Study participants were from 8 rural agricultural villages in Dodoma Province, central Tanzania (18–20). The study period was January 12, 2009–July 21, 2009 and was coincidental with a period in which rainfall was <60% of the average amount (21,22). AZT was offered to all the residents of the treatment villages, including residents not part of the follow-up investigation. Four treatment villages were selected on the basis of trachoma prevalence >10% in children 1–9 years of age (trachoma prevalence 12%, 18%, 18%, and 14%), rather than by randomization (18,19), which was consistent with World Health Organization AZT MDA guidelines for trachoma control (4). Four control villages with lower trachoma rates (8%–10%) were chosen on the basis of geographic proximity.

A complete census was conducted in the villages> In each village, 130 families with children <5 years of age were randomly selected for follow-up prevalence blood sampling; 1 child and 1 adult were randomly selected from each family. Sample size estimations were based on prestudy estimates of a malaria prevalence of 10% in children <5 years of age in the Dodoma region (23). Follow-up fingerprick blood sampling was performed at baseline and at 1, 3, 4, and 6 months later. Children were divided into 3 subgroups that were sampled once at weeks 2 (group 1), 6 (group 2), and 8 (group 3). Contemporaneous to the monthly blood sampling, staff members from villages conducted an active surveillance program in which children were visited weekly and screened for axillary temperature and questionnaires were answered regarding fever, diarrhea, and respiratory disease (18,19).

Blood samples were collected on ProteinSaver903 (GE Healthcare, Pittsburgh, PA, USA) for analysis by real-time PCR for all participants at each scheduled monthly visit (20). When a fever (temperature ≥37.5°C) in adults and children was observed or when a caregiver reported a history of fevers in children during the weekly surveillance, a rapid diagnostic test (RDT) (Paracheck-Pf; Orchid Biomedical Systems, Goa, India) and thick and thin blood films were prepared. RDTs were performed by trained study staff and interpreted in the field at the time of sampling. Two experienced microscopists at the Amani Laboratory in Muheza, Tanzania, read the slides blinded to the PCR or RDT results. Discordant results were read by a third microscopist. Persons with positive results by RDT and all febrile children <5 years of age were treated with artemether/lumefantrine according to national guidelines and removed from later analysis.

Information on bednet ownership was obtained through standardized participant interviews. Latitude, longitude, and altitude of home locations were measured by using the GPSMAP 76 unit (Garmin, Olathe, KS, USA).

DNA extraction and real-time PCR are described elsewhere (20). Five 3-mm diameter punches, equivalent to 25 μL of whole blood, were removed from the filter papers. DNA was extracted by using a commercial 96-well kit (Promega, Fitchburg, WI, USA). The DNA was concentrated by glycogen acetate and acetate/ethanol precipitation and low-speed (3,000 × g) centrifugation for 30 min.

Multiplex real-time PCR was used to amplify the 18S P. falciparum ribosomal gene with a Cy5-labeled probe (20). Samples were processed in duplicate on manually loaded 384-well plates. A 40-cycle standard PCR protocol was used in a CFX 384 real-time PCR Detection System Thermocycler (BioRad, Hercules, CA, USA). Baseline relative fluorescence units (RFUs) were readjusted by using Bio-Rad CFX manager software. The real-time PCR system also detected Borrelia spp. reported by Reller et al. (24). The sensitivity and specificity of the real-time PCR versus that of RDT or microscopy were reported by Schachterele et al. (20). The real-time PCR detected 1–100 parasites/μL and showed the highest sensitivity in latent class analysis on febrile study patients. For the monthly surveys, microscopic analysis was not performed for the 8,711 samples for which real-time PCR was conducted. The RFU cutoff used was greater than that for >50 negative blood samples from control patients at Johns Hopkins University. The cutoff of 650 RFUs for the last cycle of the highest real-time PCR replicate had a specificity of 100% for samples from the control group at Johns Hopkins University and a specificity of 94% for samples from febrile patients in Tanzania; microscopy was used as a reference method (20).

Blood samples from treatment and control villages that had higher parasite densities by real-time PCR and microscopy were subjected to PCR amplification and whole gene fragment cloning into the pCR2.1 plasmid. Amplification and cloning were followed by full PfRpL4 gene DNA sequencing of multiple bacterial clones for each patient and examination of mutations anywhere in this gene (1).

Data were analyzed from the perspective of an intention to treat in which participants from AZ MDA–treated villages who refused treatment were classified with those who accepted AZT as members of treatment villages. Univariate ratios were used to compare prevalence proportions in treated versus untreated villages. CIs for ratios that compared proportions of prevalent infection in treated versus untreated villages were estimated by using 2-sided exact binomial tests.

Maps were used to display geospatial patterns in infection prevalence and malaria clustering over time (Google Maps, Mountain View, CA, USA, and Quantum Geographic Information System Open Source Geospatial Foundation Project (www.qgis.org/en/site/). Geographic coordinates of P. falciparum–positive and –negative blood samples were projected into kilometers, and locations were smoothed by using quadratic kernel intensity estimation. Kernel intensity estimators predict spatial intensity at unsampled points from nearby sampled points by using a quadratic function to heavily weight the nearest measured points and de-emphasize points sampled at greater distances. Weighted integrals were used to estimate intensity smoothing parameters for quadratic functions, assuming a 2-dimensional (i.e., spatial), stationary, isotropic point pattern process. Spatial odds ratios (ratio of kernel intensities from positive blood samples to kernel intensities from negative blood samples) (25), were determined for P. falciparum.

To examine the effect of AZT MDA while adjusting for malaria clusters observed on maps, we used multivariate logistic models with random effects to compare odds of prevalent malaria infection for AZT MDA treatment villages with odds for control villages. Random intercepts were fit to account for the multilevel or hierarchical design that nested persons within villages. This model enabled accurate statistical inference by adjusting CIs and p values for P. falciparum clusters in villages, and this inference was apparent in prevalence maps (25). Residual spatial autocorrelation was assessed by using variograms, and residual temporal autocorrelation was eliminated by restricting models to 1 sampling interval (26). Multiple models that examined the effect of AZT MDA on prevalent malaria were fit to control for confounding by other drivers of malaria prevalence and to address the robustness of standard errors to potential spatial residual autocorrelation. Altitude and self-reported bednet ownership were held constant to control for confounding in the multivariate model. A priori, we believed any AZT MDA effect would be most perceptible between month 1 and baseline.

To distinguish between the effect of AZT on prevalence and incident infections, we performed a sensitivity analysis for only P. falciparum infections that had been preceded by a negative P. falciparum test result in the previous sampling interval. These data also underwent a second sensitivity analysis that included participants treated with artemether/lumefantrine who had been removed from the previous analyses.

Data management and analysis were conducted by using R statistical software (27). R with the SPLANCS package was used for kernel intensity estimation (28), and the LME4 package was used for the random effects model (29). All R code is available upon request.

The study site in the Kongwa District of Tanzania included 8 villages and 12,898 persons. Four villages were chosen on the basis of increased prevalence of trachoma (>10%) in children 1–9 years of age. Excluded were 34 of 66 villages that had received AZT MDA in the previous year. In the 4 villages that received AZT MDA, 6,252 persons received AZT and 642 did not receive AZT. A census in untreated control villages identified 5,991 persons. These villages were chosen because of geographic proximity to treatment villages. In each village, pairs of a parent and a child <5 years of age were randomly chosen for inclusion (1,045 persons in the AZT MDA treatment group and 1,008 persons in the control group for follow-up) (Figure 1). Blood samples for real-time PCR testing were collected from participants at baseline and at months 1, 3, 4, and 6 (4,437 patient-time samples from AZT MDA treatment villages and 4,274 patient-time samples from control villages). At any follow-up period after baseline, 1,010 (97%) study participants in AZT MDA villages who received AZT and 976 (97%) study controls were sampled for real-time PCR at least once. At baseline, treatment and control villages reported antimalarial drug use, latrine access, education (based on the highest level of education attained by the father), and home elevations. However, self-reported bednet ownership and fever histories were higher in treatment villages than in control villages (p<0.05) (Table 1).

Overall, the proportion of P. falciparum–infected participants was highest (6%) at baseline. Infections decreased sharply in treated villages and gradually in control villages throughout follow-up period (Figure 2). At the baseline evaluation, 6% (53/854) of participants from AZT MDA treatment villages were positive for P. falciparum compared with 6% (54/894) of participants from AZT MDA control villages. No differences in odds of infection were observed between control and treatment villages (odds ratio 1.03, 95% CI 0.68–1.55) (Figure 3, Table 2).

By month 1, P. falciparum prevalence in AZT MDA treatment villages decreased to 2% (14/851) and P. falciparum prevalence in AZT MDA control villages decreased to 5% (37/779). The odds ratio for AZ MDA treated villages compared with control villages was 0.34 (95% CI 0.17–0.64), which is consistent with a 66% reduction in the odds of P. falciparum infection in AZT MDA treatment participants compared with AZT MDA control participants. Beyond month 1, the association between of AZT MDA and reduced prevalence of malaria infection decreased. However, overall rates of P. falciparum infection decreased below levels for which reliable inferences could be made.

A village level comparison of changes in malaria prevalence for AZT MDA treatment villages versus prevalence in control villages indicated that 1 village in each group had no change in malaria prevalence, and 3 AZ MDA treatment villages had a significant (p<0.05, by Fisher exact test) decrease in malaria prevalence (p = 0.489, p = 0.0002, and p = 0.0005) and 2 in the control group had significant (p = 0.0008 and p = 0.001) increases. Village 5 in the control group had a significant (p = 0.0001) decrease (Figure 4; Technical Appendix Table 1).

In a subgroup of 200 children examined for malaria infection at weeks 2, 6, and 8, the malaria prevalence of 2%–5% was lower than the 10% expected (23) Thus, reliable prevalence comparisons were not relevant. During the study, febrile patients were tested by RDT and microscopy. During the critical period of 5–42 days after AZT MDA, 46 and 48 children in the treatment and control villages, respectively, were evaluated for fever. Among febrile children, only 5 children from treatment villages and 3 children from control villages were positive by real-time PCR.

At month 1, reductions in P. falciparum malaria prevalence were observed by logistic regression models that controlled for differences in self-reported bednet ownership and home altitude (Table 3) and included village level random effects. The odds of P. falciparum infection were 63% (95% CI 28%–81%) but less in treated villages than in control villages after adjusting for bednet ownership and home altitude in a model with village-level random intercepts. Because self-reported history of fever, malaria medication, latrine access, and education of the head of the household were not associated with malaria prevalence and did not appreciably alter the interpretations of the multivariate models, they were not included in the reported estimates. As was evident in univariate analysis, no differences in malaria rates existed at baseline, and the association between AZT MDA and reduced malaria rates decreased after the first month of follow up.

Spatial information on home locations enabled a thorough examination of residual spatial autocorrelation. A total of 51% (19/37) of P. falciparum infections at month 1 occurred in control village 8 (Figure 4), which raised concerns that residual spatial autocorrelation artificially decreased CIs from the reported models despite use of the random effects model. However, a variogram of standardized residuals from the model suggested that residual spatial autocorrelation was appropriately controlled by the village level random effects. The data were reanalyzed with incident infections that had been preceded by a negative real-time PCR result for P. falciparum; results did not change appreciably. We also observed an age-independent decrease in malaria prevalence at month 1 for persons 1–10 years of age (univariate analysis) and for persons >10 years of age (univariate analysis and multivariate analysis) (Technical Appendix Table 2).

Sequencing of full-length P. falciparum ribosomal L4 protein was performed for samples from 12 patients. We did not find evidence of single-nucleotide polymorphisms (SNPs) conferring AZT resistance in amino acid region 71–79 (1). A synonymous SNP at position K36 was found in all bacterial plasmid clones from samples of 1 patient and in a P. falciparum–positive control patient who was not from Tanzania. A single plasmid clone from a sample of participant from an AZT MDA control village contained a nonsynonymous SNP in the active site of the PfRpL4 gene at A78S. However, 17 other plasmid clones from the same blood sample did not contain the mutation, which suggested the aberrant sequence resulted from a Taq polymerase error during plasmid cloning of the PfRpL4 gene. Moreover, restriction enzyme HindIII, which was specific for the nonsynonymous SNP at A78S, did not digest multiple PCR amplicons from the original patient blood sample, which further suggested that the SNP on the PfRpL4 gene was present at a low concentration or may have been an artifact of the Taq polymerase amplification. A78S has not been identified in other mutant bacteria associated with AZT resistance (1). Eighteen other nonsynonymous mutations were found in samples from participants from AZT MDA treatment villages, but these mutations were not in conserved amino acid regions for PfRpL4 associated with AZT resistance in bacteria or the selected P. falciparum (except for an I39S were single occurrences among multiple clones from a single patient). Four synonymous SNPs were also detected (Table 4).