Inverse Odds Weighting for Population Transport

Why Weight a Trial to a Target Population?

A source trial and a target population can differ in their distributions of treatment effect modifiers. If the trial includes proportionally more people with one covariate profile than the target population does, an unweighted trial estimate may not represent the target. Inverse odds weighting reweights trial participants so their measured covariates represent the target distribution. The method complements standardization: standardization averages conditional outcome estimates over the target covariate distribution, while inverse odds weighting changes the contribution of observed trial participants.

Let $S=1$ indicate membership in the source trial and $S=0$ membership in the target population. Let $X$ denote measured baseline covariates, $A$ treatment assignment, and $Y$ the outcome. Write $Y^a$ for the potential outcome under treatment $a$. The target estimand for treatment $a$ is $\mu_{a,0}=E(Y^a\mid S=0)$. This tutorial focuses on using source-trial outcomes and covariates to estimate that target mean.

Definition: The inverse odds of selection weight for a source-trial participant with covariates $X=x$ is proportional to the odds of belonging to the target rather than the source, conditional on those covariates: $$ w(x)=\frac{P(S=0\mid X=x)}{P(S=1\mid X=x)}. $$ A source participant whose covariates are more common in the target than in the source receives a larger weight. A participant whose covariates are less common in the target receives a smaller weight.

Deriving the Weight

Let $f_0(x)$ and $f_1(x)$ be the covariate densities in the target and source populations, respectively. Bayes' rule gives

\frac{P(S=0\mid X=x)}{P(S=1\mid X=x)} = \frac{f_0(x)}{f_1(x)} \frac{P(S=0)}{P(S=1)}.

Thus the inverse odds weight is the target-to-source covariate density ratio multiplied by a constant that depends on the relative population sizes. The constant does not affect a ratio estimator that normalizes the weights within the treatment arm. If estimating a population total rather than a mean, the population-size factor can matter.

In practice, a selection model estimates $g(x)=P(S=1\mid X=x)$. For example, a logistic model can be fit to a combined data set containing source-trial participants and target-population members, with source membership as the outcome. The fitted inverse odds weight is $(1-\hat g(x))/\hat g(x)$. The model and sampling scheme must be appropriate: if the combined data set has deliberately chosen source and target sample sizes, those sample sizes affect the model intercept. For a normalized mean, a common multiplicative constant cancels, but the model must still represent the relevant covariate-density ratio.

Worked Example: Deriving Weights From Covariate Distributions

Suppose a hypothetical covariate $X$ has three categories. In the source trial, their proportions are $0.50,\ 0.30,\ 0.20$; in the target population, they are $0.20,\ 0.30,\ 0.50$. The target-to-source density ratios are therefore $0.20/0.50=0.40$, $0.30/0.30=1.00$, and $0.50/0.20=2.50$.

Assign source participants in the three categories weights $0.40,\ 1.00,\ 2.50$. The weighted source proportions, before normalization, are $0.50(0.40)=0.20$, $0.30(1.00)=0.30$, and $0.20(2.50)=0.50$. They sum to 1.00, so the weighted covariate distribution matches the target proportions in this example.

Suppose the source-trial mean outcomes under treatment $a$, within the three categories, are 2, 4, and 8 outcome units. If treatment assignment is independent of $X$ in the trial, the transported mean is $$ 0.20(2)+0.30(4)+0.50(8)=0.40+1.20+4.00=5.60 $$ outcome units. Equivalently, using source proportions and weights, the numerator is $0.50(0.40)(2)+0.30(1.00)(4)+0.20(2.50)(8)=5.60$, and the denominator is $0.20+0.30+0.50=1.00$.

Assumptions and the Treatment-Arm Estimator

Selection weighting requires more than a fitted selection model. The covariates used for transport must capture the relevant differences between source and target populations. In particular, conditional on $X$, the potential outcome under each treatment must be comparable between source and target: $E(Y^a\mid X,S=1)=E(Y^a\mid X,S=0)$. This is often called conditional transportability or conditional exchangeability for selection. It does not follow merely from matching the observed covariate distributions.

Consistency: a person's observed outcome under the received treatment equals that person's corresponding potential outcome.
Treatment exchangeability in the source: treatment assignment is randomized, or is conditionally independent of the potential outcomes given measured $X$. Positivity requires a nonzero probability of each treatment being used for relevant covariate profiles.
Selection positivity: covariate profiles occurring in the target must also have support in the source, so the target-to-source density ratio is defined and estimable.
Correct covariate and outcome definitions: the same relevant variables and outcome meaning must be available in source and target, with suitable measurement and follow-up.

For a source-trial participant $i$, let $e_a(X_i)=P(A_i=a\mid X_i,S_i=1)$, the probability of assignment to treatment $a$. When assignment probabilities vary with $X$, weighting only by selection odds does not generally make the participants assigned to $a$ represent the source covariate distribution. A treatment-arm estimator that accounts for both selection and assignment is

\hat\mu_{a,0} = \frac{\sum_{i:S_i=1} I(A_i=a)\,w(X_i)Y_i/e_a(X_i)} {\sum_{i:S_i=1} I(A_i=a)\,w(X_i)/e_a(X_i)}.

Here $I(A_i=a)$ is 1 when participant $i$ received $a$, and 0 otherwise. If assignment probability is constant across $X$ within the treatment arm, $1/e_a(X_i)$ is constant among those participants and cancels from the ratio. In that special case, selection weights alone give the same normalized estimator. Alternatively, estimate conditional outcome means and standardize them over the target covariate distribution; this avoids relying on the selection-weighted treatment-arm ratio when assignment probabilities vary.

Worked Example: Why Assignment Probabilities May Need Their Own Weights

Consider a hypothetical trial with three equally frequent source covariate categories. Their target-to-source ratios are $0.50,\ 1.00,\ 1.50$, and their treatment probabilities $e_1(X)$ are $0.20,\ 0.80,\ 0.50$. Suppose the potential outcome $Y^1$ is 0 in the first two categories and 1 in the third. The target proportions are $0.50/3,\ 1/3,\ 1.50/3$, so the target mean under treatment 1 is $0.50$.

Using only selection weights among participants assigned treatment 1 gives, in the large-sample limit, numerator $(1/3)(1.50)(0.50)(1)=0.25$. Its denominator is $$ \frac{1}{3}\{(0.50)(0.20)+(1.00)(0.80)+(1.50)(0.50)\} =\frac{1.65}{3}=0.55. $$ The ratio is $0.25/0.55=0.4545$, not the target mean of $0.50$. The reason is that the treatment arm overrepresents covariate categories with higher treatment-assignment probabilities.

Now include inverse treatment probabilities. The numerator becomes $$ \frac{1}{3}(1.50)(0.50)\frac{1}{0.50}=0.50, $$ since the other two categories have outcome zero. The denominator is $$ \frac{1}{3}\left\{(0.50)\frac{0.20}{0.20} +(1.00)\frac{0.80}{0.80} +(1.50)\frac{0.50}{0.50}\right\}=1.00. $$ The corrected ratio is $0.50/1.00=0.50$, the target mean. This calculation illustrates why selection weights alone require an additional condition—constant assignment probability across $X$—or a different method such as treatment-probability weighting or outcome standardization.

Normalization and Effective Sample Size

For a weighted mean, dividing all weights by the same positive constant leaves the estimate unchanged. Weights can therefore be normalized to sum to 1, or rescaled to have mean 1, to make their scale easier to interpret. Normalization does not remove the influence of unusually large relative weights. The effective sample size (ESS) summarizes how concentrated a set of weights is; for normalized weights $\tilde w_i=w_i/\sum_j w_j$, define

ESS=\frac{1}{\sum_i \tilde w_i^2} =\frac{(\sum_i w_i)^2}{\sum_i w_i^2}.

Equal weights give an ESS equal to the number of participants. Unequal weights reduce it. For a treatment-specific estimate, calculate ESS using the final weights for participants in that treatment arm, including any inverse treatment-probability factor. ESS is a useful summary of weight concentration, not a replacement for a variance estimate or a guarantee of accuracy.

Worked Example: Normalizing Weights and Calculating ESS

Suppose five hypothetical source participants have selection weights $1,\ 1,\ 2,\ 2,\ 4$, and outcomes $5,\ 7,\ 6,\ 8,\ 10$ outcome units. The weights sum to 10, so the normalized weights are $0.10,\ 0.10,\ 0.20,\ 0.20,\ 0.40$. The weighted mean is $$ \frac{1(5)+1(7)+2(6)+2(8)+4(10)}{10} =\frac{80}{10}=8.00 $$ outcome units.

The squared weights sum to $1+1+4+4+16=26$. Thus $$ ESS=\frac{10^2}{26}=\frac{100}{26}=3.846. $$ Although five participants contributed, this weight distribution has the same ESS as approximately 3.85 equally weighted participants. The largest weight, 4, contributes substantially to the estimate; the ESS signals that concentration without indicating whether the estimate is biased.

Diagnostics, Alternatives, and Limitations

Before interpreting a transported estimate, inspect the estimated selection probabilities and the resulting weights. Probabilities near zero for source membership imply very large inverse odds weights. These may indicate weak overlap, sparse covariate regions, or an overly flexible or poorly supported selection model. Check whether a small number of participants dominate the weighted estimate, report the ESS, and examine sensitivity to reasonable modeling choices. Covariate balance between the weighted source and target is an important diagnostic, but balance on measured covariates alone cannot establish conditional transportability or rule out unmeasured effect modifiers.

Inverse odds weighting and standardization use the same basic transport idea but rely on different modeling steps. Weighting models selection and then reweights observed outcomes; standardization models conditional outcomes and averages predictions over target covariates. A hybrid or doubly robust estimator can combine these strategies, but it introduces additional modeling choices. If outcome predictions are credible and target covariates are available, standardization can be more stable than extreme weights. Weighting can be attractive when the selection model is simpler or when transparent reweighting is useful.

Neither approach solves a lack of source-target overlap: extrapolating to covariate profiles absent from the trial requires assumptions not supplied by weighting. Weight truncation can reduce variance from extreme weights, but generally changes the weighting distribution and can introduce bias; it is not a cure for failed positivity. Likewise, a high ESS does not prove that the right covariates were measured, the selection model is correct, or the outcome relationship transports. The estimand and the populations represented by the data must remain clear throughout.

Common pitfall: “The trial randomized treatment, so the treated participants represent the source population; therefore, selection weights alone always recover the target treatment mean.” This is not generally true when randomization probabilities vary with $X$. The treatment arm can then have a covariate distribution different from the full source trial. Use inverse treatment-assignment probabilities as well, or estimate conditional outcome means and standardize over the target covariate distribution. Selection weights alone suffice for the normalized arm estimator when assignment probability is constant across $X$ within that arm.

Check Your Understanding

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

A covariate profile has target density 0.30 and source density 0.15. What is its target-to-source density ratio, and what does that imply about its relative weight?
Why can selection weights alone fail when treatment probabilities vary across covariates?
For weights $1,\ 3,\ 2$, calculate the normalized weights and ESS.
Name one condition that cannot be established merely by observing good measured-covariate balance after weighting.

Solutions. (1) The density ratio is $0.30/0.15=2.00$. This profile is twice as prevalent in the target as in the source, so it receives a relatively large weight. (2) The participants assigned to a treatment can be overrepresented in covariate strata with larger assignment probabilities; inverse assignment-probability weighting, or outcome standardization, is then needed. (3) The weights sum to 6, so the normalized weights are $1/6,\ 3/6,\ 2/6$. The sum of squared raw weights is $1+9+4=14$, giving $ESS=6^2/14=2.571$. (4) Balance cannot establish conditional transportability, including the absence of important unmeasured effect modifiers.