Browse/search for people

Publication - Professor Margaret May

    Methodological and statistical issues related to analysis of survival – Authors' reply

    Citation

    Trickey, A, May, M & Sterne, J, 2017, ‘Methodological and statistical issues related to analysis of survival – Authors' reply’. Lancet HIV.

    Abstract

    We are grateful for the opportunity to respond to the points raised by Alizadeh and colleagues. First, we addressed between-cohort heterogeneity (clustering) by stratifying all Cox models on cohort. This allows the baseline mortality hazard to vary between cohorts, while assuming a common effect of covariates across cohorts. We have found previously that effects of important prognostic factors such as CD4 count are consistent across cohorts. Webtable 1 (appendix p 1) of our paper shows that calendar time effects were consistent across subgroups defined by values of important prognostic factors.

    For the analyses of prognosis from 1 year after starting antiretroviral therapy, we presented analyses (see figure 2) in which we controlled for CD4 count and viral load measured at both start of treatment and at 1 year later. These measurements were all made before the start of follow-up in those analyses. We do not agree that it would have been appropriate to address the effects of these time-dependent variables through joint modelling of time-to-event and longitudinal outcomes: our analysis appropriately addressed the extent to which controlling for 1 year measurements attenuated associations with their baseline values.

    We agree that we cannot exclude selection bias as a result of informative censoring. As Alizadeh and colleagues note, we reported that some characteristics predicted loss to follow-up. However, this does not imply that censoring was informative, because our analyses conditioned on the variables that predicted loss to follow-up. Informative censoring is unlikely to have substantially biased our results because cohorts have good recording of deaths and linkage to death registries. Sensitivity analyses such as those proposed would be of interest but were outside the scope of our paper.

    We declare no competing interests.

    Full details in the University publications repository