Inverse Odds of Sampling Weights for Transport

Why Reweight Study Participants?

A treatment effect estimated in a study need not equal the effect in a target population. The difference can arise when the populations have different distributions of covariates that modify the treatment effect. Inverse odds of study participation weights address this problem by giving more influence to study participants whose covariate profiles are more common in the target population, and less influence to profiles that are less common there.

The weights change the covariate distribution represented by the study participants; they do not, by themselves, remove confounding of treatment and outcome within an observational study. The target population and the role of the study-participation indicator must therefore be defined before constructing or interpreting the weights.

Definition: Let $S=1$ indicate membership in the study sample and $S=0$ indicate membership in the target comparison population outside the study. Let $X$ be the measured covariates used for transport, and define $e(X)=P(S=1\mid X)$ in a pooled population containing study members and target members. The inverse odds of study participation weight for a study participant with covariates $X$ is $$ w(X)=\frac{1-e(X)}{e(X)}. $$ It is the conditional odds of being in the target comparison group rather than the study group.

Here, “target comparison population” means the non-study members of the target population. This definition matters: the odds weight targets that group relative to study participants. If the target instead includes the study participants as part of the full population, the corresponding density ratio generally involves inverse participation probabilities rather than these odds.

Why the Odds Weight Reweights Covariates

Write $f_1(x)$ for the covariate density among study participants and $f_0(x)$ for the density in the target comparison population. Let $\pi_1=P(S=1)$ and $\pi_0=P(S=0)$ in the pooled population used to define participation. Bayes’ rule gives

\frac{1-e(x)}{e(x)} = \frac{P(S=0\mid X=x)}{P(S=1\mid X=x)} = \frac{\pi_0 f_0(x)}{\pi_1 f_1(x)}.

Thus the odds weight is proportional to the target-to-study covariate density ratio $f_0(x)/f_1(x)$; the proportionality constant $\pi_0/\pi_1$ does not depend on $x$. Reweighting the study distribution by this ratio recovers the target covariate distribution after normalization.

Key result: Under the population and sampling definitions above, for any integrable function $g(X)$, $$ E\{w(X)g(X)\mid S=1\} = \frac{\pi_0}{\pi_1}E\{g(X)\mid S=0\}. $$ Consequently, the normalized weighted mean among study participants equals the target mean: $$ \frac{E\{w(X)g(X)\mid S=1\}}{E\{w(X)\mid S=1\}} = E\{g(X)\mid S=0\}. $$

The result follows because multiplying $f_1(x)$ by $w(x)$ gives $(\pi_0/\pi_1)f_0(x)$. Integrating that product over $x$ yields $\pi_0/\pi_1$, since $f_0$ integrates to one. Dividing the weighted integral of $g(x)$ by this normalizing constant leaves $\int g(x)f_0(x)\,dx$, the target expectation. In data, the normalized empirical weights are usually formed by dividing each study participant’s weight by the sum of study weights. That scaling makes the weights sum to one and leaves weighted means unchanged.

Requirements for a Transported Effect

Covariate reweighting identifies a target treatment effect only under assumptions connecting the study to the target. Let $Y(a)$ denote the potential outcome under treatment level $a$, for $a\in\{0,1\}$. A common target estimand is the average treatment effect among non-study target members:

\tau_T=E\{Y(1)-Y(0)\mid S=0\}.

For weighting to transport a study effect to this estimand, the covariates $X$ must be sufficient for the relevant differences between study and target. In particular, a standard condition is conditional exchangeability across study participation:

E\{Y(a)\mid X,S=0\}=E\{Y(a)\mid X,S=1\},\qquad a\in\{0,1\}.

This says that, at the same values of $X$, the mean potential outcome under each treatment is the same in the study and target groups. It is an assumption, not a consequence of fitting a participation model. The covariates should therefore include the effect modifiers needed to make this equality plausible.

Consistency: the observed outcome under a received treatment equals the corresponding potential outcome.
Study treatment identification: within the study, treatment assignment must identify the relevant causal effect. In a randomized trial, this follows from valid randomization and appropriate outcome ascertainment. In an observational study, treatment exchangeability and treatment positivity are also needed.
Participation positivity: target covariate profiles must be represented among study participants. In particular, the target density must not assign positive mass to profiles that have zero probability of study participation.
Valid participation model and sampling: the fitted participation probabilities and the data used to estimate them must adequately represent the study and target covariate distributions.

When the study is randomized with treatment assigned independently of $X$, sampling weights can be used to form weighted outcome means separately in the two treatment arms, and their difference estimates the target average treatment effect under these conditions. Sampling weights alone do not adjust treatment confounding in an observational study. There, treatment adjustment is needed in addition to transport weighting.

Worked Example: Exact Weights for Two Covariate Groups

Suppose a hypothetical study contains two covariate groups. Among study participants, $75\%$ have $X=0$ and $25\%$ have $X=1$. In the target population, the corresponding proportions are $1/3$ and $2/3$. Suppose the study identifies treatment effects of 8 outcome units for $X=0$ and 4 outcome units for $X=1$. Assume treatment is randomized in the study and that conditional exchangeability across participation holds given $X$.

The target-to-study covariate ratios, which are inverse odds weights up to a common constant, are

w(0)=\frac{1/3}{3/4}=\frac{4}{9}, \qquad w(1)=\frac{2/3}{1/4}=\frac{8}{3}.

These weights are already normalized to have mean one in the study, because

\frac{3}{4}\frac{4}{9}+\frac{1}{4}\frac{8}{3} = \frac{1}{3}+\frac{2}{3}=1.

The weighted treatment-effect estimate is the weighted sum of the group-specific effects divided by the sum of the weighted group proportions:

\frac{ \frac{3}{4}\frac{4}{9}(8)+\frac{1}{4}\frac{8}{3}(4) }{ \frac{3}{4}\frac{4}{9}+\frac{1}{4}\frac{8}{3} } = \frac{\frac{8}{3}+\frac{8}{3}}{1} = \frac{16}{3} \approx 5.333\text{ outcome units}.

The result also follows directly from the target proportions: $(1/3)(8)+(2/3)(4)=16/3$. The study’s unweighted average would instead be $(3/4)(8)+(1/4)(4)=7$ outcome units, because it gives the study’s covariate distribution rather than the target distribution. The weighted estimate is a target effect only under the stated transport and study-identification assumptions.

Estimating and Normalizing Weights in Data

In practice, $e(X)$ is usually unknown. If covariates are available for both study participants and a sample representing the target comparison population, one can fit a model for $S$ using $X$. For example, a logistic regression may estimate $e(X_i)$, the probability that person $i$ belongs to the study given their covariates. The odds weight for a study participant is then calculated from the fitted probability. The model must be fitted and interpreted with the sampling design in mind: the pooled sample’s study-to-target ratio can affect the fitted intercept and the numerical odds, even though a common multiplicative constant cancels in normalized weighted means.

A convenient empirical normalization for the study participants is

\widetilde w_i = \frac{w_i}{\sum_{j:S_j=1}w_j}, \qquad \sum_{i:S_i=1}\widetilde w_i=1.

For a variable $g$, its normalized weighted study mean is $\sum_{i:S_i=1}\widetilde w_i g_i$. For a randomized study, a weighted treatment-arm mean can be calculated as

\widehat\mu_a^{\,T} = \frac{\sum_{i:S_i=1,A_i=a}w_iY_i} {\sum_{i:S_i=1,A_i=a}w_i}, \qquad \widehat\tau_T=\widehat\mu_1^{\,T}-\widehat\mu_0^{\,T}.

The arm-specific denominators are important: each weighted mean is normalized within its own treatment arm. Under random assignment independent of $X$, the arms represent the same study covariate distribution in expectation, and weighting each arm toward the target distribution yields the target contrast. With finite samples, the arms’ observed covariate distributions can differ by chance, so the estimate may not exactly equal a target-standardized contrast.

Worked Example: Fitting Probabilities and Renormalizing

Consider a hypothetical pooled membership sample in which the estimated study probabilities are $e(0)=0.75$ and $e(1)=0.25$. In the study sample, $60\%$ have $X=0$ and $40\%$ have $X=1$. The raw odds weights are

w(0)=\frac{1-0.75}{0.75}=\frac{1}{3}, \qquad w(1)=\frac{1-0.25}{0.25}=3.

Their mean in the study sample is $0.60(1/3)+0.40(3)=0.20+1.20=1.40$. Dividing both weights by $1.40=7/5$ gives normalized weights

\widetilde w(0)=\frac{1/3}{7/5}=\frac{5}{21}, \qquad \widetilde w(1)=\frac{3}{7/5}=\frac{15}{7}.

The normalized weighted mass assigned to $X=0$ is $0.60(5/21)=1/7$. For $X=1$, it is $0.40(15/7)=6/7$. These sum to one, so the implied weighted covariate proportions are $1/7$ and $6/7$. The raw weights and normalized weights give the same weighted means; normalization changes the overall scale, not the relative influence of the two groups.

Worked Example: A Weighted Treatment Contrast

In a second hypothetical randomized study, $75\%$ of participants have $X=0$ and $25\%$ have $X=1$. The target proportions are $1/3$ and $2/3$, so use weights $4/9$ and $8/3$, respectively. Suppose the observed treatment-arm outcome means within each covariate group are as follows. Assume these means identify the corresponding group-specific causal means, and assume conditional exchangeability across study participation given $X$.

Covariate group	Study proportion	Weight	Mean if treated	Mean if control
$X=0$	$3/4$	$4/9$	10	7
$X=1$	$1/4$	$8/3$	15	11

The weighted treated mean is

\frac{ \frac{3}{4}\frac{4}{9}(10)+\frac{1}{4}\frac{8}{3}(15) }{ \frac{3}{4}\frac{4}{9}+\frac{1}{4}\frac{8}{3} } = \frac{\frac{10}{3}+10}{1} = \frac{40}{3} \approx 13.33\text{ outcome units}.

The weighted control mean is

\frac{ \frac{3}{4}\frac{4}{9}(7)+\frac{1}{4}\frac{8}{3}(11) }{ \frac{3}{4}\frac{4}{9}+\frac{1}{4}\frac{8}{3} } = \frac{\frac{7}{3}+\frac{22}{3}}{1} = \frac{29}{3} \approx 9.667\text{ outcome units}.

The transported treatment contrast is therefore $40/3-29/3=11/3\approx3.667$ outcome units. The same answer follows by weighting the within-group effects: $1/3(10-7)+2/3(15-11)=1+8/3=11/3$. The interpretation is an estimated average treatment effect for the target population, conditional on the randomization and transport assumptions—not simply a descriptive difference that automatically applies outside the study.

Common Pitfalls and Interpretation

Using participation probabilities instead of odds. The inverse odds weight is $(1-e)/e$, not $1/e$. These weights correspond to different target definitions and generally induce different covariate distributions.
Assuming normalization creates identification. Dividing weights by their sum is a scale convention. It cannot compensate for missing effect modifiers, invalid treatment comparisons, or lack of study representation for target profiles.
Calling the weighted estimate automatically causal. Sampling weights address differences in covariate distributions between study and target. They do not, on their own, eliminate confounding within an observational study.
Ignoring the model’s sampling context. The participation probability is defined relative to the pooled study and target data. If those samples are assembled with different sampling fractions, a naïvely fitted membership model may not represent population membership probabilities without accounting for the sampling scheme.
Reading a normalized weight as a probability. A normalized weight is a participant’s share of total weighted mass. It is not that participant’s probability of study participation.

The central interpretation is precise: inverse odds weights make the measured covariate distribution of study participants resemble that of the target comparison population. Transporting a causal effect additionally requires that the covariates suffice to bridge outcome differences across study participation and that the study identifies its own causal effect.

Check Your Understanding

Answer each question using the target and study definitions in the tutorial.

If $e(X)=0.80$, what is the inverse odds of study participation weight, and what does it compare?
Explain why multiplying the study covariate density by the odds weight produces a quantity proportional to the target covariate density.
Why does dividing every weight by the same positive constant leave a normalized weighted mean unchanged?
Name the cross-population assumption that allows conditional study effects to inform the target effect.
A study is observational and has valid sampling weights, but treatment is confounded within the study. Are the sampling weights alone sufficient for a causal target effect? Explain.