Practical Sampling for Impact Evaluations Cyrus Samii, Columbia University
Introduction How do we construct a sample to credibly detect a meaningful effect? Which populations or groups are we interested in and where do we find them? How many people/firms/units should be interviewed/observed from that population? How does this affect the evaluation budget? Warning! Goal of presentation is not to make you a sampling expert Goal is also not to give you a headache. Rather an overview: How do sampling features affect what it is possible to learn from an impact evaluation?
Outline 1. Sampling frame What populations or groups are we interested in? How do we find them? 2. Sample size Why it is so important: confidence in results Determinants of appropriate sample size Further issues Examples 3. Budgets
Sampling frame Who are we interested in? a) All SMEs? b) All formal SMEs? c) All formal SMEs in a particular sector? d) All formal SMEs in a particular sector in a particular region? Need to keep in mind external validity Can findings from population (c) inform appropriate programs to help informal firms in a different sector? Can findings from population (d) inform national policy? But should also keep in mind feasibility and what you want to learn Might not be possible or desirable to pilot a very broadly defined program or policy
Sampling frame: Finding the units we’re interested in Depends on size and type of experiment Lottery among applicants Example: BDS program among informal firms in a particular area Can use treatment and comparison units from applicant pool If not feasible (50,000 get the treatment), need to draw a sample to measure impact Policy change Example: A change in business registration rules in randomly selected districts To measure impact on profits, cannot sample all informal businesses in treatment and comparison districts. Will need to draw a sample of firms within districts. Required information before sampling Complete listing all of units of observation available for sampling in each area or group Tricky for units like informal firms, but there are techniques to overcome this
Outline 1. Sampling frame What populations or groups are we interested in How do we find them? 2. Sample size Why it is so important: confidence in results Determinants of appropriate sample size Further issues Examples 3. Budgets
Sample size and confidence Start with a simpler question than program impact Say we wanted to know the average annual profits of an SME in Dakar. Option 1: We go out and track down 5 business owners and take the average of their responses. Option 2: We track down 1,000 business owners and average their responses. Which average is likely to be closer to the true average?
Sample size and confidence: 5 firms 1,000 firms
Sample size and confidence Similarly, when determining program impact Need many observations to say with confidence whether average outcome of treatment group is higher/lower than in comparison group What do I mean by confidence? Minimizing statistical error Types of errors Type 1 error: You say there is a program impact when there really isn’t one. Type 2 error: There really is a program impact but you cannot detect it.
Sample size and confidence Type 1 error: Find program impact when there’s none Error can be minimized after data collection, during statistical analysis Need to adjust the significance levels of impact estimates (e.g. 99% or 95% confidence intervals) Type 2 error: Cannot see that there really is a program impact In jargon: statistical test has low power Error must be minimized before data collection Best method of doing this: ensuring you have a large enough sample Whole point of an impact evaluation is to learn something Ex ante: We don’t know how large the impact of this program is Low powered ex-post: This program might have increased firms’ profits by 50% but we cannot distinguish a 50% increase from an increase of zero with any confidence
Calculating sample size There’s actually a formula. Don’t get scared. Main things to be aware of: 1. Detectable effect size 2. Probability of type 1 and 2 errors 3. Variance of outcome(s) 4. Units (firms, banks) per treated area
Calculating sample size Detectable effect size Smallest effect you want to be able to distinguish from zero A 30% increase in sales, a 25% decrease in bribes paid Larger samples easier to detect smaller effects Do female and male entrepreneurs work similar hours? Claim: On average, women work 40 hours/week, men work 44 hours/week If statistic came from sample of 10 women & 10 men Hard to say if they are different Would be easier to say they are different if women work 30 hours/week and men work 80 hours/week But if statistic came from sample of 500 women and 500 men More likely that they truly are different
Calculating sample size How do you choose the detectable effect size? Smallest effect that would prompt a policy response Smallest effect that would allow you to say that a program was not a failure This program significantly increased sales by 40%. Great - let’s think about how we can scale this up. This program significantly increased sales by 10%. Great….uh..wait: we spent all of that money and it only increased sales by that much?
Calculating sample size Type 1 and Type 2 errors Type 1 Significance level of estimates usually set to 1% or 5% 1% or 5% probability that there is no effect but we think we found one Type 2 Power usually set to 80% or 90% 20% or 10% probability that there is an effect but we cannot detect it Larger samples higher power
Calculating sample size Variance of outcomes Less underlying variance easier to detect difference can have lower sample size
Calculating sample size Variance of outcomes How do we know this before we decide our sample size and collect our data? Ideal pre-existing data often ….non-existent Can use pre-existing data from a similar population Example: Enterprise Surveys, labor force surveys Makes this a bit of guesswork, not an exact science
Further issues 1. Multiple treatment arms 2. Group-disaggregated results 3. Take-up 4. Data quality
Further issues Multiple treatment arms Straightforward to compare each treatment separately to the comparison group To compare treatment groups requires very large samples Especially if treatments very similar, differences between the treatment groups would be smaller In effect, it’s like fixing a very small detectable effect size Group-disaggregated results Are effects different for men and women? For different sectors? If genders/sectors expected to react in a similar way, then estimating differences in treatment impact also requires very large samples
Who is taller? Detecting smaller differences is harder
Further issues Group-disaggregated results To ensure balance across treatment and comparison groups, good to divide sample into strata before assigning treatment Strata Sub-populations Common strata: geography, gender, sector, initial values of outcome variable Treatment assignment (or sampling) occurs within these groups
Why do we need strata? Geography example = T = C
Why do we need strata? What’s the impact in a particular region? Sometimes hard to say with any confidence
Why do we need strata? Random assignment to treatment within geographical units Within each unit, ½ will be treatment, ½ will be comparison. Similar logic for gender, industry, firm size, etc
Further issues Take-up Low take-up increases detectable effect size Can only find an effect if it is really large Effectively decreases sample size Example: Offering matching grants to SMEs for BDS services Offer to 5,000 firms Only 50 participate Probably can only say there is an effect on sales with confidence if they become Fortune 500 companies
Further issues Data quality Poor data quality effectively increases required sample size Missing observations Increased noise Can be partly addressed with field coordinator on the ground monitoring data collection
Example from Ghana Calculations can be made in many statistical packages – e.g. STATA, OD Experiment in Ghana designed to increase the profits of microenterprise firms Baseline profits 50 cedi per month. Profits data typically noisy, so a coefficient of variation >1 common. Example STATA code to detect 10% increase in profits: sampsi 50 55, p(0.8) pre(1) post(1) r1(0.5) sd1(50) sd2(50) Having both a baseline and endline decreases required sample size (pre and post) Results 10% increase (from 50 to 55): 1,178 firms in each group 20% increase (from 50 to 60): 295 firms in each group. 50% increase (from 50 to 75): 48 firms in each group (But this effect size not realistic) What if take-up is only 50%? Offer business training that increases profits by 20%, but only half the firms do it. Mean for treated group = 0.5* *60 = 55 Equivalent to detecting a 10% increase with 100% take-up need 1,178 in each group instead of 295 in each group
Outline 1. Sampling frame What populations or groups are we interested in How do we find them? 2. Sample size Why it is so important: confidence in results Determinants of appropriate sample size Further issues Examples 3. Budgets
Budgets What is required? Data collection Survey firm Data entry Field coordinator to ensure treatment follows randomization protocol and to monitor data collection Data analysis
Budgets How much will all of this cost? Huge range. Often depends on Length of survey Ease of finding respondents Spatial dispersion of respondents Security issues Formal vs informal firms Required human capital of enumerator Et cetera…. Firm-level survey data:$40-350/firm Household survey data: $40+/household Field coordinator: $10,000-$40,000/year Depends on whether you can find a local hire Administrative data: Usually free Sometimes has limited outcomes, can miss most of the informal sector
Summing up The sample size of your impact evaluation will determine how much you can learn from your experiment Some judgment and guesswork in calculations but important to spend time on them If sample size is too low: waste of time and money because you will not be able to detect a non-zero impact with any confidence If little effort put into sample design and data collection: See above. Questions?