Experimental Design Details
The experiment comprises 29,878 households within Atlanta neighborhoods where Relay bike share hubs were first installed during the Relay Bike Share scheme expansion in April 2017. A demonstration phase launched in June 2016 made 100 bikes available at 22 stations. The formal program launch in April 2017 saw an expansion of facilities to 500 bikes at 65 stations. In alignment with the SUTVA assumption of non-interference, we assume that individuals who receive postcards do not discuss the interventions with their neighbors. However, we will allow for within-household spillovers by studying instances in which two or more individuals sign up within the same household (in cases where we observe and track more than one sign-up at a given address).
We use address data from public tax records to perform the randomization. As such, apartment unit-level randomization is only possible for units that are owner-occupied. For all apartments where apartment units are not identifiable in tax record data, we randomize at the apartment level so that all individuals within a given apartment block received the same mailer. To ensure that we achieve balance on location and house type (e.g., single-family, duplex, multi-unit of various sizes) in the allocation of households to the various study groups, we conduct a stratified randomization, blocking first on the neighborhood in which the household resides and second on a categorical variable that captures the number of livable units at the given address (for multi-unit addresses that are not owner-occupied). A vast majority of addresses (87.9%) are identified as single-unit, and another 7.5% are identified as duplexes; these constitute the first two levels of the livable units categorical variable. The remaining categories comprise 3-4 units (1.9%), 5-8 units (1.0%), 9-20 units (1.1%), 21-50 units (0.2%), 51-100 units (0.1%), 101-200 units (0.1%), 201-300 units (0.1%), 300-500 units (0.1%). Four apartment blocks that were over 500 units were excluded from the experiment for cost purposes, since large clusters provide less benefit in terms of power. We exclude three Atlanta neighborhoods - Midtown, Downtown, and Castlebury Hill - since these neighborhoods had already been exposed to the bike share program during the demonstration phase of the Relay Bike Share scheme launched in June of 2016, when 100 bikes and 22 stations were established in these neighborhoods. We additionally exclude two university neighborhoods where the ability to randomize at the household level was limited and risk of spillovers is likely high. A key component of the analysis is the ability to match billing addresses with mailer recipient addresses, especially for the control group, which does not receive a promo code to allow for tracking. Since many students use their permanent hometown addresses as their billing addresses, there is strong potential for inability to match data outcomes to interventions received.
Mailers were sent to all households in the study on the same day in 2017 and recipients have just over a month to redeem their promo codes before the promotion expires. In the exploratory analysis, we will investigate whether there is persistence in participation levels and any identified treatment effects on the intensive margin, or if individuals merely sign up, use the promotion, and then discontinue use.