Abstract
Increasingly, the statistical and epidemiologic literature is focusing beyond issues of internal validity and turning its attention to questions of external validity. Here, we discuss some of the challenges of transporting a causal effect from a randomized trial to a specific target population. We present an inverse odds weighting approach that can easily operationalize transportability. We derive these weights in closed form and illustrate their use with a simple numerical example. We discuss how the conditions required for the identification of internally valid causal effects are translated to apply to the identification of externally valid causal effects. Estimating effects in target populations is an important goal, especially for policy or clinical decisions. Researchers and policy-makers should therefore consider use of statistical techniques such as inverse odds of sampling weights, which under careful assumptions can transport effect estimates from study samples to target populations.
Large randomized trials with complete compliance and no missing data provide internal validity in expectation as a matter of design (1). However, external validity with respect to a specific, investigator-defined target population is not similarly provided (2–7). Unless the study sample (PS) was sampled at random from the target population (PT), there is no expectation of exchangeability of the study sample and the (again, investigator-defined) target population. Yet nearly all trials are conducted among study samples that are not sampled at random from the target population, for reasons of either design (e.g., to maximize statistical power, a trial is conducted among those at highest risk of an outcome) or happenstance (e.g., if persons who exhibit health-seeking behaviors participate in the trial at higher frequencies than others). In these cases, despite having an internally unbiased sample average treatment effect, that sample average treatment effect may differ from the average treatment effect in the target population.
Given that we have internally valid trial results, we often wish to ask the question: What would happen had this trial been conducted in another, external population—the target? Above we suggested that we might want to ask about the causal effect of the treatment in the population from which the study population was sampled (albeit perhaps nonrandomly); we also might wish to address the causal effect of the treatment in a target population distinct from the study sample, that is, one which is partially or completely nonoverlapping with the study sample. In this latter case—such as when we have a randomized trial and wish to infer a causal effect in a target population—the question can be framed as one of direct standardization to the external target population. As a distinction of language, and to be consistent with the evolving literature on this topic, we refer to the former case (where the study sample is a subset of the target population) as a problem of “generalizability,” and to the latter case (where the study sample is not a subset of the target population) as a problem of “transportability.”
In either case, when externally valid estimates of effect are desired but not guaranteed by design, quantitative approaches are needed. In general, these approaches rely on assumptions which parallel the identification conditions necessary for internally valid causal effect estimation, particularly conditional exchangeability (8) with positivity (9), treatment variation irrelevance (10), no measurement error, and no misspecification of relevant parametric or semiparametric models. The last point is not necessary if nonparametric inference is possible, but in most cases the relevant space of covariates is high-dimensional, and thus a robust approach to quantitative generalizability or transportability requires some degree of modeling.
Previously, Cole and Stuart (4) introduced inverse probability weights for quantitative generalization of trial results, but they did not explain how to operationalize this approach or whether their approach was applicable to problems of both generalizability and transportability. More recently, Bareinboim and Pearl (3) highlighted several key distinctions between generalizability and transportability and introduced a method for deriving a transport formula, which relies on a detailed understanding of the causal relationships among all relevant variables. Here we integrate these 2 methods to introduce an approach to quantitative transportability which may be simpler to implement than the transport formula.
A brief additional note on terminology: Where Cole and Stuart refer to inverse probability of selection weights (4), we refer to inverse probability of sampling weights. We note, however, that in many (perhaps most) cases the study subjects were probably not formally sampled, and we do not wish to imply so with the use of that term; rather, we simply obtain a study sample through some (perhaps unclear) mechanism. For simplicity, we assume that once a study sample has been enumerated, treatment is randomized and follow-up is complete, so that there is no confounding bias in expectation and no additional missing data or selection into the analytical sample (and therefore no selection bias as a problem of internal validity).
METHODS
Preliminary issues and notation
Sampling (S) might relate to covariates (Z) in several ways, including S causing Z (or S indicating differences in distributions of Z), Z causing S, and both S and Z being caused by some additional variable U; here we restrict our attention to the first of these cases, which is closely related to Bareinboim and Pearl's term “transportability” (3). We assume that the epidemiologist has conducted a study and wishes to transport the effect estimate from that study sample to an external target population. For convenience, we assume that information on the same set of covariates has been collected in the study sample and target data, and that the epidemiologist has concatenated the 2 data sets.
In the following, i indicates a participant index i = 1, 2, . . . n, n + 1, . . . N such that the study sample comprises n participants and the target population N − n participants; study participants are designated Si = 1, while individuals in the target population are designated Si = 0; and Zi is a vector of pretreatment covariates for participant i (see the Discussion section for comments on components of Z). Yai indicates the potential outcome under some specific treatment A = a for participant i.
Method
We note that this approach differs from the inverse probability (rather than odds) of selection weights; the latter method, described by Cole and Stuart (4), is appropriate when the study sample is a subset of the target population (i.e., for generalizability rather than transportability). Inverse odds weights are appropriate when the study sample and target population are nonoverlapping; if we consider “being in the study sample” to be a kind of treatment, this method is analogous to weighting for the average treatment effect in the untreated in nonexperimental studies (11, 12).
NUMERICAL EXAMPLE
To aid intuition around this method, consider a hypothetical trial of assignment to a new antiretroviral therapy regimen for human immunodeficiency virus (HIV) compared with assignment to a reference regimen, for the outcome of virological failure at 1 year, conducted in HIV-positive people living in the United States. Suppose the study sample for the trial comprises 2,000 participants, 1,000 with single covariate Z = 1 and 1,000 with Z = 0. Among participants with Z = 1, the risk difference is −0.2 (novel treatment is protective against failure); among participants with Z = 0, the risk difference is 0.0 (no effect of intervention). The crude sample average causal risk difference is therefore −0.1, a simple average of the 2 strata.
Our target population (alternately, a random sample from our target population) comprises 2,000 persons living with HIV in the United States, of whom 80% have Z = 1 and 20% have Z = 0. In this very simple case, we can hand-calculate the (target) population average causal effect in our external setting as 0.8 × (−0.20) + 0.2 × (0.00) = −0.16. In real data, we would need to use model-based approaches to account for the joint distribution of multiple continuous and categorical variables.
We concatenate the trial data (n = 2,000; 50% with Z = 1) with the target population (n = 2,000; 80% with Z = 1), obtaining a combined population of size 4,000, including 2,600 (65%) with Z = 1 and 1,400 (35%) with Z = 0. We proceed by estimating O(S = 1 | Z) = P(S = 1 | Z)/(1 − P(S = 1 | Z)), which in this case is (1,000/2,600)/(1 − 1,000/2,600) = 1,000/1,600 where Z = 1 and (1,000/1,400)/(1 − 1,000/1,400) = 400/1,000 where Z = 0. We would use these odds to calculate a weighted pseudopopulation of 1,600 persons ((1,000 × 1,600/1,000) = 1,600) for Z = 1 and 400 persons ((1,000 × 400/1,000) = 400) for Z = 0; we would then calculate the weighted risk difference as (1,600 × −0.2 + 400 × 0.0)/(1,600 + 400) = −0.16.
This estimate coincides with the common-sense simple weighted average we derived immediately above. In addition, these inverse odds weights coincide with the intuitive explanation of how individuals from the study sample ought to be weighted so as to represent individuals in the target population, as shown in Figure 1. Notably, increasing or decreasing the size of the target population has no impact on the final estimate, which is not necessarily the case in the method proposed by Cole and Stuart (4).
Concepts of weights to map from a study sample with oversampled Z = 0 (on left) to a target population (on right).
DISCUSSION
Typical epidemiologic and biostatistical analyses emphasize internal validity of causal effects, but (as others have noted (13, 14)) a causal effect without a specified target population is poorly defined. In practice, study samples for randomized trials are rarely sampled at random directly from the target population; indeed, because consent is an ethical necessity for enrollment in a clinical trial, trial participants are effectively never a random sample of the target population. Yet this is the premise that underlies the assumption of unconditional generalizability or transportability between the study sample and the target population—a frequent (if informal) claim in randomized trials. In contrast, the methods discussed and presented here allow us to relax this questionable premise: We no longer assume that the results from the study population are unconditionally transportable from the study sample to an arbitrary target population; rather, we assume that they are transportable conditional on variables in our model.
Some people may be uncomfortable with our assumption of conditional transportability, perhaps because herein we are explicit about assumptions that are typically hidden within vague statements about how “representative” the study sample is without addressing the questions 1) representative of what target population? and 2) representative according to which characteristics Z? There is a useful conceptual parallel here with the assumption of exchangeability between treated and untreated subjects for internal validity in an observational setting. The assumption of unconditional transportability is similar (but not identical) to the assumption of unconditional exchangeability (e.g., the causal effect is unconfounded), while the assumption of transportability conditional on variables in the model is similar (but again not identical) to the notion of conditional exchangeability (e.g., the causal effect is unconfounded conditional on a set of confounders).
These parallels are useful in considering the contents of Z. In earlier work, investigators have variously described Z as comprising all effect-measure modifiers (5) or as having components identifiable from causal diagrams (3, 7, 15). The clearest guideline is that Z should be S-admissible (16)—that is, that Z should include pretreatment covariates sufficient to d-separate sampling and the outcome variable (3). This guideline is analogous to that of selecting variables for d-separation of the exposure and outcome variables for internal validity.
As with conditional exchangeability for internal validity, conditional transportability of external validity carries with it additional assumptions: namely, positivity (9) and correct model specification. For transport-positivity to hold, the probability of being included in the sample must be greater than 0 for participants in all strata defined by Z in the target population. This assumption is necessary so that the Z-specific probability of the outcome estimated in the study can “stand in” for the Z-specific probability of the outcome in the target population (see Appendix). Of course, as with positivity for internal validity, transport-positivity may be replaced by making additional assumptions (e.g., smoothing under a parametric model). As noted elsewhere, additional conditions are necessary for transportability—specifically, similar patterns of interference and similar versions of treatment between the study sample and the target population (5).
The concerns about external validity discussed here are highly relevant to observational studies as well as to trials. The results of a trial are frequently, naively assumed to be transportable to a target population. Just as often, however, an epidemiologist assumes that observational cohorts are more representative of the target population and thus that there is less need to evaluate transportability directly (much less to identify the target population explicitly). In fact, the transportability of observational studies to a particular target population of interest is not guaranteed and must be evaluated carefully.
In many clinical trials, external validity is considered only an afterthought; however, consideration of both internal validity in the study sample and external validity in a target population is crucial to providing evidence which will best improve medicine and public health (13, 17). Quantitative approaches to transportability, such as the one described here, are straightforward and should be applied more widely.
ACKNOWLEDGMENTS
Author affiliations: Department of Epidemiology, Gillings School of Global Public Health, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina (Daniel Westreich, Jessie K. Edwards, Stephen R. Cole); Department of Epidemiology, Bloomberg School of Public Health, Johns Hopkins University, Baltimore, Maryland (Catherine R. Lesko); and Departments of Mental Health, Biostatistics, and Health Policy and Management, Bloomberg School of Public Health, Johns Hopkins University, Baltimore, Maryland (Elizabeth Stuart).
This research was supported by the Eunice Kennedy Shriver National Institute of Child Health and Human Development and the Office of the Director of the National Institutes of Health (award DP2-HD084070) and the National Institute of Allergy and Infectious Diseases (grant R01 AI100654).
We thank Dr. Michael G. Hudgens for expert advice on the preparation of the manuscript.
The content of this article is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Conflict of interest: none declared.
REFERENCES
APPENDIX
Inverse Odds of Sampling Weights
Weights with which to estimate the expected value of a binary potential outcome in a target population, , using data from a study sample (S = 1) and with covariates Z, can be derived from the g-formula.
Note that the 2 exchangeability assumptions above may require different sets of covariates; thus, for convenience, Z can be thought of as the union of the sets of covariates required for both exchangeability assumptions.
Author notes
Abbreviation: HIV, human immunodeficiency virus.
