Efficacy of Wolbachia-mediated sterility for control of dengue: emulation of a cluster randomized target trial

Specification of the cluster-randomized test-negative target trial

We specified a cluster-randomized test-negative target trial¹⁰ to evaluate the efficacy of releasing wAlbB-infected Ae. aegypti male mosquitoes for dengue control via vector population suppression, from epidemiological week (EW) 1 2019 – EW 26 2022 in Singapore. The target trial was emulated using test-positive/negative databases, which comprised all patients who report to any general practitioner clinic, polyclinic or public/private hospital and were suspect of dengue illness during the trial duration in Singapore.

Emulating randomisation protocols from cluster-randomised trials for Wolbachia interventions

26 townships in Singapore which were not subject to Wolbachia interventions were considered potential locations as control sites. Towns were demarcated planning areas used by government ministries and departments for administrative purposes. While four long-term Wolbachia field trial townships were not randomly pre-selected, we emulated constrained randomisation protocols for cluster-randomised trials by randomly selecting a set of 12 control townships, such that the historical dengue incidence of the intervention arm is similar to that of the control arm in the pre-intervention period^4,11.This further prevents chance-imbalance in baseline dengue risk due to the small number of intervention (n=4) locations considered (See Supplementary Information). All locations practiced the same baseline dengue control protocol in the pre- and post-intervention periods^2,3.

Cohort

Under the Infectious Diseases Act, all laboratory-confirmed cases of dengue are legally mandated for reporting in the national dengue surveillance system. Approval from the Director General of Health, Ministry of Health, was obtained to collect and use the data of dengue-suspected patients, whose blood samples are sent for dengue tests, through a national network of diagnostic laboratories that support private clinics, public polyclinics, or public/private hospitals.

This project was exempted from formal bioethics review as it is not considered human biological research, as advised by the Ministry of Health, Singapore. All laboratory tests were performed for clinically directed reasons, and the data from these tests is routinely collected as part of routine dengue surveillance under the Infectious Disease Act, which exempts the need for informed consent.

In Singapore, 133,821 individuals reported for febrile illness and were tested for dengue at the Environmental Health Institute, hospital laboratories and commercial diagnostic laboratories, through general practitioner clinic, polyclinic or public/private hospital from EW1 2019 – EW 26 2022. All dengue-suspect patients were tested using either using an internally controlled RT-qPCR assay, dengue non-structural protein 1 (NS1) or IgM as diagnostic assays to detect dengue virus in serum samples^3,11. We excluded individuals who were tested on more than one occasion in 4 weeks, individuals who had more than one residential address in different control or intervention townships and individuals who had been tested at different labs with conflicting dengue results. We also excluded individuals who had residential addresses at intervention sites at the time of the test but had not been exposed to Wolbachia interventions for at least 3 months, based on exposure criteria described below.

Four intervention townships (Bukit Batok, Choa Chu Kang, Tampines, Yishun) had 7,049 tested individuals included in the study period. After constrained randomisation, we selected 12 control townships (Bedok, Bishan, Clementi, Geylang, Jurong West, Kallang, Pasir Ris, Punggol, Queenstown, Sengkang, Toa Payoh, Woodlands) with 69,216 tested individuals in the study period. The control arm had an average dengue incidence rate normalized by population size which was less than 5% different from the intervention arm in the pre-intervention period of EW1 2010 to EW52 2016, indicating good balance in historical dengue risk between arms.

Outcomes of interest

Analysis considered Wolbachia exposure as a binary classification based on the home address of the individual in an intervention sector within an intervention township (Wolbachia-exposed) or a control sector within the selected control townships (Wolbachia-unexposed). Sectors comprise 10 or more public housing apartment blocks and measured around 0.088 km² on average and are used for planning of surveillance and control for environmental infectious diseases in Singapore.

We subcategorised Wolbachia exposure based on whether an individual resides in a sector which experienced sustained Wolbachia releases for 3, 6 or 12 or more months, due to the time required for releases to induce noticeable vector suppression (See supplementary information). Home address was defined as the primary place of residence reported on the diagnostic test date. The intervention effect was estimated from an odds ratio comparing the exposure odds (residence in an intervention location for 3, 6 or 12 or more months) among participants who were dengue test-positive versus test-negative controls, with the use of logistic regression (See statistical analysis below). The null hypothesis was that the odds of residence in an intervention sector would be the same among participants who test-positive as that among test-negative controls. Intervention efficacy was calculated as 100×(1−odds ratio).

Characterisation of intervention

Male Wolbachia-infected Ae. aegypti were released twice weekly (weekdays, 0630–1030 hrs) at four townships in high-rise public housing estates covering 607,872 individuals as of Epidemiological Week (EW) 26 2022. Bukit Batok, Choa Chu Kang and Yishun towns were subjected to interventions which combined IIT with SIT. Tampines town used the high-fidelity sex-sorting methodology and also progressively adopted SIT protocols to release irradiated mosquitoes from January 2020⁹. To trial whether Aedes aegypti population suppression could be sustained over increasingly larger areas, an expanding release strategy was adopted in two large towns (Yishun, Tampines), where release sites were gradually expanded to adjacent neighbourhoods. In Bukit Batok and Choa Chu Kang towns, a targeted release approach was adopted, which focused releases on areas with high Aedes aegypti abundance and persistent dengue transmission. (Table 1, See supplementary information for full details). Adult Aedes aegypti populations in release and control sites were monitored using Gravitraps, with an average of six Gravitraps deployed per apartment block¹².

Covariates

We extracted a comprehensive set of spatially explicit variables to characterize environmental heterogeneity across sectors. Covariates considered prior to variable selection include, (1) vegetation maps with areas classified across multiple vegetation types including forest and managed vegetation to signify availability of natural breeding sites and nectar availability for male mosquitoes. (2) The averaged Normalized Difference Vegetation Index per sector, as an alternative measure of vegetation (3) To represent host density and urban breeding habitat availability, data on the locations of all public housing estates where over 75% of Singapore’s resident population reside were obtained. Utilising residential location and resale data, the average age of public housing residences was collected as older age is a well-established risk factor for higher mosquito abundance¹⁴.(4) Average residence price over the study duration as a proxy for household income and socioeconomic status. (5) Building height was calculated according to the number of floors with an average height of 3m. (6) The number of condominiums/landed properties was collected within each sector representing additional hosts being available. The percentage cover of built area was calculated as a sum of all residential, commercial and industrial buildings, representing the level of urbanicity, which has been associated with Aedes aegypti presence¹⁴. (7) The major open drainage network for Singapore was obtained from the Public Utilities Board and has been previously shown as a key breeding site for mosquitoes around public housing apartments¹⁵. The average distance of each block within a sector to a drain was measured as well as the length of the network within the sector. (8) Well-established meteorological variables which are known to affect mosquito survival or fecundity were collected. These included daily mean, maximum, and minimum temperature, total rainfall, and wind speed, which were obtained from 21 local weather stations. Hourly dewpoint and ambient ground air temperature were also taken from remote sensing measurements to estimate relative humidity over the time period using standard formula. These values were aggregated at a weekly level to correspond with the temporal frequency of dengue test-positive/negative data. Data sources and processing procedures were explicitly detailed in the Supplementary Information.

Statistical analysis

Baseline characteristics of the cohort were presented as mean and standard deviation or as frequency and percentage. Standardized mean differences (SMD) were used to evaluate balance between intervention and non-intervention arms, with SMD <0.1 indicating good balance.

To estimate the effect of Wolbachia interventions on the risk of dengue, we employed a doubly-robust logistic regression framework. First, to estimate propensity score models, we used the considered set of covariates described above as the independent variables and Wolbachia exposure for at least 3,6 and 12 or more months as separate outcomes of interest. Prespecified variables included age and gender, with other factors selected using high-dimensional regression and additional trimming of highly multicollinear covariates (See supplementary information). Sensitivity analyses later also indicated that variable selection procedures did not influence Wolbachia protective effect estimates.

We adjusted for differences in baseline characteristics and the propensity to be selected as a treatment site between intervention/non-intervention arms through inverse probability weighting, incorporating the selected covariates. A propensity score of belonging to the intervention arm was computed using a logistic regression that used the abovementioned covariates as explanatory terms. Inverse probability weights were computed as 1/propensity score for tested individuals who were Wolbachia-exposed; and 1/(1–propensity score) for tested individuals who were not Wolbachia-exposed. SMDs were used to assess covariate balance after inverse probability weighting. Thereafter, odds ratios (ORs) of being dengue test-positive between the intervention and control groups were estimated using a logistic regression model, with inverse probability weights applied. A doubly robust approach was employed for this model, where covariates used to construct inverse probability weights were included in each model specification as explanatory variables. This approach was used to prevent model misspecification in the generation of inverse probability weights or ORs in analyses.

To account for within-sector dependencies, we relied on cluster bootstrap based on 1,000 clustered resamples. Balanced bootstrap resampling based on sector membership can account for within-sector dependencies and has been used as a competitive approach to analyse hierarchical data¹⁶. The associated bootstrap percentile-based confidence interval was used to construct the 95% confidence interval for odds ratios, and findings were considered to be statistically significant when the 95% confidence intervals for ORs did not cross 1.

Subgroup analysis

We repeated all analysis by subsetting to intervention townships and specific years (2019, 2020, 2021, 2022) to examine any potential differences in intervention effect by location, and between epidemic and inter-epidemic years. We also repeat analysis by age (<20, 20-65, 65+) and sex (male, female) subgroups as dengue risk may be mediated by immunity levels in each age group or gender. Here, we conducted subgroup analyses by re-estimating odds ratios using the aforementioned statistical procedures, but only using individuals within that specific subgroup.

Robustness checks

We conducted a battery of sensitivity analyses to ensure the robustness of our model estimates. We (1) repeated all analysis without adjustment for covariates in the main logistic regression step (2) re-randomised our allocation of controls 1000 times and repeated our analysis by using the newly allocated controls arm, and compared our primary estimate of intervention efficacy against the empirical distribution of re-randomised intervention efficacies (3) used the full set of covariates, instead of the pre-selected covariates in our main analysis, to recompute odds ratios and intervention efficacies (4) we conducted in-space placebo checks on control sites, taking each allocated control site as the allocated placebo-intervention site and re-estimated odds ratios and intervention efficacies by comparing test-negative and positive individuals in the placebo-intervention versus other control sites (5) we conducted placebo checks on intervention sites, taking each intervention site in 52 and 104 weeks before the actual intervention as the intervention period, and recomputed intervention efficacies among allocated controls and interventions based on this placebo-intervention period. (6) We recomputed intervention efficacies using cluster bootstrap on the cluster rather than sector resolution.

Suppression of Aedes aegypti populations in field trial sites

Suppression of adult wild-type Ae. aegypti populations was demonstrated across the four field trial sites, with the Gravitrap Aedes aegypti Index (GAI) reduced as Wolbachia coverage increased across each township. When >50% coverage was achieved by EW 1 2022, the town-level GAI plunged below 0.05 for all sites (Figure 1, Supplementary Information).

• Download figure
• Open in new tab

Figure 1:

Weekly average, town level Gravitrap Aedes aegypti index (GAI) and Wolbachia coverage (%) from 2018 to 2022 in the intervention sites of (a) Bukit Batok, (b) Choa Chu Kang, (c) Tampines and (d) Yishun. The average GAI from the 12 controls is plotted for comparison. GAI is defined as the mean number of female adult Ae. Aegypti caught per functional Gravitrap per week, hence proxies for adult Ae. aegypti abundance in public housing areas in release area of each town. The geographical coverage (%) represents the percentage of areas within the town which is covered by Wolbachia interventions for at least six months and is calculated at the end of each year. Points represent coverage of Wolbachia interventions by the end of each year. The six-month mark for coverage is based on the time it takes Wolbachia release to have around 80% suppressive efficacy on Ae. aegypti abundance. The corresponding ratios of pre- and post GAIs at the sector is also plotted in (e), (f), (g) and (h), per year, on a per sector basis. The threshold of 1 indicates no difference between pre and post-GAIs for a specific sector in that specific year versus the pre-intervention period, and values above 1 indicate lower GAIs in the post-intervention period

Study Characteristics

Among 133,821 individuals who reported for febrile illness from EW1 2019 – EW26 2022, in the intervention and control arms, 76,265 (56.99%) were included in the study. Baseline demographic and spatial-temporal characteristics between Wolbachia-exposed and unexposed groups before and after inverse probability weighting were presented in Table 2. Characteristics were well-matched after inverse probability weighting (Table 2) with small differences in baseline characteristics between both groups.

Efficacy of Wolbachia-releases in reducing risk of dengue

Among dengue-tested individuals residing in areas which were Wolbachia-exposed for more than 3 months, the percentage of individuals who tested positive for dengue (13.6%, 956 of 7,049 individuals) was lower compared to the Wolbachia-unexposed (21.7%, 14,986 of 69,216 individuals) group.

In primary analysis, Wolbachia exposure for 3, 6, 12 or more months was associated to a lower risk of being test positive for dengue. Higher periods of exposure associated to greater levels of protective efficacy – at 47% (Odds ratio (OR): 0.53 [0.45-0.62]), 47% (OR: 0.53 [0.50-0.65]) and 59% (OR: 0.41 [0.39-0.50]) for 3, 6 and 12 of more months of Wolbachia exposure respectively. Protective efficacies were higher in epidemic years (2019, 2020, 2022) versus inter-epidemic years (Table 3).

By stratifying analysis across townships, years, age and sex subgroups and taking Wolbachia exposure at 12 or more months as the reference, the highest level of protective efficacies were found in Yishun township, at 77% reduction (OR: 0.23 [0.22-0.33]), and lowest in Choa Chu Kang township, at 36% reduction (OR: 0.64 [0.58-0.96]) when compared across townships. The highest level of protective efficacy was in the adolescent (62% protective efficacy, OR: 0.38 [0.31-0.61]) and adult age groups (60% protective efficacy, OR: 0.40 (0.36-0.49)) when compared across age groups and in females (71% protective efficacy, OR: 0.29 [0.27-0.37]) when compared against both sexes. Protective efficacies were found across all subgroups, with higher durations of exposure similarly associated to greater risk reductions of dengue – supporting consistent biologic replication of the intervention (Table 3).

Sensitivity analyses

We conducted a battery of sensitivity analyses to ensure the robustness of our model estimates. Repeating all analysis without adjustment for covariates, or inclusion of all available covariates in main logistic regression step did not change intervention efficacy estimates. Re-randomising our allocation of controls into the control arm 1000 times and repeating our analysis by using newly allocated controls arms did not change intervention efficacy estimates versus our primary estimate of intervention efficacy. We further did in-space placebo checks on control sites, taking each allocated control site as the allocated placebo-intervention site and re-estimated protective efficacies by comparing test-negative and positive individuals in the placebo-intervention versus other control sites. There were few significant/positive intervention efficacies demonstrated in the placebo-control sites with the grand mean of OR estimates centered closely to 1. Lastly, we conduct in-time placebo checks on intervention sites, where we took placebo-interventions in the actual intervention site before the intervention began and demonstrated that there were no significant/positive protective effects in the intervention sites in the placebo-intervention period. Re-running our cluster bootstrap inference on the town, rather than sector level, also demonstrated that we could recover the intervention efficacy point estimates, albeit with wider confidence intervals due to the coarser spatial resolution used for bootstrapping. Explicit results on all sensitivity analyses are provided in the SI.