Patterns of individual non-treatment during multiple rounds of mass drug administration for control of soil-transmitted helminths in the TUMIKIA trial, Kenya: a secondary longitudinal analysis

Summary Background Few studies have been done of patterns of treatment during mass drug administration (MDA) to control neglected tropical diseases. We used routinely collected individual-level treatment records that had been collated for the Tuangamize Minyoo Kenya Imarisha Afya (Swahili for Eradicate Worms in Kenya for Better Health [TUMIKIA]) trial, done in coastal Kenya from 2015 to 2017. In this analysis we estimate the extent of and factors associated with the same individuals not being treated over multiple rounds of MDA, which we term systematic non-treatment. Methods We linked the baseline population of the TUMIKIA trial randomly assigned to receive biannual community-wide MDA for soil-transmitted helminthiasis to longitudinal records on receipt of treatment in any of the four treatment rounds of the study. We fitted logistic regression models to estimate the association of non-treatment in a given round with non-treatment in the previous round, controlling for identified predictors of non-treatment. We also used multinomial logistic regression to identify factors associated with part or no treatment versus complete treatment. Findings 36 327 participants were included in our analysis: 16 236 children aged 2–14 years and 20 091 adults aged 15 years or older. The odds of having no treatment recorded was higher if a participant was not treated during the previous round of MDA (adjusted odds ratio [OR] 3·60, 95% CI 3·08–4·20 for children and 5·58, 5·01–6·21 for adults). For children, school attendance and rural residence reduced the odds of receiving part or no treatment, whereas odds were increased by least poor socioeconomic status and living in an urban or periurban household. Women had higher odds than men of receiving part or no treatment. However, when those with pregnancy or childbirth in the previous 2 weeks were excluded, women became more likely to receive complete treatment. Adults aged 20–25 years were the age group with the highest odds of receiving part (OR 1·41, 95% CI 1·22–1·63) or no treatment (OR 1·81, 95% CI 1·53–2·14). Interpretation Non-treatment was associated with specific sociodemographic groups and characteristics and did not occcur at random. This finding has important implications for MDA programme effectiveness, the relevance of which will intensify as disease prevalence decreases and infections become increasingly clustered. Funding Bill & Melinda Gates Foundation, Joint Global Health Trials Scheme of the Medical Research Council, UK Department for International Development, Wellcome Trust, Children's Investment Fund Foundation, and London Centre for Neglected Tropical Diseases.


Supplementary Information
Figure S1 -Flow chart of participants enumerated in the TUMIKIA baseline survey who were included in current analyses. Eligible individuals included those age 2 years or older during the baseline survey not recorded as deceased or moved away during a treatment round. The proportion of eligible individuals included in the analysis was 97.0% (36327/37458).

Reasons for Non-treatment
During MDA, community health volunteers recorded the reason for non-treatment in paper registers using a precoded list of options. Subsequently during digitisation, if no reason was recorded in the register this was digitised as a 'missing response'. For our analyses, we used the individual's age as recorded during the baseline household questionnaire, which did not exactly correspond in some cases to the age recorded in the treatment registers.

Data Censoring
For the main analysis we excluded only individuals recorded in the treatment data as either deceased or migrated in any round. We conducted additional analyses to examine the impact of different assumptions about censoring of individuals based on the availability of recorded treatment information. Prior to generating the non-treatment indicator, we flagged individuals with no linked treatment information for a given round and no linked information for any subsequent rounds. These flags were used to exclude these individuals from analyses to test the assumptions. It is important to clarify that if an individual had no recorded treatment information in a given round but had recorded information from a later round, they were classed as not treated for the earlier round and not censored. The numbers of individuals included in the main and alternate analyses or censored along with frequency of non-treatment are presented in Table S3.
All individuals with recorded information from at least one round were included in the analysis to identify predictors of non-treatment in round one (Table S4). Then, for the estimation of the association of non-treatment with subsequent non-treatment, we censored individuals if they had no treatment register record during rounds two or three and following rounds and no record at round four (Sensitivity 1). However, because there was no subsequent round after round four, this assumption meant all individuals without a record at round four were excluded (rather than being classed as 'not treated'), so we also examined the impact of ignoring this censoring at round four (Sensitivity 2). This latter approach would not affect the selection of round one predictors, only the estimation of the association of non-treatment between rounds.
We also examined the impact of these assumptions on the analysis of factors associated with overall non-treatment by excluding rounds where the individual was considered to be censored from the determination of complete, partial, or no treatment. The results of the alternate analyses are presented below (Tables S5 -S8).
Finally, as described in the main text, we examined the impact of excluding rounds where women of childbearing age were recorded to be ineligible for treatment because of recorded pregnancy, breastfeeding, or having recently given birth. The results of this analysis are presented below (Table S9).  Among 13505 children not censored at round two, previous non-treatment was associated with increased odds of non-treatment at a subsequent round (OR 2·72, 95%CI 1·79, 2·88), controlling for school attendance, head of household non-treatment, and urban household. Controlling for the same factors, but ignoring censoring at round four, among the same 13505 children, previous non-treatment was more weakly associated with increased odds of non-treatment at a subsequent round (OR 1·16, 95%CI 0·99, 1·34).

Predictors of non-treatment during round one and association of non-treatment with subsequent nontreatment
Among 16039 adults not censored at round two, those without recorded treatment during the previous round were 2·6 times more likely to not have treatment reported during the subsequent round compared to those with treatment recorded (OR 2·58, 95%CI 2·34, 2·85), controlling for age and sex. Ignoring censoring at round four, among the same adults, previous non-treatment was more weakly associated with increased odds of non-treatment at a subsequent round (OR 1·80, 95%CI 1·63, 1·98), controlling for age and sex.    Table S7 -Association of baseline individual and household characteristics with partial or no treatment (relative to complete treatment) during uncensored biannual mass drug administration rounds (Sensitivity 1) among 17009 individuals aged 15+ years in Kwale County, Kenya, 2015-2016 Table S9 -Association of baseline individual and household characteristics with partial or no treatment (relative to complete treatment) during two or more rounds of biannual mass drug administration among 20087 individuals aged 15+ years, excluding rounds for women ineligible for reported pregnancy, breastfeeding, or a recent birth, in Kwale County, Kenya, 2015-2016

Multiple testing
To select predictors of non-treatment during round one, we fit all possible subsets and selected those predictors in the model with the lowest Bayesian Information Criteria. For a sample size of 100, optimisation of the BIC corresponds to significance-based selection (i.e. comparison of two hierarchically nested models with a difference of one DF) at a significance level of 0.032. 1 This value is calculated by: ( , )=1− 2, (log( )⋅ ) As an example, our model selection process included data for 16236 children, and we calculate that this corresponds to significance-based selection at a significance level of 0.002.