Przeworski, Stokes, and Manin, 1999

Mandate view:

Elections serve to select good policy

Accountability view

Elections serve to hold governments accountable for their past actions

Taking stock

• Do we feel these mechanisms are plausible?

• What might be some issues with these characterizations of representation?

Reconsidering: Some potential pitfalls

Pitfalls in the Mandate View

• Is it plausible to think governments will do what they propose?

• Is it desirable?

• Politicians are not legally compelled to abide by their platform in any democratic system! Why?

Pitfalls in the Accountability View

• Idea is reward or punish depending on their performance

• Is it plausible to think citizens have enough information to evaluate politicians?

• What if politicians do not value getting reelected

The Vote: Our One Blunt Tool

• In the accountability view, voters are retrospective

  • They use the vote to punish

• In the mandate view, voters are prospective

  • They use the vote to select the best policy / best politicians

• In reality, voters want to do both: select good policy and punish bad behavior

• But we only have one vote. Can we achieve both goals?

Is Accountability Only About Elections?

• So far we’ve talked about voters holding elected politicians accountable

• But what about public services? Health clinics, schools, water systems?

• The people running these services are often not elected. Should they still be accountable?

Community-Based Monitoring

“Community-based monitoring, or social accountability, is an approach towards building accountability that relies on civic engagement where citizens and civil society organizations directly or indirectly participate in extracting accountability”

1. The State of the World

• In many poor countries, clinics are closed when they should be open

• Health workers are frequently absent

• Drugs and vaccines are misused, public funds are stolen

• Top-down monitoring often weak and ineffective

2. The Research Question

• Can bottom-up, community-based monitoring improve health service delivery where top-down supervision has failed?

3. The Hypotheses

4. Mechanism: The Information Barrier

H1: Information \(\rightarrow\) monitoring

“Provision of information on outcomes and performance improves citizens’ ability to challenge abuses of the system, since reliable quantitative information is more difficult for service providers to brush aside as anecdotal, partial, or simply irrelevant”

4. Mechanisms: The Organization Barrier

H2: Organization \(\rightarrow\) monitoring

“Exerting accountability (monitoring providers) is subject to potentially large free-rider problems. Elite capture further complicates the process of holding providers accountable”

5. The Research Design

• RCT across 50 communities in 9 districts in Uganda
• Treatment communities got:
  • Report cards on health facility performance (targets information barrier)
  • Facilitated community meetings to develop action plans (targets organization barrier)
• Control communities: nothing
• ~5,000 households surveyed per round

Balance

The Main Specification

\[y_{jd} = \alpha + \beta \, T_j + X_{jd}'\gamma + \theta_d + \varepsilon_{jd}\]

Results

Did monitoring increase? (H1 + H2)

• Treatment facilities adopted suggestion boxes, numbered waiting cards, posted patient rights
• Communities discussed clinic performance at local council meetings
• 70% of treatment facilities had at least one monitoring tool vs 16% of control

H3: Health Outcomes

Another Specification

B&S surveyed households before and after the intervention

They combine (“stack”) both survey rounds into one dataset and estimate:

\[y_{ijt} = \gamma \, POST_t + \beta_{DD}(T_j \times POST_t) + \mu_j + \varepsilon_{ijt}\]

What is \(T_j \times POST_t\) doing?

Conditional Hypotheses

• Sometimes we think the effect of one variable depends on the value of another variable

• Does studying more improve your grade? Probably. But does it help more if you slept well?

Example

Imagine you have data on how much your peers slept and how well they performed

Example

We fit a regular OLS, like we know how to do

summary(lm(score ~ study_hours, data = students))$coefficients |>
  round(3)

            Estimate Std. Error t value Pr(>|t|)
(Intercept)   52.523      2.733  19.219        0
study_hours    5.190      0.436  11.915        0

What Is an Interaction Term?

We add a product of the two variables to the regression:

\[\text{score} = \alpha + \beta_1 \, \text{study} + \beta_2 \, \text{sleep} + \beta_3 (\text{study} \times \text{sleep}) + \varepsilon\]

Example

Example

summary(lm(score ~ study_hours * slept_well, data = students))$coefficients |>
  round(3)

                          Estimate Std. Error t value Pr(>|t|)
(Intercept)                 50.067      1.158  43.247    0.000
study_hours                  2.904      0.185  15.678    0.000
slept_wellYes                4.815      1.605   3.000    0.003
study_hours:slept_wellYes    4.206      0.256  16.439    0.000

Reading the Output

How many values can slept_well take? Two: 0 (No) and 1 (Yes)

Reading the Output

Now plug in slept_well = 1:

\[\text{score} = \alpha + \beta_1 \, \text{study} + \beta_2 \cdot 1 + \beta_3 (\text{study} \times 1)\]

Reading the Output

summary(lm(score ~ study_hours * slept_well, data = students))$coefficients |>
  round(2)

                          Estimate Std. Error t value Pr(>|t|)
(Intercept)                  50.07       1.16   43.25        0
study_hours                   2.90       0.19   15.68        0
slept_wellYes                 4.81       1.60    3.00        0
study_hours:slept_wellYes     4.21       0.26   16.44        0

Reading the Output

• study_hours = 2.90: for students who did not sleep well, one more hour of studying raises their score by about 2.9 points on average
• slept_wellYes = 4.81: for a student who studied zero hours, going from sleeping bad to well raises their score by 4.8 points on average
• study_hours:slept_wellYes = 4.21: for students who slept well, one more hour of studying raises their score by an extra 4.2 points

Back to the Paper

\[y_{ijt} = \gamma \, POST_t + \beta_{DD}(T_j \times POST_t) + \mu_j + \varepsilon_{ijt}\]

• Outcomes change over time for everyone – people get sick, seasons change, the economy shifts. \(\gamma\) captures that

• But did outcomes change more in treatment communities? That is what \(\beta_{DD}\) tells us

• The interaction \(T_j \times POST_t\) separates what changed because of the intervention from what changed just because time passed

Difference-in-Differences

• When the interaction is between a group and time

• (plus a bunch of assumptions)

• this is a difference-in-differences

• More on this later