Mathematical model framework
The model in this study expanded on a previously developed chlamydia transmission model, described in detail elsewhere [17]. A deterministic, compartmental pair-formation modeling approach was used, in which the population living in a defined location was divided into a range of different categories, reflecting stratification by age (15-18y, 19-24y, 25-39y, 40-54y), by sex (heterosexual men and women), by sexual partnership category (never had sex, sexually active single people, and people in a long-term partnerships), and by sexual activity level (higher and lower sexual activity defined by frequency of short-term relationships). Long-term partnerships were represented as distinct compartments in the model, stratified based on the age and infection status of both partners. For this study, we expanded the modeling framework by dividing the 15–18 year-old population into those in intervention schools and those outside of the intervention schools but in the surrounding communities (Fig. 1). In the two youngest age groups (15-18y and 19-24y), the model included a fraction of the population that was not yet sexually active, with age-specific rates of transition into the sexually active population; these rates were time-varying in the youngest age group to allow for changes in the population at-risk for chlamydia. We assumed that all people ages 25 and older were sexually active. Sexually active single people of any age in the model could have short-term relationships; sexually active people older than 18 years could enter long-term partnerships. Chlamydia infection status was represented using a susceptible-infected-susceptible structure.
Settings
We applied the model in three different settings, spanning a range of reported chlamydia diagnosis rates. Philadelphia and Chicago are large cities with relatively high rates of chlamydia diagnoses and with existing high-school-based screening programs. Rural Iowa has relatively low chlamydia diagnosis rates and does not have an existing school-based screening program. Among women aged 15-24y, rates of reported chlamydia diagnoses in 2015 were 7433 per 100,000 in Philadelphia [18], 6715 per 100,000 in Chicago [19] and 2807 per 100,000 in the state of Iowa [20], compared to 3376 per 100,000 nationally [21]. For our analysis, we calculated reported diagnosis rates for Iowa excluding the 10 most populous counties to approximate the burden in more rural areas, which yielded a diagnosis rate of 1948 per 100,000 in 2015.
Philadelphia has the largest and longest-running high-school-based chlamydia screening program in the United States, established during the 2002–2003 school year. Between 2002 and 2015, there were 10,000–20,000 tests conducted annually, and the schools enrolled in the program covered approximately 55% of the 15–18-year-old population in the city. In Philadelphia’s intervention schools student participation declined over time, from approximately 30% of the student population during the 2002–2003 school year to 16% during the 2014–2015 school year. Chicago began a screening event program in 2010, following a pilot study in 2008. Between 2010 and 2013, the program was implemented in schools covering approximately 12.5% of the 15–18 year-old population in the city, and the number of tests conducted yearly increased from 2200 to 6900.
Data and model calibration
Model calibration was undertaken using a Bayesian framework [22] operationalized with Incremental Mixture Importance Sampling [23]. Each setting was calibrated independently. The calibration approach yielded a joint posterior probability distribution for the parameter values, informed by a combination of specified prior distributions and the data likelihood. For a subset of parameters, we resampled from distributions defined by posterior estimates in our previously calibrated national model [17]. Table S1 in Supplement 1 describes each parameter that was varied in the calibration and its prior and posterior ranges. To estimate community-level chlamydia transmission dynamics in the absence of school-based screening programs, we calibrated the model to three sources of setting-specific data: reported chlamydia diagnosis rates, chlamydia positivity estimates among 15–18 years olds, and the proportion of the high school population who report having ever had sex.
Sex- and age-stratified chlamydia reported diagnosis rates were obtained for the three study settings [18,19,20], and the model was calibrated to data prior to the initiation of school-based screening programs in the cases of Philadelphia and Chicago. For Philadelphia, we used the average of diagnosis rates during 2002–2015 as a proxy measure, as there were no data available prior to 2002. For Chicago, we used diagnosis rates during 2000–2009. For Iowa (excluding the 10 most populous counties), we used the diagnosis rates during 2000–2015.
Both Philadelphia and Chicago provided estimates of chlamydia positivity in intervention schools. For calibration, we used positivity estimates from the first year available for each program as a proxy for the baseline chlamydia prevalence among the population of sexually active 15–18 year-olds in each city. For Iowa, we used a chlamydia positivity estimate from rural family planning clinics [24]. To calibrate rates of sexual initiation among 15–18 year olds, we used city-level data from the Youth Risk Behavior Survey (YRBS) [25] on the percentage of high school students who reported ever having had sex (state-level data used for Iowa).
The model was initially run to equilibrium using time-invariant parameters. We introduced time-varying parameters governing rates of sexual initiation starting in 1991, to correspond to years of data availability from YRBS, and time-varying parameters relating to screening rates and completeness of reporting on diagnosed infections in the general population starting in 2000, to correspond to the first year of case report data included in the likelihood.
Analysis
We conducted two sets of analyses. In the first set of analyses, using only the models for Philadelphia and Chicago, we examined the potential influence of different observed programmatic features from the school-based screening programs in those cities on the estimated benefits from the programs. We used the calibrated models to estimate chlamydia prevalence in the 15–18 year-old population in the absence of school-based screening. We then estimated the potential impact of school-based screening using program data on the fraction of the 15–18 year-old population enrolled in intervention schools (“coverage”) and on student participation within intervention schools over time as input parameters. We used information on the highest and lowest yearly participation levels observed in Chicago and Philadelphia to determine ranges of participation levels. In Philadelphia, we modeled declining student participation over the 12 program years, from 30 to 16%, while holding population coverage constant at 55%. For Chicago, we modeled an increase in student participation over the 4 program years, from 13 to 40%, while holding coverage constant at 12.5%.
In the second set of analyses, we used the calibrated models for all three settings to explore a set of hypothetical scenarios with varying coverage levels and participation trends, and each program scenario implemented over a 12-year period. Population coverage of school-based screening among ages 15–18 years was varied from 20 to 60% across cities. Three different time trends for participation were modeled: 1) stable participation at 50%; 2) declining participation, from 50 to 8%; and 3) increasing participation, from 50 to 90%.
In both sets of analyses, parameter uncertainty was captured by running multiple model simulations, each based on sampling one set of parameter values from the joint posterior distribution of the parameters estimated through calibration. For each draw of model parameter values, we ran both a baseline (no-screening) scenario and each of the different program scenarios. Intervention effects are summarized in terms of prevalence in different scenarios among 15–18 year-olds as well as absolute differences in prevalence between intervention scenarios and the no-screening program baseline. Reduction in prevalence was calculated by taking the difference from each baseline draw and its respective counterfactual, and calculating an overall reduction across the reductions at the draw-level. We report the median and 95% credible interval for each outcome.