Pricing People into the Market: Targeting through Mechanism Design

Transcription

1 Pricing People into the Market: Targeting through Mechanism Design Terence Johnson (Notre Dame), Molly Lipscomb (University of Virginia) March 15, 2018 Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

2 Market Intermediation as a Policy solution Many key markets in developing countries (health, water, sanitation, education) suffer from imperfections (externalities, search frictions, adverse selection) and arrive at inefficient outcomes One potential policy solution is for governments to act directly as the intermediary between supply and demand, subject to budget constraints. This poses a new set of challenges: How to ensure subsidy dollars reach households who both require and will actually use the assistance? How to compete against/alongside existing markets to deliver services? How to reduce procurement costs to maximize the impact of the budget? We take a market design approach: measure the fundamentals necessary to determine optimal policy, solve for and implement the optimal indirect mechanism, and then test the solution s impact Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

3 Market Intermediation as a Policy solution Many key markets in developing countries (health, water, sanitation, education) suffer from imperfections (externalities, search frictions, adverse selection) and arrive at inefficient outcomes One potential policy solution is for governments to act directly as the intermediary between supply and demand, subject to budget constraints. This poses a new set of challenges: How to ensure subsidy dollars reach households who both require and will actually use the assistance? How to compete against/alongside existing markets to deliver services? How to reduce procurement costs to maximize the impact of the budget? We take a market design approach: measure the fundamentals necessary to determine optimal policy, solve for and implement the optimal indirect mechanism, and then test the solution s impact Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

4 Market Intermediation as a Policy solution Many key markets in developing countries (health, water, sanitation, education) suffer from imperfections (externalities, search frictions, adverse selection) and arrive at inefficient outcomes One potential policy solution is for governments to act directly as the intermediary between supply and demand, subject to budget constraints. This poses a new set of challenges: How to ensure subsidy dollars reach households who both require and will actually use the assistance? How to compete against/alongside existing markets to deliver services? How to reduce procurement costs to maximize the impact of the budget? We take a market design approach: measure the fundamentals necessary to determine optimal policy, solve for and implement the optimal indirect mechanism, and then test the solution s impact Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

5 Market Intermediation as a Policy solution Many key markets in developing countries (health, water, sanitation, education) suffer from imperfections (externalities, search frictions, adverse selection) and arrive at inefficient outcomes One potential policy solution is for governments to act directly as the intermediary between supply and demand, subject to budget constraints. This poses a new set of challenges: How to ensure subsidy dollars reach households who both require and will actually use the assistance? How to compete against/alongside existing markets to deliver services? How to reduce procurement costs to maximize the impact of the budget? We take a market design approach: measure the fundamentals necessary to determine optimal policy, solve for and implement the optimal indirect mechanism, and then test the solution s impact Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

6 Market Intermediation as a Policy solution Many key markets in developing countries (health, water, sanitation, education) suffer from imperfections (externalities, search frictions, adverse selection) and arrive at inefficient outcomes One potential policy solution is for governments to act directly as the intermediary between supply and demand, subject to budget constraints. This poses a new set of challenges: How to ensure subsidy dollars reach households who both require and will actually use the assistance? How to compete against/alongside existing markets to deliver services? How to reduce procurement costs to maximize the impact of the budget? We take a market design approach: measure the fundamentals necessary to determine optimal policy, solve for and implement the optimal indirect mechanism, and then test the solution s impact Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

7 Targeting Demand Many households who receive subsidies or subsidized goods may not use them, and many who receive subsidies might have purchased anyway This may be resolved through screening by imposing small effort costs (Dupas et al 2016, Alatas et al 2016) Large subsidies may be necessary in order to get the poorest to take up (Kremer and Miguel (2007); Cohen and Dupas (2010); Dupas (2014)) But working with the supply side can stretch the budget further (Banerjee, Hanna, Kyle, Olken and Sumarto (2016)) Using covariates to help assign treatment can increase overall take-up (Bhattacharya and Dupas, 2012) Our solution: use procurement auctions to reduce costs and target poor households with the largest subsidies, but invite relatively wealthy people to participate, charge them more, and relax the budget constraint Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

8 Mechanical versus manual desludging Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

9 Mechanical versus Manual Desludging Substantial externalities: lack of adequate sanitation associated with 10% of diarrheal disease (Mara et al 2010) There is no centralized market: households search for desludgers, starting just before the pit fills, and the difficulty of finding a desludger creates price power Median household takes up to 12 days to find a desludger, 20% report taking more than 10 days to find a desludger once their pit is already full. Substantial heterogeneity in frequency with which pits need to be desludged: 10th to 90th percentile is 3 months to 7 years, median is 12 months Households using manual typically prefer mechanical, and often switch between them Trucks are typically not capacity constrained: they say in surveys it rarely happens that they turn down jobs due to being busy Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

10 Project Summary Stage 1 (December 2014): Collect market data on 2088 households most recent desludgings and use a generalization of the second-price auction to elicit their willingness-to-pay for a desludging from an intermediary; use similar auctions to elicit firms costs Stage 2 (May 2015): Design an incentive compatible pricing rule that maximizes use of mechanical desludging subject to a budget constraint allowing $3.00 per household (about 10% of the cost of a mechanical desludging) Stage 3 (July 2015 December 2016): Survey a new set of 2944 households and make take-it-or-leave-it offers based on the optimal pricing rule from Stage 2, then measure impacts on manual utilization, the market share of mechanical, and health indicators relative to a control group Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

11 Project Summary Stage 1 (December 2014): Collect market data on 2088 households most recent desludgings and use a generalization of the second-price auction to elicit their willingness-to-pay for a desludging from an intermediary; use similar auctions to elicit firms costs Stage 2 (May 2015): Design an incentive compatible pricing rule that maximizes use of mechanical desludging subject to a budget constraint allowing $3.00 per household (about 10% of the cost of a mechanical desludging) Stage 3 (July 2015 December 2016): Survey a new set of 2944 households and make take-it-or-leave-it offers based on the optimal pricing rule from Stage 2, then measure impacts on manual utilization, the market share of mechanical, and health indicators relative to a control group Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

12 Project Summary Stage 1 (December 2014): Collect market data on 2088 households most recent desludgings and use a generalization of the second-price auction to elicit their willingness-to-pay for a desludging from an intermediary; use similar auctions to elicit firms costs Stage 2 (May 2015): Design an incentive compatible pricing rule that maximizes use of mechanical desludging subject to a budget constraint allowing $3.00 per household (about 10% of the cost of a mechanical desludging) Stage 3 (July 2015 December 2016): Survey a new set of 2944 households and make take-it-or-leave-it offers based on the optimal pricing rule from Stage 2, then measure impacts on manual utilization, the market share of mechanical, and health indicators relative to a control group Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

13 Preview of Results The market share of mechanical desludging in treatment clusters exhibits a 5.1 percentage point increase overall and a 9.4 percentage point increase among the most heavily subsidized households, relative to a control group Manual desludging exhibits a 4 percentage point decrease among the most heavily subsidized households, relative to a control group Diarrhea incidence among children exhibits a 6.1 percentage point decrease in the most heavily subsidized households, relative to a control group Due to reductions in total cost from adjusting procurement strategies, we have a realized loss of a few dollars, after accounting for subsidies Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

14 Preview of Results The market share of mechanical desludging in treatment clusters exhibits a 5.1 percentage point increase overall and a 9.4 percentage point increase among the most heavily subsidized households, relative to a control group Manual desludging exhibits a 4 percentage point decrease among the most heavily subsidized households, relative to a control group Diarrhea incidence among children exhibits a 6.1 percentage point decrease in the most heavily subsidized households, relative to a control group Due to reductions in total cost from adjusting procurement strategies, we have a realized loss of a few dollars, after accounting for subsidies Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

15 Preview of Results The market share of mechanical desludging in treatment clusters exhibits a 5.1 percentage point increase overall and a 9.4 percentage point increase among the most heavily subsidized households, relative to a control group Manual desludging exhibits a 4 percentage point decrease among the most heavily subsidized households, relative to a control group Diarrhea incidence among children exhibits a 6.1 percentage point decrease in the most heavily subsidized households, relative to a control group Due to reductions in total cost from adjusting procurement strategies, we have a realized loss of a few dollars, after accounting for subsidies Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

16 Preview of Results The market share of mechanical desludging in treatment clusters exhibits a 5.1 percentage point increase overall and a 9.4 percentage point increase among the most heavily subsidized households, relative to a control group Manual desludging exhibits a 4 percentage point decrease among the most heavily subsidized households, relative to a control group Diarrhea incidence among children exhibits a 6.1 percentage point decrease in the most heavily subsidized households, relative to a control group Due to reductions in total cost from adjusting procurement strategies, we have a realized loss of a few dollars, after accounting for subsidies Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

17 Phase One: Demand modelling Want to design pricing rules to target households based on info available to the government (ONEA) To decide what prices to quote households, we need to guess their outside option and the likelihood they get a mechanical desludging on their own Use a Roy (1951)-style model to predict selection and prices But data about past transactions can t tell us how people will respond to an intermediary: use a demand elicitation experiment to fill in this gap This delivers household-level demand curves Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

18 Phase One: Demand modelling Want to design pricing rules to target households based on info available to the government (ONEA) To decide what prices to quote households, we need to guess their outside option and the likelihood they get a mechanical desludging on their own Use a Roy (1951)-style model to predict selection and prices But data about past transactions can t tell us how people will respond to an intermediary: use a demand elicitation experiment to fill in this gap This delivers household-level demand curves Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

19 Phase One: Demand modelling Want to design pricing rules to target households based on info available to the government (ONEA) To decide what prices to quote households, we need to guess their outside option and the likelihood they get a mechanical desludging on their own Use a Roy (1951)-style model to predict selection and prices But data about past transactions can t tell us how people will respond to an intermediary: use a demand elicitation experiment to fill in this gap This delivers household-level demand curves Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

20 Phase One: Demand modelling Want to design pricing rules to target households based on info available to the government (ONEA) To decide what prices to quote households, we need to guess their outside option and the likelihood they get a mechanical desludging on their own Use a Roy (1951)-style model to predict selection and prices But data about past transactions can t tell us how people will respond to an intermediary: use a demand elicitation experiment to fill in this gap This delivers household-level demand curves Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

21 Phase One: Demand modelling Want to design pricing rules to target households based on info available to the government (ONEA) To decide what prices to quote households, we need to guess their outside option and the likelihood they get a mechanical desludging on their own Use a Roy (1951)-style model to predict selection and prices But data about past transactions can t tell us how people will respond to an intermediary: use a demand elicitation experiment to fill in this gap This delivers household-level demand curves Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

22 Model of Decision and Price Determination Latent index: ỹ i = + ε i x i δ { Observed decision: y i = manual, { mechanical, ỹ i 0 ỹ i < 0 Mechanical price: p mechanical,i =, { z i β mech + ε mech,i, ỹ i 0 ỹ i < 0 Manual price: p manual,i =, ỹ i 0 z i β man + ε man,i, ỹ i < 0 z i are observable to desludgers, while x i z i contain variables unobserved to them: exclusion restriction is based on price discrimination (Estimates) Reduced-form model predicts prices and decisions, but can t predict household behavior under counterfactual prices Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

23 Model of Decision and Price Determination Latent index: ỹ i = + ε i x i δ { Observed decision: y i = manual, { mechanical, ỹ i 0 ỹ i < 0 Mechanical price: p mechanical,i =, { z i β mech + ε mech,i, ỹ i 0 ỹ i < 0 Manual price: p manual,i =, ỹ i 0 z i β man + ε man,i, ỹ i < 0 z i are observable to desludgers, while x i z i contain variables unobserved to them: exclusion restriction is based on price discrimination (Estimates) Reduced-form model predicts prices and decisions, but can t predict household behavior under counterfactual prices Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

24 Model of Decision and Price Determination Latent index: ỹ i = + ε i x i δ { Observed decision: y i = manual, { mechanical, ỹ i 0 ỹ i < 0 Mechanical price: p mechanical,i =, { z i β mech + ε mech,i, ỹ i 0 ỹ i < 0 Manual price: p manual,i =, ỹ i 0 z i β man + ε man,i, ỹ i < 0 z i are observable to desludgers, while x i z i contain variables unobserved to them: exclusion restriction is based on price discrimination (Estimates) Reduced-form model predicts prices and decisions, but can t predict household behavior under counterfactual prices Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

25 Buyers and Switchers Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

26 Incentive Compatible Demand Elicitation Given a counterfactual price offer and the household s z i, how would they respond? Ask them: the highest-rejected-bid auction is the game where 1 Each household i faces 8 competitors, but only K < 8 will be selected to win a desludging 2 Each household i is asked to make an offer, w i, for a desludging. 3 The highest K offers are accepted, and all winners are asked to pay the K + 1-st price when they come forward to purchase a desludging. If K = 1, second-price auction; K > 1, it is still a weakly dominant strategy to bid honestly Motivate by fairness: If you won and someone else lost, you should be willing to pay at least what they would have paid. Follow-up by asking whether they would regret losing at their offer given a slightly higher price, let them revise until they are satisfied Randomize K = {2, 3, 5, 7} to test whether agents play the dominant strategy equilibrium: K isn t correlated with bids Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

27 Incentive Compatible Demand Elicitation Given a counterfactual price offer and the household s z i, how would they respond? Ask them: the highest-rejected-bid auction is the game where 1 Each household i faces 8 competitors, but only K < 8 will be selected to win a desludging 2 Each household i is asked to make an offer, w i, for a desludging. 3 The highest K offers are accepted, and all winners are asked to pay the K + 1-st price when they come forward to purchase a desludging. If K = 1, second-price auction; K > 1, it is still a weakly dominant strategy to bid honestly Motivate by fairness: If you won and someone else lost, you should be willing to pay at least what they would have paid. Follow-up by asking whether they would regret losing at their offer given a slightly higher price, let them revise until they are satisfied Randomize K = {2, 3, 5, 7} to test whether agents play the dominant strategy equilibrium: K isn t correlated with bids Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

28 Incentive Compatible Demand Elicitation Given a counterfactual price offer and the household s z i, how would they respond? Ask them: the highest-rejected-bid auction is the game where 1 Each household i faces 8 competitors, but only K < 8 will be selected to win a desludging 2 Each household i is asked to make an offer, w i, for a desludging. 3 The highest K offers are accepted, and all winners are asked to pay the K + 1-st price when they come forward to purchase a desludging. If K = 1, second-price auction; K > 1, it is still a weakly dominant strategy to bid honestly Motivate by fairness: If you won and someone else lost, you should be willing to pay at least what they would have paid. Follow-up by asking whether they would regret losing at their offer given a slightly higher price, let them revise until they are satisfied Randomize K = {2, 3, 5, 7} to test whether agents play the dominant strategy equilibrium: K isn t correlated with bids Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

29 Incentive Compatible Demand Elicitation Given a counterfactual price offer and the household s z i, how would they respond? Ask them: the highest-rejected-bid auction is the game where 1 Each household i faces 8 competitors, but only K < 8 will be selected to win a desludging 2 Each household i is asked to make an offer, w i, for a desludging. 3 The highest K offers are accepted, and all winners are asked to pay the K + 1-st price when they come forward to purchase a desludging. If K = 1, second-price auction; K > 1, it is still a weakly dominant strategy to bid honestly Motivate by fairness: If you won and someone else lost, you should be willing to pay at least what they would have paid. Follow-up by asking whether they would regret losing at their offer given a slightly higher price, let them revise until they are satisfied Randomize K = {2, 3, 5, 7} to test whether agents play the dominant strategy equilibrium: K isn t correlated with bids Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

30 Incentive Compatible Demand Elicitation Given a counterfactual price offer and the household s z i, how would they respond? Ask them: the highest-rejected-bid auction is the game where 1 Each household i faces 8 competitors, but only K < 8 will be selected to win a desludging 2 Each household i is asked to make an offer, w i, for a desludging. 3 The highest K offers are accepted, and all winners are asked to pay the K + 1-st price when they come forward to purchase a desludging. If K = 1, second-price auction; K > 1, it is still a weakly dominant strategy to bid honestly Motivate by fairness: If you won and someone else lost, you should be willing to pay at least what they would have paid. Follow-up by asking whether they would regret losing at their offer given a slightly higher price, let them revise until they are satisfied Randomize K = {2, 3, 5, 7} to test whether agents play the dominant strategy equilibrium: K isn t correlated with bids Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

31 Auction Example There are three households, A, B, and C, with values 10, 5, and 3, respectively. Suppose there are two units available: this is an HRB auction with K = 2. If everyone bids honestly, A and B win, but pay C s bid of 3. If C were to raise her bid above 5, she could win, but would get a negative payoff. If either or A or B reduce their bids, a change only occurs when they go from winning to losing. No one has profitable deviations. More generally, it is a weakly dominant strategy to bid honestly: (1) If you raise your bid above your true value, there are three cases: you might go from losing to winning and make a loss, you might still be a loser and make no gain, and you might be a winner and increasing your bid doesn t change your payment. (2) If you lower your bid below your true value, there are three cases: you might go from winning to losing and lose the benefit you would have had from winning, you might be a loser and continue to lose, and you might be a winner and increasing your bid doesn t change your payment. Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

32 Auction Example There are three households, A, B, and C, with values 10, 5, and 3, respectively. Suppose there are two units available: this is an HRB auction with K = 2. If everyone bids honestly, A and B win, but pay C s bid of 3. If C were to raise her bid above 5, she could win, but would get a negative payoff. If either or A or B reduce their bids, a change only occurs when they go from winning to losing. No one has profitable deviations. More generally, it is a weakly dominant strategy to bid honestly: (1) If you raise your bid above your true value, there are three cases: you might go from losing to winning and make a loss, you might still be a loser and make no gain, and you might be a winner and increasing your bid doesn t change your payment. (2) If you lower your bid below your true value, there are three cases: you might go from winning to losing and lose the benefit you would have had from winning, you might be a loser and continue to lose, and you might be a winner and increasing your bid doesn t change your payment. Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

33 Offers/Switching Prices Percent Offers, 1000 CFA Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

34 Computation Demand given a quote t i and observables x i : D(t i, x i ) = E ε I{ỹ i 0 t i < p mech,i } }{{} Participating always-buyers + I{ỹ i 0 t i > p mech,i } }{{} Non-participating always-buyers + I{ỹ i < 0 t i < p mech,i t i w i } }{{} x i, Switchers We (i) draw a shock ε from the distribution of residuals, (ii) this determines (ˆp mech,i + ε mech, ˆp man,i + ε man, ŷ i + ε 0 ), (iii) evaluate the indicator functions above, except for replacing the probability of switching with the probability that p mech,i > w i > t i given that ỹ 0 < 0, (iv) repeat over many values of ε Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

35 Computation Demand given a quote t i and observables x i : D(t i, x i ) = E ε I{ỹ i 0 t i < p mech,i } }{{} Participating always-buyers + I{ỹ i 0 t i > p mech,i } }{{} Non-participating always-buyers + I{ỹ i < 0 t i < p mech,i t i w i } }{{} x i, Switchers We (i) draw a shock ε from the distribution of residuals, (ii) this determines (ˆp mech,i + ε mech, ˆp man,i + ε man, ŷ i + ε 0 ), (iii) evaluate the indicator functions above, except for replacing the probability of switching with the probability that p mech,i > w i > t i given that ỹ 0 < 0, (iv) repeat over many values of ε Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

36 Estimated demands Pr[ Accept t ] Demand id 0.75 id Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

37 Phase 2: Design of the Pricing Rule The goal of our pricing rule is to increase take-up; we want to implement a particular pattern of behavior by households, that is incentive compatible with their private information w i Formulate a social planner s problem and solve it Solution quotes high prices to relatively wealthy households who are likely to get mechanical, uses that revenue to relax the budget constraint, and then quotes low prices to households that are likely to switch Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

38 Phase 2: Design of the Pricing Rule The goal of our pricing rule is to increase take-up; we want to implement a particular pattern of behavior by households, that is incentive compatible with their private information w i Formulate a social planner s problem and solve it Solution quotes high prices to relatively wealthy households who are likely to get mechanical, uses that revenue to relax the budget constraint, and then quotes low prices to households that are likely to switch Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

39 Phase 2: Design of the Pricing Rule The goal of our pricing rule is to increase take-up; we want to implement a particular pattern of behavior by households, that is incentive compatible with their private information w i Formulate a social planner s problem and solve it Solution quotes high prices to relatively wealthy households who are likely to get mechanical, uses that revenue to relax the budget constraint, and then quotes low prices to households that are likely to switch Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

40 Optimization Problem Let t i be the price we quote to household i with observables x i, D M (t i, x i ) be their expected demand for mechanical in the decentralized market, and D p (t i, x i ) be their expected demand through the platform Then we maximize total demand for mechanical max t i N i=1 D M (t i, x i ) + D P (t }{{} i, x i ) }{{} Market demand Platform demand subject to the budget constraint N }{{} S + D p (t i, x i )(t i c) 0 Subsidies i=1 } {{ } Platform profits and the linear programming constraint t i {8000, 10000, 12500, 15000, 17500, 20000} Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

41 Optimization Problem Let t i be the price we quote to household i with observables x i, D M (t i, x i ) be their expected demand for mechanical in the decentralized market, and D p (t i, x i ) be their expected demand through the platform Then we maximize total demand for mechanical max t i N i=1 D M (t i, x i ) + D P (t }{{} i, x i ) }{{} Market demand Platform demand subject to the budget constraint N }{{} S + D p (t i, x i )(t i c) 0 Subsidies i=1 } {{ } Platform profits and the linear programming constraint t i {8000, 10000, 12500, 15000, 17500, 20000} Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

42 Optimal Rule Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

43 Optimal Rule Since the original sample is random, this rule is optimal for the population: I i=1...x i... can be replaced with x i...df (x i ) in the optimization problem We convert the household-based pricing rule {(z i, t i )}N i=1 into a function that maps household characteristics into prices t (z i ) using ordered logit mapping x i into the price bin {10000, 15000, 17500, 20000} Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

44 Optimal Rule Since the original sample is random, this rule is optimal for the population: I i=1...x i... can be replaced with x i...df (x i ) in the optimization problem We convert the household-based pricing rule {(z i, t i )}N i=1 into a function that maps household characteristics into prices t (z i ) using ordered logit mapping x i into the price bin {10000, 15000, 17500, 20000} Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

45 Phase 3: Randomized Trial of the Pricing Rule We sampled 92 new neighborhoods, 2944 households to test the impact of the pricing rule. The new households z i s were gathered, and tablets computed the price t (z i ) in the background, which was quoted as a take-it-or-leave-it offer at the end of the interview Households paid a non-refundable deposit of 500 CFA if they chose to accept. Households were allowed to call any time within the 18 months (August 2015-January 2017) following the survey to redeem their desludging and pay the remainder. Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

46 (1) (2) (3) Control(SD) Diff Treat - Control(SE) Observations Number of Members of Household (4.17) (0.23) Number of Women in Household (1.70) (0.08) Indicator: Respondent completed Secondary (0.46) (0.03) Number of years lived at compound (13.63) (1.09) House Owned by Inhabitants (0.42) (0.02) Water bill more than 5,000 CFA (0.50) (0.03) Electricity Bill (in thousands of CFA) (15.43) (0.93) Latrine Pit Distance to Road (4.42) (0.31) Two tanks used last Desludging (0.15) (0.01) N Months Between Desludgings (27.19) (1.54) Last Desludging was Manual (0.44) (0.03) Has never Desludged Here (0.46) (0.03) Ever Desludged Mechanically (0.50) (0.04) Ever Desludged Manually (0.42) (0.03) Compound has 1 pit only (0.47) (0.03) (10k Group) Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

47 Table: Call Center Take Up Targeted Price Level Total Pct Offered Price Deposited Percent take-up through CC 1st 6 months Percent take-up through CC (from deposited and desludged) Modeled Take up Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

48 Reasons Households did not Call the Call Center Targeted Price Didn t need a desludging 368 Forgot about it 60 Better Outside option 59 Too Confusing/didn t understand 46 New to the compound 24 Not in charge of desludging 32 Other/refusal 20 Total 606 Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

49 Outline 1 Introduction 2 Market Design and Implementation 3 Treatment Effects Market Share Effects Was targeting successful? Impact on Manual Desludging and Health Comparison to Counter-factual Subsidies Procurement 4 Conclusion Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

50 Estimation of Key Effects: Variables of Interest 1 Do households switch from manual desludging to mechanical desludging? MarketShare = MechanicalDesludgings ManualandMechanicalDesludgings 2 Does the number of Manual Desludgings go down? Odds that a household purchases at least one manual desludging during our program. 3 What is the effect on health? Probability that a household reports that at least one child had a case of diarrhea in the past 7 days. Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

51 Empirical Strategy: Cluster Level Regressions MarketShareMechanical i = α + βtargetedpricestreatment i + γ X i + ɛ i MktShareMech PriceGrp ki = γ X ki + Σ 4 k=1 α kpricegroup ki + Σ 4 k=1 β ktargetedpricestreatment ki PriceGroup ki + ɛ ki Randomization at the neighborhood level Observations are at the neighborhood level for overall effect, neighborhood-price group level for price group effect. Outcome variable: market share of mechanical desludging Control Variables: (neighborhood averages) pit meters from road, last desludging required multiple trips, one pit in household, water bill above 5,000 CFA, above median number low walled compounds. Coefficient of Interest: β for the overall effect, β k, the impact of treatment for each of the different price groups. Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

52 Table: Overall Impact of Treatment Mkt share Pooled By Price Group Overall * (0.028) Price of 10k group * (0.052) Price of 15k group (0.049) Price of 17.5k group (0.051) Price of 20k group (0.074) N Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

53 Who received the large subsidies in practice? Pooled Phone Credit use over past week (1929) (2930) (5660) (10195) (4152) Number of Refrigerators (0.430) (0.665) (0.771) (0.726) (0.706) Number of Cars (0.248) (0.577) (0.838) (1.073) (0.683) Number of Air Conditioners (0.157) (0.359) (1.004) (1.909) (0.722) Ever Desludged Mech (0.479) (0.495) (0.486) (0.468) (0.499) Expected Price Mechanical (CFA) (4717) (4743) (5550) (7120) (5173) Last used Manual (0.500) (0.414) (0.361) (0.171) (0.440) Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

54 Empirical Strategy: Household Level Regressions Outcome i = α + βtargetedpricestreatment i + γ X i + ɛ i Outcome i = Σ 4 k=1 α kpricegroup i + γ X i +Σ 4 k=1 β ktargetedpricestreatment i PriceGroup i + ɛ i Randomization at the neighborhood level, observations are at the household level, standard errors are clustered at the neighborhood level. Outcome variable: purchased at least one manual desludging, reported at least one household child suffered from diarrhea last week. Control Variables: pit meters from road, last desludging required multiple trips, one pit in household, water bill above 5,000 CFA, above median number low walled compounds. Coefficient of Interest: β for the overall effect, β k, the impact of treatment for each of the different price groups. Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

55 Table: Effect on Manual Desludging % Manual BL Any manual Targeted price treatment 26.3% (0.010) Targeted price*10k group 51% * (0.017) Targeted price*15k group 22% (0.012) Targeted price *17.5k group 15.3% (0.018) Targeted price * 20k group 3% (0.032) N R Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

56 Table: Effect on Diarrhea Diarrhea Child had Diarrhea Prevalence BL last 7 days Targeted price treatment 13% (0.017) Targeted price*10k group 16% * (0.032) Targeted price*15k group 14% (0.021) Targeted price *17.5k group 10% (0.035) Targeted price * 20k group 7.3% (0.060) N R Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

57 Mechanisms: Take-up through the call center relatively low direct impact possible, but requires many households purchasing from us, especially from low price group to be switching from manual. Alternative mechanisms: Price offer provides outside option for households as they negotiate with desludger may even receive better prices than we offer. Households in the treatment group report having paid prices 1,120 CFA lower than the control group on average (pvalue.015 clustered at the neighborhood level). Households that have not desludged in the past update their beliefs about the affordability of manual versus mechanical/know at least one place to find a mechanical desludger. Sample size is small, but some evidence that a large part of the effect comes from this group. Program provides households with an increased interest in improving sanitation in their neighborhood more pressure to take up mechanized desludging from neighbors. Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

58 Simulated comparison with a flat subsidy Table: Simulated Impact of a Subsidy Program Mkt share Pooled Effect By Price Group Target P CF Subs Target P CF Subs Overall * (0.028) (0.029) 10k group * (0.052) (0.072) 15k group (0.049) (0.031) 17k group (0.051) (0.038) 20k group (0.073) (0.0153) N Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

59 Procurement with Auctions versus Negotiation Auction Prices by Block Negotiation Prices by Block Price FPAs begin FPAs begin Block Time Time Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

60 Table: Ex post platform budget estimates Cost Scenario: Budget (CFA) Budget (USD) Round-by-neighborhood auction price Last observed auction price Mean negotiation price Last observed negotiation price Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

61 Conclusion and Continuing Work Intermediation of the sanitation market leads to greater market share of improved service and less manual desludging and childhood diarrhea in the low price group We minimize the budget necessary to generate these improvements by targeting the poorest households for subsidies, including the wealthy households in the market and charging them higher prices, and instigating increased competition to lower procurement costs. In settings where legal institutions are weak, a successful platform can act as a policy carrot We identify market issues and test solutions based on intermediation: Senegal: generating competition through just-in-time auctions and supply-side targeting Ghana: cashless system that uses current prices to allocate future jobs Burkina supply side: paid-as-bid, lowest-rejected-bid, and structured negotiations negotiations reduce prices 9% Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

62 Demand Estimates 1 Selection Mechanical Price Manual Price Desludging frequency *** *** *** Water greater than 5,000 CFA 0.054*** *** Precarious House *** Concrete House, 1 story *** Rooming House *** Other households in compound *** *** Own House *** *** Pit distance to road *** *** Last trips greater than one *** *** Electricity bill *** Number persons in household *** Number women in household *** Household head educated *** Constant *** *** N (Return) 1 Standard errors suppressed for brevity Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38

63 (1) (2) (3) Control(SD) Diff Treat - Control(SE) Observations Number of Members of Household (4.23) (0.38) Number of Women in Household (1.62) (0.13) Indicator: Respondent completed Secondary (0.21) (0.02) Number of years lived at compound (13.78) (1.48) House Owned by Inhabitants (0.24) (0.02) Water bill more than 5,000 CFA (0.45) (0.04) Electricity Bill (in thousands of CFA) (5.44) (0.52) Latrine Pit Distance to Road (3.33) (0.29) Two tanks used last Desludging (0.08) (0.00) N Months Between Desludgings (34.92) (3.69) Last Desludging was Manual (0.50) (0.05) Has never Desludged Here (0.47) (0.05) Ever Desludged Mechanically (0.49) (0.05) Ever Desludged Manually (0.40) (0.04) Compound has 1 pit only (0.49) (0.04) (Back) Johnson, Lipscomb Targeting through Mechanism Design March 15, / 38