Primary Outcomes (end points)
The primary outcomes for this analysis are: 1) garbage weight, 2) recycling weight, 3) quality of garbage in the cart, and 4) quality of the recycling in the cart. We are interested in these outcomes as measured directly by researchers in the field and by the “Zabble Zero-Waste” app.
We will estimate the effects of the “Information Treatment” and the “Incentive Treatment” using a difference-in-differences design such as:
Y_it=β_0+β_1 T_it^1+β_2 T_it^2+δ_t+λ_i+ε_it
where Y_it is the outcome of interest for household i in time t. In this equation, T_it^1 and T_it^2 are indicators for household i being assigned to treatment 1 or 2, respectively, in a period t after the intervention begins. A time fixed effect (δ_t) for each week controls for waste changes attributable to common shocks such as holidays and other seasonal trends, and household fixed effects (λ_i) control for unobservable, time-invariant determinants of a household’s waste behavior.
To test if the treatment effects are statistically distinguishable from one another, we will perform an F-test of the hypothesis that T_it^1=T_it^2.
We will also test whether treatment households are more likely to ask for a recycling cart than control households during the study period—i.e. was there an increase in the number of households participating in recycling. To see which part of the total treatment effect is driven by the intensive versus extensive margin, we will add two interaction terms between each treatment indicator and an indicator for having no cart during the 6-week baseline period to our differences-in-differences model above. The coefficients on these terms will estimate the differential treatment effect for those that initially had no cart. That increase in recycling would be driven by the extensive margin.
In addition, we are interested in the variance of the directly collected measures of waste quality and quantity versus those collected by Zabble, and how well the two data collection methods correlate with one another. This analysis will allow us to make determinations about which measure would be higher quality and yield more statistical power for the full study.
Similarly, we will measure the variance of, and correlation between, the directly collected demographic data versus the data purchased from Data Axle, to determine if the Data Axle information is high enough quality to use in our full study.
The data from the focus group will be qualitative in nature and will inform how we design our informational and outreach materials for the full study.
We will test whether treatment effects are heterogeneous along the following dimensions, using data purchased from Data Axle:
a. political affiliation
b. age of head of household
c. educational attainment
d. household income
e. family size
f. race or ethnic group
We will also attempt to detect heterogeneous treatment effects across other dimensions by implementing the causal forests procedure described by Wagner and Athey (2018).