Presentation at "Impact Evaluation for Financial Inclusion" (January 2013)
CGAP and the UK Department for International Development (DFID) convened over 70 funders, practitioners, and researchers for a workshop on impact evaluation for financial inclusion in January 2013. Co-hosted by DFID in London, the workshop was an opportunity for participants to engage with leading researchers on the latest research methods of impact evaluation and to discuss other areas on the impact evaluation agenda.
Sukhwinder Arora - measuring progress in market development
Alexis Diamond - quasi experiments
1. An Introduction to Impact Evaluation
in Observational (Non-Experimental) Settings
Alexis Diamond
Development Impact Department
2. Goals for this Presentation
• To explain key differences between randomized experiments
(RCTs) and observational studies
• To briefly sketch some of the most important methods of
causal inference in observational studies, showing how they
might be applied to answer questions in access to finance
projects, and offering practical guidance on:
Matching (an estimator, or a tool for designing observational studies?)
Differences-in-differences
Encouragement design (Instrumental variable, or “IV” regression)
Regression discontinuity design
Synthetic control methods
2
3. Basic concepts
• Observational study: comparison of treated and control
groups in which the objective is to estimate cause and effect
relationships, without the benefit of random assignment.
Observational studies are also known as quasi-experiments or
natural experiments
• In a randomized experiment, random chance forms
comparison groups (treatment and control), making groups
comparable in terms of both measureable characteristics and
characteristics that cannot be measured.
• Generally, If assumptions are met, causal conclusions follow—
but generally only in randomized experiments do we KNOW
assumptions are met; otherwise, assumptions aren’t testable.
3
4. Experiments (RCTs) vs. observational studies
―That’s not an experiment you have there,
that’s an experience.‖
— Sir R. A. Fisher (England, 1890-1962)
Because of selection bias—the presence of confounders
Reflected in the controversy over Pitt/Khandker and Roodman/Morduch
4
5. Selection bias: ―perfect implementation‖
A microfinance project is reporting the ex-post impact
indicator $/day for participants and non-participants…
i Yi(observed) Treatment Status
1 5 Treatment
2 6 Treatment
3 4 Treatment
4 4 Control
5 2 Control
6 6 Control
5
6. Selection bias: ―perfect implementation‖
A microfinance project is reporting the ex-post impact
indicator $/day for participants and non-participants…
i Yi(observed) Treatment Status
1 5 Treatment
2 6 Treatment
3 4 Treatment
4 4 Control
5 2 Control
6 6 Control
Average for the treatment group: $5/day
6
7. Selection bias: ―perfect implementation‖
A microfinance project is reporting the ex-post impact
indicator $/day for participants and non-participants…
i Yi(observed) Treatment Status
1 5 Treatment
2 6 Treatment
3 4 Treatment
4 4 Control
5 2 Control
6 6 Control
Average for the treatment group: $5/day
Average for the control group: $4/day
7
8. Selection bias: ―perfect implementation‖
A microfinance project is reporting the ex-post impact
indicator $/day for participants and non-participants…
i Yi(observed) Treatment Status
1 5 Treatment
2 6 Treatment
3 4 Treatment
4 4 Control
5 2 Control
6 6 Control
Average for the treatment group: $5/day
Average for the control group: $4/day
Difference = +$1/day:
8
9. Selection bias: ―perfect implementation‖
How should one think about that result, +$1/day?
Does it mean the project has positive impact?
9
10. Selection bias: ―perfect implementation‖
How should one think about that result, +$1/day?
Does it mean the project has positive impact?
Impact on who?
i Yi Yi(1) Yi(0) Treatment Status Yi(1) – Yi(0)
1 5 5 ? Treatment ?
2 6 6 ? Treatment ?
3 4 4 ? Treatment ?
4 4 ? 4 Control ?
5 2 ? 2 Control ?
6 6 ? 6 Control ?
10
11. Selection bias: ―perfect implementation‖
Avg Treatment Effect for Treated (ATT) = 3
Avg Treatment Effect for Control (ATC) = -1
Avg Treatment Effect (ATE) = +3 – (-1) = 4
i Yi Yi(1) Yi(0) Treatment Status Yi(1) – Yi(0)
1 5 5 2 Treatment +3
2 6 6 3 Treatment +3
3 4 4 1 Treatment +3
4 4 3 4 Control -1
5 2 1 2 Control -1
6 6 5 6 Control -1
That simple $1/day difference we identified earlier = ATT + BIAS
11
12. Selection bias: Ignore it at your peril
Identifying impacts requires identifying Y(1) and Y(0) for the same units
BIAS can be positive/negative, big/small, observed/hidden…
i Yi Yi(1) Yi(0) Treatment Status Yi(1) – Yi(0)
1 5 5 ? Treatment ?
2 6 6 ? Treatment ?
3 4 4 ? Treatment ?
4 4 ? 4 Control ?
5 2 ? 2 Control ?
6 6 ? 6 Control ?
12
13. Observational studies: Are they credible? Yes, but…
A judgment-free method for dealing with problems of
sample selection bias is the Holy Grail of the evaluation
literature, but this search reflects more the aspirations of
researchers than any plausible reality…
—Rajeev Dehejia, ―Practical Propensity Score Matching‖
• Some have tried to set up tests for observational methods:
e.g., “Can method X (matching, regression, IV, etc.) recover
the true (experimental) benchmark?”
• Such efforts have generally failed to conclusively validate
observational studies.
13
14. Observational studies: Are they credible? Yes, but…
History abounds with examples where causality has
ultimately found general acceptance without any
experimental evidence…
The evidence of a causal effect of smoking on lung
cancer is now generally accepted, without any direct
experimental evidence to support it…
At the same time, the long road toward general
acceptance of the causal interpretation …shows the
difficulties in gaining acceptance for causal claims
without randomization.
—Guido Imbens, ―Better LATE than Nothing‖
14
15. Why bother with observational studies?
• Studies that start as perfect RCTs often end as broken RCTs,
not “gold-standard” RCTs. These broken RCTs may be better
than many observational studies, but there is no bright line
distinguishing broken RCTs from observational studies.
• Standard RCTs cannot address many important policy issues
(i.e., macroeconomic questions, or cases with general
equilibrium effects more broadly)
• Other issues are difficult to address with RCTs, setting up a
trade-off between rigor and relevance. What’s better—the
RCT in a lab setting, or the equivalent observational study?
• RCTs are often more expensive, time-consuming, and fragile
than alternatives—can be high risk and not always strategic.
15
16. More advantages of observational studies
• Sometimes you can use pre-existing data, which has time and
cost advantages (though there are clear trade-offs)
o Typical out-of-pocket time/cost of a World Bank RCT: > 1 year & $500K
o Occasionally they can be done cheaply and easily, especially in a place
like India (there are examples where it costs < $50,000)
o With administrative data, observational studies may have no (or trivial)
out-of-pocket costs, and be completed in days or weeks.
• Sometimes you want to apply observational methods to
experimental data
• Good for hypothesis-generation
• Avoids RCT’s ethical considerations
16
17. Methodology #1: Matching
You: My clients enjoy big impacts from our bank’s financing
Critic: Compared to whom? Where’s the control group?
You: Ok, I’ll go find one—and then you’ll see!
17
18. Methodology #1: Matching
You: My clients enjoy big impacts from our bank’s financing
Critic: Compared to whom? Where’s the control group?
You: Ok, I’ll go find one—and then you’ll see!
Eg: Boonperm/Haughten, “Thailand Village Fund” (2009)
Control 1
X2:
Age
Treated
Control 2
Control 3
X1: Education
18
19. Methodology #1: Matching
You: My clients enjoy big impacts from our bank’s financing
Critic: Compared to whom? Where’s the control group?
You: Ok, I’ll go find one—and then you’ll see!
Eg: Boonperm/Haughten, “Thailand Village Fund” (2009)
Control 1
X2: X2: Control 1
Age Age
Treated
Control 2 Treated
Control 2 Control 3
Control 3
X1: Education Rescale X1: Education multiplied by 2
19
20. Matching: Points to consider
• Matching is (unfortunately) as much art as science, and there
are more methodological varieties of matching than there are
flavors of ice cream
• Widespread agreement that matching is, at a minimum, a
useful pre-processing step to reduce model dependence.
Unfortunately, no consensus on balance tests/diagnostics.
• Hugely important benefit of matching is that it is performed
“blind to the answer”—comparing favorably with regression
• Matching helps with selection bias due to observed variables
(confounders)—it does not help with unobserved confounders.
For the latter, one can (and should) do sensitivity analysis.
20
21. Methodology #2: Differences-in-Differences (D-i-D)
You: My clients enjoy big impacts from our bank’s financing
Critic: Compared to whom? Where’s the control group?
You: Ok, I’ll go find one—and then you’ll see!
Critic: Too many unobservables. It’s a waste of time.
You: Well, can you assume my control group’s growth rate
(e.g., near zero), is a good proxy for the treatment
group’s counterfactual growth rate (without the loan?)
D-i-D: subtract one before/after difference from the other
Addresses observed confounders (regression assumptions) &
unobserved time-invariant confounders common to treatment
and control groups. See Kondo’s work in the Philippines (ADB).
21
22. Diffs-in-Diffs: Points to consider
Treated
Income
Estimated ATET
Treated Counterfactual
Pre-treatment Control
difference
Control
Before After
NOTE: Circles are observed, square (counterfactual) is unobserved (imputed).
22
23. Diffs-in-Diffs: Points to consider
• If matching is implausible, why would D-i-D be plausible?
Does the parallel trend assumption seem easier to believe?
• The parallel trend assumption must hold over the time
period, implying composition of two groups should remain
constant over time.
• D-i-D benefits from “placebo tests” run pre-treatment
23
24. Methodology #3: Encouragement Design
You: Well, can you assume my control group’s growth rate
(e.g., near zero), is a good proxy for the treatment
group’s counterfactual growth rate (without the loan?)
Critic: No, also not credible.
You: OK, how about a natural experiment?
Our FI established additional info kiosks in 100 villages
to encourage loan take-up—these villages were not
chosen at random, but it was ―practically‖ random.
The encouragement (“instrument”, assumed “as good as
random”) has an effect (for some) on probability of finance.
This method leverages this “exogenous” variation to overcome
potential bias from both observed and unobserved confounders.
24
25. Encouragement Design: Points to consider
• Encouragement design requires strong assumptions:
o Encouragement must really be random or almost random, and must
have no direct effect on impacts (only an indirect effect via treatment)
o The encouragement must NEVER discourage take-up (no defiers)
o Causal estimates restricted to “compliers” only… (Who?)
o Also, for credible results, encouragement had better be effective
• Strange quirk: different answers, from different models, can
all be “correct” because complier populations may differ
• Was popular, now more disparaged in observational work
• Again, sensitivity tests are available and should be run
25
26. Methodology #4: Regression Discontinuity Design
You: OK, how about a natural experiment?
Our FI established additional info kiosks in 100 villages
to encourage loan take-up—these villages were not
chosen at random, but it was ―practically‖ random.
Critic: I don’t buy it. Rollout was in fact strategic, not random.
You: Ok, I’ll try again. This bank always provides extra lines
of credit at great terms to customers with credit scores
above a certain threshold. Let’s compare results for
customers just above and below the threshold.
Treatment assumed as good as random at the threshold if the
discontinuity is sharp. RDD addresses observed and unobserved
confounders. What question will the RDD design above answer?
26
27. Regression Discontinuity Design: Points to consider
• Generally considered a very strong design: US Dept of
Education classifies it in the same category as RCT
• Only informative for those at the discontinuity threshold
• No “gaming” the threshold allowed (ideally, the threshold is
unknown to the subjects, or outside subjects’ control)
• Relatively low statistical power, requiring much larger
sample sizes than RCTs or other observational methods.
• Watch out for contamination by other treatments at the same
discontinuity
• Sensitivity tests available to probe plausibility of assumptions
27
28. Methodology #5: Synthetic control method
Critic: I don’t buy it. It must’ve been strategic, not random.
You: Ok, I’ll try again. This bank always offers extra lines of
credit at great terms to customers with credit scores
above a certain threshold. Let’s compare results for
customers just above and below the threshold.
Critic: I’m not interested in only a narrow set of borrowers.
You: Last try. How about we do an in-depth case-study of a
greenfield microfinance institution, asking about the
social welfare impact on the neighboring community?
Synthetic controls allows inference for a single treated unit.
This approach addresses observed and unobserved confounders.
28
29. Methodology #5: Synthetic control method
Estimating Average Impact on Household Consumption in a Single Village
1000000
800000
600000
400000
200000
1995 2000 2005 2010
year
Treated District (Kabil) Synthetic Control District
29
30. Synthetic controls: Points to consider
• Only method allowing for rigorous quantitative causal
inference for a single treated unit
• Enormous growth in popularity in last 5 years
• Particularly well-suited to case-studies exploring program
impacts at village/city/state/country level
• Requires time-series data and many control units
• Placebo tests are available to assess plausibility of critical
assumptions
30
31. Elaborate theories, multiple tests
When asked what can be done in observational
studies to clarify the step from association to
causation, Fisher replied: ―Make your theories
elaborate.‖ (Cochrane)
This is sage advice, but often misunderstood.
Fisher didn’t mean you should make your
theories and explanations complicated.
He meant, when constructing causal hypothesis,
envisage as many different consequences of its
truth as possible, and plan observational studies
to discover whether each holds.
• Creating/testing elaborate theories is particularly helpful for
indirectly testing for hidden biases (unconfoundedness).
31
33. Final thoughts
• Ex-ante, be clear as to standard of evidence (going to depend upon the
purpose of your inquiry, and who your audience is)
33
34. Final thoughts
• Ex-ante, be clear as to standard of evidence (going to depend upon the
purpose of your inquiry, and who your audience is)
• Also ex-ante, be clear re treatment, covariates, units, and assumptions.
34
35. Final thoughts
• Ex-ante, be clear as to standard of evidence (going to depend upon the
purpose of your inquiry, and who your audience is)
• Also ex-ante, be clear re treatment, covariates, units, and assumptions.
• Try to adjust for (eliminate) differences in observed characteristics
while remaining blind to the answer.
35
36. Final thoughts
• Ex-ante, be clear as to standard of evidence (going to depend upon the
purpose of your inquiry, and who your audience is)
• Also ex-ante, be clear re treatment, covariates, units, and assumptions.
• Try to adjust for (eliminate) differences in observed characteristics
while remaining blind to the answer.
• Run diagnostics/sensitivity tests for unobserved (hidden) bias
36
37. Final thoughts
• Ex-ante, be clear as to standard of evidence (going to depend upon the
purpose of your inquiry, and who your audience is)
• Also ex-ante, be clear re treatment, covariates, units, and assumptions.
• Try to adjust for (eliminate) differences in observed characteristics
while remaining blind to the answer.
• Run diagnostics/sensitivity tests for unobserved (hidden) bias
• Devise/test multiple“elaborate theories”. Invest in learning about the
substantive problem to be solved, and be skeptical of your own results.
37