Assessing causal treatment effect estimation when using large observational datasets.

Area of Science:

• Epidemiology
• Biostatistics

Background:

• Growing use of observational data (e.g., Electronic Health Records) necessitates robust analysis methods.
• Randomized controlled trials have limitations, increasing reliance on observational studies.
• Causal interpretation of observational data requires strict, often unverifiable, assumptions.

Purpose of the Study:

• Compare common methods for estimating treatment effects from observational data.
• Highlight the importance of justifying assumptions for causal inference.
• Evaluate methods under various confounding scenarios.

Main Methods:

• Simulation study using a chronic obstructive pulmonary disease cohort.
• Compared two-stage least squares instrumental variables, propensity score, and linear regression.
• Assessed performance with varying instrumental variable strength, unmeasured confounding, and sample sizes up to 200,000.

Main Results:

• Two-stage least squares (2SLS) can yield unbiased estimates with large samples and unmeasured confounding.
• Propensity score and linear regression methods yielded similar results in simulations.
• Weak instruments or strong unmeasured confounding increased uncertainty in 2SLS estimates.
• Violating 2SLS assumptions led to biased estimates, sometimes worse than linear regression.

Conclusions:

• Instrumental variable methods can outperform regression and propensity scores.
• Careful consideration and justification of instrumental variable assumptions are critical.
• Failure to justify assumptions may lead to worse performance than ignoring confounding.