key: cord-1000428-f9w53966
authors: Utarini, Adi; Indriani, Citra; Ahmad, Riris Andono; Tantowijoyo, Warsito; Arguni, Eggi; Ansari, M. Ridwan; Supriyati, Endah; Wardana, Dwi Satria; Metika, Yeti; Ernesia, Inggrid; Nurhayati, Indah; Prabowo, Equatori; Andari, Bekti; Green, Benjamin R.; Hodgson, Lauren; Cutcher, Zoe; Rancès, Edwige; Ryan, Peter A.; O’Neill, Scott L.; Dufault, Suzanne M.; Tanamas, Stephanie K.; Jewell, Nicholas P.; Anders, Katherine L.; Simmons, Cameron P.
title: Efficacy of Wolbachia-infected mosquito deployments for the control of dengue
date: 2021-06-10
journal: N Engl J Med
DOI: 10.1056/nejmoa2030243
sha: e88681d2ae29e3391e3da4462dd8f8637fb101b2
doc_id: 1000428
cord_uid: f9w53966

BACKGROUND: Aedes aegypti mosquitoes infected with Wolbachia pipientis (wMel strain) have reduced potential to transmit dengue viruses. METHODS: We conducted a cluster randomised trial of deployments of wMel-infected Ae. aegypti for control of dengue in Yogyakarta City, Indonesia. Twenty-four geographic clusters were randomly allocated to receive wMel deployments as an adjunct to local mosquito control measures; or to continue with local mosquito control measures only. A test-negative design was used to measure efficacy. Study participants were persons 3–45 years old attending primary care clinics with acute undifferentiated fever. Laboratory testing identified virologically-confirmed dengue cases and test-negative controls. The primary endpoint was efficacy of wMel in reducing the incidence of symptomatic, virologically-confirmed dengue, caused by any dengue virus serotype. RESULTS: Following successful introgression of wMel in intervention clusters, 8144 participants were enrolled; 3721 from wMel-treated clusters and 4423 from untreated clusters. In the ITT analysis virologically-confirmed dengue occurred in 67 of 2905 (2.3%) participants in the wMel-treated and 318 of 3401 (9.4%) in the untreated arm (OR 0.23, 95% CI, 0.15 to 0.35; P=0.004): protective efficacy of 77.1% (95% CI, 65.3 to 84.9). Protective efficacy was similar for the four serotypes. Hospitalisation for virologically-confirmed dengue was less frequent for participants resident in the wMel-treated (13/2905, 2.8%) compared to the untreated arm (102/3401, 6.3%): protective efficacy 86.2% (95% CI, 66.2 to 94.3) CONCLUSIONS: wMel introgression into Ae. aegypti populations was efficacious in reducing the incidence of symptomatic dengue, and also led to fewer dengue hospitalisations. TRIAL REGISTRATION NUMBER: ClinicalTrials.gov Identifier: NCT03055585 and INA-A7OB6TW

: Baseline characteristics of clusters (sociodemographic, historical disease incidence, area and mosquito trapping) 16 

Randomisation is a key step in the design of the AWED cluster randomised trial. Importantly, it will provide a sense of "fairness" and transparency in allocation of the intervention. It is also an opportunity to ensure "balance" between intervention and non-intervention arms.

Balance means that, overall, the intervention and non-intervention arms have similar dengue risk apart from the presence of Wolbachia. Balance aims to minimise the effect of confounding variables, so that any systematic difference in dengue incidence that occurs between Wolbachia intervention areas and untreated areas during the trial can clearly be attributed to the effects of Wolbachia. Ensuring a well-balanced trial will reduce the need for statistical adjustment during the analysis phase, increasing the "face validity" of the trial.

Balance also improves statistical power and efficiency, reducing the risk of Type II error (i.e. failure to detect an effect when the effect is, in fact, present). 1 Simple randomisation can be relied upon to produce an overall balance between intervention and non-intervention arms when there is a large number of clusters available for randomisation. However simple randomisation can't be relied on to produce balanced allocations where the sample size is small. For example, in our field efficacy trial involving just 24 clusters, it is possible simple random allocation may allocate most clusters with a high dengue risk to the same treatment arm through chance alone.

Covariate constrained (also referred to as "restricted") randomisation is the best way to achieve balance when the number of clusters is small. Table S1. 2. Generated a large number of potential random allocations (n=100,000)

3. For each allocation, calculated the value of each balancing criterion in each study arm 4. Rejected any random allocations where any one or more of the balancing criteria described in Table 1 were not met.

5. Note that for the potential confounding covariates the comparisons between study arm values and overall values were calculated in two ways, and both were applied as constraints as described above in 4:

a. Individual-level: the aggregate rate or proportion calculated across the study arm was compared with the aggregate rate or proportion calculated across the whole population b. Group-level: the mean of 12 cluster-level rates or proportions in the study arm was compared with the mean of all 24 cluster-level rates or proportions.

6. Assessed validity of the scheme:

a. Reviewed restricted number of potential allocations, ensuring the number was not too small relative to the overall number of possible allocations (as above). A minimum threshold of least 100 potential allocations was required.

b. Ensured that, within each stratum, no clusters are NEVER or ALWAYS allocated together, as this would result in an invalid randomisation scheme.

c. Examined dengue incidence correlation over time within clusters frequently randomised together, and compared against correlation within all pairwise combinations. Verified that dengue incidence in clusters frequently randomised together was not highly correlated 7. A total of 244 allocations met the balancing criteria (thus 488 possible distinct randomisations of intervention allocation). A random subset of 100 balanced allocations was selected, as the sampling frame for the final public randomisation event.

8. From this list a single allocation pattern was randomly selected, using numbered balls, at a public participatory event of community and government leaders in Yogyakarta in January 2017.

An existing colony of local Ae. aegypti containing the wMel Wolbachia strain, created for the 2016-2017 quasi-experimental study releases in Yogyakarta City, 5 was used as the founder colony for the releases described here. It was backcrossed for three generations with wildtype males collected from the study intervention area to generate the release colony, which was then maintained as described previously. 5 The insecticide resistance profile of the wMel-infected Ae. aegypti release material was matched to the local wild-type population as described previously. 5

Vector competence experiments using wMel-infected Ae. aegypti from Yogyakarta

We membrane blood-fed Yogyakarta wild-type (WT) and wMel-infected Ae. aegypti at the Oxford University Clinical Research Unit (Vietnam) using viremic blood collected from nine acute dengue inpatients at the Hospital for Tropical Diseases, Ho Chi Minh City, Vietnam using previously published protocols. 6 We measured virological differences between each mosquito line, in abdomen and saliva, at predefined times (10-21 days) after blood feeding. infections, the DENV viral load in wMel-infected mosquitoes was significantly lower than in wild-type Ae. aegypti for each of the three serotypes tested ( Figure S2 ).

As expected, wMel-infected mosquitoes were significantly less likely to have saliva containing infectious virus than their wild-type counterparts: 8% wMel-infected mosquitoes with saliva infection vs 38% wild-type [crude OR=0.14, 95% CI 0.06, 0.34; adjusted for serotype, viremia and day mosquitoes were harvested: OR = 0.07, 95% CI 0.02, 0.33] ( Figure   S1 ).

These results are consistent with a large body of work, on a variety of different genetic backgrounds, demonstrating that wMel-infected Ae. aegypti have reduced transmission potential for DENV.

Wolbachia-carrying mosquitoes were released as eggs using mosquito release containers 

Prevalence of Wolbachia in the local Ae. aegypti population was monitored by weekly collection of adult mosquitoes via a network of 348 BG Sentinel traps (Biogents, Germany). 

There are no published formulae to estimate sample size for the proposed study design, ie. hypothesis is that the average proportion of total enrolled participants that are cases is the same in treated and untreated study arms. The alternative hypothesis is that the proportion of enrolled participants that are cases is lower in the Wolbachia treated arm than the untreated arm.

Simulations were used to estimate the power to detect a range of intervention effect sizes using the two methods above, assuming 12 clusters per arm, a fixed total of 1000 true dengue cases enrolled and 4000 non-dengue controls. Empirical data on population, historical dengue incidence and incidence of other febrile illness in the 24 study clusters were used to define the baseline characteristics for the simulated scenarios. Nine 3). For each of these five 'true' effect sizes, applied to each of the 244 balanced allocations, the 'observed' effect size was calculated from the simulated data by the two methods outlined above; i) aggregated odds ratio for residence in a treated cluster among cases versus controls, and ii) t-test for comparison of the average cluster summary proportions (cases/cases+controls) between study arms. Statistical inference, from the t-test directly, or, for the odds ratios using permutation distribution approximations with standard errors adjusted to account appropriately for the clustered nature of the data, respectively, was used to calculate the proportion of constrained random allocations that yielded a significant result. This provided an estimate of Type I error at the null, and power away from the null (Table) . Both of these approaches thus are using approximations to the exact permutation distribution. 4 In practice, the appropriate reference distribution for inference will be based on the set of 244 potential balanced allocations.

These simulations estimated that approximately 1000 cases plus four times as many controls will be sufficient to detect a 50% reduction in dengue incidence with 80% power. The results show that constrained randomization is somewhat conservative at the null but generally increases power moderately. The odds ratio test is more powerful than the t-test approach, and will thus be used as the primary analysis with the additional attraction of being standard for the traditional test-negative design.

A re-estimation of sample size requirements was conducted in January 2019 after one year of recruitment. The initial power calculation used 1000 dengue cases and 4000 non-dengue controls allocated to each cluster based on historical proportions of dengue cases and other febrile illnesses, assuming no variation in the proportion of cases by cluster. This method was found to overestimate power for small samples by not taking into account randomness in the sampling. The sample size re-estimation included power estimates for 200, 400, 600, 800 and 1000 dengue cases with 4 times as many controls. Cases and controls were allocated among clusters by sampling from multinomial distributions, which incorporated added randomness by allowing the proportion allocated to each cluster to vary across simulations. The re-estimation found that 400 dengue cases plus four times as many controls would be sufficient to detect a 50% reduction in dengue incidence with 80% power.

Additional simulations were conducted in September 2019 to assess the potential impact on power if a number of untreated clusters were 'lost' to Wolbachia contamination. For the target minimum observed effect size of 50% (RR=0.5) and 400 enrolled dengue cases, contamination of 3 untreated clusters (assuming that contaminated clusters experience the full intervention effect for 1 out of the 3 years of trial recruitment) is expected to result in a ~7% loss of power, and contamination of 6 clusters to result in a ~14% loss of power.

Participants were asked about their mobility during the period 3 -10 days prior to illness onset using a structured interview administered at enrolment. This records the duration of Wolbachia prevalence of zero. The process of calculating WEI was conducted blinded to participants' case/control status, by partitioning the travel history data from the laboratory diagnostic data, to remove any possibility of observer bias.

An additional per-protocol analysis calculated WEI using only the cluster-level Wolbachia prevalence in the participant's cluster of residence (in the calendar month of participant enrolment), ignoring the participant's recent travel history. This recognises that dengue exposure risk may be higher at home versus other locations, rather than assuming an even distribution of exposure risk across daytime hours and locations visited.

Cases and controls were classified by strata of their WEI: 0-<0.2; 0.2-<0.4; 0.4-<0.6; 0.6-<0.8;

and 0.8-1. This acknowledges that the WEI is not a highly precise measure, and serves to reduce error in exposure classification. The ITT methods described were extended to allow for this individual level covariate using a generalized linear model with the cluster bootstrap based on 1,000 clustered resamples. Balanced bootstrap resampling based on cluster membership accommodates within cluster dependencies and has been shown to be a competitive alternative to GEE and mixed effects approaches in the analysis of hierarchical data. 13 Such models yield an estimate, and associated bootstrap percentile-based confidence interval, for the relative risk. Efficacy was then calculated as 100*(1-RR). The WEI strata was included as an unordered covariate to calculate stratum-specific IRRs (relative to the baseline 0-<0.2 stratum). The model results suggest a threshold effect (rather than a "linear" dose-response). For this reason, the WEI strata was not treated as an ordinal covariate, though this had been discussed in the SAP.

There exist two other modifications from the approach described in the SAP. First, the presented model results do not include adjustment for time. Second, the original analysis plan proposed the use of generalized linear mixed effects models with a random effect for cluster membership to account for hierarchical dependencies. However, the resulting point efficacy estimates were not robust to the proposed cluster random effect. As such, the generalized linear model framework is maintained, but with a clustered bootstrap approach to estimation and inference, thereby avoiding the bias and statistical inefficiency introduced from misspecification of the random effects distribution. Figure S1 : Map of study location with satellite overlay.

The map of Yogyakarta City (plus a small region of neighbouring Bantul District) is shown with wMel intervention clusters (shaded blue) and untreated clusters (no shading)

indicated. The locations of primary care clinics (red crosses) where enrolment occurred are also shown, and are numbered to correspond with the clinical-level data in Figure S5 . and Wolbachia-treated (light blue; orange) clusters. The proportion of participants who were hospitalised varied between clinics, and was higher among participants from untreated (yellow) than Wolbachia-treated (orange) clusters.

Efficacy is calculated as 100*(1-aggregate odds ratio) among participants enrolled within the first 12 months after wMel establishment, within the first 24 months, and within the full 27 month trial period. Markers show stratum-specific efficacy (and 95% confidence intervals) against VCD by quintile of Wolbachia exposure index, with WEI based on A) duration-weighted wMel frequencies in the cluster of residence and other visited locations, or B) wMel frequency in cluster of residence only. The number of VCD cases and test-negative controls with WEI values in each quintile is shown beneath the plots.

Allocation techniques for balance at baseline in cluster randomized trials: a methodological review

Cluster randomized trials

Field trials of health interventions: a toolbox

Analysis of clusterrandomized test-negative designs: cluster-level methods

Reduced dengue incidence following deployments of Wolbachia-infected Aedes aegypti in Yogyakarta, Indonesia: a quasiexperimental trial using controlled interrupted time series analysis

Field-and clinically derived estimates of Wolbachia-mediated blocking of dengue virus transmission potential in Aedes aegypti mosquitoes

Assessing quality of life-shortening Wolbachiainfected Aedes aegypti mosquitoes in the field based on capture rates and morphometric assessments

Development of a novel ozone-and photo-stable HyPer5 red fluorescent dye for array CGH and microarray gene expression analysis with consistent performance irrespective of environmental conditions

Quantitative real-time PCR detection of Zika virus and evaluation with field-caught mosquitoes

Development of a TaqMan RT-PCR assay without RNA extraction step for the detection and quantification of African Chikungunya viruses

The relative importance of innate immune priming in Wolbachia-mediated dengue interference

Validation of an internally controlled one-step real-time multiplex RT-PCR assay for the detection and quantitation of dengue virus RNA in plasma

ClusterBootstrap: An R package for the analysis of hierarchical data using generalized linear models with the cluster bootstrap