WPS4752
Policy ReseaRch WoRking PaPeR 4752
In Pursuit of Balance
Randomization in Practice
in Development Field Experiments
Miriam Bruhn
David McKenzie
The World Bank
Development Research Group
Finance and Private Sector Team
October 2008
Policy ReseaRch WoRking PaPeR 4752
Abstract
Randomized experiments are increasingly used in emerge. First, many researchers are not controlling for the
development economics, with researchers now facing method of randomization in their analysis. The authors
the question of not just whether to randomize, but show this leads to tests with incorrect size, and can result
how to do so. Pure random assignment guarantees that in lower power than if a pure random draw was used.
the treatment and control groups will have identical Second, they find that in samples of 300 or more, the
characteristics on average, but in any particular random different randomization methods perform similarly in
allocation, the two groups will differ along some terms of achieving balance on many future outcomes of
dimensions. Methods used to pursue greater balance interest. However, for very persistent outcome variables
include stratification, pair-wise matching, and re- and in smaller sample sizes, pair-wise matching and
randomization. This paper presents new evidence on stratification perform best. Third, the analysis suggests
the randomization methods used in existing randomized that on balance the re-randomization methods common
experiments, and carries out simulations in order to in practice are less desirable than other methods, such as
provide guidance for researchers. Three main results matching.
This paper--a product of the Finance and Private Sector Team, Development Research Group--is part of a larger effort in
the department to develop rigourous methodology for field experiments. Policy Research Working Papers are also posted
on the Web at http://econ.worldbank.org. The author may be contacted at mbruhn@worldbank.org.
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Produced by the Research Support Team
IN PURSUIT OF BALANCE: RANDOMIZATION IN PRACTICE
IN DEVELOPMENT FIELD EXPERIMENTS#
Miriam Bruhn, World Bank
David McKenzie, World Bank, BREAD, CReAM and IZA
Keywords: Randomized experiment; Program evaluation; Development.
JEL codes: C93, O12.
#We thank the leading researchers in development field experiments who participated in our short survey,
as well as colleagues who have shared their experiences with implementing randomization. We thank
Esther Duflo, David Evans, Xavier Gine, Guido Imbens, Ben Olken and seminar participants at the World
Bank for helpful comments, We are also grateful to Radu Ban for sharing his pair-wise matching Stata
code, Jishnu Das for the LEAPS data, and to Kathleen Beegle and Kristen Himelein for providing us with
their constructed IFLS data.. All views are of course our own.
- 1 -
1. Introduction
Randomized experiments are increasingly used in development economics.
Historically, many randomized experiments were large-scale government-implemented
social experiments, such as Moving to Opportunity in the U.S. or
Progresa/Oportunidades in Mexico. These experiments allowed for little involvement of
researchers in the actual randomization. In contrast, in recent years many experiments
have been directly implemented by researchers themselves, or in partnership with NGOs
and the private sector. These small-scale experiments, with sample sizes often comprising
100 to 500 individuals, or 20 to 100 schools or health clinics, have greatly expanded the
range of research questions that can be studied using experiments, and have provided
important and credible evidence on a range of economic and policy issues. Nevertheless,
this move towards smaller sample sizes means researchers increasingly face the question
of not just whether to randomize, but how to do so. This paper provides the first
comprehensive look at how researchers are actually carrying out randomizations in
development field experiments, and then analyzes some of the consequences of these
choices.
Simple randomization ensures the allocation of treatment to individuals or
institutions is left purely to chance, and is thus not systematically biased by deliberate
selection of individuals or institutions into the treatment. Randomization thus ensures that
the treatment and control samples are, in expectation, similar in average, both in terms of
observed and unobserved characteristics. Furthermore, it is often argued that the
simplicity of experiments offers considerable advantage in making the results convincing
to other social scientists and policymakers and that, in some instances, random
assignment is the fairest and most transparent way of choosing the recipients of a new
pilot program (Burtless, 1995).
However, it has long been recognized that while pure random assignment
guarantees that the treatment and control groups will have identical characteristics on
average, in any particular random allocation, the two groups will differ along some
dimensions, with the probability that such differences are large falling with sample size.1
1For example, Kernan et al. (1999) consider a binary variable that is present in 30 percent of the sample.
They show that the chance that the two treatment group proportions will differ by more than 10 percent is
- 2 -
Although ex-post adjustment can be made for such chance imbalances, this is less
efficient than achieving ex-ante balance, and can not be used in cases where all
individuals with a given characteristic are allocated to just the treatment group.
The standard approach to avoiding imbalance on a few key variables is
stratification (or blocking), originally proposed by R.A. Fisher. Under this approach,
units are randomly assigned to treatment and control within strata (or blocks) defined by
usually one or two observed baseline characteristics. However, in practice it is unlikely
that one or two variables will explain a large share of the variation in the outcome of
interest, leading to attempts to balance on multiple variables. One such method when
baseline data are available is pair-wise matching (Greevy et al, 2004, Imai et al. 2007).
The methods of implementing randomization have historically been poorly
reported in medical journals, leading to the formulation of the CONSORT guidelines
which set out standards for the reporting of clinical trials (Schulz, 1996). The recent
explosion of field experiments in development economics has not yet met these same
standards, with many papers omitting key details of the method in which randomization
is implemented. For this reason, we conducted a survey of leading researchers carrying
out randomized experiments in developing countries. This reveals common use of
methods to improve baseline balance, including several re-randomization methods not
discussed in print. These are (i) carrying out an allocation to treatment and control, and
then using a statistical threshold or ad hoc procedure to decide whether or not to redraw
the allocation; and (ii) drawing 100 or 1000 allocations to treatment and control, and
choosing the one amongst them which shows best balance on a set of observable
variables.
This paper discusses the pros and cons of these different methods for striving
towards balance on observables. Proponents of methods such as stratification, matching,
and minimization claim that such methods can improve efficiency, increase power, and
protect against type I errors (Kernan et al., 1999) and do not seem to have significant
disadvantages, except in small samples (Imai et al. 2008, King et al. 2007, Greevy et al.
38% in an experiment with 50 individuals, 27% in an experiment with 100 individuals, 9% for an
experiment with 200 individuals, and 2% for an experiment with 400 individuals.
- 3 -
2004, Aickin, 2001)2. However, it is precisely in small samples that the choice of
randomization method becomes important, since in large samples all methods will
achieve balance. We simulate different randomization methods in four panel data sets.
We then compare balance in outcome variables at baseline and at follow-up. The
simulations show that when methods other than pure randomization are used, the degree
of balance achieved on baseline variables is much greater than that achieved on the
outcome variable (in the absence of treatment) in the follow-up period. The simulations
show further that in samples of 300 observations or more, the choice of method is not
very important for the degree of balance in many outcomes at follow-up. In small
samples, and with very persistent outcomes, however, matching or stratification on
relevant baseline variables achieves more balance in follow-up outcomes than does pure
randomization.
We use our simulation results and theory to help answer many of the important
practical questions facing researchers engaged in randomized experiments. The results
allow us to provide guidance on how to conduct inference after stratification, matching or
re-randomization. In practice it appears that many researchers ignore the method of
randomization in inference. We show that this leads to hypothesis tests with incorrect
size. On average, the standard errors are overly conservative when the method of
randomization is not controlled for in the analysis, implying that researchers may not
detect treatment effects that they would detect if the inference did take into account the
randomization method. However, although this is the case on average, in a non-trivial
proportion of draws, it will be the case that not controlling for the randomization method
will be anti-conservative, potentially leading the researcher to find a significant effect
that is no longer significant when stratum or pair dummies are included. Moreover, we
show further that stratifying, matching, or re-randomizing and then analyzing the data
without controlling for the method of randomization results in lower power than if a pure
random draw was used to allocate treatments, except in cases where the variables that
balance is sought for have no predictive power for the future outcome of interest (in
which case there is no need to seek balance on them anyway).
2One other arguments in favor of ex-ante balancing is that, if the treatment effect is heterogeneous and
varies with observed covariates, ex-ante balancing increases the precision of subgroup analysis.
- 4 -
We therefore strongly recommend that inference account for the method of
randomization. Moreover, the results suggest that the common use of re-randomization
methods should be rethought, since the method often performs worse than pair-wise
matching in terms of balance and power, and requires more complicated statistical
analysis to account for the effect of re-randomizing.
The paper also discusses the use and abuse of tests for baseline differences in
means, the impact of balancing observables on achieving balance on unobservables, and
the issue of how many (and which) variables to use for stratifying or matching. The
downside of balancing on many variables or matching is a loss in the degrees of freedom
available for averaging out the variation coming from unobservables when estimating the
variance.
This paper draws upon a large clinical trials literature, where many related issues
have been under discussion for several decades, drawing out the lessons for development
field experiments. It complements several recent papers in development on randomized
experiments.3 The paper builds on the recent handbook chapter by Duflo, Glennerster and
Kremer (2006), which aims to provide a "how to" of implementing experiments. Our
focus differs, considering how the actual randomization is implemented in practice, and
considering matching and re-randomization approaches not discussed in this recent work.
Finally, we contribute to the existing literature through new simulations which illustrate
the performance of the different methods in a variety of situations experienced in
practice.
Whilst our focus is on field experiments in development economics, to date the
field with most active involvement of researchers in randomization, randomized
experiments are also increasingly being used to investigate important policy questions in
other fields (Levitt and List, 2008). In common with the development literature, the
extant literature in these other fields has often not explained the precise mechanism used
for randomizing. However, it does appear that re-randomization methods are also being
employed in some of these studies. The ongoing New York public schools project being
undertaken by the American Inequality Lab is one such high-profile example. The
3Summaries of recent experiments and advocacy of the policy case are found in Kremer (2003), Duflo and
Kremer (2004), Duflo (2005) and Banerjee (2007).
- 5 -
lessons of this paper will also be important in designing upcoming experiments in other
fields of economics.
The remainder of the paper is set out as follows. Section 2 provides a stocktaking
of how randomization is currently being implemented in the field, drawing on a summary
of papers and a survey of leading experts in development field experiments. Section 3
describes the data sets used in our simulations, and outlines in more detail the different
methods of randomization. Section 4 then provides simulation evidence on the relative
performance of the different methods, and on answers to key questions faced in practice.
Section 5 concludes.
2. How is randomization being implemented?
2.1. Randomization as described in papers
We begin by reviewing a selection of research papers containing randomized
experiments in development economics. We focus on relatively small-scale randomized
experiments, typically implemented via NGOs or as pilot studies. Duflo et al. (2006)
argue that such designs typically allow for more involvement by researchers, who can
often influence program design (and in particular, have input into how randomization is
implemented). The majority of such studies appear to have some baseline data available
at the time of randomization. In cases where baseline data is not available, pure
randomization seems to be used to assign units to treatment.
Table 1 summarizes a selection of randomized experiments which took place with
baseline data. This listing is not comprehensive, but is intended to cover many of the
most well-known published randomized experiments in development, and a selection of
newer working papers, and to cover papers by many of the leading proponents of
randomized evaluations.4 For each study we list the unit at which randomization occurs.
Typical sample sizes are 100 to 300 units, with the smallest sample size being 10
geographic areas used in Ashraf et al. (2006b).
Randomized experiments are often argued to provide a fair and transparent way of
allocating scarce resources when piloting or rolling out a program. This transparency is
4We do not include here experiments undertaken by the authors, such as de Mel et al. (2008), both for
objectivity reasons, and because the final write-up of these papers has been influenced by the current paper.
- 6 -
greatest to the program participants when assignment to treatment is done in public. The
column "done in public or private" therefore records whether the actual randomization
was done publicly or privately. In between lies "semi-public", where perhaps the NGO
and/or Government officials witness the randomization draw, but not the recipients of the
program. Only 2 out of the 18 papers reviewed note whether it was public or not in both
cases public lotteries (Field and Pande, 2008 and Bertrand et al. 2007). The majority of
the other randomizations we believe are private or at most "semi-public", but this is not
stated explicitly in the papers. Thus, for the most part, the idea that random assignment to
treatment provides a transparent way of allocating resources is not born out in the
existing descriptions.
Next we examine which methods are being used to reduce the likelihood of
imbalance on observable covariates. Thirteen studies use stratification, two use matched
pairs, and only three appear to use pure randomization. Ashraf et al. (2007) is the only
documented example we have found of one of the methods that the next section shows to
be in common use in our survey of experts. They note "at the time of randomization, we
verified that observable characteristics were balanced across treatments, and, in a few
cases, re-randomized when this was not the case". They do not explain the criteria used to
decide whether or not the degree of imbalance was sufficient for re-randomization to take
place.
Few papers provide the details of the method used, presumably because there has
not been a discussion of the potential importance of these details in the economics
literature. For example, stratification is common, but few studies actually give the
number of strata used in the study.5 As we will discuss, the choice of the number of strata
to use involves a trade-off between reducing residual variation and losing degrees of
freedom when estimating the standard error of the experimental estimator. Studies which
do not report the number of strata therefore make it difficult to assess this trade-off. The
number of strata can be substantial in some studies for example, Olken (2007a) uses
156 subdistricts as strata for his 608 villages. In practice there appears to be
5 For example, Banerjee et al. (2007) write "assignment was stratified by language, pretest score, and
gender". In this case pre-test score is continuous, and it is not clear how it was discretized for stratification
purposes. Likewise, Karlan and Valdivia (2006) note that "randomization was stratified by credit officer"
without stating how many credit officers there were.
- 7 -
disagreement as to whether it is necessary to include strata dummies in the analysis after
stratification more than half the studies using stratification do not include strata
dummies.
Finally, we note that all but one of the papers in Table 1 present a table for
comparing treatment and control groups, and carry out tests for imbalance.6 The number
of variables used for checking for imbalance ranges from 4 to 39. We examine the
purpose and usefulness of such tests in more detail in Section 4.
2.2 Randomization in practice according to a survey of experts
The long lag between inception of a randomized experiment and its appearance in
at least working paper form means the results above do not necessarily represent how the
most recent randomized evaluations are being implemented. We therefore decided to
survey leading experts in randomized evaluations on their experience and approach to
implementation. A short online survey was sent to 35 selected researchers in December
2007. The list was selected from members of the Poverty Action Lab, BREAD, and the
World Bank who were known to have conducted randomized experiments. We had 25 of
these experts answer the survey, with 7 out of the 10 individuals who did not respond
having worked with those who did respond, ensuring our survey also covers the methods
used by those non-responders in at least some of their experiments.
The median researcher surveyed had participated in 5 randomized experiments,
with a mean of 5.96.7 Three of those surveyed had only participated in one experiment,
while three had participated in fifteen (or more). 71 percent of the experiments these
researchers had been involved in had had baseline data (including administrative data)
that could be used at the time when randomization to treatment was done, while in the
remaining 29 percent no baseline data was available.
Preliminary discussions with several leading researchers established that in
addition to stratified randomization and matched pairs, several other methods involving
multiple random draws were being used in practice to increase the likelihood of balance
6The exception is Field and Pande (2008), who are likely limited in space in the papers and proceedings
format. They note that such a check was done and is available upon request.
7This is after top-coding the number of experiments at 15, in order to not have the responses dominated by
one researcher who had conducted more experiments than this.
- 8 -
on observed characteristics. One such approach is to take a random draw of assignment to
treatment, examine the difference in means for several key baseline characteristics, and
then re-randomize if the difference looks too large. This decision as to what is too large
could be done subjectively, or according to some statistical cutoff criteria. For example,
one survey respondent noted that they "regressed variables like education on assignment
to treatment, and then re-did the assignment if these coefficients were `too big'".
The second approach takes many draws of assignment to treatment, and then
chooses the one that gives best balance on a set of observable characteristics according to
some algorithm or rule. For example, several researchers say they write a program to
carry out 100 or 1000 randomizations, and then for each draw, regress individual
variables against treatment. They then choose the draw with the minimum maximum t-
statistic. Some impose further criteria such as requiring the minimum maximum t-statistic
for testing balance on observables to be below one. An alternative approach used by
another researcher is to regress the treatment on the set of baseline covariates and choose
the draw with the lowest R2. The number of variables used to check balance varies, but
seems to typically range from 5 to 20, and includes if possible the baseline levels of the
main study outcomes. The perceived advantage of this approach is to enable balance on
many more variables than possible with stratification, and to provide balance in means on
continuous variables.
Researchers were asked whether they had ever used a particular method, and the
method used in their most recent randomized experiment. All of the methods are often
combined with some stratification, so we examine that separately. Table 2 reports the
results. Not surprisingly, most researchers have at some point used simple randomization
(probably with some stratification) 80 percent of the full sample and 94 percent of
researchers who have carried out five or more experiments have done this. However, we
also see much more use of other methods than is apparent from the existing literature. 56
percent had used pair-wise matching, with 39 percent using it in their most recent
experiment. 32 percent of all researchers and 46 percent of the 5 or more experiments
group have subjectively decided whether to re-randomize based on an initial test of
balance. The multiple draws process described above has also been used by 24 percent of
- 9 -
researchers, and is more common amongst the more experienced researchers with 38
percent of the 5 or more experiment group using this method.
More detailed questions were asked about the most recent randomization, in an
effort to obtain some of the information not provided in Table 1. 23 of the 25 respondents
provided information on these, and none of the responses are duplicate answers for the
same experiment. First, in terms of whether the randomization is done in public or
private, 17 of the 23 were done privately, 3 were done in public (including one where
participants simply chose their own assignment from a hat), and 3 were done with the
implementing agency observing but not the participants. One of the researchers who had
carried out the randomization in public noted "carrying out the randomizations in public
with the participants in attendance is a good idea. It seems to remove most possible ill
feelings when the individuals get to participate in a "game" that determines the outcome
in terms of treatment assignment". The potential downside of public drawing is that some
of the methods used to ensure balance become more difficult, if not impossible, to
implement8.
Stratification was used in 14 out of the 15 experiments that were not employing a
matched pair design. The number of variables used in forming strata was small: 6 used
only one variable, typically geographic location; 4 used two variables (e.g. location and
gender), and 4 used four variables. Of particular note is that it appears rare to stratify on
baseline values of the outcome value of interest (e.g. test scores, savings levels, or
incomes) with only 2 of these 14 experiments including a baseline outcome as a
stratifying factor. While the number of stratifying variables is small, there is much greater
variation in the number of strata: ranging from 3 to 200, with a mean (median) of 47
(18). The number of treated observations divided by the number of strata ranges from 1 to
800, with a median of 36. Only one researcher said that stratification was controlled for
when calculating standard errors for their treatment effect.
8A further example of one of the potential downsides in public randomizations is seen in an example
provided by one study in Indonesia, in which survey respondents picked one of three balls from an opaque
bag to determine the size of a financial incentive they were to receive. Although there was one of each ball
in the bag, 80 percent of the respondents ended up receiving the highest value of the incentive. As a result,
the randomization had to be abandoned, and reassignment to treatment status was done privately, out of the
interviewers control.
- 10 -
A notable feature of the survey responses was a much greater number of
researchers randomizing within matched pairs than is apparent from the existing
literature. However, the vast majority of these matches were not done using optimal or
greedy Mahalnobis matching, but were instead based on only a few variables and
commonly done by hand. In most cases the researchers were matching on discrete
variables and their interactions only, and thus, in effect, the matching reduced to
stratification.
In terms of the follow-up period for determining the treatment effect, the most
common responses were one year and two years, followed by 6 months. Several studies
aim to follow-up at six months and one year. Only one of the studies noted a plan to
follow-up beyond two years, after taking initial results at 8 months and at 2 years. This
information on the length of time commonly used for follow-ups will be used in the next
section when discussing what we should seek balance on.
One explanation for the difference in randomization approaches used by different
researchers is that they reflect differences in context, with sample size, question of
interest, and organization one is working with potentially placing constraints on the
method which can be used for randomization. We therefore asked researchers for advice
on how to evaluate the same hypothetical intervention designed to raise the incomes of
day laborers.9 The responses varied greatly across researchers, and include each of the
methods given in Table 2. What is clear is that there appears to be no general agreement
about how to go about randomizing in practice.
3. Data, simulated methods, and variables for balancing
This section provides an overview of the four panel data sets used in this paper. It
then discusses the different randomization methods that we simulate in these data sets
and the variables considered for achieving balance.
3.1 Data
To compare the performance of the different randomization methods in practice,
we chose four panel data sets which allow us to examine a wide range of potential
9See Appendix 1 for the exact question and the responses given.
- 11 -
outcomes of interest, including microenterprise profits, labor income, school attendance,
household expenditure, test scores, and child anthropometrics.
The first panel data set covers microenterprises in Sri Lanka and comes from de
Mel et al. (2008). This data was collected as part of an actual randomized experiment, but
we keep only data for firms that were in the control group during the first treatment
round. The data set contains information on firms' profits, assets and many other firm
characteristics. It also includes detailed measures of the firm owners' entrepreneurial
ability, risk aversion, and other characteristics that are thought to be correlated with
profits. The simulations we perform for this data set are meant to mimic a randomized
experiment that administers a treatment aimed at increasing firms' profits, such as a
business training program.
The second data set is a sub-sample of the Mexican employment survey (ENE).
Our sub-sample includes heads of household between 20 and 65 years of age who were
first interviewed in the second quarter of 2002 and who were re-interviewed in the
following four quarters. We only keep individuals who were employed during the
baseline survey and imagine a treatment that aims at increasing their income, such as a
training program or a nutrition program.
The third data set comes from the Indonesian Family Live Survey (IFLS).10 We
use 1997 data as the baseline and 2000 data as the follow-up, and simulate two different
interventions with the IFLS data. First, we keep only children aged 10-16 in 1997 that
were in the 6th grade and in school. These children then receive a simulated treatment
aimed at keeping them in school (in the actual data, about 26 percent have dropped out 3
years later). Second, we create a sample of households and simulate a treatment that
increases household expenditure per capita.
The fourth data set comprises child and household data from the LEAPS project
in Pakistan (Andrabi et al. 2008). We focus on children aged 8 to 12 at baseline and
examine two child outcome variables: math test scores and height z-scores11. The
simulated treatments increase test scores or z-scores of these children. There is a wide
10See http://www.rand.org/labor/FLS/IFLS/.
11We also have data on English test scores and weight z-scores and performed all simulations with these
outcomes. The results are very close to the results using math test scores and height z-scores and are
available from the authors upon request.
- 12 -
range of policy experiments that have targeted these types of outcomes, from providing
text books or school meals to giving conditional cash transfers or nutritional supplements.
3.2 Simulated methods
For all data sets, we draw three sub-samples of 30, 100, and 300 observations
each to investigate how the performance of different methods varies with sample size. All
results are based on 10,000 bootstrap iterations of each method. The simulations
randomly split the sample into a treatment group and a control group, according to five
different methods. The first method is a single random draw.
3.2.1 Stratification
The second method is stratification. Stratified randomization is the most well-
known, and as we have seen, commonly used method of preventing imbalance between
treatment and control groups for the observed variables used in stratification. By
eliminating particular sources of differences between groups, stratification (aka blocking)
can increase the sensitivity of the experiment, allowing it to detect smaller treatment
differences than would otherwise be possible (Box et al, 2005). The most often perceived
disadvantage of stratification compared to some alternative methods is that only a small
number of variables can be used in forming strata.12
In terms of which variables to stratify on, the literature emphasizes variables
which are strongly related to the outcome of interest, and variables for which subgroup
analysis is desired. Statistical efficiency is greatest when the variables chosen are
strongly related to the outcome of interest (Imai et al., 2008). Stratification is not able to
remove all imbalance for continuous variables. For example, for two normal distributions
with different means but the same variance, the means of the two distributions between
any two fixed variables (i.e. within a stratum) will differ in the same direction as the
overall mean (Altman, 1985). In the simulations, we always stratify on the baseline
values of the outcome of interest and on one or two other variables, which either relate to
the outcome of interest or constitute relevant subgroups for ex-post analysis.
12This is particularly true in small samples. For example, considering only binary or dichotomized
characteristics, with 5 variables there are 25 = 32 strata, while 10 variables would give 210 = 1024 strata. In
our samples of 30 observations, we stratify on 2 variables, forming 8 strata. In the samples of 100 and 300
observations, we also stratify on 3 variables (24 strata), and also on 4 variables (48 strata).
- 13 -
3.2.2 Pair-wise matching
As a third method, we simulate pair-wise matching. As opposed to stratification,
matching provides a method to improve covariate balance for many variables at the same
time. Greevy et al. (2004) describe the use of optimal multivariate matching. However,
we chose to use the less computationally intensive "optimal greedy algorithm" laid out in
King et al. (2007)13. In both cases pairs are formed so as to minimize the Mahalanobis
distance between the values of all the selected covariates within pairs, and then one unit
in each pair is randomly assigned to treatment and the other to control.
As with stratification, matching on covariates can increase balance on these
covariates, and increase the efficiency and power of hypothesis tests. King et al. (2007)
emphasize one additional advantage in the context of social science experiments when
the matched pairs occur at the level of a community or village or school, which is that it
provides partial protection against political interference or drop-out. If a unit drops out of
the study or suffers interference, its pair unit can also be dropped from the study, while
the set of remaining pairs will still be as balanced as the original data set. However, the
converse of this is that if units drop out at random, the matched pair design will throw out
the corresponding pairs as well, leading to a reduction in power and smaller sample size
than if an unmatched randomization was used.14
3.2.3 Re-randomization methods
Since our survey revealed that several researchers are using re-randomization
methods, we simulate two of these methods. The first, which we dub the "big stick"
method by analogy with Soares and Wu (1983), requires a re-draw if a draw shows any
statistical difference in means between treatment and control group at the 5 percent level
or lower. The second method picks the draw with the minimum maximum t-stat out of
1000 draws.
13 The Stata code performing pair-wise Mahalanobis matching with an optimal greedy algorithm takes
several days to run in the 300 observations sample. If there is little time in the field to perform the
randomization this may thus not be an option. It is thus important to have ample time between receiving
baseline data and having to perform the randomization to have the flexibility of using matching techniques
if desired. Software packages other than Stata may be more suited for this algorithm and may speed up the
process.
14 See Greevy et al. (2004) for discussion of methods to retain broken pairs.
- 14 -
Although we are not aware of any papers which formally set out the re-
randomization methods used in practice in development, there are analogs in the
sequential allocation methods used in clinical trials (Soares and Wu, 1983; Taves, 1974;
Pocock and Simon, 1975). The use of these related methods remains somewhat
controversial in the medical field. Proponents emphasize the ability of such methods to
improve balance on up to 10 to 20 covariates, with Treasure and MacRae (1998)
suggesting that if randomization is the gold standard, minimization may be the platinum
standard. In contrast, the European Committee for Proprietary Medicinal Products
(CPMP, 2003) recommends that applicants avoid such methods, and argues that
minimization may result in more harm than good, bringing little statistical benefit in
moderate sized trials.
Why might researchers wish to use these methods instead of stratification? Imai,
King and Stuart (2008) argue that the practice of re-randomizing when the first set of
random draws is too imbalanced can be thought of as an inefficient form of blocking.
However, as noted, in small samples, stratification is only possible on one or two
variables. There may be many variables that the researcher would like to ensure are not
"too unbalanced", without requiring exact balance on each. Re-randomization methods
may be viewed as a compromise solution by the researchers, preventing extreme
imbalance on many variables, without forcing close balance on each.
3.3 Variables for balancing
In practice researchers will attempt to balance on variables they think are strongly
correlated with the outcome of interest. The baseline level of the outcome variable is a
special case of balancing on a variable believed to be correlated with the outcome. We
always include the baseline outcome variable among the variables to stratify, match or
balance on. Note that this is somewhat the exception in practice, where researchers have
often not balanced on the baseline outcome. In the matching and re-randomization
methods, we also use six additional baseline variables that are thought to affect the
outcome of interest. Stratification takes a subset of these six additional variables.15
15A list of the variables used for each dataset is in Appendix 2 (Table A2).
- 15 -
Among these balancing variables, we tried to pick variables that are likely to be
correlated with the outcome based on economic theory and existing data. There is,
however, a caveat. As we have seen, most economic interventions have impacts
measured over periods of 6 months to 2 years. While our economic models and existing
data sets often provide good information for deciding on a set of variables useful for
explaining current levels, they are often much less useful in explaining future levels of
the variable of interest. In practice over short time horizons, often we can not
theoretically or empirically explain many changes well with observed variables and
believe that these changes are the result of shocks. As a result, it may be the case in
practice that the covariates used to obtain balance on are not strong predictors of future
values of the outcome of interest.
The set of outcomes we have chosen spans a range of the ability of the baseline
variables to predict future outcomes. At one end are microenterprise profits in Sri Lanka,
where baseline profits and six baseline individual and firm characteristics explain only
12.2 percent of the variation in profits six months later. Thus balancing on these common
owner and firm characteristics will not control for very much of the variation in future
realizations of the outcome of interest. School enrolment in the IFLS data is another
example where baseline variables explain very little of future outcomes. For a sample of
300 students who were all in school at baseline, 7 baseline variables only explain 16.7
percent of the variation in school enrollment for the same students 3 years later. The
explanatory power is better for labor income in the Mexican ENE data and household
expenditure in the IFLS, with the baseline outcome and six baseline variables explaining
28-29 percent of the variation in the future outcome. The math test scores and height z-
scores in the LEAPS data have the most variation explained by baseline characteristics,
with 43.6 percent of the variation in follow-up test scores explained by the baseline test
score and six baseline characteristics.
We expect to see more difference amongst randomization methods in terms of
achieving balance on future outcomes for the variables that are either more persistent, or
that have a larger share of their changes explained by baseline characteristics. We
therefore expect to see least difference among methods for the Sri Lanka microenterprise
- 16 -
profits data and Indonesian school enrolment data, and most difference for the LEAPS
math test score and height z-score data.
4. Simulation results
Appendix 316 reports the full set of simulation results for all four data sets for 30,
100, and 300 observations. We summarize the results of these simulations in this section,
organizing their discussion around several central questions that a researcher may have
when performing a randomized assignment. We start by addressing the following core
question:
4.1 Which methods do better in terms of achieving balance and avoiding extremes?
We first compare the relative performance of the different methods in achieving
balance between the treatment and control groups in terms of baseline levels of the
outcome variable. Table 3 shows the average difference in baseline means, the 95th
percentile of the difference in means (a measure of the degree of imbalance possible at
the extremes), and the percentage of simulations where a t-test for difference in means
between the treatment and control has a p-value less than 0.10. We present these results
for a sample size of 100, with results for the other sample sizes contained in Appendix 3.
Figures 1a through 6c graphically summarize the results, plotting the densities of the
differences in average outcome variables for all three sample sizes: 30, 100, and 300
observations.
Table 3 shows that the mean difference in baseline means is very close to zero for
all methods on average all methods of randomizing lead to balance. However, Table 3
and the figures also show that stratification, matching, and especially the minmax t-stat
method have much less extreme differences in baseline outcomes, while the big stick
method only results in narrow improvements in balance over a single random draw. For
example, in the Mexican labor income data with a sample of 100, the 95th percentile of
the difference in baseline mean income between the treatment and control groups is 0.384
standard deviations (s.d.) with a pure random draw, 0.332 s.d. under the big stick method,
0.304 s.d. when stratifying on 4 variables, 0.099 s.d. with pair-wise greedy matching, and
16Appendix 3 is available on the author's website.
- 17 -
0.088 under the minmax t-stat method. The size of the difference in balance achieved
with different methods shrinks as the sample size increases asymptotically all methods
will be balanced.
The key question is then the extent to which achieving greater balance on baseline
variables translates into better balance on future values of the outcome of interest in the
absence of any treatment. The follow-up period is six months for the Sri Lankan
microenterprise data and Mexican labor income data, one year for the Pakistan test-score
and child height data, and three years for the Indonesian schooling and expenditure data.
Figures 1 to 6 show the distribution of difference in means between treatment and control
at follow-up for each method, while Table 4 summarizes how the different methods
perform in obtaining balance in follow-up outcomes.
Panel A of Table 4 shows that on average, all randomization methods give
balance on the follow-up variable, even with a sample size as small as 30. This is the key
virtue of randomization. Figures 1 to 6 and Panel B shows there are generally fewer
differences across methods in terms of avoiding extreme imbalances than with the
baseline data. This is particularly true of the Sri Lanka profit data and the Indonesian
schooling data, for which baseline variables explained relatively little of future outcomes.
With a sample size of 30, stratification and matching are reducing extreme differences
between treatment and control, but with samples of 100 or 300, there is very little
difference between the various methods in terms of how well they balance the future
outcome.
Baseline variables have more predictive power for the realizations at follow-up
for the other outcomes we consider. The Mexican labor income and Indonesian
expenditure data lie in an intermediate range of baseline predictive power, with the
baseline outcomes plus six other variables explaining about 28 percent of the variation in
follow-up outcomes. Figures 2a to 2c and 4a to 4c show that, in contrast to the Sri Lanka
and IFLS schooling data, even with samples of 100 or 300 we find matching and
stratification continue to perform better than a single random draw in reducing extreme
imbalances. Table 4 shows that with a sample size of 300, the 95th percentile of the
difference in means between treatment and control groups is 0.23 s.d. under a pure
random draw for both expenditure and labor income. This difference falls to 0.20 s.d. for
- 18 -
expenditure and 0.15 s.d. for labor income when pair-wise matching is used, and to 0.20
s.d. for both variables when stratifying or using the min-max re-randomization method.
Our other two outcomes variables, math test scores and height z-scores lie in the
higher end of baseline predictive power, with the baseline outcome and six other
variables predicting 43.6 percent and 35.3 percent of the variation in follow-up outcomes,
respectively. Figures 5a to 6c illustrate that the choice of method makes more of a
difference for these highly predictable follow-up outcomes than for the less predictable
ones. Stratifying, matching, and the minmax t-state method consistently lead to narrower
distributions in the differences at follow-up when test scores or height z-scores are the
outcomes. Nevertheless, even with these more persistent variables, the gains from
pursuing balance on baseline are relatively modest when the sample size is 300 using
pair-wise matching rather than a pure random draw reduces the 95th percentile of the
difference in means from 0.23 to 0.17 in the case of math test scores.
4.2 What does balance on observables imply about balance on unobservables?
In general, what does balancing on observables do in terms of balancing
unobservables? Aickin (2001) notes that methods which balance on observables can do
no worse than pure randomization with regard to balancing unobserved variables.17 We
illustrate this point empirically in the Sri Lanka and ENE datasets by defining a separate
group of variables from the data to be "unobservable" in the sense that we do not balance,
stratify or match on them. The idea here is that, although we have these variables in these
particular data sets, they may not be available in other data sets (such as measures of
entrepreneurial ability). Moreover, these "unobservables" are meant to capture what
balancing does to variables that are thought to have an effect on the outcome variable, but
17 To see this, consider balancing on variable X, and the consequences of this for balance on an unobserved
variable W. W can be written as the sum of the fitted value from regressing W on X, and the residual from
this regression:
W = PXW + (I - PX )W (1)
PX = X (X ' X )-1 X '
Balancing on X will therefore also balance the part of W which is correlated with X, PXW. Then, since the
remaining part of W, (I-PX)W is orthogonal to X, it will tend to balance at the same rate as under pure
randomization.
- 19 -
are truly unobservable. Table 3 indicates that the balance on these unobservables is pretty
much the same across all methods.
Rosenbaum (2002, p. 21) notes that under pure randomization, if we look at a
table of observed covariates and see balance "this gives us reason to hope and expect that
other variables, not measured, are similarly balanced". This holds true for pure random
draws, but will not be the case with methods which enhance balance on certain observed
covariates. Presenting a table which shows only the variables used in matching or for re-
randomization checks, and showing balance on these covariates, will thus overstate the
degree of balance attained on other variables that are not closely correlated with those for
which balance was pursued. For example, the 95th percentile of the difference in means in
Table 3 gives a similar level of imbalance for the unobservables as the balanced outcome
under a pure random draw, whereas under the other methods the unobservables have
higher imbalance than the outcome variable.18 We therefore recommend that if matching
or re-randomization (or stratification on continuous variables) is used, researchers clearly
separate these from other variables of interest when presenting a table to show balance.
4.3 To dummy or not to dummy?
We have seen that only a fraction of studies using stratification control for strata
in the statistical analysis. Kernan et al. (1999) state that results should take account of
stratification, by including strata as covariates in the analysis. Failure to do so results in
overly conservative standard errors, which may lead a researcher to erroneously fail to
reject the null hypothesis of no treatment effect. While the omission of balanced
covariates will not change the point estimates of the effect in linear models, leaving out a
balanced covariate can change the estimate of the treatment effect in non-linear models
(Raab et al. 2000), so that analysis of binary outcomes makes this adjustment more
important. The European Committee for Proprietary Medicinal Products (CPMP, 2003)
also recommends that all stratification variables be included as covariates in the primary
analysis, in order to "reflect the restriction on randomization implied by the
18Note the imbalance on unobservables is similar to that of a single random draw, which concurs with the
point that balancing on observables can do no worse than pure randomization when it comes to balancing
unobservables.
- 20 -
stratification". Similarly, for pair-wise matching, dummies for each pair should be
included in the treatment regression.
Furthermore, in practice, stratification is unlikely to achieve perfect balance for
all of the variables used in stratification. Whenever there is an odd number of units within
a stratum, there will be imbalance (Therneau, 1993). In addition, imbalance may arise
from units having a baseline missing value on one of the variables used in forming strata.
As a consequence, in practice, the point estimate of the treatment effect will also likely
change if strata dummies are included compared to when they are not included.
To examine whether or not controlling for stratification matters in practice, Panels
C and D of Table 4 compare the size of a hypothesis test for the difference in means of
the follow-up outcome when no treatment has been given. Panel C shows the proportion
of p-values under 0.10 when no stratum or pair dummies are included, and Panel D
shows the proportion of p-values under 0.10 when these dummies are included. Recall
that this is a test of a null hypothesis which we know to be true, so to have correct size,
10 percent of the p-values should be below 0.10. We see that this is the case for the pure
random draw, whereas failure to control for the dummies leads the stratification and pair-
wise matching tests to be too conservative on average.19 For example, with a sample size
of 30, less than 5 percent of the p-values are below 0.10 for all six outcomes when we
don't include pair dummies with pair-wise matching. For the math test score, only 0.6
percent of the p-values under stratification and none of the p-values under pair-wise
matching are under 0.10. Even with a sample size of 300, less than 5 percent of the p-
values are below 0.10 for the more persistent outcomes when stratification or matching is
used but not accounted for by adding stratum or pair dummies. In contrast, Panel D
shows that when we add stratum dummies or pair dummies, the hypothesis test has the
correct size, with 10 percent of the p-values under 0.10, even in sample sizes as small as
30.
Thus, on average, it is overly conservative to not include the controls for stratum
or pair in analysis. The resulting conservative standard errors imply that if researchers do
not account for the method of randomization in analysis, they may not detect treatment
19The child schooling in Indonesia is a binary outcome. The difference in means attending school can
therefore be only a limited number of discrete differences, and this discreteness causes the test to not have
the correct size even under a pure random draw when the sample is small.
- 21 -
effects that they would otherwise detect. However, although on average the p-values are
lower when including these dummies, Table 5 shows that this is not necessarily the case
in any particular random allocation to treatment and control. Including stratum dummies
only lowers the p-value in 58 to 88 percent of the replications, depending on sample size
and outcome variable. Thus in practice, in a non-trivial proportion of draws, it will be the
case that not including stratum dummies will be anti-conservative, potentially leading the
researcher to find a significant effect that is no longer significant when stratum dummies
are controlled for. Hence researchers can not argue that if they ignore the randomization
method, and find significant effects treating their study as if they purely randomized, that
these same treatment effects will necessarily remain significant if one were to account for
the method of randomization.
4.4 How should inference be done after re-randomizing?
While including strata or pair dummies in the ex-post analysis for the
stratification and matching methods is quite straight-forward, the methods of inference
are not as clear for re-randomization methods. In fact, the correct statistical methods for
covariate-dependent randomization schemes such as minimization are still a conundrum
in the statistics literature, leading some to argue that the only analysis that we can be
completely confident about is a permutation test or re-randomization test. Randomization
inference can be used for analysis of the method of re-randomizing when the first draw
exceeds some statistical threshold (although it requires additional programming work).
Using the rule which determines when re-randomization will take place, the researcher
can map out the set of random draws which would be allowed by the threshold rule,
throwing out those with excessive imbalance, and then carry out permutation tests on the
remaining draws20. Such a method is not possible when ad hoc criteria are used to decide
whether to redraw.
Optimal model-based inference is less clear under re-randomization, since
allocation to treatment is data-dependent. To see this, consider the data generating
processes:
20When multiple draws are used to select the allocation which gives best balance over a sequence of 100 or
1000 draws, there may be a concern that the resulting assignment to treatment is mostly deterministic. This
will be the case in very small samples (under 12 units), but is not a concern for all but the smallest trials.
- 22 -
Yi = + Treati + i (2a)
Yi = + Treati + Zi + ui (2b)
Where Treati is a dummy variable for treatment status, and Zi are a set of
covariates potentially correlated with the outcome Yi. Under pure randomization, (2a) is
used for analysis, assignment to treatment is in expectation uncorrelated with i, and the
standard error will depend on Var(i). Suppose instead that re-randomization methods are
used, which force the difference in means of the covariates in Z to be less than some
specified threshold ZTREAT - ZCONTROL < . If is invariant to sample size (e.g. difference
in proportions less than 0.10), then this condition will occur almost surely as the sample
size goes to infinity, and thus the conditioning will not affect the asymptotics. However,
in practice is usually set by some statistical significance threshold. Then if (2a) is used
for analysis (that is, the covariates are not controlled for), we will only have that i, is
independent of Treati conditional on ZTREAT - ZCONTROL < . The correct standard error
should therefore account for this conditioning, using Var(i| ZTREAT - ZCONTROL < ).
In practice this will be difficult to do, so adapting the minimization inference
recommendations of Scott et al. (2002), we recommend researchers instead include all the
variables used to check balance as covariates in the regression. Estimation of the
treatment effect in (2b) will then be conditional on the variables used for checking
balance. Note this will require a loss of degrees of freedom compared to not controlling
for these covariates, but still requires fewer degrees of freedom than pair-wise matching.
The simulation results in Table 4 suggest that this approach works in practice. Treating
the big stick or minmax t-statistic methods as if they were pure random draws results in
less than ten percent of replications having p-values under 0.10 (Panel C), whereas
including the variables used for checking balance before re-randomizing as controls
results in the correct test size (Panel D). This correction is more important for the
minmax method than the big stick method, since the minmax method achieves greater
baseline balance.
- 23 -
4.5 How do the different methods compare in terms of power for detecting a given
treatment effect?
To compare the power of the different methods we simulate a treatment effect by
adding a constant to the follow-up outcome variable for the treatment group. We simulate
constant treatments which add 1000 Rupees (25 percent of average baseline profits) to
the Sri Lankan microenterprise profits; add 920 pesos (20 percent of average baseline
income) to the Mexican labor income; add 0.4 (0.5 standard deviations) to log
expenditure in Indonesia, and add 0.25 standard deviations to the Pakistan math test
scores and child height z-scores. For the schooling treatment, we randomly set one in
three schooling drop-outs to stay in school. These treatments are all relatively small in
magnitude for the sample sizes used, so that we can see differences in power across
methods, rather than have all methods give power close to one.
Table 6 then summarizes the power of a hypothesis test for detecting the
treatment effect, taking as the t-test on the treatment coefficient in a linear regression of
the outcome variable on a constant and a dummy variable for treatment status. We report
the proportion of replications where this test would reject the null hypothesis of no effect
at the 10 percent level. Panels A and C report results when the regression model does not
include controls for the method of randomization, while Panels B and D report the power
when stratum or pair dummies, or the variables used in checking balance for re-
randomization methods are included. The results for the pure random sample in panels B
and D include these same set of seven baseline controls, to enable comparison of ex-post
controls for baseline characteristics to ex-ante balancing.
Table 6 shows that if we do not adjust for the method of randomization, the
different methods often perform similarly in terms of power, and in cases where they
differ, it is because the methods which pursue balance have less power than pure
randomization. For example, with a sample size of 30, the power for both the height and
math test-scores is approximately 0.17 under a single random draw, but can be as low as
0.018 for the math test score under pair-wise matching, and as low as 0.052 for the height
z-score with the minmax method. Adding the stratum and pair dummies or baseline
variables used for re-randomizing increases power in almost all cases. Some of the
increases in power can be sizeable the power increases from 0.018 to 0.304 for the math
- 24 -
test score with pair-wise matching when the pair dummies are added. This increase in
power is another reason to take into account the method of randomization when
conducting analysis.
Table 6 also allows us to see the gain in power from ex-ante balancing compared
to ex-post balancing. The same set of variables used for forming the match and for the re-
randomization methods were added as ex-post controls when estimating the treatment
effect for the single random draw in panels B and D. When the variables are not very
persistent, such as the microenterprise profits and child schooling, the power is very
similar whether ex-ante or ex-post balancing is done. However, we do observe some
improvements in power from matching compared to ex-post controls for some, but not
all, of the more persistent outcome variables. The power increases from 0.584 to 0.761
for the Mexican labor income when ex-ante pair-wise matching on seven variables is
done rather than a pure random draw followed by linear controls for these seven variables
ex-post. However, there is no discernable change in power from balancing for child
height, another persistent outcome variable.
4.6 Can we go too far in pursuing balance?
When using stratification, matching or re-randomization methods, one question is
how many variables to balance on and whether balancing on too many variables could be
counter-productive.
The literature is not very definitive with respect to how many variables to use in
stratification. Some call for using many variables. For example, Box et al. (2005) write
"block what you can and randomize what you cannot", while Duflo et al. (2006) state that
"if several binary variables are available for stratification, it is a good idea to use all of
them, even if some of them may not end up having large explanatory power for the final
outcome." In contrast, Kernan et al. (1999) argue that "fewer strata are better", and raise
the possibility of unbalanced treatment assignment within strata due to small cell sizes,
recommending that an appropriate number of strata is between n/50 and n/100. Finally,
Therneau (1993) shows in simulations with sample sizes of 100, that in terms of
imbalance, with a sufficient number of factors used in stratifying (so that the number of
- 25 -
strata reaches n/2), performance can actually be worse than using unstratified
randomization.
We investigate how changing the number of strata affects balance and power in
practice in our samples of 100 and 300 observations by simulating stratification with two,
three and four stratifying variables, resulting in 8, 24, and 48 strata respectively. The
results are shown in Table 7. Both the size of extreme imbalances and the power do not
vary much with the number of strata for any of the six outcomes. In most cases there is
neither much gain, nor much loss, from including more strata. However, we do note that
for a sample size of 100, when strata dummies are included, power is always slightly
lower when 4 stratifying variables (and 48 strata) are included than when 3 stratifying
variables (and 24 strata) are used. For example, with the math test score, power falls from
0.464 to 0.399 when the number of strata is doubled.
A question related to the choice of how many variables to balance on is what
happens when one balances on irrelevant covariates. Greevy et al. (2004, p. 264) claim
that blocking or pairing on irrelevant covariates does not harm statistical efficiency or
power relative to not-matching. We argue (and show), however that there is a cost of
balancing on irrelevant variables. Since statistical analysis after balancing requires
controlling for the covariates used in balancing, the potential cost is the loss of degrees of
freedom from controlling for these variables. This can be offset by the reduction in
variation in the outcome variable that is explained by the variables balanced on. To see
this, compare the variance of the estimate of the treatment effect estimated using (2a)
versus (2b), where Z is generalized to be a k-dimensional set of covariates which
balancing was carried out on (including interactions between covariates used in forming
strata):
n
2
Var(withcontro )
ls n - 2 u^ i
i=1 (3)
Var( purerandom =
n
ization) n - k - 2^ 2
i
i=1
Equation (3) shows the trade-off involved in balancing. Balancing on relevant variables
n
means that the residual sum of squares u^ 2 will (in expectation) be less than the
i
i=1
- 26 -
n
residual sum of squares ^ 2, lowering the standard error. However, controlling for
i
i=1
covariates involves losing k degrees of freedom.
Consider then the worse case scenario, where none of the variables balanced on
have any predictive power for the outcome of interest. In this case the two residual sum
of squares in (3) will be equal, and the variance ratio reduces to (n-2)/(n-k-2). Standard
errors can then be much larger with these covariate controls when many covariates are
used with a small sample. For example, using 5 covariates to balance on with 10
observations leads to 63 percent greater standard errors if none of these covariates are
correlated with the outcome of interest. However, by a sample size of 100, even 10
irrelevant covariates could at most increase standard errors by 5.5 percent, equivalent to a
reduction in sample size from 100 to 90. With 200 or 400 as the sample size, it is indeed
the case that even in the most unlikely situation that all covariates are uncorrelated with
the outcome of interest, balancing on 5 or 10 covariates will not increase standard errors
by more than 3 percent.
However, balancing on irrelevant variables will continue to have repercussions
for standard errors if the number of variables balanced on increases at the same rate as the
sample size. This is true in matching, and in some cases when geographical variables are
used for forming strata. In pair-wise matching, the number of covariates used as controls
in the treatment regression is n/2. If the variables used to form matches do not have any
role in explaining the outcome of interest, we see that the ratio of standard errors will
approach 2 , that is, can be 41 percent higher under pair-wise matching than pure
randomization.
In our simulations, we address the issue of balancing on irrelevant variables by
stratifying and matching based on i.i.d. noise. The last two columns of Table 6 show the
power of the stratified and matching estimators when pure noise is used. Once we control
for stratum dummies, power is clearly less when irrelevant variables are used for
stratifying or matching than when relevant variables are used. For example, the power
with a sample size of 300 for household expenditure under pair-wise matching is 0.574
when relevant baseline variables are used to form the match compared to 0.356 when
- 27 -
i.i.d. noise is used in the matching. Thus the choice of variables used in stratifying or
matching does play an important role in determining power.
However, if we wish to compare the impact of matching or stratifying on
irrelevant variables to a pure random draw, we should compare the power for a single
random draw in panels A and C to the power for matching and stratifying on i.i.d. noise
in panels B and D which contain controls for stratum or pair dummies. The power is very
similar for all sample sizes. In practice, any given draw of i.i.d. noise is likely to have
some small correlation with the outcome of interest, reducing the residual sum of squares
when controlled for in a regression. It seems this small correlation is just enough to offset
the fall in degrees of freedom, so that the worst-case scenarios discussed above don't
come to pass.21. Hence in practice, it seems that stratifying on i.i.d. noise does not do any
worse than a simple random draw in terms of power when sample sizes are not very
small.
Finally, Table 6 shows that when stratification or matching is done purely on the
basis of i.i.d. noise, treating the randomization as if it was a pure random draw does not
lower power compared to the case when a single random draw is used. This is in contrast
to the case when matching or stratification is done on variables with strong predictive
power. Intuitively, when pure noise is used for stratification, it is as if a pure random
draw was taken. However, this does not mean that ex-post one can check whether the
variables used for matching or stratification have predictive power for the future
outcome, and if not, ignore the method of randomization. Ignoring the matching or
stratification is only correct if the baseline variables are truly pure noise if there is any
signal in these stratifying or matching variables then ignoring the randomization method
will result in incorrect size for hypothesis tests. Since in practice it will almost surely not
be the case that the correlation between the baseline variables and future outcome is
exactly zero (even if it is not statistically different from zero), one should control for the
method of randomization.22
21Note though that even our smallest sample size of 30 is larger than the cases Martin et al. (1993) study
where a loss of power can occur.
22Indeed, even in our simulations where we know the baseline variables are i.i.d. noise, any particular draw
from an i.i.d. distribution has some small correlation with the future outcome variables, and in Appendix 2
we see that controlling for the strata or pair dummies does tend to move the size of a hypothesis test of no
difference in means closer to the true level although the differences between controlling for strata or pair
- 28 -
4.7 What is the meaning of the standard Table 1 (if any)?
Section 2 points out that most research papers containing randomized experiments
feature a table (usually the first in the paper) that tests whether there are any statistically
significant differences in the baseline means of a number of variables across treatment
and control groups. The unanimous use of such tests is interesting in light of concern in
the clinical trials literature about both the statistical basis for such tests, and their
potential for abuse.23 Altman (1985, p. 26) writes that when "treatment allocation was
properly randomized, a difference of any sort between the two groups...will necessarily
be due to chance...performing a significance test to compare baseline variables is to
assess the probability of something having occurred by chance when we know that it did
occur by chance. Such a procedure is clearly absurd." Altman (1985, p. 26) goes on to
add that "statistical significance is immaterial when considering whether any imbalance
between the groups may have affected the results". In particular, it is wrong to infer from
the lack of statistical significance that the variable in question did not affect the outcome
of the trial, since a small imbalance in a variable highly correlated with the outcome of
interest can be far more important than a large and significant imbalance for a variable
uncorrelated with the variable of interest.
A particular concern with the use of significance tests is that researchers may
decide whether or not to control for a covariate in their treatment regression on the basis
of whether it is significant. Permutt (1990) shows that the resulting test's true
significance level is lower than the nominal level. Instead greater power is achieved by
always adjusting for a covariate that is highly correlated with the outcome of interest,
regardless of its distribution between groups. There seem to be some instances in the
literature where these balancing tests are being used to guide which variables to include
in robustness tests of the results. For example, Olken (2007a) notes three variables are
individually significant at the 10 percent level, but that the main results of the paper do
not change substantially if these variables are included.
dummies and not doing so are much smaller than when the baseline variables used for stratification did
predict future outcomes.
23See also Imai, King and Stuart (2008) for discussion on this issue in social science field experiments, and
for their suggestions as to what should constitute a proper check of balance.
- 29 -
A final concern with the use of significant tests for imbalance is their potential for
abuse. For example, Schulz and Grimes (2002) report that in the clinical trials literature,
researchers who use hypothesis tests to compare baseline characteristics report fewer
significant results than expected by chance. They suggest one plausible explanation is
that some investigators may not report some variables with significant differences,
believing that doing so would reduce the credibility of their reports. We have no evidence
to suggest this is occurring in the development literature, but one interpretation for the
repeated randomization methods discussed earlier in this paper is a desire by researchers
to show no significant differences between groups when such tests are used.
So how should we interpret such tables? The first question of interest in practice
is, given that such a test shows a statistically significant difference in baseline means,
does this make it more likely that there is also a statistically significant difference in
follow-up means in the absence of treatment? The answer is yes, provided that the
baseline data have predictive power for the follow-up outcomes. Figures 7 and 8 illustrate
this for the cases of Sri Lankan microenterprise profits, which have little predictive power
for future profits, and Mexican labor income which have more predictive power.
Appendix 424 shows similar figures for the other variables. For each dataset, these graphs
are based on the 10,000 simulations of a single random draw (without any balancing).
The x-axis shows the p-value for a test of difference in baseline means for the outcome of
interest. The y-axis shows the same p-value for a test of difference in follow-up means in
the absence of treatment. We divided the values on the x-axis into 100 bins and
calculated that 10th, 50th, and 90th percentile of the follow-up p-value within each bin.
For the Sri Lanka data, the values of the percentiles are flat across the whole x-
axis, suggesting that there is no relationship between p-value at baseline and p-value at
follow-up. For the other datasets, however, as illustrated by the Mexican labor income,
where the outcome variables show more persistence over time, the percentiles display an
upward-sloping pattern. That is, a statistically significant difference in baseline means
makes it more likely to also see a statistically significant difference in follow-up means.
The follow-up question of interest in practice is then: If we observe statistical
imbalance at baseline, but control for baseline variables in our analysis, are we any more
24Appendix 4 is available on the author's website.
- 30 -
likely to observe imbalance at follow-up than if we had obtained a random draw which
didn't show baseline imbalance? To examine this question, we take the 10,000
simulations of a single random draw and divide them into two sets. The first set includes
all draws that had a statistically significant difference at the 5 percent level in at least one
of our 7 baseline variables. We call this the "unbalanced" set. The second set is the
"balanced" set and includes all other draws. The top panels of Figure 9a and 9b show the
distribution of the differences in means between treatment and control for baseline labor
income and baseline math test scores are more tightly concentrated around zero in the
balanced set than the unbalanced set.25 The middle panels show that these differences are
less pronounced, but still persist at follow-up, again showing that imbalance in baseline
makes it more likely to have imbalance at follow-up. However, once we control for the 7
baseline variables, the distributions of a test of no treatment effect in the follow-up
outcome (when no treatment was given) is identical regardless of whether or not there
was baseline imbalance.
Intuitively, when randomization is used to allocate units into treatment and
control groups, if we do find unbalanced baseline characteristics, once we control for
them, the remaining unobservables are no more or less likely to be unbalanced than if we
did not find unbalanced baseline characteristics. However, as recommended by Altman
(1985), we should choose which baseline characteristics to control for not on the basis of
statistical differences, but on the strength of their relationship to the outcome of interest.
Overall, in most randomized settings, we therefore recommend reporting the point
estimates of the differences across groups, but not reporting the p-values on these
differences. If these differences are in variables thought to influence the outcome of
interest, one should control for them, regardless of whether or not the difference is
statistically significant. Note, however, that there are two exceptions where carrying out a
test of statistical significance is meaningful. First, statistically testing the difference
between treatment and control groups at baseline can be relevant if there was possible
interference in the randomization. This may be relevant when random assignment is
carried out in the field by survey enumerators, but should not be a concern when the
25Appendix A3 presents the same figures for other outcome variables and sample sizes. They all show the
same patterns as in Figure 9.
- 31 -
researcher does the randomization by computer. Second, a related common use for these
significant tests is seen in Ashraf et al. (2006a), who are only able to survey 1777 of the
4000 microfinance clients allocated to treatment and control. They test whether there are
differences between the treatment and control groups amongst those surveyed.
5. Conclusions
Our surveys of the recent literature and of the most experienced researchers
implementing randomized experiments in developing countries finds that most
researchers are not relying on pure randomization, but are doing something to pursue
balance on observables. In addition to stratification, we find pair-wise matching and re-
randomization methods to be used much more than is apparent from the existing
literature. The paper draws out implications from the existing statistical, clinical, and
social science literature on the pros and cons of these various methods of seeking
balance, and compares the performance of the different methods in simulations.
Our simulation results show the method of randomization matters more in small
sample sizes, such as 30 or 100 observations, and matters more for relatively persistent
outcome variables such as health and test scores than for less persistent outcome
variables such as microenterprise profits or household expenditure. Overall we find pair-
wise matching to perform best in achieving balance in small samples, provided that the
variables used in forming pairs have good predictive power for the future outcomes.
Stratification and re-randomization using a minmax method also lead to some
improvements over a pure random draw, but in the majority of our simulations are
dominated by pair-wise matching. With sample sizes of 300 we find that the method of
randomization matters much less, although matching still leads to some improvement in
balance for the persistent outcomes.
Our analysis of how randomization is being carried out in practice suggests
several areas where the practice of randomization can be improved or better reported.
This leads us to draw out the following recommendations:
1) Better reporting of the method of random assignment is needed. Researchers need
to describe clearly their choice of method, the reason for this choice, and whether
or not the randomization was carried out in public or private. This is particularly
- 32 -
important for experiments done on small samples, where the choice of
randomization method makes more difference.
2) "As ye randomize, so shall ye analyze" (Senn, 2004): Researchers should
account for the method of randomization when performing statistical analysis.
Since the majority of inference in economics is model-based, rather than
randomization inference, this means adding controls for all covariates (and
interactions between covariates) used in seeking balance. In particular, strata
dummies should be included when analyzing the results of stratified
randomization. Our simulations show that while on average failure to account for
the method of randomization generally results in overly conservative standard
errors, there are also a substantial number of draws in which standard errors
which do not account for the method of randomization overstate the significance
of the results. Moreover, failure to control for the method of randomization results
in incorrect test size.
3) Re-think the common use of re-randomization. Our simulations find pair-wise
matching to generally perform better than re-randomization in terms of balance
and power, and like re-randomization, matching allows balance to be sought on
more variables than possible under stratification. Adjusting for the method of
randomization is statistically cleaner with matching or stratification than with re-
randomization.
4) Be cautious in seeking balance on too many variables, since generally our models
and data have poor predictive power for changes. The baseline of the outcome
variable and variables desired for subgroup analysis are obvious candidates for
balancing on. However, seeking to balance on many other covariates involves a
downside in terms of loss in degrees of freedom when estimating standard errors,
possibly more cases of missing observations, a potentially weaker match in
matching methods in terms of the main covariates of interest, and odd-numbers
within strata when stratification is used. Thus, contrary to some claims, it is
possible to over-stratify or seek balance on too many variables.
5) Statistical tests of imbalance between treatment and control groups should only
be performed when there is reason to suspect interference, or when only a sample
- 33 -
of those in the experiment are surveyed, and such tests should not be used to
decide which variables to control for in treatment regressions. The common
practice of testing for significant differences between the two groups is otherwise
testing whether something that we know was due to chance was due to chance.
Researchers should control for variables believed to strongly influence the
outcome of interest, regardless of whether the difference between treatment and
control groups is significant or not.
6) Acknowledge that the different goals of randomization can conflict with one
another in small samples. The idea of randomization as a valid basis for inference
(through permutation analysis), the desire for comparable groups, and the fairness
and transparency involved in one-off public assignment present trade-offs for the
researchers in terms of choice of randomization method.
References
Aickin, Mikel (2001) "Randomization, balance, and the validity and efficiency of design-
adaptive allocation methods", Journal of Statistical Planning and Inference 94: 97-
119.
Altman, Douglas A. (1985) "Comparability of randomized groups", The Statistician 34:
125-36
Andrabi, Tahir, Jishnu Das, Asim Khwaja and Tristan Zajonc (2008) "Do Value-added
Estimates Add Value? Accounting for Learning Dynamics", www.leapsproject.org
Ashraf, Nava, James Berry and Jesse M. Shapiro (2007) "Can Higher Prices Stimulate
Product Use? Evidence from a Field Experiment in Zambia", Poverty Action Lab
Paper No. 41.
Ashraf, Nava, Dean Karlan and Wesley Yin (2006a) "Tying Odysseus to the Mast:
Evidence from a Commitment Savings Product in the Philippines", Quarterly Journal
of Economics 635-672
Ashraf, Nava, Dean Karlan and Wesley Yin (2006b) "Deposit Collectors", Advances in
Economic Analysis and Policy 6(2), article 5
Banerjee, Abhijit (2007) Making Aid Work. The MIT Press, Cambridge, MA.
Banerjee, Abhijit, Shawn Cole, Esther Duflo, and Leigh Linden (2007) "Remedying
Education: Evidence from Two Randomized Experiments in India", Quarterly Journal
of Economics 1235-1264.
Bertrand, Marianne, Simeon Djankov, Rema Hanna and Sendhil Mullainathan (2007)
"Obtaining a Driver's License in India: An Experimental Approach to Studying
Corruption", Quarterly Journal of Economics 1639-76.
Björkman, Martina and Jakob Svensson (2007) "Power to the People: Evidence from a
Randomized Field Experiment of a Community-Based Monitoring Project in
Uganda", World Bank Policy Research Working Paper 4268
- 34 -
Bobonis, Gustavo, Edward Miguel and Charu Puri Sharma (2006) "Iron Deficiency
Anemia and School Participation", Journal of Human Resources, 41(4), 692-721
Box, George E.P., J. Stuart Hunter and William G. Hunter (2005) Statistics for
Experimenters: Design, Innovation and Discovery, Second Edition. Wiley-
Interscience, New Jersey.
Burtless, Gary (1995) "The Case for Randomized Field Trials in Economic and Policy
Research", The Journal of Economic Perspectives, 9(2): 63-84.
Committee for Proprietary Medicinal Products (2003) "Points to consider on adjustment
for baseline covariates" CPMP/EWP/2863/99
(http://www.emea.europa.eu/pdfs/human/ewp/286399en.pdf [accessed February 6,
2008]
De Mel, Suresh, David McKenzie and Christopher Woodruff (2008) "Returns to capital:
Results from a randomized experiment", Quarterly Journal of Economics,
forthcoming.
Duflo, Esther (2005) "Evaluating the Impact of Development Aid Programmes: The Role
of Randomised Evaluations", pp. 205-247 in Development Aid: Why and How?
Towards Strategies for Effectiveness. Proceedings of the AFD-EUDN Conference,
2004, Agence Française de Développement.
Duflo, Esther, Rachel Glennerster, and Michael Kremer (2007) "Using Randomization in
Development Economics: A Toolkit", CEPR Working Paper No. 6059.
Duflo, Esther, Rema Hanna and Stephen Ryan (2007) "Monitoring Works: Getting
Teachers to Come to School", Mimeo. MIT.
Duflo, Esther and Michael Kremer (2004) "Use of Randomization in the Evaluation of
Development Effectiveness" pp. 205-232 in Osvaldo Feinstein, Gregory K. Ingram
and George K. Pitman, editors Evaluating Development Effectiveness (World Bank
Series on Evaluation and Development, Vol 7) Transaction Publishers: New
Brunswick, NJ
Dupas, Pascaline (2006) "Relative Risks and the Market for Sex: Teenagers, Sugar
Daddies, and HIV in Kenya", Mimeo. Dartmouth College
Field, Erica and Rohini Pande (2008) "Repayment Frequency and Default in Micro-
Finance: Evidence from India", Journal of European Economic Association Papers
and Proceedings, forthcoming.
Glewwe, Paul, Albert Park and Meng Zhao (2006) "The Impact of Eyeglasses on the
Academic Performance of Primary School Students: Evidence from a Randomized
Trial in Rural China", Mimeo. University of Minnesota.
Glewwe, Paul, Michael Kremer, Sylvie Moulin and Eric Zitzewitz (2004) "Retrospective
vs. prospective analyses of school inputs: the case of flip charts in Kenya", Journal of
Development Economics 74: 251 268
Greevy, Robert, Bo Lu, Jeffrey H. Silver, and Paul Rosenbaum (2004). "Optimal
multivariate matching before randomization", Biostatistics, 5: 263275.
He, Fang, Leigh Linden and Margaret MacLeod (2007) "Teaching What Teachers Don't
Know: An Assessment of the Pratham English Language Program", Mimeo. Columbia
University.
Imai, Kosuke, Gary King and Clayton Nall (2007) "The Essential Role of Pair Matching
in Cluster-Randomized Experiments, with Application to the Mexican Universal
Health Insurance Evaluation", Mimeo. Harvard.
- 35 -
Imai, Kosuke, Gary King and Elizabeth Stuart (2008) "Misunderstandings among
Experimentalists and Observationalists about Causal Inference" Journal of the Royal
Statistical Society, Series A Vol. 171: 481-502
Karlan, Dean and Martin Valdivia (2006) "Teaching Entrepreneurship: Impact of
Business Training on Microfinance Clients and Institutions",Mimeo. Yale University.
Kernan, Walter N., Catherine M. Viscoli, Robert W. Makuch, Lawrence M. Brass, and
Ralph I. Horwitz (1999) "Stratified Randomization for Clinical Trials", Journal of
Clinical Epidemiology 52(1): 19-26.
King, G., Gakidou, E., Ravishankar, N., Moore, R. T., Lakin, J., Vargas, M., T´ellez-
Rojo,M. M., ´ Avila, J. E. H., ´ Avila, M. H., and Llamas, H. H. (2007). A `politically
robust' experimental design for public policy evaluation, with application to the
mexican universal health insurance program. Journal of Policy Analysis and
Management 26, 3, 479506.
Kremer, Michael (2003) "Randomized Evaluations of Educational Programs in
Developing Countries: Some Lessons" The American Economic Review Papers and
Proceedings, 93(2): 102-106
Kremer, Michael, Jessica Leino, Edward Miguel and Alix Peterson Zwane (2006)
"Spring Cleaning: A Randomized Evaluation of Source Water Quality Improvement",
Mimeo Harvard University
Levitt, Steven and John List (2008) "Field Experiments in Economics: The Past, The
Present, and the Future", NBER Working Paper No. 14356.
Martin, Donald C.,Paula Diehr, Edward B. Perrin, and Thomas D. Koepsill (1993) "The
effect of matching on the power of randomized community intervention studies",
Statistics in Medicine 12: 329-338.
Miguel, Edward and Michael Kremer (2004) "Worms: Identifying Impacts on Education
and Health in the Presence of Treatment Externalities", Econometrica 72(1): 159-217.
Olken, Benjamin (2007a) "Monitoring Corruption: Evidence from a Field Experiment in
Indonesia", Journal of Political Economy 115(2): 200-49.
Olken, Benjamin (2007b) "Political Institutions and Local Public Goods: Evidence from
a Field Experiment", Mimeo. Harvard University.
Permutt, Thomas (1990) "Testing for Imbalance of Covariates in Controlled
Experiments", Statistics in Medicine 9: 1455-62.
Pocock, SJ and R Simon (1975) "Sequential Treatment Assignment with Balancing for
Prognostic Factors in the Controlled Clinical Trial", Biometrics 31:103-15.
Raab, Gilliam M., Simon Day, and Jill Sales (2000) "How to Select Covariates to Include
in the Analysis of a Clinical Trial", Controlled Clinical Trials 21:330-42.
Rosenbaum, Paul R. (2002) Observational Studies: Second Edition. Springer Series in
Statistics, Springer: New York
Schulz, Kenneth (1996) "Randomised Trials, Human Nature, and Reporting Guidelines",
The Lancet 348: 596-598
Schulz, Kenneth and David Grimes (2002) "Allocation concealment in randomized trials:
defending against deciphering", The Lancet 359: 614-618
Scott, Neil W., Gladys McPherson, Craig R. Ramsay and Marion Campbell (2002) "The
method of minimization for allocation to clinical trials: a review", Controlled Clinical
Trials 23: 662-74.
- 36 -
Senn, Stephen (2004) "Added values: Controversies concerning randomization and
additivity in clinical trials", Statistics in Medicine 23(24): 3729-53.
Skoufias, Emmanuel (2005) "PROGRESA and its Impacts on the Welfare of Rural
Households in Mexico", IFPRI Research Report 139, IFPRI, Washington D.C.
Soares, J.F. and C.F. Wu (1983) "Some restricted randomization rules in sequential
designs", Communications in Statistics Theoretical Methods A 12: 2017-34.
Taves, DR (1974) "Minimization: A new method of Assigning Patients to Treatment and
Control Groups", Clinical Pharmacology and Therapeutics 15: 443-53.
Therneau, Terry M. (1993) "How many stratification factors is "too many" to use in a
randomization plan?", Controlled Clinical Trials 14(2): 98-108.
Treasure, Tom and Kenneth MacRae (1998) "Minimisation: the platinum standard for
trials?", British Medical Journal 317(7155): 362-63.
- 37 -
Table 1: Summary of Selected Randomized Experiments in Developing Countries
Table for # variables Test
Randomization Sample Number Public or Stratification Matched Number of Strata or pair assessing used to of significance
Paper Unit Size Treated Private Used? pairs? Strata dummies used? balance? check balance for balance?
Published/forthcoming Papers
Ashraf et al. (2006a) Microfinance clients 1777 710 n.a. No No Yes 12 Yes
Ashraf et al. (2006b) Barangay (area) 10 5 n.a. No Yes Yes Yes 12 Yes
Banerjee et al. (2007) School 98 49/49 n.a. Yes No n.a. No Yes 4 Yes
School 111 55/56 n.a. Yes No n.a. No Yes 4 Yes
School 67 32/35 n.a. Yes No n.a. No Yes 4 Yes
Bertrand et al. (2007) Men wanting a 822 268/264 Public Yes (C). No 23 Yes Yes 22 Yes
Driver's license
Bobonis et al. (2006) Preschool cluster 155 59/51/45 n.a. No No Yes 24 Yes
Field and Pande (2008) Microfinance group 100 38/30 Public No No No (A)
Glewwe et al. (2004) School 178 89 n.a. Yes No n.a. No Yes 8 Yes
Miguel and Kremer (2004) School 75 25*3 n.a. Yes No n.a. No Yes 21 Yes
Olken (2007a) Village 608 202/199 n.a. Yes No 156 Yes Yes 10 Yes
Subdistrict 156 n.a. n.a. Yes No 50 Yes Yes 10 Yes
Working Papers
Ashraf et al. (2007) Household 1260 6 groups n.a. Yes No 5 Yes Yes 14 Yes
Björkman and Svensson (2007) Community 50 25 n.a. Yes No n.a. Yes 39 Yes
Duflo et al. (2007) School 113 57 n.a. Yes No n.a. No (E) Yes 15 Yes
Dupas (2006) School 328 71 n.a. Yes No n.a. No Yes 17 Yes
Glewwe et al. (2006) Township 25 12 n.a. No Yes Yes 4 Yes
He et al. (2007) School division 194 97 n.a. Yes No n.a. No Yes 22 Yes
Karlan and Valdivia (2006) Microfinance group 239 104/84 n.a. Yes No n.a. No (D) Yes 14 Yes
Kremer et al. (2006) Spring 200 50/50/100 n.a. Yes No n.a. No Yes 28 Yes
Olken (2007b) Village 48 17 n.a. Yes No 2 Yes Yes 8 Yes
Notes:
n.a. denotes information not available in the paper.
A: Paper says check was done on a number of variables and is available upon request.
C: It appears randomization was done within recruitment session, but the paper was not clear on this.
D: Dummies for location are included, but not for credit officer which was the other stratifying variable.
E. Dummies for district are included, but not for the number of households in the area which were also used for stratifying within district.
Table 2: Survey Evidence on Randomization Methods Used by Leading Researchers
% WHO HAVE EVER USED % Using Method in
5+ experiment Most Recent
Unweighted Weighted Group Experiment
Single Random Assignment to Treatment (possibly with stratification) 80 84 92 39.1
Subjectively deciding whether to redraw 32 52 46 4.3
Using a statistical rule to decide whether to redraw 12 15 15 0.0
Carrying out many random assignments, and choosing best balance 24 45 38 17.4
Explicitly matching pairs of observations on baseline characteristics 56 52 54 39.1
Number of Researchers 25 25 13 23
Notes:
Methods described in more detail in the paper.
Weighted results weight by the number of experiments the researcher has participated in
5+ experiment group refers to researchers who have carried out 5 or more randomized experiments
Table 3: How do the different methods compare in terms of Baseline Balance?
Simulation Results for 100 Observation Sample Size
Single Stratified Stratified Pairwise Big Draw with
Random on 2 on 4 Greedy Stick minmax t-stat
Draw variables variables Matching Rule out of 1000 draws
Panel A: Average difference in BASELINE between treatment and control means (in std. dev.)
Microenterprise profits (Sri Lanka) 0.001 0.000 -0.001 0.000 0.001 0.000
Household expenditure (Indonesia) -0.002 0.001 -0.001 0.000 -0.001 -0.002
Labor income (Mexico) 0.000 0.000 0.000 0.000 -0.001 0.000
Height z-score (Pakistan) 0.001 0.001 0.000 0.000 -0.001 0.000
Math test score (Pakistan) 0.003 0.000 -0.001 0.000 0.002 0.000
Baseline unobservables (Sri Lanka) 0.000 0.000 0.000 0.000 0.000 0.001
Baseline unobservables (Mexico) 0.000 0.000 0.000 0.000 0.000 0.000
Panel B: 95th percentile of difference in BASELINE between treatment and control means (in std. dev.)
Microenterprise profits (Sri Lanka) 0.386 0.195 0.241 0.312 0.324 0.091
Household expenditure (Indonesia) 0.390 0.145 0.191 0.266 0.328 0.107
Labor income (Mexico) 0.384 0.280 0.304 0.099 0.332 0.088
Height z-score (Pakistan) 0.395 0.160 0.206 0.119 0.319 0.089
Math test score (Pakistan) 0.392 0.164 0.237 0.074 0.328 0.106
Baseline unobservables (Sri Lanka) 0.434 0.417 0.414 0.434 0.434 0.434
Baseline unobservables (Mexico) 0.457 0.448 0.439 0.457 0.457 0.457
Panel C: Proportion of p-values <0.1 for testing difference in BASELINE means
Microenterprise profits (Sri Lanka) 0.097 0.000 0.005 0.036 0.045 0.000
Household expenditure (Indonesia) 0.102 0.000 0.000 0.011 0.049 0.000
Labor income (Mexico) 0.100 0.015 0.029 0.000 0.053 0.000
Height z-score (Pakistan) 0.100 0.000 0.001 0.000 0.038 0.000
Math test score (Pakistan) 0.100 0.000 0.006 0.000 0.048 0.000
Baseline unobservables (Sri Lanka) 0.101 0.096 0.095 0.084 0.098 0.091
Baseline unobservables (Mexico) 0.108 0.095 0.093 0.103 0.102 0.110
Notes:
Statistics are based on 10,000 simulations of each method. Details on methods and variables are in Table A2.
Table 4: How do the different methods compare in terms of Balance on Future Outcomes?
Sample Size of 30 Sample Size of 300
Single Stratified Pairwise Big Draw with Single Stratified Stratified Pairwise Big Draw with
Random on 2 Greedy Stick minmax Random on 2 on 4 Greedy Stick minmax
Draw variables Matching Rule t-stat Draw variables variables Matching Rule t-stat
Panel A: Average difference in FOLLOW-UP between treatment and control means (in std. dev.)
Microenterprise profits (Sri Lanka) 0.001 0.000 0.002 -0.003 0.002 0.000 0.001 0.001 0.000 0.000 0.000
Child schooling (Indonesia) -0.005 -0.010 -0.005 0.004 -0.006 0.002 0.003 -0.001 0.000 -0.002 -0.002
Household expenditure (Indonesia) 0.000 0.002 -0.001 0.000 -0.006 -0.001 -0.001 0.000 -0.001 -0.001 -0.001
Labor income (Mexico) -0.003 0.000 0.003 0.003 -0.002 0.001 0.000 0.001 -0.001 0.001 -0.002
Height z-score (Pakistan) 0.007 0.001 0.001 -0.003 0.001 -0.001 0.000 0.000 0.000 0.002 0.000
Math test score (Pakistan) 0.001 0.002 -0.001 -0.003 0.005 -0.001 0.000 0.000 -0.001 -0.001 0.001
Panel B: 95th percentile of difference in FOLLOW-UP between treatment and control means (in std. dev.)
Microenterprise profits (Sri Lanka) 0.713 0.627 0.592 0.705 0.708 0.220 0.210 0.209 0.211 0.216 0.224
Child schooling (Indonesia) 0.834 0.745 0.556 0.556 0.556 0.213 0.219 0.212 0.227 0.227 0.196
Household expenditure (Indonesia) 0.721 0.643 0.503 0.677 0.590 0.226 0.194 0.196 0.200 0.219 0.198
Labor income (Mexico) 0.755 0.546 0.642 0.705 0.529 0.227 0.196 0.198 0.149 0.213 0.195
Height z-score (Pakistan) 0.710 0.620 0.568 0.620 0.443 0.222 0.186 0.189 0.189 0.212 0.225
Math test score (Pakistan) 0.717 0.448 0.361 0.648 0.525 0.227 0.180 0.184 0.167 0.209 0.175
Panel C: Proportion of p-values <0.1 for testing difference in FOLLOW-UP means with inference as if
pure randomization was used (e.g. no adjustment for strata or match dummies)
Microenterprise profits (Sri Lanka) 0.105 0.059 0.045 0.101 0.109 0.100 0.080 0.080 0.085 0.092 0.103
Child schooling (Indonesia) 0.052 0.113 0.033 0.041 0.010 0.121 0.087 0.082 0.098 0.111 0.096
Household expenditure (Indonesia) 0.102 0.069 0.011 0.083 0.046 0.101 0.056 0.052 0.064 0.092 0.059
Labor income (Mexico) 0.106 0.012 0.049 0.029 0.009 0.100 0.056 0.062 0.011 0.087 0.028
Height z-score (Pakistan) 0.097 0.056 0.031 0.059 0.007 0.097 0.044 0.049 0.049 0.081 0.097
Math test score (Pakistan) 0.101 0.006 0.000 0.072 0.022 0.101 0.038 0.042 0.028 0.076 0.032
Panel D: Proportion of p-values <0.1 for testing difference in FOLLOW-UP means with inference which
takes account of randomization method (i.e. controls for stratum, pair, or re-randomizing variables)
Microenterprise profits (Sri Lanka) 0.103 0.091 0.104 0.103 0.122 0.098 0.103 0.133 0.103 0.102 0.101
Child schooling (Indonesia) 0.103 0.117 0.033 0.098 0.108 0.098 0.102 0.104 0.098 0.104 0.104
Household expenditure (Indonesia) 0.102 0.098 0.102 0.101 0.094 0.099 0.100 0.099 0.101 0.105 0.100
Labor income (Mexico) 0.083 0.107 0.101 0.079 0.067 0.100 0.095 0.101 0.104 0.100 0.112
Height z-score (Pakistan) 0.100 0.097 0.104 0.100 0.103 0.094 0.097 0.097 0.098 0.095 0.102
Math test score (Pakistan) 0.099 0.102 0.106 0.098 0.098 0.101 0.097 0.099 0.097 0.100 0.102
Notes:
Panels A and B coefficients are for specifications without controls for stratum or pair dummies.
Statistics are based on 10,000 simulations of each method. Details on methods and variables are in Table A2.
Table 5: Is it always conservative to ignore the method of randomization?
Proportion of replications where controlling for stratum or pair dummies lowers the
p-value on a test of difference in means between treatment and control groups
Stratified Stratified Pairwise Big Draw with
on 2 on 4 Greedy Stick minmax
variables variables Matching Rule t-stat
Panel A: Sample Size 30
Microenterprise profits (Sri Lanka) 0.690 . 1.000 0.493 0.555
Child schooling (Indonesia) 0.373 . 0.699 0.567 0.854
Household expenditure (Indonesia) 0.622 . 1.000 0.523 0.657
Labor income (Mexico) 0.820 . 1.000 0.546 0.773
Height z-score (Pakistan) 0.579 . 1.000 0.537 0.825
Math test score (Pakistan) 0.684 . 1.000 0.522 0.740
Panel B: Sample Size 300
Microenterprise profits (Sri Lanka) 0.668 0.731 1.000 0.526 0.689
Child schooling (Indonesia) 0.705 0.634 1.000 0.506 0.674
Household expenditure (Indonesia) 0.869 0.733 1.000 0.522 0.738
Labor income (Mexico) 0.874 0.712 1.000 0.525 0.725
Height z-score (Pakistan) 0.860 0.655 1.000 0.522 0.754
Math test score (Pakistan) 0.882 0.735 1.000 0.533 0.776
Notes:
Statistics are based on 10,000 simulations of each method. Details on methods and variables are
in Table A2.
Table 6: How do the different methods compare in terms of Power in detecting a given treatment effect?
Sample Size of 30
Single Stratified Pairwise Big Draw with Stratified Matching
Random on 2 Greedy Stick minmax on on
Draw variables Matching Rule t-stat i.i.d noise i.i.d. noise
Panel A: Proportion of p-values<0.10 when no adjustment is made for method of randomization
Microenterprise profits (Sri Lanka) 0.144 0.106 0.100 0.139 0.154 0.119 0.086
Child schooling (Indonesia) 0.123 0.146 0.106 0.115 0.066 0.133 0.144
Household expenditure (Indonesia) 0.390 0.382 0.340 0.382 0.360 0.396 0.387
Labor income (Mexico) 0.172 0.097 0.154 0.157 0.097 0.150 0.218
Height z-score (Pakistan) 0.174 0.134 0.127 0.134 0.052 0.213 0.194
Math test score (Pakistan) 0.167 0.051 0.018 0.139 0.087 0.176 0.131
Panel B: Proportion of p-values<0.10 when adjustment is made for randomization method
(and for the single random draw controls for the seven baseline variables are added to the regression)
Microenterprise profits (Sri Lanka) 0.130 0.135 0.158 0.131 0.167 0.158 0.164
Child schooling (Indonesia) 0.109 0.131 0.115 0.112 0.095 0.111 0.144
Household expenditure (Indonesia) 0.409 0.424 0.574 0.419 0.461 0.382 0.356
Labor income (Mexico) 0.204 0.226 0.190 0.220 0.243 0.175 0.151
Height z-score (Pakistan) 0.246 0.201 0.200 0.251 0.281 0.162 0.157
Math test score (Pakistan) 0.183 0.313 0.304 0.187 0.217 0.158 0.170
Sample Size of 300
Single Stratified Stratified Pairwise Big Draw with Stratified Matching
Random on 2 on 4 Greedy Stick minmax on on
Draw variables variables Matching Rule t-stat i.i.d noise i.i.d. noise
Panel C: Proportion of p-values<0.10 when no adjustment is made for method of randomization
Microenterprise profits (Sri Lanka) 0.288 0.274 0.278 0.267 0.280 0.280 0.289 0.279
Child schooling (Indonesia) 0.606 0.585 0.562 0.607 0.597 0.600 0.563 0.610
Household expenditure (Indonesia) 0.999 0.999 1.000 1.000 0.999 1.000 0.998 0.999
Labor income (Mexico) 0.494 0.486 0.480 0.475 0.489 0.474 0.490 0.484
Height z-score (Pakistan) 0.728 0.757 0.756 0.766 0.743 0.767 0.757 0.728
Math test score (Pakistan) 0.615 0.654 0.650 0.655 0.619 0.657 0.631 0.624
Panel D: Proportion of p-values<0.10 when adjustment is made for randomization method
(and for the single random draw controls for the seven baseline variables are added to the regression)
Microenterprise profits (Sri Lanka) 0.301 0.305 0.343 0.290 0.302 0.309 0.283 0.338
Child schooling (Indonesia) 0.608 0.596 0.589 0.602 0.619 0.595 0.559 0.607
Household expenditure (Indonesia) 1.000 1.000 1.000 1.000 1.000 1.000 0.998 0.998
Labor income (Mexico) 0.584 0.561 0.541 0.761 0.584 0.582 0.493 0.602
Height z-score (Pakistan) 0.863 0.849 0.854 0.853 0.867 0.866 0.741 0.721
Math test score (Pakistan) 0.812 0.792 0.781 0.829 0.816 0.826 0.630 0.603
Notes:
Statistics are based on 10,000 simulations of each method. Details on methods and variables are in Table A2.
Simulated treatment effects are as follows
Microenterprise profits: A 1,000 Sri Lankan Rupee increase in profits (about 25% of average baseline profits)
Child schooling: One in three randomly selected children in the treatment group who would have dropped out don't
Household expenditure: An increase of 0.4 in ln household expenditure per capita, which corresponds to about one half a
standard deviation or moving a household from the 25th to the 50th percentile.
Labor income: A 920 Peso increase in income (about 20% of average baseline income)
Height z-score: An increase of one quarter of a standard deviation in the z-score, where the z-score is defined as standard
deviations from mean US height for age
Math test score: An increase of one quarter of a standard deviation in the test score
Table 7: How does stratification vary with the number of Stratum?
Simulation results
Sample Size 100 Sample Size 300
Stratified Stratified Stratified Stratified Stratified Stratified
on 2 on 3 on 4 on 2 on 3 on 4
variables variables variables variables variables variables
(8 strata) (24 strata) (48 strata) (8 strata) (24 strata) (48 strata)
Panel A: Imbalance - 95th percentile of difference in follow-up means
Microenterprise profits (Sri Lanka) 0.322 0.338 0.338 0.210 0.213 0.209
Child schooling (Indonesia) 0.399 0.346 0.369 0.219 0.211 0.212
Household expenditure (Indonesia) 0.337 0.335 0.343 0.194 0.193 0.191
Labor income (Mexico) 0.335 0.327 0.344 0.196 0.198 0.198
Height z-score (Pakistan) 0.297 0.299 0.310 0.186 0.191 0.189
Math test score (Pakistan) 0.285 0.298 0.316 0.180 0.181 0.184
Panel B: Power: Proportion of p-values<0.10 when no strata dummies included
Microenterprise profits (Sri Lanka) 0.129 0.138 0.144 0.274 0.281 0.278
Child schooling (Indonesia) 0.303 0.267 0.273 0.585 0.574 0.562
Household expenditure (Indonesia) 0.852 0.850 0.845 0.999 1.000 1.000
Labor income (Mexico) 0.170 0.161 0.180 0.486 0.480 0.480
Height z-score (Pakistan) 0.286 0.295 0.297 0.757 0.757 0.756
Math test score (Pakistan) 0.236 0.245 0.254 0.654 0.649 0.650
Panel C: Power: Proportion of p-values<0.10 when strata dummies included
Microenterprise profits (Sri Lanka) 0.186 0.273 0.242 0.305 0.327 0.343
Child schooling (Indonesia) 0.278 0.301 0.255 0.596 0.596 0.589
Household expenditure (Indonesia) 0.904 0.914 0.876 1.000 1.000 1.000
Labor income (Mexico) 0.204 0.212 0.199 0.561 0.541 0.541
Height z-score (Pakistan) 0.487 0.463 0.457 0.849 0.843 0.854
Math test score (Pakistan) 0.464 0.464 0.399 0.792 0.790 0.781
Notes:
Statistics are based on 10,000 simulations of each method. Details on methods and variables are in Table A2.
Figures 1-6: Distribution of Differences in Means between the Treatment and Control Groups
and Baseline and Follow-up
Figure 1a: Sri Lanka Microenterprise profits sample size 30
3
2
eniles
Ba
1
0
5
1.
re
atl 1
hs
ontm6 .5
0
-1 -.5 0 .5 1
Difference in average profits (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 1b: Sri Lanka Microenterprise profits sample size 100
6
4
eniles
Ba
2
0
2
re 5
1.
atl
hs 1
ontm6 .5
0
-1 -.5 0 .5 1
Difference in average profits (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 1c: Sri Lanka Microenterprise profits sample size 300
10
8
eniles 6
Ba 4
2
0
5
2.
2
ertalshtno 5
1.
1
m6
.5
0
-.4 -.2 0 .2 .4
Difference in average profits (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 2a: ENE Labor Income Data sample size 30
4
3
eniles 2
Ba
1
0
5
1.
ertalshtno 1
m6 .5
0
-1 0 1
Difference in average income (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 2b: ENE Labor Income Data Sample Size 100
6
4
eniles
Ba
2
0
3
ertalshtno 2
m6 1
0
-.5 0 .5
Difference in average income (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 2c: ENE Labor Income Data Sample size 300
10
8
eniles 6
Ba 4
2
0
5
4
ertalshtno 3
2
m6
1
0
-.4 -.2 0 .2 .4
Difference in average income (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 3a: IFLS School Data Sample Size 30
52.
2
retal 51.
earsy3 1
.5
0
-1 -.5 0 .5 1
Difference in average in-school dummy (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 3b: IFLS School Data Sample Size 100
4
3
reatls 2
eary3
1
0
-1 -.5 0 .5 1
Difference in average in-school dummy (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 3c: IFLS School Data Sample Size 300
8
6
reatls 4
eary3
2
0
-1 -.5 0 .5 1
Difference in average in-school dummy (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 4a: IFLS Expenditure Data Sample Size 30
5
2.
2
eniles 5
1.
Ba 1
.5
0
51.
retalsraey3 1
.5
0
-2 -1 0 1 2
Difference in average ln hh expenditure p cap (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 4b: IFLS Expenditure Data Sample Size 100
5
4
eniles 3
Ba 2
1
0
52.
2
retalsraey3 5
1.
1
.5
0
-1 -.5 0 .5 1
Difference in average ln hh expenditure p cap (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 4c: IFLS Expenditure Data Sample Size 300
10
8
eniles 6
Ba 4
2
0
4
retalsraey3 3
2
1
0
-.5 0 .5
Difference in average ln hh expenditure p cap (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 5a: LEAPS Math Test Score Data Sample Size 30
3
2
eniles
Ba
1
0
2
retalr 51.
1
eay1
.5
0
-1 -.5 0 .5 1
Difference in average math test score (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 5b: LEAPS Math Test Score Data Sample Size 100
8
6
eniles 4
Ba
2
0
3
ertalraey1 2
1
0
-1 -.5 0 .5 1
Difference in average math test score (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 5c: LEAPS Math Test Score Data Sample Size 300
10
8
eniles 6
Ba 4
2
0
5
4
ertalraey1 3
2
1
0
-1 -.5 0 .5 1
Difference in average math test score (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 6a: LEAPS Height Z-Score Data Sample Size 30
4
3
eniles 2
Ba
1
0
2
5
retalraey1 1.
1
.5
0
-1 -.5 0 .5 1
Difference in average height z-score (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 6b: LEAPS Height Z-Score Data Sample Size 100
6
4
eniles
Ba
2
0
3
retalraey1 2
1
0
-1 -.5 0 .5 1
Difference in average height z-score (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 6c: LEAPS Height Z-Score Data Sample Size 300
10
eniles 5
Ba
0
4
retalraey1 3
2
1
0
-1 -.5 0 .5 1
Difference in average height z-score (weighted by std dev)
1 Draw 8 Strata Matched Big Stick Minmax
Figure 7: Sri Lanka Data Figure 8: ENE Data
P-Values on Difference in Outcome Variable P-Values on Difference in Outcome Variable
at Follow-up vs. Baseline at Follow-up vs. Baseline
30 observations sample 30 observations sample
1 1
.8 .8
up-wololfta .6 up-wololfta .6
uelav-P .4 uelav-P .4
.2 .2
0 0
0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1
P-value at baseline P-value at baseline
10th percentile 50th percentile 90th percentile 10th percentile 50th percentile 90th percentile
100 observations sample 100 observations sample
1 1
.8 .8
pu-w pu-w
llooftaeluav-P .6
.4 llooftaeluav-P .6
.4
.2 .2
0 0
0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1
P-value at baseline P-value at baseline
10th percentile 50th percentile 90th percentile 10th percentile 50th percentile 90th percentile
300 observations sample 300 observations sample
1 1
.8 .8
up-wololfta .6 up-wololfta .6
uelav-P .4 uelav-P .4
.2 .2
0 0
0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1
P-value at baseline P-value at baseline
10th percentile 50th percentile 90th percentile 10th percentile 50th percentile 90th percentile
Figure 9: If we observe baseline imbalance, and control for baseline variables, is there any
difference in follow-up balance?
9a: ENE Labor Income Data 100 Observations
4
000.
3
eniles 000.
2
000.
Ba 1
000.
0
3
000.
2
pu-woll 000.
1
Fo
000.
slor 0
3
ntoc/w 000.2
pu-wololF 000.1
000.
0
-4000 -2000 0 2000 4000
Difference in average outcome
Balanced Unbalanced
9b: LEAPS Math Test Score Data 300 Observations
025.
2
.0
eniles 015.
1
Ba .0
005.
0
2
.0
pu-woll 015.
1
.0
Fo
005.
slor 0
ntoc/w 025.2.0015.1
pu-wololF .0
005.0
-100 -50 0 50 100
Difference in average outcome
Balanced Unbalanced
Appendix 1: How would leading field experiment experts approach randomization for the same
intervention?
Our survey of experts presented researchers with the following question:
Consider the following (hypothetical) pilot experiment being carried out. An intervention is being carried out with the goal of raising
the incomes of day laborers by helping them gain new interview and employment skills. The pilot group consists of 100 men and 100
women aged 20 to 45, all in the same geographic area. Baseline data include age, current income, current weekly hours of work,
education, age, marital status, and household size. A colleague asks for your advice on how to assign 100 of these 200 individuals to
the treatment. Follow-ups will be at 6 months and 1 year. Please describe how you would recommend that they carry out this
randomization. If your answer depends on other information not provided here, state the conditions under which you would do one
method vs another. Please be specific in terms of what variables if any they should stratify on, what you recommend they should do to
check for balance, and whether (and how) they should take multiple random draws if you recommend doing this to ensure balance on
particular variables.
The responses to these questions were as follows:
· I would just randomize. Stratifying on such a small sample will cause weird things in the data.
· Random assignment with multiple draws; pick the one where the R-squared of the baseline data has the
least explanatory power. My thinking here is that 200 units seems "small", so the potential for covariate
imbalance in a single random draw large
· I would recommend that they choose the characteristics most likely to be correlated with the outcome of
interest and then stratify based on those characteristics. Afterwards, they can check the balancedness of
the sample if they have the available data.
· I would stratify on gender, education and current income. Then take a random draw after stratification
and then check balance.
· Given the relatively small sample size, I would want to make sure to stratify on variables that might
interact with the treatment. Thus, I'd certainly want to stratify on age categories, education categories
and possibly gender. If there is reason to believe, as a result of previous work on the efficacy of
interview and employment skills training (I am not familiar with this literature), that there are other
strong factors that would lead to heterogenous treatment effects, then I would add those factors as well
(that's why gender is a maybe for me here). After stratifying, I would make multiple random draws,
checking for which draw yields the best balance across all the variables I have.
· I would stratify on gender, current income and current weekly hours of work, since these would all have
1st-order effects on the estimated program effect. I would do this by generating variables for an
individual's place in the income and weekly hours of work distributions (say by quartile) and then
stratifying by these variables. After performing the randomization, I would check differences between
the treatment and control groups in the other variables, (as well as in income level and hours worked
since the levels might be slightly off if randomization is by quartile). If these differences were very
large, I would re-run the randomization. If I had a bigger sample, I would stratify by more variables. The
baseline variables listed in the example are all "important," so I would use all or a large subset of them
in the randomization. But with only 200 people, I would keep to the variables listed above and check the
rest.
· I would definitely create matched pairs and then randomly select treatment/control among each of the
matched pairs. The single-most important variable to match upon is the lagged dependent variable
(current income), but the other variables are important too. I would perform an "optimal matching"
algorithm per Biostatistics (2004), 5, 2, pp. 263275.
· Think about the covariates that are likely to affect the treatment effect, and use them to stratify. For
example, if in the context you study you have good reasons to think that the treatment effect will be
different for women than for men, or if you have no prior on gender differences but you want to be able
to check for heterogeneity in the treatment effect across gender groups, then you should really stratify by
gender. My guess is that, given the intervention you describe, you should stratify by gender and
education. Also by current income if there is a lot of heterogeneity in income at baseline, although if the
intervention is well targeted I'd expect only little variation in current income and so there would be no
need to stratify. After you've stratified by the covariates that might most interact with the treatment, take
a random draw and check, for each available baseline variable, that the difference between the two
groups is small and cannot be distinguished from zero. If the difference is significant (say at the 10%)
for at least one variable, take a new random draw and check again. Keep doing this until you find one
with no significant differences. If that's not possible, then choose one where the difference is in
household size and/or marital status (unless you think that in your context these variables are likely to
affect the treatment effect). Since you have baseline outcome data, your difference-in-difference
estimator will take care of that difference.
· I'd stratify based on sex, education, and weekly hours of work at baseline and have one random draw.
· I would divide each into above/below median. Then do block randomization within each unique cell.
Then program a loop which did that repeatedly and check the average t-statistic and max t-statistic in a
regression or series of mean comparisons of continuous variables. I would also check to see if there are
any particularly large outliers in any of the variables such that it is going to cause imbalance any way I
look at it (and with only 100 obs, that is plausible). If there are perhaps 2 or 4 outliers (hopefully not an
odd number!), then I may block randomize on the outliers to ensure they are evenly distributed (i.e.,
instead of above/below median, do three blocks, below-median, outlier, and above-median-below-
outlier.
· I would stratify on outcome and sex, ranking men and women by income and randomizing within those
pairs. I would do a single draw.
· Stratify: really depends on on (i) underlying variation in data/outcomes and (ii) whether one expects
treatment to vary much by strata. If not much treatment heterogeneity and low underlying variation I
wouldn't stratify (except maybe gender given that seems implied); if treat heterogeneity but not too
much underlying variation within strata then I would stratify by ex-ante most salient such dimension of
heterogenity (if lots of underlying variation may just want to change sample to increase observations in
the strata of interest and just stick to that). I generally prefer a simple random draw (within strata) But
there is lots of underlying variation then given power considerations would pair-wise match using
baseline data and then pick one in each pair randomly to treat assuming this does not worsen spillovers.
· Stratify on all baseline variables, and randomize within each cell without subsequently checking for
balance.
· Obviously if there is a single covariate across which the researchers require perfect balance, they should
stratify on that. If there are multiple covariates, then they encounter a dimensionality problem which is
analogous to that found in matching estimators; how to weight differences in one dimension versus
another. With few discrete categories this problem can be overcome by blocking & sub-stratification,
but with numerous continuous covariates this is harder to do. Hence the common practice of writing
loops which re-run the randomization until balance on a pre-specified set of characteristics has been
reached. This pre-specified balancing criterion then becomes analogous to a stratification criterion,
except that the standard errors on a simple t-test of pre-treatment means is no longer the correct test
statistic because it is the result of many draws rather than just one. However in a difference-in-
differences test, the two are very similar except that a pre-defined stratification criterion with a single
draw is simpler and so probably preferable.
· I would probably encourage them to find an additional dataset with the outcome of interest, run a
regression of the outcome of interest on the characteristics in the baseline group, and then stratify the
sample into roughly 20 bins of 10 people each based on the characteristics that predict the outcome of
interest. Barring any pre-period data on the outcome, this is not feasible, so you need to make those
judgements based a priori on what outcomes you expect to predict the outcome of interest. My guess
would be to stratify on something like gender, 2 age bins, 3 income bins, 2 education bins, for a total of
2*2*3*2 = 24 bins of about 9 people each. Note that one factor that severely can constrain this is the
way in which they do randomization. In many cases the complex stratification is not feasible, in which
case, I would not do it.
Appendix 2: Variables and Methods used in the Simulations (Table A2)
Panel A: Variables
Microenterprise profits in Sri Lanka (de Mel et al, 2007)
Baseline control variables: Profits, hours worked, female dummy, sales, capital, asset index, and "saw tsunami" dummy
"Unobservable" baseline variables: Household size, dirt floor dummy, age, married dummy, migrant dummy, internal migrant dummy,
relative abroad dummy, years of education, Muslim dummy, Tamil speaker dummy, risk taking index,
relative risk aversion, digit span recall index, time taken to solve a maze, entrepreneurial self-efficacy,
financial literacy, father owned a business dummy, mother owned a business dummy, going into
business to care for family members dummy, age of business, business run out of home dummy,
registered with District Secretariat dummy, registered with local government dummy, bank loan
dummy, keeps records dummy, 2 industry dummies
Stratification categories (2 variables): Gender and quartiles of baseline profits
Stratification categories (3 variables): Gender, quartiles of baseline profits, and three groups of hours worked (<=38, 39-58, >58)
Stratification categories (4 variables): Gender, quartiles of baseline profits, three groups of hours worked (<=38, 39-58, >58), and asset
index below and above median
ENE (Mexican Labor Market Survey)
Baseline control variables: Income, hours worked, female dummy, rural dummy, number of rooms in home, business owner (or
self-employed) dummy, and 1 to 5 employees dummy
"Unobservable" baseline variables: Dirtfloor dummy, has phone at home dummy, owns home dummy, age, married dummy, more than
one job dummy, social security dummy, 4 region dummies, 8 industry dummies, 4 education dummies
Stratification categories (2 variables): Gender and quartiles of baseline income
Stratification categories (3 variables): Gender, quartiles of baseline income, and three groups of hours worked (<=42, 43-48, >48)
Stratification categories (4 variables): Gender, quartiles of baseline income, three groups of hours worked (<=42, 43-48, >48), and business
owner (or self-employed) dummy
IFLS school data
Baseline control variables: Female dummy, age, government school dummy, mother education years, household size, ln
household expenditure per capita, urban dummy
Stratification categories (2 variables): Gender and quartiles of baseline ln household expenditure per capita
Stratification categories (3 variables): Gender, quartiles of baseline ln household expenditure per capita, and three groups of mothers
education years (0-2, 3-6, and 7+)
Stratification categories (4 variables): Gender, quartiles of baseline ln household expenditure per capita, three groups of mothers education
years (0-2, 3-6, and 7+), and urban vs. rural
IFLS expenditure data
Baseline control variables: Ln household expenditure per capita, household size, number of kids below 5, household head
education years, male household head dummy, household head age, urban dummy
Stratification categories (2 variables): Urban vs. rural and quartiles of baseline ln household expenditure per capita
Stratification categories (3 variables): Urrban vs. rural, quartiles of baseline ln household expenditure per capita, and three groups of head of
household education years (0-2, 3-6, and 7+)
Stratification categories (4 variables): Urban vs. rural, quartiles of baseline ln household expenditure per capita, three groups of mothers
education years (0-2, 3-6, and 7+), and head of household age above or below median
LEAPS math test score data
Baseline control variables: Math test score, english test score, age, gender, private school dummy, mother educated beyond
elementary dummy, PCA asset wealth index
Stratification categories (2 variables): Gender and quartiles of baseline math test score
Stratification categories (3 variables): Gender, quartiles of baseline math test score, and 3 categories of the PCA assets wealth index (<=-
0.4, >-0.4 & <=0.8, >0.8)
Stratification categories (4 variables): Gender, quartiles of baseline math test score, 3 categories of the PCA assets wealth index (<=-0.4, >-
0.4 & <=0.8, >0.8), and mother educated beyond elementary or not
LEAPS height z-score data
Baseline control variables: Height z-score, weight z-score, gender, mother educated beyond elementary dummy, PCA asset
wealth index, 2 district dummies
Stratification categories (2 variables): Gender and quartiles of baseline height z-score
Stratification categories (3 variables): Gender, quartiles of baseline height z-score, and 3 districts
Stratification categories (4 variables): Gender, quartiles of baseline math test score, 3 districts, and mother educated beyond elementary or
not
Panel B: Methods For each dataset,...
Pairwise Greedy Matching: ...the algorithm matches on seven baseline control variables listed in Panel A.
Big Stick Rule: ...the method calculates p-values on difference in seven baseline variables listed in Panel A.
Draw with Minmax T-Stat: ...the method calculates t-stats on difference in seven baseline variables listed in Panel A.