Designing schemes for improved water quality

The RCT tests the effects of different payment methods on farm- and scheme- level outcomes. An individual farmer can modify their behaviour in a number of ways to reduce the flow of phosphorus from the land into the water course. These are:
• Low inputs on permanent grassland
• No P fertiliser on high P-index soils
• Loosening compacted soils

In each catchment, the control group is assigned to a fixed price scheme which offers each farmer the choice to change their behaviour to include at least one of the three practices on at least some portion of at least one field. If the target for phosphorus reduction is more than met, the scheme allocates funding to the offers which offer best value per predicted phosphorus saving. In catchment 1, the treatment group is instead assigned to a sealed-bid pay-as-bid auction. Each bid is ranked on by cost per predicted phosphorus saving, and bids are accepted from least costly to most expensive until the target is just met. All winning bids are then paid at the amount that the bid specified.

In catchment 2, the control treatment is the same. However, the treatment is different. Instead of a Pay-as-bid auction, the fixed price control is compared to a uniform price auction. The uniform price auction is run as a descending auction, with three prices announced sequentially if more than the targeted phosphorus saving offered by farmers exceeds the target.

2019-11-15

2020-01-31

The primary variable of interest is the predicted level of phosphorus abatement that could be purchased by the regulator from each farmer. For the auctions, this is the level of abatement offered at the reserve prices (opening offers in the descending price auction); for the fixed price schemes it is the amount that is offered. Note that if rationing occurs in the scheme, then the amount that could be purchased may well be greater than the amount that actually is purchased.

The predicted level of phosphate is calculated by the funder and varies according to field size and which activity is chosen. They are calculated in the same way independent of treatment.

Participation – we are interested in the probability that a farmer offers any abatement in each scheme. For the auctions, this is restricted to those farmers who offer abatement at no more than the reserve prices.

Participation of each farmer is a binary variable indicating whether that farmer offered any abatement under the fixed price, and in the case of the auctions if they offered any abatement with a bid below the reserve prices.

Our experiment uses a between subject design, randomly assigning a farmer to either a fixed price scheme or an auction. Random assignment happens at the level of clustered neighbouring farms. We observe outcomes at the farm level. We use a clustered approach to limit the degree of information spill-over between neighbouring farms.

The funder was concerned with the difficulty in implementing a scheme which randomised assignment at the level of the individual farmer. There was also a concern that doing so would increase the possibility of information spill-overs between neighbouring farmers. Additionally, there are a number of constraints that the funders placed on the randomisation process. Specifically, they wanted roughly equal: representation across sub-catchments, number of farms, and distribution of farm size. For the Catchment 2, which neighbours the 1, they had two further considerations. 1) They wanted to ensure that there was equal representation across farm type (arable versus livestock; almost all famers are livestock farmers in Catchment 1). 2) For the 20 farmers in Catchment 2 (out of 368) we had to ensure that treatment assignment matched that in Catchment 1.
The procedure that we implemented therefore accounts for these concerns. The algorithm first clusters farms within sub-catchments using a k-means clustering process, then randomly assigns these clusters to each treatment. The final step selects the randomisation which gave best balance across farm number and distribution of farm size. For Catchment 2, the final step was modified to also balance on agricultural activity, and constrained for farms with land in Catchment 1.

Clusters of neighbouring farmers, as described in “Randomisation method” section.

Yes

26 clusters in Catchment 1
24 clusters in Catchment 2

400 farmers in Catchment 1 368 farmers in Catchment 2

Catchment 1: 13 clusters, 197 farms in treatment 1; 13 clusters, 203 farms in treatment 2
Catchment 1: 12 clusters, 186 farms in treatment 1; 12 clusters, 182 farms in treatment 2