Reanalysis of cluster randomized trial data to account for exposure misclassification using a per-protocol and complier-restricted approach

The intention-to-treat analysis of a cluster randomized trial of Wolbachia (wMel strain) mosquito releases for control of dengue (the Applying Wolbachia to Eliminate Dengue [AWED] trial) estimated a protective efficacy of 77.1% for participants resident in areas randomized to receive releases of wMel-infected Aedes aegypti mosquitoes. The limiting assumptions of ITT analyses in cluster randomized trials and the mobility of both mosquitoes and humans across cluster boundaries mean that the primary analysis is likely to have underestimated the full public health benefit. Using spatially and temporally resolved data on the distribution of Wolbachia mosquitoes and on trial participants' mobility collected during the AWED trial, this work performs a complier-restricted and per protocol re-examination of the efficacy of the Wolbachia intervention, taking explicitly into account human mobility and the uneven establishment and spillover of wMel mosquitoes. By applying the same estimation procedures as those performed in the primary analysis, we are able to identify the impact of exposure misclassification on estimated efficacy. Increased intervention efficacy was estimated in all analyses by the refined exposure measures. The complier-restricted analysis resulted in an estimated efficacy of 80.7% (95% CI: 65.9, 89.0). The effect of the refined exposure estimation on per-protocol results was even more striking, with an estimated 82.7% (71.7, 88.4) efficacy when comparing participants with an estimated wMel exposure of 80% or higher compared to those with 20% or lower. These reanalyses provide a case study of how human and mosquito movement can lead to underestimation of the intervention effect in trials of vector interventions, and indicate that the protective efficacy of Wolbachia is even higher than reported in the primary trial results.


Introduction
A breakthrough in efforts to curtail the global spread of dengue has recently come in the form of the intracellular bacterium Wolbachia (w Mel strain), which increases the resistance of Aedes aegypti mosquitoes -the primary vector of the dengue virus -to the replication and onward transmission of a number of arboviral diseases including dengue, Zika, chikungunya, and yellow fever [11,16,1,3].Encouraging results 1 from quasi-experimental field trials indicated that introgression of w Mel into local Ae.aedes populations was associated with reduced incidence of notified dengue cases [9,12,5].The efficacy of Wolbachia for dengue control was demonstrated experimentally in a gold standard parallel-arm cluster randomized trial (CRT) in Indonesia (the AWED trial, 'Applying Wolbachia to Eliminate Dengue'), which reported a protective efficacy of 77.1% (95% CI: 65.3%, 84.9%) against virologically-confirmed dengue in the primary intention-to-treat (ITT) analysis.[15] While highly promising, the ITT estimate is unlikely to capture the full intervention effect due to aspects of the study design that are not well handled by an ITT analysis, resulting in an underestimated intervention efficacy [4].In cluster randomized trials of community-delivered interventions, such as the w Mel deployments in the AWED trial, individual participants' true exposure status may be different to the cluster-allocated intervention status due to mobility of humans and spillover of the intervention across cluster boundaries.w Mel coverage is heterogeneous across time and space, within "intervention" clusters themselves and with observed spillover into "untreated" clusters during the later portion of the study period.Further, humans move outside of their cluster of residence in their daily routines.When such mosquito and human mobility extends the protective effect of the intervention to individuals in the control clusters, and dilutes the exposure of ostensibly 'treated' individuals who spend time in control clusters, it can result in traditional ITT analyses that underestimate the full public health benefit.
A first effort at moving beyond the ITT to examine the full protective effect of the Wolbachia intervention was a pre-specified secondary per-protocol analysis of the AWED trial data in which participants' exposure status was reclassified from a cluster-level binary to an individual-level weighted 'Wolbachia Exposure Index' (WEI) to account for self-reported mobility and measured differences in w Mel establishment across the study area.In this initial 'per-protocol' analysis, reported together with the ITT analysis in the publication of trial results, [15] the WEI was estimated in two distinct ways: 1) based solely on the measured cluster-level w Mel prevalence in Ae. aegypti mosquitoes in the participant's cluster of residence during the calendar month of participant enrolment, and 2) based on the cluster-level w Mel prevalence in each cluster the participant reported visiting in the week prior to illness onset, weighted by the proportion of total observed time the participant spent in each cluster.The methods are further described in the study protocol [2].The WEI based on cluster of residence alone resulted in very little exposure variability (i.e.most participants had WEI values near 1 or 0), making it difficult to perform reliable risk comparisons at intermediate levels of exposure.Dengue risk was significantly reduced in only the highest WEI stratum (WEI ≥ 80%) compared to the lowest (WEI < 20%), with efficacy estimated at 77.4% (95% CI: 62.1%, 85.6%), very similar to the ITT efficacy estimate.The WEI based on w Mel and travel history had a bit more variation in exposure for participants in each arm, and estimated a dose-response effect, suggesting statistically significant differences in protective efficacy for individuals with at least a WEI of 40% as compared to those with WEI less than 20%.[15] Both definitions of WEI used a somewhat crude estimation of time spent under protective w Mel cover, relying on cluster-level w Mel proportions in the month of enrollment; a more geographically and temporally localized w Mel proportion may more accurately capture an individual's protective exposure.Simulation work by others demonstrates that a greater protective efficacy is expected than observed in the initial ITT and per protocol analyses [4].Human and mosquito mobility occur on much finer scales than the 1 km 2 geographic regions ("clusters") randomized to treatment allocation.Ignoring the contamination effect of this movement has been shown to bias the efficacy estimate towards the null [4].
The present study re-estimates AWED trial participants' individual-level w Mel exposure using spatially and temporally resolved data on the distribution of Wolbachia mosquitoes and on trial participants' mobility collected during the AWED trial, in order to perform an improved per-protocol and a complier-restricted re-examination of Wolbachia efficacy in the AWED trial.This granular exposure data accounts explicitly for human mobility, as well as the spatial heterogeneities in w Mel coverage particularly arising from spillover of w Mel mosquitoes across the boundaries between treated and untreated clusters.By reducing misclassification in individuals' Wolbachia exposure status, these reanalyses of data from the AWED cluster randomized trial provide the most unbiased experimental estimate to date of the true efficacy of Wolbachia for dengue control.

Human mobility
w Mel-infected Ae. aegypti mosquitoes were released into 12 intervention clusters (of 24 total clusters) in Yogyakarta between March and December 2017.Patients presenting to local health clinics (puskesmas) between January 2018 and March 2020 with undifferentiated acute febrile illness of 1 to 4 days duration and aged between 3 and 45 years were invited to enroll in the AWED trial [2].Those who consented to enroll had a blood sample collected for dengue diagnostic testing and were asked about their movements during daytime hours (5am -9pm) in each of the 3 to 10 days prior to their illness onset, corresponding to the incubation period of a dengue virus infection.The 6306 participants included in the AWED analysis dataset reported spending the majority of their time at home (Fig. 1, median 68.8%, interquartile range 57.8-85.2%),consistent with the results of a baseline mobility survey in Yogyakarta [7].
Though 87.2% of participants left their residence at least once during the 8-day period (Fig S1A ), typically to attend school (56.8%) or work (19.2%) (Fig S2), 86% of the reported locations visited were within 1km of a participant's residence (Fig S3).The number of unique locations visited did not differ between adults and children, between study arm of residence, or between dengue cases and test-negative controls (Fig S1B -D).However, the aggregate time that participants spent at increasing distances from home increased with age, with only 28.5% (276/970) of young children 3-5 years and 32.1% (571/1778) children 6-10 years reporting any time spent further than 1km from home, compared to 54.5% (821/1506) adolescents 11-18 years and 64.9% (1330/2050) adults.More than a third of participants (39%) stayed within their cluster of residence throughout all 8 days.A median of 98.4% (IQR 68.0-100.0%) of participants' time was spent under the same intervention arm as their cluster of residence.The proportion of time under the intervention assignment showed no apparent differences for individuals in the intervention versus untreated arm.
w Mel heterogeneity Adult mosquitoes were collected from a fixed network of 455 traps approximately monthly throughout the 27-month trial period in order to monitor w Mel prevalence in Ae. aegypti over time in the intervention clusters and contamination in untreated clusters.Between January 2018 and March 2020, there were a total of 10,432 mosquito trapping events, only 10 of which had no Ae.aegypti present.A total of 87,679 Ae. aegypti mosquitoes were screened; 49,266 (56.19%) of which were detected to have w Mel.The majority of BG traps were located within the AWED RCT study boundaries (Fig. 2A), with 181 traps in the intervention region and 186 traps in the untreated region.An additional 11 traps were located in an untreated area on the southeast boundary of the RCT study site and 77 traps in a Wolbachia-treated area on the northwest boundary of the RCT study site; these areas had served as the untreated control and intervention area, respectively, in a quasi-experimental Wolbachia release study prior to the AWED randomized trial [9].Overall, 96.1% and 93.6% of mosquitoes screened in the quasi-intervention and RCT intervention areas contained w Mel, respectively.Only 9% of the mosquitoes captured in the quasi-untreated and 15% of mosquitoes from the RCT untreated areas had w Mel present.In the RCT intervention and quasiintervention areas, the marginal monthly capture and detection of w Mel in mosquitoes was fairly consistent across time.In the RCT untreated area, there was a marked increase in the proportion and absolute number of w Mel-infected mosquitoes detected in the final year of the study (Fig. 2B).

Complier-restricted reanalysis of Wolbachia efficacy
To examine the efficacy of Wolbachia among those who received the intervention as randomly assigned, we restricted the analysis to those participants who reported spending all of the potential DENV exposure period under the intervention assignment concordant with their cluster of residence ("compliers").Among the 6,306 participants in the AWED primary analysis data set, 3,114 participants (49%) stayed strictly within the RCT study area and within clusters in the same study arm as their cluster of residence, during the 3 to 10 days prior to illness onset: 1,359 test-negatives and 33 dengue cases in the intervention arm and 1,530 test-negatives and 192 dengue cases in the untreated arm.Applying the same modified odds ratio approach used in the primary analysis of the AWED trial [15] and described in an earlier methods paper [10], intervention efficacy in the complier-restricted subgroup was estimated at 80.6% (95% confidence interval: 65.9%, 89.0%), about 3 percentage points higher than that estimated in the primary analysis.
If we expand the definition of compliers to include participants who left the RCT study area but remained in areas with concordant treatment assignment to their cluster of residence, an additional 772 participants are included: 64 additional test-negatives in the intervention arm and 651 test-negatives and 57 dengue cases in the untreated arm.Intervention efficacy in this subgroup was similarly estimated at 79.7% (95% confidence interval: 67.2%, 87.4%).

Per-protocol reanalysis of Wolbachia efficacy
To examine the protective effect of Wolbachia exposure on dengue risk, regardless of compliance with intervention assignment, we constructed an interpolated w Mel surface to estimate each individual participant's Wolbachia exposure level.Figure 3 shows the inverse density weighted maps of w Mel coverage aggregated monthly across a 100×100 m grid overlaid on the AWED study site.Interpolating based on proximity in both space and time helps to stabilize the highly variable individual trap events.Across each panel, it is evident that w Mel is established at a high level (light regions) in intervention areas and, as suggested by Figure 2, begins to creep across the boundaries between the intervention and untreated regions by years 2 and 3 of the trial, leaving the centers of the untreated clusters uncontaminated (dark regions).The results from the leave-one-out cross-validation method used to determine the optimal hyperparameters for interpolation are provided in Appendix 2.
We approach per-protocol re-estimation of efficacy in two ways, mirroring the previous efforts towards obtaining individual-level WEI [2,15].We first estimate an individual's Wolbachia-exposure index (WEI) based on the spatiotemporally interpolated proportion of w Mel mosquitoes at the individual's residence on the reported day of illness onset, referred to as 'interpolated WEI (residence)'.Second, we use a weighted sum of the individual's interpolated WEI over their travel history and residence in the 3 to 10 days prior to illness onset, referred to as 'interpolated WEI (activity)'.A moving sum of the ultimate and penultimate trap collections (relative to a participant's illness onset) was used in both interpolation procedures in order to stabilize the highly variable individual trap collections.
The per-protocol reanalysis based on the interpolated WEI (residence) estimated the Wolbachia efficacy to reach 82.0% (95% CI: 72.3%, 87.9%) when comparing those whose residential WEI was above 0.8 to those whose residential WEI was below 0.2 (Figure 4). Figure 4 further shows that a monotonic doseresponse relationship with efficacy was observed as the interpolated WEI (residence) increased.Statistical significance was found at every contrasting WEI (residence) strata when WEI was at or above 0.4.A similar trend was observed with interpolated WEI (activity), where statistically significant effects were observed in every WEI stratum, with a maximum efficacy estimate of 82.7% (95%CI: 71.7%, 88.4%) for WEI greater than 0.8 and minimum efficacy estimate of 33.7% (95% CI: 21.7%, 46.5%) observed at WEI within [0.2, 0.4), compared to WEI less than 0.2 (Figure 4).Compared to the original per-protocol results based on clusteraggregate WEI, the reanalysis using interpolated values of individual WEI resulted in an increase in the estimated intervention efficacy and (for the WEI based on residence only) a more pronounced dose-response relationship.After accounting for human movement and intervention contamination across cluster boundaries, our reanalysis of data from a cluster randomised trial of Wolbachia-infected mosquito releases for the control of dengue showed an increased estimate of intervention efficacy compared to the primary analyses published previously.Participants' observed movement patterns were strongly age-dependent, highlighting the need for consideration of human mobility in both the design and -where possible -the analysis of trials for interventions delivered at a community level.Substantial spillover of Wolbachia mosquitoes into several untreated clusters during the 27-month period of clinical enrolment complicated the measurement of the intervention effect, but also suggests that small initial gaps in Wolbachia coverage under programmatic deployment could fill themselves in with time without further intervention.
In randomized controlled trials, misclassification of participants' exposure can occur if they do not receive -or are non-compliant with -the intervention they were randomly allocated to receive.This exposure misclassification produces an estimate of intervention efficacy that is biased towards the null when the comparison groups are, in reality, more similar in their exposure status than dictated by the randomized treatment allocation.In the specific case of cluster-randomized trials of area-level interventions without buffer areas between clusters, contamination by human movement and/or spillover of the intervention across cluster boundaries can lead to an underestimate of efficacy in the ITT analysis.In the AWED trial of Wolbachia-infected mosquito releases in Yogyakarta, Indonesia, most participants reported spending the majority of their time prior to illness onset under their intervention assignment, however half of participants spent at least 15% of their time either outside the trial area or in a location with a discordant intervention assignment.In addition, a marked increase in w Mel-infected mosquito spillover into untreated clusters was observed in the final year of the trial.These contamination risks were recognised at the design stage in the AWED trial, in which 12 of 24 contiguous clusters were randomised to receive releases of Wolbachia-infected mosquito releases for control of dengue [2], but could not be accounted for fully in the initial intentionto-treat or per-protocol analyses, which relied on cluster-level summary values of Wolbachia prevalence in Ae. aegypti mosquitoes.By deriving spatiotemporally interpolated values of Wolbachia prevalence at participants' homes and other visited locations, and combining these with detailed travel history data to calculate individual-level WEI in order to reduce biases in exposure measurement, we have estimated the intervention efficacy to be more than five percentage points higher than in the ITT analysis (82.7% [95%CI 71.7, 88.4] vs 77.1% [65.3, 84.9]), among participants with a Wolbachia exposure index >80% .
The results of our reanalysis suggest that exposure misclassification from both human mobility and spillover of w Mel mosquitoes contributed to the underestimate of Wolbachia intervention effect in the ITT analysis.In the complier-restricted analysis, the point estimate of intervention effect increased by three percentage points compared to the ITT analysis after removing only the biasing effect of human mobility by restricting the analysis to those participants who spent all of their time within the assigned intervention arm, while still considering cluster-level Wolbachia status as a binary (treated vs untreated).The further increase in efficacy measured in the per-protocol reanalysis -both relative to the ITT analysis and to the original per-protocol analysis based on cluster-level w Mel prevalence -indicates that within-cluster variability in w Mel prevalence was also contributing to the misclassification of individual participants' w Mel exposure status.The estimated intervention effect for participants in the highest stratum of w Mel exposure (80-100%), relative to the lowest stratum (0-20%) was very similar regardless of whether or not an individual's exposure was based on interpolated w Mel prevalence at their primary residence only, or weighted by time spent at other visited locations.This observation further supports the home as the primary location of dengue transmission risk, at least in Yogyakarta city.
Recent work [6] used geolocated residence, date of illness onset, and DENV serotype information to estimate small-scale spatiotemporal dependence among the dengue cases detected in the AWED trial and showed, for the first time, that the focal clustering of dengue cases is interrupted by the presence of w Melinfected mosquitoes.Those results support the proposition that the true w Mel intervention effect was larger than measured in the primary analysis of the AWED trial.
While this work aimed to reduce exposure misclassification by taking a finer spatiotemporal snapshot of an individual's WEI, lingering complications in estimation persist.Fogelson et al (forthcoming) find that estimates of intervention efficacy from mixed effects models, rather than the generalized lineal models applied here, can be biased when within-cluster exposure levels are highly homogeneous.When this estimator bias is properly handled, their work estimates a greater than 90% intervention effect from the WEI activity and residence individual-level exposures.However, the Fogelson et al analysis modelled WEI as a continuous variable such that the efficacy estimate is based on a contrast of 100% versus 0% WEI, whereas our analysis contrasts the highest (80-100%) versus the lowest (<20%) WEI stratum.
Theoretical and empirical studies of w Mel invasion into Ae.aegypti populations have predicted slow but steady spread at a rate of 100-200 metres per year in northern Australia [14,13].w Mel spread was predicted to be slower in tropical regions with higher Ae.aegypti densities, with variation in mosquito population density and geographical barriers contributing to variations in dispersal.This is consistent with the observed spread of w Mel across the boundaries of several untreated clusters by 27 months after the end of releases.
Following the successful demonstration in the AWED trial of the efficacy of w Mel-infected mosquito deployments in reducing dengue incidence, w Mel deployments into the untreated areas of Yogyakarta city were completed between October 2020 and January 2021.Between August 2021 and November 2022, w Mel deployments were expanded to the districts of Sleman and Bantul adjacent to Yogyakarta city, reaching a cumulative estimated population of 1.8 million people in a contiguous area of 540 km2.A recent secondary analysis of the AWED trial [8] has shown that a Wolbachia intervention effect equivalent to the AWED result was measurable from interrupted time series analysis of routine dengue case notification data.The area-wide coverage of Wolbachia in and around Yogyakarta city now provides an opportunity to evaluate the feasibility for w Mel to lead to sustained suppression, or even local elimination (Bannister-Tyrrell et al 2023, in preparation), of dengue in Yogyakarta.

Data source
The participant population leveraged for this work comes from the Applying Wolbachia to Eliminate Dengue (AWED) trial, described in detail elsewhere [2,15].Briefly, the AWED trial was a cluster randomized, parallel arm study with test-negative sampling carried out in Yogyakarta, Indonesia for a period of 27 consecutive months, beginning in January 2018 and ending during the emergence of the global coronavirus pandemic in March 2020.The city of Yogyakarta was divided into 24 contiguous clusters each approximately 1km 2 , twelve of which were randomized via constrained randomization to receive releases of w Mel-infected mosquitoes between March and December 2017.The remaining twelve clusters served as controls.Routine vector control efforts continued throughout the study.
Entomological data was also included for two regions contiguous to the AWED trial site.These regions had been a part of a quasi-experimental study where one region, comprised of seven kelurahans (urban villages) on the northwestern perimeter of Yogyakarta City, received Wolbachia-infected mosquito releases between August 2016 and March 2017 and the second region, comprised of three kelurahans on the southeastern perimeter, served as an untreated control area (Fig. 2A).The study and its results have been described elsewhere [9].

Participant data
Participants enrolled in the AWED trial provided demographic information (age and sex), a geolocated residential address, and a detailed travel history for the hours between 5 AM and 9PM over the three to ten days before illness onset.A total of 116,473 unique movement events were reported by the 6,306 individuals in the analysis sample.The majority of these reported locations fell within the RCT study area (92.9%).There were 8,277 movement events outside of the RCT area: 1,091 (<1%) and 533 (<1%) fell in the quasiexperimental intervention and untreated areas, respectively.The remaining 6,653 (5.7%) events fell outside of any study area.

Entomological monitoring
A network of 367 BG Sentinel traps (Biogents, Germany) collected adult mosquitoes throughout the AWED trial area between January 2018 and March 2020, such that median trap density was 16.4 BG/km 2 in the intervention clusters and 15.3 BG/km 2 in the untreated clusters [15].An additional 88 BG Sentinel traps monitored Wolbachia prevalence in the quasi-experimental regions outside of the AWED randomized trial area.The median trap density was 15.7 BG/km 2 in the quasi-experimental intervention area and 3.6 BG/km 2 in the quasi-experimental non-release area (Figure 2A) [9].All traps were monitored on staggered weekly schedules.The number of mosquitoes caught in each BG trap was recorded by species, sex, and in total.Ae. aegypti were stored at -20°C in 80% ethanol until testing for w Mel infection.[15,9] Descriptive analyses

Human mobility
Descriptive analyses of participant mobility were performed examining 1) the proportion of cumulative observation time at varying distances from home by age group (Fig. 1), 2) number of unique locations visited (Fig. S1), and 3) type of location (Fig. S2).
w Mel Spatiotemporal Heterogeneity In the previously reported per protocol analysis, w Mel spatiotemporal heterogeneity was incorporated into the efficacy estimate by taking the proportion of trapped Ae. aegypti with w Mel Wolbachia detected, in either the cluster of residence or as a time-weighted average across the clusters visited in the month of enrollment [2].Here, we leverage the extensive and frequently monitored BG trap network to obtain estimated w Mel proportions at much greater proximity to residential and visited locations.
Before interpolating the w Mel surface across the study area, the trap-level proportions of w Mel mosquitoes were stabilized by summing the observed mosquito counts and observed mosquito counts with w Mel detected for the current and penultimate trap event.Spatiotemporal inverse density weighting (IDW) was then used to construct interpolated w Mel surfaces (Supplemental Materials).Leave-one-out cross validation was used to perform a grid search to specify the optimal value of nuisance parameters C (a scalar denoting the relative importance of time to geographical distance), m j (number of nearest neighbors), and p (the inverse distance weighting factor) such that the set minimizes the root mean square prediction error (RMSPE).Further description of this approach can be found in the Supplemental Materials.

Complier-restricted analysis
The complier-restricted analysis restricted the analytic dataset to those who spent all reported time under the intervention assignment determined by their cluster of residence (n = 3,114).Individuals who moved outside of their cluster of residence, but remained within clusters assigned to the same intervention were included in the analysis.Individuals who reported spending time in the alternate intervention arm were excluded.Time spent outside of the AWED study area was handled two ways.First, individuals who left the AWED trial area were excluded.Second, the analysis was broadened to include individuals whose time inside and outside of the AWED study area fell under their assigned intervention arm.For individuals in the intervention arm, this means including those who only spent time in the AWED intervention area and the quasi-experiment intervention area.For those in the untreated arm, this includes individuals who spent time anywhere outside of the AWED study area except for in the quasi-experiment intervention area.
Intervention efficacy was estimated with the modified odds ratio approach used in the primary analysis from the NEJM manuscript [15], as described in the study protocol [2] and earlier methods paper [10].

Per-protocol reanalysis
We used the interpolated w Mel surface to estimate each individual's level of exposure.Just as in the initial secondary analysis of the AWED trial data, [15] we first estimated an individual's Wolbachia-exposure index (WEI) based on the spatiotemporally interpolated proportion of w Mel mosquitoes at the individual's residence on the reported day of illness onset.Second, we used a weighted sum of the individual's interpolated WEI over their travel history and residence in the 3 to 10 days prior to illness onset.
Intervention efficacy was estimated with the same approach as in the initial report, [15] using a generalized linear model and balanced bootstrap resampling approach based on cluster residence.

Spatiotemporal Inverse Density Weighting
The general formula for inverse density weighting (IDW) is given in Equation 1, where Ẑ(•) is the estimated proportion of mosquitoes with w Mel at a specified location s 0 and time t 0 .Let s i denote the ith observed location {s i : i = 1, 2, . . ., m T } which has two spatial components (x i , y i ), for instance, longitude and latitude.Let t j denote the time of the jth observation where observations were made at {t j : j = 1, 2, . . ., T }.

Z(s
The multiplier λ ij represents the inverse density weight (Eq.2), which is a function of distance in space-time (d ij,0 ) from the observed data Z(s i , t j ) and the space-time location of interest Ẑ(s 0 , t 0 ).
Distance in this setting was estimated using Equation 3. d((s i , t j ), (s 0 , t 0 )) = (x i − x 0 ) 2 + (y i − y 0 A grid search minimizing the cross-validation root mean square prediction error on a random sample of fourteen months worth of trap data determined the values for parameters m j , C, and p.
3 Changes in WEI

Figure 1 :
Figure 1: Mobility of AWED trial participants in Yogyakarta City during the 3-10 days prior to illness onset, based on self-reported travel (5AM -9PM).The graph shows the median and interquartile range of the time participants within each age group spent at home and cumulatively at increasing distances from home (children aged 3-5 years, 6-10 years, 11-18 years, and greater than 18 years).

Figure 2 :
Figure 2: (A) Map of Yogyakarta City with geolocated trap locations.Shape is used to distinguish intervention areas where w Mel releases occurred ('X') and untreated areas where w Mel releases did not occur ('+').Traps within the AWED RCT study boundaries are denoted in black, while those in the quasi-experimental areas are marked in gray.(B) The total number of mosquitoes screened (solid line) and the number of mosquitoes with w Mel detected (dashed line) per month by study area.

Figure 3 :
Figure 3: Spatiotemporal inverse density weighted maps of w Mel% interpolated across a 100x100 meter grid overlaid on the AWED trial surface on January 1 each year of (A) 2018, (B) 2019, and (C) 2020.

Figure 4 :
Figure 4: Forest plot comparing Wolbachia efficacy estimates from the primary analysis and reanalysis of the AWED trial in Yogyakarta, Indonesia.Included are the original ITT estimates [15] (light blue diamond), the complier-restricted analysis results performed here (dark blue square), the new per-protocol estimates based on spatio-temporally interpolated WEI (navy, closed circle), and the original per-protocol estimates based on the cluster-aggregate WEI (light blue, open circle).

Figure S1 :
Figure S1: Number of distinct locations AWED trial participants self-reported visiting between 5am and 9pm during the 3-10 days prior to illness onset, including their primary residence.A) Frequency distribution of number of visited locations among 6306 participants in the analysis dataset.B-D) Relative frequency distribution of number of visited locations among participants <15 vs ≥15 years (B), participants resident in intervention vs untreated clusters (C), and virologically-confirmed dengue cases vs test-negative controls (D).

Figure S2 :
Figure S2: The top locations sorted by the number of participants who reported visiting each location.

Figure S3 :
Figure S3: Changes in participant-level WEI values based on the method of estimation.A) The individuallevel WEI values estimated in the original per-protocol analysis based on cluster-level w Mel prevalence during the month of enrollment, estimated either as a weighted average of prevalences in the clusters visited ("activity", left axis) or based on cluster-aggregated wMel levels at the cluster of residence ("residence", right axis).B) The original per-protocol individual-level WEI (activity) values (left) are compared to the individual-level WEI (activity) values from the interpolated surface (right).C) The original per-protocol individual-level WEI (residence) values (left) are compared to the individual-level WEI (residence) values from the interpolated surface (right).C) A recreation of the plot in A, but using the interpolated WEI (activity) and interpolated WEI (residence) values.