Inverse Probability Weighting for External Controls

Why Reweight External Controls?

External controls may differ from trial participants in baseline characteristics that predict outcomes. Comparing their unadjusted outcomes can therefore mix differences in patient composition with differences in outcomes under the control strategy. Inverse probability weighting addresses this problem by assigning more influence to external controls whose measured covariate profiles resemble those of trial participants, and less influence to profiles that are relatively common in the external cohort.

The weights in this tutorial concern source membership: they estimate how the distribution of baseline covariates in the external cohort must be reweighted to represent the trial population. They are not treatment-propensity weights. The earlier material on estimating propensity scores provides the basic idea of modeling a probability from covariates; here the outcome of that model is trial versus external-cohort membership.

This density ratio is large for covariate profiles that are more prevalent in the trial than in the external cohort. Weighting by it changes the external cohort's covariate distribution toward the trial's. Stabilization multiplies every external-control weight by the same factor so that, under the population model, its mean among external controls is one.

Deriving the Weight

The pooled source-membership probability depends both on the covariate distributions and on the proportions of trial and external participants in the pooled fitting sample. Bayes' rule gives the source-membership odds as follows:

Rearranging isolates the density ratio needed to reweight external controls:

Thus the mean-one stabilized weight for external control \(i\) is \(w_i=(q_E/q_T)\hat e(X_i)/(1-\hat e(X_i))\), using the fitted source-membership score \(\hat e\). The factor is \(q_E/q_T\), not its inverse. The correction matters because the fitted odds reflect the pooled sample's source proportions as well as the covariate-density ratio.

Setting \(g(X)=1\) shows that the expected unstabilized weight among external controls is one. The stabilized expression above equals that same density ratio, so it also has expected value one. In finite samples, its observed average need not be exactly one, particularly when scores are estimated or the model is imperfect.

The derivation requires the pooled source model to represent the relevant sampling process, and the estimated score to capture the covariate differences needed for reweighting. In practice, a logistic regression or another suitable model can estimate \(e(X)\). The model should allow important nonlinearities and interactions when needed; a simple model is not automatically adequate merely because it produces scores between zero and one.

Using Weights to Estimate a Control Outcome

For a measured outcome \(Y\), a normalized weighted mean among external controls is

Under the density-ratio identity, this estimates the mean control outcome standardized to the trial's measured covariate distribution, provided the outcome is comparable across sources conditional on the adjusted covariates. The normalized form also makes the estimate invariant to multiplying every weight by the same positive constant. That invariance does not make the correct stabilization factor irrelevant: the mean-one convention remains useful for interpreting and reporting weights, and other analyses may depend on their scale.

Assumptions and Interpretation

Weighting can align measured covariate distributions, but it does not by itself establish that trial and external outcomes are exchangeable. A causal comparison using a trial treatment group and weighted external controls needs several substantive conditions:

• Consistency and comparable strategies: the control strategy and outcome have the same relevant meaning in both sources. Differences in endpoint definitions, follow-up, or treatment versions can undermine this condition.
• Conditional source exchangeability for control outcomes: given the measured covariates used for adjustment, the expected control outcome for eligible trial participants is represented by the external controls. Important unmeasured prognostic differences violate this condition.
• Positivity: covariate profiles in the target trial population must have representation in the external cohort. Formally, the external density must be positive wherever the trial density is positive, \(f_E(x)>0\) whenever \(f_T(x)>0\).
• Adequate source-score estimation: the fitted score must capture the covariate distribution differences relevant to the weighting. A misspecified model can leave important imbalance.

Under these assumptions, the weighted external mean can serve as an estimate of the control outcome mean in the trial population. A contrast with the trial's treatment-group outcome mean may then estimate a treatment effect for that population, subject to the trial's treatment-assignment design and outcome analysis. The estimate is not automatically causal merely because weights were calculated.

Extreme Weights: A Practical Tradeoff

A score near one for an external control makes \(\hat e/(1-\hat e)\) large. Such a control has a covariate profile estimated to be much more common in the trial than in the external cohort. It may be essential for representing part of the trial population, but a very large weight can also make the result depend heavily on one or a few observations. Stabilization changes the overall scale; it does not remove large relative weights.

Analysts should inspect the estimated score and weight distributions, identify which records carry the largest weights, and examine whether those records have plausible covariate values and reliable outcome measurements. They should also check whether weighted covariate summaries move toward the trial's summaries. These checks can reveal poor overlap, model problems, or data errors; they cannot prove that unmeasured confounding is absent.

Weight truncation or capping should therefore be treated as a sensitivity analysis, not a universal repair. If used, the rule should be justified and specified transparently; report the uncapped estimate and how the result changes under the chosen rule. When extreme weights indicate that important trial profiles are scarcely represented in the external cohort, no numerical cap can create the missing information. The later material on overlap weighting addresses a different target under limited comparability; it should not be silently substituted for trial-population weighting.

Alternatives and Limitations

Matching, introduced in the earlier material on matching trial participants to external controls, selects or pairs controls to improve comparability. Weighting instead retains external observations and changes their contributions to a summary. Matching can discard observations and may not represent the full trial covariate distribution; weighting can retain them but may be unstable when some profiles receive very large weights.

Outcome regression is another adjustment strategy: it models the outcome conditional on covariates rather than reweighting the control cohort. It relies on its own model assumptions and is treated separately in the later material on outcome regression. These approaches are not interchangeable by default. In particular, weighting does not repair incompatible endpoints, unmeasured confounding, source-specific measurement error, or severe lack of covariate support.

Finally, weights are estimated quantities. Treating them as fixed when calculating uncertainty can understate uncertainty, and treating weighted records as literal replicated patients can misrepresent the data. In Bayesian analyses, weighting must be incorporated in a way appropriate to the model and inferential target; a weight is not itself a likelihood contribution or a probability. The present calculations establish the adjustment construction, not a complete inferential procedure.