^{1}

^{2}

^{3}

^{2}

^{2}

^{2}

^{1}

^{3}

The authors have declared that no competing interests exist.

Conceived and designed the experiments: VJ HCG RMF PME DH. Performed the experiments: VJ. Analyzed the data: VJ. Contributed reagents/materials/analysis tools: RMF. Wrote the paper: VJ HCG RMF PME DH. Maintained database: RMF.

Blinding trachoma is targeted for elimination by 2020 using the SAFE strategy (Surgery, Antibiotics, Facial cleanliness, and Environmental improvements). Annual mass drug administration (MDA) with azithromycin is a cornerstone of this strategy. If baseline prevalence of clinical signs of trachomatous inflammation – follicular among 1-9 year-olds (TF1-9) is ≥10% but <30%, the World Health Organization guidelines are for at least 3 annual MDAs; if ≥30%, 5. We assessed the likelihood of achieving the global elimination target of TF1-9 <5% at 3 and 5 year evaluations using program reports.

We used the International Trachoma Initiative’s prevalence and treatment database. Of 283 cross-sectional survey pairs with baseline and follow-up data, MDA was conducted in 170 districts. Linear and logistic regression modeling was applied to these to investigate the effect of MDA on baseline prevalence. Reduction to <5% was less likely, though not impossible, at higher baseline TF1-9 prevalences. Increased number of annual MDAs, as well as no skipped MDAs, were significant predictors of reduced TF1-9 at follow-up. The probability of achieving the <5% target was <50% for areas with ≥30% TF1-9 prevalence at baseline, even with 7 or more continuous annual MDAs.

Number of annual MDAs alone appears insufficient to predict program progress; more information on the effects of baseline prevalence, coverage, and underlying environmental and hygienic conditions is needed. Programs should not skip MDAs, and at prevalences >30%, 7 or more annual MDAs may be required to achieve the target. There are five years left before the 2020 deadline to eliminate blinding trachoma. Low endemic settings are poised to succeed in their elimination goals. However, newly-identified high prevalence districts warrant immediate inclusion in the global program. Intensified application of the SAFE strategy is needed in order to guarantee blinding trachoma elimination by 2020.

Trachoma, the world’s leading infectious cause of blindness, is scheduled for elimination by 2020. Reaching this elimination target depends on successful implementation of the SAFE strategy (Surgery, Antibiotics, Facial cleanliness, and Environmental improvements). Annual mass antibiotic distributions are key to breaking the cycle of transmission in a community. However, it is not clear how many annual mass treatments need to be carried out in order to achieve elimination. Our study analyzes the effect of mass antibiotic distribution on different baseline prevalence levels of trachoma, in order to assess factors that affect the success of reaching elimination goals. We find that the prevailing belief, which suggests that 3 annual mass treatments can achieve local elimination of trachoma at prevalences between 10–30%, and 5 annual mass treatments for districts above this benchmark, is probably incorrect. In fact, much longer intervals may be required with “business as usual” programmatic strategies, which often include skipped years of treatment. Districts with high prevalence levels may require more intense treatment strategies to eliminate trachoma. Intensified recommendations must be implemented without delay in order to reach the 2020 elimination deadline.

Trachoma remains the world’s leading infectious cause of blindness, although it has disappeared from much of the developed world due to advances in hygiene and sanitation. The World Health Organization (WHO) has classified it amongst the neglected tropical diseases (NTDs), as where it remains, it is concentrated among the world’s poorest populations. These communities live “at the end of the road,” beyond the reach of development infrastructure, and lack access to the basic sanitation measures that prevent disease transmission. Currently, WHO estimates that 232 million people live in endemic areas, 21.4 million have active trachoma, and 7.3 million suffer from trachomatous trichiasis (TT) and are at immediate risk of becoming blind [_{1–9}) [

In order to achieve sustainable elimination, effective implementation of each component of the SAFE strategy is essential. Treatment with Zithromax (azithromycin) successfully clears individual infections [_{1–9} prevalence is 10–29%, and at least five MDAs before an impact survey when baseline TF_{1–9} prevalence is ≥30% [

Many perceive these benchmarks to suggest that a certain number of years of treatment “guarantee” elimination, but this may be incorrect. Even in relatively low-endemic regions, elimination may take more than three annual MDAs [_{1–9} prevalence [_{1–9} prevalence is ≥50%, five years of annual treatment is likely not enough [

Given the increase in available research and programmatic data, these recommendations can be assessed and refined to allow trachoma control programs to appropriately plan and budget for elimination. In this study, we used a global dataset of baseline and impact surveys to assess the evidence base for the effect of MDA on trachoma prevalence, with the goal of determining whether improved recommendations can be developed in order to improve programmatic efficiency and ensure continuous progress towards elimination.

In order to effectively coordinate the Zithromax donation on behalf of Pfizer, the International Trachoma Initiative (ITI) maintains a comprehensive database of trachoma prevalence and Zithromax treatments performed around the world. This database allows ITI to effectively allocate drugs, and conduct forecasting and planning of programmatic scale-up [

Each observation in the database includes the following information, if available: active trachoma prevalence and the clinical sign used as an active indicator (TF or TF/TI), trachomatous trichiasis (TT) prevalence, age range of individuals surveyed for TF and TT, survey location, survey year, survey design and sampling methodology, and data source. Where multiple surveys were conducted at a given location, they were coded to indicate if they preceded or followed treatment. Where treatment was conducted, some entries include estimates of district population, reported antibiotic distribution in doses, and coverage (estimated as doses distributed divided by total population).

There is substantial variation between some of the surveys represented in the database. For example, the indicator used for active trachoma is a measure of circulating disease in a community. Though the WHO standard is to measure trachomatous inflammation—follicular (TF) among children aged 1–9 years (TF_{1–9}), some surveys assessed TF among school-aged children or children under 6 years old. All surveys included in the database used the simplified clinical grading system for trachoma [

While cross-sectional population-based prevalence surveys (PBPS) are considered the gold standard for assessing trachoma prevalence at a given location [_{1–9} prevalence is below 10% at district level [

We assessed the factors affecting change in prevalence over time in pairs of surveys collected at the same location. The database initially contained 2365 surveys. These represented 29 countries and were performed between the years 1992–2013. We censored 156 TRAs and 46 ASTRAs. Of the 2157 remaining surveys, 353 represented follow-up after treatment, 1318 represented baseline that preceded treatment, and the remaining 486 represented surveys that did not prompt treatment. All 1671 surveys that preceded or followed treatment were assigned unique IDs by location and matched. Matches were parsed into pairs corresponding to two prevalence surveys in the same location and ordered chronologically. Matched pairs were merged with data on treatment and coverage that used the same unique IDs by location. In areas where follow-up assessment was conducted at a smaller implementation level than the baseline survey (e.g. district surveys following a zonal survey), the follow-up data was averaged across the original unit of implementation to allow comparison.

We investigated adjustment factors where active disease prevalence was not measured as TF_{1–9}. In settings with TF prevalence exceeding 20%, the age-prevalence peak may shift such that younger individuals are more likely to have a greater share of disease burden [

Pairs were identified as representing MDA if any treatment was recorded between the survey dates, or if ITI coding indicated that MDA had taken place. All other pairs were considered to represent “background” prevalence change. Variables were created representing annual MDAs between treatment (number of MDAs that took place between baseline and follow-up surveys), number of years between surveys, number of years before treatment (years between baseline survey and first MDA), number of years since treatment (years between first MDA and follow-up survey), total annual MDAs (number of MDAs before the follow-up survey, regardless of whether they took place after the baseline survey), skips between (“treatment holiday,” or skipped years between annual MDAs), and total skipped years (any years without treatment before the follow-up survey and after the beginning of treatment). See

This represents Years Before (here, 3 years after baseline survey and before MDA starts), Rounds Between (here, 3 MDA rounds between surveys), and Years Since (here, 5 years since treatment began). Skipped years are coded as follows: Skips Between (here, 1 skipped year between treatment rounds), and Total Skips (any years without treatment since treatment began, here, 2).

As all temporal information in the database is based on calendar years, discrimination between time intervals smaller than a year was not possible. Thus, a given “year” could be as short as 12 months or as long as 23 (e.g., if a baseline survey took place at the beginning of one calendar year and an MDA took place at the end of the next calendar year). Coding proceeded on the assumption that baseline surveys would be followed by treatment, while impact surveys followed treatment. Instances of anomalous code were manually inspected and cleaned. The final dataset had 170 pairs of surveys corresponding to baseline and follow-up after MDA, and 112 pairs that did not correspond to MDA. All of these represented population-based prevalence surveys.

In order to perform ordinal logistic regression modeling (described below), we created a categorized ordinal variable for TF_{1–9}. TF_{1–9} categories were specified based on the thresholds that define current WHO recommendations for treatment [

Categories are based on WHO recommendations for elimination.

Coverage data, applicable only to the treatment dataset (since the background dataset did not by definition involve MDA), was only reported in 2010–2012. Therefore, coverage data was available for the end of the treatment cycle for only those survey pairs whose treatment interval included at least one of these years: this was true of just 52 (approximately 31%) of the survey pairs in the treatment dataset. We therefore omitted this variable from modeling.

The final dataset contained 282 pairs of surveys, which were conducted between 1996–2013. We used SAS 9.4 (SAS Institute, Cary, NC, USA) to produce descriptive statistics of the dataset (_{1–9} prevalence at follow-up. Stepwise selection and backwards elimination strategies, with entry and stay criteria of α = 0.10, respectively, were used for model building, with all possible variables included at the outset. Aikake Information Criterion (AIC) was used to compare models. The assumption of linearity was confirmed using an overall F test, as well as by plotting the residuals of the explanatory variables. Univariate and multivariate logistic regression models were fitted to banded TF_{1–9} prevalence at follow-up (see _{1–9} prevalence. Stepwise selection and backwards elimination were again used to determine final model candidates. Maximum likelihood was used to estimate the coefficients for model predictors [

“Treatment” (with MDA)^{a} |
“Background” (no MDA)^{a} |
|||
---|---|---|---|---|

No. | % | No. | % | |

Burkina Faso | 18 | 10.8 | 37 | 33.0 |

Ethiopia | 25 | 15.0 | 10 | 8.8 |

Ghana | 23 | 13.8 | 0 | 0.0 |

Mauritania | 20 | 12.0 | 0 | 0.0 |

Nigeria | 4 | 2.4 | 50 | 44.6 |

Vietnam | 25 | 13.2 | 0 | 0.0 |

Burundi, Guinea, Guinea Bissau, Kenya, Malawi, Mali, Morocco, Mozambique, Nepal, Niger, Sudan, Tanzania, The Gambia | 55 | 32.4 | 15 | 13.4 |

Category 0: <5% | 14 | 8.2 | 14 | 12.5 |

Category 1: 5–9.9% | 27 | 15.9 | 22 | 19.6 |

Category 2: 10–29.9% | 96 | 56.5 | 48 | 42.9 |

Category 3: 30–49.9% | 23 | 13.5 | 26 | 23.2 |

Category 4: >50% | 10 | 5.9 | 2 | 1.8 |

Category 0: <5% | 89 | 52.4 | 46 | 41.1 |

Category 1: 5–9.9% | 30 | 17.7 | 21 | 18.8 |

Category 2: 10–29.9% | 37 | 21.8 | 36 | 32.1 |

Category 3: 30–49.9% | 14 | 8.2 | 8 | 7.1 |

Category 4: >50% | — | — | 1 | 0.9 |

1–2 years | 25 | 14.7 | 1 | 0.9 |

3–4 years | 51 | 30.0 | 13 | 11.6 |

5–6 years | 43 | 25.3 | 36 | 32.1 |

7–9 years | 35 | 20.6 | 14 | 12.5 |

>10 years | 16 | 9.4 | 48 | 42.9 |

0 Rounds | — | — | 112 | 100 |

1–3 Rounds | 91 | 53.5 | — | — |

4–5 Rounds | 36 | 21.2 | — | — |

>5 Rounds | 18 | 10.6 | — | — |

^{b} |
||||

^{c} |
||||

0 Years | 56 | 32.9 | — | — |

1–2 Years | 58 | 34.1 | — | — |

3–6 Years | 26 | 15.3 | — | — |

7+ Years | 5 | 2.9 | ||

^{b} |
— | — | ||

1–3 Years | 27 | 15.9 | — | — |

4–5 Years | 80 | 47.1 | — | — |

6+ Years | 38 | 22.4 | — | — |

^{b} |
— | — | ||

Any data, 2010–2012 | 52 | 28.4 | — | — |

— | — |

^{a}Pairs were sorted into the "treatment" dataset if any MDA had occurred in the interval between them.

^{b}41 pairs in the treatment dataset were missing data on when treatment occurred.

^{c}These values represent the time interval between the baseline survey and the actual start of treatment.

In the treatment dataset, 28 observations coded as representing MDA but missing data on treatment were dropped from the linear and logistic models due to missing predictor values. Pairs dropped included data from Ghana, Nigeria, Tanzania, The Gambia, and Vietnam.

Using several selection strategies in generalized linear modeling, we included the following variables in the final model for the treatment dataset: baseline TF_{1–9} prevalence (0.13, 95% CI: -0.17, 0.43), rounds of MDA (-2.59, 95% CI: -4.47, -0.71), years since treatment began (1.80, 95% CI: 0.67, 2.93), years before treatment began (-0.94, 95% CI: -1.79, -0.17), and the interaction between rounds of MDA and baseline TF_{1–9} prevalence (0.062, 95% CI: 0.003, 0.12). These were significant at the 0.05 level, with the exception of baseline prevalence, which also exhibited collinearity with the interaction term but had to be retained for a hierarchically well-formulated model. The final multivariate model, specified below, had an r^{2} value of 0.40:

In contrast, the best model fit to the background dataset (without MDA) accounted for only about 8% of the variation in the data, demonstrating that these model parameters do not do a good job of accounting for TF_{1–9} prevalence change in the absence of treatment.

Univariate ordinal logistic regression performed on the treatment dataset (_{1–9} prevalence was significantly associated with reduced likelihood of achieving lower categories of follow-up TF_{1–9} prevalence. Years since treatment began and total skipped years since treatment began were also significant. Increased number of annual MDAs and years skipped between annual MDAs also showed a non-significant trend towards association with reduced likelihood of reduction.

“Treatment” (with MDA)^{a} |
||||
---|---|---|---|---|

No. (%) | OR | 95% CI | ||

Baseline TF Prevalence | ||||

Years Between Surveys | 170 (100%) | 1.02 | (0.94,1.11) | 0.635 |

Years Before Treatment Start | 145 (85%) | 1.13 | (0.96, 1.32) | 0.142 |

Rounds Between Surveys | 145 (85%) | 0.91 | (0.76,1.10) | 0.343 |

Years Skipped During Treatment Interval | 149 (88%) | 0.81 | (0.64,1.03) | 0.0850 |

^{a}TF_{1–9} prevalence at follow-up is measured in five ordered categories (<5%, 5–9.9%, 10–29.9%, 30–39.9%, 40–49.9%, and >50%).

A multivariate ordinal regression model fitted to the treatment dataset was used to model the odds of reduction to a lower category of follow-up TF_{1–9} prevalence. The proportional odds assumption was satisfied for this model. An increase in the following was associated with significantly lower odds of TF_{1–9} prevalence reduction (see _{1–9} prevalence (OR = 0.92, 95% CI 0.89–0.94), and years since treatment began (0.77, 95% CI = 0.61–0.97). However, an increase in annual MDAs (OR 1.56, 95% CI 1.16, 2.10) and years before treatment (OR 1.30, 95% CI = 1.08, 1.57) were associated with significantly increased odds of TF_{1–9} prevalence reduction. Censoring of the “super-district” observations, which used mean follow-up TF_{1–9} prevalence to account for baselines measured at the zonal level, did not have a significant effect on these ORs.

The unmodeled data demonstrated a general trend of reduction from baseline to follow-up in the treatment dataset, though this was more pronounced at lower baseline prevalence levels (see _{1–9} prevalence levels in the multivariate logistic model (see _{1–9} prevalence, the probability of reduction to below 10% given a 30% baseline TF_{1–9} prevalence was 56%. At higher baseline endemicities, the point estimate for probabilities became lower, and the error increased. So while a 56% probability of reduction was predicted for a baseline TF_{1–9} prevalence of 30% given 3 annual MDAs, this was not statistically significant. As number of MDAs increased, the confidence interval narrowed, such that a 64% chance of reduction from 30% baseline was predicted for 5 treatment rounds. Even if the number of MDAs was increased to 10 for an area at 50% endemicity, the probability of reduction (estimated at 42%) was non-significant.

Data is unmodeled.

Each probability is cumulative over the lower ordered values, such that the probability of being <30% includes the probability of being <10% and <5%. Predictions not significant at α = 0.10 are in gray.

Although various simple measures of skipped years were not significant in the multivariate model, an increase in years since treatment began was significantly associated with reduced odds of prevalence reduction, such that adding a year to the treatment cycle (without a corresponding increase in treatment rounds) led to about a 5% reduction in the probability of success achieving reduction below 10%. The model also predicts increasing success with a waiting period before implementing treatment.

In this study, using data collected in a programmatic context over ITI’s 15-year history, we have demonstrated that the context in which mass drug administration for trachoma is conducted may be as important as the number of annual rounds implemented. Hyperendemic districts (baseline TF_{1–9} prevalence >50%) should implement at least seven MDAs before considering an impact survey, while relatively low-endemic districts (<20% baseline TF_{1–9} prevalence) likely could resurvey after three annual MDAs. However, our models are built using data that represents the imperfect world in which trachoma control programs have operated, with skipped treatment years and little data on antibiotic coverage and improvements in hygiene and sanitation. The context in which MDA is implemented is also crucial, and is likely key to successful elimination of trachoma.

Some of the principles demonstrated by our models regarding treatment context are well recognized. Trachoma tends to decline slowly on its own, probably due to the effects of gradual development and improvements in hygiene and sanitation [_{1–9} prevalence demonstrates that the effect of MDA varies at different endemicities. We had limited ability to investigate interactions between variables due to insufficient power and a small number of potential variables. As such, although the interaction term in the linear models shows that a higher baseline prevalence is less responsive to treatment, neither this term nor a potentially interesting interaction between baseline prevalence and skipped years could be included in the logistic models due to unacceptable levels of multicollinearity. However, in all the models, skipped years, or additional years since treatment began made reduction less likely. We see this effect despite the fact that a single “year” in our data may represent anywhere from 12 to 23 months, given that reporting is agnostic to timing of surveys and treatment during the calendar year.

We assessed the combined effects of these variables by generating predictions for various treatment schemes. The multivariate logistic model predicts that increasing the number of annual MDAs leads to a higher probability of TF_{1–9} prevalence reduction. No matter how many continuous MDAs are conducted, achievement of the elimination target levels becomes less likely as baseline prevalence increases. Of the ten districts in the treatment dataset with baseline TF_{1–9} prevalence >50%, none showed reduction to below 5%, and only one achieved reduction to below 10%, despite the application of up to seven annual MDAs (see _{1–9} prevalences between 30–50%, only about half of the districts achieved reduction below 10%.

The model suggests, therefore, that low endemic districts (<20%) are likely to achieve reduction to below 10% after three annual MDAs, and should be resurveyed at that time. However, at 30% baseline TF_{1–9} prevalence, the model predicts a 56% chance of reduction to below 10%. This probability dwindles as baseline TF_{1–9} prevalence increases. From the limited available evidence, even 7 annual MDAs were insufficient in hyperendemic districts (>50% TF_{1–9} prevalence) to make a meaningful public health difference. In such programmatic contexts, over 7 annual MDAs may be necessary to achieve the target. These findings are supported by other studies: in a programmatic context in Mali, three annual rounds of MDA were not sufficient at baseline prevalences of close to 30% [

Once again, our models do not represent the effect of MDA conducted in controlled conditions. It is likely that many of the districts in our dataset did not achieve their prevalence reduction goals due to inconsistent application of the SAFE strategy. For example, most of the high endemic districts experienced discontinuous treatment. As described, skipped treatment years significantly decrease the probability of TF_{1–9} prevalence reduction. Our models also omit data on other factors known to influence the effect of MDA, such as treatment coverage [

We also lack measures of hygiene and environmental factors, the F and E components of the SAFE strategy. Reduction in trachoma has been associated with clean faces and hygiene indicators [

Despite these omissions, our results are valuable precisely because they represent the effect of MDA as it is conducted by trachoma control programs. Although low endemic districts are likely to succeed in their elimination goals under the current WHO recommendations, we must consider carefully how to support the remaining districts with baseline TF_{1–9} prevalence over 30%. With just under five years left before the 2020 elimination goal, those districts must plan for intensified treatment programs. They may consider alternatives such as targeted treatment [

In order to continue our progress towards trachoma elimination, we must emphasize the WHO recommendations that call for programmatic continuity, which should be attainable even in countries where program implementation is difficult, given increased donor support. We must also emphasize the importance of antibiotic coverage, hygiene education, and sanitation improvements. This should start at the level of the data we collect. We cannot track progress, measure success, or even understand what success looks like for variables we do not measure.

Trachoma serves as an object lesson that antibiotic interventions, such as azithromycin mass treatment, can only go so far in the context of poor development. With increasing rounds of MDA, we may eventually reduce TF_{1–9} prevalence to below 5%, even in the most high-endemic districts remaining. Our data suggests that such districts ought to prepare for extended MDA timelines. However, we should not rely on antibiotics alone to achieve trachoma elimination. The most effective and efficient solution is likely to implement all aspects of the SAFE strategy, which recognizes that though high-coverage, continuous MDAs are essential, clean water and good hygiene may be as important. For programs seeking real and sustainable elimination, it may be that no amount of time is long enough to achieve trachoma elimination without lasting change of the environment in which it persists.

Data on age and prevalence from the PRET and ASANTE trials was used for standardization. The collection and maintenance of the database necessary for this study has been made possible by the dedication of numerous ITI staff members. Finally, we are indebted to the numerous survey teams who have, over the years, collected the data used in this analysis while distributing sight-saving antibiotics.