Abstract
In causal mediation analysis, nonparametric identification of the natural indirect effect typically relies on, in addition to no unobserved pre-exposure confounding, fundamental assumptions of (i) so-called “cross-world-counterfactuals” independence and (ii) no exposure-induced confounding. When the mediator is binary, bounds for partial identification have been given when neither assumption is made, or alternatively when assuming only (ii). We extend existing bounds to the case of a polytomous mediator, and provide bounds for the case assuming only (i). We apply these bounds to data from the Harvard PEPFAR program in Nigeria, where we evaluate the extent to which the effects of antiretroviral therapy on virological failure are mediated by a patient’s adherence, and show that inference on this effect is somewhat sensitive to model assumptions.
1 Introduction
Causal mediation analysis seeks to determine the role that an intermediate variable plays in transmitting the effect from an exposure to an outcome. An indirect effect refers to the effect that goes through the intermediate variable; a direct effect is a measure of the effect that does not. The study of causal mediation has enjoyed an explosion in popularity in recent years [1, 2, 3, 4, 5], not only in terms of theoretical developments, but also in practice. This has been most notable in the fields of epidemiology and social sciences. This strand of work is based on ideas originating from Robins and Greenland [6] and Pearl [7] grounded in the language of potential outcomes [8, 9, 10] to give nonparametric definitions of effects involved in mediation analysis, allowing for settings where interactions and nonlinearities may be present.
Consider an intervention which sets the exposure of interest for all subjects in the population to one of two possible values: a reference value or an active value. The total effect of such an intervention corresponds to the change of the counterfactual outcome mean if the exposure were set to the active value compared with if it were set to the reference value. Robins and Greenland [6] formalized the concept of effect decomposition of the total effect into direct and indirect effects by describing pure direct and indirect effects using counterfactual language. Pearl [7] further formalized this concept, giving general definitions using counterfactual notation to what he termed natural direct and indirect effects, as well as general identification results. The pure direct effect (PDE) corresponds to the change in the counterfactual outcome mean under an intervention which changes a person’s exposure status from the reference value to the active value, while maintaining the person’s mediator at the value it would have had under the exposure reference value. In contrast, the natural indirect effect (NIE) corresponds to the change in the average counterfactual outcome under an intervention that sets a person’s exposure value to the active value, while changing the value of the mediator from the value it would have had under the reference exposure value, to its value under the active exposure value. The PDE and NIE sum to give the total effect.
Identification of these natural effects has been somewhat controversial as it requires assumptions that may be overly restrictive for many applications. First, identification invokes a so-called cross-world-counterfactuals-independence assumption, which by virtue of involving counterfactuals under conflicting interventions on the exposure, cannot be enforced experimentally [7, 11]. Secondly, a necessary assumption for identification rules out the presence of exposure-induced confounding of the mediator’s effect on the outcome, even if all confounders are observed. While this assumption is in principle testable provided no unmeasured confounding, more often than not, post-exposure covariates are altogether ignored in routine application, in which case mediation analyses may be invalid. These issues have been considered recently, and some work has been done on partial or point identification under a weaker assumption. Specifically, on the one hand Robins and Richardson [11] and Tchetgen Tchetgen and VanderWeele [12] provide conditions for point identification of the PDE and NIE when a confounder is directly affected by the exposure. On the other hand, Robins and Richardson [11] give bounds for the PDE and NIE for binary mediator without making the cross-world-counterfactual-independence assumption, but assuming no exposure-induced confounding of the mediator-outcome relation, and Tchetgen Tchetgen and Phiri [13] extend these bounds to account for exposure-induced confounding. Bounds are commonly employed in causal inference when structural assumptions are not sufficiently strong to give point identification of a causal parameter of interest [14, 14, 15, 16, 17, 18, 19, 20, 21]. We build on this previous work to provide a number of new nonparametric bounds for the PDE and NIE allowing for a polytomous mediator under relaxations of the assumptions of (i) cross-world-counterfactuals independence, and (ii) no exposure-induced confounding, both separately and jointly. In particular, we relax assumption (ii) to allow for exposure-induced confounders when these confounders are measured and discrete. We apply these bounds to data from the Harvard PEPFAR program in Nigeria, where we evaluate the extent to which the effects of antiretroviral therapy on virological failure are mediated by a patient’s adherence.
2 Preliminaries
For a directed acyclic graph (
are mutually independent; the
are mutually independent for each
To view the
The
The graph in Figure 1(A) illustrates the simplest possible mediation setting, where
This
Richardson and Robins [24] propose another form of causal graphs, known as Single-World Intervention Graphs (
These graphs manage to clear up some of the ambiguity inherent to
For both full and partial identification of the
We will consider as well defined the nested counterfactual
The terms
This model then implies that for all
Cross-world counterfactual independence statements, however, are not experimentally enforceable [11]. This issue has been discussed extensively [11, 24], and in large part motivated the development of
In Section 3, we extend this result to the setting of a polytomous
As previously mentioned, another often-overlooked condition required for identification of
Generally, even under an
Clearly the joint probability term can never be identified from observed data, since we will never be able to observe
which is in fact identical to the identification formula under the
A few conditions for identification in this setting have been proposed. Robins and Richardson [11] give two. The first is that
It seems biologically unlikely, however, that in a scenario in which
The above assumption is implied by rank preservation [11], which is unlikely to hold in social and health sciences as it rules out individual-level effect heterogeneity [12]. As none of these conditions are experimentally verifiable, the authors themselves “do not advocate blithely adopting such assumptions in order to preserve identification of the
Tchetgen Tchetgen and VanderWeele [12] give two testable conditions for identification of
where
Their second condition is no
for all levels
Eschewing the cross-world-counterfactual assumptions of the
We extend these bounds as well to allow for polytomous
3 New partial identification results
We begin by extending the bounds of Robins and Richardson [11] and Tchetgen Tchetgen and Phiri [13] to settings with discrete mediator and outcome. Proofs can be found in the Appendix.
Theorem 1
Under the
The upper and lower bounds coincide when
Corollary 1
For polytomous
The second part of the corollary continues to hold even when there is a hidden common cause of
Whereas the previous results invoked no cross-world-counterfactual independences under the
with
and
The following result states that bounds for
Theorem 2
Under the
where
Similar to the previous result, these bounds coincide if either
As mentioned, all results given here can be extended to settings with observed pre-exposure confounders, which we denote
The identification formulas in Corollary 1 are the same, but conditional on
When
4 Application to Harvard PEPFAR data set
We now consider an application to a data set collected by the Harvard President’s Emergency Plan for AIDS Relief (PEPFAR) program in Nigeria. The data set consists of HIV-1 infected adult patients who had not previously received antiretroviral therapy (ART), began ART in the program, and were followed at least one year following initiation. Patients without reliable viral load data at two of the hospitals were excluded. Only complete cases initially prescribed to either TDF+3TC/FTC+NVP or AZT+3TC+NVP[1] were considered for this analysis. Thus, the data set we consider consists of 6,627 patients, 1,919 of whom were prescribed to TDF+3TC/FTC+NVP, and the remaining 4,708 prescribed to AZT+3TC+NVP.
There has accumulated evidence of a differential effect on virologic failure between these two first-line antiretroviral treatment regimens [28]. Virologic failure is defined by the World Health Organization as repeat viral load above 1,000 copies/mL. We base this on measurements at 12 and 18 months of ART duration in our analysis. A natural question of scientific interest is what role adherence plays in mediating this differential effect. We are primarily interested in learning about the scientific mechanism of this effect on the individual level. The natural indirect effect best captures this mechanism in that it captures an isolated effect difference mediated by adherence by, in a sense, deactivating effect differences along all other possible causal pathways. We specifically examine the effect through adherence over the second six months since treatment assignment, i.e., the six months prior to the first viral load measurement. Identification is complicated by the presence of treatment toxicity, which clearly affects adherence directly, and has the potential to modify the effect of the treatment assignment on virologic failure. Thus, toxicity measured at six months after treatment assignment is an exposure-induced confounder of the effect of the mediator on the outcome. Further, toxicity and virologic failure are likely to be rendered dependent by unobserved underlying biological common causes as in Figure 3, where
Let
Here we estimate the natural indirect effect of
It is immediately apparent that the range of uncertainty for the
Another investigator unwilling to impose cross-world-counterfactual independence assumptions is left with little to say as the bounds are considerably wider, regardless of how toxicity is handled. These bounds easily contain the null hypothesis of no
5 Discussion
We have shown that PEPFAR results are sensitive to the choice of assumptions made, consequently, we counsel investigators employing mediated effects to exercise caution in considering the basis for point identification and to explicitly state the assumptions required for validity. Where assumptions are empirically untestable, they should be argued for on the basis of scientific understanding, and ideally the alternative should be explored by employing partial identification bounds given both here and elsewhere. While some work has been done to develop sensitivity analyses for unmeasured confounding of the mediator [3, 34, 35], sensitivity analyses for ranges of plausible associations between cross-world counterfactuals remain undeveloped. Further development of sensitivity analyses of both forms would be highly beneficial for practical use, and is fertile ground for future work. Additionally, interest is growing in mediation analysis in longitudinal settings with repeated measures of the exposure, confounders, and mediator. Extending this work to such settings is also a fruitful direction for future research. We hope that the work presented here will inspire deeper consideration and transparency regarding underlying identifying assumptions in the practice of mediation analysis.
Funding statement: This work was funded, in part, by the US Department of Health and Human Services, Health Resources and Services Administration (U51HA02522), the Centers for Disease Control and Prevention (CDC) through a cooperative agreement with the AIDS Prevention Initiative in Nigeria (APIN) (PS 001058), and by the National Institutes of Health (R01AI104459-01A1).
Acknowledgments
The authors gratefully acknowledge the hard work and dedication of the clinical, data, and laboratory staff at the PEPFAR supported Harvard/AIDS Prevention Initiative in Nigeria (APIN) hospitals that provided secondary data for this analysis. The contents are solely the responsibility of the authors and do not represent the official views of the funding institutions. We thank the anonymous referees for their helpful comments, which greatly improved the clarity of this article.
Appendix
Proofs of theorems
Proof
Proof ofTheorem 1. Applying the (sharp) Fréchet inequalities
to each summand in
yields the result. □
Proof
of Theorem 2. Since
References
1. Petersen ML, Sinisi SE, van der Laan MJ. Estimation of direct causal effects. Epidemiology. 2006;17:276–284.10.1097/01.ede.0000208475.99429.2dSearch in Google Scholar PubMed
2. Imai K, Keele L, Tingley D. A general approach to causal mediation analysis. Psychol Methods. 2010;15:309.10.1037/a0020761Search in Google Scholar PubMed
3. Tchetgen Tchetgen EJ, Shpitser I. Semiparametric theory for causal mediation analysis: Efficiency bounds, multiple robustness and sensitivity analysis. Ann Stat. 2012;40:1816–1845.10.1214/12-AOS990Search in Google Scholar PubMed PubMed Central
4. Shpitser I. Counterfactual graphical models for longitudinal mediation analysis with unobserved confounding. Cognit Sci. 2013;37:1011–1035.10.1111/cogs.12058Search in Google Scholar PubMed
5. VanderWeele T. Explanation in causal inference: Methods for mediation and interaction. New York, NY: Oxford University Press, 2015.Search in Google Scholar
6. Robins JM, Greenland S. Identifiability and exchangeability for direct and indirect effects. Epidemiology. 1992;3:143–155.10.1097/00001648-199203000-00013Search in Google Scholar PubMed
7. Pearl J. Direct and indirect effects. In: Breese J, Koller D, editors. Direct and indirect effects. Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence Morgan Kaufmann, San Francisco, CA, 2001:411–420.10.1145/3501714.3501736Search in Google Scholar
8. Splawa-Neyman J, Dabrowska D, Speed T. On the application of probability theory to agricultural experiments. Essay on principles. Section 9. Stat Sci. 1990;5:465–472.10.1214/ss/1177012031Search in Google Scholar
9. Rubin DB. Estimating causal effects of treatments in randomized and nonrandomized studies. J Educ Psychol. 1974;66:688.10.1037/h0037350Search in Google Scholar
10. Rubin DB. Bayesian inference for causal effects: The role of randomization. The Annals of Statistics. 1978;34–58.10.1214/aos/1176344064Search in Google Scholar
11. Robins JM, Richardson TS. Alternative graphical causal models and the identification of direct effects. Shrout p, and Keyes K, M and Ornstein K editors. Causality and psychopathology: Finding the determinants of disorders and their cures, New York, NY Oxford University Press; 2011: 103–158.10.1093/oso/9780199754649.003.0011Search in Google Scholar
12. Tchetgen Tchetgen EJ, VanderWeele TJ. On identification of natural direct effects when a confounder of the mediator is directly affected by exposure. Epidemiology. 2014;25:282.10.1097/EDE.0000000000000054Search in Google Scholar
13. Tchetgen Tchetgen EJ, Phiri K. Bounds for pure direct effect. Epidemiology. 2014;25:775–776.10.1097/EDE.0000000000000154Search in Google Scholar
14. Robins JM. The analysis of randomized and non-randomized AIDS treatment trials using a new approach to causal inference in longitudinal studies. Health Service Res Methodology: A Focus on AIDS. 1989;113:159.Search in Google Scholar
15. Balke AA, Pearl J. Probabilistic counterfactuals: Semantics, computation, and applications. Technical Report DTIC Document 1997.Search in Google Scholar
16. Zhang JL, Rubin DB. Estimation of causal effects via principal stratification when some outcomes are truncated by “death”. J Educ Behav Stat. 2003;28:353–368.10.3102/10769986028004353Search in Google Scholar
17. Kaufman S, Kaufman JS, MacLehose RF, Greenland S, Poole C. Improved estimation of controlled direct effects in the presence of unmeasured confounding of intermediate variables. Stat Med. 2005;24:1683–1702.10.1002/sim.2057Search in Google Scholar
18. Cheng J, Small DS. Bounds on causal effects in three-arm trials with non-compliance. J R Stat Soc Ser B Stat Methodol. 2006;68:815–836.10.1111/j.1467-9868.2006.00568.xSearch in Google Scholar
19. Cai Z, Kuroki M, Pearl J, Tian J. Bounds on direct effects in the presence of confounded intermediate variables. Biometrics. 2008;64:695–701.10.1111/j.1541-0420.2007.00949.xSearch in Google Scholar
20. Sjölander A. Bounds on natural direct effects in the presence of confounded intermediate variables. Stat Med. 2009;28:558–571.10.1002/sim.3493Search in Google Scholar
21. Taguri M, Chiba Y. A principal stratification approach for evaluating natural direct and indirect effects in the presence of treatment-induced intermediate confounding. Stat Med. 2015;34:131–144.10.1002/sim.6329Search in Google Scholar
22. Pearl J. Causality: Models, reasoning, and inference, 2nd ed. New York: Cambridge University Press; 2009.10.1017/CBO9780511803161Search in Google Scholar
23. Robins JM. A new approach to causal inference in mortality studies with a sustained exposure period-application to control of the healthy worker survivor effect. Math Modell. 1986;7:1393–1512.10.1016/0270-0255(86)90088-6Search in Google Scholar
24. Richardson TS, Robins JM. Single world intervention graphs (SWIGs): A unification of the counterfactual and graphical approaches to causality. Center for the Statistics and the Social Sciences, University of Washington, Series Working Paper 2013.Search in Google Scholar
25. Nelsen RB. An introduction to copulas. New York, NY: Springer Science & Business Media; 2007.Search in Google Scholar
26. Klee V, Minty GJ. How good is the simplex algorithm. Technical report, DTIC Document 1970.Search in Google Scholar
27. Schrijver A. Theory of linear and integer programming. Chichester, UK: John Wiley & Sons; 1998.Search in Google Scholar
28. Tang MW, Kanki PJ, Shafer RW. A review of the virological efficacy of the 4 World Health Organization–recommended tenofovir-containing regimens for initial HIV therapy. Clin Infect Dis. 2012;54:862–875.10.1093/cid/cir1034Search in Google Scholar PubMed PubMed Central
29. Tamer E. Partial identification in econometrics. Ann Rev Econ. 2010;2:167–195.10.1146/annurev.economics.050708.143401Search in Google Scholar
30. Chernozhukov V, Hong H, Tamer E. Estimation and confidence regions for parameter sets in econometric models. Econometrica. 2007;75:1243–1284.10.1111/j.1468-0262.2007.00794.xSearch in Google Scholar
31. Romano JP, Shaikh AM. Inference for identifiable parameters in partially identified econometric models. J Stat Plan Inference. 2008;138:2786–2807.10.1016/j.jspi.2008.03.015Search in Google Scholar
32. Andrews DW, Guggenberger P. Validity of subsampling and “plug-in asymptotic” inference for parameters defined by moment inequalities. Econ Theory. 2009;25:669–709.10.1017/S0266466608090257Search in Google Scholar
33. van der Vaart AW, Wellner JA. Weak convergence and empirical processes. New York: Springer 1996.10.1007/978-1-4757-2545-2Search in Google Scholar
34. Tchetgen Tchetgen EJ. On causal mediation analysis with a survival outcome. Int J Biostat. 2011;7:1–38.10.2202/1557-4679.1351Search in Google Scholar PubMed PubMed Central
35. Vansteelandt S, VanderWeele TJ. Natural direct and indirect effects on the exposed: Effect decomposition under weaker assumptions. Biometrics. 2012;68:1019–1027.10.1111/j.1541-0420.2012.01777.xSearch in Google Scholar PubMed PubMed Central
© 2017 Walter de Gruyter GmbH, Berlin/Boston
This article is distributed under the terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.