CAPÍTULO 2: SOLUCIÓN DE SOFTWARE PROPUESTA
2.5. Descripción textual de los casos de uso del sistema
Next I estimate the overall average treatment effect (ATE) of providing price informa-tion on monthly utility bills using a difference-in-differences strategy, where households served by the low information utility before the merger are in the treatment group,
house-holds served by the high information utility are in the control group, and treatment occurs when the merger takes place. The standard difference-in-differences approach regresses the log of average daily water consumption for household i in billing period t on a set of household fixed effects, month-by-year fixed effects, and an indicator for low information households after the merger as follows:
ln(Yit) = β 1Low Info, Post+ αi+ λt + Witγ + εit (1.7)
where Yit is monthly water consumption divided by the days on the water bill for house-hold i in billing period t, 1Low Info, Postis an indicator variable equal to one for low informa-tion households in billing periods after the merger and zero otherwise, αi are household fixed effects, λt are a full set of month-by-year fixed effects, and Wit are a vector of weather controls including average temperature and precipitation in inches, and εit is the idiosyncratic error term. I allow for robust standard errors, which I cluster at the meter reading route level.15
In addition to the standard exogeneity assumption, this approach assumes that low in-formation and high inin-formation households have parallel trends in average consumption.
The merger causes a parallel shift in consumption trends, which allows for identification of the ATE, β, as the difference between consumption trends after the merger relative to the difference before the merger. There are several concerns for this identification strategy, which include differential trends between low information and high information households and other confounding policies. I discuss these issues in detail and enhance my empirical strategy to address these concerns.
15Previous studies have used meter reading route fixed effects to control for unobserved neighborhood-specific characteristics that may influence water consumption (Ferraro, Miranda and Price, 2011). Stan-dard errors are also likely to be correlated within meter reading routes. Some possible sources of corre-lation could include spatial variation in water use due to regional weather patterns as well as economic shocks to consumers residing in the same housing developments. There are 95 meter reading routes, which ensures that there are a sufficiently high number of clusters.
Figure 1.8: Differences in Pretreatment Seasonality and Trends
−2−1.5−1−.50Log Daily Consumption Jan2010 Jul2010 Jan2011 Jul2011 Jan2012 Jul2012 Jan2013 Jul2013 Jan2014 Jul2014
Low Info High Info
Difference in Seasonality: Treatment vs. Control
(a) Average Daily Water Use: Control vs. Treatment
0.02.04.06.08Log Daily Consumption Jan2010 Jul2010 Jan2011 Jul2011 Jan2012 Jul2012 Jan2013 Jul2013 Jan2014 Jul2014
Low Info High Info
Difference in Linear Trends
(b) Differences in Estimated Linear Trends: Control vs. Treatment
Note: These figures demonstrate differences in pre-treatment trend and seasonality between low and high information households. Figure (a) plots average daily consumption (monthly consumption divided by the number of days on the water bill) for both groups. Figure (b) plots estimated linear time trends for low and high information households from a model that regresses the log of consumption divided by the number of days on the water bill on separate linear time trends, separate month fixed effects, weather controls, and household fixed effects. The sample is limited to pre-merger time periods. Robust standard errors are clustered at the meter reading route level.
There are indeed differences in pre-treatment trends as well as seasonality between the low information and high information households. Figure 1.8 (a) shows average monthly consumption divided by the days on the water bill for low information and high information households, grouped by billing months over the pre-merger period. Although both utilities have highly seasonal water consumption patterns, there are differences in seasonal water consumption between the two utilities. Average consumption is much higher for low information households during the summer and exhibits more dramatic peaks in the height of the irrigation season. This motivates the addition of separate month fixed effects for the low information households. Next, I allow for linear differences in trend between treatment and control. Figure A.2 in Appendix A estimates a more flexible difference in trends over the pre-treatment period and indicates that a linear difference in trends is probably appropriate. Figure 1.8 (b) shows the estimated linear trends for each utility over the pre-merger periods. Consumption for the High Information group has a small, but insignificant increasing trend over this period. By contrast, low information consumption has a significant increasing linear trend. Given these significant differences, a model that does not allow for separate linear trend and seasonality would understate the impact of the merger on water consumption.
The identification assumption in a difference-in-differences model that allows for a separate linear trend for the low information households is that low information and high information households have parallel growth in average consumption rather than parallel paths of average consumption. Furthermore, the identification assumptions could be further relaxed to allow for differences in seasonality between low and high information households that is fixed over time. The relaxed difference-in-differences model is as
follows:
ln(Yit) = β 1Low Info, Post+ αi+ λt+ µ 1Low Info∗ t +
12
X
s=1
νs1Low Info,Month s+ Witγ + εit (1.8) where 1Low Info ∗ t is an indicator for low information households interacted with a linear time trend and 1Low Info,Month s is an indicator equal to one for low information households if their billing month occurs during the sth calendar month. Conditional on these relaxed assumptions β is the causal effect of providing price information to low information households on monthly utility bills.