Previous research has demonstrated that antibiotic resistance often comes with a fitness cost, the magnitude of which may depend on growth conditions, mechanism of resistance, genetic background, and the presence of compensatory mutations 
. In vitro
and animal model experiments are capable of analyzing the influence of these variables, but in humans, even quantifying an average fitness cost of drug resistance is made difficult by the increased complexity and decreased control introduced by working with field data. As a result, most epidemiological studies interested in the fitness cost of drug resistance in S. pneumoniae
have looked at correlations between naturally occurring changes in prescription rates and the prevalence of corresponding antibiotic resistances 
. Although this has demonstrated a significant relationship between the two factors, lack of an underlying transmission model and relatively small changes in antibiotic consumption have hindered the estimation of relative fitness values.
In this study, we have used data gathered as part of a longitudinal trachoma elimination study to examine the fitness cost of macrolide resistance in S. pneumoniae
. Observed macrolide resistance rates rose from 28.2% following four rounds of biannual mass antibiotic distribution, to 76.8% after two subsequent rounds, followed by a drop to 20.6% twenty-four months after the conclusion of the program. Fitting a mathematical model to these data, we estimated the relative fitness for strains harboring two different mechanisms of macrolide resistance, and showed how these estimates may depend on assumed transmission parameters and antibiotic effectiveness. Calculated fitness costs increased significantly with higher antibiotic efficacy against drug-sensitive strains, and longer durations of infection (1/
). These results agree with theory. Respectively, a larger available ecological niche, or fewer infection cycles, would both be expected to magnify the competitive disadvantage of drug resistance given identical data.
The clinical significance of these differing fitness costs can be measured, in part, by antibiotic-resistant strains' probability of survival versus time. Our model predicts that, in the base scenario, it may take up to 5 years for macrolide resistant strains of S. pneumoniae
to reach high probabilities of elimination by competition alone. In fact, this number may be an underestimate of the true value, as our model does not allow for the stable coexistence of pneumococcal strains. Previous modeling studies have shown that coexistence can be explained by simultaneous carriage of multiple strains 
, a phenomenon we were unable to model due to a lack of data on superinfection. However, the decreased complexity of the model allowed us to follow the probability of all community-level infection states over time, better capturing the stochastic effects that are important in small populations.
It should also be noted that due to bias inherent in our phenotyping method, any individual strains with both mechanisms of resistance would be classified as erm
B by our protocol. A study from a different area of Ethiopia observed these strains to constitute 6.6% of resistant samples after 4 rounds of azithromycin treatment 
. This misclassification would most likely increase the estimated fitness cost for the erm
B strain, although low prevalence of mef
B+ strains would attenuate this effect.
In addition, our model does not allow for strain reintroduction, or for the amelioration of fitness costs by compensatory mutation, an adaptation that has been observed to develop quickly in other organisms 
. The rate of compensatory mutation is only one part of the picture, however. In pneumococcal, both mechanisms of macrolide resistance are mediated by entire genes rather than single nucleotide polymorphisms (SNPs). As a result, the genetic target for mutations that attenuate or even eliminate resistance is comparatively much larger, consequently making these events far more likely. In other words, in the absence of antibiotic selection against pneumococcus, we would expect to see two opposing effects not captured in our model. Mutation could reduce antibiotic resistance, as well as the fitness cost associated with it.
Notwithstanding uncertainty in the long-term prediction of resistance rates, the existence of an epidemiologically verified fitness cost of macrolide resistance in S. pneumoniae suggests that intraspecific competition may be exploited to at least partially reclaim antibiotic efficacy after the development of high levels of resistance. In practice, however, this process may be highly variable and take years even under the most optimistic scenarios. Although it is quite possible that the fitness cost of macrolide resistance is sufficient to ensure its eventual elimination in the absence of antibiotic selection, this process takes time, and prevention is likely the best policy in the fight against resistance.