Experimental Design
The intervention will be evaluated through a randomized control trial with 76,324 participants, stratified by gender and neighborhood. Participants were randomly selected from the e-government platform (SP156) database that initially contained 755,585 records, and assigned to one of six arms: five treatment arms and one control arm in which participants do not receive any messages. After the text messages were sent, citizens will be contacted on their cellphones to answer a survey designed to measure beliefs, recall, awareness, and self-reported behavior related to the epidemic.
Given the experiment includes five treatment arms that will be compared with the control group, to maximize statistical power, the number of participants assigned to the control was 2.24 (square root of 5) times larger than each treatment group. As a result, the division of the total number of phone number records in our database was the following: (i) control: 233,488, (ii) treatment arm 1: 104,419, (iii) treatment arm 2: 104,420, (iv) treatment arm 3: 104,419, (v) treatment arm 4: 104,420, (iv) treatment arm 5: 104,419.
Because of limited resources, we only sent text messages to a representative sample of each arm. Considering our power calculations, we aimed to have 2,572 post-treatment completed interviews from the control group and 2,538 from each treatment arm. And given the telephone center provided us with an estimate of 20% response rate for the phone calls, we sent text messages to a sample five times bigger than the aimed number of interviews: around 12,860 in the control group, and around 12,690 in each treatment arm. The intervention was delivered to 63,465 records between June 2 and June 8, and 12,859 records were kept in the control group. Due to budget constraints, we were unable to send a placebo SMS to a group of people or to do a pre-treatment survey.
Our strategy to assess the impact of the messages is based on a telephone survey that will be conducted by the SP156 telephone center. The survey instrument (questionnaire) included a small number of items and was designed to be completed in only 6 minutes. The questionnaire included four batteries of questions that measure: (i) beliefs, (ii) self-reported behavior, (iii) comprehension/ recall, and (iv) sociodemographic variables. The first three sections comprise the study outcomes. As such, we will combine items per section to create additive indices, and then evaluate the impact of the messages on each of the first three dimensions (outcomes). The sociodemographic information collected in the last part of the instrument will be used as controls.
We will estimate an intention-to-treat (ITT) effect. We are solely estimating the ITT, and not the Average Treatment Effect (ATE) because we cannot know the compliance rates, that is, whether the person opened the text messages, only if one has received it. We will conduct a one-tailed test, with α = 5%, β = 80%, and Dunnet’s correction for multiple comparisons (family-wise error). With this set, we have two percentage points of minimum detectable effect.