RANDOMIZED CONTROL TRIALS RCTS KEY CONSIDERATIONS AND TECHNICAL

RANDOMIZED CONTROL TRIALS (RCTS): KEY CONSIDERATIONS AND TECHNICAL COMPONENTS Making Cents 2012 -Washington DC

Outline What makes an RCT Common Questions and Concerns Key Considerations Technical Components Evaluation Example

What makes an RCT Impact evaluation is one type of program evaluation Randomized Control Trial is one type of impact evaluation Monitoring &Evaluation Program Evaluation Impact Evaluation Randomized Evaluation

What makes an RCT Goal: to document the impact of a program or intervention Did the program change lives? What would have happened if the program hadn’t existed? Compare: what happened and what would have happened without the program aka the counterfactual

What makes an RCT Defining feature: random assignment of units individual beneficiaries, schools, health clinics, etc to treatment and control groups allows us to assess the causal effects of an intervention

Questions and Concerns Expensive Sample size Program started Ethics Technical capacity

Key Considerations What types of questions? Specific – targeted, focused: test a certain hypothesis Testable - has outcomes that can be measured Important - will lead to lessons that will affect the way we plan or implement programs

Key Considerations Right project to evaluate? Important, specific and testable question Timing--not too early and not too late Program is representative not gold plated Time, expertise, and money to do it right Results can be used to inform programming and policy

Key Considerations Right project to evaluate? Plan and randomize before the program starts Randomly assign who gets the program Evaluation design should occur with project design Need sampling frame on intended subjects Randomize after the baseline survey

Key Considerations Not the right project? Premature and still requires considerable “tinkering” to work well Too small of a scale to compare into two “representative groups” Unethical or politically unfeasible to deny a program while conducting evaluation, ie if a positive impact proven Program has already begun and not expanding elsewhere Impact evaluation too time-consuming or costly and therefore not cost-effective

Key Considerations Primary Outcome Intervention starts Impact ual act f r e t Coun Have to do randomization before program starts Timing

Technical Components: Sample size and unit of randomization Sample size determined size What effect by sizeeffect if feasible? Small program impact . 1 standard deviation change Moderate program impact . 2 greater standard deviation change Large program impact . 3 or greater standard deviation change

Technical Components: Effect sizes Define effect size that is “worthwhile” What is the smallest effect that should justify the program being adopted? If the effect is smaller than that, it might as well be zero: we are not interested in proving that a very small effect is different from zero In contrast, if any effect larger than that would justify adopting this program: we want to be able to distinguish it from zero Statistical significance vs policy significance

Technical Components: Effect sizes Program manager: “We think our program will increase average income by 20%” Researcher: “OK. That’s equivalent to an effect size of… (frantically calculates)… 0. 4 standard deviations… which means you need a sample size of… 2200 beneficiaries. ” (Time and money is spent on data collection, monitoring intervention, etc. One year passes. ) Researcher: “Well, we did not find a 20% increase in income. Maybe you should scrap the program. ” Program manager: “WAIT!! The program is still worthwhile if it only increases income by 10%!!” Researcher: “Ooops. We don’t have the power to detect that. We would have needed a sample size of 6000. . ” Punchline: Define effect size that is “worthwhile”… NOT

Technical Components: Effect sizes An effect size of… Is considered… …and it means that… 0. 2 Modest The average member of the treatment group had a better outcome than the 58 th percentile of the control group 0. 5 0. 8 Large Whoa…that’s a big effect size! The average member of the treatment group had a better outcome than the 69 th percentile of the control group The average member of the treatment group had a better outcome than the 79 th percentile of the control group 0. 4 0. 2 0 -4 -3 -2 -1 0 1 2 3 4 5 6

Technical Components: Sample Individual level If expecting . 2 effect size one intervention as compared to control . 8 power – 770 people . 9 power – 1040 people two interventions as compared to control . 8 power – 1150 people If expecting. 1 effect size one intervention as compared to control . 8 power – 3150

Technical Components: Sample Cluster level If expecting. 2 effect size - one intervention With Clusters of 10 people. 8 power – 115 clusters With Clusters of 20 people. 8 power – 80 clusters lf expecting. 2 effect size - two interventions With Clusters of 30 ppl. 8 power – 65 clusters Clusters of 50 ppl. 8 power – 56 clusters

Technical Components: Randomization Clients/beneficiaries are randomly assigned to receive the program or different program models Everyone has an equal chance of being assigned Potential Beneficiaries to all groups Randomization Treatment (receive program) Control (no or delayed program) The only difference between the two groups is whether they are assign to receive the new service

How do we normally select participants HQ Closest to the main roads? Biggest advocate for program services? Greatest need? Not served by other organizations?

Random assignment Randomization allows us to be sure we have a reliable and similar comparison group. When everyone had an equal chance of getting the program and randomization worked – we have our counterfactual. We know what would have happened Pre-program without the program. We know outcomes in blue villages are similar to outcomes in red villages.

But how can we be sure they are similar Income person, per day in leones, before the program 10, 057 5000 0 0 Treat Compare

What non random assignment might look like Income person, per day, leones HQ 12470 8990 5000 0 Treat Compare Blue villages don’t make a good counterfactual. These villages are different. They are better off.

Pre-program income – randomized Income person, per day in leones, before the program 10, 057 5000 0 0 Treat Compare

Post-program income – measure impact Income person, per day, after the program IMPACT 14590 10570 5000 0 Treat Compare Post-program and impact

Technical Components: Randomization Primary Outcome Intervention Impact ual act f r e t Coun Time Randomization is unique – it gives us this reliable counterfactual

Random Sampling vs. Random Assignment Random Sampling: each individual has the same probability of being included in the sample • • • Only survey/interview some households out of a community May select representative sample of villages out of district Can select random sample and conduct needs assessment § Random Assignment: each individual is as likely as any other to be assigned to the treatment or control group –

Random sampling and random assignment Randomly sample from area of interest (select some eligible participants/ villages) Randomly assign to treatment and control out of sample Randomly sample for surveys (from both treatment and control)

Mechanics of randomized assignment Need pre-existing list of all potential beneficiaries Many methods of randomization Public lottery In office private Selection from hat/bucket Random number generator Computer program code Staggered/phase-in Rotation Pulling names/communities out of a hat to select program beneficiaries is random assignment

Is random fair?

Multiple treatments: comparing programs Treatment 1 Treatment 2 Treatment 3

Phase-in design: slow program roll out 3 Round 1 2 Treatment: 1/3 Control: 2/3 3 3 1 Round 2 Treatment: 2/3 Control: 1/3 3 1 3 2 2 1 2 3 3 2 Randomized evaluation ends 2 1 2 2 3 1 2 1 3 1 1 2 2 2 1 3 1

Technical Components: Ethics Program proven to work Sufficient resources Doing harm Clear expectations Transparent process

Technical Components: Budget Impact evaluations require a significant budget line $30, 000 - $400, 000 The budget is influenced by Sample size Numbers of waves of data collection Logistical costs in-country Methods of data collection PDAs vs. paper-based tools Length of tool Qualitative vs. quantitative Staffing

Overall Goal: Evidence Based Policy Cost Benefit Analysis Needs Assessment Process Evaluation Impact Evaluation Opportunity costs of money Quantification of costs Benefits Cost-Benefit Analysis

Overall Goal: Evidence Based Policy 35

Example: Cash transfers for vocational training The Northern Uganda Social Action Fund (NUSAF)

NUSAF “Youth Opportunity Program” Cash grants of about $7000 per group ($377/person) Intended for acquiring vocational skills and tools Goals: 1. Raise incomes and employment 2. Increase community cohesion and reduce conflict 3. Build capacity of local institutions Experiment: Groups randomly assigned to receive the grant

Average age: 25 Average education: 8 th grade Average cash earnings: $0. 48/day PPP Average employment: 10 hours/week

Program allocated by lottery among eligible applicants 535 groups, with 18, 000 youth 265 treatment groups receive grant 270 groups assigned to a control group

Timeline of events 2006 Tens of thousands apply, hundreds of groups funded 2007 Funds remain for 265 groups in 10 districts Government selects, screens and approves 535 groups 2/2008 Baseline survey with 5 people per group Randomization at group level 7 -9/2008 Government transfers funds to treatment groups 10/2010 Mid-term survey commences roughly 2 years after transfer Effective attrition rate of 8% 2/2012 Next survey planned

The Results • Economic Outcomes: Improved Employment, Incomes and High Returns on Investment • Social Cohesion, Reconciliation and Conflict: Improvements for Men, Mixed Results for Women • Governance and Corruption: No Evidence of Widespread Leakage

Thank you QUESTIONS kbidwell@poverty-action. org Poverty-action. org