Course Level: Intermediate
Format Requirements: The course will consist of lectures, exercises drawn from the published literature and interactive discussion. The intended audience includes researchers from all substance matter fields, statisticians, epidemiologists, and decision analysts interested either in methods of causal analysis or causal interpretation of results based on the underlying method. Requirements: Basic knowledge in epidemiologic methods (confounding).
Background: One of the most important tasks of decision makers is to derive causal interpretations using both statistical analyses of original datasets and decision analysis. Often an intervention, action or risk factor is modeled to have a "causal effect" on one or more model parameters (e.g., probability, rate, or mean of outcome). Therefore, both the biostatistician and the decision analyst need tools to check: (1) when effect estimates have a causal interpretation and when they do not; and (2) the appropriate methods to derive causal effects instead of merely statistical associations.
Description and Objectives: The Objectives of this course are to:
(1) Define causal interventions and actions, draw and interpret causal diagrams, and apply the rules of causal diagrams to distinguish causal from non-causal statistical associations.
(2) Decide which biostatistical/epidemiological methods must be used in different situations to derive causal effect parameters.
(3) Use causal diagrams to estimate the direction of bias in "non-causal" models.
This course will provide an introduction to the principles of causation and causal diagrams, with focus on Directed Acyclic Graphs (DAG) and a brief introduction to methods for causal inference including g-formula, marginal structural models (inverse probability of treatment weighting), and structural nested models (lecture - exercises - discussion). Published cardiovascular, HIV, nutrition and obstetrics examples will be used to:
- Adjust for compliance in randomized clinical trials, where both "intention to treat" and "per protocol" analyses can fail to yield the true causal intervention effect;
- Assess the “fallibility of estimating direct effects” (i.e., adjusting for intermediate steps);
- Adjust for time-independent confounders in observational studies (i.e., confounder affects both risk factor and disease), where standard stratification or regression analysis yield valid causal effects if all confounders are measured, and
- Adjust for time-dependent confounding in observational studies (i.e., the confounder simultaneously acts as an intermediate step in the causal chain between risk factor and disease), where standard regression analysis fails and "causal methods" such as marginal structural models or g-estimation must be used.