Potential herd protection against Plasmodium falciparum infections conferred by mass antimalarial drug administrations.

The global malaria burden has decreased over the last decade and many nations are attempting elimination. Asymptomatic malaria infections are not normally diagnosed or treated, posing a major hurdle for elimination efforts. One solution to this problem is mass drug administration (MDA), with success depending on adequate population participation. Here, we present a detailed spatial and temporal analysis of malaria episodes and asymptomatic infections in four villages undergoing MDA in Myanmar. In this study, individuals from neighborhoods with low MDA adherence had 2.85 times the odds of having a malaria episode post-MDA in comparison to those from high adherence neighborhoods, regardless of individual participation, suggesting a herd effect. High mosquito biting rates, living in a house with someone else with malaria, or having an asymptomatic malaria infection were also predictors of clinical episodes. Spatial clustering of non-adherence to MDA, even in villages with high overall participation, may frustrate elimination efforts.

Introduction

Mass drug administration (MDA) is the provision of medications to entire target populations and the approach has been used for many infectious diseases, including lymphatic filariasis, soil-transmitted helminths, onchocerciasis, schistosomiasis, and trachoma (Keenan et al., 2013). MDA has historically been used for P. falciparum malaria (Poirot et al., 2013) and has recently been trialed in several locations in Africa (Gitaka et al., 2017; Mwesigwa et al., 2018; Shekalaghe et al., 2011) and Asia (Manning et al., 2018; Tripura et al., 2018; Nguyen et al., 2018; Pongvongsa et al., 2018; Landier et al., 2017a). It is being considered by several nations as a tool (to be used in unison with other interventions) for elimination (World Health Organization, 2017; Zuber and Takala-Harrison, 2018), and has already been implemented as an operational strategy in at least one modern setting (Parker et al., 2017; Landier et al., 2018a).

Given that drug pressure (through provision of antimalarial drugs) provides a survival advantage for resistant parasites, there has been some hesitance in using MDA for malaria. One historical malaria eradication campaign relied on the inclusion of sub-therapeutic levels of antimalarials distributed in table salt across large populations (Pinotti et al., 1955). This program likely led to the emergence of parasite resistance in the same regions (Wootton et al., 2002) and this has in part led to hesitance among some institutions (i.e. the World Health Organization and ministries of health) to implement MDA for malaria (World Health Organization, 2015a). MDA should ideally be used in settings with strong public health infrastructure, including easy access to diagnosis and treatment; an up-to-date and responsive surveillance system; and effective community engagement. Used appropriately, MDA can quickly reduce or eliminate parasite reservoirs and can act as a catalyst for subregional elimination of P. falciparum malaria (Landier et al., 2018a).

While antimalarials are usually administered following diagnosis (confirmed or presumed) or used as a prophylactic, MDA is used because of an intended population- or community-level effect. The rationale is that the transmission potential or reproductive rate of malaria is so high that a sufficient amount of the parasite reservoir needs to be removed in order to disrupt transmission. This group-level effect is also referred to as a ‘herd effect’ (John and Samuel, 2000; Pollard et al., 2015) and the concept applies to most communicable diseases. If a sufficient amount of the population participates in MDA, transmission chains cannot be sustained and transmission will cease, ultimately leading to a reduction in malaria morbidity and mortality (World Health Organization, 2017). There is likely to be a context-specific critical threshold for MDA coverage, below which the reduction of the parasite reservoir is not sufficient to halt ongoing transmission. Some literature has suggested that at least 80% coverage and adherence of MDA in the targeted population is necessary in order for the MDA to be successful (World Health Organization, 2015b). If the aim for antimalarial MDA is to interrupt transmission, the notion of a herd effect providing additional levels of population protection is plausible but has not been examined empirically (Cotter et al., 2013).

Drawing from detailed micro-epidemiological and spatial data from an MDA trial in Kayin State, Myanmar (Figure 1), we describe geographic and epidemiological patterns of clinical and subclinical P. falciparum malaria in villages undergoing MDA. We investigate associations between individual- and group-level participation in MDA (potential direct and indirect effects, respectively); subclinical infections; and clinical episodes of P. falciparum post-MDA. Such empirical research is important for providing an evidence base for further research, intervention, and policy work.

Figure 1.

Map indicating the locations of the study villages along the Myanmar-Thailand border; and the distribution of houses, mosquito catch sites and malaria posts within study sites.

Results

3229 villagers (1689 male) were included in this study. During the study period, 80 study participants were diagnosed with clinical P. falciparum. 201 participants were found to have P. falciparum by uPCR. 325 uPCR-positive participants had Plasmodium infections not identifiable at the species level and thus were not included in these analyses. Total numbers of clinical episodes and uPCR-detected infections were higher than the total number of infected individuals because some participants had multiple infections.

After MDA, the vast majority of clinical P. falciparum episodes occurred in only one of the study villages (TOT). 66 out of the 80 participants who had a clinical P. falciparum episode were from TOT village (three from HKT, seven from TPN and four from KNH).

Eleven of the 80 participants (14%) who had a clinical P. falciparum episode during the study period had repeated clinical episodes and 19 of the 80 (24%) participants who had a clinical episode were found to have a uPCR-detected P. falciparum infection in at least one of the surveys. uPCR-detected P. falciparum infections were more prevalent in males than females (UOR: 2.03; CI: 1.50–2.76).

Spatiotemporal patterns in clinical episodes, uPCR-detected infections, and MDA adherence

uPCR-detected P. falciparum infections were widespread in all villages at baseline (Figure 2). These infections were significantly reduced following MDA in all villages. The prevalence of uPCR-detected P. falciparum infections had reduced in two control villages (villages TPN and HKT) prior to MDA.

Figure 2.

Clinical P.falciparum episodes (yellow square points) and uPCR-detected P. falciparum infections (blue dots) at house level over time (by survey month; month 0 (M0) through month 24 (M24)) for each of the four study villages.

Statistically significant clusters (detected using SaTScan) are indicated for both clinical episodes (underlying yellow circles) and uPCR-detected infections (underlying blue circles). Grey points indicate house locations for houses with no infections or episodes in a given time. In the maps clinical episodes are aggregated to align with surveys (i.e. M1, M2 and M3 aggregated into M3), though they were recorded and analyzed by individual month. The study was conducted from May 2013 through June 2015 (KNH began in June; TPN in May; HKT in July; and TOT in May of 2013).

There were statistically significant clusters of uPCR-detected P. falciparum infections in each village at baseline but subsequently no significant clusters were detected (Figure 2). Clusters of clinical P. falciparum episodes occurred in two villages (KNH and TOT). The cluster in KNH occurred from M5 through M7 but included only four episodes. There were two separate clusters in village TOT. A cluster in the western portion of the village began in M12 and lasted until M18 (with a total of 35 episodes). A single-house cluster occurred in the eastern portion of the village (M15 through M18) with five episodes among four house members (2 in a 10 yo male, 1 in a 48 yo male, 1 in a 16 yo male, and 1 in a 48-year-old female).

There were significant clusters of non-participation in the MDAs in three of the study villages (TPN, HKT and TOT (Appendix 1—figure 2)). The non-participation cluster in TOT made up a large portion of the western part of the village and included 115 individuals not participating in the MDA (out of 919 total individuals in TOT). The non-participation clusters in HKT and TPN included 206 and 15 individuals respectively.

Appendix 1— figure 2.

Spatial clusters (detecting using SatScan) of non-participation in MDA.

Sporadic clinical P. falciparum episodes occurred in village TOT following MDA (MDA was completed by M3), followed by a small outbreak beginning in M12 (Figures 2 and 3). The first clinical P. falciparum episodes during this outbreak occurred among villagers who lived in the cluster of non-MDA participation (Figure 3). By M15 the clinical episodes were occurring through much of the village (Figure 3).

Figure 3.

Spatiotemporal distribution of clinical P.falciparum episodes (yellow square points), uPCR-detected P. falciparum infections (blue dots), and a cluster of non-participation in MDA (grey circle/ochre border, detected using SatScan) in TOT village.

Season is indicated by colored squares in the top right corner of each map. A measure of the spread of clinical P. falciparum cases is given by the standard distance deviation (‘SDD’), indicated by the hollow circle with dark grey outline. One standard deviation is shown, indicating that roughly 68% of all cases lie inside of the circle. After MDA (M3), clinical episodes began occurring in the westernmost portion of the village. By month 15 (M15), clinical episodes were occurring throughout the village.

Cumulative hazards plots of clinical P. falciparum episodes in village TOT illustrate the temporal patterns in infections according to neighborhood MDA non-adherence (aggregated into terciles) and individual participation in MDA (Figure 4). The proportion of individuals who had acquired a clinical P. falciparum episode began consistently increasing in M12 for those living in either mid or high MDA non-adherence neighborhoods. P. falciparum episodes among low MDA non-adherence neighborhoods began increasing approximately 1 month after the increase in high non-adherence neighborhoods but never reached the level experienced in either mid or high MDA non-adherence neighborhoods. 4.4% of all individuals in low MDA non-adherence neighborhoods had at least one clinical P. falciparum episode by the end of the study period, in comparison to 7.6% in mid and 9.6% in high MDA non-adherence neighborhoods (log-rank test p-value=0.0485; Figure 4A).

Figure 4.

Cumulative hazard of having a clinical P. falciparum episode by MDA adherence.

Cumulative hazard for clinical P. falciparum episodes in village TOT by (A) neighborhood MDA adherence (‘low non-adhere’ is a neighborhood with a low proportion of non-adherents; ‘high non-adhere’ is a neighborhood with a high proportion of non-adherents), (B) neighborhood adherence (same as in A) and individual adherence (‘individual non-participation’ indicates individuals who took no MDA while ‘individual participation’ indicates individuals who took at least 1 round of MDA). Figure 4B indicates that individuals who participated in MDA and lived in a neighborhood with low adherence had the highest risk of having a clinical episode post-MDA. Individuals who took no rounds of MDA but lived in a neighborhood with a high proportion of adherents had the lowest risk of acquiring a clinical episode post-MDA.

The increase in clinical P. falciparum episodes in M12 also coincided with an increase in HBR in village TOT (Appendix 1—figure 3).

Appendix 1— figure 3.

Human biting rate (HBR) for primary vectors by study month.

Longitudinal multivariable analysis of clinical P. falciparum episodes

After MDA, clinical P. falciparum episodes in village TOT were most likely to occur among 5 to 14 year olds (AOR: 3.41; CI: 1.33–8.77, compared to 0 to 4 year olds) and participants who lived in a house with someone else who had a clinical P. falciparum episode during the same month (AOR: 3.43; CI: 1.52–7.72), after adjusting for other covariates (Table 1). Individuals who lived in a neighborhood with a high proportion of people who did not adhere to MDA had 2.8 times the odds of having a clinical episode (AOR: 2.85; CI: 1.28–6.37) compared to people who lived in neighborhoods where most people adhered to MDA (Table 1). The human biting rate was also associated with increased odds of having a clinical P. falciparum episode, with a 10% increase in odds for every one unit increase in HBR (AOR: 1.09; CI: 1.05–1.13).

Table 1.

Multivariable mixed effects logistic regression for odds of having a clinical P. falciparum episode (village TOT only).

The model includes a random intercept for individual participants, with repeat observations occurring within individuals over the study period.

Covariate	AOR	p-Value
Age 0 to 4	Comparison
Age 5 to 14	3.41 (1.33–8.77)	0.0104
Age 15 plus	2.17 (0.86–5.46)	0.1053
Female	Comparison
Male	1.19 (0.66–2.11)	0.5612
Participated in no rounds of MDA	Comparison
Participated in MDA (at least one round)	1.43 (0.73–2.78)	0.2994
No house member with clinical episode	comparison
House member with clinical episode	3.43 (1.52–7.72)	0.0004
Low neighborhood non-adherence to MDA	comparison
Mid neighborhood non-adherence to MDA	2.00 (0.87–4.60)	0.0879
High neighborhood non-adherence to MDA	2.85 (1.28–6.37)	0.0098
Mean village HBR	1.09 (1.05–1.13)	<0.0001
Study month	1.19 (1.10–1.27)	<0.0001

Discussion

The primary objective of this research was to look for a potential herd effect and at the impact of non-adherence with regard to MDA for P. falciparum malaria. There was an apparent group level effect from MDA adherence, suggesting herd protection and evident from three lines of evidence.

First, clinical episodes decreased among all groups for a longer period than the prophylactic effect (approximately 1 month) of the administered antimalarials. This group level protective effect from MDA was also evident in the rainy season following MDA (Figure 3) which corresponded to a surge in vector activity (Appendix 1—figure 3). Once the P. falciparum outbreak began (M12), there was a lag of approximately 1 month between the onset of clinical episodes in neighborhoods with mid and low MDA adherence and then occurred in neighborhoods with high MDA adherence (Figure 4A).

Second, neighborhoods with high MDA adherence never experienced the same levels of infection as those with mid or low MDA adherence (Figure 4A). Individuals who participated in MDA but lived in a neighborhood with low adherence had the highest risk of having a clinical episode whereas those who did not participate in MDA but lived in a neighborhood with high adherence had the lowest risk of having a clinical P. falciparum episode (Figure 4B). As has been described in other settings, this individual-level finding may be related to relative perceptions of risk; with potential complacency among individuals living in areas with lower levels of malaria (Koenker et al., 2013).

Third, the results from the multivariable logistic regression also suggest that living in a neighborhood with a high proportion of people who did not adhere to MDA was a significant risk factor for acquiring a clinical P. falciparum episode, after adjusting for individual MDA adherence and other important predictors of having a clinical episode (Table 1). To our knowledge, this is the first documentation of a herd effect conferred by MDA for P. falciparum malaria.

The increase in clinical P. falciparum episodes post-MDA also corresponded to an increase in village HBR. HBR also peaked in one other village (HKT) at the same time as in village TOT (Appendix 1—figure 3), but occurred in the absence of a detectable parasite reservoir and the HBR did not persist at high levels. Evidence also suggests that the MP in TOT was not functioning well in the first year of the study (reported in Landier et al., 2017a). The combination of a persisting parasite reservoir and persistently high HBR (from M13 – M18) in TOT likely explains the drastically different results between the study villages with regard to P. falciparum malaria elimination (Figure 2). A better functioning MP in TOT would likely have reduced the size of the outbreak.

Clustering of uPCR-detected P. falciparum infections across houses occurred for limited periods of time only prior to MDA (Figure 2). This clustering suggests that interventions such as reactive case detection would have resulted in the detection of extra cases (of both clinical and uPCR-detected P. falciparum) when searching within houses and occasionally in neighboring houses, but these would have only been a small proportion of all infections within the villages (Parker et al., 2016) and would not have halted transmission. Conversely, community based early diagnosis and treatment and MDA with high participation, targeted at the village scale or larger, appear effective at reducing prevalence, incidence, and transmission of P. falciparum (Landier et al., 2018b).

There are several limitations to this work. Individuals who did not participate in MDA also did not participate in blood screenings immediately after MDA (i.e. M3 in village TOT). uPCR-detected infections are therefore likely to be underdiagnosed for these individuals, and it is likely that such infections clustered and overlapped with the clusters of non-adherence to MDA. While no genetic analyses were done with samples from the village, it is likely that these underdiagnosed infections, combined with the high HBR, led to a resurgence of clinical episodes. There is also evidence of a poorly functioning MP in this village, which could have led to undiagnosed clinical episodes, especially during the beginning of the study. Some infections are likely to be acquired outside of the village, leading to complex spatial patterns in infections that are mapped at the house level. Within-household clustering can be the result of within-household transmission, or shared exposure outside of the household or village among household members. Finally, these data come from a limited number of villages (total of 4), with analysis of P. falciparum episodes coming from the sole village that continued to have P. falciparum after MDA. Given that clinical P. falciparum episodes post-MDA were only possible to analyze in a single village, and that neighborhoods were not discrete and overlapped (100 m buffer around each house), a neighborhood-level effect was not included in this analysis. It is possible that the confidence intervals around the neighborhood MDA adherence variable are therefore too small and would not have been statistically significant had a neighborhood effect been included. This work would benefit from analyses with larger datasets.

This work has relevance with regard to further research and practice concerning MDA. While participation is obviously crucial to success, there is unlikely to be a single adherence proportion that can be applied in all situations. In this study, the elimination efforts were successful in three out of four villages even though one of those villages (HKT) had a similar overall adherence to the village with P. falciparum remaining after MDA (TOT). Likewise, this work points toward the need for considering spatial scale in MDA and in MDA adherence. Most current MDA trials and programs in the GMS and elsewhere consider a single village or community as the target unit. In some cases, especially when high prevalence villages are spatially clustered across a landscape, it may be necessary to target units above the village level (i.e. groups of villages).

Materials and methods

Study location and design

The study site consisted of four villages (KNH, TPN, HKT, and TOT) along the Myanmar-Thailand border, in Kayin (Karen) State, Myanmar (Landier et al., 2017b). The villages were selected based on P. falciparum malaria prevalence surveys using ultrasensitive quantitative PCR (uPCR) (Imwong et al., 2015) and were part of a MDA pilot study (Landier et al., 2017b). The northernmost village is approximately 105 km from the southernmost and the two closest villages, KNH and TPN are within 10 km of each other (Figure 1). The study was conducted from May 2013 through June 2015.

A full population census was completed in each of the four study villages at baseline May – June 2013. Everyone enumerated in the census was given a unique identification code. Geographic coordinates were collected for all houses in the four study villages and a unique identification code was assigned to each house. All individuals were then linked to their respective houses.

Blood surveys were conducted at baseline in each village, aiming to screen all individuals above an age of 9 months. Venous blood (3 mL) was drawn from each participant, transported to a central laboratory and analyzed using a highly sensitive quantitative PCR (uPCR) assay with a limit of detection of 22 parasites per mL. (Imwong et al., 2014). Infections detected through these blood screenings are hereafter referred to as uPCR-detected infections. Most (86%) uPCR-detected infections were subclinical (Landier et al., 2017a) and clinical infections were provided the standard treatment (see below).

A community-based malaria clinic (referred to as a malaria post or MP) was established in each village at the beginning of the project, as part of the malaria intervention. Village health workers were trained to diagnose malaria using rapid diagnostic tests (RDTs) and to treat RDT positive infections with dose based on weight and age. The ID code of each participant who self-presented at the MP was recorded, along with RDT results, and these cases are hereafter referred to as clinical malaria episodes of either P. falciparum or P. vivax. Malaria episodes were treated with dihydroartemisinin-piperaquine (DHA + P) for P. falciparum and chloroquine for P. vivax. Radical cure for P. vivax was not provided because the absence of G6PD (Glucose-6-phosphate dehydrogenase) tests required to prevent hemolysis in G6PD individuals (Bancone et al., 2014; Chu et al., 2017).

MDAs were conducted in two villages at the beginning of the study (month 0, M0) and extended to the two control villages beginning in month 9 (M9). Restricted randomization was used to decide which villages received early or deferred MDA. MDA consisted of 3 days of DHA + P, with a single low dose of primaquine on the third day, repeated over three months (M0, M1, M2 for the first group and M9, M10, M11 for the control group). Follow-up blood surveys were conducted in each village every third month after M0 until M18. A final full blood survey was completed in each village at M24.

Mosquitoes were collected monthly using human landing catches to estimate the human biting rate (HBR). Mosquito catching teams were based at five sites (both indoors and outdoors) within each of the four study villages (total of 20 catch sites) for five consecutive nights during the study period M0 through M20. Mosquitoes were caught using glass tubes and later identified morphologically (Ya-Umphan et al., 2017).

The locations of study villages, MPs, catch sites, and village houses are indicated in Figure 1.

Analysis

Variables

All individuals recorded in the census with a house address in the four study villages were included in this analysis.

The data were aggregated into 1 month time steps and individuals were coded with a ‘1’ for any month in which they presented at the village MP and were diagnosed with P. falciparum. Likewise, individuals who did not have a clinical episode within a given month were coded with a ‘0’ for that respective month. Individuals who were ever diagnosed with a clinical episode or uPCR-detected P. falciparum infections were likewise coded as a ‘1’ for analyses of having ever been detected by uPCR for an infection or having ever had a clinical episode during the study period.

Predictor variables (covariates) are listed in Table 2. Individual-level predictors included age group, gender, infection status, and adherence to MDA. Household-level predictor variables included a binary variable for whether or not another household member had a clinical episode and whether or not another household member had a uPCR-detected infection.

Table 2.

Table of predictor variables (covariates) used in regressions.

Covariate	Level	Description
Age group	Individual	Ordinal; age groups: 0 to 4; 5 to 14; and 15 and above
Gender	Individual	Binary; male or female
Individual adherence to MDA	Individual	Binary; whether an individual participated in MDA or not (at least one full round)
Household member with clinical episode	Household	Binary; one if another house member had a clinical episode and 0 if not
Household member with uPCR-detected infection	Household	Binary; one if another house member had a uPCR-detected infection and 0 if not
Neighborhood MDA non-adherence	Household/neighborhood	Ordinal (split into tertiles); proportion of people within 100 m radius who did not complete all three rounds of MDA
Human biting rate (HBR)	Village	Continuous; average number of bites per person per night
Study month	Village	Continuous; 1–26 (from May 2013 through June 2015); included as a control

Neighborhood MDA non-adherence was calculated as the proportion of people who took no rounds of MDA within 100 m radius of each house in the study population. This proportion was calculated for each house in the study villages and non-adherence proportions were then attributed to individuals based on the house to which they were attributed.

The human biting rate (HBR) for primary vectors (Anopheles minimus s.l., An. maculatus s.l., and An. dirus s.l.) was calculated for each month.

Exploratory spatial and temporal analyses

All predictor variables were explored in bivariate analyses. Unadjusted odds ratios (UOR) were calculated for binary predictors and Wilcox rank sum tests were calculated for continuous variables. Cumulative hazards curves were used to analyze temporal patterns in clinical episodes. uPCR-detected infections (from surveys) and clinical episodes (from the MPs) were mapped at the house level across time. Maps were created for each village and each survey time point (months 0 through 24: M0 – M24), with clinical episodes aggregated to align with surveys (i.e. M1, M2 and M3 aggregated into M3).

The weighted standard distance deviation (SDD) was used to visually analyze the distribution of clinical P. falciparum episodes for each survey time point in the one village with sufficient P. falciparum episodes (TOT) post-MDA. Clinical episodes were aggregated to align with survey time points (i.e. months 1 through 3 were aggregated and plotted in month 3). One SDD was calculated, corresponding to approximately 68% of all points falling inside of the resulting circle.

Scan statistics were used to test for clustering of uPCR-detected P. falciparum infections across survey months; clinical P. falciparum episodes across all months of the study period; and MDA non-participation (Kulldorff, 1997). The scan statistics used a moving window (a circle) that centered on each point in the village, testing for the relative risk of cases given a population size within the circle in comparison to the risk outside of the circle (Kulldorff, 1997). The circle increased in size until it included half of the population and then moved to the next geographic reference point. For Plasmodium infections and malaria episodes the space-time discrete Poisson model was used, whereas for MDA participation a purely spatial Poisson model was used (Kulldorff, 1997) (as MDAs were completed within a 3-month time period).

Multivariable logistic regression

A multivariable mixed effects logistic regression was used to estimate model adjusted odds ratios (AOR) and confidence intervals for predictor covariates with regard to the individual odds of having a clinical P. falciparum malaria episode. Since clinical P. falciparum episodes were almost exclusively limited to a single village post-MDA (village TOT), the analysis of clinical P. falciparum episodes focuses on this village alone.

Covariates for these models are listed in Table 2 and included individual, household, neighborhood, and village-level predictors. Model and variable selection are described in detail in Appendix 1. Results for uPCR-detected P. falciparum infections are listed in Appendix 1. While this study focuses on P. falciparum, P. vivax data were also collected and are included in secondary analyses in the Appendix.

Software

Exploratory statistics and regressions were calculated using R (version 3.4.3; https://cran.r-project.org/) and the ‘epiR’, ‘lme4’, and ‘survival’ packages. All maps were created using ArcGIS 10.5 (https://www.arcgis.com/). Exploratory spatial data analysis was conducted using ArcGIS 10.5 and SatScan v9.5 (https://www.satscan.org/). The neighborhood participation variable was created using ArcGIS and the Python programming language (version 3.5.2; https://www.python.org/).

Ethics approval

The study protocol was reviewed and approved by the Oxford Tropical Research Ethics Committee (reference no. 1017–13 and 1015–13), the Tak Province Community Ethics Advisory Board (T-CAB), and by village committees in each of the four study villages. Survey and mass drug administration participation were voluntary and all participants provided written informed consent. Participants < 18 yo provided assent to participate, along with consent from their guardian(s). Potential participants received study information in their mother language (Karen or Burmese) both through community engagement activities on through one-on-one meetings. The ethics committees require consent to publish when individuals are identifiable. Since no individuals are identifiable through this publication, no consent to publish was obtained.

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Herd protection against Plasmodium falciparum infections conferred by mass antimalarial drug administrations" for consideration by eLife. Your article has been reviewed by Prabhat Jha as the Senior Editor, a Reviewing Editor, and three reviewers. The following individual involved in review of your submission has agreed to reveal her identity: Gillian Stresman (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

This paper presents an interesting dataset on the spatial and temporal patterns of malaria before and after a mass drug administration intervention in 4 villages. Understanding the effect of community-level protection that results from mass drug administration campaigns is currently a key gap in planning for malaria elimination activities. The data are impressively detailed.

Essential revisions:

This paper presents a very interesting dataset on the spatial and temporal patterns of malaria before and after a mass drug administration intervention in 4 villages. Understanding the effect of community-level protection that results from mass drug administration campaigns is currently a key gap in planning for malaria elimination activities. The data are impressively detailed.

However, we have concerns around the presentation of ideas, results, and conclusions with the overall narrative lacking cohesion/focus. For example, the aim of the work is to look at the presence of herd immunity as a result of MDA but the majority of the introduction section focuses on justifying the use of MDA instead of the literature around adherence being a critical factor in the success of MDA, the concept of herd immunity, how critical thresholds for coverage do not exist but potential ways of estimating them etc. Also, the subsequent sections jump between outcome of temporal and spatial analysis, between Pf vs. Pv and infection vs. clinical case making it hard to follow. Overall, the aims, results, and discussion do not match the title and conclusions of paper. We suggest that the authors take time to carefully frame the story they are trying to tell. One suggestion would be to include the Pv work as supplemental information. The outcome of the MDA on Pv is essentially a null result with the MDA implementation not being designed to have an impact on transmission (no radical cure). The findings are important but not surprising and therefore could be included as supplemental so it doesn't detract from the main and more interesting outcome of the impact on Pf.

A second major concern is that the work does not appear to be placed within the larger research and programmatic context. A lot of the research from the GMS region come from this institution, but degree of self-citation is concerning with ~50% of references coming from the research group. Furthermore, the authors appear to be drawing conclusions achieved in other papers without referencing instead of focusing on the aims of this specific piece of work. For example, the Discussion section comments on individual-level results, that were not presented in this paper but have been highlighted in other papers, on the same data without referencing (e.g. "no protective effect.[…] at the individual level"; "those who lived in […] regardless of their individual participation in MDA").

Third, the approach taken to estimate HBR, a key factor suggested to contribute to reintroduction of Pf transmission is questionable. Although a pragmatic approach, what level of spatial bias has been introduced by attributing household-level HBR values based on the nearest trapping site? Given the known high-degree of spatial heterogeneity at small scales in mosquito abundance, would this add an additional level of bias when just selecting the nearest sampling point? Would it be more appropriate to aggregate HBR according to neighborhood as has been defined by the adherence variable to ensure spatial congruence? This can be achieved by creating prediction surfaces using model based geostatistical methods with the HBR variable as outcome and then taking the mean predicted HBR using the same window size as for adherence.

That said, there are many strong points about the analysis e.g. the possibility to look at both individual level and area-level covariates, and the measurement of both symptomatic and asymptomatic infections and relationship between the two.

An interesting result was the apparent clustering of resurgent cases in the area with lower adherence to MDA in one village (apparent smaller herd effect). But the evidence for this was rather weak – it was based mainly on one cluster of non-adherence in one village, which also happened to experience a higher number of cases. This seems potentially explained by the higher biting rate of mosquitoes in this part of the village rather than the non-adherence, but it was not clear whether there was multivariate adjustment to check this? Also, the asymptomatic infections did not seem to follow the same pattern. Including this finding in the title and as a key result elsewhere in the paper places too much confidence in this result, and the conclusions should be more cautious. Perhaps generally the limitations of spatial analysis within only 4 villages could be further acknowledged.

We also suggest the authors explore further the effect of individual level adherence vs area-level average adherence, specifically are non-adherent individuals living in high-adherence neighbourhoods protected, compared with non-adherent individuals in low-adherence neighbourhoods? (and vice versa) It is not really clear at the moment whether this is a spatial effect, or whether it's just perhaps that non-adherent people are both more likely to be infected and to live in non-adherent neighbourhoods (subsection “Spatiotemporal patterns in uPCR-detected infections, clinical episodes and MDA adherence” and subsection “Longitudinal analysis of clinical P. falciparum and P. vivax episodes”). If a critical cutoff value for adherence levels to achieve a 'herd immunity' during MDA campaigns could be estimated using these data, it would be a major finding and significant improvement to this work. This would be possible to do with these data using ROC analysis with the continuous adherence variables.

Last, in general it was not clear whether multivariate analysis was done – if so, it would be important to see the effect of each variable after adjustment. If not, the authors need to explain why multivariate analysis was not performed.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Herd protection against Plasmodium falciparum infections conferred by mass antimalarial drug administrations" for further consideration at eLife. Your revised article has been favorably evaluated by Neil Ferguson (Senior Editor), a Reviewing Editor, and three reviewers.

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

Reviewer #1:

The rewritten paper is much clearer and easier to read than the initial version. Most of the points raised in the previous reviews have been well addressed.

Reviewer #2:

The authors have made substantial revisions based on the previous reviews and most points are thoroughly addressed. I like the new Figure 4.

The idea of a herd effect is interesting and certainly worth reporting, but my only remaining concern is that the conclusions and title are too strong based on the limited data from one village and also analytical problems. The authors have clarified in response to another reviewer that their multivariate model (Table 2) does not include random neighbourhood effects, which is not an appropriate analysis with which to test neighbourhood variables since it will overestimate the statistical significance of the herd effect result.

Ideally an alternative analysis needs to be tried, for example separating the village into discrete spatial units and adding random effects at this level. At the very least the authors should consider wording the Abstract and Title more cautiously – to clarify that this herd effect was only seen in one village and more investigation is needed to verify whether this is a real phenomenon.

Reviewer #3:

This paper is an interesting and very timely assessment of the community effect of a mass drug administration trial. The revised version of the paper is much improved but could be further strengthened by focusing the Introduction and Discussion section around the objective of the paper, instead of the more general role of MDA as is currently written. Specifically, I suggest that the authors should address the following comments.

Introduction: Are the authors trying to say that MDA should supplement existing malaria programs/used when the full package doesn't eliminate transmission (as is already stated)? There is an argument that this intervention could bypass weak health systems to interrupt transmission and it has been used under alternative scenarios (e.g. Ebola epidemic response). Should this be 'should ideally be used in settings…'

The potential risks associated of using MDA are missing from the introduction (i.e. developing drug resistance). Why has the WHO avoided recommending it for so long? What's changed their stance?

Introduction: The justification that MDA is used to achieve a population level effect is missing important nuances (and offers some repetition). The rationale of MDA is typically argued from the biological perspective whereby the transmission potential/reproductive rate of malaria is so high, to achieve interruption, enough of the parasitic reservoir needs to be removed before it has the chance to fuel subsequent infections. The concept of a "herd effect" applies in most communicable disease systems: If enough of the population is protected, transmission chains cannot be sustained and transmission with die out. What is 'enough' is unknown as stated and is affected by the propensity for diseases to cluster in populations (defined either spatially or demographically). If the aim of MDA for malaria is to interrupt transmission chains, the notion of a herd effect providing additional levels of population protection is plausible but has never been examined empirically.

Subsection “Analysis”: what was the rationale for using a 100m radius. Are the results affected according to size of buffer?

Subsection “Analysis”: is there an impact when considering individual level (e.g. proportion of individuals non-compliant) vs household level (e.g. households with 80%? compliance) non-compliance?

Discussion section: MDA adherents in low-adherence areas had highest risk? Even compared to those in the same neighborhoods with low-adherence? This is an interesting finding consistent with what has been observed with bednets: those who have the most malaria are more likely to adhere to interventions, potentially related to perceptions of risk.

Discussion section: I would encourage authors to split this paragraph into 3 sections to ensure that these key observations can be discussed in depth.

Discussion section: do the authors have evidence to support this counterfactual? Or do they mean would have likely/may have resulted in… ?

Discussion section: I'm not sure of the relevant of this paragraph to the objectives of the work. The focus is on the herd protection or am I missing a connection between adherence and the more general rational for MDA? What are the potential limitations to MDA when considering the potential impacts of herd protection that may be specific to this setting?

Discussion section: Similar to point above, the objective is the work is related to herd immunity and the concluding remarks seem to focus on the general impact of MDA. Here, I would encourage the authors to include some insight revolving around the potential that comes with an improved understanding of the herd effect for malaria control and elimination activities. Next steps for further exploring this phenomenon?

Discussion section: The objective of the study was to look at the herd effect and the impact of non-adherence. I would encourage the authors to re-write this first paragraph of the discussion to focus on this instead of what has been reported in other publications.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Potential herd protection against Plasmodium falciparum infections conferred by mass antimalarial drug administrations" for further consideration at eLife. Your revised article has been favorably evaluated by Neil Ferguson (Senior Editor), a Reviewing Editor, and 3 reviewers.

The manuscript has been improved but there are some minor remaining issues that need to be addressed before acceptance, as outlined below in the individual reports.

Reviewer #1:

I had no substantive concerns with the previous version and feel that the manuscript is essentially suitable for publication.

However, in the course of the review process the authors have made incremental amendments that understate the very real concerns that exist around the potential for selection of resistance by MDA. In particular, reference 18 (one of the authors' own publications, which is not so easy to access because of a paywall) is cited to support the contention that there is "no evidence of resistance emerging as a result of controlled, targeted MDA with the use of therapeutic drug levels". Arguable, such evidence is exactly what was reported by Trigg et al. for pyrimethamine/sulfadoxine in Tanzania (Trig.et al., 1997). Even if there were no evidence, the absence of evidence is not good evidence for absence of an effect, and the case for no effect seems to be summarised in the Abstract of von Seidlein and Dondorp 2015 by the statement "it is difficult to conceptualize how targeted malaria elimination could contribute to artemisinin resistance…". The wording here is critical: it is rather easy to conceptualise how failed attempts at elimination using MDA could lead to resistance, especially if drug use is not well controlled. Since this paper is not about drug resistance, I think that the claims about resistance in the Introduction should be removed.

Reviewer #2:

I copy here my previous review and the authors' response, with my current response below:

“The authors have made substantial revisions based on the previous reviews and most points are thoroughly addressed. I like the new Figure 4.

The idea of a herd effect is interesting and certainly worth reporting, but my only remaining concern is that the conclusions and title are too strong based on the limited data from one village and also analytical problems. The authors have clarified in response to another reviewer that their multivariate model (Table 2) does not include random neighbourhood effects, which is not an appropriate analysis with which to test neighbourhood variables since it will overestimate the statistical significance of the herd effect result.

Ideally an alternative analysis needs to be tried, for example separating the village into discrete spatial units and adding random effects at this level. At the very least the authors should consider wording the abstract and title more cautiously – to clarify that this herd effect was only seen in one village and more investigation is needed to verify whether this is a real phenomenon.”

“Yes, this is an approach used, for example, by Ali et al. (2005, Lancet, "Herd immunity conferred by killed oral cholera vaccines in Bangladesh") and in some ways it would be ideal. With our data some problems would remain though because any discrete spatial units ("neighborhoods") that are imposed on the villages would be arbitrary and since the villages are quite small, there would only be 4-5 each (unless using very small "neighborhoods"). For the one village with Pf after MDA, we'd only have 4 or 5 neighborhoods and we would question if a random intercept would actually be useful, given that there are assumptions about the mean and variance of the random intercept.

There is also a spatial argument to be made in that an arbitrary boundary or line that would split such areas within the village would be creating a boundary between houses that are extremely close to each other. If geographic proximity is important here, and we believe it is, then we would actually lose a lot of information by using this approach.

Regardless, the reviewer is correct that we need to be cautious in our wording here. We have taken the suggestion to word the Title and Abstract more cautiously, and we discuss some of these issues in the Discussion section.”

Thank you for addressing these points very well in terms of altering the abstract and title. For my last suggestion, please can the authors either (a) explicitly state that the confidence interval reported for the paper's main result is likely to be much too narrow, because it tests for the effect of a neighbourhood level variable (coverage within 100m of a household) using individual level data without allowing for neighbourhood effects. Or (b) remove the confidence interval entirely and explain why it could not be calculated. Or (c) try an alternative analysis.

I appreciate the difficulties of doing an analysis which incorporates neighbourhood effects, but the confidence interval of the main result in the Abstract, shown in subsection “Longitudinal multivariable analysis of clinical P. falciparum episodes” as it stands must therefore be incorrect (AOR: 2.85; CI: 1.28 – 6.37). It would be unlikely to be a significant result if neighbourhood effects were included. It is still an interesting result, which could be explored further in other larger datasets.

Otherwise the manuscript is very interesting and useful, and I have no further comments.

Reviewer #3:

This work is much improved. I have no additional comments and recommend the work for publication.

Essential revisions:

[…] 1) However, we have concerns around the presentation of ideas, results, and conclusions with the overall narrative lacking cohesion/focus. For example, the aim of the work is to look at the presence of herd immunity as a result of MDA but the majority of the introduction section focuses on justifying the use of MDA instead of the literature around adherence being a critical factor in the success of MDA, the concept of herd immunity, how critical thresholds for coverage do not exist but potential ways of estimating them etc. Also, the subsequent sections jump between outcome of temporal and spatial analysis, between Pf vs. Pv and infection vs. clinical case making it hard to follow. Overall, the aims, results, and discussion do not match the title and conclusions of paper. We suggest that the authors take time to carefully frame the story they are trying to tell. One suggestion would be to include the Pv work as supplemental information. The outcome of the MDA on Pv is essentially a null result with the MDA implementation not being designed to have an impact on transmission (no radical cure). The findings are important but not surprising and therefore could be included as supplemental so it doesn't detract from the main and more interesting outcome of the impact on Pf.

We have now refocused and rewritten the manuscript in order to address this valid critique. The main goal of this work was to analyze clinical P. falciparum malaria after MDA had been administered, looking at both individual and group-level effects. Since the work was focused on clinical P. falciparum we’ve now removed all analysis of P. vivax and uPCR-detected P. falciparum and offer these analyses in the supplementary material. We’ve rewritten the Introduction to focus more generally on MDA for P. falciparum malaria and have revised much of the rest of the manuscript in order to improve readability. We believe that these changes make the manuscript much more clear and concise.

2) A second major concern is that the work does not appear to be placed within the larger research and programmatic context. A lot of the research from the GMS region come from this institution, but degree of self-citation is concerning with ~50% of references coming from the research group. Furthermore, the authors appear to be drawing conclusions achieved in other papers without referencing instead of focusing on the aims of this specific piece of work. For example, the Discussion section comments on individual-level results, that were not presented in this paper but have been highlighted in other papers, on the same data without referencing (e.g. "no protective effect […] at the individual level"; "those who lived in [.…] regardless of their individual participation in MDA").

In our refocusing of this work, we now begin by discussing MDA in general rather than the GMS context. We’ve now attempted to appropriately reference other research teams in this revision.

Regarding results from other manuscripts – these data have been previously analyzed (Landier et al., 2017). However, the effects that we mention (for example individual versus group effects from MDA) are from this analysis. The analysis of risk factors for acquiring a clinical P. falciparum episode after MDA (results in Table 2) uses a multivariable regression which includes both individual and group (“neighborhood”) level variables for MDA adherence. We hope that this is now made clear through our revision of the manuscript.

3) Third, the approach taken to estimate HBR, a key factor suggested to contribute to reintroduction of Pf transmission is questionable. Although a pragmatic approach, what level of spatial bias has been introduced by attributing household-level HBR values based on the nearest trapping site? Given the known high-degree of spatial heterogeneity at small scales in mosquito abundance, would this add an additional level of bias when just selecting the nearest sampling point? Would it be more appropriate to aggregate HBR according to neighborhood as has been defined by the adherence variable to ensure spatial congruence? This can be achieved by creating prediction surfaces using model based geostatistical methods with the HBR variable as outcome and then taking the mean predicted HBR using the same window size as for adherence.

The reviewer is correct, both about micro-heterogeneities in HBR and that our specification of this variable is overly simplistic. We agree that a smoothed surface would be better, however we don’t believe that these data are appropriate for making prediction surfaces. There were only 5 catch sites per village and most villages are not in geographic proximity to each other (therefore not appropriate to make a single surface for all villages). The catch sites are not systematically distributed through the villages and their geographic extent is not as broad as the geographic extent of houses in each village. We would therefore have to create interpolations from only 5 points and then extrapolate outside of the 5 points to meet the geographic range of the villages. To address this problem in our analysis we have therefore opted to use HBR calculations at the village level in the regressions. This is still simplistic and doesn’t address micro-scale heterogeneities. In the absence of detailed human movement data and HBR outside of the village, we believe that this specification for HBR is the best possible approach with these data.

4) An interesting result was the apparent clustering of resurgent cases in the area with lower adherence to MDA in one village (apparent smaller herd effect). But the evidence for this was rather weak – it was based mainly on one cluster of non-adherence in one village, which also happened to experience a higher number of cases. This seems potentially explained by the higher biting rate of mosquitoes in this part of the village rather than the non-adherence, but it was not clear whether there was multivariate adjustment to check this? Also, the asymptomatic infections did not seem to follow the same pattern. Including this finding in the title and as a key result elsewhere in the paper places too much confidence in this result, and the conclusions should be more cautious. Perhaps generally the limitations of spatial analysis within only 4 villages could be further acknowledged.

Please see our response to point 2 above. We hope that the revised manuscript now clarifies that the models were multivariable and are looking at both individual and group level MDA adherence as well as HBR.

The reviewer is correct that these are a limited set of data and that we should be cautious in over-interpreting the results (which mostly come from a single village that had remaining Pf after MDA). We have added the following statement to the Discussion section:

“Finally, these data come from a limited number of villages (total of 4), with analysis of P. falciparum episodes coming from the sole village that continued to have P. falciparum after MDA.”

5) We also suggest the authors explore further the effect of individual level adherence vs area-level average adherence, specifically are non-adherent individuals living in high-adherence neighbourhoods protected, compared with non-adherent individuals in low-adherence neighbourhoods? (and vice versa) It is not really clear at the moment whether this is a spatial effect, or whether it's just perhaps that non-adherent people are both more likely to be infected and to live in non-adherent neighbourhoods (subsection “Spatiotemporal patterns in uPCR-detected infections, clinical episodes and MDA adherence” and subsection “Longitudinal analysis of clinical P. falciparum and P. vivax episodes”).

Please see our response to point 2 above. We hope that the revised manuscript now clarifies that the models were multivariable and are looking at both individual and group /area level adherence and effects.

6) If a critical cutoff value for adherence levels to achieve a 'herd immunity' during MDA campaigns could be estimated using these data, it would be a major finding and significant improvement to this work. This would be possible to do with these data using ROC analysis with the continuous adherence variables.

Our goal in this paper is to show that there is a measurable herd effect, but we believe that it is drastically complicated by context and that a single proportion of adherence won’t be applicable in all target populations. This analysis will soon be followed by a mathematical modelling paper using these and other data in order to further explore this topic.

7) Last, in general it was not clear whether multivariate analysis was done – if so, it would be important to see the effect of each variable after adjustment. If not, the authors need to explain why multivariate analysis was not performed.

Please see our response to point 2 above. We hope that the revised manuscript now clarifies that the models were multivariable and are looking at both individual and group /area level adherence and effects.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Reviewer #2:

[…] The idea of a herd effect is interesting and certainly worth reporting, but my only remaining concern is that the conclusions and title are too strong based on the limited data from one village and also analytical problems. The authors have clarified in response to another reviewer that their multivariate model (Table 2) does not include random neighbourhood effects, which is not an appropriate analysis with which to test neighbourhood variables since it will overestimate the statistical significance of the herd effect result.

Ideally an alternative analysis needs to be tried, for example separating the village into discrete spatial units and adding random effects at this level. At the very least the authors should consider wording the Abstract and Title more cautiously – to clarify that this herd effect was only seen in one village and more investigation is needed to verify whether this is a real phenomenon.

Yes, this is an approach used, for example, by Ali et al., (2005) and in some ways it would be ideal. With our data some problems would remain though because any discrete spatial units (“neighborhoods”) that are imposed on the villages would be arbitrary and since the villages are quite small, there would only be 4-5 each (unless using very small “neighborhoods”). For the one village with Pf after MDA, we’d only have 4 or 5 neighborhoods and we would question if a random intercept would actually be useful, given that there are assumptions about the mean and variance of the random intercept.

There is also a spatial argument to be made in that an arbitrary boundary or line that would split such areas within the village would be creating a boundary between houses that are extremely close to each other. If geographic proximity is important here, and we believe it is, then we would actually lose a lot of information by using this approach.

Regardless, the reviewer is correct that we need to be cautious in our wording here. We have taken the suggestion to word the title and abstract more cautiously, and we discuss some of these issues in the Discussion section.

Reviewer #3:

[…] Introduction: Are the authors trying to say that MDA should supplement existing malaria programs/used when the full package doesn't eliminate transmission (as is already stated)? There is an argument that this intervention could bypass weak health systems to interrupt transmission and it has been used under alternative scenarios (e.g. Ebola epidemic response). Should this be 'should ideally be used in settings…'

Yes. We have now changed the wording of this statement (Introduction).

The potential risks associated of using MDA are missing from the introduction (i.e. developing drug resistance). Why has the WHO avoided recommending it for so long? What's changed their stance?

We have now added a few sentences discussing historical problems and concerns regarding MDA:

Introduction

“Given that drug pressure (through provision of antimalarial drugs) provides a survival advantage for resistant parasites, there has been some hesitance in using MDA for malaria. One historical malaria eradication campaign relied on the inclusion of sub-therapeutic levels of antimalarials distributed in table salt across large populations [15]. This program likely led to the emergence of parasite resistance in the same regions [16] and this has in part led to hesitance among some institutions (i.e. the World Health Organization and ministries of health) to implement MDA for malaria [17]. There is, however, no evidence of resistance emerging as a result of controlled, targeted MDA with the use of therapeutic drug levels [18].”

Introduction: The justification that MDA is used to achieve a population level effect is missing important nuances (and offers some repetition). The rationale of MDA is typically argued from the biological perspective whereby the transmission potential/reproductive rate of malaria is so high, to achieve interruption, enough of the parasitic reservoir needs to be removed before it has the chance to fuel subsequent infections. The concept of a "herd effect" applies in most communicable disease systems: If enough of the population is protected, transmission chains cannot be sustained and transmission with die out. What is 'enough' is unknown as stated and is affected by the propensity for diseases to cluster in populations (defined either spatially or demographically). If the aim of MDA for malaria is to interrupt transmission chains, the notion of a herd effect providing additional levels of population protection is plausible but has never been examined empirically.

We have now attempted to incorporate these nuances and some of this text (reworded) into this part of the manuscript:

Introduction

“While antimalarials are usually administered following diagnosis (confirmed or presumed) or used as a prophylactic, MDA is used because of an intended population- or community-level effect. […] If the aim for of antimalarial MDA is to interrupt transmission, the notion of a herd effect providing additional levels of population protection is plausible but has not been examined empirically [22].”

Subsection “Analysis”: what was the rationale for using a 100m radius. Are the results affected according to size of buffer?

The 100m radius was relatively arbitrary, but we did do a sensitivity analysis on the buffer size. We’ve added this detail in the Appendix now:

“The sensitivity of the model results with regard to the neighborhood MDA non-adherence buffer size was also tested. Neighborhood MDA-non-adherence was tested with varying buffer sizes, beginning with a radius of 20m and going up to a radius of 220m, by 40m intervals. Neighborhood buffers from 60m to 140m were statistically significant and had comparable model adjusted odds ratios (at 60m: AOR = 2.33; CI: 1.03 – 5.25; at 100m: AOR = 2.84; CI: 1.29 – 6.27; at 140m: AOR = 2.35; CI: 1.06 – 5.19). Small buffers (i.e. 20m) included only one or very few households. At larger buffer sizes (> 180m) most houses were included within a buffer (furthest distance between any two houses was approximately 1km) leading to little spatial heterogeneity in participation across the village.”

Subsection “Analysis”: is there an impact when considering individual level (e.g. proportion of individuals non-compliant) vs household level (e.g. households with 80%? compliance) non-compliance?

For this analysis we’ve only considered the individual level. We prefer to not do the analysis using the household level because of the variation in household size (histogram attached here). For example, a house with 10 members and 80% adherence could still have 2 members with parasites in their blood while a house with 5 members and only 60% adherence would still have 2 members with parasites in their blood. Arguably, these 2 individuals are equally infective for others in the neighborhood, but would be treated as though they were completely different if we analyzed based on the house level.

Discussion section: MDA adherents in low-adherence areas had highest risk? Even compared to those in the same neighborhoods with low-adherence? This is an interesting finding consistent with what has been observed with bednets: those who have the most malaria are more likely to adhere to interventions, potentially related to perceptions of risk.

We have now added the following statement along these lines:

Discussion section

“As has been described in other settings, this individual-level finding may be related to relative perceptions of risk; with potential complacency among individuals living in areas with lower levels of malaria [23].”

Discussion section: I would encourage authors to split this paragraph into 3 sections to ensure that these key observations can be discussed in depth.

Done.

Discussion section: do the authors have evidence to support this counterfactual? Or do they mean would have likely/may have resulted in….

We have softened the statement as requested (Discussion section).

Discussion section: I'm not sure of the relevant of this paragraph to the objectives of the work. The focus is on the herd protection or am I missing a connection between adherence and the more general rational for MDA? What are the potential limitations to MDA when considering the potential impacts of herd protection that may be specific to this setting?

We have now dropped this paragraph.

Discussion section: Similar to point above, the objective is the work is related to herd immunity and the concluding remarks seem to focus on the general impact of MDA. Here, I would encourage the authors to include some insight revolving around the potential that comes with an improved understanding of the herd effect for malaria control and elimination activities. Next steps for further exploring this phenomenon?

We have now rewritten the conclusion to focus more specifically on herd effects from MDA for P. falciparum, general implications of this work, and future directions.

Discussion section: The objective of the study was to look at the herd effect and the impact of non-adherence. I would encourage the authors to re-write this first paragraph of the discussion to focus on this instead of what has been reported in other publications.

Done.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Reviewer #1:

[…] However, in the course of the review process the authors have made incremental amendments that understate the very real concerns that exist around the potential for selection of resistance by MDA. In particular, reference 18 (one of the authors' own publications, which is not so easy to access because of a paywall) is cited to support the contention that there is "no evidence of resistance emerging as a result of controlled, targeted MDA with the use of therapeutic drug levels". Arguable, such evidence is exactly what was reported by Trigg et al. for pyrimethamine/sulfadoxine in Tanzania (Trig et al., 1997). Even if there were no evidence, the absence of evidence is not good evidence for absence of an effect, and the case for no effect seems to be summarised in the Abstract of von Seidlein and Dondorp 2015 by the statement "it is difficult to conceptualize how targeted malaria elimination could contribute to artemisinin resistance…". The wording here is critical: it is rather easy to conceptualise how failed attempts at elimination using MDA could lead to resistance, especially if drug use is not well controlled. Since this paper is not about drug resistance, I think that the claims about resistance in the Introduction should be removed.

We were unsure of whether or not the reviewer wanted us to discard all discussion/claims about resistance in the paragraph – but we agree with a previous suggestion that the hesitancy for using MDA should at least be mentioned. We felt that the most contentious statement was the one referred to above (“with no evidence of resistance emerging as a result of controlled, targeted MDA…”) and we have therefore removed this statement from the manuscript.

Reviewer #2:

[…] For my last suggestion, please can the authors either (a) explicitly state that the confidence interval reported for the paper's main result is likely to be much too narrow, because it tests for the effect of a neighbourhood level variable (coverage within 100m of a household) using individual level data without allowing for neighbourhood effects. Or (b) remove the confidence interval entirely and explain why it could not be calculated. Or (c) try an alternative analysis.

I appreciate the difficulties of doing an analysis which incorporates neighbourhood effects, but the confidence interval of the main result in the Abstract, shown in subsection “Longitudinal multivariable analysis of clinical P. falciparum episodes” as it stands must therefore be incorrect (AOR: 2.85; CI: 1.28 – 6.37). It would be unlikely to be a significant result if neighbourhood effects were included. It is still an interesting result, which could be explored further in other larger datasets.

Done. We have included the following statement at the end of the Discussion section, immediately before stating our conclusions:

“Given that clinical P. falciparum episodes post-MDA were only possible to analyze in a single village, and that neighborhoods were not discrete and overlapped (100m buffer around each house), a neighborhood-level effect was not included in this analysis. It is possible that the confidence intervals around the neighborhood MDA adherence variable are therefore too small and would not have been statistically significant had a neighborhood effect been included. This work would benefit from analyses with larger datasets.”

Appendix 1—table 1.

Terciles (lower, middle and upper 1/3) of MDA non-adherence (% taking no rounds of MDA) in TOT

High non-adherence	>0.294
Mid non-adherence	>0.20 and <0.294
Low non-adherence	<0.20

Appendix 1—table 2.

Multivariable mixed effects logistic regression for odds of having a clinical P. falciparum episode.

The model includes a random intercept for individual participants, with repeat observations occurring within individuals over the study period. Unlike the model in the main text (Table 1) neighborhood non-adherence to MDA and individual adherence to MDA are continuous covariates. Study month was included as a control (a linear specification was used, but polynomial specifications were also tested). The covariates for human biting rate (HBR) and having a house member with a clinical episode in the same month were specified as time-varying covariates.

Covariate	AOR	p-value
Age 0 to 4	comparison
Age 5 to 14	3.5 (1.3–9.0)	0.0112
Age 15 plus	2.2 (0.9–5.6)	0.1014
female	comparison
Male	1.2 (0.7–2.2)	0.5147
no house member with clinical episode	comparison
house member with clinical episode	3.8 (1.8–7.7)	0.0003
proportion of MDA doses complete	1.0 (1.0–1.1)	0.4488
proportion of non-adherers in neighborhood	1.4 (1.0–1.8)	0.0380
mean village HBR	1.1 (1.1–1.1)	<0.0001
study month	1.2 (1.1–1.3)	<0.0001

Appendix 1—table 3.

Logistic regression for the odds of having a uPCR-detected P. falciparum infection after MDA.

Individuals in the data were coded as having an infection if they were ever determined by uPCR to have an infection through blood screenings in full village blood surveys after MDA. Almost all P. falciparum episodes occurred in a single village (TOT) and the analysis for P. falciparum was only conducted on data from that village. HBR is not included in this regression as it varies across time.

Covariate	AOR	p-value
0 to 4	comparison
five to 14	6.1 (1.1–113.9)	0.0877
15 plus	5.5 (1.1–99.3)	0.1017
female	comparison
male	1.3 (0.6–2.7)	0.5217
participated in no rounds of MDA	comparison
participated in MDA (at least one round)	0.4 (0.2–1.1)	0.0955
no house member with uPCR infection	comparison
house member with uPCR infection	1.1 (0.4–2.7)	0.8039
no house member with clinical episode	comparison
house member with clinical episode	1.7 (0.7–3.8)	0.2029
low neighborhood non-adherence to MDA	comparison
mid neighborhood non-adherence to MDA	1.1 (0.4–3.2)	0.8743
high neighborhood non-adherence to MDA	2.6 (1.0–7.1)	0.0488
number of surveys attended	1.3 (1.1–1.6)	0.0075

Appendix 1—table 4.

Multivariable mixed effects logistic regression for odds of having a clinical P. vivax episode.

The model includes a random intercept for individual participants, with repeat observations occurring within individuals over the study period. Study month was included as a control (a linear specification was used, but polynomial specifications were also tested). The covariates for human biting rate (HBR) and having a house member with a clinical episode in the same month were specified as time-varying covariates.

Covariate	AOR	p-value
Age 0 to 4	comparison
Age 5 to 14	1.0 (0.4–2.4)	0.9945
Age 15 plus	0.4 (0.2–0.9)	0.0358
female	comparison
male	0.8 (0.4–1.5)	0.4687
participated in no rounds of MDA	comparison
participated in MDA (at least one round)	1.7 (0.7–4.3)	0.2678
no house member with clinical episode	comparison
house member with clinical episode	5.8 (3.4–9.9)	<0.0001
low neighborhood non-adherence to MDA	comparison
mid neighborhood non-adherence to MDA	1.6 (0.3–7.6)	0.5546
high neighborhood non-adherence to MDA	1.8 (0.3–9.8)	0.5205
mean village HBR	1.0 (1.0–1.1)	0.1447
village KNH	comparison
village TPN	0.6 (0.2–1.8)	0.3666
village HKT	0.3 (0.1–1.5)	0.1329
village TOT	0.6 (0.1–3.4)	0.6074
study month	1.0 (1.0–1.1)	0.0126

Appendix 1—table 5.

Multivariable logistic regression for the odds of having a uPCR detected P. vivax infection after MDA.

Individuals in the data were coded as having an infection of either species if they were ever determined by uPCR to have an infection through blood screenings in full village blood surveys after MDA.

Covariate	AOR	p-value
0 to 4	comparison
five to 14	2.6 (1.7–3.9)	<0.0001
15 plus	2.1 (1.4–3.2)	0.0002
female	comparison
male	1.8 (1.4–2.2)	<0.0001
participated in no rounds of MDA	comparison
participated in MDA (at least one round)	0.6 (0.4–0.8)	0.0027
no house member with uPCR infection	comparison
house member with uPCR infection	1.5 (1.2–1.9)	0.0022
no house member with clinical episode	comparison
house member with clinical episode	1.3 (1.0–1.7)	0.0284
low neighborhood non-adherence to MDA	comparison
mid neighborhood non-adherence to MDA	0.6 (0.4–1.1)	0.0994
high neighborhood non-adherence to MDA	0.6 (0.4–1.1)	0.0945
KNH	comparison
TPN	1.0 (0.7–1.5)	0.8617
HKT	1.4 (0.8–2.5)	0.2614
TOT	5.4 (2.9–9.8)	<0.0001
number of surveys attended	1.4 (1.3–1.5)	<0.0001

Appendix 1—table 6.

Household and population counts for study villages

Data used in this analysis are available following the Mahidol-Oxford Tropical Medicine Research Unit data access policy. Both the policy and application form are available at: http://www.tropmedres.ac/data-sharing.

Village	Households	Population
KNH	86	504
TPN	75	468
HKT	176	1338
TOT	149	919
total	486	3229