We would like to thank Haine et al. for their interest in our study and to provide answers to the comments made in their letter to the editor [1,2]. First, we would like to stress that all of the analyses performed in our study were undertaken by an independent statistical expert team (not linked to the principal investigator) and that the authors all declared no conflicts of interest.
We agree with Haine et al. on the determination of the number of variables that should be tested in a multiple logistic regression model. The appropriate ratio of events to variables that should be used in predictive models remains poorly defined, according to the guidelines provided by Hosmer and Lemeshow [3,4] and also those of Feinstein for using multivariable logistic regression. However, there is general agreement that this important assumption could have a negative effect upon the statistical results. The issue of model, called overfit or underfit, has been reported in the statistical literature but its implications are not widely recognized or indeed well understood.
This issue is more often discussed in studies examining genetic testing for a specific susceptibility. Our approach is not strictly similar. We aimed to test the true relationship between the delta value of CD34+KDR+ and the target lesion revascularization (TLR) event. The multivariate model that we performed was only used to adjust our results for controlled variables. These six other variables identified as controlled variables are well-known predictors of TLR in the literature. However, to answer this comment more fully, we performed six new logistic regression models testing two parameters per model (delta CD34+KDR+ (%) and age, delta CD34+KDR+ (%) and sex, delta CD34+KDR+ (%) and active smoker, delta CD34+KDR+ (%) and diabetes, delta CD34+KDR+ (%) and stent diameter, delta CD34+KDR+ (%) and stent length). In all of these analyses, the P-value of delta CD34+KDR+ (%) was lower than 0.05 (0.018, 0.019, 0.022, 0.019, 0.016 and 0.018 respectively). These findings strongly support the results of our study and suggest a link between delta CD34+KDR+ (%) and TLR in patients receiving a bare metal stent. Regarding the number of cellular variables that were tested, we could provide further details. We agree with Haine et al. that multiplicity should be prevented. But we would like to underline that we did not test 16 variables, but only four that were related to the delta parameters. Indeed, the analyses were focused on the dynamic aspect (i.e. the dynamic change of markers between baseline and the peak value [H6 or H24]). Table 2 in our paper provided only descriptive data for the two groups for each parameter at each time [H0, H6 and H24]. Therefore, only four parameters were really tested against the presence/absence of TLR. We agree that we have tested many hypotheses on the same dataset. It would be reasonable to carry out multiple tests and we could provide such analysis. The P-value adjustments were provided using the Proc Multtest of SAS 9.2 software (SAS Institute Inc, Cary, NC, USA). Regarding the delta CD34+KDR+ (%), the adjusted P-value was equal to 0.03 using two adaptive methods (Holm adaptive and Hochberg adaptive). The univariate approach does not allow one to conclude that delta CD34+KDR+ (%) is independently linked to the TLR event and, therefore, a multivariate approach was required to correct ‘false positive’ and ‘false negative’ effects.
Finally, selecting the in-stent late luminal loss instead of a dichotomous primary endpoint to define restenosis might resolve the problem of low events in a similarly sized population; however, in-stent luminal loss is not a clinical endpoint.