Epidemiologists must ask causal questions. Describing patterns of disease and exposure is not sufficient to improve health. Instead, we seek to understand why such patterns exist and how we can best intervene to change them. The crucial role of causal thinking in this process has long been acknowledged in our field’s historical focus on confounding.
Major advances in formal causal frameworks have occurred over the past decades. Several specific applications, such as the use of causal graphs to choose adjustment variables1 or the use of counterfactuals to define the effects of longitudinal treatments,2 are now common in the epidemiologic literature. However, the tools of formal causal inference have the potential to benefit epidemiology much more extensively.
We argue that the wider application of formal causal tools can help frame sharper scientific questions, make transparent the assumptions required to answer these questions, facilitate rigorous evaluation of the plausibility of these assumptions, clearly distinguish the process of causal inference from the process of statistical estimation, and inform analyses of data and interpretation of results that rigorously respect the limits of knowledge. We, together with others, advocate for a systematic approach to causal questions that involves (1) specification of a causal model that accurately represents knowledge and its limits; (2) specification of the observed data and their link to the causal model; (3) translation of the scientific question into a counterfactual quantity; (4) assessment of whether, under what assumptions, this quantity is identified–whether it can be expressed as a parameter of the observed data distribution or estimand; (5) statement of the resulting statistical estimation problem; (6) estimation, including assessment of statistical uncertainty; and (7) interpretation of results (Figure 1).3,4 We emphasize how causal models can help navigate the ubiquitous tension between the causal questions posed by public health and the inevitably imperfect nature of available data and knowledge.
A GENERAL ROADMAP FOR CAUSAL INFERENCE
1. Specify knowledge about the system to be studied using a causal model.Of the several models available, we focus on the structural causal model,5–10which provides a unification of the languages of counterfactuals,11,12structural equations,13,14and causal graphs.1,7Structural causal models provide a rigorous language for expressing both background knowledge and its limits.
Causal graphs represent one familiar means of expressing knowledge about a data-generating process; we focus here on directed acyclic graphs (Figure 2). Figure 2A provides an example of a directed acyclic graph for a simple data-generating system consisting of baseline covariates W, an exposure A, and an outcome Y. Such graphs encode causal knowledge in several ways. First, graphs encode knowledge about the possible causal relations among variables. Knowledge that a given variable is not directly affected by a variable preceding it is encoded by omitting the corresponding arrow (referred to as “an exclusion restriction”). Figure 2A reflects the absence of any such knowledge; baseline covariates W may have affected the exposure A, and both may have affected the outcome Y. In other cases, investigators may have knowledge that justifies exclusion restrictions. For example, if A represents adherence to a randomly assigned treatment R (Figure 2B), it might be reasonable to assume that random assignment (if effectively blinded) had no effect on the outcome other than via adherence. Such knowledge is represented by omission of an arrow from R to Y. Second, omission of a double-headed arrow between two variables assumes that any unmeasured “background factors” that go into determining the values that these variables take are independent (or, equivalently, that the variables do not share an unmeasured cause, referred to as “an independence assumption”). Figure 2A makes no independence assumptions, whereas Figure 2B reflects the knowledge that, because R was randomly assigned, it shares no unmeasured common cause with any other variables.
The knowledge encoded in a causal graph can alternatively be represented using a set of structural equations, in which each node in the graph is represented as a deterministic function of its parents and a set of unmeasured background factors. The error term for a given variable X (typically denoted as UX) represents the set of unmeasured background factors that, together with variable X’s parents (nodes with arrows pointing to X), determine what value X takes (Figure 2). The set of structural equations, together with any restrictions placed on the joint distribution of the error terms (expressed on the graph as assumptions about the absence of unmeasured common causes between two nodes) together constitute a structural causal model.5,6
Such a structural causal model provides a flexible tool for encoding a great deal of uncertainty about the true data-generating process. Specifically, a structural causal model allows for uncertainty about the existence of a causal relationship between any two variables (through inclusion of an arrow between them), uncertainty about the distribution of all unmeasured background factors that go into determining the value of these variables (frequently, no restrictions are placed on the joint distribution of the errors, beyond any independence assumptions), and uncertainty about the functional form of causal relationships between variables (the structural equations can be specified nonparametrically). If knowledge in any of these domains is available, however, it can be readily incorporated. For example, if it is known that R was assigned independently to each subject with probability 0.5, this knowledge can be reflected in the corresponding structural equation (Figure 2B), as can parametric knowledge on the true functional form of causal relationships.
In sum, the flexibility of a structural causal model allows us to avoid many (although not all) unsubstantiated assumptions and thus facilitates specification of a causal model that describes the true data-generating process. Alternative causal models differ in their assumptions about the nature of causality and make fewer untestable assumptions.15–19
2. Specify the observed data and their link to the causal model.Specification of how the observed data were generated by the system described in our causal model provides a bridge between causal modeling and statistical estimation.
The causal model (representing knowledge about the system to be studied) must be explicitly linked to the data measured on that system. For example, a study may have measured baseline covariates W, exposure A, and outcome Y on an independent random sample of n individuals from some target population. The observed data on a given person thus consist of a single copy of the random variable O = (W, A, Y). If our causal model accurately describes the data-generating system, the data can be viewed as n independent and identically distributed draws of O from the corresponding system of equations.
More complex links between the observed data and the causal model are also possible. For example, study participants may have been sampled on the basis of exposure or outcome status. More complex sampling schemes such as these can be handled either by specifying alternative links between the causal model and the observed data or by incorporating selection or sampling directly into the causal model.6,20
The structural causal model is assumed to describe the system that generated the observed data. This assumption may or may not have testable implications. For example, the systems described by Figures 2A, D, and E could generate any possible distribution of O = (W, A, Y). We thus say that these causal models place no restrictions on the joint distribution of the observed data, implying a nonparametric statistical model. In contrast, the system described by Figure 2B can generate only distributions of O = (W, R, A, Y) in which R is independent of W (a testable assumption). This causal model thus implies a semiparametric statistical model. Independence (and conditional independence) restrictions of this nature can be read from the graph using the criterion of d-separation.5,7 (The set of possible distributions for the observed data may also be restricted via functional or inequality constraints.)21–23
The statistical model should reflect true knowledge about the data-generating process, ensuring that it contains the true distribution of the observed data. Although in some cases parametric knowledge about the data-generating process may be available, in many cases a causal model that accurately represents knowledge is compatible with any possible distribution for our observed data, implying a nonparametric statistical model.
3. Specify the target causal quantity.The formal language of counterfactuals forces explicit statement of a hypothetical experiment to answer the scientific question of interest.
Specification of an ideal experiment and a corresponding target counterfactual quantity helps ensure that the scientific question drives the design of a data analysis and not vice versa.24 This process forces the researcher to define exactly which variables would ideally be intervened on, what the interventions of interest would look like, and how the resulting counterfactual outcome distributions under these interventions would be compared.
A causal model on the counterfactual distributions of interest can be specified directly (and need not be graphical)2,11,12,25 or, alternatively, it can be derived by representing the counterfactual intervention of interest as an intervention on the graph (or set of equations). The initial structural causal model describes the set of processes that could have generated (and thus the set of possible distributions for) the observed data. The postintervention causal model describes the set of processes that could have generated the counterfactual variables we would have measured in our ideal experiment and thus the set of possible distributions for these variables. Figure 2C provides an illustration of the postintervention structural causal model corresponding to Figure 2A, under an ideal experiment in which exposure A is set to 0 for all persons.
One common counterfactual quantity of interest is the average treatment effect: the difference in mean outcome that would have been observed had all members of a population received versus not received some treatment. This quantity is expressed in terms of counterfactuals as E(Y1−Y0), where Ya denotes the counterfactual outcome under an intervention to set A = a. The corresponding ideal experiment would force all members of a population to receive the treatment and then roll back the clock and force all to receive the control.
In many cases, the ideal experiment is quite different. For example, if the exposure of interest were physical exercise, one might be interested in the counterfactual risk of mortality if all subjects without a contraindication to exercise were assigned the intervention (an example of a “realistic” dynamic regime, in which the treatment assignment for a given person depends on that person’s characteristics).26–28 Furthermore, if the goal is to evaluate the impact of a policy to encourage more exercise, a target quantity might compare the existing outcome distribution with the distribution of the counterfactual outcome under an intervention to shift the exercise distribution, while allowing each person’s exercise level to remain random (an example of a stochastic intervention).17,29–31 Additional examples include effects of interventions on multiple nodes and mediation effects.2,32–36Figure 3 lists major decision points when specifying a counterfactual target parameter and provides examples of general categories of causal questions that can be formally defined using counterfactuals.
4. Assess identifiability.A structural causal model provides a tool for understanding whether background knowledge, combined with the observed data, is sufficient to allow a causal question to be translated into a statistical estimand, and, if not, what additional data or assumptions are needed.
Step 3 translated the scientific question into a parameter of the (unobserved) counterfactual distribution of the data under some ideal intervention(s). We say that this target causal quantity is identified, given a causal model and its link to the observed data, if the target quantity can be expressed as a parameter of the distribution of the observed data alone—an estimand. Structural causal models provide a general tool for assessing identifiability and deriving estimands that, under explicit assumptions, equal causal quantities.5,37–39
A familiar example is provided by the use of causal graphs to choose an adjustment set when estimating the average treatment effect (or other parameter of the distribution of Ya). For observed data consisting of n independent identically distributed copies of O = (W, A, Y), when preintervention covariates W block all unblocked back-door paths from A to Y in the causal graph (the “back-door criterion”),5,7 then the distribution of Ya is identified according to the “G-computation formula” (given in Equation 1 for discrete valued O).6 The same result also holds under the “randomization assumption” that Ya is independent of A given W18:
Equation 1 equates a counterfactual quantity (the left-hand side) with an estimand (the right-hand side) that can be targeted for statistical estimation. The back-door criterion can be straightforwardly evaluated using the causal graph; it fails in Figure 2A and B but holds in Figure 2D and E. Single world intervention graphs allow graphical evaluation of the randomization assumption for a given causal model.17
Although the estimand in Equation 1 might seem to be an obvious choice, use of a formal causal framework can be instrumental in choosing an estimand. For example, in Figure 2F,
The use of a structural causal model in this case warns against adjustment for Z, even if it precedes A. In other cases, adjustment for interventions that occur after the exposure may be warranted.
Many scientific questions imply more complex counterfactual quantities (Figure 3), for which no single adjustment set will be sufficient and alternative identifiability results are needed. Causal frameworks provide a tool for deriving these results, often resulting in new estimands and thereby suggesting different statistical analyses. Examples include effect mediation,19,32–34 the effects of interventions at multiple time points,2,18,41 dynamic interventions,17,18,42 causal and noncausal parameters in the presence of informative censoring or selection bias,2,6,20 and the transport of causal effects to new settings.27,43–45
5. Commit to a statistical model and estimand.A causal model that accurately represents knowledge can help to select an estimand as close as possible to the wished-for causal quantity, while emphasizing the challenge of using observational data to make causal inferences.
In many cases, rigorous application of a formal causal framework forces us to conclude that existing knowledge and data are insufficient to claim identifiability—in itself a useful contribution. The process can also often inform future studies by suggesting ways to supplement data collection.46 However, in many cases, better data are unlikely to become available or “current best” answers are needed to questions requiring immediate action.
One way to navigate this tension is to rigorously differentiate assumptions that represent real knowledge from assumptions that do not, but which if true, would result in identifiability. We refer to the former as “knowledge-based assumptions” and the latter as “convenience-based assumptions.” An estimation problem that aims to provide a current best answer can then be defined by specifying: (1) a statistical model implied by knowledge-based assumptions alone (and thus ensured to contain the truth); (2) an estimand that is equivalent to the target causal quantity under a minimum of convenience-based assumptions; and (3) a clear differentiation between convenience-based assumptions and real knowledge. In other words, a formal causal framework can provide a tool for defining a statistical estimation problem that comes as close as possible to addressing the motivating scientific question, given the data and knowledge currently available, while remaining transparent regarding the additional assumptions required to endow the resulting estimate with a causal interpretation.
For example, measured variables are rarely known to be sufficient to control confounding. Figure 2A may represent the knowledge-based causal model, under which the effect of A on Y is unidentified; however, under the augmented models in Figure 2D and E, result (1) would hold. If our goal is to estimate the average treatment effect, we might thus define a statistical estimation problem in which: (1) the statistical model is nonparametric (as implied by Figure 2A and by Figure 2D and E); and (2) we select
as the estimand. We can then proceed with estimation of Equation 2, while making explicit the conditions under which the estimand may diverge from the causal effect of interest.
6. Estimate.Choice between estimators should be motivated by their statistical properties.
Once the statistical model and estimand have been defined, there is nothing causal about the resulting estimation problem. A given estimand, such as Equation 2, can be estimated in many ways. Estimators of Equation 2 include those based on inverse probability weighting,47 propensity score matching,40 regression of the outcome on exposure and confounders (followed by averaging with respect to the empirical distribution of confounders), and double robust efficient methods,48,49 including targeted maximum likelihood.4
To take another example, marginal structural models (Figure 3) are often used to define target causal quantities, particularly when the exposure has multiple levels.2,35 Under causal assumptions, such as the randomization assumption (or its sequential counterpart), this quantity is equivalent to a specific estimand. However, once the estimand has been specified, estimation itself is a purely statistical problem. The analyst is free to choose among several estimators; inverse probability–weighted estimators are simply one popular class.
Any given class of estimator itself requires, as “ingredients,” estimators of specific components of the observed data distribution. For example, one approach to estimating Equation 2 is to specify an estimator of E(Y|A, W). (P(W = w) is typically estimated as the sample proportion.) In many cases, the true functional form of this conditional expectation is unknown. In some cases, E(Y|A, W) can be estimated nonparametrically using a saturated regression model; however, at moderate sample sizes or if W or A are high dimensional or contain continuous covariates, the corresponding contingency tables will have empty cells and this approach will break down. As a result, in many practical scenarios, the analyst must either rely on a parametric regression model that is likely to be misspecified (risking significant bias and misleading inference) or do some form of dimension reduction and smoothing to trade off bias and variance.
Similar considerations arise when specifying an inverse probability–weighted or propensity score–based estimator. In this case, estimator consistency depends on consistent estimation of the treatment mechanism or propensity score P(A = a|W). An extensive literature on data-adaptive estimation addresses how best to approach this tradeoff for the purposes of fitting a regression object such as E(Y|A, W) or P(A = a|W).50 An additional literature on targeted estimation discusses how to update resulting estimates to achieve an optimal bias-variance tradeoff for the final estimand, as well as how to generate valid estimates of statistical uncertainty when using data-adaptive approaches.4
In sum, there is nothing more or less causal about alternative estimators. However, estimators have important differences in their statistical properties, and these differences can result in meaningful differences in performance in commonly encountered scenarios, such as strong confounding.51–53 Choice among estimators should be driven by their statistical properties and by investigation of performance under conditions similar to those presented by a given applied problem.
7. Interpret.Use of a causal framework makes explicit and intelligible the assumptions needed to move from a statistical to a causal interpretation of results.
Imposing a clear delineation between knowledge-based assumptions and convenience-based assumptions provides a hierarchy of interpretation for an analysis (Figure 4). The use of a causal framework to define a problem in no way undermines purely statistical interpretation of results. For example, an estimate of Equation 2 can be interpreted as an estimate of the difference in mean outcome between exposed and unexposed subjects who have the same values of observed baseline covariates (averaged with respect to the distribution of the covariates in the population). The same principle holds for more complex analyses—for example, if inverse probability weighting is used to estimate the parameters of a marginal structural model.
The use of a formal causal framework ensures that the assumptions needed to augment the statistical interpretation with a causal interpretation are explicit. For example, if we believe that Figure 2D or E represents the true causal structure that generated our data, then our estimate of Equation 2 can be interpreted as an estimate of the average treatment effect. The use of a causal model and a clear distinction between convenience-based and knowledge-based assumptions makes clear that the true value of an estimand may differ from the true value of the causal effect of interest. The magnitude of this difference is a causal quantity; choice of statistical estimator does not affect it. Statistical bias in an estimator can be minimized through data-driven methods, while evaluating the likely or potential magnitude of the difference between the estimand and the wished-for causal quantity requires alternative approaches (sometimes referred to as sensitivity analyses).54–58
There is nothing in the structural causal model framework that requires the intervention to correspond to a feasible experiment.6,59,60 In particular, some parameters of substantive interest (such as the natural direct effect) cannot be defined in terms of a testable experiment but can nonetheless be defined and identified under a structural causal model.32 If, in addition to the causal assumptions needed for identifiability, the investigator is willing to assume that the intervention used to define the counterfactuals corresponds to a conceivable and well-defined intervention in the real world, interpretation can be further expanded to include an estimate of the impact that would be observed if that intervention were to be implemented in practice (in an identical manner, in an identical population, or, under additional assumptions, in a new population).27,43–45 Finally, if the intervention could be feasibly randomized with complete compliance and follow-up, interpretation of the estimate can be further expanded to encompass the results that would be seen if one implemented this hypothetical trial. The decision of how far to move along this hierarchy can be made by the investigator and reader based on the specific application at hand. The assumptions required are explicit and, when expressed using a causal graph, readily understandable by subject matter experts.
The debate continues as to whether causal questions and assumptions should be restricted to quantities that can be tested and thereby refuted via theoretical experiment.15–17,19,61 Given that most such experiments will be done, if ever, only after the public health relevance of an analysis has receded,17,19 the extent to which such a restriction will enhance the practical impact of applied epidemiology seems unclear. However, while we have chosen to focus on the structural casual model, our goal is not to argue for the supremacy of a single class of causal model as optimal for all scientific endeavor. Rather, we suggest that the systematic application of a roadmap will improve the impact of applied analyses, irrespective of formal causal model chosen.
Epidemiologists continue to debate whether and how to integrate formal causal thinking into applied research. Many remain concerned that the use of formal causal tools leads both to overconfidence in our ability to estimate causal effects from observational data and to the eclipsing of common sense by complex notation and statistics. As Levine articulates this position, “the language of ‘causal modeling’ is being used to bestow the solidity of the complex process of causal inference upon mere statistical analysis of observational data… conflating statistical and causal inference.”62 We argue that a formal causal framework, when used appropriately, represents a powerful tool for preventing exactly such conflation and overreaching interpretation.6,63
Like any tool, the benefits of a causal inference framework depend on how it is used. The ability to define complex counterfactual parameters does not ensure that they address interesting or relevant questions. The ability to formally represent causal knowledge as graphs or equations is not a license to exaggerate what we know. The ability to prove equivalence between a target counterfactual quantity and an estimand under specific assumptions does not make these assumptions true, nor does it ensure that they can be readily evaluated. The best estimation tools can still produce unreliable statistical estimates when data are inadequate.
Good epidemiologic practice requires us to learn as much as possible about how our data are generated, to be clear about our question, to design an analysis that answers this question as well as possible using available data, to avoid or minimize assumptions not supported by knowledge, and to be transparent and skeptical when interpreting results. Used appropriately, a formal causal framework provides an invaluable tool for integrating these basic principles into applied epidemiology. We argue that the routine use of formal causal modeling would improve the quality of epidemiologic research, as well as research in the innumerable disciplines that aim to use statistics to learn how the world works.
1. Greenland S, Pearl J, Robins JM. Causal diagrams for epidemiologic research. Epidemiology. 1999;10:37–48
2. Robins JM, Hernán MA, Brumback B. Marginal structural models and causal inference in epidemiology. Epidemiology. 2000;11:550–560
3. Pearl JHoyle RH. The Causal Foundations of Structural Equation Modeling. Handbook of Structural Equation Modeling. 2012 New York Guilford Press:68–91
4. van der Laan M, Rose S Targeted Learning: Causal Inference for Observational and Experimental Data. 2011 Berlin, Heidelberg, New York Springer
5. Pearl J Causality: Models, Reasoning, and Inference. 2000 New York Cambridge University Press
6. Pearl J. Causal diagrams for empirical research. Biometrika. 1995;82:669–710
7. Pearl J. Causal inference in statistics: an overview. Statistics Surveys. 2009;3:96–146
8. Strotz RH, Wold HO. Recursive vs. nonrecursive systems: an attempt at synthesis (part I of a triptych on causal chain systems). Econometrica. 1960;28:417–427
9. Haavelmo T. The statistical implications of a system of simultaneous equations. Econometrica. 1943;11:1–12
10. Wright S. Correlation and causation. J Agric Res. 1921;20:557–585
11. Neyman J. On the application of probability theory to agricultural experiments. Essay on principles. Section 9. Stat Sci. 1923;5:465–472
12. Rubin D. Estimating causal effects of treatments in randomized and nonrandomized studies. J Educational Psychol. 1974;66:688–701
13. Duncan O Introduction to Structural Equation Models. New York: Academic Press;. 1975
14. Goldberger A. Structural equation models in the social sciences. Econometrica: Journal of the Econometric Society. 1972;40:979–1001
15. Spirtes P, Glymour C, Scheines R Causation, Prediction, and Search. Number 81 in Lecture Notes in Statistics. 1993 New York/Berlin: Springer-Verlag
16. Dawid AP. Causal inference without counterfactuals. J Am Stat Association. 2000;95:407–424
17. Richardson TS, Robins JM. Single World Intervention Graphs (SWIGs): A Unification of the Counterfactual and Graphical Approaches to Causality. Technical Report 128
. Seattle, WA: Center for Statistics and the Social Sciences, University of Washington; 2013
18. Robins J. A new approach to causal inference in mortality studies with a sustained exposure period. Mathem Mod. 1986;7:1393–512
19. Robins JM, Richardson T. Alternative graphical causal models and the identification of direct effects. In: Shrout P, Keyes K, Ornstein K, eds. Causality and psychopathology: Finding the determinants of disorders and their cures
. Oxford, UK: Oxford University Press; 2010:103–58
20. Hernán MA, Hernández-Díaz S, Robins JM. A structural approach to selection bias. Epidemiology. 2004;15:615–625
21. Kang C, Tian J Inequality Constraints in Causal Models with Hidden Variables. 2006 Arlington, VA AUAI Press
22. Shpitser I, Pearl J. Dormant independence. 2008Proceedings of the 23rd National Conference on Artificial Intelligence. Vol. 2 Menlo Park, CA: AAAI Press:1081–1087
24. Hernan MA, Alonso A, Logan R, et al. Observational studies analyzed like randomized experiments: an application to postmenopausal hormone therapy and coronary heart disease. Epidemiology. 2008;19:766–779
25. Robins JM. Structural nested failure time models. In: Encyclopedia of Biostatistics. 1998 Chichester, UK John Wiley and Sons:4372–4389 Armitage P and Colton T, eds.
26. Hernán MA, Lanoy E, Costagliola D, Robins JM. Comparison of dynamic treatment regimes via inverse probability weighting. Basic Clin Pharmacol Toxicol. 2006;98:237–242
27. Robins J, Orellana L, Rotnitzky A. Estimation and extrapolation of optimal treatment and testing strategies. Stat Med. 2008;27:4678–4721
28. van der Laan MJ, Petersen ML. Causal effect models for realistic individualized treatment and intention to treat rules. Int J Biostat. 2007;3:Article 3
29. Taubman SL, Robins JM, Mittleman MA, Hernán MA. Intervening on risk factors for coronary heart disease: an application of the parametric g-formula. Int J Epidemiol. 2009;38:1599–1611
30. Muñoz ID, van der Laan M. Population intervention causal effects based on stochastic interventions. Biometrics. 2012;68:541–549
31. Cain LE, Robins JM, Lanoy E, Logan R, Costagliola D, Hernán MA. When to start treatment? A systematic approach to the comparison of dynamic regimes using observational data. Int J Biostat. 2010;6:Article 18
32. Pearl J. Direct and indirect effects. 2001In: Proceeedings of the 17th Conference on Uncertainty in Artificial Intelligence San Francisco, CA: Morgan Kaufmann:411–420
33. Petersen ML, Sinisi SE, van der Laan MJ. Estimation of direct causal effects. Epidemiology. 2006;17:276–284
34. Robins JM, Greenland S. Identifiability and exchangeability for direct and indirect effects. Epidemiology. 1992;3:143–155
35. Robins J. Association causation and marginal structural models. Synthese. 1999;121:151–179
36. Robins J, Hernan MAFitzmaurice G, Davidian M, Verbeke G, Molenbergh G. Estimation of the causal effects of time-varying exposures. Longitudinal Data Analysis. 2009 London: Chapman and Hall/CRC:566–568
37. Shpitser I, Pearl J. Complete identification methods for the causal hierarchy. J Mach Learn Res. 2008;9:1941–1979
38. Tian J, Pearl J. A general identification condition for causal effects. In: Proceedings of the 18th National Conference on Artificial Intelligence. 2002 Menlo Park, CA: AAAI Press:567–573
39. Tian J, Shpitser IDechter R, Geffner H, Halpern J. On identifying causal effects. In: Heuristics, Probability and Causality: A Tribute to Judea Pearl. 2010 UK College Publications:415–444
40. Rosenbaum P, Rubin D. The central role of the propensity score in observational studies. Biometrika. 1983;70:41–55
41. Robins J. Addendum to: “A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect”. Comput Math Appl. 1987;14:923–945
42. Tian J. Identifying dynamic sequential plans. 2008In: Proceeedings of the 24th Conference on Uncertainty in Artificial Intelligence Helsinki, Finland: AUAI Press:554–561
43. Petersen ML. Compound treatments, transportability, and the structural causal model: the power and simplicity of causal graphs. Epidemiology. 2011;22:378–381
44. Hernán MA, VanderWeele TJ. Compound treatments and transportability of causal inference. Epidemiology. 2011;22:368–377
45. Pearl J, Bareinboim E Transportability Across Studies: A Formal Approach. 2011 Los Angeles, CA Computer Science Department, University of California
46. Geng EH, Glidden DV, Bangsberg DR, et al. A causal framework for understanding the effect of losses to follow-up on epidemiologic analyses in clinic-based cohorts: the case of HIV-infected patients on antiretroviral therapy in Africa. Am J Epidemiol. 2012;175:1080–1087
47. Hernán MA, Robins JM. Estimating causal effects from epidemiological data. J Epidemiol Community Health. 2006;60:578–586
48. Robins J, Rotnitzky A, Zhao L. Estimation of regression coefficients when some regressors are not always observed. J Am Stat Assoc. 1994;89:846–866
49. Scharfstein DO, Rotnitzky A, Robins JM. Adjusting for nonignorable drop-out using semiparametric nonresponse models. J Am Stat Assoc. 1999;94:1096–120
50. Hastie T, Tibshirani R, Friedman J The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 20092nd ed New York Springer-Verlag
51. Kang J, Schafer J. Demystifying double robustness: a comparison of alternative strategies for estimating a population mean from incomplete data. Stat Science. 2007;22:523–539
52. Petersen ML, Porter KE, Gruber S, Wang Y, van der Laan MJ. Diagnosing and responding to violations in the positivity assumption. Stat Methods Med Res. 2012;21:31–54
53. Moore KL, Neugebauer R, Laan MJ, Tager IB. Causal inference in epidemiological studies with strong confounding. Statistics in Medicine. 2012;31:1380–404
54. Robins J, Rotnitzky A, Scharfstein DHalloran M, Berry D. Sensitivity analysis for selection bias and unmeasured confounding in missing data and causal inference models. Statistical Models in Epidemiology: The Environment and Clinical Trials. 1999 New York, NY Springer:1–92
55. Imai K, Keele L, Yamamoto T. Identification, inference, and sensitivity analysis for causal mediation effects. Stat Sci. 2010;25:51–71
56. Greenland S. Multiple-bias modelling for analysis of observational data. J Royal Stat Soc A. 2005;168:267–306
57. Vanderweele TJ, Arah OA. Bias formulas for sensitivity analysis of unmeasured confounding for general outcomes, treatments, and confounders. Epidemiology. 2011;22:42–52
58. Díaz I, van der Laan MJ. Sensitivity analysis for causal inference under unmeasured confounding and measurement error problems. Int J Biostat. 2013;9:149–160
59. Hernán MA. Invited commentary: hypothetical interventions to define causal effects–afterthought or prerequisite? Am J Epidemiol. 2005;162:618–620; discussion 621
60. van der Laan M, Haight T, Tager I. Respond to “ Hypothetical interventions to define causal effects.” Am J Epidemiol. 2005;162:621–622
61. Pearl J. Causal analysis in theory and practice: on mediation, counterfactuals and manipulations. Posting to UCLA Causality Blog 3 May, 2010. Available at: http://www.mii.ucla.edu/causality/?p=133
. UCLA Causality Blog2010. Accessed October 28, 2013.
62. Levine B. Causal models [letter]. Epidemiology. 2009;20:931
63. Hogan JW. Causal Models: the author responds. Epidemiology. 2009;20:931–932
64. Petersen ML, Wang Y, van der Laan MJ, Guzman D, Riley E, Bangsberg DR. Pillbox organizers are associated with improved adherence to HIV antiretroviral therapy and viral suppression: a marginal structural model analysis. Clin Infect Dis. 2007;45:908–915
65. Bodnar LM, Davidian M, Siega-Riz AM, Tsiatis AA. Marginal structural models for analyzing causal effects of time-dependent treatments: an application in perinatal epidemiology. Am J Epidemiol. 2004;159:926–934
66. Hernán MA, McAdams M, McGrath N, Lanoy E, Costagliola D. Observation plans in longitudinal studies with time-varying treatments. Stat Methods Med Res. 2009;18:27–52
67. Rosenblum M, Jewell NP, van der Laan M, Shiboski S, van der Straten A, Padian N. Analyzing direct effects in randomized trials with secondary interventions: an application to HIV prevention trials. J R Stat Soc Ser A Stat Soc. 2009;172:443–465
68. Gsponer T, Petersen M, Egger M, et al. The causal effect of switching to second-line ART in programmes without access to routine viral load monitoring. AIDS. 2012;26:57–65
69. Petersen ML, van der Laan MJ, Napravnik S, Eron JJ, Moore RD, Deeks SG. Long-term consequences of the delay between virologic failure of highly active antiretroviral therapy and regimen modification. AIDS. 2008;22:2097–2106
70. Westreich D, Cole SR, Tien PC, et al. Time scale and adjusted survival curves for marginal structural Cox models. Am J Epidemiol. 2010;171:691–700
71. Wester CW, Stitelman OM, deGruttola V, Bussmann H, Marlink RG, van der Laan MJ. Effect modification by sex and baseline CD4+ cell count among adults receiving combination antiretroviral therapy in Botswana: results from a clinical trial. AIDS Res Hum Retroviruses. 2012;28:981–988
72. Shpitser I, Pearl J. Effects of treatment on the treated: identification and generalization. 2009In: Proceeedings of the 24th Conference on Uncertainty in Artificial Intelligence Arlington, VA: AUAI Press:514–521