Causal Inference 2
Presenter: Karla DiazOrdaz
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
INSTRUMENTAL VARIABLE APPROACHES FOR ESTIMATING COMPLIER AVERAGE CAUSAL EFFECTS ON BIVARIATE OUTCOMES IN RANDOMISED TRIALS WITH NON-COMPLIANCE
Non-compliance is a common problem in Randomised Controlled Trials (RCTs), as some participants depart from their randomised treatment, by for example switching from the experimental to the control regimen. In the presence of non-compliance, a complimentary estimand of interest is the causal effect of treatment received. Instrumental variable (IV) methods can be used to obtain the complier average causal effect (CACE). Cost-effectiveness analyses (CEA) are an important source of evidence for informing clinical decision-making and health policy. CEA commonly report an ITT estimand, however, policy-makers may require additional estimands, such as the relative cost-effectiveness for compliers. We extend existing IV approaches to multivariate settings, such as cost-effectiveness analyses. We propose a three-stage least squares (3sls) regression approach, which combines 2sls with a seemingly unrelated regression system of equations. We also consider two `unadjusted' Bayesian models that ignore the correlation between the errors of the treatment received and outcome models, but allow the use of non-normal distributions. Finally, we develop a Bayesian full likelihood approach (BFL), which jointly models the effects of random assignment on treatment received and the outcomes. We investigate the methods' performance in estimating CACE in cost-effectiveness studies with a simulation and a re-analysis of an RCT. We contrast these methods to applying separate 2sls to estimate CACE for each endpoint. Costs are assumed to follow Normal, Gamma or Inverse Gaussian distributions, to have positive or negative correlation with health outcomes, with different levels of non-compliance (30% or 70%), and sample sizes (n=100 or 1000). With 30% non-compliance, each method provides unbiased estimates. However, with high levels of non-compliance, both unadjusted Bayesian methods provide biased estimates. The 2sls approach, while unbiased, reports Confidence Interval (CI) coverage above (positive correlation), and below (negative correlation) nominal levels. By contrast, the BFL and 3sls methods provide estimates with low levels of bias and CI coverage rates which are close to nominal levels throughout.
Causal Inference 2
Presenter: Yuying Xie
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
A Model Averaging Approach for Estimating Propensity Scores by Optimizing Balance
Many approaches, including traditional parametric modelling and machine learning techniques, have been proposed to estimate propensity scores. This paper describes a new model averaging approach to propensity score estimation in which parametric and nonparametric estimates are combined to achieve covariate balance. Simulation studies are conducted across different scenarios varying in the degree of interactions and nonlinearities in the treatment model. The results show that the proposed method produces less bias and smaller standard error than existing approaches. It also shows that the combined approach with the objects of minimizing average standardized mean difference leads to the best performance. The proposed approach is also applied to a real data set in evaluating the causal effect of formula or mixed feeding versus exclusive breastfeeding on a child's BMI Z-score at age 4. The data analysis shows that formula or mixed feeding is more likely to lead to obesity at age 4, compared to exclusive breastfeeding.
Causal Inference 2
Presenter: Samantha Noreen
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
Improving Covariate Balancing Propensity Score for Continuous Treatment Regimes
The propensity score plays an essential role in causal inference using observational data. However, a number of challenges arise when using the propensity score to deal with non-binary treatment regimes. In this work, we propose an approach to improve the covariate balancing propensity score (CBPS; Imai and Ratkovic, 2014) for continuous treatment regimes that can handle both continuous and discrete covariates. The proposed approach is shown to outperform the original CBPS in simulations and is further illustrated through analysis of the clinical data from the Emory Amyotrophic Lateral Sclerosis (ALS) Center.
Causal Inference 2
Presenter: Michael Wallace
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
Model Assessment and Dynamic Treatment Regimens
Personalized medicine is a rapidly expanding area of health research wherein subject level information is used to inform treatment. Dynamic treatment regimens (DTRs) are one means by which personalized medicine can be studied theoretically and applied in practice. DTRs are sequences of decision rules which take subject information as input and provide treatment recommendations as output. Such regimens therefore tailor each treatment decision to a patient's unique circumstances, but can also identify management plans which optimize long term outcomes by accommodating potentially obscure delayed treatment effects and other complex interactions. However, taking such factors into account can complicate the problem of causal inference in this context. One approach considers the blip: a structural nested mean model of the expected difference in the (potentially counterfactual) outcome when using a baseline treatment instead of the observed treatment. DTR estimation in this context therefore relies on estimating blip parameters and numerous methods have been proposed for this purpose. Some of these methods, including G-estimation and dynamic weighted ordinary least squares, are implemented through the specification of three models: one for the blip, and two other 'nuisance' models relating treatment and outcome to covariates. One supposed benefit of these approaches is that they are doubly-robust, leading to consistent blip parameter estimators as long as at least one of the two nuisance models is correctly specified. However, few options exist for selection (or even validation) of these nuisance models, or indeed the all-important blip. We present two new approaches to model assessment in this setting. First, we illustrate a quasi-likelihood information criterion within the DTR setting, allowing straightforward selection of blip models. We then turn our focus to the two nuisance models, and show how the double robustness property itself may be used to perform not only model assessment, but also model validation. The methods are illustrated through simulations as well as application to data from studies of heart disease and depression.
Causal Inference 2
Presenter: Takashi Yanagawa
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
Reappraisal of odds ratios for validating observational comparative studies
Conditions of balancing score and strongly ignorable were introduced by Rosenbaum and Rubin (Biometrika, 1983) to validate an observational study. Under these two conditions they showed that the difference between treatment and control at each value of a balancing score is an unbiased estimate of the average treatment effect in a randomized follow-up study. However, the strongly ignorable condition is not easy to verify. In this paper we replace the average treatment effect and unbiased estimate by odds ratios and consistent estimate, respectively, to validate an observational comparative study based on the following ideas. First, define confounding by means of odds ratios as in my previous paper (Environmental Health Perspectives,1979; Biometrika, 1984). Second, introduce the concept of stratified matching so that odds ratios throughout strata are stable (homogeneous). It will be shown that that the Mantel-Haenszel estimator (J. National Cancer Institute 1959) based on an observational study is a consistent estimator of the common odds ratio of a randomized follow-up study.
Causal Inference 2
Presenter: Mats Stensrud
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon A Carson Hall (Level 2)
Session Synopsis:
Epidemiological paradoxes explored by causal frailty models - The qualitative DAG needs quantitative support
A wide range of associations in epidemiology are claimed to be paradoxical. However, reversal of associations could have plausible explanations. Directed acyclic graphs (DAGs) are frequently drawn to explore causal structures, and DAGs often reveal spurious effects. Indeed, by drawing a DAG, a potential bias will probably be identified in most real-life scenarios. The DAG, however, is purely qualitative. Hence, quantitative estimates are needed to better understand the practical relevance of a conceivable bias. In epidemiology, DAGs are frequently displayed without assessing the magnitude of the biases. Spurious effects are often unmeasured, but quantitative estimates of unmeasured heterogeneity can be derived from simple frailty models. An extensive framework for frailty modeling exists in survival analysis, and frailty analyses have been widely used in related fields such as demography and bioanthropology. By combining causal DAGs and quantitative frailty models, we improve the understanding of counter-intuitive associations in epidemiology. First, we discuss scenarios where a treatment effect is examined over time, and we point to a time-dependent Simpson's paradox. For example, we show that a treatment with constant effect can appear beneficial at time t=0, but harmful at a time point t>0. Then, we explore a competing risk setting, where being at increased risk of one event may falsely reduce the risk of another event. Finally, we derive mathematical expressions to reveal spurious effects in studies of a diseased population (index-event studies). In particular, analyses estimating the effect of a risk factor (e.g. obesity) on an outcome (e.g. mortality) in a population with a chronic disease (e.g. kidney failure), will be prone to the index-event bias. In all of these frequently occurring scenarios, ignoring the unknown heterogeneity may lead to erroneous conclusions. Hence, applied researchers should recognize these biases, and assessing the frailty effect will often be appropriate.
