Use of Cox regression with treatment status as a time- dependent covariate to re-analyze survival benefit excludes immortal time bias effect in patients with glioblastoma who received prolonged adjuvant treatment with valganciclovir Cecilia S€ oderberg-Naucler 1 , Inti Peredo 1,2 , Afsar Rahbar 1 , Fredrik Hansson 3 , Anders Nordlund 3 and Giuseppe Stragliotto 1,4 1 Department of Medicine, Solna, Unit for Experimental Medicine, Cell and Molecular Immunology, Center for Molecular Medicine, Karolinska Institute, Stockholm, Sweden 2 Department of Neurosurgery, Karolinska University Hospital, Stockholm, Sweden 3 Norma, Traktorv€ agen 11, 226 60 Lund, Sweden 4 Department of Neurology, Karolinska University Hospital, Stockholm, Sweden In a letter to the Editor, Liu and Hu 1 commented on our studies demonstrating a potential survival benefit in glioblas- toma patients who received prolonged treatment with valgan- ciclovir for cytomegalovirus (CMV) infection in their tumor 2,3 as add on to their standard oncology treatment. Liu and Hu raised the question of whether so-called immortal- time bias could explain the highly improved survival observed in treated patients. Indeed, with a retrospective design and a definition of exposure that requires a minimum time of exposure, there is a risk of immortal time bias, since those qualifying for exposure by definition are given a certain degree of immortal time. 4 Hence, Liu and Hu’s call for re- analysis of our data using Cox regression with treatment sta- tus as a time-dependent covariate is justified. Here we pres- ent the outcome of that re-analysis. First, we reanalyzed our data by splitting the follow-up time for the treated group into treated and untreated follow- up time. A Cox regression model allowing for delayed entry in the treatment group yielded an overall hazard ratio (HR) of 0.41 (95% confidence interval [CI]: 0.28–0.61) for treated versus untreated patients. Figure 1, to be compared with graph A in the original letter, 2 shows that the effect of immortal time introduced in the previous analysis is rela- tively small (note that for reasons of comparison the HR of untreated versus treated patients is presented in Fig. 1). The main reason for the limited effect of immortal time bias is that treatment in most cases started relatively soon after operation (median, 53 days; lower quartile, 21 days; upper quartile, 154 days) and that relatively few deaths occurred in controls or treated patients within the range of time before treatment was initiated. When estimating the HR for patients starting treatment before and after 53 days after operation (i.e., median) the HRs were 0.42 (95% CI: 0.26–0.70) and 0.40 (95% CI: 0.23–0.69), respectively. Thus, timing of treat- ment start did not substantially modify the HR, further limit- ing the potential effect of immortal time bias. When analyzing the treatment association as function of time since treatment start, divided into half year intervals we observed that the treatment association seems most pro- nounced within the first 12 months and thereafter diminishes with time since treatment start (Supporting Information in Table 1). The potential for immortal-time bias is greater in the subgroup of patients who received at least 6 months of valganciclovir treatment. Calculating the rate ratio can serve as an illustration of potential immortal-time bias. 5,6 Support- ing Information in Table 2 gives estimated rate ratios before and after adjustment for potential immortal time of subjects treated for at least 6 months and for those treated for less than 6 months. Both rate ratios are shifted toward a lesser treatment effect. The rate ratio for the 10 patients treated for less than 6 months was greater than unity, indicating a worse outcome than in untreated controls. The shifts in rate ratio for those treated for at least 6 months is less dramatic and remains well below unity, even though immortal time Figure 1. Survival curves with respect to treatment estimated from a treatment-stratified Cox regression model allowing for delayed entry in the treatment group. (Note that the x-axis starts at day 9 since the first patient receives treatment on that day and it will therefore constitute the effective origin with respect to the treatment effect.) Letter to the Editor Int. J. Cancer: 135, 248–249 (2014) V C 2013 UICC International Journal of Cancer IJC