Marginal Structural Models and Causal Inference in Epidemiology

Robins, James M.; Hernán, Miguel A.; Brumback, Babette

doi:10.1097/00001648-200009000-00011

Cited by 4,486 publications

(4,908 citation statements)

References 16 publications

Supporting

Mentioning

4,785

Contrasting

Unclassified

Order By: Relevance

“…In reality, when a factor impacts an outcome through both direct and indirect pathways, we cannot observe the direct effect in absence of the indirect effect, and vice versa; their estimation depends on counterfactual comparisons (Robins 2003). A general counterfactual model has been proposed that permits the estimation of total and direct effects of fixed and time-varying exposures in longitudinal studies whether randomized or observational in design (Robins et al 2000). However, a more detailed discussion is beyond the scope of this paper.…”

Section: Discussionmentioning

confidence: 99%

“…In Figure 1D, PCBs are assumed to affect both serum lipids and the outcome, creating a spurious association (Robins et al 2000). Here, only an unadjusted model is appropriate for risk estimation.…”

Section: Methodsmentioning

confidence: 99%

“…Here we demonstrate the impact of model (mis) specification and its effect on the interpretation of study findings. We used directed acyclic graphs (DAGs) to define a causal framework among exposure, lipids, and health outcome and values for parameters as informed by the literature (Hernan et al 2002; Robins et al 2000). Using DAGs to supply a causal framework and parameter values informed by the literature, we present the results of a simulation study.…”

mentioning

confidence: 99%

See 2 more Smart Citations

Lipid Adjustment in the Analysis of Environmental Contaminants and Human Health Risks

Schisterman

Whitcomb

Louis

et al. 2005

Environ Health Perspect

346

252

View full text Add to dashboard Cite

The literature on exposure to lipophilic agents such as polychlorinated biphenyls (PCBs) is conflicting, posing challenges for the interpretation of potential human health risks. Laboratory variation in quantifying PCBs may account for some of the conflicting study results. For example, for quantification purposes, blood is often used as a proxy for adipose tissue, which makes it necessary to model serum lipids when assessing health risks of PCBs. Using a simulation study, we evaluated four statistical models (unadjusted, standardized, adjusted, and two-stage) for the analysis of PCB exposure, serum lipids, and health outcome risk (breast cancer). We applied eight candidate true causal scenarios, depicted by directed acyclic graphs, to illustrate the ramifications of misspecification of underlying assumptions when interpreting results. Statistical models that deviated from underlying causal assumptions generated biased results. Lipid standardization, or the division of serum concentrations by serum lipids, was observed to be highly prone to bias. We conclude that investigators must consider biology, biologic medium (e.g., nonfasting blood samples), laboratory measurement, and other underlying modeling assumptions when devising a statistical plan for assessing health outcomes in relation to environmental exposures.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

mentioning

confidence: 99%

See 1 more Smart Citation

Lipid Adjustment in the Analysis of Environmental Contaminants and Human Health Risks

Schisterman

Whitcomb

Louis

et al. 2005

Environ Health Perspect

346

252

View full text Add to dashboard Cite

show abstract

“…This weighted pooled logistic regression model approximates a marginal structural Cox regression model when certain conditions are met (short intervals and low event rates) 33, 34. The weights from the four components above were multiplied to obtain the single time‐updated weight incorporated into this model.…”

Section: Methodsmentioning

confidence: 99%

Increased non‐AIDS mortality among persons with AIDS‐defining events after antiretroviral therapy initiation

et al. 2018

View full text Add to dashboard Cite

Introduction HIV‐1 infection leads to chronic inflammation and to an increased risk of non‐AIDS mortality. Our objective was to determine whether AIDS‐defining events (ADEs) were associated with increased overall and cause‐specific non‐AIDS related mortality after antiretroviral therapy (ART) initiation.MethodsWe included HIV treatment‐naïve adults from the Antiretroviral Therapy Cohort Collaboration (ART‐CC) who initiated ART from 1996 to 2014. Causes of death were assigned using the Coding Causes of Death in HIV (CoDe) protocol. The adjusted hazard ratio (aHR) for overall and cause‐specific non‐AIDS mortality among those with an ADE (all ADEs, tuberculosis (TB), Pneumocystis jiroveci pneumonia (PJP), and non‐Hodgkin's lymphoma (NHL)) compared to those without an ADE was estimated using a marginal structural model.ResultsThe adjusted hazard of overall non‐AIDS mortality was higher among those with any ADE compared to those without any ADE (aHR 2.21, 95% confidence interval (CI) 2.00 to 2.43). The adjusted hazard of each of the cause‐specific non‐AIDS related deaths were higher among those with any ADE compared to those without, except metabolic deaths (malignancy aHR 2.59 (95% CI 2.13 to 3.14), accident/suicide/overdose aHR 1.37 (95% CI 1.05 to 1.79), cardiovascular aHR 1.95 (95% CI 1.54 to 2.48), infection aHR (95% CI 1.68 to 2.81), hepatic aHR 2.09 (95% CI 1.61 to 2.72), respiratory aHR 4.28 (95% CI 2.67 to 6.88), renal aHR 5.81 (95% CI 2.69 to 12.56) and central nervous aHR 1.53 (95% CI 1.18 to 5.44)). The risk of overall and cause‐specific non‐AIDS mortality differed depending on the specific ADE of interest (TB, PJP, NHL).ConclusionsIn this large multi‐centre cohort collaboration with standardized assignment of causes of death, non‐AIDS mortality was twice as high among patients with an ADE compared to without an ADE. However, non‐AIDS related mortality after an ADE depended on the ADE of interest. Although there may be unmeasured confounders, these findings suggest that a common pathway may be independently driving both ADEs and NADE mortality. While prevention of ADEs may reduce subsequent death due to NADEs following ART initiation, modification of risk factors for NADE mortality remains important after ADE survival.

show abstract

“…It was especially important to include the adjustment variable “time from diagnosis to intervention,” because large differences in this variable (which may also serve as a proxy for residual disease) may confound the results. The inverse probability of treatment weighting method used propensity scores to build stabilized weights that balanced the distribution of covariates between intervention groups while preserving sample size 12, 13…”

Section: Introductionmentioning

confidence: 99%

Overall survival with ponatinib versus allogeneic stem cell transplantation in Philadelphia chromosome‐positive leukemias with the T315I mutation

Nicolini

Basak

Kim

et al. 2017

Cancer

View full text Add to dashboard Cite

BACKGROUNDEffective treatment options for patients with chronic myeloid leukemia (CML) or Philadelphia‐positive (Ph+) acute lymphoblastic leukemia (ALL) who have the threonine to isoleucine mutation at codon 315 (T315I) are few. The objective of this study was to compare overall survival (OS) between patients with CML and those with Ph+ ALL who received treatment with ponatinib versus allogeneic stem cell transplantation (allo‐SCT).METHODSA post hoc, retrospective, indirect comparison of OS among patients who received single‐agent ponatinib in the Ponatinib Ph+ ALL and CML Evaluation (PACE) trial with those who underwent allo‐SCT as reported to the European Bone Marrow Transplant registry, stratified by CML disease phase and Ph+ ALL, was conducted. Kaplan‐Meier survival curves and multivariate Cox proportional‐hazards models were used to compare OS between intervention groups, adjusting for time from diagnosis to intervention, age, sex, and geographic region; 24‐month and 48‐month OS rates and median OS were reported.RESULTSAfter adjustment for potential confounders, 24‐month and 48‐month OS rates were significantly higher in patients with chronic‐phase CML (CP‐CML) who received ponatinib compared with those who underwent allo‐SCT (24 months: 84% vs 60.5%, respectively; P = .004; 48 months: 72.7% vs 55.8%, respectively; P = .013), with a hazard ratio (HR) of 0.37 (95% confidence interval [CI], 0.16‐0.84; P = .017). In patients who had accelerated‐phase CML, OS rates were not significantly different between the groups (HR, 0.90; 95% CI, 0.20‐4.10; P = .889). In patients who had blast‐crisis CML and those with Ph+ ALL, ponatinib was associated with shorter OS compared with allo‐SCT (blast‐crisis CML: HR, 2.29 [95% CI, 1.08‐4.82; P = .030]; Ph+ ALL: HR, 2.77 [95% CI, 0.73‐10.56; P = .146]).CONCLUSIONSAlthough allo‐SCT remains an important treatment option for patients with T315I‐positive advanced CML and Ph+ ALL, ponatinib represents a valuable alternative for patients with T315I‐positive CP‐CML. Cancer 2017;123:2875–80. © 2017 American Cancer Society.

show abstract

Marginal Structural Models and Causal Inference in Epidemiology

Cited by 4,486 publications

References 16 publications

Lipid Adjustment in the Analysis of Environmental Contaminants and Human Health Risks

Lipid Adjustment in the Analysis of Environmental Contaminants and Human Health Risks

Increased non‐AIDS mortality among persons with AIDS‐defining events after antiretroviral therapy initiation

Overall survival with ponatinib versus allogeneic stem cell transplantation in Philadelphia chromosome‐positive leukemias with the T315I mutation

Contact Info

Product

Resources

About