$ (0.50)(0.12)+(0.50)(0.24)=0.06+0.12=0.18. $$ For \(A=1\), it is $$ (0.50)(0.09)+(0.50)(0.16)=0.045+0.080=0.125. $$
The estimated 3-year risk difference is \(0.125-0.18=-0.055\), a 5.5-percentage-point lower event-\(k\) risk under \(A=1\). This is a total effect on the event-\(k\) cumulative incidence. If treatment also changes the competing-event incidence, that change can contribute to the event-\(k\) risk contrast. The estimand does not describe a hypothetical world in which competing events are removed.
Interpretation, Checks, and Alternatives
A standardized risk is a marginal quantity for the selected target population. A regression coefficient is usually conditional on the model’s covariates and may describe a different scale, such as a log odds or a hazard. It should not be substituted for the standardized risk difference or risk ratio. Likewise, a conditional treatment association within covariate strata does not automatically equal the population contrast; averaging predicted risks is the step that makes the target distribution explicit.
Before interpreting a result, check whether the prediction procedure actually estimates cumulative incidence for event \(k\). A cause-specific hazard describes an instantaneous event rate among people still event-free. It is not itself the probability of event \(k\) by time \(t\). Treating competing events as ordinary censoring and then interpreting the resulting estimate as the observed-world cumulative incidence generally answers a different question. A competing-risk model must account for the competing event when constructing the predicted risk.
Useful diagnostics include comparing predicted risks with observed outcomes across covariate and treatment groups, checking calibration at the chosen horizon, and examining whether predictions are plausible (between 0 and 1, with sensible time behavior). Inspect treatment overlap across covariate profiles in the target population. Sparse or absent treatment groups in some profiles mean that the model is extrapolating; a precise-looking average does not make unsupported predictions reliable. Also assess whether censoring is substantial or differs across groups, and whether the method’s censoring assumptions are credible.
Standardization is one way to obtain a population contrast from an outcome model. Crude observed risks are simpler but can be confounded when treatment groups have different covariate distributions. Conditional regression summaries preserve covariate-specific information but may not directly answer the target-population question. Weighting approaches provide another route to marginal contrasts, but rely on their own modeling and positivity requirements. Regardless of approach, the estimand and target population should be held fixed when comparing results.
Standardization can be sensitive to outcome-model misspecification and to the chosen target distribution. It does not eliminate unmeasured confounding, guarantee overlap, or make censoring assumptions unnecessary. It also does not identify effects in covariate profiles absent from the data without extrapolation. Report the event type, time horizon, target population, treatment strategies, risk scale, and uncertainty interval so the result has a clear interpretation.
Check Your Understanding
Answer each question, then compare with the short solutions below.
- A target population is 70% group 1 and 30% group 2. Predicted risks under treatment 0 are 0.08 and 0.28; under treatment 1 they are 0.06 and 0.18. Calculate both standardized risks and the risk difference.
- Why must the same target covariate distribution be used for both treatment-specific risks in a standardized contrast?
- Does a cause-specific hazard ratio directly give the risk ratio for event \(k\) by a fixed time? Explain.
- Name two assumptions needed for a standardized contrast to have a causal interpretation.
- If one target-population covariate stratum has no treated participants, what is the concern?
Solutions. (1) Risk under treatment 0: \((0.70)(0.08)+(0.30)(0.28)=0.056+0.084=0.140\). Risk under treatment 1: \((0.70)(0.06)+(0.30)(0.18)=0.042+0.054=0.096\). The risk difference is \(0.096-0.140=-0.044\), or 4.4 percentage points lower under treatment 1. (2) A common distribution ensures the contrast compares treatments in the same target population rather than mixing treatment effects with differences in population composition. (3) No. A hazard ratio compares instantaneous rates among those still event-free; it is not a cumulative-incidence risk ratio, which also reflects competing events over time. (4) Examples include consistency, conditional exchangeability, positivity, and appropriate censoring conditions. (5) Positivity may fail; the treated risk for that stratum cannot be learned directly from observed data, so estimation requires unsupported extrapolation.