Search

Regression by composition is a new and flexible toolkit for building and understanding statistical models. Focusing here on regression models for a binary outcome conditional on a binary treatment and other covariates, we motivate the need for regression by composition. We do this first by exhibiting—using L'Abbé plots—the families of relationships between untreated and treated conditional outcome risks that emerge from generalized linear models for many different link functions. These are compared with the relationships (between untreated and treated risks) that arise from mechanistic sufficient component cause models, which are first principles causal models for binary outcomes. By considering mechanistic models that allow for non-monotone causal effects and by allowing sufficient causes to be associated, we expand upon similar discussions in the recent literature. We discuss conditions under which commonly used statistical models for binary data, such as logistic regression, arise from mechanistic models where the sufficient causes are associated in a particular way, as well as other situations in which the statistical models arising do not correspond to a generalized linear model but can be naturally expressed as a regression by composition model. [ABSTRACT FROM AUTHOR]

The ICH E9 addendum introduces the term intercurrent event to refer to events that happen after treatment initiation and that can either preclude observation of the outcome of interest or affect its interpretation. It proposes five strategies for handling intercurrent events to form an estimand but does not suggest statistical methods for estimation. In this article we focus on the hypothetical strategy, where the treatment effect is defined under the hypothetical scenario in which the intercurrent event is prevented. For its estimation, we consider causal inference and missing data methods. We establish that certain "causal inference estimators" are identical to certain "missing data estimators." These links may help those familiar with one set of methods but not the other. Moreover, using potential outcome notation allows us to state more clearly the assumptions on which missing data methods rely to estimate hypothetical estimands. This helps to indicate whether estimating a hypothetical estimand is reasonable, and what data should be used in the analysis. We show that hypothetical estimands can be estimated by exploiting data after intercurrent event occurrence, which is typically not used. for this article are available online. [ABSTRACT FROM AUTHOR]

When an entire cohort of patients receives a treatment, it is difficult to estimate the treatment effect in the treated because there are no directly comparable untreated patients. Attempts can be made to find a suitable control group (e.g. historical controls), but underlying differences between the treated and untreated can result in bias. Here we show how negative control outcomes combined with difference-in-differences analysis can be used to assess bias in treatment effect estimates and obtain unbiased estimates under certain assumptions. Causal diagrams and potential outcomes are used to explain the methods and assumptions. We apply the methods to UK Cystic Fibrosis Registry data to investigate the effect of ivacaftor, introduced in 2012 for a subset of the cystic fibrosis population with a particular genotype, on lung function and annual rate (days/year) of receiving intravenous (IV) antibiotics (i.e. IV days). We consider 2 negative control outcomes: outcomes measured in the pre-ivacaftor period and outcomes among persons ineligible for ivacaftor because of their genotype. Ivacaftor was found to improve lung function in year 1 (an approximately 6.5–percentage-point increase in ppFEV1), was associated with reduced lung function decline (an approximately 0.5–percentage-point decrease in annual ppFEV1 decline, though confidence intervals included 0), and reduced the annual rate of IV days (approximately 60% over 3 years). [ABSTRACT FROM AUTHOR]

Dynamic prediction models provide predicted survival probabilities that can be updated over time for an individual as new measurements become available. Two techniques for dynamic survival prediction with longitudinal data dominate the statistical literature: joint modelling and landmarking. There is substantial interest in the use of machine learning methods for prediction; however, their use in the context of dynamic survival prediction has been limited. We show how landmarking can be combined with a machine learning ensemble—the Super Learner. The ensemble combines predictions from different machine learning and statistical algorithms with the goal of achieving improved performance. The proposed approach exploits discrete time survival analysis techniques to enable the use of machine learning algorithms for binary outcomes. We discuss practical and statistical considerations involved in implementing the ensemble. The methods are illustrated and compared using longitudinal data from the UK Cystic Fibrosis Registry. Standard landmarking and the landmark Super Learner approach resulted in similar cross‐validated predictive performance, in this case, outperforming joint modelling. [ABSTRACT FROM AUTHOR]

Estimation of causal effects of time-varying exposures using longitudinal data is a common problem in epidemiology. When there are time-varying confounders, which may include past outcomes, affected by prior exposure, standard regression methods can lead to bias. Methods such as inverse probability weighted estimation of marginal structural models have been developed to address this problem. However, in this paper we show how standard regression methods can be used, even in the presence of time-dependent confounding, to estimate the total effect of an exposure on a subsequent outcome by controlling appropriately for prior exposures, outcomes, and time-varying covariates. We refer to the resulting estimation approach as sequential conditional mean models (SCMMs), which can be fitted using generalized estimating equations. We outline this approach and describe how including propensity score adjustment is advantageous. We compare the causal effects being estimated using SCMMs and marginal structural models, and we compare the two approaches using simulations. SCMMs enable more precise inferences, with greater robustness against model misspecification via propensity score adjustment, and easily accommodate continuous exposures and interactions. A new test for direct effects of past exposures on a subsequent outcome is described. [ABSTRACT FROM AUTHOR]