The Sixteenth Conference of the International Society for Environmental Epidemiology (ISEE): Abstracts
Case-crossover analyses have proved a useful method for studying the acute effects of air pollution whilst controlling for season and temperature. The only drawback has been the choice of the set of control days, which is complicated by the need to balance between overmatching and the adequate controlling of confounders. We overcame the need for control selection by using a partitioned regression of time-matched differences in mortalities against the associated differences in exposures and confounders. The advantages of taking such differences are: the co-linearity between exposures is generally reduced, as is the autocorrelation within exposures; and the outcome variable can be characterised using a Normal distribution.
The stages of the method are as follows: Partition the daily time series into consecutive non-overlapping 28-day windows Calculate the differences between the mortalities, exposures and confounders and their respective means in each window Regress the differences in mortalities against the differences in exposures and confounders. The null hypothesis of no relationship between mortality and exposure is tested using bootstrap data, created by randomly permuting mortality, exposure and confounders in the four-week windows. Two-hundred bootstrap surrogates are created and run through steps 1-3 to give estimates (and bootstrap confidence limits) of the regression parameters under the null hypothesis. We compared the performance of partitioned regression to case-crossover using simulated data. The case-crossover controls were chosen using the 28-day windows with a 4-day exclusion period from the case day. The simulated data was: three years long, had a mean daily mortality of 22, included a long-term linear trend and seasonal confounder, and had a true absolute risk of between 0 and 0.40 for every unit increase in exposure.
Power and coverage of a partitioned regression and case-crossover analysis from 100 simulations (using a 5% significance level).
Our method regresses the differences in time-matched mortalities against related differences in exposures and confounders. The method has an equal ability to control for seasonal confounding as the case-crossover, but has improved power and coverage, and overcomes the need to choose a control-referent method.