In this study we used NGS to screen for superinfection in 129 high-risk women and identified 9 cases of superinfection. Combined with previous studies
, a total of 21 cases of superinfection were detected among 146 women screened in this cohort. There was a statistically significant difference between the incidence of superinfection (2.61 per 100pys) and initial infection (5.75 per 100 pys), with a hazard ratio of 0.47 after adjusting for potential confounding factors. This suggests that HIV infection provides partial protection from subsequent infection.
The relatively large size of this cohort and high number of superinfection cases enabled us to detect for the first time a statistically significant difference between the incidence of initial infection and superinfection. This possibility has been proposed previously, though the studies were not designed and/or powered to detect a difference 
. In the largest incidence study prior to the present study, Redd et al.
screened a comparable number of individuals (149) in a lower-risk cohort and identified 7 cases of superinfection. The incidence of superinfection was not found to differ significantly from initial infection, but there was a trend for lower incidence of superinfection when controlling for baseline sociodemographic differences between the groups at risk of initial and superinfection. Analysis of our data using the same methods as Redd et al.
– Poisson regression with propensity score matching 
– was consistent with the results of our Andersen-Gill analysis, showing a significant difference in incidence, with an estimated incidence ratio of 0.48 (p
0.011) comparing superinfection to initial infection.
In addition to sample size, two strengths of our incidence analysis were our specification of infection timing to within a few months on average and our comparison of initial and superinfection risk within the same cohort. These enabled us to adjust for the same potential confounding factors in both the initial infection and the superinfection risk sets, using frequently collected time-varying covariate data. Particularly important, given the sequential nature of superinfection, was adjustment for calendar year to control for decline in infection risk in the cohort over time. The distributions of initial and superinfection events over calendar time were similar (Fig. S3
), suggesting community-level changes over time did not severely bias our analysis.
The ~two-fold reduction we found in the incidence of superinfection has a number of possible interpretations. First, it may indicate that the adaptive immune response elicited by initial infection provides partial protection from second infection. If this were the case, superinfection might preferentially occur early in infection, before the response has matured 
. In support of this idea, we found that, although superinfection occurred throughout the course of first infection, the incidence of superinfection was significantly lower than initial infection after the first 6 months of infection, but not earlier. This suggests that susceptibility to superinfection decreased over time, coincident with broadening and strengthening of HIV-specific immunity. Indeed, this has been suggested by two earlier studies, each documenting three cases of superinfection that occurred within the first year after initial infection 
If the difference in incidence we observed is due to a partially protective adaptive immune response, we would anticipate superinfection would preferentially occur with more distantly related viruses, more likely to escape the response. Using viral subtype and pairwise amino acid distance as surrogate measures of antigenic distance, our data provided no evidence of this effect. The majority of the 21 superinfection events we detected were intrasubtype, and the proportion of subtype A, C and D viral sequences was similar for the initial and superinfecting viruses, consistent with the subtype distribution in this cohort 
. The pairwise distance between initial and superinfecting variants was no higher than the distribution of distances between random pairs of singly-infected individuals from the Mombasa cohort. This may potentially be explained by limited sample size or insufficient simultaneously circulating subtypes. It also may be that sequence relatedness is a poor indicator of susceptibility to the immune response or the genome regions we analyzed are not critical antigenic determinants of protection.
Alternatively, it is possible that protective immune responses are not driving the protective effect we observed. Another potential explanation for the lower risk of superinfection is that HIV infection itself may reduce infection risk by depleting permissive target cells. On the other hand, chronic immune activation and immunodeficiency following HIV infection could increase susceptibility, potentially blunting protective effects 
. Thus, there may be a complex interplay of biological factors impacting HIV risk in an HIV-positive individual.
So far, studies of immune correlates of superinfection have yielded variable results – some suggesting neutralizing antibody deficits in superinfection 
, while others, including studies in the Mombasa cohort, detected no differences in antibody 
or cellular 
responses. A major challenge in these studies has been the identification and analysis of large enough numbers of superinfection cases: the small sample sizes in studies to date (three to twelve superinfected individuals) would restrict detection to only very large effects. Small sample size is just one factor that has made detecting immune deficits associated with superinfection challenging and contributed to variable results among studies. There has also been variation among published studies in the control groups used for comparison, including the time at which the response was analyzed relative to the time of superinfection and initial infection. Given the dynamic nature of the immune response, sample timing could impact measures in both controls and cases. Furthermore, precision in the estimated timing of superinfection varies between studies, and between cases, providing an additional variable. Divergent findings between studies may also reflect differences in the assays used and subtleties in the immune parameters they capture.
Our finding of lower risk of superinfection than initial infection provides greater impetus for larger-scale comprehensive analysis of multiple immune mechanisms, including both those analyzed in the smaller studies to date and, perhaps of more interest, those not characterized in prior studies. If the discrepancies in earlier studies reflect the fact that multiple immune parameters are at play, then examining a variety of immune responses in the same individuals in a larger cohort may be needed to define responses that contribute to HIV susceptibility following initial infection.
Like all studies, the study presented here has a number of limitations. Firstly, while our screening methods are among the most sensitive developed, it remains possible that some cases of superinfection were missed. In particular, reinfection by the same source partner is not captured by any existing methods. Additionally, our specification of the timing of superinfection was limited by the samples available to us. While follow-up was generally frequent in this study population, there were six superinfection cases where sample availability limited our ability to define the time of superinfection to within a one-year period. This uncertainty in superinfection timing did not affect our findings, as we found that whether we assumed in the incidence analysis that the true timing of superinfection was at the start, midpoint or end of the timing window, the results indicated that the incidence of superinfection was significantly lower than that of initial infection. Finally, as in all observational studies, residual confounding of our incidence estimate by behavioral changes and sexual network-level factors not measured or accounted for in our analyses remains a possibility. However, the fact that we compared initial and superinfection risk within the same cohort and collected covariate data at frequent intervals enabled us to minimize this issue to an extent not possible in previous studies.
This study provides the first robust evidence that HIV infection reduces the risk of subsequent infection. The underlying mechanism remains unclear, but this finding prompts exploration of correlates of protection from HIV in high-risk individuals who continue to be exposed after first infection. Furthermore, this study reinforces that superinfection occurs at a considerable rate, calling for studies of its impact on the clinical progression, transmission, and epidemiology of HIV.