[Skip to Navigation]
Sign In
July 2012

Error in Data Analysis in: Adolescent Use of School-Based Health Centers and High School Dropout

Arch Pediatr Adolesc Med. 2012;166(7):679. doi:10.1001/archpediatrics.2012.754

We were notified by a reader (Sharkansky S. School-based health center use and high school dropout rates. Arch Pediatr Adolesc Med. 2012;166[7]:675-676) that the article by Kerns and colleagues (Kerns SEU, Pullman MD, Walker SC, Lyon AR, Cosgrove TJ, Bruns EJ. Adolescent use of school-based health centers and high school dropout. Arch Pediatr Adolesc Med. 2011;165[7]:617-623) might be in error because of a mistake in the analysis. Our statistical consultant confirmed that this reader might be correct and we thus asked the authors to reexamine their analytic strategy and repeat their analysis if appropriate. The authors discovered they had erred by inadvertently introducing a systematic time-dependent bias (“immortal-time”), in which events at one point affect events in the past, in this case graduation affecting frequency of clinic use. Although this was a major error, we decided not to retract the article but instead to correct it because this was an unintentional error in analysis. We apologize for this error and thank the perceptive reader for identifying it.

Error in Data Analysis. In the Article “Adolescent Use of School-Based Health Centers and High School Dropout”(2011;165[7]:617-623), published in the July 2011 issue of the Archives, we operationalized quantity of school-based health center (SBHC) use as the number of times students visited the SBHC, divided by the number of semesters they were enrolled in school, and then categorized use into no, low-, moderate-, and high-use groups. However, we now realize the error with this approach is that the longer students were enrolled in school, the more opportunities they had to use the SBHC. In other words, it could be that attending school through graduation had an effect on measured SBHC use rather than SBHC use impacting graduation.

We repeated our analysis with modifications to avoid time-dependent bias, ie, restricting the exposure to events that occurred before the outcome. Two Cox regression time-to-event models with time-varying covariates were used to determine significant predictors of time until dropout. Our predictor of interest for model 1 was the time-varying cumulative average amount of SBHC use for all semesters up to and including the previous semester. As in our prior analyses, this was categorized into none, low (0.125-0.5 visits/semester), moderate (0.51-2.5 visits/semester), and high (>2.51 visits/semester) use. Contrary to our previous findings, SBHC use was not significantly related to dropout. Our predictor of interest for model 2 was the time-varying amount of SBHC use during the previous semester only, categorized into none, low/moderate (1-2 visits), and high (3 visits or more). Again, SBHC use was not significantly related to dropout. We apologize for this mistake.