Transcript Econometric analysis

Econometric analysis
informing policies
UNICEF workshop, 13 May 2008
Christian Stoff
Statistics Division, UNESCAP, [email protected]
Outline
•
•
•
•
•
Causality in experiments
Confounding factors
Quasi experiments: Difference estimators
Difference-in-difference estimators
Possible questions
A quick intro: The ideal situation
• Causality means a specific action leads to a
specific, measurable consequence
• Ideal: Randomized controlled experiment
• Subjects are randomly allocated to either control
or treatment group
• ONLY difference between two groups is
treatment
The reality
• In practice experiments are rare
• Subjects are NOT randomly assigned so that
sorting out of other relevant factors is difficult
• Econometrics provides the tools for controlling
these other factors
Challenge: Confounding factors
• Regress test score on student-teacher ratio
• But what about number of students in class still learning
English? – Omitted variable?
• Omitted variable correlated with explanatory and dependent
variable (low student-teacher ratios -> high % English learners
-> bad scores)
• Thus a policy of increasing no. teachers may not increase test
scores because high English learners (%) are the real problem
• Solution: Control for differences in English learners (%), i.e.
regress test score on student-teacher ratio AND English
learners (%)
Limits to controlling these factors
• Many years of cross-country research
• However, countries often have such different settings
and the “causal” relationships are only specific to the
country and the time period
• Therefore in search of quasi or natural experiments
between units that are “not too different”
• Danger lies in…
– Possible correlation between error term and explanatory variable
(i.e. treatment not assigned at random)
– Teachers try especially hard in areas with programs
– General equilibrium effects: when program is enlarged additional
factors may arise (external validity)
Difference estimators:
Using MICS3
• Define unit of analysis (households, districts, provinces, countries)
• Selected units gone through policy program (i.e. treatment) AND
assignment was “as if” random
Yi   0  1 X i  ui
• If X i is binary, then no functional form assumption needed; it is
simply the difference in the conditional expectations
• If X i can take on multiple values, then the above regression
assumes linearity
• But often there are pre-treatment differences between control and
treatment group…
Difference-in-difference estimators:
Using MICS2 and MICS3
• Types of datasets: Cross-section, panel and time-series
• Includes observations on same units before and after experiment
Yi   0  1 X i  ui
ˆ
• OLS estimator  1 is the difference in the group means of Y
• Control for district-level context constant over time through fixed or
random effects or adjust standard errors for clustering
• Advantage over difference estimator: 1. More efficient; 2. Eliminates
pre-treatment differences
Some possible questions
•
•
•
Education research:
–
Study drop-out rates and relate it to child labour questions
–
Study effect of different child disciplining strategies (punishment,
praise, etc.) on a child’s “success” in school
–
Combine MICS data with GIS disaster data and study effect of
disasters on school attendance
–
Combine with policy data between 2000-2005 and evaluate the
effectiveness of policies aimed at promoting higher school attendance
Child health:
–
Effect of different fuel types for cooking on child-health indicators?
–
Effects of different types of access to water and sanitation on a child’s
probability of having diarrhoea or succeeding in school?
–
How does Vitamin A affect a child’s health?
–
Impact of different health service facilities
Adult’s knowledge and attitude towards violence:
–
What is the effect of having information access (TV or radio) on
knowledge about HIV or contraception?
–
What is the effect of having information access (TV or radio) on
education methods or attitudes towards domestic violence?