Incentives and the Allocation of Authority in Organizations: A Field Experiment with Bureaucrats. LSE May 24, 2018

Size: px
Start display at page:

Download "Incentives and the Allocation of Authority in Organizations: A Field Experiment with Bureaucrats. LSE May 24, 2018"

Transcription

1 Incentives and the Allocation of Authority in Organizations: A Field Experiment with Bureaucrats Oriana Bandiera LSE Adnan Khan LSE Michael Carlos Best Columbia Andrea Prat Columbia LSE May 24, 2018

2 Outline Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion

3 Agency Problems: Rules and Incentives Agency problems can be addressed by 1. Incentives that encourage the right behaviour 2. Rules that regulate behavior directly Different organisations use different combinations of 1&2 Bureaucracies mostly use rules, common reasons bureaucrats would abuse discretion to their personal advantage private returns higher than any reward the govt can pay Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 2 / 71

4 Multi-level Agency Monitoring of adherence to rules creates a second set of agents: Auditors, inspectors, etc. These are also subject to an agency problem (Shleifer & Vishny 1993; Barron & Olken 2009) Solution to the downstream agents problem might exacerbate the monitor problem more rules > more authority to the monitor > can exploit it to personal advantage Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 3 / 71

5 Autonomy, Incentives and the Allocation of Control Autonomy shifts control from the monitor to the agent > If agents interests are better aligned with the organisation, then autonomy improves performance allows agent to exercise initiative (Banerjee 1992) & leverage local information (Aghion & Tirole 1997) autonomy correlated with higher project completion rates (Rasul & Rogger 2017) and better targeting (Duflo, Greenstone, Hanna & Ryan 2016) Performance pay incentivises the agent but will also affect the strategy of the monitor Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 4 / 71

6 This paper Theoretical framework to make precise how effect of incentives and autonomy depends on the relative strength of the agency problem of the monitor vs. delivery agents Field experiment to provide causal evidence on the effect of incentives and autonomy Context: Procurement of generic goods in Punjab, Pakistan: agents are procurement officers monitors are staff at the accountant general office Treatments: 1. Autonomy, 2.Incentives, 3. Both Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 5 / 71

7 Why procurement of generic goods? important part of government activity notoriously wasteful tangled in red tape measurable (with some effort!) and comparable performance evidence that waste due to excess rules dwarfs corruption (Bandiera, Prat, Valletti 09) Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 6 / 71

8 Preview Theory predicts that if monitor is honest, autonomy (incentives) will increase (decrease) prices if monitor is dishonest, autonomy will decrease prices, while the effect of incentives is ambiguous We find that autonomy reduces prices by 7pp throughout the fiscal year incentives reduce prices at the start but increase at the end > average effect is nil Results consistent with dishonest monitor case Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 7 / 71

9 Outline Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion

10 Model Setup A mass of procurement decisions need to be made 2 players move simultaneously 1. An agent chooses a mark-up level x a 2. For a share γ of decisions a supervisor is active. She chooses a markup level x s γ: supervisory power (agent s autonomy) Final price: p = Average price: { c + x a + x s c + x a if the purchase is audited otherwise p = c + x a + γx s Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 9 / 71

11 Payoffs Agent s payoff: Supervisor s payoff: Key Parameters: 1. α: Agent s honesty 2. β: Supervisor s honesty 3. γ: Agent s autonomy V a = u (x a ) αg ( p) γx a V s = u (x s ) βg ( p) Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 10 / 71

12 Agent s Reaction Function ˆx a = arg max xa u (x a ) αg (c + x a + γx s ) γx a Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 11 / 71

13 Supervisor s Reaction Function ˆx s = arg max xs u (x s ) βg (c + x a + γx s ) Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 12 / 71

14 Supervisor s Reaction Function Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 13 / 71

15 Equilibrium Markups Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 14 / 71

16 Autonomy Treatment: Decrease in γ Lower auditing probability Agent markup increases Higher expected price Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 15 / 71

17 Autonomy Treatment: Decrease in γ Lower auditing probability Agent markup increases Higher expected price Lower supervisor markup opportunity Average supervisor markup decreases Lower expected price Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 16 / 71

18 Autonomy Treatment: Decrease in γ Lower auditing probability Agent markup increases Higher expected price Lower supervisor markup opportunity Average supervisor markup decreases Lower expected price The combined price effect is ambiguous. But... Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 17 / 71

19 Autonomy Treatment: Decrease in γ Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 18 / 71

20 Autonomy Treatment: Decrease in γ Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 19 / 71

21 Incentive Treatment: Increase in α Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 20 / 71

22 Incentive Treatment: Increase in α Lower markup x a decreases Supervisor markup x s increases Average price c + x a decreases Unaudited price c + x a + x s decreases Audited price c + x a + γx s decreases or increases Depends on how honest the supervisor is Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 21 / 71

23 Incentive Treatment: Increase in α Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 22 / 71

24 Incentive Treatment: Increase in α Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 23 / 71

25 Time Variation in Supervisor Power Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 24 / 71

26 Time Variation in Supervisor Power Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 25 / 71

27 Time Variation in Supervisor Power Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 26 / 71

28 Autonomy Treatment Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 27 / 71

29 Incentives Treatment Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 28 / 71

30 Relative Time Variation Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 29 / 71

31 Summary of Predictions on Prices Treatment Honest Supervisor Dishonest Supervisor All Early Late All Early Late Autonomy Incentives or = - + Combined??? - -? Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 30 / 71

32 Outline Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion

33 Outline Context & Data Procurement in Punjab, Pakistan Measuring Value for Money

34 Procurement in Punjab, Pakistan The legal authority for public procurement is vested in Procurement Officers (POs) POs manage Public Bodies, allocated budget under different accounting heads (salary, repairs, etc.), including procurement, by the Finance Department POs are required to submit all their expenditures to an independent federal agency - the office of the Accountant General, or AG office - for pre-audit before payment can be processed. AG has offices in each district, responsible for POs in that district Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 33 / 71

35 A Typical Procurement Process 1. A demand for an item goes to the PO for approval 2. PO surveys the market for vendors and rates for the items 3. PO receives the goods from the vendors 4. PO sends a request for payment (bill/voucher) to the AG office 5. AG sanctions payment to the vendor or demands more paperwork. Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 34 / 71

36 Outline Context & Data Procurement in Punjab, Pakistan Measuring Value for Money

37 Measuring Value for Money We focus on the universe of generic goods. these account for a large share of government budget bought by many consumers and produced by several suppliers most are sold in competitive markets, so everybody should pay the same price. And yet.... Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 36 / 71

38 Different POs pay very different prices for exactly the same good Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 37 / 71

39 We saw this already Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 38 / 71

40 Together with Punjab Procurement Regulatory Authority (PPRA) and Punjab Information Technology Board (PITB), we set up an E-Governance platform: Punjab Online Procurement System (POPS) Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 39 / 71

41 POPS Collects Detailed Spending Data Through POPS, office staff enter detailed data on what they are buying Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 40 / 71

42 Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 41 / 71

43 A Rich Dataset on Public Procurement We now have a dataset on purchases of 21 goods, from pencils to chairs. 10,702 purchases. Trim top/bottom 1% of unit prices for each item 10,553 obs Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 42 / 71

44 Outline Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion

45 Subjects 664 Procurement offices in charge of procurement of 750 Public Bodies (88% in charge of 1 PB, 10% 2, 2% 3 or more) take-up 80% > sample contains 520 POs 26 Districts (out of 36) - cover over 80% of the population (110million) 4 Departments: Agriculture (254 PBs) Higher Education (404 PBs) Health (32 PBs) Communication and Works (60 PBs) Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 44 / 71

46 Design We assign the POs to 4 equal sized groups 1. Incentives 2. Autonomy 3. Incentives + Autonomy 4. Control Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 45 / 71

47 Location Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 46 / 71

48 Incentive Treatment Twice per year, an independent commission awards 3 prizes 1. gold : 2 months wages, to the top 7.5% 2. silver : 1 month wages, to the next 22.5% 3. bronze : 0.5 month wages, to the next 45% 4. nothing to remaining 25% Commissioners: senior private sector auditor & head PPRA (co-chair), representatives of all departments (10 members) Data on quality adjusted prices provided by us Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 47 / 71

49 Autonomy Treatment? 30.0 These are potential reasons for why DDOs don t achieve good value for money. In your experience how important is each of these? Only a limited number of vendors are willing to wait for delayed payment Vendors charge higher prices for delayed payment DDOs have nothing to gain by improving value for money DDOs are worried that if they change vendors to achieve better value for money this might raise red flags Budgets are released late so DDOs cannot plan appropriately AG/DAO requirements are not clear and they do not clear bills without inside connections or paymentof speed money DDOs do not have enough petty cash to make purchases quickly. DDOs & office staff do not receive enough training on procurement procedures Cost centers cannot roll their budget over into the following year Other 0.0 Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 48 / 71

50 Autonomy Treatment 1. Cash in hand, Rs100k ($1k ) 2. Budget released in two timely installments instead of four 3. Transparent and predictable audit procedure PO better able to plan but also to hide PO has access to more funds in one shot -> more flexibility to strike deals but also more chances to embezzle Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 49 / 71

51 Year 1: July 2015 June 2015 Timeline 06/14 Cost Centers allocated to treatment arms 07 08/14 Trainings on POPS and treatment brochures 08 09/14 Follow-up trainings on POPS 03 04/15 Baseline Survey Year 2: July 2015 June /15 Refresher trainings on treatments and POPS 10/15 Cash in Hand rolled out 03 04/16 Midline Survey 04/16 Performance Evaluation Committee Midline Meeting 06/16 Experiment Ends Post-Experiment 08-09/16 Endline Survey Part 1 & Missing Data Collection 02/17 Performance Evaluation Committee Endline Meeting 02 03/17 Endline Survey Part 2 Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 50 / 71

52 Measuring Outcomes: Value for Money We measure value for money by regressing unit prices on full set of items attributes in the control group ln (p igto ) = X igto β g + ρ g ln (q igto ) + γ g + δ 1 Department o + δ 2 District o price attributes ˆβ C + δ 3 NCCs o + λ 1g t + ε igto Use these attribute prices to control for quality ( ) quality-adjusted prices r igto = ln (p igto ) ln ˆp C igto Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 51 / 71

53 Outline Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion

54 Summary of Predictions on Prices Treatment Honest Supervisor Dishonest Supervisor All Early Late All Early Late Autonomy Incentives or = - + Combined??? - -? Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 53 / 71

55 Identification We estimate r igto = α + 3 η k Treatment k o k=1 + δ 1 Department o + δ 2 District o + γ g + ε igto η k measures the causal effect of treatment k if treatment does not affect control POs Violated if AG behaves differently toward control offices because of treatment 1. Experimental POs are a small fraction of total POs supervised 2. AGs are district specific > compare estimates with and without District FE Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 54 / 71

56 Outline Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion

57 Results Average Effects Heterogeneity Outline

58 Treatment Effects - Purchase level Table: Effect on Quality-Adjusted Prices (1) (2) (3) (4) Incentives (0.029) (0.030) (0.029) (0.030) Autonomy ** ** ** *** (0.026) (0.025) (0.026) (0.026) Both * * * (0.031) (0.031) (0.030) (0.029) Constant * * ** *** (0.054) (0.053) (0.058) (0.070) Item FEs Yes Yes Yes Yes # of CCs No Yes Yes Yes Department FEs No No Yes Yes District FEs No No No Yes p(incentives = Autonomy) p(incentives = Both) p(autonomy = Both) p(both = Sum) Observations Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 57 / 71

59 No Difference in Item Composition E [r igb T ] E [r igb C] = (E [X igb T ] E [X igb C]) β C + ( β T β C) E [X igb T ] }{{}}{{} Item Composition Price per Item Impact Incentives Autonomy Both Treatment Item Composition Price per Item Total Effect Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 58 / 71

60 No Drop in Item Quality Impact Incentives Autonomy Both Treatment Attribute Composition Price per Attribute Total Effect Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 59 / 71

61 Results Incentives do not affect prices, either alone or with autonomy Autonomy leads to: quality adjusted prices decrease by 7pp composition of items purchased unchanged quality increase price paid per attribute decreases by 12pp Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 60 / 71

62 Results Average Effects Heterogeneity Outline

63 Dynamic Treatment Effects Theory predicts that effect of autonomy is constant while incentives reduce prices early on and increase them at the end of the fiscal year Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 62 / 71

64 Dynamic TE: Autonomy Lowers Prices Throughout Average Quality-Adjusted Price η Q1-3 = (0.023) η Q4 = (0.033) 01jul oct jan apr jul2016 Date Autonomy Control Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 63 / 71

65 Dynamic TE: Incentives Lower and then Raise Average Quality-Adjusted Price η Q1-3 = (0.025) η Q4 = (0.034) 01jul oct jan apr jul2016 Date Incentives Control Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 64 / 71

66 Average Quality-Adjusted Price Dynamic TE: Combination η Q1-3 = (0.024) η Q4 = (0.034) 01jul oct jan apr jul2016 Date Both Control Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 65 / 71

67 Dynamic Treatment Effects: All Average Quality-Adjusted Price jul oct jan apr jul2016 Date Incentives Both Autonomy Control Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 66 / 71

68 Proxying for AG type We do not observe type (β) directly, but we can use data from previous fiscal year to back out average district prices both correlated with β Theory predicts autonomy should lead to a larger reduction in prices when β is higher Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 67 / 71

69 Heterogeneity by AG type Treatment Effect Below Median District Above Median District Incentives Autonomy Both Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 68 / 71

70 Outline Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion

71 Conclusion Organizations often use rules and monitoring to deal with agency issues Creates two sets of agents: implementing agents and monitoring agents. Rules allocate authority between the two agents Incentives to one agent offset by response of the other agent Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 70 / 71

72 Conclusion Experimental results from procurement bureaucrats in Punjab, Pakistan show Incentives to implementing agents largely offset by response of monitors improvements in early part of year offset by much worse outcomes at end of year Shifting authority to implementing agents improves outcomes prices go down on average monitor-driven seasonality disappears Allocation of authority and incentives to different sets of agents needs to depend on relative alignment with principal s preferences. Motivation Theory Context & Data Experimental Design From theory to data Results Conclusion 71 / 71