Development Papers Flashcards
Berg and Zia
Entertainment media to see impact on financial decisions, soap Opera Scandal
Methodological challenges
Counterfactual group, impossible to restrict everyone to watch. Financial incentives to watch another programme during same time
helps also lower survey attrition
possible greater attention (not good thing)
results
- Increase likelihood of borrowing through fromal channels
- borrow for productive purpose
- lower gambling
- short term effect on usage of debt helpline
depature from classical literature on financial education (useless)
Emotions
emotional connection to chrcaters play a role, knowledge outcomes related to main protagonist higher than smaller roles
key difference length of exposure
post experiment group discussions, people recall protagonist more often
Tutorial1
Symmetric Encouragement Design
Give financial incentives to both
Treatment : more people watch Scandal
Control : more people watch Muhvango
high compliance
min exposure to both (same on TV)
min spillovers (2 streets between households)
Stratified randomization
example: you want same amount of men and women in control and treatement (balanced groups) -> u do stratified randomization
Dupas
See if for experienced goods a one-off subsidy could also boost long-run adoption through learning, more advanced ben-nets
learning might spread to others
Problems with subsidies
- sunk cost: people who pay more, use it more
- screening effect: if free also people with zero WTP get product
- acnhoring: take previous price as good unwilling to pay higher
if product learning effects this might change
Results
- Low subsidy beggining (increase short-run adoption) -> learning -> get it also in phase 2 increase future WTP
- Higher exposure (spatial to nighbors with nets) -> positive spill over effect
- no evidence phase1 price taken anchor phase 2 price
possible with bad learning (water disinfectant) results change
Tutorial2
Spillovers
- Direct : lower chance of infection
- Indirect: lower chance of transimssion
SUTVA when person i does not influence person j (not hold)
We need to consider vriable spillover in regression
Fixed effects
captures variations between areas of houshold location
school farm ecc all under one vector v
in regression estimate captures area specific effects, both observed and unobserved
Banerjee Duflo (education)
Hiring young women to help poor perfomring students, increase in learning outcomes. After one year initial gains remained but they fade out
Education issues
- High suplly, low attendance
- Very low quality and scarce results
first generation learners (stuff just too difficult to begin with)
additional teachers, school books don’t do shit
Intervention
- Balsakhi Women from community to help with basic skills
- Computer learning program
Tutorial3
Clustered standard errors
- assumption for OLS : i.i.d error terms
- Often not like this, students in same grade are more like each other than from same grade but another school (unobserved errors correlated)
- standard errors not accurate (no impact on coefficient)
- Cluster-level at same level of randomization
Fixed effects different (capture heterogeneity between diff. groups)
Results
positive effect (short-run), waker students gained the most (higer test scores), also computer thing effective.
Normalized difference
they do this in the paper for test results
can distort treatment effects, inflates effectiveness of intervetion done one more homogeneous groups (smaller standard deviation)
why still use it ?
Allows for easy comaparbility (meta-analysis)
cost-effective calculations
Kate Orkin
Bernard et al. , (Aspirations)
documentary of successful peope in etiopia to influence aspirations of people. 5 years later people increase future oriented investments and in children’s education
Experiment
Treatment watched doc on role models
Control some bullshit entertainment
After 5 years
Results
- higher investments in agriculture ecc
- extra working hours
- spend more on children education
- improved lifestyle (housing, durable goods )
behaviour starte quick 6 month after
tutorial4
Ancova
Estimate ture effect of treatment
** include baseline outcome variable** why ?
persistence ( outcome variable aspiration, more or less the same as yesterday)
past outcomes affect the outcome variable ( dream big as a kid, aspirations inherently higher)
decreases its standard error
no change in treatement estimate
Randomized Saturation Design
Does treatment intensity lead to more spillovers, if we expect no change estimates for HI and LI should be the same
otherwies HI > LI
in paper no spillovers
Banerjee Duflo (Microcredit)
Lot of backlash of micro credit (farmes debt pressure suicide), need to estimate average effect of microcredit ->
evalutation of effect of group-lending model.
Results
Households borrow more from microcredit institutions, informal declines (less demand than expected)
Consumption no difference but more buygin durable goods.
Not more likely to be entrepeuners, more investing in already business.
Increase average profit -> driven by upper-tail businesses.(95th percentile)
After 3 years treatment more assets (also 85th percintile increas in profits)
Accorgimenti
- Low take-up , lowers power and precision, impact of microcredit driven by marginal borrowers
- evaluation run in a context of high economic growth
- for-profit microfinance model
- no complementary srvices (traning)
- marginal neighborhoods
Reassurance similar results in 5 other studies in diff. contexts
tutorial5
Differential Attrition
Participants drop out of studies (happens natural)
but if Treatment housholds morke likely to drop out problem.
Remaining treatment and control not comparable (bias treatment effects)
Run regression (outcome variable attrition) see if estimate of being in treatment significantly different from 0
solutions to Differential Attirtion
- sampling weights: assign lower weight to obs. with charactstc which reduces attrition
- Lee bounds : range for treatment effects, random assingment to treatment, monotonicity (treatement affects attrition in one way)
Outliers
Average results driven by execptions (outliers), treatemtn effect super strong only on one person.
Deal with outliers
- TRIM : delete outliers (lose obs. reduce statistical power)
- Winsorizing: replace outliers by percentile (raggruppa tutto sotto 10th percentile at 10th percentile)
- Quantile treatment: celebrate outliers, divide into smaller groups, re-estimate regression for each group (concern: sample per quantile very small)
Lowe (caste integration)
Man from different casts in cricket teams, homogenous caste teams and mxied to see collaborative and adversial intergroup contact .
prejudice depnd on different type of contact(contact hypothesis) , lacks evidence
results
intuitively, cross-caste integration more likely hostile with adversierie than with teammates.
both contacts reduce ability-based statistical discrimination
collaborative increases cross-caste trade.
no less effective when teams assigned to pay structure (increases within team competition)
Tutorial6
Type I and II errors
Type I : False positive (claiming intervention had effect when it didn’t)
Type II : False negative (failing to detect an effect when there is one)
Maximize statistical power for Type II error
Overlap between distributions, the more overlap the lower the power
1. Have a larger treatment effect
2. decrease variance of outcome variable
3. larger sample size
Linear regression
- Linearity of treatment effect (cam be not hold)
- extrapolation (line trattegiatache segue la shit, could be not true)
de Mel (Microenterprises)
Some firms super high interest rate for borrowing (evidence for high marginal returns), need to estimate returns to capital. Challenge in estimating this, optimal level of capital depends on entrepreneur ability, way around problem exploit excogenous shock (self-selection problem firms applying to credit). -> Randomized grants, shocks to capital stock for Sri Lankan microentreprises.
Results
4.6 to 5.3% per month returns, higher than market interest rates, different between firms.
returns Higher for More constrained entrepreneurs.
Higher for man than for women (although normally more credit-constrained)
tutorial7
Levels vs. Logs
logs dampen outliers
but also change interpretation of coeff (ricordati EOR fo eco)
il log è come fare * 100
Instrumental variable
relevance : treatment impacts capital
exclusion restriction: treatment impacts profit only via capital (not ture , increase in labor suplly which affects profits)
HOW TO SOLVE -> adjust profits only via capital (substract extra labor)
Lowes (Kuba Kingdom)
Empirical evidence of how institutions affect cultural norms (obey rules and laws), kuba kingdom good for natural experiment as establishement of kuba unrelated with previous pop or geo diff.
kingdom had good professional institutions. Comparing people with ancestors inside the kingdom to just outside
results
Kuba ancestry is associated with more rule breaking and more theft.
Consistent with the model of Tabellini -> Kuba state institutions crowd out internal norms of rule following.
Survey show Kuba parents believe it is less important to teach rule following values to their children than non-Kuba parents.
Tutorial8
Natural Experiment
- Treatment occurs naturally (as good as random)
treatment in the past, outcome in the present - Clear and credible counterfactual
* baseline balance previous pop culturally homogenous
* exogenous borders (rivers gave borders, stable and defined)
* similar external setting (now all sample people live in same city Kananga)
very hard to provide a good natural experiment
Potential confounders
- Selective migration (migration decision for Kuba ad non-kuba differ)
- different people surveyed
- Geography (geographical diff could affect cultural evolution)
- Trust in researchers, understanding of game
- Colonial and post-colonial (Kuba - and non_kuba treated differently)