RICE AND ANDERSON RESPOND

Abstract

Elliot et al. expressed their concern of selection bias owing to the inclusion of vehicles carrying 1 child with 1 or more adults in our study of child safety seat effectiveness.1 Elliot et al. point out that an assumption of similar underlying within-vehicle death risk is required. We examined this issue during the design of the matched-cohort study. We concluded that no bias would result because, after adjustment for seating position, gender, age splines, collision type, rollover status, and numerous “matched” factors like vehicle type and emergency vehicle response time, the assumption would be reasonable. To examine this possible bias, we fitted our analysis model to restricted sets of the data (Table 1). We started with 1 of our study's imputed data sets and estimated a child safety seat risk ratio (RR) of 0.33 (95% confidence interval [CI] = 0.29, 0.37) for children aged 3 years or younger. This estimate was based on 5732 matched sets of occupants. Then, we restricted the data to children aged 8 years or younger and adult drivers and estimated an RR of 0.35 (95% CI = 0.30, 0.40) based on 4503 sets. Finally, we further restricted the data to child-only sets and estimated an RR of 0.30 (95% CI = 0.22, 0.41) based on 1188 sets. While much information is lost by using the restricted data sets, the effectiveness estimates are essentially unchanged. This examination does not support the concern of bias. The child-only analysis produces similar point estimates with wider CIs. The estimated seat belt RRs also were not meaningfully altered during this exercise. They were 0.39, 0.41, and 0.32 for the full, partially restricted, and fully restricted data sets, respectively. TABLE 1 Estimated Safety Seat and Safety Belt Use Risk Ratios (RRs), by Data Set: United States, 1996–2005 Elliot et al. states that our analysis can be likened to comparing children with adults who never use child safety seats and that the analysis is thus an extrapolation beyond the data. We assert that our study is better conceptualized as both a direct comparison of restrained and unrestrained children in some sets and an indirect comparison of restrained and unrestrained children each with unrestrained adults in other sets. It can also be viewed as a regression analog of the widely used double-pair comparison.2 The assessment we provide here demonstrates the validity of the method used in our study.