Observational Studies Overview

Introduction

• Observational studies are used to investigate cause and effect relationships when controlled experiments are not feasible.
• Examples of observational studies include cohort, cross-sectional, and case-control studies.
• Observational studies are important for evaluating interventions when randomized controlled trials are unnecessary, inappropriate, impossible, or inadequate.

Challenges with Observational Studies

• Lack of control:
  • Confounding bias: When a variable is associated with both the exposure and the outcome, leading to an incorrect estimate of the causal effect.
  • Selection bias: When the characteristics of the study groups are not comparable, leading to an incorrect estimate of the causal effect.
• Unmeasured confounders: Variables not accounted for in the analysis can lead to bias.
• Time-varying confounders: Variables that change over time and can influence both the exposure and the outcome.

Advantages of Randomized Studies

• Randomization: Ensures that treatment assignment is independent of baseline characteristics and outcomes, minimizing selection bias.
• Control: Allows researchers to control for confounding variables by randomly assigning participants to treatment groups.

Conditions for Causal Inference

• Exchangeability: Also known as "conditional exchangeability," assumes that treatment is independent of potential outcomes within strata defined by measured confounders. This assumption is not verifiable in observational studies.
• Positivity: Requires a non-zero probability of receiving each treatment level at all values of the measured confounders. This can be empirically verified in observational studies.
• Consistency: Assumes that a defined standardized treatment exists without variation, preventing multiple versions of the same treatment affecting outcomes.

Approaches to Estimate Causal Effects in Observational Studies

• Methods that mimic randomization:
  • Stratification: Dividing participants into groups based on shared characteristics to reduce confounding.
  • Matching: Pairing participants with similar characteristics to ensure comparability.
  • Propensity score matching: Using a statistical model to estimate the probability of receiving each treatment level based on measured characteristics, then matching participants with similar propensity scores.
• Instrumental variable analysis: Using a variable that influences the exposure but not the outcome directly to estimate the causal effect.
• Sensitivity analysis: Exploring the robustness of the findings to unmeasured confounders.

Example: Causal Effect of Smoking Cessation on Weight Gain

• The example uses the "NHANES" data to investigate the causal effect of quitting smoking on weight gain.
• Study design: Observational study analyzing factors like age, sex, race, university, weight, smoking intensity, and smoking years.
• Target Outcome: Weight gain ( measured as the difference in weight between two time points).
• Exposure: Smoking cessation (coded as 1 for quitting and 0 for continuing smoking).
• Confounders: Factors associated with both smoking status and weight gain, including age and smoking intensity.

Estimating the Associational Effect

• Associational effect: The observed difference in weight gain between those who quit smoking and those who did not.
• The associational effect does not represent the true average causal effect because of the presence of confounders.

Estimating the Average Causal Effect

• Average causal effect: The difference in weight gain that would occur if both groups had received the same treatment (either quitting smoking or continuing to smoke).
• To estimate the average causal effect, the study needs to address the confounding bias introduced by variables like age and smoking intensity.

Methods for Estimating the Average Causal Effect

• Parametric G-formula:
  • A statistical method that uses a regression model to estimate the average causal effect. It adjusts for confounders by estimating the expected outcome for each participant under both treatment levels.
  • The G-formula relies on the consistency assumption, which assumes that the treatment received is the same as the treatment assigned.

Key takeaways

• Observational studies offer valuable insights for understanding cause-and-effect relationships when experimental studies are not feasible.
• The identification of confounding bias and selection bias is crucial in observational studies.
• Methods like propensity score analysis, matching, and the G-formula can be used to adjust for confounding and estimate causal effects.
• Understanding the strengths and weaknesses of observational studies is crucial when interpreting findings and drawing conclusions about causality.