Big Data, Data Science, and Causal Inference: A Primer for Clinicians

Abstract

Clinicians handle a growing amount of clinical, biometric, and biomarker data. In this "big data" era, there is an emerging faith that the answer to all clinical and scientific questions reside in "big data" and that data will transform medicine into precision medicine. However, data by themselves are useless. It is the algorithms encoding causal reasoning and domain (e.g., clinical and biological) knowledge that prove transformative. The recent introduction of (health) data science presents an opportunity to re-think this data-centric view. For example, while precision medicine seeks to provide the right prevention and treatment strategy to the right patients at the right time, its realization cannot be achieved by algorithms that operate exclusively in data-driven prediction modes, as do most machine learning algorithms. Better understanding of data science and its tasks is vital to interpret findings and translate new discoveries into clinical practice. In this review, we first discuss the principles and major tasks of data science by organizing it into three defining tasks: (1) association and prediction, (2) intervention, and (3) counterfactual causal inference. Second, we review commonly-used data science tools with examples in the medical literature. Lastly, we outline current challenges and future directions in the fields of medicine, elaborating on how data science can enhance clinical effectiveness and inform medical practice. As machine learning algorithms become ubiquitous tools to handle quantitatively "big data," their integration with causal reasoning and domain knowledge is instrumental to qualitatively transform medicine, which will, in turn, improve health outcomes of patients.

Keywords: big data; causal inference; data science; machine learning; the ladder of causation.

Figure 1

Examples of causal directed acyclic graph that encodes a priori domain knowledge and causal structural hypothesis. (A) Birth-weight paradox. There is no direct arrow from maternal smoking (exposure) to infant mortality (outcome), representing no causal effect. However, association/prediction-mode machine learning algorithm would automatically adjust for variables that are associated both with smoking and mortality (e.g., low birth-weight). Graphically, a rectangle placed around the low-birth weight variable represents adjustment. However, this adjustment for the collider (a node on which two directed arrows “collide”; Table 1) opens the flow of association from exposure → collider → covariates (e.g., structural anomaly) → outcome, which leads to a spurious (non-causal) association. (B) Simple example of causal diagram, consisting of exposure (biologic agent), outcome (asthma control), and covariates (e.g., baseline severity of illness). The presence of edge from a variable to another represents our knowledge on the presence of a direct effect. (C) Example of confounding. While there is no causal effect (i.e., no direct arrow from exposure to outcome), there is an association between these variables through the paths involving a common-cause covariate (i.e., a confounder), leading to a non-causal association between the exposure and outcome (i.e., confounding; Table 1). (D) Example of de-confounding. This confounding can be addressed by adjusting for the confounder by blocking the back-door path. Graphically, a rectangle placed around the confounder blocks the association flow through the back-door path. (E) Example of mediation. The causal relation between the exposure (systemic antibiotic use), mediator (airway microbiome), and outcome (asthma development). The confounders (e.g., acute respiratory infections) between the exposure, mediator, and outcome should be adjusted. The indirect (or mediation) effect is represented by the path which passes through the mediator. The direct effect is represented by the path which does not pass (the broken line; Table 1). (F) Example of mendelian randomization. Genetic variants that are strongly associated with the exposure of interest (mental illnesses) function as the instrument variable. Note that there is no association (or path) between the genetic variants and unmeasured confounders (i.e., independent condition) and that the genetic variants affect the outcome only through their effect on the exposure (i.e., exclusion restriction condition; Table 3).

Figure 2

Identification and estimation of heterogenous treatment effects. In this hypothetical example, suppose, we investigate treatment effects of systemic corticosteroids on hospitalization rates among preschool children with virus-induced wheezing. (A) Randomized control trial (RCT) to investigate the average treatment effect of systemic corticosteroids (conventional 1:1 RCT). (B) Investigating heterogeneous treatment effects using tree-based machine learning models. In each of the branches (e.g., subgroup A children have specific virus infection and a history of atopy), children have a comparable predicted probability of receiving systemic corticosteroids. Children within each subgroup function as if they came from an RCT with eligibility criteria stratified by clinical characteristics.

Figure 3

Integration of “big data,” data science, and domain knowledge toward precision medicine. Development of precision medicine requires an integration of “big data” from expanded data sources and capture with robust data science methodologies and analytics that encode domain causal knowledge and counterfactual causal reasoning.