Standardizing Risk To A Target Population

From Individual Predictions to a Population Risk

A predicted risk for one covariate profile answers a conditional question: what is the probability of a particular cause of failure by a specified time for someone with those covariates? A population-average prediction answers a different question: what is that risk, on average, across the covariate patterns in a defined target population? The distinction matters whenever people in that population have different predicted risks.

Standardization makes the population being described explicit. First obtain a predicted competing-risk probability for each target-population member at the same time horizon. Then average those predictions using weights that represent the target population. The resulting estimate is an absolute risk for that population and horizon, not a prediction for a single “average” person.

Definition: Let $T$ be time to the first event, $J$ its cause, and $k$ one specified cause among $K$ competing causes. For covariates $Z$, the cause-specific cumulative incidence by horizon $\tau$ is $F_k(\tau\mid Z)=P(T\leq \tau,\ J=k\mid Z)$. For a target population with covariate distribution $G_T$, its population-average risk is $$ R_k^T(\tau)=E_T\{F_k(\tau\mid Z)\}. $$ It is the average probability of experiencing cause $k$ by $\tau$, averaging over the covariate distribution of the target population.

In practice, a model fitted to source data provides predictions for the target members. If the target consists of $M$ records with covariates $Z_1,\ldots,Z_M$, and record $i$ represents $a_i\geq 0$ members of the target population, define the normalized weight $\omega_i=a_i/\sum_{j=1}^M a_j$. The standardized estimate is

\widehat R_k^T(\tau) =\sum_{i=1}^{M}\omega_i\widehat F_k(\tau\mid Z_i), \qquad \sum_{i=1}^{M}\omega_i=1.

When each record represents one target member, $a_i=1$, so this is the ordinary arithmetic mean of the $M$ predicted risks. With sampling or post-stratification weights, the weighted mean instead reflects the specified target composition. The horizon $\tau$, event definition, and cause $k$ must be common across all predictions. A prediction at one year cannot be averaged with a prediction at five years and interpreted as a single-horizon risk.

Why Averaging Gives the Target Risk

The standardization identity follows by conditioning on covariates. A member of the target population has some covariate value $Z$. The probability of cause $k$ by $\tau$, averaged over that population, is the average of the conditional probabilities for the covariate values that occur there. In expectation notation this is precisely $E_T\{F_k(\tau\mid Z)\}$. For a finite target population, the expectation is the weighted sum over its members.

This argument also explains the role of the target distribution: the conditional risk function comes from the prediction model, while the distribution over which it is averaged comes from the target records and their weights. Changing the target covariate distribution can change the standardized risk even if the fitted prediction model is unchanged.

The method assumes that the individual predictions estimate the relevant conditional cumulative incidences. The risk model must be suitable for the data and event process, and its estimation must address censoring under the conditions required by that model. If predictions are transported from a source population to a different target population, the conditional risk relationships must apply in the target after accounting for the covariates used, and the target records must represent the target covariate distribution. Standardization does not repair a misspecified model, unmeasured differences that alter risk, or target covariate patterns for which the model has no credible support.

Worked Example: Equal-Weight Target Population

Suppose a fitted model gives hypothetical two-year predictions for four members of a target population. The cause of interest is a first cardiovascular event, with other causes of failure treated as competing events. The predicted risks are 0.10, 0.20, 0.05, and 0.15. Each record represents one member, so equal weights are appropriate for this finite target.

The standardized risk is the mean of the four predictions:

\widehat R_k^T(2) =\frac{0.10+0.20+0.05+0.15}{4} =\frac{0.50}{4} =0.125.

Thus the estimated population-average risk of the specified cause by two years is 12.5%. This describes the average predicted risk across these four target members; it does not mean each member has a 12.5% predicted risk. The calculation assumes all four predictions refer to the same two-year horizon and the same cause definition.

Weights and the Target Composition

Weights are not merely a computational adjustment. They encode how target members contribute to the population average. A record with a larger weight contributes more because it represents more members or is assigned more influence by the chosen target-population design. Weights should therefore be chosen to represent the target population, not selected to make the estimate agree with a preferred answer.

Worked Example: Unequal Representation

Consider four hypothetical target records with predicted three-year risks of 0.04, 0.10, 0.16, and 0.22. Suppose their target-population weights are 1, 1, 2, and 2, respectively. These weights represent six target members in total: the third and fourth records each represent two members.

The normalized weights are $1/6,1/6,2/6,2/6$. Therefore,

\widehat R_k^T(3) =\frac{1(0.04)+1(0.10)+2(0.16)+2(0.22)}{1+1+2+2} =\frac{0.04+0.10+0.32+0.44}{6} =\frac{0.90}{6} =0.150.

The estimated three-year target-population risk is 15.0%. The unweighted mean would be $(0.04+0.10+0.16+0.22)/4=0.130$, which describes an equal-weight average of the four records, not the population composition represented by the stated weights. The weighted calculation is appropriate here because the weights have an explicit target-population interpretation.

For predictions with multiple competing causes, standardize each cause-specific risk using the same target records and weights. This keeps the cause-specific estimates aligned to one population. Because the weights are nonnegative and sum to one, averaging also preserves the probability constraints satisfied by the individual predictions.

Worked Example: Standardizing Risks for Three Causes

Suppose three hypothetical target members have five-year predicted risks for three mutually exclusive causes as shown below. The fourth outcome is being event-free at five years. Each member has equal weight.

Member	Cause 1	Cause 2	Cause 3	Sum of cause risks
A	0.10	0.05	0.02	0.17
B	0.20	0.10	0.04	0.34
C	0.30	0.10	0.06	0.46

The standardized estimates are computed cause by cause:

\widehat R_1^T(5)=\frac{0.10+0.20+0.30}{3}=0.200, \qquad \widehat R_2^T(5)=\frac{0.05+0.10+0.10}{3}=0.0833, $$ $$ \widehat R_3^T(5)=\frac{0.02+0.04+0.06}{3}=0.0400.

The standardized probability of failure from any of the three causes is the sum, $0.200+0.0833+0.0400=0.3233$, or 32.33%. Equivalently, the mean of the member-specific sums is $(0.17+0.34+0.46)/3=0.3233$. The corresponding average event-free probability is $1-0.3233=0.6767$, or 67.67%. These calculations are coherent because all cause-specific predictions refer to the same members, horizon, and competing-event definition.

Checks, Comparisons, and Limitations

Before reporting a standardized risk, check that the target records and weights correspond to the population named in the interpretation, that every prediction uses the same horizon, and that the event and competing-cause definitions agree across model outputs. Confirm that each predicted cause-specific risk is between zero and one and that the cause risks for an individual do not sum to more than one. If the individual predictions violate these probability constraints, averaging them will not make the underlying predictions valid.

Also inspect whether target covariate values are represented in the data used to fit the model. Predictions for covariate patterns far outside the fitted data range depend heavily on extrapolation. A numerical average can still be calculated in that situation, but it may not be reliable. Missing target covariates, incorrect weights, or a mismatch between the source and target risk relationships can likewise undermine the result.

Quantity	What it summarizes	Key distinction
Profile-specific prediction	Risk for one specified covariate profile at a stated horizon	Conditional on that profile; not generally a population average
Standardized prediction	Weighted average of individual predictions over a defined target population	Depends on both the prediction model and the target covariate distribution
Crude observed risk estimate	Observed event experience in a data set, subject to its follow-up and censoring	Does not automatically estimate the model-based risk for a separately defined target population

Different modeling approaches may supply the conditional predictions being averaged; standardization is the population-averaging step, not a replacement for choosing or checking the prediction model. Nor does the point estimate alone communicate its precision. Inference for standardized risks requires accounting for the uncertainty in model fitting and, when applicable, in the target weights; that uncertainty is not represented by the arithmetic average itself.

A Common Mistake

A common error is to report the prediction for one selected covariate profile as if it were the target-population average. The mistaken reasoning is that the profile seems typical, so its predicted risk should represent everyone. But a profile-specific prediction is conditional on that profile; it averages over no target covariate distribution. The correction is to predict for the target members and average those predictions with the weights that define the target. This produces the stated population average without implying that every member has the same risk.

Key takeaway: A standardized competing-risk estimate is the weighted average of cause-specific predicted cumulative incidences for target-population members at one common horizon. Its interpretation is determined jointly by the cause, horizon, prediction model, target records, and weights.

Check Your Understanding

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

Three equally weighted target members have predicted cause-specific risks of 0.08, 0.14, and 0.20 by four years. What is the standardized risk?
Four records have predictions 0.05, 0.10, 0.15, and 0.20 and weights 1, 1, 1, and 3. Calculate the weighted standardized risk.
Why must predictions being averaged refer to the same time horizon and cause definition?
For each target member, predicted risks across all competing causes sum to no more than one. Explain why averaging each cause with the same nonnegative weights summing to one preserves this constraint.

Solutions. 1. The mean is $(0.08+0.14+0.20)/3=0.14$, so the standardized four-year risk is 14%. 2. The weighted estimate is $\{1(0.05)+1(0.10)+1(0.15)+3(0.20)\}/6=0.90/6=0.15$, or 15%. 3. Otherwise the average combines different probability events, so it does not describe one cause by one common horizon. 4. For each member, the sum across causes is at most one. Multiplying by nonnegative weights and summing gives a weighted total at most $\sum_i\omega_i=1$, so the standardized cause risks also sum to no more than one.