We thank Cyprian Mostert for his interest in our work (Jun 2018). We assessed hospitalization rates among patients starting antiretroviral treatment (ART) in South Africa. We found that “hospitalization rates peaked during the first year since ART” but then “declined significantly over time on ART.” We interpreted the peak as reflecting HIV-related morbidity resulting from late treatment initiation and the decline as reflecting the health-improving effects of ART. Contrary to Mostert’s description of our study, we did not empirically measure the effect of early treatment but rather simulated how many hospital days could have been avoided if ART had been initiated earlier.

Mostert suggests that the “toxicity of stavudine” could explain the high rate of hospitalizations around treatment initiation. Stavudine was replaced by the less toxic tenofovir after April 2010. We therefore repeated our analysis using only data for after April 2010. Our results remained essentially unchanged: Hospitalization rates declined by a factor of 5.59 (p = 0.007) rather than by a factor of 5.47 (p < 0.001) from ART initiation to four years later.

Mostert also points out that Kenneth Land and colleagues showed that incorrect specification of the error term distribution in Poisson models can lead to underestimation of the uncertainty of regression coefficients and thus false significant findings. Therefore, we repeated our analysis with robust standard errors, one of the solutions described by Land and colleagues. Again, our results were virtually unchanged: All but one of the coefficients remained significant at their original levels.

Our additional robustness checks thus support our original conclusions: ART substantially reduces hospitalizations in South Africa, and earlier ART initiation could strengthen this effect.