Are Nudges Effective to Contain COVID-19?  A Randomized Control Trial in Rural Bangladesh

We have a single control group, and two treatment groups; and all villages in our household panel are divided into these groups (see Sections ‘Experimental Design’ and ‘Experiment Characteristics’ below for details).

Treatment households receive information and monetary incentives to adhere to social distancing measures and to prevent infection with the Coronavirus and further spread of COVID-19.

2020-04-28

2020-05-06

In addition to primary outcomes gathered in two surveys, we aim to obtain mobile phone mobility data.
Mobile phone mobility data will be used on a daily basis, ideally from a point prior to the intervention, from the time of the intervention until two weeks after the intervention, and from the time of the intervention until three months after the intervention (or once the epidemic is over). However, at this point in time, we cannot be sure that the data we might obtain spans this time frame.

We consider the same primary outcomes from surveys at two points in time: The first collection will be two weeks after the intervention. The second collection will be three months after the intervention, or once the epidemic is over.
Primary outcomes will be collected from households in control villages, and from treated households in treatment villages, and additionally from their untreated neighbors in our sample. Moreover, a village survey might be conducted to inform about the situation in the village.

To assess success of our treatment and behavior of the treated households, we consider the following information: Knowledge and beliefs about the Coronavirus and COVID-19, compliance with preventative measures (e.g. social distancing) as collected by survey questions asked to households themselves and their neighbors, and as measured by mobile phone tracking.
To assess behavior of the untreated households (in treatment and control villages), we consider the following information: Knowledge and beliefs about COVID-19, compliance with preventative measures (e.g. social distancing) as collected by survey questions asked to households themselves and their neighbors, and as measured by mobile phone tracking.
To assess the overall situation of the pandemic in a household and a village, we consider the following information: Infections and symptoms, deaths.

Symptoms in the last two weeks will be measured for the respondent of the survey and the household (averaging over individuals) by the number of symptoms from a list of 12 symptoms (cont. measure); a classification into high risk symptoms ([a] having fever, cough, or shortness of breath, or [b] any other 2 symptoms combined) will be performed for the individual (binary measure) and the household (averaging over individuals).
Similarly, we measure diagnosis, staying in quarantine and death for the respondent (binary measure) and for the household, averaging over binary measures (cont. measure; a proportion), with exception of death, where we have a binary measure for the household.
All this information will be aggregated to an individual and household health scale, by standardizing above measures, and possibly weighting them (e.g. from a PCA on the whole sample or the village), for building an average.
On the village level, we measure the number of people showing symptoms, number of people/households in self-quarantine, the number of people/households being diagnosed with Covid-19, and the number of deaths related with Covid-19.

Adherence to preventative measures in the last two weeks will be measured for the respondent by aggregation of the days in which they self-reportedly adhered to 8 measures from a list. Aggregation is performed by a) unweighted mean b) weighted mean, using weights e.g. from a PCA on the whole sample or the village; always accounting for reversed items (cont. measure). Adherence to social distancing measures will be measured for the respondent by aggregation of the numbers they state corresponding to 11 social distancing measures. Aggregation is performed by taking the maximum (cont. measure). Adherence to social distancing measures in the recent past will be measured for the respondent by aggregation of the numbers they state corresponding to the 3 social distancing measures. Aggregation is performed by a) unweighted mean b) weighted mean, using weights e.g. from a PCA on the whole sample or the village; always accounting for reversed items (cont. measure). Reduction of social contact and close social contact with neighbors will be measured by aggregation of the self-reported numbers they state corresponding to 6 typical situations of social contact, but also by the corresponding numbers neighbors state (on average) about the respective household. Aggregation over the situations is performed by a) unweighted mean b) weighted mean, using weights e.g. from a PCA on the whole sample or the village; always accounting for reversed items (cont. measure). Moreover, from the mobile phone location data, we will measure the reduction of daily movement.
All these measures capturing behavior will be aggregated to a behavior scale, by standardizing above measures, and possibly weighting them (e.g. from a PCA on the whole sample or the village), for building an average. Depending on the data quality and the percentage of households, for which we can gather information from their neighbors, we might have to eliminate information from neighbors from the so computed main outcome on adherence to preventative measures/behavioral reaction to the situation.

Correct beliefs as to whether or not everybody can make a change will be measured for the respondent by checking whether this answer is among the given answers, when asking “Who can make a change in fighting the novel Coronavirus?” Beliefs with respect to the degree of influence individuals have to fight the Coronavirus will be measured for the respondent by their answer to a five point scale. Knowledge about Coronavirus will be measured for the respondent, by the number of correct answers when going through a list with 12 possible answers, 5 of them correct (cont. measure).
Beliefs and knowledge will be aggregated to a beliefs and knowledge scale, by standardizing above measures, and possibly weighting them (e.g. from a PCA on the whole sample or the village), for building an average.

All of the above measures will be enriched with measures on centrality and distance between households, to perform network and spatial analysis.

Higher Order Risk Preferences as measured by the method of Schneider, Ibanez and Riener (2020), using three certainty equivalents (for the normalized utilities of .25, .5, and .75) via the staircase method/bisection approach with 3 questions per certainty equivalent. Additionally, from previous data collection, we have measures on risk preferences, time preferences, and social preferences.

Higher Order Risk Preferences, in particular prudence, have been found to explain health behavior (Schneider & Sutter, 2020), which empirically regularly fails to be explained by risk aversion. In order to investigate the relation of higher order risk preferences and health behavior in an epidemic, in particular their interplay with our treatments, we elicit higher order risk preferences.

We have two treatments and a single control group, consisting of 45, 45, and 60 villages, respectively (for details, see section “Intervention” above).

For assignment of villages to treatment and control groups, we use the minMSE method by Schneider & Schlather (2020), which is a re-randomization method. Treatment assignment for villages was performed in an office on a computer using the ‘minMSE’ R package (Schneider & Baldini, 2019) with 1000 iterations and default parameters except for initial temperature (set at t0 = -10 for 10 percentage of initial value). The pre-treatment information considered for treatment assignment were collected in a village survey (see ‘Supporting Documents & Materials’ for details). We produced 1000 alternative treatment vectors to ensure treatment assignment was not deterministic, and selected one randomly. For assignment of individuals within treatment villages, we first built pairs of households that live close to each other, and then randomly assigned one to treatment and the other to the within-village control group using the ‘nbpMatching’ R package (Beck C, Lu B, Greevy R., 2015) following Lu et al. (2011).

Two-step randomization: First, villages are assigned to treatment and control groups. Second, households within the treatment villages are selected randomly for treatment from our sample of households.

Yes

We plan to use all 150 villages in our household panel.

On average, we plan to treat between 10 and 13 households per village, while further 10 to 13 households per village will serve as within village control households; we aim at the higher number for higher power, which, however, might not be feasible for capacity constraints as timely implementation and a short intervention period is of highest importance in this project. Thus, on average we address 20 to 26 households per village in 90 treatment villages, and 20 to 26 households in our 60 control villages. Our total sample size will thus consist of 3000 to 3900 households.

1. Information Treatment: 45 villages with 10-13 households (treated) and 10-13 households untreated: 900-1170 households
2. Information Treatment plus monetary incentive: 45 villages with 10-13 households (treated) and 10-13 households untreated: 900-1170 households
3. Control Treatment: 60 villages with 20-26 households: 1200-1560 households