Randomized Evaluation:
Start to Finish
Freida Siregar
J-PAL SEA
Course Overview
1. What is evaluation? Why Evaluate?
2. Why randomize?
3. How to randomize
4. Threats and Common Pitfalls
5. Evaluation from Start to Finish
6. Cost Effectiveness Analysis
Course Overview
1. What is evaluation? Why Evaluate?
2. Why randomize?
3. How to randomize
4. Threats and Common Pitfalls
5. Evaluation from Start to Finish
6. Cost Effectiveness Analysis
Evaluation Overview
4
Evaluation Design
• Context
• Research/ Policy question
• Treatments
• Randomization
Baseline
• Survey: Collect indicators
• Random assignment
• Quality checks
Treatment
• Treatment implementa-tion
• Internal controls
• Coordination
• Field visit
Midline and/or Endline
• Survey: Collect indicators
• Quality checks
Analysis
• Policy findings
• Policy outreach
Evaluation from Start to Finish
5
1. Background and Design
2. Implementation: Treatment
3. Implementation: Data Collection
4. Analysis and Scale up
Evaluation background
6
Background and Design
7
Context
• Raskin program
• Partner: TNP2K
Research/ Policy question
• Research/ Policy question
• Theory of change
• Log frame
Treatments Randomization
• Unit of randomization
• Sample frame
• Stratification
Background and Design
8
Context
• Raskin program
• Partner: TNP2K
Research/ Policy question
• Research/Policy question
• Theory of change
• Log frame
Treatments Randomization
• Unit of randomization
• Sample frame
• Stratification
The program: Raskin
9
• One of Indonesia’s largest social assistance programs – US $1.5 billion spending each year
– 53% of national public expenditure on social assistance (World Bank 2012)
• Provides poor and vulnerable households with subsidized rice – 15 kg of rice sold for Rp. 1,600 per kg at Distribution Point
(Titik Distribusi)
– Targeted to Indonesia’s poorest 30% (PPLS’10)
• In practice, eligible households only received 32% of the intended subsidy (Raskin HH survey)
The partner: TNP2K
10
• Tim Nasional Percepatan Penanggulangan Kemiskinan (TNP2K),
• Established by and positioned under Vice President Boediono
• TNP2K’s mandate: – Broker evidence-based policies for
improving effectiveness of social assistance programs,
– Coordinate multiple government actor towards this purpose
• Improved program targeting and delivery is a major priority for TNP2K. Raskin falls under TNP2K’s Cluster I
Background and Design
11
Context
• Raskin program
• Partner: TNP2K
Research/ Policy question
• Research/Policy question
• Theory of change
• Log frame
Treatments Randomization
• Unit of randomization
• Sample frame
• Stratification
Evaluation Objectives
12
Build evidence about whether TNP2K’s proposed
distribution of Raskin ID cards will improve the Raskin
program
The pilot aims to answer three key policy questions:
1. Will Raskin cards improve Raskin targeting of and
take-up rates for eligible households (RTS-PM), and
reduce price?
2. Will Raskin cards be socially accepted?
3. What is the most effective way to implement cards?
Theory of Change
Distribution of Raskin ID cards to Raskin
beneficiaries
Reception of Raskin ID cards
Beneficiaries obtain more of
their Raskin subsidies
Increase effectiveness of social protection
programs
Sample assumption: Beneficiaries understand use of cards, use cards, do not confuse with previously existing Raskin cards/coupons
Sample assumption: Beneficiaries demand reduced prices, village Raskin authorities listen to beneficiary demands and have ability to make change
Sample assumption: Successful delivery to household, no “bypassing” of cards
Sample assumption: Fair(er) practices sustained, poor program effectiveness due to lack of transparency
Log Frame
14
Objectives Hierarchy Indicators Sources of Verification Assumptions / Threats
Impact (Goal/ Overall
objective)
Increase
transparency and
effectiveness of
social protection
programs
Quantity and price of Raskin purchased
Household survey Fair(er) practices sustained, poor program effectiveness due to lack of transparency
Outcome (Project
Objective)
Beneficiaries
obtain more of their
Raskin subsidies
Quantity and price of Raskin purchased
Household survey Beneficiaries demand reduced prices, village Raskin authorities listen to beneficiary demands and have ability to make change
Outputs Reception of Raskin ID cards
Whether or not beneficiaries receive Raskin cards
Household survey Beneficiaries understand use of cards, use cards, do not confuse with previously existing Raskin cards/coupons
Inputs (Activities)
Distribution of Raskin ID cards to Raskin beneficiaries
Whether or not cards are sent
Household survey, administrative data from PT Pos*
Successful delivery to household, no “bypassing” of cards
Background and Design
15
Context
• Raskin program
• Partner: TNP2K
Research/ Policy question
• Research/Policy question
• Theory of change
• Log frame
Treatments Randomization
• Unit of randomization
• Sample frame
• Stratification
Intervention Overview
16
Raskin ID Card
Variation 1: Design
Variation 2: Content
Variation 3: Distribution
Variation 4: Socialization
Coupons
No coupons
Price at TD
No Price
All beneficiaries
Bottom 10%
Standard
Enhanced
Example of Raskin ID Cards
17
Raskin card with coupon and price Raskin card without coupon, with price
Enhanced Socialization Posters
18
Treatment variations
19
Card Variations Standard
socialization Enhanced
Socialization
All beneficiary
Price Coupon Group 1 Group 2
No Coupon Group 3 Group 4
No Price Coupon Group 5 Group 6
No Coupon Group 7 Group 8
Bottom 10%
Price Coupon Group 9 Group 10
No Coupon Group 11 Group 12
No Price Coupon Group 13 Group 14
No Coupon Group 15 Group 16
Control (No card, no socialization)
Treatment Implementation Sept – Nov ‘12
Data Collection
Midline Survey Oct – Dec ‘12
Endline Survey Mar – May ‘13
Project Implementation
Baseline Jan – Feb ‘12
Dec 2012 presentation
June 2013 presentation
Process & Impact Evaluation
20
Design Evaluation
Background and Design
21
Context
• Raskin program
• Partner: TNP2K
Research/ Policy question
• Research/Policy question
• Theory of change
• Log frame
Treatments Randomization
• Unit of randomization
• Sample frame
• Stratification
Identifying unit of randomization
22
• What is the smallest
administrative unit at which
Raskin distribution is
conducted?
• Kecamatan? Gudang bulog?
Village? Dusun?
Source: www.bulog.co.id
Province
City/Regency
Village
Sub-city /regency
Sub-city /regency
Governor
Mayor/Regent
Perum Bulog (Divre/Subdivre/Kansilog)
Gudang (Satgas Raskin)
Titik Distribusi (Pelaksana Distribusi)
Pokja Warung Desa Pokmas
Raskin HH beneficiaries (Paid in cash, Rp 1.600/kg netto at the
Distribution Point)
Head of National Raskin Team (Kemeko Bid Kesra)
Sample frame
23
• 600 villages (including control)
28 dropped due to risk and remoteness
• 572 villages within 6 Kabupaten
Pemalang and Wonogiri (Central Java),
Palembang and Ogan Komering Ilir (South Sumatera),
Bandar Lampung and Central Lampung (Lampung)
• Raskin sample
identical to
previous project’s
(Targeting II)
Stratification
24
• Treatment stratified by
– Kabupaten,
– Targeting II Treatment Group,
– Kecamatan and
– Urban to rural ratio of 2:3
What the randomization looks like
25
Treatment 1: Raskin ID Card
Treatment 2: Raskin ID Card + Enhanced Socialization
Control: No treatment
26
Treatment and Control villages were statistically
equivalent prior to the pilot
Baseline data from February 2012
Control CardsN = 5643
03
69
By Treatment Status
Raskin purchased (in kg)
IMPLEMENTATION:
TREATMENT
27
Implementation: treatment
28
ID how to implement
treatment on ground
Establish strong internal controls
Coordination with
government
Prep, launch treatment
(cards, enh soc) Visit field
Implementation Plan
29
• Identify how to implement treatment on ground
– Develop detailed implementation plan with facilitation
specialist and government
– Consider scale up: utilize government’s vendors of choice
• Establish strong internal controls
– Random checks on card printing
– Recruitment standards (for facilitators)
– Standardized forms to document treatment implementation
– Clear reporting procedure
Implementation Plan
30
• Coordination with government – Jakarta workshop for reps from province, kabupaten/kota
– Pre-field work coordination with regency/city, district, and village heads
• Prepare, launch treatment
– Penjadwalan, pelatihan, koordinasi logistik
• Visit field: qualitatively observe the responses to treatment
Treatment variations
31
Card Variations Standard
socialization Enhanced
Socialization
All beneficiary
Price Coupon Group 1 Group 2
No Coupon Group 3 Group 4
No Price Coupon Group 5 Group 6
No Coupon Group 7 Group 8
Bottom 10%
Price Coupon Group 9 Group 10
No Coupon Group 11 Group 12
No Price Coupon Group 13 Group 14
No Coupon Group 15 Group 16
Control (No card, no socialization)
32
Recap: Raskin ID Card Variations
Cards randomly varied by:
• Design: coupon or no coupon
• Content: price or no price
• Distribution: to all Raskin beneficiaries or only the
poorest 10%
Distribute to Beneficiaries
• Sept. to mid-Oct. ‘12
• 378 villages receive cards;
194 villages control
33
Cards villages randomly varied to receive:
• Standard Socialization: letter, DPM; or
• Enhanced Socialization:
• + 3 DPM per dusun
• + 3 information posters per dusun
• Socialization to village leaders
• Announcement through speakers
Recap: Raskin ID Card Variations
34
• Late-Sept. to mid-Nov. ’12
• The 378 villages that received cards
received socialization
• 186 villages: Standard Socialization
• 192 villages: Enhanced Socialization
• The 194 control villages did not receive
socialization
Recap: Raskin ID Card Variations
A facilitator explaining Raskin Cards to village leaders in OKI, Central Lampung
Challenges in treatment implementation
35
• Some treatment villages were
unsafe or very remote
• Many cards did not reach
beneficiaries
Challenges in treatment implementation
36
• Facilitators did not receive enough Raskin
beneficiary posters
• During meetings, participants complained about other aspects of Raskin/ government programs
IMPLEMENTATION: DATA
COLLECTION
37
Implementation: data collection
38
Draft questionnaire
Internal Piloting
Coordination with
government
Prepare, launch data collection
teams
Quality checks, Field visits
Data Collection Plan
39
• Survey instruments: household and community survey
• Baseline—use previous project’s endline – Ensure control and treatment groups are statistically equivalent
• Identify respondents – Respondents:
• Raskin beneficiaries (poor )
• Raskin beneficiaries (very poor; 10% paling miskin)
• Non-Raskin beneficiaries
– Listing, to identify non-Raskin beneficiary households
– Use PPLS’10 data, to identify potential Raskin beneficiary households
Challenges in data collection
40
• Matching administrative data with field
– Human error, change in poverty status/address
• Accommodating changes to administrative areas (e.g.
pemekaran)
• Time constraints
• Obtaining sufficient HR to conduct data collection work
• Other concerns: how far behind can respondents
remember? How do we phrase/explain “Titik Distribusi”?
Etc.
Data Collection
41
Survey Data source Respondents surveyed
Data collected
Baseline 2011
Endline from Targeting II, a previous project
PKH eligibles, non-poor
Main purpose: ensure control and treatment groups statistically equivalent
Midline Oct-Dec’12
5,148 HH, through household surveys and community surveys (targeted to village head)
Mix between non-poor and Raskin beneficiaries (poor and very poor)
Amount and price of Raskin rice purchased, awareness on Raskin program, satisfaction levels with Raskin program, overall household consumption, subjective wealth/standing, etc.
Endline Mar-May ’13
6,292 HH, through HH and community surveys
Ibid Ibid
Challenges in program evaluation
42
• Attrition: when evaluators fail to collect data on
individuals who were selected as part of the original
sample
– ML: 9% replaced (418/4,572), EL: 9.8% (561/5,706)
– Replacement of respondents integrated into data collection
process
– Dropped 28 villages
43
ANALYSIS
44
Analysis
45
1. Write analysis plan 2. Write STATA do.file 3. Run data through program
4. Generate graphs 5. Analyze: Connect and reference with qualitative field observations
6. Disseminate findings
Higher satisfaction with Raskin by
eligible households in card villages
relative to the control
Increase in purchase of Raskin by
eligible compared to control
Eligible households in treatment villages
purchase more rice than those in control group
46
Control CardsN = 2770
02
04
06
08
0
Eligible
Control CardsN = 1362
Ineligible
By Treatment Status and Sample
HH bought raskin (pooled since oct)
Control CardsN = 2769
01
.93.8
5.7
7.6
Eligible
Control CardsN = 1362
Ineligible
By Treatment Status and Sample
Amount of raskin (kg, pooled since oct)
Can we generate this “subsidy received” graph?
1.4kg (25%) more rice
4.5pp
MID
LIN
E R
ESU
LTS
47
Eligible households in treatment villages
receive more in subsidy
No difference in household satisfaction
between treatment and control
Eligible households in treatment villages
purchase more rice
Eligible households in treatment villages
pay a smaller markup price
0.9 kg (19%) more rice
Rp. 77/kg (13%) smaller markup price
approx. Rp. 6000/HH/month (21%)
So did the ineligibles in the card villages also enjoy a smaller markup price? How much?
END
LIN
E R
ESU
LTS
Evolution of Results
48
Treatment → Outcome Midline Endline Cards → take-up, purchase, and subsidy
(+) (+)
Cards → HH satisfaction (+) eligible no effect
Enhanced Socialization (ES)
→ purchases, subsidy (+) eligible (+) eligible
ES → satisfaction (+) HHs (+) HHs
(−) leaders no effect leaders
Printing price → price markup (−) (−), only in ES
Coupons → subsidy (−) ineligible (+) eligible
Distribution to poorest 10%
→ rice purchases no effect (+), due to ES
49
• Distributing Raskin Cards improves the program
• Beneficiary take-up and Raskin purchases increase
• Price markup decreases
• On net, subsidy of Rp. 6000 for eligible with no
decrease for ineligibles
• Enhanced socialization improved beneficiary subsidy
and satisfaction
• Achieved with only 2-3 person-days of external
facilitation and 3 posters per dusun.
Insights for Policy
50
• Printing price on ID cards improved their effectiveness
• Short-run effect persists in the medium-run, when
combined with enhanced socialization
• Distributing cards only to the poorest 10% may be an
effective way to improve targeting
• Initially, poorest decile reported lower satisfaction and no
difference in take-up
• However, in the latest survey they report a higher subsidy
and higher satisfaction, in villages with enhanced
socialization
Insights for Policy (cont.)
51
• Coupon usage potentially effective with enhanced
socialization
• In the medium run, card variation effects depend on
enhanced socialization of the program.
• This applies to printing the price, coupons and
targeting the poorest 10%
Insights for Policy (cont.)
Policy Scale Up
52
Social Protection Cards (KPS)
53
• TNP2K scaled up Raskin Cards as Social Protection Card (KPS).
• As of June 2013, KPS cards have been distributed nationally to 15.5
million households (65.6 million people).
• The cards can be used access a temporary unconditional cash transfer
(BLSM); a cash transfer for poor students (BSM); and Raskin.
TERIMA KASIH
54
APENDIX
55
The Research Team
56
Principal
Investigators Design treatment,
randomization, analysis
Research Analyst Randomization, generate
respondent list, STATA
coding
Research Associate Develops implementation,
manages project (budget,
timeline, relationships,
reports), quality checks
Data Manager De-identifies data, generate
respondent list, cleans data,
matching, data quality
checks