Heckle and Chide: Empowering matatu passengers to enforce better driving behavior in Kenya James Habyarimana Georgetown University and William Jack Georgetown University
Motivation Accidents happen! he says, with a cheerful shrug. "Maniacs? Maybe we are a little bit - but you've got to drive fast to get the money! –A matatu driver in Kenya Taxi drivers put money first and passengers' and pedestrians' lives second –Patrick Ayumu, Ghana
Objectives of the project Evaluate a randomized intervention aimed at reducing matatu accidents by –empowering passengers to…. –.…enforce better driving behavior Using evocative messages placed inside the matatu
WHY road safety? Major cause of injury and death –Rising share in global deaths Economically costly – 2% of national income in Kenya –Vulnerable population: working age (15-44) accounts for 75% of RT fatalities (Odero (2003))
Source: Mathers and Loncar (2006)
WHY matatus ? They account for a large share of inter-city passenger transport –Vulnerable population in road traffic injuries They are involved in 20% of recorded crashes –But larger share of injuries/fatalities They are well suited to our intervention
WHY so many crashes? Road conditions Vehicle conditions Behavior of other road users Behavior of matatu drivers Focus of study
WHO can affect driver behaviour? Matatu drivers Owners / Operators Government (incl. police) Passengers Focus of study
HOW do we empower passengers? Tell them to speak up! Heckle and Chide Insert stickers with messages inside matatus
WHICH stickers: heckle and chide imperatives?
WHICH stickers: The soft touch?
WHICH stickers: Shock therapy?
Sticker Placement Plan Side door Front of matatu Drivers seat Ajali Foot Leg Sit Vibaya
HOW do we evaluate impact of the intervention? RCT –compare randomly selected matatus with stickers to a control group of matatus without stickers Outcome measures –Crash rates Associated injuries/fatalities –Survey results of passenger and driver behavior
Motivating the intervention Are accident rates efficient? –Collective action problems inside matatus If not, what is the role of regulation? –Enforcement problems in public regulation Stickers could either: –increase perceived benefit of action – if people underestimate the effects of accidents; or –reduce the cost of taking action stickers legitimize heckling Focal point for passenger action
Matatu-land, Nairobi
Recruitment
A challenging research environment
Outcome variable
Timeline August 2007 March 2008 May 2008 Recruitment Pilot recruitment Weekly Raffles Accident Data Collection Follow up surveys Trip observations January 2000
Data Sample of 2,276 matatus from 21 SACCOs* –6 SACCOs account for about 50% of the sample account for vehicles 40% of sample had been assigned during pilot phase –Random assignment from SACCO lists to treatment status: p= % new matatus –assignment based on last digit of plate number –Odd Treatment –Even Control * Savings and Credit Cooperatives
Consent and Compliance Informed consent obtained from drivers Consent from owners very difficult Better compliance in pilot sample Shares of matatus that at least one sticker inserted AssignmentEntire SampleOld SampleNew Sample Control (no stickers) Treatment (stickers) Total 0.52
Partial compliance Percent receiving each treatment Number of StickersControlTreatment
Sample Balance CovariatesControlTreatmentDifference Significant Owns cell phone 0.89 (0.01) 0.91 (0.01) No Odometer reading 356,506 (7,236) [327,365] 361,386 (6,350) [343,602] No Capacity (passengers) (0.05) (0.05) No Uses tout 0.45 (0.02) 0.48 (0.01) No Number of weekly trips (0.36) [14] (0.30) [14] No Average daily distance (6.14) [400] (5.33) [400] No Has speed governor 1.00 (0.00) 1.00 (0.00) No Share owned by large Cooperative 0.49 (0.02) 0.51 (0.01) No Involved in accident in last 12 months (0.002) (0.004) Yes
Selected CovariatesCOMPLIANTNON-COMPLIANT ControlTreatmentControlTreatment Owns a cell phone 0.87 (0.01) 0.92 (0.01) 0.97 (0.01) 0.84 (0.03) Odometer Reading ( ) [ ] ( ) [ ] ( ) [401230] ( ) [ ] Passenger Capacity (0.05) (0.05) (0.13) (0.09) Proportion use tout 0.45 (0.02) 0.50 (0.02) 0.40 (0.04) 0.39 (0.04) Age, years 2.29 (0.10) 2.78 (0.10) 3.00 (0.27) 2.67 (0.23) Number of weekly trips (0.40) [14] (0.32) [14] (0.87) [14] (0.74) [14] Average daily distance, km (6.41) [400] (5.86) [400] (18.62) [400] (12.78) [400] Proportion large Saccos 0.46 (0.02) 0.51 (0.02) 0.65 (0.04) 0.51 (0.04) Proportion had accident in last 12 mths (0.002) (0.004) (0.008) What drives selection into actual treatment?
Outcome data: accidents Main outcome of interest is accidents –Accident occurrence –Severity - # injured, killed per accident Collected data from two sources –Insurance companies All vehicles are required to have minimal coverage In theory all accidents should be observable – submission of claims endogenous –Our own data collection efforts
Other outcome data Survey data from drivers and passengers to assess behavior of both –Safety of drivers –Heckling and chiding by passengers Direct observation of driver behavior –Send anonymous passengers on matatu trips?
Empirical Specification Difference-in-differences strategy to estimate –Parallel trends assumption Main concern is that treatment status is potentially endogenous Estimate intent-to-treat parameter Use assignment to treatment as instrument –IV estimates
Actual Treat status Before (2007) After (2008) Difference Control.045 (.007).041 (.006) (.009) Treatment.057 (.006).025 (.005) (.007) Difference.012 (.009) (.007)* (.012)* Average treatment effect Standard errors in parentheses, + significant at 10%; * significant at 5%; ** significant at 1%
AssignmentBeforeAfterDifference Control.045 (.007).040 (.006) (.008) Treatment.057 (.007).026 (.005) (.008) Difference.012 (.01) (.008) (.011)* Intent-to-Treat Estimator Standard errors in parentheses, + significant at 10%; * significant at 5%; ** significant at 1%
Actual TreatmentIntent-to-TreatInstrumental Variables (1)(2)(3)(4)(5)(6) Post (0.588)(0.662)(0.580)(0.654)(0.928)(0.840) Treatment status (0.227)(0.116)(0.189)(0.155)(0.189)(0.161) Post * Treatment (0.026)*(0.021)*(0.032)*(0.025)*(0.032)*(0.025)* Constant (0.000)**(0.001)**(0.000)**(0.001)**(0.000)**(0.007)** Management Controls XXX Observations R-squared Average Treatment Effects: LPM P-values in parentheses, + significant at 10%; * significant at 5%; ** significant at 1%
Results so far Claims rate (% p.a.) % Change over baseline Accidents avoided Deaths/ injuries avoided Baseline 5.1 Sticker effect Average treatment effect %?? ITT /Reduced form %?? IV %??
Next Steps Examine data on possible mechanisms Collect more detailed claims data from insurance companies –Includes data on injuries –Types of events being affected by intervention Direct observation