Nothing to See Here? A Non-Inferiority Approach to Parallel Trends.

Area of Science:

• Health policy evaluation
• Econometrics
• Causal inference

Background:

• Difference-in-differences (DID) is widely used for observational health policy evaluation.
• DID relies on the parallel trends assumption: treatment and comparison groups would have similar trends without intervention.
• Current parallel trends tests have low power and can introduce bias when used for screening.

Purpose of the Study:

• To propose a novel non-inferiority/equivalence framework for assessing the parallel trends assumption in DID.
• To offer a statistically robust method that controls the probability of missing significant trend violations.
• To provide a higher-power alternative to existing methods, potentially reducing bias in treatment effect estimation.

Main Methods:

• Developed a non-inferiority/equivalence approach to evaluate parallel trends, measured on the treatment effect scale.
• Demonstrated the framework's ability to nest common use cases like linear trend tests and event studies.
• Analyzed the potential for bias when used as a screening step under various error structures.

Main Results:

• The proposed non-inferiority approach tightly controls the probability of missing substantial parallel trend violations.
• When used as a screening step, the method may introduce no or minimal bias under common error assumptions.
• The framework offers a higher-power alternative for testing treatment effects in flexible models compared to traditional methods.

Conclusions:

• The non-inferiority/equivalence approach provides a more powerful and less biased method for assessing parallel trends in DID analyses.
• This framework enhances the reliability of causal inference in health policy evaluations.
• The study illustrates the practical application of the method using a case study on the Affordable Care Act.