Evidence of high bednet usage from a list randomization experiment in rural Gambia

Background Recording behaviours that have the potential to impact health can be doubly challenging if the behaviour takes place in private spaces that cannot be observed directly and where there is potential for social desirability bias, i.e. where the respondent may give an answer that they think the recorder wants to hear. Sleeping under a long-lasting insecticidal net (LLIN) is an important intervention for malaria prevention, yet it is difficult to gauge the extent to which coverage (how many nets are in the community) differs from usage (how many people sleep under a net). We employed a novel method, list randomization, which partially obscures respondents' answers to sensitive questions, as a mean to provide an accurate estimate of LLIN usage in The Gambia. Methods We surveyed 196 residents from 196 households recruited into a randomised controlled trial assessing the effect of a housing intervention on malaria. In the current experiment, 98 of the 196 study participants were randomly assigned to the control group and received a four-question list about non-sensitive behaviours; the other participants in the intervention group, received the same list, with the addition of one question on a sensitive behaviour; whether or not they had used a bednet the previous night. Participants were read the list of questions and then said how many of the statements were true. We estimated bednet usage by calculating the difference in means between the total number of affirmative items between the two groups, and quantified uncertainty using a t-test. Results The mean number of affirmative responses in the control group was 2.60 of four statements (95% confidence interval, 95% CI, = 2.50-2.70), compared with 3.68 (95% CI = 3.59-3.78) in the intervention group. Such difference (1.08; 95% CI = 94.9-100%) suggests approximately 100% bednet usage. Conclusions Our findings suggest complete universal bednet usage in the study area. Further validation of the list randomization method in areas with lower net coverage is required.

Scale up of LLINs and prompt and effective case management contributed averting 68% of cases (Bhatt et al. 2015). LLINs protect users by providing a physical barrier to night-time biting mosquitoes and killing them upon contact with the net. At around $2 dollars a net, they are a highly cost-effective intervention (2)(3)(4). LLIN coverage in sub-Saharan Africa has never been higher, with 80% of households having at least one net in 2016, and 43% of households having one or more nets for every two people (5). However, there are concerns about how 'coverage' is measured and whether 80% coverage translates to 80% of the population being protected by an LLIN.
Coverage is defined as the number of nets relative to the number of people in the household, a metric that can be verified by counting the number of nets compared to the number of sleeping places. However, assessing whether an individual sleeps 4 under a net (bed net use) is more difficult than assessing coverage. For example, while a household may still own a LLIN (coverage) its use may diminish after mass distribution campaigns due to product "wear and tear" (6), changes in use by season, social events and other factors. Assessments of LLIN usage often occur immediately after they are distributed, which may result in overestimating usage rates that are likely to decline over time. For example, a recent multi-country study suggested a 50% reduction in usage in the 23 months following LLIN distribution (7).
Perhaps of greater concern is the difficulty of estimating LLIN usage without directly observing people sleeping; it usually relies on questionnaires or observing the presence of a hung net in a house. A recent meta-analysis estimated that selfreported rates of LLIN use were 13.6% greater than directly observed rates (8). The extent of the gap between observed and reported usage is highly variable by country and social group (9) but exists in all studies examined.
The existence of this gap suggests that LLIN usage is potentially a sensitive behaviour. Reporting sleeping under a net the previous night is likely to be biased, since most recipients have been told that the net is protective and have been given the net as a gift. As a matter of politeness, or perhaps even fear of negative repercussions (in the case of government-organized LLIN campaigns), respondents are likely to say they have used the net, even when they have not, so as to avoid giving offence. One method by which social desirability bias can be reduced is list randomization (10,11). In a list randomization experiment, participants are divided into "control" and "treatment" groups. The control group is given a series of yes/no questions about everyday activities (communication, transportation, eating, and work), but instead of answering each question individually, participants simply tally the number of "yes" responses and report that number to the researcher. The experiment group is provided with the same questions, but with an additional question on sensitive behavior. Using list randomization, participants from both groups are able to hide their item-specific responses from the researcher, but the data generated from the process allows for aggregate comparison between the groups, with the difference in total "yes" items approximating the population-level "yes" prevalence.
List randomization has previously been used to reduce data bias pertaining to sensitive topics such as personal finance (12), intimate partner violence (13), illegal migration (14), and attitudes regarding homosexuality and gender (15). To our knowledge, despite its relevance to public health campaigns and its sensitive nature, there are no published reports on list randomization applied to the question of LLIN use. Here we employ list randomization to estimate LLIN use in an area of seasonal malaria transmission in The Gambia.

Study location
The study was carried out in the Upper River Region (URR) of The Gambia (13.

Study design
The study was carried out from the end of November 2017 (after the September peak) to mid-January 2018 (when mosquito density was very scarce).
For the purposes of the list randomization experiment, we classified the "control" and "non-trial" households into one group (i.e., a total of 133 non-intervention houses), since at the time of the list randomization experiment both groups' houses were unimproved. A free LLIN distribution campaign organized by the trial investigators took place in 2016 before the trial commenced and a national campaign carried out by the National Malaria Control Programme (NMCP) distributed free LLIN in July 2017, just before the list randomization experiment Randomization to select participants was stratified by river bank (north vs south), ethnic group (with the purposely selection of one, Jagajari, for being the only Serrehule village in the trial), and size. 91 villages were enrolled in a householdclustered randomized controlled trial assessing the impact of house improvement on malaria outcomes (18). Fifteen of the 91 trial villages were randomly selected for an ancillary longitudinal socio-economic study. A total of 196 household heads were enrolled in this study. Each household head was associated with a house in one of three groups: (1) 63 "intervention" houses that were thatch-roofed houses modified to have metal roofs and ventilated, (2) 65 "control" houses that were unmodified thatch-roof houses and (3) 68 "non-trial" households, which included thatch and metal-roofed houses..
Enrolled households were randomly assigned to one of two "question lists": 97 received the control questionnaire and 99 the experiment questionnaire. The control questionnaire contained four questions about daily activities; the experiment questionnaire contained the same questions, with an additional question asking whether the participant slept under a LLIN the previous night ( Table 1). The order of statements in all questionnaires was randomized at the individual level, thus, every 7 household had its own randomly assigned questionnaire. The maximum number of statement combinations [N*(N-1)] was 12 (4*3) in the control group and 20 (5*4) in the experimental group. Therefore, 12 and 20 different typologies of questionnaires were randomized to the households. The statement in bold represents the "experimental" item.
A trained field worker administered the question list to the household heads.
Participants were asked not address individual statements, but instead to count on their fingers (held behind their backs, so as to block the view of the interviewer) the number of statements which were true for them. The reported number (0 to 4 for the control group, 0 to 5 for the experimental group) was recorded.

Statistical analysis
The sample size of 196 participants was calculated to be sufficient for the calculation of a 95% confidence interval on a LLIN usage point estimate of 85%. We with the experimental statement. We tested the hypothesis that the two groups had the same means using an unequal variances t-test. We assume that the introduction of the experimental statement has no effect on the answers to the nonexperimental statements (i.e. the "no design effect assumption") and do not require that the agreements with non-experimental statements be true, but rather that they be equally truthful between groups. The mean difference between the two groups provides an estimate of the proportion of bednet use in the study population.
Since true bednet usage could not be directly observed, we exploited data gathered on self-reported bednet use among children residing in houses with our study participants, performing a robustness check of our results by assessing the correlation between child bednet use (via trial survey response) and adult bednet use (via list randomization). For this group, we compared, at the aggregate level, the association between household children sleeping under a bednet (per selfreport) and the adult head of household sleeping under a bednet (per list randomization), with the assumption that a high level of correlation suggests a high level of list randomization reliability (since it is reasonable to assume that intrahousehold adult and child net usage is likely correlated).

Ethics
Participants were enrolled into the study provided they gave their full and informed consent.  (Table 2).  Though, for feasibility reasons, we did not carry out a direct estimation to assess the validity of our results as some other studies have done (19), the high figures suggest high LLIN usage, at least during the period in which the study took place (towards the end of the transmission season, when coverage is likely to be highest).
In fact, the near 100% coverage found in this study was significantly higher than previous studies (16). Importantly, among our 196 participants, none stated that they did the minimum (zero) or maximum (four or five, depending on the group) number of activities from our list. There are two possible explanations for this finding. First, it is possible that our list items were poorly chosen and did not provoke sufficient heterogeneity in responses. Second, the list randomization method, though meant to anonymize item-specific responses, may not work in certain contexts. We believe that the second explanation is more likely. The study was carried out in an area where a great deal of health research had been carried out previously, and the field worker was known to be an employee of a health research facility. One potential explanation is that some study participants may have inflated the number of affirmative responses to reduce suspicion by the fieldworker that they were sleeping without a mosquito net. In other words, social desirability bias may have occurred in our study, despite our best efforts to avoid it. Further implementation research is required on both (1) the correct selection of list items, and (2) how the context (who the fieldworker is, etc.) may affect results.
Our approach mirrors that of a recent analysis of sexual behaviours (19) and intimate partner violence (13). However, there are number of limitations to this study. Firstly, it was carried out in an area where considerable research has taken place, and therefore may not be generalizable to the population at large, particularly areas with less research activity. The Medical Research Council Unit in The Gambia have been active in the study area for many years, which might affect knowledge and practices around LLINs, as well as social desirability bias. Our fidelity test between whether children were reported to sleep under a bednet and whether an adult reported (via list randomization) to do the same lacks obvious comparability due to the likely differences in behaviour by age and relatively low number of households where children's status could be identified. Additionally, we 13 could not validate the reliability of either method (self-reporting or list randomization) through direct observation, due to privacy concerns. Though a movement logger could theoretically be used as a validation tool, the awareness of the logger itself might also bias results, since one can assume that an individual is more likely to use an LLIN if they know their use is being directly monitored.
Finally, though our sample size was sufficient for an overall assessment of bednet coverage based on a two-group comparison, we did not have sufficient statistical power to identify the potential determinants of bednet use, such as gender, ethnicity, age, or socioeconomic status. Nonetheless the list randomization test offers a promising new method for exploring LLIN use in study communities.
Our list randomization experiment indicates near universal bednet use by the adult population of eastern Gambia, in the year following an LLIN distribution campaign.
Though, for feasibility reasons, we did not carry out a direct estimation to assess the validity of our results as some other studies have done (19), the high figures suggest high LLIN usage, at least during the period in which the study took place (towards the end of the transmission season, when coverage is likely to be highest).
In fact, the near 100% coverage found in this study was significantly higher than previous studies (16 The Gambia have been active in the study area for many years, which might affect knowledge and practices around LLINs, as well as social desirability bias. Our fidelity test between whether children were reported to sleep under a bednet and whether an adult reported (via list randomization) to do the same lacks obvious comparability due to the likely differences in behaviour by age and relatively low number of households where children's status could be identified. Additionally, we could not validate the reliability of either method (self-reporting or list randomization) through direct observation, due to privacy concerns. Though a movement logger could theoretically be used as a validation tool, the awareness of the logger itself might also bias results, since one can assume that an individual is more likely to use an LLIN if they know their use is being directly monitored.
Finally, though our sample size was sufficient for an overall assessment of bednet coverage based on a two-group comparison, we did not have sufficient statistical power to identify the potential determinants of bednet use, such as gender, ethnicity, age, or socioeconomic status. Nonetheless the list randomization test offers a promising new method for exploring LLIN use in study communities. declarations

Ethics approval and consent to participate
The study was approved by the Gambia Government and Medical Research Council's joint ethics committee. All participants gave informed consent.

Consent for publication
Not applicable.

Availability of data and materials
The datasets generated and analysed during this study are not publicly available since they include identifiable protected health information of a sensitive nature.
Researchers interested in accessing the data can contact any of the authors, who will put them in touch with the Medical Research Council's ethics committee, and will facilitate access to raw data in the case of ethical approval being obtained