Part 1

Draw a causal diagram, as in class. Include an outcome variable, a treatment variable, an observed confounding variable (control variable), an unobserved confounding variable, an instrumental variable, and a bad control variable.

alt

 

If I want to estimate the (linear) causal relationship between the treatment variable and outcome variable under the assumption that there are no unobserved confounders, what linear regression model should I estimate?

A treatment variable has a causal effect on the outcome if and only if there are no unobserved confounders, so under the assumption that there are none, an Ordinary Least Squares regression would suffice to find the causal relationship between the two variables.

 

Suppose I am concerned that there are unobserved confounders that affect both the treatment and the outcome, and as such I cannot assume that the relationship between them is causal. How should I go about identifying the causal effect?

If there are unobserved confounders that affect both the treatment and the outcome, a causal relationship may be identified by using controls. This could mean finding data that you believe to be closely correlated with the effect of the unobserved confounders and including it in the model. Another approach is finding a natural experiment which ideally acts as a sorting mechanism that is effectively random in that whether an item of interest is assigned to the control group or the treatment group does not depend on any factors that could influence the outcome.

 

Describe the two properties that an instrumental variable must have. What is the exclusion restriction? What is the relationship between an instrumental variable and a natural experiment?

An instrumental variable must (a) be correlated with the endogenous explanatory variable(s), i.e. the treatment, and (b) must not be correlated with the error term. The exclusion restriction is that the instrumental variable should have no effect on the outcome other than its effect through the treatment.

A natural experiment is a situation that essentially provides us with an implicit instrumental variable which is the result of a ‘natural’ (and hopefully) random occurrence, whereas typically instrumental variables are the product of intentional design and are influenced by the researcher. In both cases arguments to support the proposition of a causal relationship between the treatment and the outcome are necessary.

 

Part 2

Divide the following between your teammates (put a name next to each submission, one paper each): describe two natural experiments that researchers have used to examine the permanent income hypothesis and two natural experiments used to examine the fiscal multiplier. You may need to read the original papers. What data do they use? What is the natural experiment? What are their exclusion restrictions? Do you believe their exclusion restrictions? What are their findings?

1. Mafia and Public Spending: Evidence on the Fiscal Multiplier from a Quasi-experiment

by A. Acconcia, G. Corsetti, S. Simonelli (2014)

(Sam)

 

What data do they use?

The main data used is made up of measures of spending and output in 95 Italian provinces for each year between 1990 and 1999.

Data on the number and details of council dismissals in each province during 1991-2012 is also used.

What is the natural experiment?

The paper exploits the existence of an Italian law to curb mafia activity that gave the central government the power to remove elected officials in the event of evidence their decisions were influenced by a mafia and appoint external commissioners who would govern in their place for the next 18 months. The authors found that this would invariably lead to a subsequent contraction of local government spending, which was on average comparable in size to the change in fiscal variables in leading empirical analyses of multipliers.

What are their exclusion restrictions?

The experiment relies on the assumption that council dismissals resulting from the law had no effect on economic output other than through the impact of a reduction in public spending. The authors identified two ‘channels’ that could lead to this not being the case.

First, the impact of legal action on mafia activity could have a significant direct effect on output. The authors argue that they sufficiently control for this by including measures of the outcome of police investigation. However, they say that when these controls are omitted they find that legal action against the mafia has a positive effect on output, and therefore even if their controls are not sufficient then the multiplier effect they identify would have a downward, rather than upward, bias.

Second, it is possible that council dismissals have a negative effect on output due to a reduction in administrative productivity. However, it is shown that analysis of council dismissals that were unrelated to the anti-mafia law and did not necessarily correspond with a fiscal contraction showed no evidence of council dismissals leading to a drop in output.

Do you believe their exclusion restrictions?

The arguments presented are quite compelling. For each of the potential problems identified they provide not only a theoretical reason to believe the restriction to hold but also empirical backing. They also ensure that those analyses deal with problems they may have, such as the fact that police investigation would drop if mafia operations decreased during the period of the new council, meaning that they had to ensure the number of arrests was correlated with the scale of mafia activities. Another precaution they took was ensuring that there was no relationship between output and the council dismissals in that the output growth rate was either above or below average in preceding years.

One possible avenue they neglected was to discuss whether there was anything about the provinces in which the law was enforced other than having relatively high mafia activity that may have attracted the mafia and could also have had some effect on the size of the multiplier.

What are their findings?

The results of the first specification they used (which didn’t include lagged values of spending and output) gave the one-year multiplier a point estimate of 1.17 that was statistically significant at the 5% level, meaning a 1% decrease in spending was estimated to decrease output by 1.17%. The inclusion of dynamic effects in the model led to estimates of the multiplier being between 1.24 and 1.8. They also estimated the spillover effect of the spending shock on nearby provinces, finding no evidence.

 

2. Estimating Local Fiscal Multipliers

by Juan Carlo Suarez Serrato and Phillippe Wingender (2014)

(Raaj)

 

What data do they use?

Population Data:

They use contemporaneous county population estimates for over 3000 counties from 1970 to 2009. As there was no data released during census years using this methodology released in the census years, they replicate the methodology to create estimates. They do this so they can compare the estimates to the census counts for the census years.

Federal Spending Data:

They use annual data from the Consolidated Federal Funds Report (CFFR) to ascertain the geographical distribution of federal spending. The spending data is then disaggregated by agency and spending program, allowing them to restrict their analysis to the following categories: - Direct Payments to Individuals - Direct Pay- ments for Retirement and Disability - Grants (Medicaid transfers to states, Highway Planning and Construction, Social Services Block Grants, etc.) - Procurement and Contracts (both Defense and non-Defense) - Salaries and Wages of federal employees - Direct Loans.

Finally they exclude the Insurance and Guaranteed Loans categories on the basis that they represent liabilities and not actual spending.

Personal Income Data:

They take data on income, salaries and wages from the Bureau of Economic Analysis’ Regional Economic Information System (REIS). This dataset is collated from various other sources. In order to make this data comparable across counties they use income, earnings and employment in per capita terms. All dollar values are expressed in 2009 dollars calculated using the national Consumer Price Index, published by the US Bureau of Labour Statistics.

What is the natural experiment?

Serrato and Wingender (2014) attempt to estimate the local fiscal multiplier by measuring the causal impact of government spending on income and employment growth.

Because of the endogeneity of government spending - its correlation with the error term - it is difficult to draw an accurate causal relationship between it and income using an OLS model. For this reason they use an Instrumental Variable (IV) model to get a better estimation of the true size of the fiscal multiplier.

Government spending at the local level is allocated according to estimates of the local population size. The census takes place every ten years and as such, estimates of the local population size between census’ use a different estimation method. The two different methods of measurement/estimation produce different results and as such there is a significant discrepancy in the measured local population the year before the census and in the census year. The population revision the in census years are referred to as the census “error of closure”. These errors can be significant. For example, in 2000, the census counted 6.8 million more people than the estimated population level, based on the 1990 census.

As such, they use the “error of closure” to instrument the change in federal spending on the local level in the affected years.

What are their exclusion restrictions?

The exclusion restriction is that census shocks only affect local economic growth through the effect that it has on the allocation of federal funding. That is, alterations in the county population estimates only impact economic growth through the effect it has on the amount of federal funding that is spent within a county.

Do you believe their exclusion restrictions?

Serrato and Wingender (2014) address the following three issues which may inhibit the exclusion restriction.

  • There is no significant correlation between census shocks across nearby counties within a given region.
  • There is no serial correlation of shocks between counties. An example of a serially correlated shock could be the case if, for example, counties had significant influxes of migrant workers, which would affect both economic growth and the census shock. This provides evidence against the confounding factors that could drive variation across areas.
  • There is no positive correlation between the census shock and economic growth in the years before the error of closure should actually affect government spending in a county. They include this as the level of government spending - which is based on the population findings in the census - are not announced until two years after the census is conducted.

Their exclusion restrictions are believable as it is difficult to see how the instrument would affect income through means other than government spending, other than those which have been successfully addressed above.

What are their findings?

The estimates from their instrumental variables imply that government spending has a local income multiplier of 1.57 and an estimated cost per job of $30,000 per year. They also find that government spending has a positive spillover on neighbouring countries and that government spending has a higher impact in low growth areas.

This means that fiscal policy can be a viable lever to effect changes in income.

 

3. The Effects of Lottery Prizes on Winners and their Neighbors: Evidence from the Dutch Postcode Lottery

by Peter Kuhn, Peter Kooreman, Adriaan R. Soetevent, Arie Kapteyn (2011)

(Adjin)

What data do they use?

The data concerned with this natural experiment is that of the Dutch Postcode Lottery (PCL). The authors were able to get a grant from the Netherlands Organization for Scientific Research (NWO).

Their sample size consisted of 301 Non-0winning postcode participants and 223 winning postcode participants.

What is the natural experiment?

The natural experiment in this scenario is that there is an unbiased distribution of wealth to the participating individuals, seeing as a large amount of wealth is given to the winner, the authors can then test the permanent income hypothesis and see if consumption behavior changes.

The lottery wins in their episode amount to 12,500 Euros, which is equal to eight monthly average household incomes in the Netherlands. In line with the Permanent Income Hypothesis.

What are their exclusion restrictions?

The exclusion restrictions were that they controlled for codegroup fixed effects and for any differences in observed pre-lottery household characteristics(For example income) that are present despite randomization, which then would increase the statistical precision.

The authors also mention that even though the winning code is randomly determined, the purchase of lottery tickets through higher risk adverse individuals could hold some correlation between the winnings, the authors control for this with the number of tickets purchased.

They also mention that the lottery winners who did not reply to their surveys could reduce the statistical precision, so the authors control for this by relying on codegroup fixed effects.

Do you believe their exclusion restrictions?

Yes, i believe so - this gives us a good indication that the lottery winners are chosen randomly and have no bias effect such as previous income or number of tickets purchased. This gives us a good, unbiased conclusion on the consumer behavior after receiving a lottery winning and thereby showing us that it holds to the permanent income hypothesis.

What are their findings?

The authors conclude that according to their study’s, the effects of a lottery prize on winners are confined largely to cars and other consumer durables. They note that the permanent income model in which households adjust the timing of their durables purchases to smooth consumption is consistent with their findings.

The authors also find some encouraging fiscal policy news that ‘stimulus’ policies aimed specifically at durables (I.e Cars), may have substantial own effects, as well as significant social multiplier effects (I.e Neighbors increasing consumption as a result of the winnings through sale of used-cars or income transfers if family members occupy these neighboring homes).

 

4. The Economics Stimulus Payments of 2008 and the Aggregate Demand for Consumption

by Christian Broda and Jonathon A. Parker (2014)

(Ariel)

The data that Broda and Parker used came from Nielsen’s Consumer Panel. They looked at weekly purchases of almost 30 000 households from supermarkets, drugstores and mass-merchandise sectors.

What Broda and Parker set out to find in this experiment is whether consumer demand increases in response to receiving an Economic Stimulus Payment (ESP). In other words, what is the effect of a preannounced income increase on consumption growth? The researchers include a variable, which absorbs any seasonal or average changes in spending for each group of recipients separately in each week. Then they discuss the differences in the way people responded based on their liquidity.

Broda and Parker’s findings reject the Permanent Income Hypothesis as they reach the conclusion that there is some growth in household spending on receipt of an Economic Stimulus Payment. Moreover, they find that spending by household does not significantly change when they learn of the stimulus payments.

 alt

Forlorn Keynes is still forlorn