Design of trials for interrupting the transmission of endemic pathogens

Abstract

Background: Many interventions against infectious diseases have geographically diffuse effects. This leads to contamination between arms in cluster-randomized trials (CRTs). Pathogen elimination is the goal of many intervention programs against infectious agents, but contamination means that standard CRT designs and analyses do not provide inferences about the potential of interventions to interrupt pathogen transmission at maximum scale-up.

Methods: A generic model of disease transmission was used to simulate infections in stepped wedge cluster-randomized trials (SWCRTs) of a transmission-reducing intervention, where the intervention has spatially diffuse effects. Simulations of such trials were then used to examine the potential of such designs for providing generalizable causal inferences about the impact of such interventions, including measurements of the contamination effects. The simulations were applied to the geography of Rusinga Island, Lake Victoria, Kenya, the site of the SolarMal trial on the use of odor-baited mosquito traps to eliminate Plasmodium falciparum malaria. These were used to compare variants in the proposed SWCRT designs for the SolarMal trial.

Results: Measures of contamination effects were found that could be assessed in the simulated trials. Inspired by analyses of trials of insecticide-treated nets against malaria when applied to the geography of the SolarMal trial, these measures were found to be robust to different variants of SWCRT design. Analyses of the likely extent of contamination effects supported the choice of cluster size for the trial.

Conclusions: The SWCRT is an appropriate design for trials that assess the feasibility of local elimination of a pathogen. The effects of incomplete coverage can be estimated by analyzing the extent of contamination between arms in such trials, and the estimates also support inferences about causality. The SolarMal example illustrates how generic transmission models incorporating spatial smoothing can be used to simulate such trials for a power calculation and optimization of cluster size and randomization strategies. The approach is applicable to a range of infectious diseases transmitted via environmental reservoirs or via arthropod vectors.

Keywords: Cluster randomization; Elimination; Stepped wedge design; Transmission model; Vector control.

Fig. 1

Schematics of three SWCRT design sequences on a grid of 81 equal-area clusters. Clusters are numbered in the order of the design rollout sequence. In the diagram as shaded, clusters 1–20 have received the intervention and clusters 21–81 have not yet received the intervention. All sequences begin at one randomly selected cluster. a Hierarchical SWCRT sequence: the sequence begins at one randomly selected meta-cluster. Clusters within that meta-cluster are filled in a random order until the meta-cluster is complete, then the next meta-cluster is selected. b Oil-drop SWCRT sequence: the sequence begins at one randomly selected cluster and spreads across adjacent clusters until the grid is filled. c Random SWCRT sequence: clusters are selected at random until the grid is completely filled. For all SWCRT designs, by the end of the intervention rollout there will be an equal number of cluster-days with and without the intervention. However, at almost all time points during the rollout, the populations in the two study arms will be unequal. SWCRT stepped wedge cluster randomized trial

Fig. 2

Example single random SWCRT sequence runs of the transmission simulator. Incidence of clinical events in intervened (red) and non-intervened (blue) populations as modeled by the transmission simulator for three levels of community radius, a 0.5 km, b 1.0 km, and c 1.5 km. The cluster width is held constant at 1 km, corresponding to an area of 1 km². The transmission model input efficacy is 80 %. During the first 40 time steps of each simulation, the incidence of clinical events is an auto-regressive moving average process that oscillates around the initial incidence value of 20 %. The intervention commences at time step 41 and from time steps 41 to 121, the incidence of the pathogen decreases sharply in both arms due to the direct effect of the intervention and the community effect. The community effect has more impact at greater radii. SWCRT stepped wedge cluster randomized trial

Fig. 3

Relationships for the six effectiveness measures from Table 1 during a single random SWCRT sequence run of the transmission simulator. ${ê}_1\left(t\right)$ (filled green circles) is a direct comparison between outcomes in the intervened group versus the status at baseline, ${ê}_2\left(t\right)$ (filled pink triangles) is a direct comparison between outcomes in the non-intervened group versus the status at baseline, and ${ê}_3\left(t\right)$ (filled blue squares) is an overall comparison of the entire study area versus baseline. ${ê}_4\left(t\right)$ (bright green squares) is a direct comparison between the intervened and all non-intervened, ${ê}_5\left(t\right)$ (gold circles) is a direct comparison between the intervened and those remote from the intervention, and ${ê}_6\left(t\right)$ (dark green triangles) is a direct comparison between non-intervened populations close to and remote from the intervention. SWCRT stepped wedge cluster randomized trial

Fig. 4

The three contemporaneous effectiveness measures over time. ε ₄ (bright green squares), ε ₅ (gold circles), and ε ₆ (dark green triangles). The horizontal lines correspond to the simulated efficacy E _s=30 %

Fig. 5

Power over time of contemporaneous effectiveness measure ε ₅(t) to detect a difference between the intervened treatment arm and non-intervened arm remote from the intervention. All simulations are based on a cluster diameter of 1 km and one-sided type I error rate α=5 %. Open circle: hierarchical ordering; plus sign: oil-drop ordering; open triangle: random ordering. a Intervention efficacy of 30 %; b intervention efficacy of 80 %

Fig. 6

Power over time of contemporaneous effectiveness measure ε ₆(t) to detect a difference between the naive individuals close to the intervention and those remote from the intervention. All simulations are based on a cluster diameter of 1 km and one-sided type I error rate α = 5 %. Open circle: hierarchical ordering; plus sign: oil-drop ordering; open triangle: random ordering. a Intervention efficacy of 30 %; b intervention efficacy of 80 %

Fig. 7

Sequence selection. Three hierarchical sequences applied across Rusinga Island for two levels of intervention efficacy, 30 % and 80 %. Results are color coded by the meta-cluster membership. Reading from the left, the meta-cluster sequences are [V, II, VII, III, IV, VIII, I, VI, IX], [V, IV, IX, III, VI, I, VIII, V, II], and [IV, VII, II, IX, III, VIII, I, V, VI]. Coverage intervals widen upon introduction of the intervention to meta-cluster VIII, located at the base of the peninsula in the north-east of the island, in the right two sequences. Coverage intervals were off the scale for the last meta-cluster of sequence 298. Of the three cluster sequences presented, only sequence 296 met the criteria for entry into the pool for the SolarMal randomization sequence selection lottery (coverage intervals of the primary effectiveness measure, ε ₅, less than 10 % from time points 60 to 100, and no single village rollout greater than 6 months’ duration)