Use of three summary measures of pediatric vaccination for studying the safety of the childhood immunization schedule. Issue 10 (28th February 2019)

• Record Type: Journal Article
• Authors: Xu, Stanley; Newcomer, Sophia R.; Kulldorff, Martin; Daley, Matthew F.; Fireman, Bruce; Glanz, Jason M.
• Is Part Of: Vaccine. Volume 37:Issue 10(2019)
• Page Start: 1325
• Page End: 1331
• Publication Date: 2019-02-28
• Subjects: Summary measures -- Pediatric vaccination -- Immunization safety -- Collinearity -- Variance inflation factor; Vaccines -- Periodicals; 615.372
• DOI: 10.1016/j.vaccine.2019.01.040
• Languages: English
• ISSNs: 0264-410X

Abstract

Highlights: In a Vaccine Safety Datalink dataset, summary measures such as cumulative exposure to vaccine doses, vaccine antigens, and aluminum in vaccines were highly correlated. The variance inflation factors were 7.14, 6.25 and 2.17 when range-scaled number of vaccine antigens, number of vaccines, and vaccine aluminum exposure were each regressed against the other two measures, respectively. In simulation where one of the three summary measures was the true risk factor, we showed that it was very likely that an association would be detected if a summary measure other than the true risk factor was used. In the simulations, we also showed that the empirical power to detect the association between the true risk factor and the outcome significantly decreased if the other two measures were in the model. If one measure was a strong risk factor and the other two were weak risk factors, a multiple logistic regression model with the three summary measures would be appropriate for detecting the strong risk factor. Correctly specifying the relationship between risk factors and the outcome was crucial. Abstract: Background: Summary measures such as number of vaccine antigens, number of vaccines, and vaccine aluminum exposure by the 2nd birth day are directly related to parents' concerns that children receive too many vaccines over a brief period. High correlation among summary measures could cause problems in regression models that examine their associations with outcomes. Objectives: To evaluate the performance of multiple regression models using summary measures as risk factors to simulated binary outcomes. Methods: We calculated summary measures for a cohort of 232, 627 children born between 1/1/2003 and 9/31/2013. Correlation and variance inflation factors (VIFs) were calculated. We conducted simulations (1) to examine the extent to which an association can be detected using a summary measure other than the true risk factor; (2) to evaluate the performance of multiple regression models including true and redundant risk factors; (3) to evaluate the performance of multiple regression models when all three were risk factors; (4) to examine the performance of multiple regression models with incorrect relationship between risk factors and outcome. Results: These summary measures were highly correlated. VIFs were 7.14, 6.25 and 2.17 for number of vaccine antigens, number of vaccines, and vaccine aluminum exposure, respectively. In simulations, an association would be detected if a summary measure other than the true risk factor was used. The power to detect the association between the true risk factor and outcome significantly decreased if redundant risk factors were included. When all three were risk factors, multiple regression model was appropriate to detect the stronger risk factor. Correctly specifying the relationship between risk factors and the outcome was crucial. Conclusions: Multiple regression models can be used to examine the association between summary measures and outcome despite of high correlation among summary measures. It is important to correctly specify the relationship between risk factors and outcome.