Author response to “Impact of survival bias on opioid‐related outcomes when using death as an exclusion criterion”

Thank you for the opportunity to respond to Vicki Osborne’s letter,1 which correctly points out the need for additional details in reporting our methods and results.2 We also agree and acknowledged in our original report the potential for survival bias that could lead to under ascertainment of our secondary outcome measures such as emergency department (ED) and opioid-related adverse events. As part of this letter, we have included Table 1 that describes the study sample after the application of each inclusion and exclusion criteria described in our original report to allow readers to assess the potential impact of these restrictions on our sample. The restriction requiring each recipient to survive to the end of the post-periods was imposed in order to assess opioid exposure measures over a uniform time period of 1year for all recipients, which was the primary focus of our original analysis. As presented in Table 1, there was a modest proportion of veterans (3.5–8.3%) excluded because of mortality, but there were many more veterans excluded from the historical control group than the policy-affected group (1234 vs. 388). Upon revisiting this issue, we discovered that when the survival criteria were initially implemented, persons in both groups had to survive to the end of the entire study period, which explains the difference in the number of subjects excluded in the groups. Those in the historical controls had to survive at least 3years, whereas those in the policy group had to survive 2years. To evaluate the influence of requiring all subjects to survive to the study end period and the potential survivor bias, we reran our analyses for our secondary outcome measures without requiring subjects to survive to the study end date. As seen in Table 2, the number of acetaminophen overdoses was unchanged and the number of opioid-related events changed very little indicating that the potential for survival bias to influence our reported findings on these measures was non-existent or minimal. There were substantial increases in the number of ED visits between the two samples; there were approximately 5–15% more ED visits occurring in the larger sample that did not require subjects to be alive at study end. To determine if these differences in ED visits impact our findings, we reran the repeated measures ordinary least squares model without excluding those who died in the post-periods. In the original analysis, we saw that ED visits significantly increased in the post-periods as compared with the pre-periods (B=0.049, 95%CI: 0.014, 0.084), while no significant association existed between the policy groups nor with the GROUP*PERIOD interaction term (B=0.037, 95%CI: 0.016, 0.090). With the revised analysis, ED visits still significantly increased in the post-periods (B=0.055, 95%CI: 0.021, 0.089), while no significant association existed between the policy groups, and the GROUP*PERIOD interaction term remained non-significant (B=0.029, 95%CI: 0.023, 0.081) indicating that propoxyphene withdrawal had no detectable association with changes in ED visits. Although survival bias could have potentially influenced the findings of our original secondary analyses, the revised analyses indicate that this is unlikely, and we continue to find no association between the propoxyphene withdrawal and acetaminophen overdoses, opioid-related adverse events, or ED visits in chronic users of propoxyphene. Just prior to the publication of our original paper, another report using claims data from a large US health insurer found a significant decrease in the number of opioid overdoses after propoxyphene was withdrawn and Oxycontin was reformulated as abuse deterrent.3 However, it is unclear how much, if any, of that decrease was attributable to the propoxyphene withdrawal, and when viewed in conjunction with our data, one might speculate that the reformulation of oxycodone may be more likely. Further research would be needed to clarify these associations using more similar samples. It should be noted that these revised analyses do not account for variable follow-up time between the cohorts; however, the vast majority of subjects (>90%) survived throughout the 1year follow-up period, and any differences in person-time are unlikely to be a major source of potential bias in these revised analyses.