Causal Inference

A technique that uses a third variable—the instrument—to estimate causal effects in the presence of unobserved confounding. The instrument must influence the treatment (drug exposure) but have no direct effect on the outcome except through the treatment.

• Mendelian Randomization (MR): Uses genetic variants as instruments to mimic a randomized controlled trial.
• Example: A gene variant that naturally lowers a protein's level can reveal if that protein is a causal drug target for a disease.
• Key Assumption: The instrument must not be associated with confounders.

The process of answering 'what if' questions at the individual level. A counterfactual predicts what would have happened to a specific patient had they received a different treatment.

• Unit-Level Inference: Unlike average treatment effects, this targets a single person.
• Example: 'Would this patient's tumor have shrunk if they had received Drug B instead of Drug A?'
• Structural Equation Models (SEMs): The mathematical backbone used to compute counterfactuals from observational data.

Visual and mathematical representations of causal assumptions. Nodes represent variables (drug, outcome, confounder), and directed edges represent direct causal effects. The absence of a cycle ensures no variable causes itself.

• Confounder Identification: A common cause of both treatment and outcome that must be controlled for.
• Collider Bias: Mistakenly controlling for a variable caused by both treatment and outcome, which can open non-causal paths.
• Backdoor Criterion: A graphical rule to determine which variables must be adjusted to isolate a causal effect.

A statistical technique to reduce bias in observational studies by balancing treated and control groups. The propensity score is the probability of receiving a treatment given observed covariates.

• Matching: Pairing treated and untreated patients with similar scores.
• Inverse Probability of Treatment Weighting (IPTW): Creating a pseudo-population where treatment assignment is independent of measured confounders.
• Limitation: Cannot account for unmeasured confounders.

A quasi-experimental design that compares the change in outcome over time between a treatment group and a control group. It removes biases from permanent differences between groups and common time trends.

• Parallel Trends Assumption: In the absence of treatment, the difference between the groups would have remained constant.
• Application: Analyzing the effect of a new regulatory approval on drug utilization by comparing regions with and without the policy change.

Uses genetic variants as instrumental variables to mimic a randomized controlled trial. By analyzing whether a gene variant that mimics a drug's effect influences disease risk, researchers can predict the causal outcome of targeting that protein before a clinical trial begins.

• Example: PCSK9 gene variants that lower LDL cholesterol were shown to reduce heart disease risk, validating PCSK9 as a target and leading to repurposing strategies for existing antibodies.
• Key Insight: This method overcomes confounding by using the random assortment of genes at conception as a natural randomization mechanism.

Applies causal inference to spontaneous reporting systems like the FDA Adverse Event Reporting System (FAERS) to identify unexpected beneficial side effects. Algorithms must disentangle the true drug-event signal from confounding by the underlying disease.

• Example: Sildenafil was originally developed for hypertension and angina. Observational reports of a specific side effect, analyzed through causal lenses, led to its repurposing for erectile dysfunction.
• Method: Disproportionality analysis, refined with causal Bayesian networks, filters out noise from polypharmacy and comorbidity biases.

Mines longitudinal patient data to emulate target trials. By structuring observational EHR data to mimic a randomized study protocol—defining time zero, eligibility criteria, and treatment arms—analysts can estimate the causal effect of a drug on a secondary outcome.

• Example: Analysis of diabetic patient records showed that metformin use was causally associated with reduced cancer incidence, a finding now in prospective repurposing trials.
• Challenge: Requires rigorous handling of time-varying confounding and indication bias, often using inverse probability of treatment weighting.

Leverages the Connectivity Map (CMap) to find drugs that causally reverse a disease's gene expression pattern. The causal logic is that if a drug's effect on gene expression is the inverse of the disease's effect, the drug may treat the disease.

• Example: CMap analysis revealed that the HDAC inhibitor vorinostat, an anticancer drug, reversed gene signatures of inflammatory bowel disease, suggesting a repurposing opportunity.
• Mechanism: This moves beyond simple correlation by requiring a mechanistic, counterfactual match between the drug-induced and disease-associated transcriptional states.

A specialized form of Mendelian randomization that focuses on genetic variants within or near a drug target gene to predict the on-target effects of modifying that protein's function. This directly informs whether an existing drug for one indication could work for another.

• Example: cis-MR analyses of IL-6 receptor variants predicted that tocilizumab, an arthritis drug, would be effective against severe COVID-19, a hypothesis later confirmed in the RECOVERY trial.
• Advantage: Reduces pleiotropy (a gene affecting multiple traits) by instrumenting the specific drug target, providing cleaner causal estimates.

Uses a physician's prescribing preference as an instrumental variable to address confounding by indication in observational datasets. This method exploits the quasi-random variation in treatment assignment driven by a doctor's habit, rather than patient characteristics.

• Example: Used to determine the causal effect of COX-2 inhibitors versus traditional NSAIDs on gastrointestinal bleeding risk, controlling for the fact that high-risk patients were preferentially prescribed COX-2s.
• Key Assumption: The instrument (prescribing preference) must affect the outcome only through its effect on the treatment received, not through any other pathway.