Where have all the patients gone?

Evaluations of health-care programmes are only as good as the data and methods used, and as in all longitudinal research, loss to follow-up is among the prime spoilers. For cohort studies including randomised controlled trials, loss to follow-up attenuates the ability to examine outcomes or the eff ect of an intervention and predictors of these outcomes. Patient-level data from health facilities are subject to the same biases. Most patients are probably not lost at random and reasons for loss to follow-up could be inherently linked to the eff ectiveness of clinical services. A range of statistical options are available (eg, imputation, censoring weights) to correct for loss to follow-up; however, the absence of randomness among those lost can reduce the validity of such methods. Fleming’s suggestion that “the preferred and often only satisfactory approach to addressing missing data is to prevent it”, is a particularly tall order in view of the exigencies of HIV clinical care, especially in subSaharan African settings with high HIV prevalence and suboptimum infrastructure. In fact, the costs and eff orts required to intensively track all missing patients is likelyprohibitive in most settings. In this issue of The Lancet HIV, Elvin Geng and colleagues off er a viable alternative for sizeable, chronic disease programmes; they investigate intensive tracing of a manageable random sample of patients loss to follow-up and incorporate their weighted fi ndings to those in the available clinic sample. This study includes the largest and most diverse group of sites in which this approach has been implemented. The pronounced diff erence between estimates of mortality and associated risk factors with and without this correction provides clear evidence of the power of this approach. Perhaps most telling were diff erences in mortality by clinic. The authors report results from 14 clinics in east Africa among fi ve programmes. Sample-corrected estimates of 3 year mortality in all clinics ranged from two-times to more than ten-times higher than in the naive (unadjusted) data. Additionally, the mortality rankings of four of the fi ve programmes changed when the corrected estimates were applied. If mortality is one of the crucial outcomes by which performances of sites and programmes should be judged, and we believe it is, valid facility-level estimates of patients’ outcomes are imperative. Interestingly, Cochrane and Grade guidelines established to weigh the evidence of diff erent health interventions set a low-bar, and consider randomised controlled trials with greater than 15% or even 20% loss to follow-up not to off er robust evidence. Notably, three of the fi ve programmes had loss to follow-up of less than 15%, yet nevertheless, correcting the mortality estimates with outcomes from those who were lost made tangible diff erences in mortality-based clinic performance. The sampling-based approach can benefi t further from assessments of its incremental cost and costeff ectiveness, and a more systematic approach to selection of clinics that are representative of districts, provinces, or regions. These issues are being addressed, along with the inclusion of point-of-care diagnostics and qualitative techniques among identifi ed patients in the ongoing Better Information for Health in Zambia (BetterInfo) study, funded by The Bill and Melinda Gates Foundation, and being done in a representative sample of sites in four provinces in the country. Even as The President’s Emergency Plan For AIDS Relief and national programmes make important advances in the collection of patient clinical and laboratory data, a great risk of inadequate and biased understanding of overall programme performance remains if those loss to follow-up are not systematically assessed. This threat is also present for macroevaluations of the overall eff ectiveness of the HIV response dependent on programme and modelled data. Geng and colleagues’ sampling-based method is a light-touch, secondgeneration moni toring and evaluation approach Published Online January 27, 2015 http://dx.doi.org/10.1016/ S2352-3018(15)00015-6