Sensitivity Analysis Methods to Know for Causal Inference

Sensitivity analysis methods help assess how unmeasured confounding affects causal inferences. These techniques, like the E-value method and Rosenbaum bounds, provide tools to evaluate the robustness of treatment effects and identify potential biases in observational studies.

1. E-value method

   • Provides a quantitative measure of how strong an unmeasured confounder must be to negate the observed treatment effect.
   • Useful for assessing the robustness of causal conclusions in observational studies.
   • The E-value is calculated based on the ratio of the risk of the outcome in treated versus untreated groups.

2. Rosenbaum bounds

   • Offers a way to assess the sensitivity of treatment effect estimates to unobserved confounding.
   • Establishes bounds on the treatment effect under different assumptions about the degree of unobserved confounding.
   • Particularly useful in matched observational studies to evaluate the impact of hidden biases.

3. Cornfield conditions

   • A set of conditions that must be satisfied for a causal inference to be valid in the presence of unmeasured confounding.
   • Focuses on the relationship between the treatment, outcome, and confounders to determine the plausibility of causal claims.
   • Helps in identifying potential biases that could affect the validity of causal estimates.

4. Ding and VanderWeele's bounding factor

   • Provides a framework for bounding the causal effect of a treatment when there are unmeasured confounders.
   • Allows researchers to express the potential impact of unmeasured confounding on causal estimates.
   • Facilitates sensitivity analysis by quantifying the degree of confounding needed to alter conclusions.

5. Multiple bias modeling

   • Involves modeling several potential biases simultaneously to assess their collective impact on causal estimates.
   • Helps in understanding how different biases can interact and influence the results of a study.
   • Provides a more comprehensive view of the robustness of causal inferences.

6. Probabilistic bias analysis

   • Uses probability distributions to quantify the uncertainty associated with unmeasured confounding.
   • Allows researchers to incorporate expert opinions or prior knowledge about potential biases into the analysis.
   • Aids in making more informed decisions about the validity of causal claims.

7. Tipping point analysis

   • Identifies the threshold at which an unmeasured confounder would change the direction or significance of the treatment effect.
   • Helps in understanding the robustness of causal conclusions by determining how sensitive they are to unmeasured biases.
   • Useful for guiding future research and data collection efforts.

8. Instrumental variable sensitivity analysis

   • Evaluates the strength of the instrumental variable in the presence of unmeasured confounding.
   • Assesses how robust the causal estimates are to violations of the instrumental variable assumptions.
   • Provides insights into the reliability of causal inferences drawn from instrumental variable analyses.

9. Mediation sensitivity analysis

   • Examines how sensitive mediation effects are to unmeasured confounding of the mediator-outcome relationship.
   • Helps in understanding the robustness of causal pathways and the role of mediators in causal inference.
   • Aids in identifying potential biases that could affect the interpretation of mediation analyses.

10. Regression discontinuity sensitivity analysis

    • Assesses the robustness of causal estimates derived from regression discontinuity designs to potential biases.
    • Evaluates how changes in the cutoff point or assumptions about the functional form can impact results.
    • Provides insights into the validity of causal claims made using regression discontinuity methods.