Revealing and Addressing Length-Bias Under Null, and Heterogeneous Effects, in Case Crossover Studies: a Causal Inference Approach.

• Activity Number: 259
• Type: Contributed
• Date/Time: Tuesday, August 13, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section*
• Author(s): Ravi Varadhan*+ and Constantine Frangakis
• Affiliation(s): Johns Hopkins University and Johns Hopkins University
• Keywords: case-crossover ; causal inference ; heterogeneity ; length-bias

Abstract

Designs that use only cases to assess exposure effects are increasingly used when finding appropriate controls is impractical, but analyses of even simple such designs present problems. In this work, we study new methods for the case crossover design of Maclure (1991). The goal with this design is to assess the hypothesis that a periodic exposure (e.g., coffee) triggers an event (e.g., myocardial infarction), using only subjects who have the event. The standard types of analysis then calculate: (i) the number of exposed periods having an event, as a fraction of total exposed periods; (ii) the number of unexposed periods having an event, as a fraction of total unexposed periods; and (iii) a "risk or odds ratio", using (i) and (ii) across subjects, compared to unity. We show that such standard analyses are length-biased with the result that: even if there is trully no exposure-event relation (causal null), the standard "risk or odds ratios" are lower than 1, thus misleading. We provide a new method of analysis that: (i) is valid under the causal null; and also (ii) takes account of heterogeneous effects under alternatives, and compare our methods with the standard approaches.