BACKGROUND: Case-crossover studies can be used to assess the effects of transient exposures in acute events by comparing cases' exposures at the time of the event against their usual exposure. Cases from a case-control study can also be analysed by this method if appropriate information is obtained. METHODS: We apply the case-crossover method to data collected in a case-control study of acute coronary events, to estimate the risk of alcohol consumption in the 24 h before a coronary event. Two questions, concerning the usual frequency of consuming alcohol and how much was consumed, were used in the analysis. The sensitivity of effect estimates to the two questions and to different ways of eliciting probabilities from them, are examined. The cases and controls were both analysed in the same way, assigning controls a 'pseudo-event' time at random. It is suggested that controls may provide a benchmark to establish whether bias exists in the case-series analysis. RESULTS: We find that the case-crossover analysis is sensitive to how exposure probabilities are assigned. Adopting a Poisson model for exposures leads to inflated risk estimates. The bias is removed if the control series is used to adjust estimates. There appears to be a small protective effect from alcohol consumption. CONCLUSION: Case-crossover designs have considerable potential, but reliable information on usual exposure limits their applicability. More research is required on how best to elicit exposure probabilities.
BACKGROUND: Case-crossover studies can be used to assess the effects of transient exposures in acute events by comparing cases' exposures at the time of the event against their usual exposure. Cases from a case-control study can also be analysed by this method if appropriate information is obtained. METHODS: We apply the case-crossover method to data collected in a case-control study of acute coronary events, to estimate the risk of alcohol consumption in the 24 h before a coronary event. Two questions, concerning the usual frequency of consuming alcohol and how much was consumed, were used in the analysis. The sensitivity of effect estimates to the two questions and to different ways of eliciting probabilities from them, are examined. The cases and controls were both analysed in the same way, assigning controls a 'pseudo-event' time at random. It is suggested that controls may provide a benchmark to establish whether bias exists in the case-series analysis. RESULTS: We find that the case-crossover analysis is sensitive to how exposure probabilities are assigned. Adopting a Poisson model for exposures leads to inflated risk estimates. The bias is removed if the control series is used to adjust estimates. There appears to be a small protective effect from alcohol consumption. CONCLUSION: Case-crossover designs have considerable potential, but reliable information on usual exposure limits their applicability. More research is required on how best to elicit exposure probabilities.
Authors: Elizabeth Mostofsky; Harpreet S Chahal; Kenneth J Mukamal; Eric B Rimm; Murray A Mittleman Journal: Circulation Date: 2016-03-08 Impact factor: 29.690