Search

In assessing the efficacy of a time-varying treatment structural nested models (SNMs) are useful in dealing with confounding by variables affected by earlier treatments. These models often consider treatment allocation and repeated measures at the individual level. We extend SNMMs to clustered observations with time-varying confounding and treatments. We demonstrate how to formulate models with both cluster- and unit-level treatments and show how to derive semiparametric estimators of parameters in such models. For unit-level treatments, we consider interference, namely the effect of treatment on outcomes in other units of the same cluster. The properties of estimators are evaluated through simulations and compared with the conventional GEE regression method for clustered outcomes. To illustrate our method, we use data from the treatment arm of a glaucoma clinical trial to compare the effectiveness of two commonly used ocular hypertension medications.

Various indicators of progression of chronic kidney disease (CKD) have been used as outcomes in clinical research studies. The effect of using varying measures on the association of risk factors with CKD progression has not been well characterized.

Abstract: Statistical analysis of longitudinal outcomes is often complicated by the absence of observable values in patients who die prior to their scheduled measurement. In such cases, the longitudinal data are said to be “truncated by death” to emphasize that the longitudinal measurements are not simply missing, but are undefined after death. Recently, the truncation by death problem has been investigated using the framework of principal stratification to define the target estimand as the survivor average causal effect (SACE), which in the context of a two-group randomized clinical trial is the mean difference in the longitudinal outcome between the treatment and control groups for the principal stratum of always-survivors. The SACE is not identified without untestable assumptions. These assumptions have often been formulated in terms of a monotonicity constraint requiring that the treatment does not reduce survival in any patient, in conjunction with assumed values for mean differences in the longitudinal outcome between certain principal strata. In this paper, we introduce an alternative estimand, the balanced-SACE, which is defined as the average causal effect on the longitudinal outcome in a particular subset of the always-survivors that is balanced with respect to the potential survival times under the treatment and control. We propose a simple estimator of the balanced-SACE that compares the longitudinal outcomes between equivalent fractions of the longest surviving patients between the treatment and control groups and does not require a monotonicity assumption. We provide expressions for the large sample bias of the estimator, along with sensitivity analyses and strategies to minimize this bias. We consider statistical inference under a bootstrap resampling procedure.

Objective.To evaluate the outcome of tumor necrosis factor-α inhibition (anti-TNF) for pediatric uveitis.Methods.We retrospectively assessed children (age ≤ 18 yrs) with noninfectious uveitis receiving anti-TNF at 5 uveitis centers and 1 pediatric rheumatology center. Incident treatment success was defined as minimal or no uveitis activity at ≥ 2 consecutive ophthalmological examinations ≥ 28 days apart while taking no oral and ≤ 2 eyedrops/day of corticosteroids. Eligible children had active uveitis and/or were taking higher corticosteroid doses.Results.Among 56 eligible children followed over 33.73 person-years, 52% had juvenile idiopathic arthritis (JIA) and 75% had anterior uveitis (AU). The Kaplan-Meier estimated proportion achieving treatment success within 12 months was 75% (95% CI 62%–87%). Complete absence of inflammatory signs with discontinuation of all corticosteroids was observed in an estimated 64% by 12 months (95% CI 51%–76%). Diagnoses of JIA or AU were associated with greater likelihood of success, as was the oligoarticular subtype among JIA cases. In a multivariable model, compared to those with JIA-associated AU, those with neither or with JIA or AU alone had a 75%–80% lower rate of achieving quiescence under anti-TNF, independent of the number of immunomodulators previously or concomitantly prescribed. Uveitis reactivated within 12 months of achieving quiescence in 14% of those continuing anti-TNF (95% CI 6%–31%). The incidence of discontinuation for adverse effects was 8%/year (95% CI 1%–43%).Conclusion.Treatment with anti-TNF was successful and sustained in a majority of children with noninfectious uveitis, and treatment-limiting toxicity was infrequent. JIA-associated AU may be especially responsive to anti-TNF.

Glomerular filtration rate (GFR) is considered the best measure of kidney function, but repeated assessment is not feasible in most research studies.

Excessive daytime sleepiness (EDS) is associated with increased mortality in older adults, yet sleep disordered breathing (SDB), a common cause of sleepiness, has not been shown to increase mortality in older adults. This study examined the relationship between daytime sleepiness, SDB, self-report sleep parameters, and mortality in older adults.

Coronary artery calcification (CAC) and metabolic syndrome (MS) have been associated with increased cardiovascular risk. The study objective was to examine the association of MS with CAC presence and progression in renal transplant recipients.

We consider a method for extending instrumental variables methods in order to estimate the overall effect of a treatment or exposure. The approach is designed for settings in which the instrument influences both the treatment of interest and a secondary treatment also influenced by the primary treatment. We demonstrate that, while instrumental variables methods may be used to estimate the joint effects of the primary and secondary treatments, they cannot by themselves be used to estimate the overall effect of the primary treatment. However, instrumental variables methods may be used in conjunction with approaches for estimating the effect of the primary on the secondary treatment to estimate the overall effect of the primary treatment. We consider extending the proposed methods to deal with confounding of the effect of the instrument, mediation of the effect of the instrument by other variables, failure-time outcomes, and time-varying secondary treatments. We motivate our discussion by considering estimation of the overall effect of the type of vascular access among hemodialysis patients.

Delayed time to antibiotic administration has been linked with higher mortality for patients with community-acquired pneumonia, but the impact of antibiotic resistance on clinical outcomes has been controversial. In the current study we assess the combined impact of antibiotic resistance and antibiotic timing on outcomes, including inhospital mortality, complications, length of stay, and time to stability, for patients hospitalized with community-acquired bacteremic pneumococcal pneumonia. We conducted a retrospective cohort study in 43 hospitals in the Southeastern Pennsylvania region from 2001 to 2004. Eligible adult patients had pneumococcal bacteremia and radiographic evidence of pneumonia. Outcomes were assessed based on medical record review. Multivariable regression was used to adjust for severity of illness and sequentially assess the impact of antibiotic resistance and time to active antibiotic therapy. The overall inhospital mortality was 10%. Overall, levels of macrolide, cephalosporin, and fluoroquinolone resistance were low and did not adversely impact the time to administration of active antibiotic therapy. Receipt of at least 1 active antibiotic within 4 hours was associated with reduced mortality (odds ratio OR, 0.47; 95% confidence interval CI, 0.2-1.0) and shortened length of stay (OR, 0.77; CI, 0.6-1.0) but did not reduce the risk of other adverse outcomes. We conclude that early antibiotic administration reduces the risks of mortality in patients with bacteremic pneumococcal pneumonia. Current patterns of drug resistance did not lead to delays in administration of active antimicrobial therapy.

Polymorphisms of genes such as angiotensin-converting enzyme (ACE), angiotensinogen (AGT), and angiotensin receptor type I (AGTR1) have been associated with hypertension. Hypertension, in turn, has been associated with decreased renal allograft survival. Therefore, this study investigated whether single nucleotide polymorphisms (SNPs) in these genes are associated with decline in renal function posttransplantation.

Background:Failure to complete the medical evaluation for renal transplantation may impede access to transplantation and preclude the possibility of preemptive transplantation. We sought to (1) characterize completion rates of the transplantation medical evaluation and (2) determine factors associated with completion of the evaluation. We hypothesized that patients not on dialysis therapy complete the evaluation process more quickly than patients receiving dialysis. Methods:Between September 2002 and September 2003, a total of 175 patients who were evaluated for renal transplantation at the Hospital of the University of Pennsylvania were enrolled in a prospective cohort study. Patients completed a self-administered questionnaire. The progress of patients’ medical evaluations, including completion of requested tests and evaluations, was extracted from the electronic medical record. Results:During follow-up, 100 patients (57.1%) completed the evaluation, including tests and evaluations requested by the transplant team, whereas 49 patients (28.0%) had tests still pending. The remaining patients died (2.3%), lost interest in transplantation (1.1%), or were immediately (7.4%) or later (4.0%) declared medically ineligible for transplantation. In the multivariable Cox proportional hazards model, black race (adjusted hazard ratio, 0.63; 95% confidence interval, 0.40 to 1.00; P= 0.05) was associated with time to completion of the transplantation evaluation, but receiving maintenance dialysis at the time of the initial transplantation evaluation was not (adjusted hazard ratio, 0.92; 95% confidence interval, 0.60 to 1.42; P= 0.72). Conclusion:Completion of the medical evaluation for transplantation is slower in blacks than nonblacks. We were unable to detect a significant difference between dialysis and nondialysis patients in rates of completion of the evaluation.

Background:Understanding the factors associated with successful arteriovenous fistula (AVF) maturation may provide an opportunity to improve vascular access outcomes. Methods:This prospective cohort study enrolled 348 hemodialysis patients receiving an AVF and followed them until ascertainment of successful AVF maturation. Multivariable logistic regression was used to develop (1) explanatory models of factors associated with AVF success and (2) models for predicting success based on sociodemographic and comorbidity profiles. Receiver operating characteristic (ROC) curve analysis was performed to evaluate the discriminative ability of our predictive model. Predictive probabilities of AVF maturation under the assumption of optimal surgical technique defined by explanatory models were calculated. Results:A total of 55.5% of AVFs successfully matured. A history of stroke or transient ischemic attack, increasing age, and dependence on dialysis at the time of access placement were associated with lower probabilities of maturation. Of potentially modifiable variables, maturation was associated with greater intraoperative doses of heparin, use of large-diameter veins, and mean arterial pressure of 85 mm Hg or greater. Using optimal surgical technique, the probability of successful AVF maturation would have been as high as 84%. A predictive logistic regression model had moderate ability to predict AVF maturation under optimal surgical therapy with an area under the ROC curve of 0.69. Conclusion:Surgical technique potentially may be modified to yield a greater probability of successful AVF maturation. Predictive probabilities of successful AVF maturation under optimal surgical technique reinforce the view that functioning AVFs can be achieved in a large proportion of hemodialysis patients, consistent with experience in selected regions of the United States and abroad.

The compliance score in randomized trials is a measure of the effect of randomization on treatment received. It is in principle a group-level pretreatment variable and so can be used where individual-level measures of treatment received can produce misleading inferences. The interpretation of models with the compliance score as a regressor of interest depends on the link function. Using the identity link can lead to valid inference about the effects of treatment received even in the presence of nonrandom noncompliance; such inference is more problematic for nonlinear links. We illustrate these points with data from two randomized trials.

Objective:Synthetic vascular grafts, the most common type of permanent dialysis graft used in the United States, are associated with great morbidity and expense. In this cohort study of patients undergoing hemodialysis and receiving a new synthetic vascular graft, detailed intraoperative procedural data were examined as predictors of graft dysfunction and failure. Methods:A prospective cohort study of patients receiving hemodialysis who had undergone synthetic (polytetrafluoroethylene) graft placement over 4½ years were followed up until either the graft was no longer usable, kidney transplantation was performed, the patient died, or the study ended. The principal methods of analysis used were the Kaplan-Meier method, to compute survival function for primary and secondary graft patency, and Cox proportional hazards regression, to model associations between predictor variables and graft failure. Results:Of 284 patients enrolled in the study, 172 patients (61%) had at least one graft-related event, ie, temporary graft dysfunction or graft failure, during follow-up. Three-year cumulative graft survival was 54% (95% confidence interval, 45%-62%). Using proportional hazards analysis, history of claudication (rate ratio [RR], 2.14 [range, 0.97-4.73]; P=.06), number of previous permanent grafts (1 graft: RR, 1.49 [range, 0.88-2.51]; 2 or more grafts: RR, 2.85 [range, 1.43-5.69]; P=.01), dialysis dependency at surgery (RR, 2.96 [range, 1.23-7.12]; P=.02), and use of arterial clamps in construction of the graft (RR, 2.32 [range, 1.14-4.73]; P=.02) were associated with lower survival, even after accounting for medical history. Type of graft material, ie, Gore-Tex versus other material (RR, 0.28 [range, 0.16-0.50]; P<.01) and use of the axillary vein (RR, 0.61 [range, 0.36- 1.02]; P=.06) for the access site were associated with a lower rate of graft failure. Similar results were found with Poisson analysis of all graft-related events. In addition, acute arterial anastomosis, ie, arterial angle less than 90 degrees (RR, 0.63 [range, 0.45-0.91]; P=.01) and use of the brachial artery versus the radial artery (RR, 0.54 [range, 0.33-0.54]; P=.01) were associated with a lower rate of graft-related events. Conclusion:These findings suggest possible areas for intervention to improve synthetic vascular graft outcome. (J Vasc Surg 2003;37:1036-42.)

The relationship between transplantation prior to chronic dialysis initiation and the pattern of acute rejection of kidneys from living donors (LDKT) has not been fully explored.

Administrative censoring, in which potential censoring times are known even for subjects who fail, is common in clinical and epidemiologic studies. Nonetheless, most statistical methods for failure‐time data do not use the information contained in these potential censoring times. Robins has proposed two approaches for using this information to estimate parameters in an accelerated failure‐time model; the methods generally require the analyst to treat as censored some subjects whose failure time is observed. This paper provides a rationale for this ‘artificial censoring’, discusses some of its consequences, and illustrates some of these points with data from a randomized trial of breast cancer screening. Copyright © 2001 John Wiley & Sons, Ltd.

Administrative censoring, in which potential censoring times are known even for subjects who fail, is common in clinical and epidemiologic studies. Nonetheless, most statistical methods for failure-time data do not use the information contained in these potential censoring times. Robins has proposed two approaches for using this information to estimate parameters in an accelerated failure-time model; the methods generally require the analyst to treat as censored some subjects whose failure time is observed. This paper provides a rationale for this ‘artificial censoring’, discusses some of its consequences, and illustrates some of these points with data from a randomized trial of breast cancer screening. Copyright © 2001 John Wiley & Sons, Ltd.

Previous investigators have suggested that screening-related biases may explain associations between postmenopausal hormone use and breast cancer. To investigate these biases, we studied postmenopausal women in the Nurses’ Health Study from 1988 to 1994. Hormone use is associated with increased subsequent screening. Among women not screened in the previous 2 years, the probability difference, comparing current hormone users with others, for having mammography in the following 2 years is 19.5; among women previously screened, the difference is 4.9. These differences persist after control for other factors. If the increase in screening is causal, screening by mammogram could be intermediate in the causal pathway to breast cancer diagnosis. To deal with this problem, we restrict attention to a subset of the cohort in which the effect of postmenopausal hormone use on screening is small (women previously screened). In this subset, the rate ratio comparing breast cancer rates among current postmenopausal hormone users with others is 1.28. In a sensitivity analysis, the bias could not by itself plausibly account for the associations in our data. Our data provide evidence of an association between postmenopausal hormone use and breast cancer that is not solely the product of a detection bias.

To review conceptual issues regarding confounding by indication in the context of studies of calcium channel blockers (CCBs). Review of literature, with special attention to two articles in the current issue. Conflicting arguments about the presence of uncontrollable confounding by indication in studies of CCBs are reviewed and criticized. Studies with potential confounding by indication can benefit from appropriate analytic methods, including separating the effects of a drug taken at different times, sensitivity analysis for unmeasured confounders, and instrumental variables and G-estimation. Whether confounding by indication accounts for observed associations is often difficult to determine; this is the case in studies of CCBs. When confounding by indication is suspected, a variety of methods to deal with it may be useful. Copyright © 2000 John Wiley & Sons, Ltd.

Researchers commonly express scepticism about using observational data to estimate the effect of a treatment on an outcome the treatment is intended to affect. In this paper, we consider using data from the Multicenter AIDS Cohort Study (MACS) to determine whether zidovudine prevents the development of Kaposi's sarcoma among HIV‐positive gay men. Several methodologic issues common to observational data characterized the study: information on potentially important confounders was missing at some study visits; investigators did not always know the time of changes in treatment level, nor the value of confounders at that time, and the censoring process depended strongly on time‐varying covariates related to outcome. We describe application to our data of Robins' paradigm for defining, modelling and estimating the effect of a time‐varying treatment and show how to modify his approach to deal with the methodologic issues we have mentioned. Further, we demonstrate that relative risk regression is less well equipped to deal with these issues. We compare our results to the findings from randomized trials, and conclude that observational studies may sometimes be useful in evaluating the effect of treatment on an intended outcome. © 1998 John Wiley & Sons, Ltd.

The direct effect of a treatment on some outcome is that part of the treatment's effect not referred through a specified covariate intermediate on the pathway between treatment and outcome. Such direct effects are often of primary interest in a data analysis. Unfortunately, standard methods of analysis (for example, stratification or modelling) do not, in general, produce consistent estimates of direct effects whether or not the covariate is ‘controlled’. Robins and co‐authors have proposed two methods for estimation of direct effects applicable when reliable information is available on the covariate. We propose a third approach for reducing bias: data restriction. By restricting the analysis to strata of the data in which the effect of treatment on the covariate is small, we can (under certain assumptions) reduce bias in estimating treatment's direct effect. We discuss these points with reference to difference and ratio measures of treatment effect. The approach will sometimes be applicable even with an unmeasured or poorly measured covariate. We illustrate these points with data from an observational study of the effect of hormone replacement therapy on breast cancer. © 1998 John Wiley & Sons, Ltd.

The direct effect of a treatment on some outcome is that part of the treatment’s effect not referred through a specified covariate intermediate on the pathway between treatment and outcome. Such direct effects are often of primary interest in a data analysis. Unfortunately, standard methods of analysis (for example, stratification or modelling) do not, in general, produce consistent estimates of direct effects whether or not the covariate is ‘controlled’. Robins and co-authors have proposed two methods for estimation of direct effects applicable when reliable information is available on the covariate. We propose a third approach for reducing bias: data restriction. By restricting the analysis to strata of the data in which the effect of treatment on the covariate is small, we can (under certain assumptions) reduce bias in estimating treatment’s direct effect. We discuss these points with reference to difference and ratio measures of treatment effect. The approach will sometimes be applicable even with an unmeasured or poorly measured covariate. We illustrate these points with data from an observational study of the effect of hormone replacement therapy on breast cancer. © 1998 John Wiley & Sons, Ltd.

Researchers commonly express scepticism about using observational data to estimate the effect of a treatment on an outcome the treatment is intended to affect. In this paper, we consider using data from the Multicenter AIDS Cohort Study (MACS) to determine whether zidovudine prevents the development of Kaposi's sarcoma among HIV-positive gay men. Several methodologic issues common to observational data characterized the study: information on potentially important confounders was missing at some study visits; investigators did not always know the time of changes in treatment level, nor the value of confounders at that time, and the censoring process depended strongly on time-varying covariates related to outcome. We describe application to our data of Robins’ paradigm for defining, modelling and estimating the effect of a time-varying treatment and show how to modify his approach to deal with the methodologic issues we have mentioned. Further, we demonstrate that relative risk regression is less well equipped to deal with these issues. We compare our results to the findings from randomized trials, and conclude that observational studies may sometimes be useful in evaluating the effect of treatment on an intended outcome. © 1998 John Wiley & Sons, Ltd.

We address several questions relating to the use of standard regression and Structural Nested Mean Model (SNMM) approach (e. g., Ten Have et al. 2007) to analyze post-randomization effect modifiers of the intent-to-treat effect of a randomized intervention on a subsequent outcome, which has not been well examined. We show through simulations that the SNMM performs better with respect to bias of estimates of the intervention and interaction effects than does the corresponding standard interaction approach when the baseline intervention is randomized and the post-randomization factors are subject to confounding, and even when there is no association between the intervention and effect modifier. However, causal inference under the SNMM makes untestable assumptions that the causal contrasts do not vary across observed levels of the intervention and post-randomization factor. In addition, the precision of the SNMM-based estimators depends on the effect of the randomized intervention on the post-randomization factor varying across baseline covariate combinations. These issues and methods are illustrated with the application of the standard and causal methods to a randomized cognitive therapy (CT) trial, for which there is a conceptual model of negative cognitive styles or distortions impacted by CT but then in turn modifying the effect of CT on subsequent suicide ideation and social problem solving outcomes.

In randomized controlled trials, the evaluation of an overall treatment effect is often followed by effect modification or subgroup analyses, where the possibility of a different magnitude or direction of effect for varying values of a covariate is explored. While studies of effect modification are typically restricted to pretreatment covariates, longitudinal experimental designs permit the examination of treatment effect modification by intermediate outcomes, where intermediates are measured after treatment but before the final outcome. We present a novel application of generalized structural mean models (GSMMs) for simultaneously assessing effect modification by post-treatment covariates and accounting for noncompliance to assigned treatment status. The proposed approach may also be used to identify post-treatment effect modifiers in the absence of noncompliance. The methods are evaluated using a simulation study that demonstrates that our approach retains consistent estimation of effect modification by intermediate variables that are affected by treatment and also predict outcomes. We illustrate the method using a randomized trial designed to promote re-employment through teaching skills to enhance self-esteem and inoculate job seekers against setbacks in the job search process. Our analysis provides some evidence that the intervention was much less successful among subjects that displayed higher levels of depression at intermediate post-treatment waves of the study. We also compare the assumptions of our approach and principal stratification as alternatives to account for differences in effects by intermediate variables.

Pearl's article provides a useful springboard for discussing further the benefits and drawbacks of principal stratification and the associated discomfort with attributing effects to post-treatment variables. The basic insights of the approach are important: pay close attention to modification of treatment effects by variables not observable before treatment decisions are made, and be careful in attributing effects to variables when counterfactuals are ill-defined. These insights have often been taken too far in many areas of application of the approach, including instrumental variables, censoring by death, and surrogate outcomes. A novel finding is that the usual principal stratification estimand in the setting of censoring by death is by itself of little practical value in estimating intervention effects.

Pearl’s article provides a useful springboard for discussing further the benefits and drawbacks of principal stratification and the associated discomfort with attributing effects to post-treatment variables. The basic insights of the approach are important: pay close attention to modification of treatment effects by variables not observable before treatment decisions are made, and be careful in attributing effects to variables when counterfactuals are ill-defined. These insights have often been taken too far in many areas of application of the approach, including instrumental variables, censoring by death, and surrogate outcomes. A novel finding is that the usual principal stratification estimand in the setting of censoring by death is by itself of little practical value in estimating intervention effects.

Most attempts at causal inference in observational studies are based on assumptions that treatment assignment is ignorable. Such assumptions are usually made casually, largely because they justify the use of available statistical methods and not because they are truly believed. It will often be the case that it is plausible that conditional independence holds at least approximately for a subset but not all of the experience giving rise to one's data. Such selective ignorability assumptions may be used to derive valid causal inferences in conjunction with structural nested models. In this paper, we outline selective ignorability assumptions mathematically and sketch how they may be used along with otherwise standard G-estimation or likelihood-based methods to obtain inference on structural nested models. We also consider use of these assumptions in the presence of selective measurement error or missing data when the missingness is not at random. We motivate and illustrate our development by considering an analysis of an observational database to estimate the effect of erythropoietin use on mortality among hemodialysis patients.

Most attempts at causal inference in observational studies are based on assumptions that treatment assignment is ignorable. Such assumptions are usually made casually, largely because they justify the use of available statistical methods and not because they are truly believed. It will often be the case that it is plausible that conditional independence holds at least approximately for a subset but not all of the experience giving rise to one's data. Such selective ignorability assumptions may be used to derive valid causal inferences in conjunction with structural nested models. In this paper, we outline selective ignorability assumptions mathematically and sketch how they may be used along with otherwise standard G-estimation or likelihood-based methods to obtain inference on structural nested models. We also consider use of these assumptions in the presence of selective measurement error or missing data when the missingness is not at random. We motivate and illustrate our development by considering an analysis of an observational database to estimate the effect of erythropoietin use on mortality among hemodialysis patients.

We consider a method for extending instrumental variables methods in order to estimate the overall effect of a treatment or exposure. The approach is designed for settings in which the instrument influences both the treatment of interest and a secondary treatment also influenced by the primary treatment. We demonstrate that, while instrumental variables methods may be used to estimate the joint effects of the primary and secondary treatments, they cannot by themselves be used to estimate the overall effect of the primary treatment. However, instrumental variables methods may be used in conjunction with approaches for estimating the effect of the primary on the secondary treatment to estimate the overall effect of the primary treatment. We consider extending the proposed methods to deal with confounding of the effect of the instrument, mediation of the effect of the instrument by other variables, failure-time outcomes, and time-varying secondary treatments. We motivate our discussion by considering estimation of the overall effect of the type of vascular access among hemodialysis patients.

Marginal structural models (MSM) provide a powerful tool for estimating the causal effect of a treatment. These models, introduced by Robins, model the marginal distributions of treatment-specific counterfactual outcomes, possibly conditional on a subset of the baseline covariates. Marginal structural models are particularly useful in the context of longitudinal data structures, in which each subject's treatment and covariate history are measured over time, and an outcome is recorded at a final time point. However, the utility of these models for some applications has been limited by their inability to incorporate modification of the causal effect of treatment by time-varying covariates. Particularly in the context of clinical decision making, such time-varying effect modifiers are often of considerable or even primary interest, as they are used in practice to guide treatment decisions for an individual. In this article we propose a generalization of marginal structural models, which we call history-adjusted marginal structural models (HA-MSM). These models allow estimation of adjusted causal effects of treatment, given the observed past, and are therefore more suitable for making treatment decisions at the individual level and for identification of time-dependent effect modifiers. Specifically, a HA-MSM models the conditional distribution of treatment-specific counterfactual outcomes, conditional on the whole or a subset of the observed past up till a time-point, simultaneously for all time-points. Double robust inverse probability of treatment weighted estimators have been developed and studied in detail for standard MSM. We extend these results by proposing a class of double robust inverse probability of treatment weighted estimators for the unknown parameters of the HA-MSM. In addition, we show that HA-MSM provide a natural approach to identifying the dynamic treatment regimen which follows, at each time-point, the history-adjusted (up till the most recent time point) optimal static treatment regimen. We illustrate our results using an example drawn from the treatment of HIV infection.

Marginal structural models (MSM) provide a powerful tool for estimating the causal effect of a treatment. These models, introduced by Robins, model the marginal distributions of treatment-specific counterfactual outcomes, possibly conditional on a subset of the baseline covariates. Marginal structural models are particularly useful in the context of longitudinal data structures, in which each subject's treatment and covariate history are measured over time, and an outcome is recorded at a final time point. However, the utility of these models for some applications has been limited by their inability to incorporate modification of the causal effect of treatment by time-varying covariates. Particularly in the context of clinical decision making, such time-varying effect modifiers are often of considerable or even primary interest, as they are used in practice to guide treatment decisions for an individual. In this article we propose a generalization of marginal structural models, which we call history-adjusted marginal structural models (HA-MSM). These models allow estimation of adjusted causal effects of treatment, given the observed past, and are therefore more suitable for making treatment decisions at the individual level and for identification of time-dependent effect modifiers. Specifically, a HA-MSM models the conditional distribution of treatment-specific counterfactual outcomes, conditional on the whole or a subset of the observed past up till a time-point, simultaneously for all time-points. Double robust inverse probability of treatment weighted estimators have been developed and studied in detail for standard MSM. We extend these results by proposing a class of double robust inverse probability of treatment weighted estimators for the unknown parameters of the HA-MSM. In addition, we show that HA-MSM provide a natural approach to identifying the dynamic treatment regimen which follows, at each time-point, the history-adjusted (up till the most recent time point) optimal static treatment regimen. We illustrate our results using an example drawn from the treatment of HIV infection.

Despite much investigation of zidovudine, little has been reported regarding its effect on the development of most individual AIDS-defining illnesses, including Kaposi's sarcoma (KS). We used observational data from the Multicenter AIDS Cohort Study (MACS) to estimate the effect of zidovudine use on the subsequent incidence of KS. To do this, we examined and adjusted for predictors of zidovudine use. CD4 lymphocyte counts, the development of HIV-related symptoms and AIDS, and changes in these factors were important predictors of zidovudine use. We used these associations to control for confounding by these and other factors with the G-estimation approach. We found no evidence that zidovudine use affected the time to KS in the MACS; the point estimate (95% confidence interval [CI]) for increase in time to KS was zero (−28%–68%). The relative risk was 1.0 (95% CI, 0.54–1.84). Randomized trials suggest that zidovudine may prevent KS. We discuss possible explanations for differences between results.