NBER WORKING PAPER SERIES

MECHANISM EXPERIMENTS AND POLICY EVALUATIONS

Jens Ludwig
Jeffrey R. Kling
Sendhil Mullainathan

Working Paper 17062

[Link]

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
May 2011

This paper is forthcoming in the Journal of Economic Perspectives as part of a symposium on field
experiments. For excellent research assistance we thank Laura Brinkman and Michael Reddy. We
thank Nava Ashraf, David Autor, Iwan Barankay, Jon Baron, Howard Bloom, Lorenzo Casaburi, Philip
Cook, Stefano DellaVigna, John DiNardo, Elbert Huang, Chad Jones, Lawrence Katz, Supreet Kaur,
John List, Stephan Meier, David Moore, Steve Pischke, Harold Pollack, Dina Pomeranz, David Reiley,
Frank Schilbach, Robert Solow, Tim Taylor and conference participants at the University of Pennsylvania’s
Wharton School of Business and the American Economic Association for helpful comments. For financial
support we thank the Russell Sage Foundation (through a visiting scholar award to Ludwig). Any errors
and all opinions are our own. The views expressed here are those of the authors, and should not be
interpreted as those of the Congressional Budget Office or National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been peer-
reviewed or been subject to the review by the NBER Board of Directors that accompanies official
NBER publications.

© 2011 by Jens Ludwig, Jeffrey R. Kling, and Sendhil Mullainathan. All rights reserved. Short sections
of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full
credit, including © notice, is given to the source.
Mechanism Experiments and Policy Evaluations
Jens Ludwig, Jeffrey R. Kling, and Sendhil Mullainathan
NBER Working Paper No. 17062
May 2011
JEL No. C93

ABSTRACT

Randomized controlled trials are increasingly used to evaluate policies. How can we make these experiments
as useful as possible for policy purposes? We argue greater use should be made of experiments that
identify behavioral mechanisms that are central to clearly specified policy questions, what we call
“mechanism experiments.” These types of experiments can be of great policy value even if the intervention
that is tested (or its setting) does not correspond exactly to any realistic policy option.

Jens Ludwig Sendhil Mullainathan

University of Chicago Department of Economics
1155 East 60th Street Littauer 208
Chicago, IL 60637 Harvard University
and NBER Cambridge, MA 02138
jludwig@[Link] and NBER
mullain@[Link]
Jeffrey R. Kling
Congressional Budget Office
and NBER
 I. INTRODUCTION

Randomized controlled trials are increasingly used to evaluate policies. To cite

one example, in 2002, the U.S. Department of Education founded the Institute of

Education Sciences with a primary focus on running experiments; its annual budget is

now about $700 million (U.S. Department of Education, 2010). This trend has been

spurred in part by numerous independent groups—the Coalition for Evidence-Based

Policy, the Campbell Collaboration, an international network of researchers hosted by the

Norwegian Knowledge Centre for the Health Services, the Poverty Action Lab, and

Innovations for Poverty Action—that promote policy experimentation. Others however

question the wisdom of this trend. A vigorous debate has arisen around the value of

experimental methods for informing policy (e.g., Angrist and Pischke, 2009, 2010;

Banerjee and Duflo, 2009; Deaton 2010; Heckman, 2010; Imbens, 2010). We argue this

debate has often been framed too narrowly on experimental versus non-experimental

methods. An important distinction between experimental methods has been overlooked.

Suppose a policy maker has already decided on using an experiment. She faces a

design problem. Given a fixed budget, how should she design her experiment to

maximize policy-relevant information? The answer seems obvious: replicate the policy as

it would be implemented at scale, and randomly assign units (people or sites of the sort

that would be targeted by the policy) to treatment and control conditions. The design

challenges involve selecting the most cost effective units of randomization and the data

collection strategies. We call the resulting experiments policy evaluations. In practice

most policy experimentation involves policy evaluations. Yet in some (practically

3
relevant) situations, these are not the best experiments to use—even if the sole goal is to

help inform policy decisions.

A simple example illustrates our point. Suppose the U.S. Department of Justice

(DOJ) wanted to help local police chiefs decide whether to implement “broken windows”

policing, which is based on the theory that police should pay more attention to enforcing

minor crimes like graffiti or vandalism because they can serve as a “signal that no one

cares” and thereby accelerate more serious forms of criminal behavior (Kelling and

Wilson 1982, p. 31). Suppose that there is no credibly exogenous source of variation in

the implementation or intensity of broken windows policing across areas, which rules out

the opportunity for a low-cost study of an existing “natural experiment” (Meyer, 1995,

Angrist and Pischke, 2009). To an experimentally-minded economist, the most obvious

next step goes something like: DOJ should choose a representative sample of cities,

randomly select half of their high-crime areas to receive broken windows policing (or

perhaps randomly assign half the cities to get citywide broken windows policing), and

carry out a traditional policy evaluation.

Now consider an alternative experiment: Buy a small fleet of used cars. Break the

windows of half of them. Park the cars in a randomly selected subset of neighborhoods,

and then measure whether more serious crimes increase in response. What might seem

like a fanciful example is actually the basic research design used in the 1960s study by

Stanford psychologist Philip Zimbardo that helped motivate the broken windows theory

(Kelling and Wilson, 1982, p. 31),1 which in turn led to the implementation of broken

windows policing at scale in New York City during the 1990s. One can of course perform

1
The same design was used more recently by a Dutch team for a study published in Science (Keizer et al.,
2008).

4
variants with other small crimes; for example, one could hire young men to wear the

standard-issue uniform for drug distribution (plain white t-shirt, baggy jeans, Timberland

boots) and have them loiter at randomly selected street corners. This mechanism

experiment does not test a policy: it directly tests the causal mechanism that underlies the

broken windows policy.

Which experiment would be more useful for public policy? Partly it’s an issue of

staging. Suppose the mechanism experiment failed to find the causal mechanism

operative. Would we even need to run a policy evaluation? If (and this is the key

assumption) the mechanism experiment weakened policy makers’ belief in broken

windows policing, then we can stop. Running the (far cheaper) mechanism experiment

first serves as a valuable screen. Conversely, if the mechanism experiment found very

strong effects, we might now run a policy evaluation to calibrate magnitudes. Or,

depending on the costs of the policy evaluation, the magnitudes found in the mechanism

experiment, and what else we think we already know about the policing and crime

“production functions,” we may even choose to adopt the policy straightaway.

Mechanism experiments more carefully incorporate prior knowledge and can be

designed to maximize information in the places where the policy maker needs to know

the most. In our broken windows example, suppose there is general agreement about the

list of minor offenses that might plausibly accelerate more serious crimes (that is, the list

of candidate mediating mechanisms M in Figure 1). Suppose (from previous work) we

also know the elasticity of minor offenses with respect to policing (P→M in Figure 1).

What policymakers do not know is the accelerator: by how much will reducing minor

offenses cascade into reducing other offenses. The mechanism experiment estimates the

5
parameter about which there is the greatest uncertainty or disagreement (M→Y in Figure

1). In contrast, a policy evaluation that measures the policy’s impact on serious crimes,

P→Y, also provides information about the crime accelerator, but with more noise

because it combines the variability in crime outcomes with the variability in the impact of

policing on minor crimes in any given city / year combination. With enough sample (that

is, money), one could recover the (M→Y) link. In a world of limited resources,

mechanism experiments concentrate resources on estimating the parameters that are most

decision relevant.

We argue that mechanism experiments should play a more central role in the

policy process. The broken windows example is not an isolated case: many policies have

theories built into them, even if they are sometimes just implicit. Often these theories can

be tested more cost effectively and precisely with experiments that do not mimic real (or

even feasible) policies. Our argument runs counter to the critique leveled by some

economists against the large-scale government social experiments of the 1970s and 1980s

for “not necessarily test[ing] real policy options” (Harris, 1985, p. 161). We argue that

some of these experiments, because they highlight mechanisms, could have far-reaching

policy value. Social scientists already value mechanism experiments because they

contribute to building knowledge. Our argument is that even if the sole goal were

informing policy, mechanism experiments play a crucial, under-appreciated role.

This distinction between mechanism experiments and “policy evaluations could

also change the debate between about the use of experimentation to guide policy. We feel

many of the criticisms of experimentation are really criticisms of policy evaluations—

particularly “black box” policy evaluations where the mechanisms through which the

6
policy may affect outcomes are numerous or unclear. Deaton (2010, p. 246) for example

fears experimentation generates information that is too “narrow and local” to be of much

use for policy. While this can also be true for mechanism experiments, because of their

emphasis on how programs work, the knowledge gained can extend to a broader range of

situations.

The next section of the paper provides a brief review of how economists have

thought about experimentation and the problem of forecasting the effects of different

types of policies in different settings – that is, the challenge of external validity. We then

discuss what mechanism experiments can teach us. We use the randomized Moving to

Opportunity (MTO) residential mobility experiment (and several hypothetical extensions)

as an extended example. We then suggest a framework to help think about the conditions

under which the most policy-relevant information comes from a mechanism experiment,

a policy evaluation, or both, and close with some suggestions for future research.

7
 II. POLICY EXPERIMENTS AND EXTERNAL VALIDITY

Policymaking is inevitably about prediction. What is the effect of some familiar

policy when implemented in the future, or in some new setting? What is the effect of

some entirely new policy? A useful way to think about the types of research activities

that help answer these questions comes from Wolpin (2007) and Todd and Wolpin

(2008). They distinguish between ex post policy evaluation – understanding what

happened as the result of a policy or program that was actually implemented – and ex

ante policy evaluation, which DiNardo and Lee (2010, p. 2) describe as beginning “with

an explicit understanding that the program that was actually run may not be the one that

corresponds to a particular policy of interest. Here, the goal is not descriptive, but instead

predictive. What would be the impact if we expanded eligibility of the program? What

would the effects of a similar program be if it were run at a national (as opposed to a

local) level? Or if it were run today (as opposed to 20 years ago)? It is essentially a

problem of forecasting or extrapolating, with the goal of achieving a high degree of

external validity.”

The challenge in making policy forecasts from ex post evaluations stems from the

possibility that treatments may interact with characteristics of the policy’s setting –

including the target population, time period, or other contextual factors. The effects of

broken windows policing in Evanston, an affluent North Shore suburb of Chicago, may

differ from the policy’s effects when implemented in distressed neighborhoods on the

south side of Chicago. Those features of a policy’s setting (or of the policy itself) that

may influence the policy’s impacts are what the research literature outside of economics

calls moderators. We argue that the type of experiment that is most useful for addressing

8
this challenge and informing policy forecasts depends on what researchers believe they

know about a policy’s mechanisms, which non-economists sometimes also call

mediators.

At one extreme are situations in which researchers do not know very much about

a policy’s candidate mechanisms -- the list of plausible mechanisms might be

overwhelmingly long, or we might have little sense for whether the mechanisms

potentially interact or even work at cross purposes, or what (if any) aspects of the

relevant causal chain operate in ways that are invariant across policy settings. The

standard approach has been to carry out policy evaluations in as many settings as possible

of the sort in which the policy might actually be implemented. As Cook and Campbell

(1979) note, “tests of the extent to which one can generalize across various kinds of

persons, settings and times are, in essence, tests of statistical interactions… In the last

analysis, external validity … is a matter of replication” (p. 73, 78). Angrist and Pischke

(2010, p. 23-24) argue “a constructive response to the specificity of a given research

design is to look for more evidence, so that a more general picture begins to emerge …

the process of accumulating empirical evidence is rarely sexy in the unfolding, but

accumulation is the necessary road along which results become more general.”

As mentioned above, there is an active debate within the economics profession

about the value of this type of research program, focused largely on issues of external

validity. Under this approach we forecast a policy’s effects in some setting using previous

tests of the policy in similar settings – that is, we try to match on the policy’s candidate

moderators (see, for example, Hotz, Imbens and Mortimer, 2005, Cole and Stuart, 2010,

Imbens, 2010, Stuart et al., 2011). One challenge is that without some understanding of a

9
policy’s mechanisms, how do we decide which aspects of the policy or its setting is a

potentially important moderator? Another challenge comes from the fact that policy

evaluations are costly, and so we will never be able to carry out evaluations of ever

candidate policy of interest in every potentially relevant setting. We have nothing new to

add to this debate, which we view as largely orthogonal to the main argument we advance

in this paper.

The type of situation to which our paper is relevant arises when researchers have

some beliefs about the mechanisms through which a policy influences social welfare.

One way researchers currently use such beliefs is by interpreting the results of

randomized experiments through the lens of a particular structural model (Wolpin, 2007;

Todd and Wolpin, 2008; Heckman, 2010; Imbens, 2010). This approach takes the policy

experiment that is run as given, imposes some assumptions after the fact about the

policy’s mechanisms, fits the model, then forecasts the effects of a wide range of policies

and settings. This approach can make sense when we have sufficiently sharp prior beliefs

about the way the world works to be confident that we have the right structural model,

and that the key structural parameters really are structural (that is, invariant across

settings). The structural model substitutes assumptions for data, traded off against the risk

that our assumptions are incorrect.

But if we believe we know something about the mechanisms through which the

policy might operate, why limit ourselves to using this information only after a policy

evaluation has been designed and carried out? Why not use this information to help

inform the design of the experiment that is being run? Why not design experiments that

are explicitly focused on isolating the effects of candidate mechanisms? Once our focus

10
shifts to identifying mechanisms, the importance of having close (or even any)

correspondence between the interventions we test and the specific policy applications we

seek to inform is diminished. The change that this way of thinking implies for the design

of our policy experiments is not just cosmetic. The change can be drastic, as we illustrate

in the next section.2

Mechanism experiments can help with the policy forecasting or external validity

problem in two ways. First, improved understanding of a policy’s mechanisms can help

us predict what aspects of the policy or its setting may moderate the policy’s impacts.

Second, by taking advantage of what researchers believe they already know mechanism

experiments can be less costly than policy evaluations. This means that we can carry out

relatively more mechanism experiments in different settings, and help prioritize the types

of policies and settings in which we should carry out full-scale policy evaluations.

III. MECHANISM EXPERIMENTS

In what follows we illustrate some of the ways in which mechanism experiments

can help generate policy-relevant information. We use the Moving to Opportunity (MTO)

residential-mobility experiment as an extended example. As context for this example, we

note that economists have become increasingly interested in the role of social interactions

in affecting people’s choices and behavioral outcomes (Becker and Murphy, 2000;

Manski, 2000). Housing policy affects the social interactions that people experience (as

well as the quality of their local public goods) by affecting the geographic concentration

of poverty in America.

2
Another relevant observation here is that if we really believe that the key structural parameters in our
model are structural, then there is no intrinsic reason that we would need to test a real policy to identify
their value.

11
 To learn about the effects of concentrated neighborhood poverty on poor families,

in the early 1990s the U.S. Department of Housing and Urban Development launched the

MTO demonstration. Since 1994, MTO has enrolled around 4,600 low-income public

housing families with children and randomly assigned them into three groups: 1) a

traditional voucher group, which received a standard housing voucher that subsidizes

them to live in private-market housing; 2) a low-poverty voucher group that received a

standard housing voucher that is similar to what was received by the traditional voucher

group, with the exception that the voucher could only be redeemed in Census tracts with

1990 poverty rates below 10 percent; and 3) a control group, which received no

additional services.

Assignment to the low-poverty voucher group led to more sizable changes in

neighborhood poverty and other neighborhood characteristics than did assignment to the

traditional voucher group (Ludwig et al., 2008). The traditional voucher treatment did not

have many detectable impacts on the outcomes of MTO parents or children 4-7 years

after baseline (Kling, Ludwig and Katz, 2005, Sanbonmatsu et al., 2006, and Kling,

Liebman and Katz, 2007, Fortson and Sanbonmatsu, 2010). The low-poverty voucher

treatment generated a more complicated pattern of impacts. The low-poverty voucher did

not affect children’s schooling outcomes, perhaps because MTO moves wound up

generating only modest changes in school quality, but did cause sizable reductions in

youth violence. The low-poverty voucher had effects on other youth behaviors that

differed by gender, with girls doing better and boys doing worse as a result of the moves

(Clampet-Lundquist et al., 2011). For adults, the low-poverty voucher treatment did not

product detectable changes in labor market or other economic outcomes, but did have

12
important effects on mental health and some physical health outcomes, including

obesity.3

Two of us (Kling and Ludwig) have worked on MTO for many years, and have

often heard the reaction that the traditional voucher treatment is more policy-relevant and

interesting than the low-poverty voucher treatment, because only the former corresponds

to a realistic policy option. But it was the low-poverty voucher that generated a

sufficiently large “treatment dose” to enable researchers to learn that something about

neighborhood environments can matter for important outcomes like adult obesity, a fact

that would not have been discovered if MTO’s design had only included the more

realistic traditional voucher treatment. For this reason, findings from the low poverty

voucher have been very influential in housing policy circles.

To illustrate the different ways in which mechanism experiments might be useful

for policy, we focus on one of the key findings from MTO —the reduction in adult

obesity. Imagine New York City policymakers were interested in reducing the prevalence

of obesity, which contributes to health problems like diabetes and heart disease. One

pattern in the epidemiological data that would immediately be obvious to policy planners

would be the large disparities in obesity prevalence across neighborhoods. Data from

2003-7 show that the share of adults who are obese ranges from 8 percent on the Upper

East Side to 30 percent in the directly adjacent neighborhood of East Harlem (Black and

Macinko, 2010).

These patterns would presumably lead policymakers to think about the

“production function” that produces obesity, in the spirit of Sen et al. (2009). Of course

3
The public health community officially defines obesity as having a body mass index, which is weight in
kilograms divided by height in meters squared, or BMI = kg/m2, of 30 or more.

13
the prevalence of obesity is a function of calories consumed and time spent in energetic

activity, but also of other factors that interact with diet and physical activity, and that

vary across neighborhoods. For example, some neighborhoods may be “food deserts” that

have few grocery stores that sell fresh fruits and vegetables, which might lead people to

eat higher-calorie, less healthy foods (Wehunt, 2009). Some neighborhoods may have

few parks or places to exercise. Neighborhood conditions might also affect levels of

psychosocial stress, which in turn can affect diet and exercise.

Depending on our prior beliefs about certain aspects of the framework discussed

above, mechanism experiments could be useful for guiding policy in this area by helping

us: 1) rule out candidate policies; 2) expand the set of policy options for which we can

forecast effects; 3) prioritize available research funding; 4) concentrate resources on

estimating parameters about which we have the most uncertainty or disagreement; and 5)

strengthen causal inference with either randomized or “natural” experiments.

1. Ruling out policies

Twenty-five years ago the distinguished sociologist Peter H. Rossi (1987, p. 4)

considered the discouraging results of the policy-evaluation literature of the 1970s and

1980s and formulated his Iron Law of Evaluation: “the expected value of any net impact

assessment of any large scale social program is zero.”4 This pessimistic assessment is

presumably motivated by the difficulty of consistently implementing social programs

well, and by our limited understanding about what combination of mechanisms is most

4
Rossi’s Stainless Steel Law of Evaluation holds that “the better designed the impact assessment of a social
program, the more likely is the resulting estimate of net impact to be zero.” Rossi’s Zinc Law of
Evaluation is somewhat less pessimistic in its way: “only those programs that are likely to fail are
evaluated.”

14
important for improving people’s life chances. Sometimes mechanism experiments can

be used to provide an initial indication of whether Rossi’s law holds in a given context,

and to do so at reduced cost compared to carrying out a series of full-scale policy

evaluations.

For example, policymakers concerned that distressed neighborhoods may be

“food deserts” might consider a policy of subsidizing grocery stores to locate into such

areas as a way to reduce obesity. Carrying out a policy evaluation of location incentives

for grocery stores would be expensive because the unit of randomization is the

community, the cost per community is high, and the number of communities needed to

have adequate statistical power is large.

Now consider the following mechanism experiment that could be carried out

instead: Enroll a sample of low-income families, and randomly assign some of them to

receive free weekly delivery of fresh fruits and vegetables to their homes. By using

individuals as the unit of randomization, rather than communities, this mechanism

experiment would be much less expensive than the more “realistic” policy evaluation.

Randomizing people rather than neighborhoods also lets us test a “treatment dose” that is

much more intensive than what could be obtained with any realistic policy intervention.

Imagine we found that several hundreds of dollars’ worth of free fruits and

vegetables delivered to someone’s door each month had no effect on obesity. Suppose we

also believed eating habits adapt rapidly to changes in food availability, that social

interactions are not very important in shaping eating habits, and that reducing the price of

accessing fruits and vegetables never reduces the chances of eating them (that is, there is

a monotonic relationship between the treatment dose and the treatment response). In that

15
case null results from our mechanism experiment would lead us to predict that any sort of

policy that tried to address the “food desert” problem would (on its own) be unlikely to

diminish the obesity problem.

If we had more uncertainty about the role of social interactions or time in

affecting eating habits, then different mechanism-experiment designs would be required.

If we believed that social interactions might be important determinants of people’s eating

habits, then we would need a more costly experiment with three randomized arms, not

just two – a control group, a treatment arm that received free food delivery for

themselves, and a treatment arm that received food delivery for themselves and for a

limited number of other households that the family designated (“buddy deliveries”).5 If

we thought that eating habits were determined at a still larger macro-level, we would

have to randomly assign entire communities to receive free home food delivery. A

community-level test of home fruit and vegetable delivery could still wind up being less

expensive than a policy evaluation of incentive locations for grocery stores, because of

the large guarantees that would be required to entice a grocery store to incur the start-up

costs of establishing a new location in a neighborhood. But if we thought that eating

habits changed very slowly over time, and at the community level, then we would have to

commit to providing home food delivery for entire communities for extended periods of

time – at which point there might be little cost advantage compared to a policy evaluation

of grocery-store subsidies.

5
Duflo and Saez (2003) discuss a cleverly designed experiment that used individuals as the unit of analysis
but was designed to identify spillover effects. In their experiment, some people in some departments within
a company received incentives to visit a benefit fair to learn more about savings plans. They assessed both
direct effects of the information, and effects of information spillovers (from comparisons of the outcomes
of the non-incentivized individuals in incentivized departments to individuals in non-incentivized
departments). The information diffused through the experiment had a noticeable impact on plan
participation.

16
 The possibility of using a mechanism experiment to learn more about “food

deserts” is not an isolated example. Consider that the 2010 health care legislation

included $9.5 billion in funding over five years to double the number of people served by

local community health centers. Expanding community health care centers could

potentially reduce total health care spending by improving access to regular preventive

health care services, thereby reducing the need for emergency care (National Association

of Community Health Centers. 2010). However, this hypothesis has not been subject to

any sort of rigorous empirical test that we know of, because it is hard to imagine running

a large-scale experiment that randomly assigned areas to get community health care

centers. But we could run a mechanism experiment that randomly assigned individual

low-income families in medically under-served neighborhoods to receive improved

access to routine, preventive health care using a van that pulled right up in front of their

home. Indeed, this was the same basic idea behind the Washington, D.C. Mobile Unit for

Child Health Care experiment that was run 45 years ago (Gutelius et al., 1977).

As a final (non-health) example, consider the question of whether smaller high

schools improve student achievement. Possible mechanisms through which this might

occur include stronger relationships between students and school staff, having students

spend more time around peers who share their interests, and providing school

administrators with more autonomy (Bloom et al., 2010). Instead of immediately carrying

out a full-scale, very costly policy evaluation of small schools, why not first carry out a

mechanism experiment focused on bonding instead? Take a representative sample of

charter schools, which already provide administrators with autonomy. Randomly assign

some teachers to be offered the chance to earn overtime pay by working after school and

17
weekends with small, randomly-selected groups of students in an effort to promote

faculty-student and student-to-student bonding. Evidence that this intervention was

capable of promoting student engagement and academic outcomes would suggest the

value of carrying out a large-scale policy evaluation. But evidence that even what H.L.

Mencken (1948) would call a “horse-doctor’s dose” of extra bonding did not affect

student outcomes would greatly reduce the motivation to carry out a large-scale policy

evaluation of smaller schools.

2. Expand the set of policies and settings for which we can forecast policy impacts.

Ruling out entire classes of policy interventions is easier when our experiments

test interventions that are as intensive (or more) as anything that could be accomplished

by actual policies. Testing unrealistically intensive treatment arms also has the benefit of

letting us forecast the effects of a wide range of more realistic policy options in those

cases when, in spite of Rossi’s Iron Law, our policy experiments do identify successful

interventions. As Hausman and Wise (1985, p. 194-5) noted a quarter-century ago: “If,

for policy purposes, it is desirable to estimate the effects of possible programs not

described by treatments, then interpolations can be made between estimated treatment

effects. If the experimental treatments are at the bounds of possible programs, then of

course this calculation is easier.”

The policy impact that this type of study can have is illustrated by the RAND

Health Insurance Experiment, which included many treatment arms that do not

correspond to any sort of health insurance policy one could buy today (Newhouse et al.,

1993). Yet this remains one of our most important sources of information about how the

18
generosity of health insurance plans affects the demand for health care and subsequent

health outcomes. With a total cost of $285 million in 2010 dollars, the RAND experiment

also holds the record – for now – as the most expensive mechanism experiment of all

time (Greenberg and Shroder, 2004, p. 181).

3. Prioritize research funding

If a mechanism experiment tested the most intensive imaginable intervention to

address the problem of “food deserts” and found no effect on obesity, we would rule out

not only the value of policies to address food deserts but also, obviously, the value of

policy evaluations to test those types of policies. Null results from a policy evaluation of

a more realistic but less-intensive intervention would not let us shut down an entire line

of research inquiry in the same way, since it would always be possible to imagine that a

slightly more intensive intervention might yield more promising results.

Encouraging results from a mechanism experiment would help us decide where to

invest additional research funding, and might also help shape the types of policies that we

subjected to full-scale policy evaluations. Suppose, for example, we found that delivering

hundreds of dollars worth of free fruit and vegetables to the doorsteps of low-income

families each month only changed the consumption of, say, apples. This finding might

lead policymakers to conclude that the right policy to evaluate is not just a costly effort to

incentivize new grocery stores to move into high-poverty areas, but also (or perhaps

instead) a lower-cost program to subsidize bodegas and convenience stores just enough to

make it profitable for them to stock this one type of fruit.

19
4. Concentrate resources on estimating parameters about which we are most uncertain.

In the introduction we noted that mechanism experiments can help us concentrate

resources on estimating parameters about which we have the most uncertainty or

disagreement. As another example along these lines, suppose policymakers are concerned

about the secondary consequences of psychosocial stress on poor families, including

health impacts. For families in poor urban areas, one of the most important sources of

stress is crime – particularly gun crime (Cook and Ludwig, 2000; Kling, Liebman and

Katz, 2005; Kling, Ludwig and Katz, 2005). Policymakers could sponsor a full-scale

evaluation of targeted police patrols against illegal guns in high-crime areas, then test the

impacts on obesity and other health outcomes. But previous work already tells us

something about this intervention’s effects on crime (Cohen and Ludwig, 2003), and

perhaps also about the effect of crime on stress (Buka et al., 2001). The new information

from this experiment is primarily about the stress→obesity link. But for a given budget

we could learn more about the stress→obesity pathway (and how that might vary across

settings) by carrying out a mechanism experiment that enrolled residents of high-crime

areas and assigned some to a meditation-based stress-reduction program (Kabat-Zinn et

al., 1992).

In other situations we might be most uncertain about the link between our policy

levers and key mediating mechanisms (P→M). For example, in the case of obesity we

already understand the connection of body mass with diet and exercise (Cutler, Glaeser,

and Shapiro, 2003). But we might not understand the effects of our policies on diet and

exercise. In situations like this, diet and exercise become what medical researchers call

“surrogate clinical endpoints,” which then become the dependent variables of interest for

20
our experiments. The idea of focusing selectively on testing individual links in a causal

chain also raises the possibility of using mechanism experiments to compress the

timetable required to learn about the long-term effects of some policy, by testing different

sequential links in a causal chain simultaneously.

Perhaps less obvious is the value of carrying out multiple experiments that use

different policy levers to manipulate the same mechanism, given the great difficulty of

determining what is the true mediating mechanism that links a policy to an outcome,

rather than just a proxy for the mediating variable that really matters.6 Showing that the

effects of reduced stress on obesity is the same regardless of whether stress levels are

modified through a meditation program or by some sort of anti-gun policing program

would be informative about whether the mediating mechanism of stress is “non-

implementation specific,” to use John DiNardo’s term, or what Heckman (2010) calls

“policy invariant.”

A final non-health example about the ability of mechanism experiments to focus

research resources comes from the possibility of avoiding the need to carry out full-

blown “synergy” (or “kitchen sink”) experiments of the sort that the federal government

regularly sponsors, like Jobs Plus. This experiment tested the combined effects of

providing public housing residents with financial incentives for work (relief from the

“HUD tax” on earnings that comes from setting rent contributions as a fixed share of

income), employment and training services, and efforts to improve “community support

for work.” Previous studies have already examined the effects of the first two program

6
Some simple notation suggested to us by Steve Pischke helps illustrate the problem. Let P be the policy,
M be the mediator, Y be the outcome (with P→M→Y as in Figure 1), with M=U+V, cov(U,V)=0,
cov(U,Y)=0, and cov(V,Y)>0. That is, only the V part of M is causally related to Y. In population data we
see cov(M,Y)>0. In this example, M is an implementation specific mediator because policies that change
the V part of M will change Y, but policies that change only the U part of M will not influence Y.

21
ingredients when administered independently, while the potential value of community

support for work is suggested by the Wilson (1987, 1996) among others. The key

program theory of Jobs Plus is that these three mechanisms interact, and so have more-

than-additive effects on labor market outcomes (Bloom, Riccio and Verma, 2005).

Across six cities, Jobs Plus randomly assigned entire housing projects to either a control

group, or a program group in which residents received the bundle of Jobs Plus services.

We could have instead carried out a mechanism experiment that enrolled a

slightly less disadvantaged (and hence slightly less directly policy-relevant) study sample

that needed one or two but not all three of the mechanisms the Jobs Plus theory suggests

are needed for labor market success. Imagine enrolling people who applied for means-

tested housing assistance, which in some cities is rationed using randomized lotteries

(Jacob and Ludwig, 2011), and are already living in neighborhoods with high

employment rates. Then we randomly assign some of them to receive employment and

training services. A test of the Jobs Plus “synergy” theory comes from comparing the

response to these services for those who were versus were not lucky enough to be

randomly assigned a housing subsidy. Our proposed mechanism experiment conserves

resources by reducing the dimensionality of the experimental intervention.

5. Help strengthen causal inference

Mechanism experiments can help us interpret the results of policy evaluations,

including null findings. Once we know that some mechanism is linked to an outcome, the

first thing we would check upon seeing a zero impact in a full-scale policy evaluation is

whether the policy successfully changed the mediator. Evidence that the mediator was

22
unchanged would suggest the potential value of testing other policies that might generate

larger changes in the mediator. Without the mechanism experiment, we wouldn’t be sure

whether it would be worth following up a null impact from a policy evaluation with more

research in that area.

Mechanism experiments can also strengthen the basis for causal inference with

“natural experiment” policy evaluations. Imagine a simple pre-post study of aggregate

U.S.-level data of a change in Medicaid policies that reduced out-of-pocket costs to poor

adults from having, say, bariatric surgery, in which part of someone’s stomach is

removed or reduced in size by a gastric band in order to reduce appetite and food

consumption. Suppose the study found that after the policy change, bariatric surgery rates

among low-income people increase and obesity rates decline. Absent any additional

information, this study design would not provide compelling evidence about the link

between bariatric surgery and obesity, given the large number of other factors that are

changing over time that influence obesity. But there would seem to be far fewer

confounding threats to estimating the effect of the policy on bariatric surgery rates (the

P→Y link) from a simple pre-post comparison. Additional evidence about the mechanism

(the M→Y link between bariatric surgery and obesity) would enable us to infer how

much of the time trend in obesity prevalence was due to the Medicaid policy change.

Evidence for the mechanism-outcome link here happens to come from a medical trial

rather than an experimental test of an unrealistic policy, but the example nonetheless

highlights our key point.

New mechanism experiments could even be designed with the explicit goal of

better understanding existing natural experiment findings. For example, numerous studies

23
of compulsory schooling laws document the causal relationship of educational attainment

with earnings, crime, health, and other outcomes (Oreopoulos and Salvanes, 2009). Less

well understood are the mechanisms behind this relationship. Is it that schooling affects

academic skills? Or specific vocational skills? Or social-cognitive skills? The answer is

relevant for thinking about how we should deploy the $485 billion the U.S. spends each

year on K-12 public schooling (U.S. Census Bureau. 2011, Table 258). Why not spend a

few million dollars on a mechanism experiment that assigns youth to curricula or

supplemental activities that emphasize different specific skills, to better understand the

mechanisms behind the effects of compulsory schooling laws?

IV. WHEN CAN MECHANISM EXPERIMENTS BE USEFUL?

The purpose of our paper is not to argue that economists should only carry out

mechanism experiments, or that mechanism experiments are “better” than policy

evaluations. Our main point is that given the current paucity of mechanism experiments

designed to help answer policy questions, on the margin we think that economists should

be doing more of them.

Table 1 presents a framework for thinking about the conditions under which

mechanism experiments can help inform policy decisions. Under a very particular set of

conditions, mechanism experiments may by themselves be sufficient to guide policy

decisions. More common are likely to be scenarios in which mechanism experiments and

traditional policy evaluations (which could include “natural” as well as randomized

experiments) are complementary. Under some circumstances mechanism experiments

24
might not even be that helpful, and a more useful approach would be to just go right to

running a black-box policy evaluation.

1. When Mechanism Experiments Can Be Helpful

In order for a mechanism experiment to make any sense at all, we need to believe

that we know at least something about the candidate mechanisms through which a policy

might affect the outcomes of ultimate policy concern (the right-hand column of Table 1).

Under some circumstances mechanism experiments might be sufficient to guide

policy design. We need to believe that the list of candidate mechanisms through which a

policy might affect outcomes is fairly short, or that the long list of potentially relevant

mechanisms do not interact or work at cross purposes (a short list of candidate

mechanisms that could interact would not by itself preclude a mechanism experiment).

Depending on the application we might need to know something already about other

parts of the causal chain. At the very least we would need to be confident that existing

systems are capable of reliably delivering the policies that activate key mechanisms.

Even then, if the cost of carrying out a policy evaluation were low enough relative to the

policy stakes, we would probably still wish to carry out a policy evaluation to improve

our policy forecast. We would settle for just a mechanism experiment if the costs of

carrying out a policy evaluation were prohibitive, or the policy stakes were low.

2. Do Mechanism Experiments plus Policy Evaluations

One reason it would make sense to follow a mechanism experiment that had

encouraging results with a full-blown policy evaluation would be to learn more about

25
other parts of the causal chain, such as when there is implementation uncertainty. For

example, medical researchers distinguish between “efficacy trials,” which are small-scale

research trials of model programs carried out with high fidelity, and “effectiveness trials”

that test the effects of some intervention carried out under field conditions at scale.

Efficacy trials can be thought of as a type of mechanism experiment, since having a

bespectacled, laptop-toting professor loom over the program’s implementation is not

usually standard operating procedure. Compared to efficacy trials, larger-scale

effectiveness trials often have more program attrition, weaker training for service

providers, weaker implementation monitoring, and smaller impacts (Lipsey et al., 2007).

As noted above, prior evidence from mechanism experiments can enhance the

efficiency of our portfolio of policy evaluations by helping us figure out which

evaluations are worth running. This includes carrying out mechanism experiments in

different settings to determine in where it is worth trying a policy evaluation.

Learning about the mechanisms through which a policy affects outcomes can also

help predict which aspects of the policy or its settings will moderate the policy’s impacts,

although it is worth keeping in mind that the correspondence between mechanisms and

moderators is far from perfect. For example, the well-known Tennessee STAR

experiment found that reducing class sizes in elementary school improved learning

outcomes (Krueger, 1999; Schanzenbach, 2007). Lazear (2001) argues that a key

mechanism for these class-size effects is the reduced chance that instructional time in a

given classroom is diverted by disruptive students, which helps explain why in the STAR

experiment lower-income and minority students seemed to benefit the most. But when

California had to hire a large number of teachers to enact class-size reduction statewide

26
average teacher quality seemed to decline, particularly in those schools serving

disproportionately low-income and minority students (Jepsen and Rivkin, 2009). Thus,

teacher quality turned out to be a surprise mechanism (at least to California

policymakers). Student background wound up being a moderator that influenced different

parts of the causal chain in different ways.

3. Do Some Combination of “Basic Science” and Policy Evaluation

In some situations researchers do not yet know enough to narrow down the list of

candidate mechanisms through which a policy operates, or worry that a policy’s long list

of candidate mechanisms might interact (or if some might work at cross purposes) –

represented by the first column of Table 1. The debate within the economics profession is

about whether it is best under these circumstances to carry out “basic science” studies or

to carry out policy evaluations. The extreme position is that policy evaluations can never

be useful for policy purposes, which strikes us as unlikely to be correct.

In the case of Moving to Opportunity, for example, observational studies going

back to the 1920s had shown that neighborhood attributes are correlated with behavioral

outcomes, even after controlling for individual- and family-level factors. Policymakers

need to know whether such correlations reflect an underlying causal relationship, which

is relevant to decisions like whether to whether to devote resources to building public

housing or to private-market rent subsidies, or whether to allow suburban townships to

limit zoning approval for low-cost housing. Given the large number of potentially-

interacting mechanisms through which residential location might affect behavior and

27
well-being, it is not clear that anything short of a black-box policy evaluation would have

much value in guiding these policy decisions.

One common criticism of black-box policy evaluations is that we cannot

understand the characteristics that explain heterogeneity of treatment effects (that is, a

policy’s moderators) without understanding the policy’s key mediating mechanisms.

While there is no question that evidence about mechanisms is tremendously valuable, we

believe it is not correct that black-box evaluations are never useful.

Consider the example of statins, which have been used since the late 1980s7 and

shown in numerous randomized clinical trials to reduce the risk of heart disease and

overall mortality (Ross et al., 1999, Gotto, 2003). Statins were originally thought to

prevent heart attacks by lowering cholesterol levels in the blood, which in turn reduced

the chance of plaque build-up. But meta-analyses of black-box clinical trials showed that

statins improved health outcomes even among people who already had relatively low

levels of blood cholesterol at baseline (Golomb et al., 2004, Wilt et al., 2004,

Thavendiranathan et al., 2004). Moreover, these meta-analyses showed that the

cardiovascular benefits of statins seemed to occur too rapidly after onset of treatment to

be explained by the effects of statins on plaque accumulation (Golomb et al., 2008). The

leading hypothesis – at least for now – is that statins reduce heart attacks partly by

reducing inflammation (Zhang, 2010) or blood pressure (Golomb et al., 2008).

The key point for present purposes is that right now we don’t really know exactly

why statins reduce heart attacks. Yet meta-analyses of black-box clinical trial studies

show that they clearly do improve health, and can also tell us something about how their

effects on health are moderated by patient characteristics such as age, gender, and
7
Thanks to Elbert Huang and Harold Pollack for this example.

28
baseline health status. Our limited understanding of the mechanisms through which

statins work has not prevented them from becoming one of the world’s top-selling drug

classes, to the extent that some medical experts have suggested should be “put into the

water supply” (Golomb et al., 2004, p. 154).

A similar point was made during Congressional testimony in 1971 by Sidney

Farber, the “godfather of cancer research,” who argued (as quoted in Fortune, 2007): “We

cannot wait for full understanding; the 325,000 patients with cancer who are going to die

this year cannot wait; nor it is necessary, to make great progress in the cure of cancer, for

us to have the full solution of all the problems of basic research… The history of

medicine is replete with examples of cures obtained years, decades, and even centuries

before the mechanism of action was understood for these cures – from vaccination, to

digitalis, to aspirin.”8

This is not to say that later understanding of mechanisms does not generate

tremendous benefits to society. For example, learning more about how chemotherapy

works has dramatically increased the benefit/cost ratio of such treatments over time. But

evidence that an intervention works, even if we don’t understand why, is better than not

having access to that intervention at all. Repeated black-box experiments can eventually

help us learn something about the policy’s moderators, and, as in our statins example, can

also inform our theorizing about candidate mechanisms as well.

8
Thanks to Harold Pollack for suggesting this quotation. At the risk of over-emphasizing the point, one
more example comes from two of the most important mental health drug discoveries – lithium, which is
used to treat bipolar disorder, and Thorazine, which is used to treat psychosis. Modern medicine has very
little understanding of why either medicine works in helping patients (Harris, 2011).

29
 V. CONCLUSIONS

It seems like common sense that the best way to use experiments to inform policy

is to test policies. However, we argue here for increased use of randomized experiments

that identify behavioral mechanisms that are central to clearly specified policy questions,

even if the specific interventions that are tested (or their settings) do not correspond

exactly to what policymakers would implement in practice. While our suggestion might

seem obvious once articulated, mechanism experiments that are designed to help answer

specific policy questions remain rare. We hasten to add that mechanism experiments and

traditional policy evaluations are as a general proposition best thought of as

complements, rather than substitutes. We need to make greater use of mechanism

experiments, on the margin, without fetishizing mechanisms.

The larger question of how to structure experiments to maximize the ability to

apply the findings more generally in other contexts opens up a number of potentially

fruitful lines of additional research beyond what we have considered here. For example,

many people seem to have the intuition that evidence about either the link between policy

levers and mechanisms (P→M from Figure 1) or between mechanisms and outcomes

(M→Y) is more generalizable than evidence about the link between policies and ultimate

outcomes of interest (P→Y). It is not hard to think of situations in which this is true, but

this need not be true in all cases.9 It would be useful to learn more about how the causal

9
Imagine a case with a single candidate mediator and outcome of interest. Whether either of the individual
links in this causal chain (P→M or M→Y) is more stable across contexts than is the total effect of the
policy on the outcome, P→Y, depends in part on how P→M and M→Y co-vary across contexts. It is not
hard to imagine cases in which the two relationships negatively co-vary, so the effect of the policy on the
outcome is more stable across situations than the link between the policy and mediator or the mediator and
outcome. Suppose that in neighborhoods where adoption of broken windows policing leads to relatively
larger increases in arrests for minor offenses, the stigma of arrest declines, and so the deterrent effect of the
prospect of being arrested goes down. Or suppose that in areas where local residents are not very easily

30
links from P→M and M→Y co-vary across contexts, and the extent to which those links

reinforce each other or may tend to offset each other.

A second line of investigation that seems worth exploring more is the benefits and

costs of policy field experiments (both mechanism experiments and policy evaluations)

versus “natural experiment” studies. Sometimes natural experiment studies have designs

that are as good as random assignment of treatment because they actually involve random

assignment (see for example, Angrist 1990, Kling, 2006, and Jacob and Ludwig, 2011).

But more often, natural experiment studies necessarily rely on research designs that

generate information that may be more local than that obtained from an experiment (such

as regression discontinuity), or that may be more vulnerable to omitted variables bias. On

the other hand, natural experiment studies circumvent the external validity concerns

raised by either randomization bias (the self-selection of people willing to sign up for a

randomized experiment; see Heckman, 1992 and Malani, 2006) or selection-partner bias

(the willingness of organizations to participate in experiments; see Alcott and

Mullainathan, 2011). But with few exceptions little is currently known about the extent of

randomization or selection-partner bias in practice. Alternatively, policy field

experiments and natural experiments may be complements in a broader program of

research on an issue that involves multiple stages (Kling 2007).

A final question worth considering is the issue of when and how to export results

across contexts. While statistical matching of estimates obtained from similar

interventions and contexts is fine, as far as it goes, a broader framework would let us

incorporate behavioral models, parameters, and prior beliefs into the policy forecasting

deterred by the prospect of being arrested, policymakers respond by implementing this policing strategy in
a way that leads to relatively larger numbers of minor arrests.

31
exercise. This type of policy forecasting, or ex ante policy evaluation, will inevitably

require more assumptions, theory and guesswork than ex post studies of previous policies

(see also Harrison and List, 2004, p. 1033). But policy forecasting is in the end at least as

important for public policy. As the distinguished physicist Richard Feynman (1964) once

argued, “The moment you make statements about a region of experience that you haven’t

directly seen, then you must be uncertain. But we always must make statements about the

regions that we haven’t seen, or it’s no use in the whole business.”

32
REFERENCES

Allcott, Hunt and Sendhil Mullainathan (2010) “External validity and partner selection
bias.” Working Paper, Harvard University Department of Economics.

Angrist, Joshua D. (1990) “Lifetime earnings and the Vietnam Era Draft Lottery:
Evidence from Social Security Administrative Records.” American Economic Review.
80: 313-335.

Angrist, Joshua D. and Jorn-Steffen Pischke (2009) Mostly Harmless Econometrics.

Princeton, NJ: Princeton University Press.

Angrist, Joshua D. and Jorn-Steffen Pischke (2010) “The credibility revolution in

empirical economics: How better research design is taking the con out of econometrics.”
Journal of Economic Perspectives. 24(2): 3-30.

Banerjee, Abhijit V. and Esther Duflo (2009) “The experimental approach to

development economics.” Annual Review of Economics. 1: 151-178.

Beatty, Alexandra (2009) Strengthening Benefit-Cost Analysis for Early Childhood

Interventions: Workshop Summary. Washington, DC: National Academy of Sciences
Press.

Becker, Gary S. and Kevin M. Murphy (2003) Social Economics: Market Behavior in a
Social Environment. Belknap Press of Harvard University Press.

Black, Jennifer L. and James Macinko (2010) “The changing distribution and
determinants of obesity in the neighborhoods of New York City, 2003-7.” American
Journal of Epidemiology.

Bloom, Howard S., James A. Riccio, and Nandita Verma (2005) Promoting Work in
Public Housing: The Effectiveness of Jobs-Plus. New York: MDRC.

Bloom, Howard S., Saskia Levy Thompson, and Rebecca Unterman (2010) Transforming
the High School Experience: How New York City’s Small Schools Are Boosting Student
Achievement and Graduation Rates. New York: MDRC.

Buka, Stephen L., Theresa L. Stichick, Isolde Birdthistle, and Felton J. Earls (2001)
“Youth exposure to violence: Prevalence, risks, and consequences.” American Journal of
Orthopsychiatry. 71(3): 298-310.

Clampet-Lundquist, Susan, Kathryn Edin, Jeffrey R. Kling, and Greg J. Duncan (2011)
“Moving At-Risk Youth Out of High-Risk Neighborhoods: Why Girls Fare Better Than
Boys.” American Journal of Sociology, 116(4): 1154-1189.

33
Cohen, Jacqueline and Jens Ludwig (2003) “Policing crime guns.” In Evaluating Gun
Policy, Jens Ludwig and Philip J. Cook, Eds. Washington, DC: Brookings Institution
Press. pp. 217-250.

Cole, Stephen R. and Elizabeth A. Stuart (2010) “Generalizing evidence from

randomized clinical trials to target populations: The ACTG 320 Trial.” American Journal
of Epidemiology. 172(1): 107-115.

Cook, Philip J. and Jens Ludwig (2000) Gun Violence: The Real Costs. New York:
Oxford University Press.

Cook, Thomas D. and Donald T. Campbell (1979) Quasi-Experimentation: Design and

Analysis Issues for Field Settings. Wadsworth.

Cutler, David M., Edward L. Glaeser, and Jesse M. Shapiro (2003) “Why have
Americans become more obese?” Journal of Economic Perspectives. 17(3): 93-118.

Deaton, Angus (2010) “Instruments, randomization, and learning about development.”

Journal of Economic Literature. 48: 424-455.

DiNardo, John and David S. Lee (2010) “Program evaluation and research designs.”
Cambridge, MA: NBER Working Paper 16016.

Duflo, Esther, and Emmanuel Saez (2003) “The Role of Information and Social
Interactions in Retirement Plan Decisions: Evidence From a Randomized Experiment,”
Quarterly Journal of Economics, 118(3): 815-842.

Feynman, Richard. 1964. “The Great Conservation Principles.” The Messenger Series.
Quotation starts at 38:48. Available at
<[Link]
9050-7ea34f236045||>.

Fortson, Jane G. and Lisa Sanbonmatsu (2010) “Child health and neighborhood
conditions: Results from a randomized housing voucher experiment.” Journal of Human
Resources. 45(4): 840-864.

Fortune, Clifton Leaf. 2007. “Why we’re losing the war on cancer (and how to win it).”
CNN Health. January 9. At <[Link]
09/health/[Link].waroncancer_1_gamma‐rays‐testicular‐cancer‐national‐cancer‐
act/6?_s=PM:HEALTH>.

Golomb, Beatrice A., Michael H. Criqui, Halbert White, and Joel E. Dimsdale (2004)
“Conceptual foundations of the UCSD statin study.” Archives of Internal Medicine. 164:
153-162.

34
Golomb, Beatrice A., Joel E. Dimsdale, Halbert L. White, Janis B. Ritchie, and Michael
H. Criqui (2008) “Reduction in blood pressure with statins.” Archives of Internal
Medicine. 168(7): 721-727.

Gotto, Antonio M. (2003) “Safety and statin therapy.” Archives of Internal Medicine.
163: 657-659.

Greenberg, David and Mark Shroder (2004) The Digest of Social Experiments, 3rd.
Edition. Washington, DC: Urban Institute Press.

Gutelius, Margaret F., Arthur D. Kirsh, Sally MacDonald, Marion R. Brooks, and Toby
McErlean (1977) “Controlled study of child health supervision: Behavioral results.”
Pediatrics. 60: 294-304.

Harris, Jeffrey E. (1985) “Macro-experiments versus micro-experiments for health

policy.” In Jerry Hausman and David Wise, Eds. Social Experimentation. Chicago:
University of Chicago Press. pp. 145-185.

Harris, Gardiner. 2011. “Federal Research Center Will Help to Develop Vaccines.” New
York Times. January 23, p. A1. At
<[Link]

Harrison, Glenn W. and John A. List (2004) “Field experiments.” Journal of Economic
Literature. 42(4): 1009-1055.

Hastings, Justine. S., and Jeffrey M. Weinstein (2008) “Information, School Choice, and
Academic Achievement: Evidence from Two Experiments.” Quarterly Journal of
Economics,123(4): 1373-1414.

Hastings, Justine. S., and Lydia Tejeda-Ashton (2008) “Financial Literacy, Information,
and Demand Elasticity: Survey and Experimental Evidence.” Cambridge, MA: NBER
Working Paper No. 14538.

Hausman, Jerry A. and David A. Wise (1985) Social Experimentation. Chicago:

University of Chicago Press.

Heckman, James J. (1992) “Randomization and social policy evaluation.” In Evaluating

Welfare and Training Programs, Edited by Charles Manski and Irwin Garfinkel.
Cambridge, MA: Harvard University Press. pp. 201-230.

Heckman, James J. (2010) “Building bridges between structural and program evaluation
approaches to evaluating policy.” Journal of Economic Literature. 48(2): 356-398.

Hotz, V. Joseph, Guido W. Imbens, and Julie H. Mortimer (2005) “Predicting the
efficacy of future training programs using past experiences at other locations.” Journal of
Econometrics. 125: 241-270.

35
Imbens, Guido S. (2010) “Better LATE than nothing: Some Comments on Deaton (2009)
and Heckman and Urzua (2009).” Journal of Economic Literature. XLVIII(2): 399-423.

Jacob, Brian A. and Jens Ludwig (2011) “The effects of housing assistance on labor
supply: Evidence from a voucher lottery.” American Economic Review.

Jepsen, Christopher and Steven Rivkin (2009) “Class size reduction and student
achievement: The potential tradeoff between teacher quality and class size.” Journal of
Human Resources. 44(1): 223-250.

Kabat-Zinn, J., AO Massion, J Kristeller, LG Peterson, KE Fletcher, L Pbert, WR

Lenderking and SF Santorelli (1992) “Effectiveness of a meditation-based stress
reduction program in the treatment of anxiety disorders.” American Journal of
Psychiatry. 149: 936-943.

Keizer, Kees, Siegwart Lindenberg, and Linda Steg (2008) “The spreading of disorder.”
Science. 322: 1681-1685.

Kelling, George L. and James Q. Wilson (1982) “Broken windows.” The Atlantic
Monthly. (March).
[Link]

Kling, Jeffrey R. (2006) “Incarceration length, employment and earnings.” American

Economic Review. 96(3): 863-876.

Kling, Jeffrey R. (2007) “Methodological Frontiers of Public Finance Field

Experiments.” National Tax Journal 60(1): 109-127.

Kling, Jeffrey R., Jeffrey B. Liebman, and Lawrence F. Katz (2005) “Bullets don’t got no
name: Consequences of fear in the ghetto.” In Discovering Successful Pathways in
Children’s Development: New Methods in the Study of Childhood and Family Life,
Edited by Thomas S. Weisner. Chicago: University of Chicago Press. pp. 243-281.

Kling, Jeffrey R., Jeffrey B. Liebman, and Lawrence F. Katz (2007) “Experimental
analysis of neighborhood effects.” Econometrica. 75(1): 83-119.

Kling, Jeffrey R., Jens Ludwig, and Lawrence F. Katz (2005) “Neighborhood Effects on
Crime for Female and Male Youth: Evidence from a Randomized Housing Voucher
Experiment,” Quarterly Journal of Economics. 120(1): 87-130.

Krueger, Alan B. (1999) “Experimental estimates of education production functions.”

Quarterly Journal of Economics. 114(2): 497-532.

Lazear, Edward (2001) “Educational production.” Quarterly Journal of Economics.

116(3): 777-803.

36
Lipsey, Mark W., Nana A. Landenberger, and Sandra J. Wilson (2007) Effects of
Cognitive-Behavioral Programs for Criminal Offenders. Campbell Systematic Reviews.

Ludwig, Jens, Jeffrey Liebman, Jeffrey Kling, Greg J. Duncan, Lawrence F. Katz, Ronald
C. Kessler and Lisa Sanbonmatsu (2008) “What can we learn about neighborhood effects
from the Moving to Opportunity experiment?” American Journal of Sociology. 114(1):
144-88.

Malani, Anup (2006) “Identifying placebo effects with data from clinical trials.” Journal
of Political Economy.

Manski, Charles F. (2000) “Economic analysis of social interaction.” Journal of

Economic Perspectives. 14(3): 115-136.

Mencken, H .L. 1948 [1998]. “Stare Decisis,” The New Yorker. Reprinted and abridged
in the Wall Street Journal, December 24, 1998, as “A Bum’s Christmas.” At
<[Link]

Meyer, Bruce D. (1995) “Natural and quasi-experiments in economics.” Journal of

Business and Economic Statistics. 13(2): 151-161.

National Association of Community Health Centers. 2010. “Expanding Health Centers

Under Health Care Reform: Doubling Patient Capacity And Bringing Down Costs.”
June. At <[Link]

Newhouse, Joseph P. and the Insurance Experiment Group (1993) Free for All? Lessons
from the RAND Health Insurance Experiment. Cambridge, MA: Harvard University
Press.

Oreopoulos, Philip and Kjell G. Salvanes (2009) “How large are returns to schooling?
Hint: Money isn’t everything.” Cambridge, MA: National Bureau of Economic Research
Working Paper 15339.

Ross, Susan D., I. Elaine Allen, Janet E. Connelly, Bonnie M. Korenblat, M. Eugene
Smith, Daren Bishop, and Don Lou (1999) “Clinical outcomes in statin treatment trials:
A meta-analysis.” Archives of Internal Medicine. 159: 1793-1802.

Rossi, Peter H. (1987) “The iron law of evaluation and other metallic rules.” Research in
Social Problems and Public Policy. 4: 3-20.

Sanbonmatsu, Lisa, Jeffrey R. Kling, Greg J. Duncan, and Jeanne Brooks-Gunn (2006)
“Neighborhoods and academic achievement: Results from the Moving to Opportunity
experiment.” Journal of Human Resources. 41(4): 649-691.

37
Schanzenbach, Diane Whitmore (2007) “What have researchers learned from Project
STAR?” Brookings Papers on Education Policy.

Sen, Bisakha, Stephen Mennemeyer, and Lisa C. Gary (2009) “The relationship between
neighborhood quality and obesity among children.” Cambridge, MA: NBER Working
Paper 14985.

Stuart, Elizabeth A., Stephan R. Cole, Catherine P. Bradshaw, and Philip J. Leaf (2011).
The use of propensity scores to assess the generalizability of results from randomized
trials. The Journal of the Royal Statistical Society, Series A. 174(2): 369-286.

Thavendiranathan, Paaladinesh, Akshay Bagai, M. Alan Brookhart, and Niteesh K.

Choudhry (2006) “Primary prevention of cardiovascular diseases with statin therapy.”
Archives of Internal Medicine. 1666: 2307-2313.

Todd, Petra E. and Kenneth I. Wolpin (2008) “Ex ante evaluation of social programs.”
Working Paper, University of Pennsylvania Department of Economics.

U.S. Census Bureau. 2011. Statistical Abstract of the United States. At

<[Link]

U.S. Department of Education. 2010. “Fiscal Year 2011 Budget Summary — February
1, 2010. Section III. F. Institute of Education Sciences.” At
<[Link]

Wehunt, Jennifer. 2009. “The Food Desert.” Chicago Magazine. July. At

<[Link]

Wilson, William J. (1987) The Truly Disadvantaged: The Inner City, the Underclass, and
Public Policy. Chicago: University of Chicago Press.

Wilson, William J. (1997) When Work Disappears: The World of the New Urban Poor.
Vintage.

Wilt, Timothy J., Hanna E. Bloomfield, Roderick MacDonald, David Nelson, Indulis
Rutks, Michael Ho, Gregory Larson, Anthony McCall, Sandra Pineros, and Anne Sales
(2004) “Effectiveness of statin therapy in adults with coronary heart disease.” Archives of
Internal Medicine. 464: 1427-1436.

Wolpin, Kenneth I. (2007) “Ex ante policy evaluation, structural estimation, and model
selection.” American Economic Review. 97(2): 48-52.

Zhang, Lei, Shuning Zhang, Hong Jiang, Aijun Sun, Yunkai Wang, Yunzeng Zou, Junbo
Ge, and Haozhu Chen (2010) “Effects of statin therapy on inflammatory markers in
chronic health failure: A meta-analysis of randomized controlled trials.” Archives of
Medical Research. 41: 464-471.

38
Figure 1 – Logic Model for Broken Windows Policing

39
 Table 1
Policy Experiment Check-List

Prior beliefs / understanding of mechanisms

Low High
Implications for Run a policy evaluation Run a mechanism
experimental design experiment to rule out
policies (and policy
evaluations)

OR OR

Do more basic science; Run mechanism experiment

multiple methods to to help rule in policies
uncover mechanisms
Either follow with full
policy evaluation
(depending on costs of
policy evaluation, and
potential program benefits /
scale), or use results of
mechanism experiment for
calibration and structural
estimation for key
parameters for benefit-cost
calculations.
Implications for policy Run multiple policy Use mechanism knowledge
forecasting / external evaluations; carry out to measure characteristics
validity policy forecasting by of policy and setting
matching to estimates (moderators) for policy
derived from similar forecasting.
policies and settings
(candidate moderators)
Can run new mechanism
Debate: Which experiments to test in
characteristics to match on? different settings prior to
Where do these come from? carrying out policy
evaluations in those
settings.

40