PUBLIC SECTOR ECONOMICS SUMMARY
WEEK 1: INTRODUCTION
Public Economics is the study of the role of the government in the economy.
This involves 4 questions:
- When should the government intervene in the economy?
- How might a government intervene?
- What are the effects of those interventions on economic outcomes?
- Why do governments choose to intervene in the way that they do?
When should the government intervene in the economy?
- The competitive market equilibrium is efficient
- The government should intervene when there are market failures that leads to
inefficiencies.
o When markets are not competitive
o When consumers are boundedly rational
o When there is information is asymmetries, moral hazard, externalities
- Redistribution may be necessary if we also care about equality (trade-off with
equality and efficiency in the economy)
How might governments intervene?
- Influence the price mechanism through taxes or subsidies
- Restrict or mandate transactions
- Public provision of goods and services
- Public financing of private provision
What is the effect of interventions on outcomes?
- The field of different policy options it's called empirical public economics
o Direct effects: effects that arise because no agents change behavior
o Indirect effects: effects that arise because of behavioral changes.
- Both have to be taken into account and for the second one we also need insights
from behavioral economics.
Why do governments intervene in the way that they do?
- This is different from how government should act.
- Positive question (instead of normative like the 2nd one).
- Finding out what people want and what the average preferences are is difficult.
governments may have different preferences as well.
- How government make public policy decisions = political economy.
- This helps to enter answer why some policies are different across countries.
Importance of public sector economics
- Public spending is a large share of GDP and it doesn’t get less important.
- Public economics includes all kind of topics that we often see in the news.
o Health insurance and policies
o Education
o Pension systems
o Labor economics
o Fiscal (de)centralization
To study the effects of certain interventions, we need theoretical and empirical tools.
Theoretical tools
, - Decreasing “huurtoeslag” (subsidy towards your rent if you earn a low income) in the
Netherlands to encourage people to work more?
- We usually rely on utility-maximizing behavior of agents.
- Indifference curve: all combinations of consumption baskets which generate the
same utility. The more to the North-East, the higher the utility (utility functions are
convex since it represents diminishing marginal returns).
- We cannot get the highest utility possible always; we are bounded by a budget,
which is called the budget constraint. The intersection of the utility function and the
budget constraint (tangency condition) should be the point to which we are striving.
- Huurtoeslag working more means more consumption (blue lines). But… when the
huurtoeslag is implemented, the budget constraint goes up a little: it is like you get
money for free.
-
- The likelihood of corner solutions is bigger (we move more to the right-lower corner).
- A large discontinuity almost always leads to a corner solution.
- When the huurtoeslag is decreased/slashed, people start working more (less leisure,
so move to the left).
- We can calculate how much labor supply changes and look at the effects on welfare
as well (we need elasticities).
Empirical tools
- When theory doesn’t give us good and accurate predictions, we need to study how
people actually behave, without having too make so much assumptions.
- We can identify the mechanisms of observed effects.
- Correlation does not always imply causation people often think that when two
things move together, they are causally identified, but this is just correlation
causation fallacy. When 2 variables move together, it doesn’t always imply that the
movement of one variable causes the movement of the other variable.
- 3 reasons why correlation might be spurious (endogeneity = term for violating
assumption that the error term is uncorrelated with the independent variable):
o Reverse causality: Y causes X as well.
o Chance: by randomness, 2 variables might be related
o Omitted-variable bias: unobserved variable Z causes both X and Y. biased
and inconsistent estimates. We often lack a counterfactual and therefore we
cannot always correctly see the causal relationship.
- 3 solutions to this:
o Randomized Control Trial (RCT): random subset of statistical units is going to
be examined.
OLS is consistent and unbiased when assignment to control group is
random (and observations are large enough), because the zero-
conditional mean assumption holds.
Policy makers will not be happy to do a treatment among some
people, but not on others (may make them less popular by the public)
, o Natural (or quasi)-experiment: changes in economic environment generate
nearly identical treatment and control groups (someone else randomized for
us).
Evaluated ex post without help of policy makers
We need more involved statistical tools (is treatment really random?)
o Statistical tool for natural experiments:
IV: a variable Z partly determines X but does not directly influence Y
(X is endogenous to Y).
First stage: estimate what X should look like based on changes in Z.
Second stage: use these X values to see how it affects Y.
Assumptions: Z should be correlated with X (good predictor for X), but
Z should not be correlated with any other unobserved terms (exclusion
restriction).
Second tool = Difference-in-differences (DiD).
Compare time series of treated units before and after the treatment
with time series of untreated units before and after treatment.
Also, if all units are treated but at different times.
Parallel trends assumption = in absence of treatment, both treated
and untreated units would have seen the same evolution.
No other events occurred at the same time as treatment.
Third tool: Regression discontinuity design (RDD)
Compare statistical units close to a cut-off, as they should be very
similar in terms of background characteristics.
Landing above or below the cut-off involves uncertainty and can’t be
targeted easily.
Assumption = treatment assignment at the cut-off is as good as
random.
o This all delivers us effects of one particular treatment.
o Structural estimation = estimate parameters from theoretical model
Can predict effects for other treatments
But it relies on a model with assumptions.
Tutorial: more information regarding DiD and regression discontinuity:
DiD
- In 1992, the minimum wage in New Jersey rose from $4.25 to $5.05.
- If you want to establish the causal labor market effects of such a policy change, you
want to estimate: E(difference) = E(U1 – U0).
- where U1 denotes unemployment when the minimum wage rises, and U0 denotes
unemployment had the minimum wage stayed constant.
- The problem here is that we do not have a counterfactual (we do not know U0) in
New Jersey. We need to find a different state that did not have an increase in the
minimum wage, but that is different in characteristics to New Jersey Pennsylvania.
- We can’t simply compare differences across locations, because we cannot isolate
the causal effect (M) will create a selection bias since the characteristics already
differ before treatment.
- We can’t simply compare differences across time, because the time effect is not
eliminated temporal bias (M cannot be isolated as well).
- Idea is to combine both approaches to eliminate both these biases.
- Compare time differences of Pennsylvania with New Jersey M is isolated now.
, - We can also do a regression.
- DiD rests on 2 assumptions:
o T is the same across all units (countries would have been identical in the
absence of the treatment) = parallel trends assumption untestable because
we cannot observe the outcome of the treated country had it not been treated.
o There are no time-variant state-specific unobservables (i.e. nothing
unobserved varies across time in NJ that also determines unemployment).
o We need to verify by reasoning that the parallel trends assumption is
satisfied. We can argue that if the data is more similar before treatment, they
would also be more likely to be the same after the treatment (but, this is not
the same as testing it). Or, we can run separate regressions each time period
before and after treatment to detect time-varying effect (= event study).
- A problem with DiD can be that there is serial correlation: today’s unemployment rate
is correlated with last year’s invalid SE’s.
o Solve this with:
o Block bootstrapping standard errors: The block bootstrap is used when the
data, or the errors in a model, are correlated. The block bootstrap tries to
replicate the correlation by resampling blocks of data.
o Cluster standard errors at the group level (e.g. state): when number of
observations is small, you should use Wild Cluster Bootstrap
o Aggregate data into one pre- and one post-period
- There are four existential threats to identification when using DiD
o Non-parallel trends
o Compositional differences
o Long-term effects vs. reliability: the longer you study DiD, the more able you
are to detect long-term effects, and you are more vulnerable to picking up
diverging trends (trade-off).
o Functional form dependence: try out different models
- Non-parallel trends: treatment is endogenous
o Event study can solve it
o Or, falsification tests: repeat on different control group.
o Time-variant heterogeneity: alternative control group use triple D
specification. Compare DD to other variable that suffers from the same time-
variant heterogeneity. This isolates the causal treatment effect (when we
make some additional assumptions).
- Compositional differences
o Repeated cross-sections (instead of panel data) can risk changes in
composition of observations.
Regression discontinuity design
- When there is selection into treatment, comparing treatment and control may not
yield causal effects if treatment status is correlated with anything observed or
unobserved that correlates with the outcome variable.
- We need to find a discontinuity (a jump in probability of treatment assignment as you
move along some running variable).
WEEK 1: INTRODUCTION
Public Economics is the study of the role of the government in the economy.
This involves 4 questions:
- When should the government intervene in the economy?
- How might a government intervene?
- What are the effects of those interventions on economic outcomes?
- Why do governments choose to intervene in the way that they do?
When should the government intervene in the economy?
- The competitive market equilibrium is efficient
- The government should intervene when there are market failures that leads to
inefficiencies.
o When markets are not competitive
o When consumers are boundedly rational
o When there is information is asymmetries, moral hazard, externalities
- Redistribution may be necessary if we also care about equality (trade-off with
equality and efficiency in the economy)
How might governments intervene?
- Influence the price mechanism through taxes or subsidies
- Restrict or mandate transactions
- Public provision of goods and services
- Public financing of private provision
What is the effect of interventions on outcomes?
- The field of different policy options it's called empirical public economics
o Direct effects: effects that arise because no agents change behavior
o Indirect effects: effects that arise because of behavioral changes.
- Both have to be taken into account and for the second one we also need insights
from behavioral economics.
Why do governments intervene in the way that they do?
- This is different from how government should act.
- Positive question (instead of normative like the 2nd one).
- Finding out what people want and what the average preferences are is difficult.
governments may have different preferences as well.
- How government make public policy decisions = political economy.
- This helps to enter answer why some policies are different across countries.
Importance of public sector economics
- Public spending is a large share of GDP and it doesn’t get less important.
- Public economics includes all kind of topics that we often see in the news.
o Health insurance and policies
o Education
o Pension systems
o Labor economics
o Fiscal (de)centralization
To study the effects of certain interventions, we need theoretical and empirical tools.
Theoretical tools
, - Decreasing “huurtoeslag” (subsidy towards your rent if you earn a low income) in the
Netherlands to encourage people to work more?
- We usually rely on utility-maximizing behavior of agents.
- Indifference curve: all combinations of consumption baskets which generate the
same utility. The more to the North-East, the higher the utility (utility functions are
convex since it represents diminishing marginal returns).
- We cannot get the highest utility possible always; we are bounded by a budget,
which is called the budget constraint. The intersection of the utility function and the
budget constraint (tangency condition) should be the point to which we are striving.
- Huurtoeslag working more means more consumption (blue lines). But… when the
huurtoeslag is implemented, the budget constraint goes up a little: it is like you get
money for free.
-
- The likelihood of corner solutions is bigger (we move more to the right-lower corner).
- A large discontinuity almost always leads to a corner solution.
- When the huurtoeslag is decreased/slashed, people start working more (less leisure,
so move to the left).
- We can calculate how much labor supply changes and look at the effects on welfare
as well (we need elasticities).
Empirical tools
- When theory doesn’t give us good and accurate predictions, we need to study how
people actually behave, without having too make so much assumptions.
- We can identify the mechanisms of observed effects.
- Correlation does not always imply causation people often think that when two
things move together, they are causally identified, but this is just correlation
causation fallacy. When 2 variables move together, it doesn’t always imply that the
movement of one variable causes the movement of the other variable.
- 3 reasons why correlation might be spurious (endogeneity = term for violating
assumption that the error term is uncorrelated with the independent variable):
o Reverse causality: Y causes X as well.
o Chance: by randomness, 2 variables might be related
o Omitted-variable bias: unobserved variable Z causes both X and Y. biased
and inconsistent estimates. We often lack a counterfactual and therefore we
cannot always correctly see the causal relationship.
- 3 solutions to this:
o Randomized Control Trial (RCT): random subset of statistical units is going to
be examined.
OLS is consistent and unbiased when assignment to control group is
random (and observations are large enough), because the zero-
conditional mean assumption holds.
Policy makers will not be happy to do a treatment among some
people, but not on others (may make them less popular by the public)
, o Natural (or quasi)-experiment: changes in economic environment generate
nearly identical treatment and control groups (someone else randomized for
us).
Evaluated ex post without help of policy makers
We need more involved statistical tools (is treatment really random?)
o Statistical tool for natural experiments:
IV: a variable Z partly determines X but does not directly influence Y
(X is endogenous to Y).
First stage: estimate what X should look like based on changes in Z.
Second stage: use these X values to see how it affects Y.
Assumptions: Z should be correlated with X (good predictor for X), but
Z should not be correlated with any other unobserved terms (exclusion
restriction).
Second tool = Difference-in-differences (DiD).
Compare time series of treated units before and after the treatment
with time series of untreated units before and after treatment.
Also, if all units are treated but at different times.
Parallel trends assumption = in absence of treatment, both treated
and untreated units would have seen the same evolution.
No other events occurred at the same time as treatment.
Third tool: Regression discontinuity design (RDD)
Compare statistical units close to a cut-off, as they should be very
similar in terms of background characteristics.
Landing above or below the cut-off involves uncertainty and can’t be
targeted easily.
Assumption = treatment assignment at the cut-off is as good as
random.
o This all delivers us effects of one particular treatment.
o Structural estimation = estimate parameters from theoretical model
Can predict effects for other treatments
But it relies on a model with assumptions.
Tutorial: more information regarding DiD and regression discontinuity:
DiD
- In 1992, the minimum wage in New Jersey rose from $4.25 to $5.05.
- If you want to establish the causal labor market effects of such a policy change, you
want to estimate: E(difference) = E(U1 – U0).
- where U1 denotes unemployment when the minimum wage rises, and U0 denotes
unemployment had the minimum wage stayed constant.
- The problem here is that we do not have a counterfactual (we do not know U0) in
New Jersey. We need to find a different state that did not have an increase in the
minimum wage, but that is different in characteristics to New Jersey Pennsylvania.
- We can’t simply compare differences across locations, because we cannot isolate
the causal effect (M) will create a selection bias since the characteristics already
differ before treatment.
- We can’t simply compare differences across time, because the time effect is not
eliminated temporal bias (M cannot be isolated as well).
- Idea is to combine both approaches to eliminate both these biases.
- Compare time differences of Pennsylvania with New Jersey M is isolated now.
, - We can also do a regression.
- DiD rests on 2 assumptions:
o T is the same across all units (countries would have been identical in the
absence of the treatment) = parallel trends assumption untestable because
we cannot observe the outcome of the treated country had it not been treated.
o There are no time-variant state-specific unobservables (i.e. nothing
unobserved varies across time in NJ that also determines unemployment).
o We need to verify by reasoning that the parallel trends assumption is
satisfied. We can argue that if the data is more similar before treatment, they
would also be more likely to be the same after the treatment (but, this is not
the same as testing it). Or, we can run separate regressions each time period
before and after treatment to detect time-varying effect (= event study).
- A problem with DiD can be that there is serial correlation: today’s unemployment rate
is correlated with last year’s invalid SE’s.
o Solve this with:
o Block bootstrapping standard errors: The block bootstrap is used when the
data, or the errors in a model, are correlated. The block bootstrap tries to
replicate the correlation by resampling blocks of data.
o Cluster standard errors at the group level (e.g. state): when number of
observations is small, you should use Wild Cluster Bootstrap
o Aggregate data into one pre- and one post-period
- There are four existential threats to identification when using DiD
o Non-parallel trends
o Compositional differences
o Long-term effects vs. reliability: the longer you study DiD, the more able you
are to detect long-term effects, and you are more vulnerable to picking up
diverging trends (trade-off).
o Functional form dependence: try out different models
- Non-parallel trends: treatment is endogenous
o Event study can solve it
o Or, falsification tests: repeat on different control group.
o Time-variant heterogeneity: alternative control group use triple D
specification. Compare DD to other variable that suffers from the same time-
variant heterogeneity. This isolates the causal treatment effect (when we
make some additional assumptions).
- Compositional differences
o Repeated cross-sections (instead of panel data) can risk changes in
composition of observations.
Regression discontinuity design
- When there is selection into treatment, comparing treatment and control may not
yield causal effects if treatment status is correlated with anything observed or
unobserved that correlates with the outcome variable.
- We need to find a discontinuity (a jump in probability of treatment assignment as you
move along some running variable).
