Aliprantis - When Should Children Start School

When Should Children Start School?
Dionissi Aliprantis
Federal R eserve Bank of Cleveland
This paper studies causal effects informative for deciding the age when children
should start kindergarten. I present evidence from the Early Childhood Longitu-
dinal Study, Kindergarten Class of 1998–99 ðECLS-KÞ that standard instrumental
variable strategies do not identify effects of delaying kindergarten entry for any
subpopulation of interest. I propose and implement a new strategy for identifying
individual-level education production function parameters. Estimates indicate that
there can be decreasing and even negative returns to relative age: For the oldest
children in a cohort, educational achievement in third grade decreases as their age
relative to that of their classmates increases.
I. Introduction
Investing in early childhood education appears to be good policy from
several perspectives. Empirical evidence from studies such as the Perry
Preschool Experiment and the Abecedarian Project indicate that invest-
ments in early childhood can significantly improve outcomes over the life
cycle ðCurrie 2001; Heckman et al. 2010bÞ. Even if one ignores the nor-
mative arguments that such interventions could help move our society
toward greater equality of opportunity, there are still compelling argu-
ments for investing in early childhood education. There is reason to
believe that educational investments made at early ages yield higher
returns than similar ones made later in the life cycle ðCunha et al.
2005Þ, and researchers have calculated the returns to society from such
investments to be extremely high ðRolnick and Grunewald 2003; Heck-
man et al. 2010aÞ. Policy makers have responded to this research by
increasing investments in early childhood education during recent years
I thank Becka Maynard, Francisca G.-C. Richter, Ken Wolpin, Petra Todd, Fabienne
Doucet, Guhan Venkatu, Russ Cole, Rob Grunewald, Kimber Bogard, Tim Dunne, Julie
Cullen, Isaac Ehrlich, Ed Vytlacil, participants at the Penn Institute of Education Sciences
Seminar and Penn Empirical Micro Club, and two anonymous referees for helpful com-
ments. The research reported here was supported by the Institute of Education Sciences,
US Department of Education, through grant R305C050041-05 to the University of Pennsyl-
vania. The opinions expressed are those of the author and do not represent views of the
Federal Reserve Bank of Cleveland, the Board of Governors of the Federal Reserve System, or
the US Department of Education.
[ Journal of Human Capital, 2014, vol. 8, no. 4]

© 2014 by The University of Chicago. All rights reserved. 1932-8575/2014/0804-0003$10.00
481
This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

All use subject to University of Chicago Press Terms and Conditions (http://www.journals.uchicago.edu/t-and-c).
482 Journal of Human Capital
ðKirp 2007Þ, and these investments figure only to increase after the
Obama adminstration’s recent proposal to provide preschool for all low-
and moderate-income 4-year-old children ðWhite House 2013Þ.
The coordination of related policy will likely determine the effective-
ness of new investments in early childhood education ðBogard and
Takanishi 2005Þ. One related policy widely believed to have long-run
effects, and one that has changed considerably in recent years, is the age
at which children start kindergarten. In the United States, children are
eligible to enroll in kindergarten if they turn 5 before a specific date set
by their state of residence, known as an entrance cutoff date. In recent
decades many state governments have chosen to move entrance cutoff
dates earlier, increasing the average age at which children start kindergar-
ten in the United States.1 Parents and schools have also helped to increase
the average entrance age during recent decades by “redshirting” children,
or choosing to delay their entry into kindergarten ðWeil 2007; Elder and
Lubotsky 2009Þ.
This paper presents empirical evidence related to three causal effects
that are crucial for determining the age when children should start school:
ðiÞ the effect of delaying a child’s kindergarten entry by 1 year, also known
as the effect of redshirting a child; ðiiÞ the effect of a state changing its
entrance cutoff date; and ðiiiÞ the effect of a child’s birth ðand conceptionÞ
being moved to an earlier date. While effect i is relevant for parents’ and
schools’ decisions, effect ii is relevant for entrance age policy makers, and
effects ii and iii are relevant for gaining a general understanding of
children’s development process. The empirical analysis uses data from
the Early Childhood Longitudinal Study, Kindergarten Class of 1998–99
ðECLS-KÞ to understand these effects on educational achievement. The
strength of this data set is that it contains an unusually rich set of individual-
level variables, allowing for the identification of effects unidentified by
state-level data. The weakness of the ECLS-K, however, is that it allows us
to study only short-term effects.
The paper makes four key contributions in the process of investigating
the aforementioned effects. The first is to show that the effects of delaying
entry identified in the literature may be less informative than is currently
appreciated. I empirically test and reject the hypothesis of treatment effect
homogeneity for academic achievement. When combined with the failure
of the instrument monotonicity condition known to occur in this context,
such evidence of essential heterogeneity of treatment effects suggests that
most instrumental variable ðIVÞ strategies do not identify the effect of
choosing to delay a child’s entry for any subpopulation of interest ðBarua
and Lang 2009; Aliprantis 2012Þ. Thus the best evidence on the effects of
redshirting is likely to come from regression discontinuity designs, which
1
For example, 22 states moved their cutoff dates to an earlier point in the school year
between 1975 and 2000 ðStipek 2002Þ. See fig. 1 in Elder and Lubotsky ð2009Þ for a summary
of the evolution of these laws during the past 40 years.

When Should Children Start School? 483
avoid this problem at the cost of identifying effects pertaining to a more

restrictive subpopulation of children than IV parameters.
A second contribution of the paper is to present a strategy for identify-
ing education production function ðEPFÞ parameters that imposes a set of
assumptions alternative to those imposed by value-added EPF strategies.
The central assumption of the identification strategy is that ability is
distributed independently of race. Using race as an instrument for family
inputs before entering school, under additively separable EPF specifica-
tions, it is possible to identify ability with test scores obtained upon
entering school. The IV assumptions identifying ability are some of the
most likely assumptions to be violated in the analysis, and I discuss their
limitations so that future work might improve on them. The assumptions
necessary for regression coefficients obtained via ordinary least squares
ðOLSÞ to identify EPF parameters are likely to be satisfied when an
estimate of ability is included as an independent variable.
This identification strategy facilitates the third contribution, which is
obtained by estimating individual-level EPFs in which a student’s age at
enrollment ði.e., entrance ageÞ and age relative to that of his or her
classmates ði.e., relative ageÞ are EPF inputs. Estimation results indicate
that increasing the relative age of a student’s classmates would actually
increase their achievement in the spring of third grade, at least for the
older children in a class. The effect associated with this counterfactual is
policy relevant because it tells us the effect on a given child of the state
moving its entrance cutoff date to an earlier point in the year.2
The fourth and final contribution of this analysis is to show that effects
are remarkably heterogeneous by subpopulations. With respect to gen-
der, relative age effects on boys are much larger than the effects on girls.
Effects are also heterogeneous by home environment. Children with few
books at home perform much worse when entering at an older entrance
age and benefit from being relatively older than their peers. Both of these
effects are reversed for children with many books at home, although their
magnitudes are much smaller.
These empirical results are relevant for comparing alternative views of
early childhood development. One view of early childhood development
is that resources devoted toward a child’s learning and therefore the
child’s development trajectory are determined when children start school
through a zero-sum game ðDawid and Muehlheusser 2012Þ. In an institu-
tional setting in which relative advantage in initial achievement is what
helps a child to develop, relative age is a margin along which parents can
increase their children’s “readiness” for school. It is through this institu-
tional lens that some parents and teachers have interpreted evidence
of positive entrance age effects on academic achievement ðBedard and
2
This counterfactual contrasts with what could be estimated with a state-level identifica-
tion strategy and data. Such a counterfactual would represent a combination of this effect
and that from redshirting for various subpopulations.

Dhuey 2006; Datar 2006Þ, grade progression ðMcEwan and Shapiro 2008;
Elder and Lubotsky 2009; Dobkin and Ferreira 2010Þ, and attainment
ðFredriksson and Öckert 2005; Puhani and Weber 2005Þ to imply that
redshirting the youngest children in a cohort is optimal whenever possible
ð“Redshirting” episode on 60 Minutes, CBS News, March 4, 2012Þ.3
The central result of this paper, that initial age advantages need not
translate into superior development trajectories, is more consistent with
the view that normal development trajectories often cross each other,
being heterogeneous over individuals ðShonkoff et al. 2000, 28–29Þ, age
ð29–30Þ, and initial achievement ðWillett 1988Þ. More specifically, the
results in this paper are consistent with the literature finding negligible
effects from increased entrance age on IQ and earnings ðBlack, Devereux,
and Salvanes 2011Þ and negligible effects from relative age on achieve-
ment ðCascio and Schanzenbach 2007Þ or adult outcomes such as employ-
ment, wages, and home ownership ðDobkin and Ferreira 2010Þ. Most
consistent with the results in this paper, Elder and Lubotsky ð2009Þ also
find that being relatively younger within a cohort improves academic
achievement.4
The paper finishes by discussing several challenges to determining
optimal entrance age policy that are unlikely to be overcome. Most im-
portant is that causal effects estimated in the literature, including in this
analysis, are unlikely to represent inherent or unaltering relationships
between age and achievement. Thus strong assumptions are necessary in
order to predict effects from future changes to entrance and relative ages
using effects identified in past data ðAliprantis 2014bÞ.
Viewing the empirical results together with these challenges suggests
that attention should be given to designing a broad class of policies that
can accommodate the significant heterogeneity of entrance and relative
ages in any cohort of children. An obvious and existing policy focus would
be to improve the environment before entering kindergarten, which
should reduce the disparities in what kids know and can do upon entering
school ðShonkoff et al. 2000, 386–88Þ. Another obvious proposal would be
to focus on supporting children’s development once they arrive at school,
making sure that each classroom is ready for its children rather than vice
versa. This might be implemented through mixed-age classrooms, per-
haps following Montessori or similar models ðLillard 2008; Khan 2012Þ.
This proposal might also be implemented in just the opposite way, by al-
locating young children to homogeneous classrooms using demographic
characteristics such as age ðinterview with Malcolm Gladwell, 60 Minutes,
March 4, 2012Þ and/or gender. Further research into the effects of various
3
In addition to the literature on the effects entrance age has on academic outcomes, there
is also a literature documenting the important role of relative age in competitive sports
ðMuscha and Grondin 2001Þ and leadership ðDhuey and Lipscomb 2008Þ. Appendix A
discusses the background of entrance age policy in greater detail.
4
Elder and Lubotsky ð2009Þ additionally find evidence that age-related differences in
academic performance dissipate as children advance in school, attributing most of the initial
differences to the accumulation of skills before children enter kindergarten.

entrance age policies should be helpful for designing policies that can
support skills more efficiently begetting skills ðCunha and Heckman 2007Þ
without advantage necessarily begetting advantage.
The remainder of the paper is organized as follows. Section II presents
empirical evidence that redshirting choices are made under essential
heterogeneity and discusses the implications of this evidence. Both the
ECLS-K data set used in the analysis and some variable definitions are
introduced in this section. Section III presents the strategy for identifying
EPF parameters, with Section III.B first describing the selection of the
sample used in estimation. Section III.C specifies the EPFs to be estimated
and explicitly states the identifying assumptions necessary to identify their
parameters. Section III.D discusses the variation in the data used to
identify the EPF parameters, which also merits attention. Estimation re-
sults are presented in Section III.E, and Sections IV and V use these
estimated EPFs to determine the effects of counterfactual entrance age
policies and individual choices. Section VI discusses broad policy implica-
tions, and Section VII presents conclusions.
II. Effects of Delaying Entry

A. Methodological Overview
Following the terminology in Bedard and Dhuey ð2006Þ, I refer to the
relative age at which a child would be observed if he entered kindergarten
when first eligible as assigned relative age and the child’s actual age relative
to the school’s cutoff date as observed relative age. A central feature of the
identification strategy in Section III will be to select a sample for which
observed and relative ages are identical. Thus when relative age is dis-
cussed in the analysis, this term will refer to observed relative age unless
otherwise specified.
A key methodological difficulty in identifying effects of observed en-
trance and relative ages is that their allocation tends to be nonrandom. In
response to the nonrandom determination of observed entrance and
relative ages, many analyses in the literature use date of birth and entrance
age cutoffs to instrument for observed entrance age. Barua and Lang
ð2009Þ and Aliprantis ð2012Þ show that since redshirting creates a violation
of the monotonicity assumption from Imbens and Angrist ð1994Þ, no
interpretable parameter can be recovered from the data unless treatment
effects are homogeneous. However, since adults choose when to enroll
children, there is likely to be a correlation between the unobserved factors
determining the redshirting decision and the unobserved factors deter-
mining age effects.5
5
For example, parents might make redshirting decisions on the basis of information and/
or objectives relating to sports ðMuscha and Grondin 2001Þ, leadership ðDhuey and Lips-
comb 2008Þ, educational achievement ðMcEwan and Shapiro 2008Þ, or notions about so-
cial maturity ðGraue 1993Þ. Economic constraints are likely to play a role in these decisions
ðAliprantis 2012Þ.

In this scenario of essential heterogeneity, to be formally defined in

Section II.B, the entire IV framework breaks down without monotonicity,
so that no interpretable parameter can be recovered from the data
ðHeckman and Vytlacil 2005Þ. The key intuition for this application is that
when monotonicity fails to hold, the intent-to-treat parameter is no longer
a weighted average of effects on compliers and no effect on never-takers.
Therefore, the intent-to-treat parameter can no longer be simply scaled by
the share of compliers to obtain the local average treatment effect. The
intuition discussed here is formally stated in Heckman and Vytlacil ð2005Þ
and Aliprantis ð2012Þ.
In the absence of treatment effect homogeneity, the best evidence on
the effects of redshirting is likely to be found in studies employing regres-
sion discontinuity designs. In exchange for avoiding the above identifica-
tion problems, regression discontinuity designs identify a parameter per-
taining to a more specific subpopulation of children than IV parameters
ðchildren who are born on the entrance cutoff date and who would not be
redshirted if deemed eligible; see theorem 3 in Hahn, Todd, and van der
Klaauw ½2001Þ.
B. Model and Test of Redshirting Effect Homogeneity

Since the crucial treatment effect homogeneity assumption allowing for IV
techniques to identify the effects of redshirting in the absence of mono-
tonicity is empirically testable, I proceed to conduct some of the relevant
hypothesis tests developed in Heckman, Schmierer, and Urzúa ð2010Þ.
The effect homogeneity assumption can be studied using a joint model
of selection into treatment and potential outcomes. Two features of the
model used in this section merit special attention and are discussed in
detail in Appendix B. First, the model does not allow for monotonicity to
be violated. This feature of the model will be accommodated by choosing
a subsample of children who all turn 5 on one side of the entrance cutoff
date, avoiding the nonmonotonicity pointed out in Barua and Lang
ð2009Þ and Aliprantis ð2012Þ. Second, redshirting is modeled as a binary
treatment representing a very specific combination of changes to a stu-
dent’s age relative to that of his classmates ði.e., their relative ageÞ and
their age at enrollment ði.e., their entrance ageÞ.6
These features of the model are important because the model in
Section III.C allows for continuous treatments and because the evidence
considered here about essential heterogeneity in a restricted sample for
6
Thus the redshirting treatment is not exchangeable with all the other counterfactuals
considered in this paper and throughout the literature. It is important to note that this
counterfactual also represents manipulation of treatments occurring later in the life cycle
that are not studied in this analysis. One such treatment is the years of schooling students
must complete before becoming eligible to drop out of school, which was originally used in
Angrist and Krueger ð1991Þ and is thought to be the main mechanism affecting the adult
outcomes studied in Cascio and Lewis ð2006Þ, Dobkin and Ferreira ð2010Þ, McCrary and
Royer ð2011Þ, and Cook and Kang ð2013Þ. Another treatment manipulated is the age at which
students first enter the labor market ðDeming and Dynarski 2008Þ.

which monotonicity must hold is taken as evidence about essential het-

erogeneity in the full sample for which the instrument monotonicity as-
sumption does not hold. Furthermore, the model analyzed in Section III.
C and beyond allows for a complementary set of changes in entrance and
relative ages to those represented by redshirting.
Since the specific treatment of interest in this section is that of delaying
entry by an entire year, define a binary treatment variable as

1 if child i is redshirted
Di ; ð1Þ
0 if child i enters when first eligible:
Here it is worth discussing the precise combination of changes to en-

trance and relative ages represented by the binary redshirting treatment
variable. To illustrate, consider a child who turns 5 on August 31 and
faces a September 1 entrance cutoff date. If enrolled when first eligible,
this child will be the youngest eligible child in his cohort. If school starts
on September 1, the child will have an entrance age of 5 years and 1 day
and will be 12 months younger relative to the oldest eligible child in the
cohort. If, in contrast, the child is redshirted, the child will have an en-
trance age of 6 years and 1 day and will be 1 day older relative to the old-
est eligible child in the cohort. The importance of highlighting the spe-
cific counterfactual represented by redshirting will become clearer in
later sections that examine counterfactual changes to entrance cutoff
dates and birth dates, which represent alternative changes to children’s
entrance and relative ages.
Returning to the model, assume that potential outcomes in each treat-
ment state, Yð0Þ and Yð1Þ, are functions of observable characteristics XD
and some treatment-level specific unobservable component Uj for j ∈
f0, 1g:
←
Y ð0Þ 5 m0 ðX 0 Þ 1 U0 ;
← ð2Þ
Y ð1Þ 5 m1 ðX1 Þ 1 U1 :
←
Throughout the analysis I use “5” to denote statistical equations and “5 ”
to denote structural equations to avoid the confusion discussed in Chen
and Pearl ð2012Þ. Statistical equations convey information about observed
changes in the outcome and treatment variables, while structural equa-
tions convey information about controlled changes in these variables. This
distinction is similar to the one between setting and conditioning discussed
in Heckman ð2005Þ, and further discussion can be found in Chen and
Pearl ð2012Þ and Aliprantis ð2014bÞ. I am interested in identifying fea-
tures of the joint distribution of treatment Di and the treatment effect
bi ; Yi ð1Þ 2 Yi ð0Þ, and in this analysis the outcome of interest is educa-
tional achievement.
Suppose further that individuals select into treatment on the basis of
a latent index:

n
← 1 if D * ≥ 0
D5 ð3Þ
0 otherwise:
Assuming X 0 5 X1, each individual’s latent index D * depends on observ-
able characteristics X, an instrument Z, and some unobserved component
V as follows:7
←
D * 5 mD ðX ; Z Þ 2 V
← ð4Þ
5 mðX Þ 1 gZ 2 V :
Define the propensity score conditional on the instrument Z to be
P ðX; Z Þ ; PrðD 5 1jX; Z Þ ; FV ðmD ðX; Z ÞÞ; ð5Þ
with the conditional quantiles of the unobservable characteristics deter-
mining selection into treatment denoted by UD 5 FV jX ðV jX Þ. I maintain
the following additional assumptions adapted from Heckman and Vytlacil
ð2005Þ.
EH1. gi 5 g for all i and g ≠ 0.
EH2. ðU0i ; U1i ; UDi Þ
j j Z jX .
i i
EH3. The distribution of UD is absolutely continuous.
EH4. E½|Yð0Þ| < ` and E½|Yð1Þ| < `.
EH5. 0 < PðX, ZÞ < 1 for all X.
While assumption EH3 is made for exposition, EH4 ensures that the
model’s treatment effect parameters are well defined, and EH5 ensures
overlap between treatment and control groups for all values of X ðHeckman
et al. 1998Þ. Most important for this analysis are EH1 and EH2. An impor-
tant implication of EH2 is that Z affects the outcome variable only through
treatment, and EH1 ensures that the instrument Z has a monotonic im-
pact on the probability of selecting into treatment. The monotonicity
assumption EH1 does not face the challenges described in Barua and
Lang ð2009Þ and Aliprantis ð2012Þ because the analysis in this section will
be focused on a subsample of children who all turned 5 before the en-
trance cutoff date and because entrance age is here modeled as a binary
variable. Further discussion of the monotonicity assumption EH1 can be
found in Appendix B, and Heckman and Vytlacil ð2005Þ contains further
general discussion of these assumptions.
An additional assumption often made in the statistics and economics
literature is that agents select into treatment only on the basis of the ob-
served characteristics measured in X. This assumption, typically referred
to as strong ignorabilityÞ, implies treatment effect homogeneity ðTEHÞ. In
our additive model, TEH can be stated as follows.
TEH. U1 2 U0
jj U .
D
7
Although assuming X 0 5 X 1 is standard throughout the literature, this is a nontrivial
assumption. See Aliprantis ð2014bÞ for an analysis of selection into covariates in response to
treatment assignment.

Note that nowhere in EH1–EH5 is any restriction imposed regarding

the relationship between the distributions of the unobserved components
of potential outcomes and selection into treatment, ðU0i, U1iÞ and UDi, so
that treatment effect homogeneity need not characterize the model. As
established in Imbens and Angrist ð1994Þ, Heckman and Vytlacil ð2005Þ,
and Heckman, Urzua, and Vytlacil ð2006Þ, treatment effects can still be
identified in models with essential heterogeneity or in which selection
into treatment is determined in part by anticipated gains from treatment
ðU1i 2 U0iÞ that depend on individual-level characteristics that are unob-
served by the econometrician ðUDiÞ. To be precise, Heckman et al. ð2006Þ
define essential heterogeneity as the scenario in which treatment effect
heterogeneity is correlated with selection into treatment even conditional
on observables, which in our additive model can be stated as follows.
EH. CovðU1 2 U0 ; UD Þ ≠ 0.
If we define bi ; Yi ð1Þ 2 Yi ð0Þ, then note that essential heterogeneity

can alternatively be written as Covðbi ; Di Þ ≠ 0jX i , and treatment effect ho-
mogeneity can alternatively be written as bi j j D jX .
i i
The case in which the unobserved characteristics determining selec-
tion are not related to treatment effect heterogeneity, as assumed under
TEH, has a testable implication that all instruments identify a homoge-
neous parameter. Heckman et al. ð2010Þ show that the null hypothesis H 0:
bj j DjX can be stated equivalently as
H0IV : bIV;½p ;p 1 5 bIV;½p ;p 2 ; ð6Þ

1 2
where ½p ; p 1 and ½p ; p 2 represent two disjoint intervals of the support of

1 2
PðX, ZÞ and
CovðY; P ðX; Z ÞjP ðX; Z Þ ∈ ½p ; p j Þ
bIV;½p ;pj 5 ð7Þ
j
j VarðP ðX; Z ÞjP ðX; Z Þ ∈ ½p ; p j Þ
j
for j 5 1, 2. The Wald test statistic of H0IV is asymptotically distributed as

a x21 random variable, and Heckman et al. ð2010Þ characterize the prop-
erties of this test under a variety of scenarios.
The null hypothesis in equation ð6Þ is based on the fact that under
treatment effect homogeneity, all instruments identify the same homo-
geneous parameter. Using the propensity score as an instrument over
different intervals of its support should thus identify the same param-
eter in the absence of essential heterogeneity. That is, under treatment
effect homogeneity, exogenous variation in the propensity score should
have the same impact on the outcome variable, regardless of where in the
support of the distribution of propensity scores this variation occurs. See
Imbens and Angrist ð1994Þ, Heckman and Vytlacil ð2005Þ, Heckman et al.
ð2010Þ, or Aliprantis ð2014aÞ for more intuition on why TEH implies H0IV .

C. Data
The analysis presented in this paper utilizes data from the Early Child-
hood Longitudinal Study, Kindergarten Class of 1998–99 ðECLS-KÞ data
set, which is a nationally representative sample of over 22,000 children
enrolled in over 1,200 schools who started kindergarten in the fall of
1998. Data were collected during the fall and the spring of kindergarten
ð1998–99Þ, the fall and spring of first grade ð1999–2000Þ, the spring of
third grade ð2002Þ, fifth grade ð2004Þ, and eighth grade ð2007Þ from the
children, their parents/guardians, teachers, and school administrators.8
Since the survey in round 3 ðfall of first gradeÞ was designed to gather
data from only a subsample of about 30 percent, this round will not be
considered for the analysis. To account for the sampling scheme used
to collect the ECLS-K, weights will be utilized in all estimation.
Variables.—We define monthly groups for observed relative age as the
month relative to the entrance cutoff date ðof the year when entering
kindergartenÞ during which a child turns 5. For the cohort of children
we observe entering school in the fall of 1998, we will observe children
entering who are too young to be eligible but who enter anyway, chil-
dren who enter when first eligible, and children who were eligible to
begin school in the fall of 1997 but whose parents decided to enter them
in kindergarten in the fall of 1998. We define the youngest of these
children to be in the relative age group M –12 and the oldest to be in the
relative age group M24. These relative age groups can be seen in figure 1.
In order to assign children in the ECLS-K to these relative age groups, I
first construct their observed entrance and relative ages. Let TðCalendar
DateÞ denote a function of dates from the calendar to a time line in which
day 1 is January 1, 1990. I define each child’s observed entrance age EA
to be her age in months over 5 years at the start of the school year when
she began school, and I implement this definition empirically as the child’s
age over 5 on September 1, 1998. I also define a child’s observed relative
age RA for the entire sample to be the amount of time before the en-
trance cutoff date she turned 5 years of age:
EA ; T ðEntrance Date; Year of EntryÞ 2 T ð5th BDayÞ
ð8Þ
; T ðSept: 1; 1998Þ 2 T ð5th BDayÞ;
RA ; T ðCutoff Date; Year of EntryÞ 2 T ð5th BDayÞ

ð9Þ
; T ðCutoff Date; 1998Þ 2 T ð5th BDayÞ:
Definitions ð8Þ and ð9Þ help to illustrate that since redshirting changes
the year of a child’s entry by 1 year, counterfactual changes to year of entry
change both entrance and relative ages. With regard to relative age, for
children who enter on time when first eligible, the youngest children
will have a relative age of 0, and the oldest children will have a relative
8
Eighth grade will be the last round of data collection because of sample attrition.

Figure 1.—Identifying variation in entrance and relative ages: A, children facing a Septem-
ber 1, 1998, cutoff date; B, children facing a November 1, 1998, cutoff date. Date/location a : a
child turning 5 on January 1, 1998, facing a September 1, 1998, cutoff will have EA 5 8 and
R A 5 8. Date/location b : a child turning 5 on March 1, 1998, facing a September 1, 1998,
cutoff will have E A 5 6 and R A 5 6. Date/location c : a child turning 5 on January 1, 1998,
facing a November 1, 1998, cutoff will have E A 5 8 and R A 5 10. Date/location d : a child
turning 5 on March 1, 1998, facing a November 1, 1998, cutoff will have EA 5 6 and R A 5 8.
age of 12 months. Redshirted children will have relative ages between 12

and 24 months, and children who enter early will have relative ages
between 212 and 0 months. These variables are shown in figure 1 and
will be discussed in depth in Section III.
I use school-level entrance cutoff dates to construct entrance and rela-
tive age variables. In the ECLS-K, 6 percent of school-level cutoff dates are
not ascertained. For an additional 1 percent of children, a school-level
cutoff date is reported to be not applicable, and an additional 7 percent of
children have an implausible cutoff date ð1995 or earlierÞ. I consider the
data to be missing for all these children, a total of 14 percent of the
children in the ECLS-K.
I use school-level entrance cutoff dates rather than state-level ones
because of increased accuracy, but these measures are almost identical for
the sample used in the ensuing analysis.9 For the sample, the median
difference between school- and state-level measures is 1.7 days, and the
90th percentile of the difference is 2.5 days. Private school students make
up 19 percent of the sample, and figures 2 and 3 show that most of the
9
This sample is described in the next section, but it is worth noting that it explicitly ex-
cludes states with local education authority options in order to limit the endogeneity of en-
trance cutoff dates to state-level decisions. This restriction helps to ensure that these vari-
ables match closely.

Figure 2.—Public school students in the sample: differences in school- and state-level en-
trance cutoff dates ðin monthsÞ.
Figure 3.—Private school students in the sample: differences in school- and state-level
entrance cutoff dates ðin monthsÞ.

large differences in the entrance cutoff date measures are found among
this population.
The outcome measures used in this analysis are math and reading item
response theory ðIRTÞ test scores.
D. Empirical Specification and Test Results

In order to test H0IV with data from the ECLS-K, I follow Heckman et al.
ð2010Þ and first estimate a probit model to obtain predicted propensity
scores for redshirting. I estimate this probit on the sample of kindergart-
ners in relative age groups fM1, M2, M3, M13, M14, M15g where the instru-
ment Z is assigned relative age and its square. The selected subsample
generates important restrictions when combined with the simple model
in equations ð1Þ–ð5Þ. Most importantly, since the subsample consists only
of children on one side of the entrance cutoff date, the violation of mono-
tonicity detailed in Barua and Lang ð2009Þ and Aliprantis ð2012Þ is not
possible.
Time-invariant demographic characteristics X are used to estimate pro-
pensity scores. Included in X are the number of children’s books at home,
whether the child ever received benefits from the Special Supplemental
Nutrition Program for Women, Infants, and Children ðWICÞ, and whether
the child’s mother is present and works 35 hours or more as measured
in the spring of kindergarten; the socioeconomic status ðSESÞ quintile of
the child and the mother’s education level measured in the fall of kinder-
garten; whether the father was present in the household during the fall
of first grade; and gender and race dummies.
Figure 4 shows a histogram of the estimated propensity scores p^ðx; zÞ.
We see the expected pattern that children who were redshirted tend
to have higher estimated propensity scores. Also consistent with expecta-
tions is that the empirical estimates in figure 5 show that the instrument
induces fewer individuals into treatment at lower values of m ^ðX Þ ðfrom
eq. ½5Þ than at higher values.
In order to test H0IV , I must divide the sample into two subsamples using
the estimated propensity score. As noted in Heckman et al. ð2010Þ, there is
currently no a priori way to select the intervals to be used to split the
sample. As seen in figure 4, a cut point should be selected that can ac-
commodate the relatively smaller sample of redshirted students while
also ensuring common support ðHeckman et al. 1998Þ. With these con-
siderations in mind, I choose to divide the sample into those with pro-
pensity scores below the 33rd percentile of the redshirted sample and
those with propensity scores between the 33rd and 66th percentiles of
the redshirted sample. Thus if ½p ; p A is the interval above the split point,
A
p 5 t0:33 and p A 5 t0:66 , where t0:33 and t0:66 are determined using the de-
A
layed sample. Analogously, if ½p ; p B is the interval below the split point,
B
p 5 t0:0 and p B 5 t0:33 for the redshirted sample.
B

Figure 4.—Distribution of estimated propensity scores ðby initial enrollment decisionÞ.
Propensity scores are predicted using the parameters of a probit model estimated on assigned
relative age and its square, the number of children’s books at home, whether the child ever
received WIC benefits, and whether the child’s mother is present and works 35 hours or more
as measured in the spring of kindergarten; the SES quintile of the child and the mother’s
education level measured in the fall of kindergarten; whether the father was present in the
household during the fall of first grade; and gender and race dummies.
Figure 5.—Predicted probability of redshirting ðconditional on mðX Þ and percentiles of Z Þ

After partitioning the propensity scores into these intervals, I then esti-
mate a regression of test scores on the predicted propensity score p^ðx; zÞ
as well as the covariates in X with separate slopes for p^ðz; xÞ ∈ ½p ; p A and
p^ðz; xÞ ∈ ½p ; p B . The test of H0IV is empirically implemented as a Wald
A
B
test of the hypothesis that these slope coefficients are equal over the
different intervals of p^ðx; zÞ.
Tables 1 and 2 report the results of these tests together with the esti-
mates of the two coefficients at each test date. It is most reasonable to
view these as four tests of the hypothesis that selection into redshirting
is characterized by treatment effect homogeneity with respect to math
test scores and four separate tests of the hypothesis that selection into
redshirting is characterized by effect homogeneity with respect to read-
ing test scores. The final column presents corrected critical values to con-
trol the familywise error rate arising from the multiple comparisons made
here. These corrected critical values are displayed so that the ranked
p-values are compared to
a a a
; ; :::;
n n21 1
in a sequentially rejective Bonferroni test, as developed in Holm ð1979Þ,
where n is the number of tests performed and a is the desired size.
Tables 1 and 2 present strong evidence against treatment effect ho-
mogeneity: All tests reject the null of effect homogeneity. This evidence
TABLE 1
Testing H0IV for Math: bIV;½p ;p A 5 bIV;½p ;p B
A B
^ ðx; zÞ of Redshirted Subsample

Above/Below Split Using p
Adjusted Critical
Â
b ^B
b p -Value Value a 5 :05
Fall kindergarten 17 29 .000 .013
ð4Þ ð2Þ
Spring kindergarten 19 210 .000 .016
ð6Þ ð3Þ
Spring 1st grade 24 213 .000 .025
ð10Þ ð6Þ
Spring 3rd grade 29 2 .035 .05
ð14Þ ð9Þ
Note.—bA 5 bIV;½A;A is the IV coefficient estimated for the subsample whose characteristics
rank them above the cut point A; bB is defined analogously for the values below the cut point.
The adjusted critical value corresponds to comparing the ranked p-values to a=n, a=ðn 2 1Þ, . . . ,
a=1 in a sequentially rejective Holm-Bonferroni test, where n is the number of tests per-
formed and a is the desired size. Instruments in Z are assigned relative age and its square.
Covariates in X also used to estimate the propensity score include the number of children’s
books at home, whether the child ever received WIC benefits, and whether the child’s
mother is present and works 35 hours or more as measured in the spring of kindergarten;
the SES quintile of the child and the mother’s education level measured in the fall of
kindergarten; whether the father was present in the household during the fall of first grade;
and gender and race dummies. IV parameters are estimated by regressing test scores on
predicted propensity scores and covariates X using the subsample of kindergartners in rela-
tive age groups fM13, M14, M15, M25, M26, M27g.

TABLE 2
Testing H0IV for Reading: bIV;½p ; pA 5 bIV;½p ; pB
A B
^ ðx; zÞ of Redshirted Subsample

Above/Below Split Using p
Adjusted Critical
Â
b ^B
b p -Value Value a 5 :05
Fall kindergarten 17 21 .000 .013
ð4Þ ð3Þ
Spring kindergarten 16 24 .007 .016
ð7Þ ð4Þ
Spring 1st grade 10 218 .013 .05
ð13Þ ð8Þ
Spring 3rd grade 33 1 .010 .025
ð14Þ ð10Þ
Note.—bA 5 bIV;½A;A is the IV coefficient estimated for the subsample whose characteristics
rank them above the cut point A; bB is defined analogously for the values below the cut
point. The adjusted critical value corresponds to comparing the ranked p -values to a=n,
a=ðn 2 1Þ, . . . , a=1 in a sequentially rejective Holm-Bonferroni test, where n is the number
of tests performed and a is the desired size. Instruments in Z are assigned relative age and
its square. Covariates in X also used to estimate the propensity score include the number
of children’s books at home, whether the child ever received WIC benefits, and whether the
child’s mother is present and works 35 hours or more as measured in the spring of kin-
dergarten; the SES quintile of the child and the mother’s education level measured in
the fall of kindergarten; whether the father was present in the household during the fall of
first grade; and gender and race dummies. IV parameters are estimated by regressing test
scores on predicted propensity scores and covariates X using the subsample of kindergart-
ners in relative age groups fM13, M14, M15, M25, M26, M27g.
against effect homogeneity is a rather negative result as it suggests that

redshirting decisions are made in settings of essential heterogeneity with
respect to educational achievement in early childhood. That is, the benefit
of delaying entry for achievement in early childhood is not entirely pre-
dicted by observed characteristics at the time of entry.10
Effect heterogeneity and homogeneity are fundamentally assumptions
about how much the econometrician can observe. The ECLS-K is an ex-
tremely rich data set, and in many data sets, less information will be ob-
servable. Combined with the theoretical result on the violation of mono-
tonicity in Barua and Lang ð2009Þ and Aliprantis ð2012Þ, these empirical
results indicate that date of birth and entrance cutoff dates ði.e., assigned
relative ageÞ should be used as an instrument for observed entrance age
only with great caution in the presence of redshirting.
III. Identifying Education Production Function Parameters

A. Identification Strategy: Overview
The evidence in Section II discourages us from using assigned relative
age as an instrument for observed entrance age in the presence of red-
10
This need not be true when the benefit of delaying is cast in terms of the counterfactual
outcomes related to educational achievement at an older age, competitive sports ðMuscha and
Grondin 2001Þ, or leadership ðDhuey and Lipscomb 2008Þ. Also, this statement about coun-
terfactual outcomes is compatible with the recent finding that cognitive ability measured
before children enter school is not predictive of redshirting ðBassock and Reardon 2013Þ.

shirting. However, this evidence does not preclude the possibility of

estimating entrance age effects using an identification scheme inspired
by Angrist and Krueger ð1991Þ if we can find a sample that is not af-
fected by redshirting. This is precisely the strategy taken for identifying
the effects of a state counterfactually changing its entrance cutoff date, as
well as effects of counterfactual changes to a child’s conception and birth
dates.
The identification strategy in the ensuing analysis proceeds in two
major steps. In the first step a subsample is selected to be as free from
redshirting as possible, so that estimation is free from the threats to iden-
tification posed by redshirting ðsee Sec. IIÞ. However, even in a perfectly
randomized sample, as ideally obtained in this first step, several hurdles
remain to the identification of the EPF parameters allowing us to under-
stand the causal effects of interest. In the second step, EPFs are specified
and estimated with a particular focus on assumptions related to ability,
which is unobserved. The critical feature of the second step is the assump-
tion that ability is distributed independently of race. This allows us to use
race as an instrument for family inputs, with ability identified in an ad-
ditively separable model by the variation in test scores remaining after
controlling for family inputs. Since a more common approach to the
second step of the identification strategy would be to estimate a value-
added specification, the assumptions of these alternative identification
strategies are compared in Appendix B.
B. Identification Strategy: Sample Selection

If we assume that parents’ decision rule for determining observed entry
age does not change over time, cutoff dates stayed the same between 1997
and 1998, and any seasonal patterns in number of births are repeated
every year, then we may use the number of children in each birth cohort
to approximate the percentage of children in each observed relative age
group who enter early, when first eligible, or after redshirting. The
numbers of observations in each relative age group are presented in ta-
ble 3, and table 4 shows these estimates aggregated to the level of quar-
ters. Table 5 shows that assigned relative age is by far the strongest pre-
dictor of entering late.
TABLE 3
Entering Rates by Month in the ECLS-K ð%Þ
Month before Cutoff Turned 5
Entering 12 11 10 9 8 7 6 5 4 3 2 1
Early 6 2 1 0 0 0 0 0 0 0 0 0
On time 93 97 98 98 97 97 96 93 90 87 81 74
Late 1 1 1 1 3 3 4 6 11 14 19 26

TABLE 4
Entering Rates by Quarter in the ECLS-K ð%Þ
Quarter before Cutoff Turned 5
Entering 4 3 2 1
Early 3 0 0 0
On time 96 97 93 80
Late 1 3 7 20
The data presented in table 3 and figure 6 guide the selection of a sam-
ple relatively unaffected by redshirting. Note that over a quarter of chil-
dren who turned 5 within 1 month of their school’s cutoff date, 26 per-
cent, were redshirted. While this rate does decline as children become
relatively older, it is still 11 percent for children who turned 5 in the
fourth month before their school’s cutoff date. Also note that a negli-
gible number of children enter early, except for 6 percent in the oldest
assigned relative age group. Finally, turning our attention to children
who turned 5 between 6 and 11 months before their school’s entrance
cutoff date, over 96 percent of each cohort entered kindergarten when
first eligible.
The sample used for the remaining analysis in this paper is composed
of those kindergartners in the observed relative age groups M 6 through
TABLE 5
Entering Rates by Demographic Characteristics in the ECLS-K
Percent
Entering Late
Overall 8
Race:
Black 5
White 10
Hispanic 6
Asian 3
Gender:
Male 10
Female 5
Home environment:
Many books ð≥100Þ 9
Few books ð<30Þ 7
Father in household 8
No father in household 6
Foreign language spoken at home 4
No foreign language spoken at home 8
Socioeconomic status:
High SES 8
Low SES 7
Mother:
High school dropout 7
BA holder 9
Employed 7
Not employed 8

Figure 6.—Initial enrollment decisions ðby assigned relative age cohorts BjÞ
M11 because they are the least affected by redshirting. Furthermore, the
sample is restricted to children living in states that set an entrance cut-
off date between August 31 and January 1. Children from states with local
education authority options ðColorado, Massachusetts, New Hampshire,
New Jersey, New York, Pennsylvania, and VermontÞ, from the three states
with entrance cutoff dates before August 31 ðAlaska, Indiana, and Mis-
souriÞ, and from states with fewer than 30 respondents ðArkansas, Dela-
ware, District of Columbia, Hawaii, Idaho, Montana, Nebraska, Nevada,
North Dakota, Oregon, South Carolina, South Dakota, West Virginia, and
WyomingÞ are omitted from our sample. The 27 states included in our
sample are Alabama, Arizona, California, Connecticut, Florida, Georgia,
Illinois, Iowa, Kansas, Kentucky, Louisiana, Maine, Maryland, Michi-
gan, Minnesota, Mississippi, New Mexico, North Carolina, Ohio, Okla-
homa, Rhode Island, Tennessee, Texas, Utah, Virginia, Washington, and
Wisconsin.
This analysis will not consider outcomes after the third grade because
of the severe attrition in the ECLS-K data set. Tables 6 and 7 show the
sample sizes for math and reading IRT test scores by survey round. The
tables reveal heavy attrition in grades 5 and 8 in the ECLS-K. In our
sample, 41 percent of initial respondents do not have reading IRT scores
by the fifth grade, and 51 percent attrite by the eighth grade. The prob-
lem is even worse for math IRT scores.
Figures 7 and 8 show the distributions of math and reading IRT test
scores of children in the sample. These figures show how these distribu-
tions changed for all children between the fall of kindergarten and the

TABLE 6
Attrition in the ECLS-K: Math IRT Score Present
Sample All ECLS-K
Round Attrition ð%Þ N Attrition ð%Þ N
Fall kindergarten 0 3,950 5 18,640
Spring kindergarten 3 3,850 0 19,650
Fall 1st grade ... ... ... ...
Spring 1st grade 17 3,270 15 16,630
Spring 3rd grade 28 2,830 27 14,370
Spring 5th grade 44 2,230 43 11,270
Spring 8th grade 54 1,830 53 9,290
Note.—Fall 1st grade is not used in the analysis since this wave sampled only about 30 per-
cent of the entire sample.
spring of third grade. Math scores are more smoothly distributed than
reading scores.
C. Identification Strategy: Education Production Function Specification

1. The General Model of Educational Achievement
Given the sample free from redshirting just defined, I now proceed to
specify EPFs to be estimated on that sample. Since I am interested in very
general age effects, I specify that at each point in time t, individuals possess
a whole vector of ages A iðtÞ, each element of which affects their test scores.
Define that vector to be
Ai ðtÞ ; ½aðtÞ; EAi ; RAi ;
where aðtÞ represents the age of the youngest first-time eligible child in
the sample at time t, EAi is child i’s entrance age, and RAi is child i’s rela-
tive age as defined in Section II.C. Note that while aðtÞ, which can be
thought to represent the age at test, increases over time, EAi and RAi stay
fixed.
I now specify a generalization of equation ð3Þ from Todd and Wolpin
ð2003Þ that allows for more general age effects. Assume that the test score
TABLE 7
Attrition in the ECLS-K: Reading IRT Score Present
Sample All ECLS-K
Round Attrition ð%Þ N Attrition ð%Þ N
Fall kindergarten 0 3,740 7 17,620
Spring kindergarten 1 3,710 0 18,940
Fall 1st grade ... ... ... ...
Spring 1st grade 14 3,230 14 16,340
Spring 3rd grade 25 2,820 25 14,280
Spring 5th grade 41 2,220 41 11,270
Spring 8th grade 51 1,820 51 9,230

Figure 7.—The distributions of math IRT test scores
Figure 8.—The distributions of reading IRT test scores

of student i in school j at age A iðtÞ is a function of the entire history of

family and school inputs up until time t ðFij½AiðtÞ and Sij½AiðtÞÞ, innate
ability ðmij0 Þ, and an unobserved factor that is independent of i, j, t, and
other inputs ðeijt Þ:
←
Yijt 5 YAðtÞ fFij ½Ai ðtÞ; Sij ½Ai ðtÞ; mij0 ; eijt g: ð10Þ
Equation ð10Þ is specified to allow for certain age-related features

not accommodated by standard specifications. To begin, equation ð10Þ
does not assume that test scores are measured at the same age for all in-
dividuals but rather at the same point in time. We observe test scores
measured at various times t determined by the grade in which a student
is enrolled ðbetween fall of kindergarten and spring of third gradeÞ, not
his calendar age. This fact, together with variation in birth dates and en-
trance cutoff dates, generates variation in AiðtÞ.
Relatedly, both Fij½AiðtÞ and Sij½AiðtÞ are functions of AiðtÞ and not t
alone because the age at which these inputs are applied is likely to mat-
ter. For example, in the empirical specifications used in the analysis, ele-
ments of AiðtÞ are interpreted directly as elements of Sij½AiðtÞ since a
child’s own biological development process is influenced by that of other
children through direct social interactions between children, as well as
indirectly through social interactions with teachers and the broader
school environment. All these social interactions in the school are likely
to depend on the entrance and relative ages of the child.
As examples, consider two mechanisms through which relative age could
affect school inputs. One such mechanism has led to concerns about read-
iness. As a child becomes older relative to other children in his classroom,
he may receive additional inputs at school, perhaps in the form of addi-
tional attention or encouragement from his teacher or from tracking
with a higher-achieving peer group. Stimulated from the extra input,
the child’s achievement goes up. This mechanism could also work in
reverse for children receiving fewer inputs as a result of being relatively
younger than their classmates. If this mechanism were to dominate, we
would expect relatively older children to perform better on achievement
tests.
A second mechanism through which relative age might serve as a school
input is also related to peer effects and social interactions and could help
explain why children are not repeatedly redshirted. If a teacher teaches
to the average level of achievement in his or her classroom, school in-
puts will be tailored toward the average age in a classroom. Inputs aimed
at higher-achieving peers could cause children to accelerate their learn-
ing to keep pace with their peers, raising their test scores.11 In this case,
11
Alternatively, it might be the case that if children cannot keep pace, such inputs might
also leave children behind, decreasing their test scores.

we would expect relatively younger children to perform better on achieve-

ment tests.
There are also clear mechanisms through which entrance age might
affect family or school inputs. For example, entrance age might affect out-
comes independently of relative age or in combination with relative age.
Suppose that the biological development process were precisely deter-
mined by biological age. In this case, entrance age would have a large
impact on how children responded to school inputs, regardless of the
relative age of their classmates. Entrance age could also be interpreted as
an element of Fij½A iðtÞ since it governs the intensity of the exposure to fam-
ily inputs by determining the amount of time spent with the family rather
than at school.
Finally, mij0 is interpreted as genetic endowment, or ability at concep-
tion, which is assumed to be fixed over time. The assumption that mij0 is
fixed over time may seem severe, but the AðtÞ subscripting on the produc-
tion function allows for mij0 to have age-varying impacts on achievement
reflecting changing efficiency of utilization. Thus the specified produc-
tion function that includes a fixed ability mij0 is observationally equivalent
to allowing ability to change over time.
2. Empirical Specification and Identifying Assumptions

According to Todd and Wolpin ð2003Þ, the empirical implementation of
equation ð10Þ has floundered on two basic problems.
Problem i. mij0 is unobserved.

Problem ii. Data sets typically do not include the entire vectors of
Fij½AiðtÞ and Sij½AiðtÞ.
This analysis imposes four key additional assumptions in order to over-

come problems i and ii. Problem ii is overcome in this analysis by estimat-
ing a tractable ði.e., linear and separableÞ specification of equation ð10Þ
that includes variables measuring lagged inputs. Problem i is overcome
in this analysis by imposing slightly stronger assumptions. The variation in
the fall of kindergarten test scores is assumed to be generated by varia-
tion in family inputs and ability. Assuming that race is distributed inde-
pendently of ability, race indicators can be used to instrument for fam-
ily inputs, with the variation in test scores remaining after controlling
for family inputs is assumed to be due to ability mij0.
To be precise, throughout the remainder of the analysis the follow-
ing assumptions are maintained.
MA1. The production function for educational achievement at age
AðtÞ is a linear and separable function of family inputs, school inputs, abil-
ity, and an identically and independently distributed unobserved factor:
←
Yijt 5 F ij ½Ai ðtÞaaðtÞ 1 S ij ½Ai ðtÞbaðtÞ 1 daðtÞ mij0 1 eijt : ð11Þ

Furthermore, it is assumed that in the fall of kindergarten ðt 5 1Þ, school

inputs have not yet affected the production of educational achieve-
ment:12
←
Yij1 5 F ij ½Ai ð1Þaað1Þ 1 z að1Þ EAi 1 dað1Þ mij0 1 eij1 : ð12Þ
MA2. Family and school inputs are unchanging over time, so that in-
put measures at the fall of kindergarten capture the entire history of in-
puts. The specific implementation of MA2 estimated here can be stated
as assuming Fij½AiðtÞ 5 Fij for all t > 1 and that school inputs are en-
trance age, relative age, and a vector of school dummies:
S ij ½Ai ðtÞ ; ½EAi ; RAi ; EAi RAi ; Sij4 ; : : : ; Sijk :
Assumptions MA1 and MA2 are maintained to overcome problem ii.

Under MA1 and MA2, equation ð10Þ can be specified as
←
Yijt 5 Fij aaðtÞ 1 baðtÞ;1 EAi 1 baðtÞ;2 R Ai 1 baðtÞ;3 EAi R Ai
ð13Þ
1 baðtÞ;4 Sij4 1 1 baðtÞ;k Sijk 1 daðtÞ mij0 1 eijt :
A natural question related to MA2 is whether input measures at the fall

of kindergarten actually capture the entire history of inputs at dates rang-
ing from kindergarten until the spring of third grade. A first issue is
whether these measures capture inputs between conception and the time
at which they were measured. All the home input variables included in
estimation can be thought of as measures of inputs between concep-
tion and the time at which they were measured. The home inputs included
in estimation are income, mother’s education, birth weight, whether a
foreign language is spoken at home, whether the child was enrolled in a
prekindergarten program, mother’s employment between birth and kin-
dergarten, whether the father was living in the household, and whether
the child ever received WIC benefits.
A separate measurement issue related to MA2 is whether it is valid to
use variables measured in the fall of kindergarten as inputs until the spring
of third grade. One reason for doing this is that school fixed effects are
used as school inputs. As a way to deal with mobility and to increase power,
it is assumed that the school fixed effect measured in the fall of kinder-
garten represents those schools attended afterward. Other school inputs
include entrance age, relative age, the interaction of the two ages, state
fixed effects, gender, and race. Gender and race are interpreted as school
inputs since they determine the types of social interactions a child will
have at school, although they could also capture components of home
inputs not measured by the home input variables used in estimation.
Since kindergarten to third grade is a short time horizon over which most
12
Note that here the coefficient on entrance age has changed to a z , indicating that effects
from entrance age at this point in time represent effects from the biological development
process in isolation from any school effects.

of the input variables will vary little, assumption MA2 is not considered
to be a strong assumption.
As implemented, assumption MA2 interprets entrance age and rela-
tive ages as school inputs. The expression aaðtÞ;0 1 baðtÞ;0 sets the level for
the cohort at each point in time based on the overall age of the cohort,
aðtÞ. The deviation from that average level based on the additional en-
trance age of the child is determined by baðtÞ;1, and the deviation due to
the relative age of the child’s classmates is determined by baðtÞ;2. Thus aaðtÞ;0
1 baðtÞ;0 and baðtÞ;1 represent the average growth in achievement due sim-
ply to the biological development process. The combination of baðtÞ;1 , baðtÞ;2 ,
and baðtÞ;3 determines the precise way the child’s own biological develop-
ment process is influenced by that of other children through direct so-
cial interactions between children, as well as indirectly through social
interactions with teachers and the broader school environment. That is,
EAi and RAi can be interpreted as school inputs because they determine
the age at which one is exposed to educational instruction in school and
also help to determine the type of classroom and teacher interactions
students will have as a result of their own development and their develop-
ment relative to that of their classmates. Finally, it is also worth noting
that EAi will determine the amount of time spent in the home environ-
ment rather than school, so its effect parameters are likely to be sum-
maries representative of heterogeneous individual circumstances.
Two further assumptions are maintained in order to overcome prob-
lem i and obtain estimates of ability mij0. With ability measures in our pos-
session, equation ð13Þ should satisfy the standard OLS assumptions and
therefore be straightforward to estimate. To be precise, the additional
maintained assumptions are as follows.
MA3. Define a set of race/ethnicity dummies Ri ; ðRi1 ; Ri2 ; Ri3 Þ, where
Ri is a binary indicator for being black, Ri2 is an indicator for being His-
1
panic, and Ri3 is an indicator for being Asian. It is assumed that race/
ethnicity is correlated with family inputs: E½Rik Fij ≠ 0 for at least one k
for each element Fij of Fij.
MA4. For each subject, unobserved factors are identically and in-
dependently distributed, eijt ∼ iidN ð0; j2et Þ, with ability being normally dis-
tributed ðmij0 ∼ N ð0; j2m ÞÞ independently of race: mij0 j j R k for all k.
i
Estimates of ability mij0 are obtained by recalling that under MA1, fall
of kindergarten ðt 5 1Þ school inputs have not yet affected the produc-
tion of educational achievement:
←
Yij1 5 F ij ½Ai ð1Þaað1Þ 1 z að1Þ EAi 1 dað1Þ mij0 1 eij1 : ð14Þ
Owing to a limited number of instruments, the t 5 1 vector Fij½Aið1Þ is as-

sumed to include only household income, mother’s educational attain-
ment, and birth weight. It is assumed that these variables capture all
home inputs between conception and fall of kindergarten. Under these
assumptions, equation ð14Þ can be rewritten as

←
Yij1 5 F ij aað1Þ 1 z að1Þ EAi 1 hij1 ; ð15Þ
where
hij1 ; dað1Þ mij0 1 eij1 :
Equation ð15Þ can be estimated via two-stage least squares ð2SLSÞ using
the three race dummies in R i as instruments for the three endogenous
variables in Fij. The residuals obtained from this estimation are
←
hij1 5 dað1Þ mij0 1 eij1 : ð16Þ
If Rik is a race dummy, then MA3 and MA4 imply the standard IV
assumptions.
MA3. E½Rik Fij ≠ 0 for at least one k for each element Fij of Fij.
MA4*. E½Rik hij1 5 0 for all k.
Assumption MA3 is clearly true; the first-stage regressions of inputs

on race dummies have F-statistics ranging between 20 and 89. Assump-
tion MA4* follows from MA4 and requires the data-generating process to
have two important features. First, household income, mother’s educa-
tional attainment, and birth weight are the only home inputs to the
production of educational achievement before entering school. This as-
sumption is clearly not true in a strict sense and should be relaxed if
possible in future analyses. Nevertheless, it facilitates the empirical analy-
sis and is likely to be a reasonable approximation to reality. The other
requirement is that there are no inherent racial differences in innate
ability, which itself is likely to be untestable because of the fundamental
identification issue in determining the causes of between-group differ-
ence in outcomes like IQ or other “ability” measures ðGoldberger and
Manski 1995Þ. The subsequent empirical results are all conditional on
MA4 holding.13
After obtaining estimates ^hij1 5 daðtÞ m
îj0 1 êij1 from the test scores at time
t 5 1 ðfall of kindergartenÞ, equation ð13Þ is estimated by OLS under the
following specification for the remaining points in time t > 1 ðspring of
kindergarten, spring of first grade, and spring of third gradeÞ:
←
Yijt 5 Fij aaðtÞ 1 baðtÞ;1 EAi 1 baðtÞ;2 R Ai 1 baðtÞ;3 EAi RAi
ð17Þ
1 baðtÞ;4 Sij4 1 1 baðtÞ;k Sijk 1 baðtÞ;m ½daðtÞ m
îj0 1 êij1 1 eijt :
13
Although the extensive debate over assumption MA4 is best left for other studies
such as Goldberger and Manski ð1995Þ, here it is at least worth mentioning two points
about MA4. One is that first-order mechanisms such as the history of racial discrimina-
tion ðBlackmon 2008; Aliprantis and Carroll 2013Þ and secular changes in the labor mar-
ket ðWilson 1987Þ can explain important between-group differences in outcomes under
MA4. Another is that the violation of MA4 is an idea that has been used to justify malicious
behavior and policies ðsee Zuberi ½2001 or Washington ½2006, chap. 6 for important
historical examplesÞ as well as policy fallacies ðManski 2011Þ.

The OLS estimation of equation ð17Þ brings two issues to mind. First,
the random assignment of entrance age ðEAÞ and relative age ðRAÞ does
not ensure that other inputs are not chosen in response to these vari-
ables. Aliprantis ð2014bÞ shows that ðbÔLS ; b ÔLS ; b ÔLS Þ do identify the EPF
aðtÞ;1 aðtÞ;2 aðtÞ;3
parameters of equation ð17Þ as long as any other input chosen in response
to entrance or relative ages is included as an observed input in estima-
tion.14 Second, attenuation bias is a concern because our measure of ability
includes measurement error ð^hij1 5 daðtÞ m îj0 1 êij1 Þ. This is a general prob-
lem when using identification strategies that rely on noisy measures. Since
this issue is only beginning to receive attention for value-added models
ðKoedel, Leatherman, and Parsons 2012Þ, fully accommodating this issue
is left for future research.
D. Identification Strategy: Identifying Variation

Figure 3 helps to illustrate the variation in the sample used to identify
the parameters of the specified education production functions. Points a,
b, c, and d represent hypothetical combinations of birth dates and en-
trance cutoff dates a child could be facing. Counterfactually moving a
child from one point to another represents either varying her birth
date, the entrance cutoff date she faces, or both. For example, moving
from point b to a is the same as being born at an earlier date, all else
being constant. We can see in table 8 that this increases a child’s en-
trance age EA as well as her relative age RA. Thus, date of birth alone is
not enough to separately identify the effects of these variables.
Moving from point b to d is the same as moving a child to a state with
a different entrance cutoff date, all else being constant. In this case the
child will have the same entrance age but will be a different age relative
to her classmates. Thus variation in state entrance cutoff dates identifies
relative age effects, fbaðtÞ;2 g.
Now consider moving a child from point a to d. This is the same as
moving the child’s birth date to a later date but also moving the child to
a state whose entrance cutoff date is earlier by the same amount of time.
In this case, the child will be the same age relative to her classmates, but
she will enter kindergarten at a younger age. Thus variation in birth date,
together with variation in state entrance cutoff dates, identifies the ef-
fects of entrance age, fbaðtÞ;1 g.
Although entrance age and relative age are linearly dependent within
individuals, this is not the case between individuals. The variation in these
two ages generated by differences in date of birth across individuals and
entrance cutoff dates across states is shown in scatterplots in figures 9
and 10. Figure 9 shows the variation in the entire ECLS-K in compari-
son to the variation in the sample as shown in figure 10. Note that owing
←
14
The structural equation notation ð5Þ is used to clarify that the EPF parameters of
eq. ð17Þ correspond to the direct effects discussed in Aliprantis ð2014bÞ.

TABLE 8
Entrance and Relative Ages as a Function
of Birth Date and State Cutoff
Birth Date/State Entrance Relative
Combination Age Age
a 8 8
b 6 6
c 8 10
d 6 8
to the sample selection conducted to ensure that these ages vary ran-
domly, we observe only children between the relative ages of 5 and
11 months. It must be remembered that the subsequent analysis can
speak only to this sample of children and is uninformative about age
effects on children starting kindergarten at other relative or absolute ages.
If the identification scheme just described is to provide unbiased esti-
mates of the entrance and relative age effects using the variation in ages
shown in figure 10, date of birth must be random conditional on observ-
ables ðfamily inputs, school inputs, school fixed effects, and state fixed
effectsÞ. However, considerable evidence indicates that date of birth is
not random but rather is the product of parents’ choices, biological fac-
tors, or both. Seasonal birth patterns have been demonstrated to be re-
lated to maternal characteristics such as age, marital status, education,
and birth order ðBobak and Gjonca 2001; Buckles and Hungerman 2008Þ;
Figure 9.—Variation in entrance age and relative age in the entire ECLS-K

Figure 10.—Variation in entrance age and relative age in the sample
demographic characteristics such as income and race ðBound and Jaeger

2000Þ; and even tax schedules ðDickert-Conlin and Chandra 1999Þ. It is
currently unclear how these socioeconomic characteristics interact with
biological factors ðRizzi and Dalla-Zuanna 2007Þ, temperature ðLam and
Miron 1996Þ, and geographic location ðBobak and Gjonca 2001Þ in de-
termining seasonal birth patterns. Although this process has been for-
mally modeled ðLam, Miron, and Riley 1994Þ, the parameters of such a
model of seasonal birth patterns have yet to be identified. This leaves
seasonality of births as a serious concern for our identification strategy.
Most studies of the effects of entrance age have dealt with concerns
related to seasonal birth patterns by simply assuming that date of birth is
exogenous. In support of this approach, Dickert-Conlin and Elder ð2009Þ
present compelling evidence that while birth date is clearly manipulated
within short windows for nonmedical purposes, this does not appear to
be the case in the windows around school entrance cutoff dates. Similar
evidence is found in McEwan and Shapiro ð2008Þ. However, these stud-
ies focus on the manipulation of birth timing at the level of days or weeks
through cesarean deliveries and the inducement of labor. Hence, it is
still possible that birth timing is planned around a less precise period such
as a month or a quarter, leaving the identification strategy presented
here open to the problems arising if parents make nonrandom choices,
whether consciously or unconsciously, about when to have children.
I examine some evidence from the ECLS-K related to the seasonality
of birth in table 9. This table presents results of F-tests on null hypotheses

TABLE 9
p -Values of F -Tests of Equality of Characteristic Means by Quarter and Month
All Children Sample
Characteristic Quarter Month Quarter Month
Race:
Black .98 .07 .00 .00
White .41 .42 .00 .00
Hispanic .44 .79 .00 .00
Asian .38 .91 .03 .22
Gender:
Female .71 .18 .89 .92
Mom’s education level:
Mom’s highest grade completed <12 .03 .31 .73 .09
Mom’s highest degree received
5 high school diploma .42 .61 .33 .64
Mom’s highest degree received ≥ BA .81 .82 .63 .40
Home characteristics:
Number of books at home .30 .31 .99 .52
Live with father .65 .91 .00 .04
Birth characteristics:
Parents married at birth .18 .37 .62 .26
Birth weight .30 .07 .11 .12
SES 5 1 .03 .08 .47 .11
SES 5 2 .54 .69 .75 .15
SES 5 3 .96 .73 .39 .48
SES 5 4 .64 .91 .78 .19
SES 5 5 .14 .59 .08 .11
of equal means across seasonality of birth for several characteristics.

Apart from race and living with one’s father, there does not appear to be
strong evidence of birth seasonality in our sample. In a comparison of the
means in tables 10, 11, 12, and 13, some of the differences in the sample
seem to be driven mechanically by sampling error and the fact that over
90 percent of children in the sample were born between October and
April. Thus it may be reasonable to follow convention and assume that
date of birth is exogenous, while at the same time acknowledging that
the current analysis will be greatly improved by further research into sea-
sonal birth patterns.
In addition to seasonal birth patterns, the endogeneity of state entrance
cutoff dates is also a concern for our identification strategy, and we use
2000 census data from the National Historical Geographic Information
System ðNHGIS; Minnesota Population Center 2004Þ to investigate this
issue. Although the EPFs we estimate all include state and school fixed
effects, unobserved characteristics within states with particular entrance
cutoff dates could still be driving results. Strong correlations between en-
trance cutoff dates and demographic characteristics would increase this
concern. Figures 11, 12, 13, and 14 ease this concern by showing that the
correlations between entrance cutoff date and state demographic

TABLE 10
Seasonality of Characteristics: Means and Differences
from Reference Group by Quarter
All Children
Quarter 1 Quarter 2 Quarter 3 Quarter 4
Characteristic ðMeanÞ ðDiff.Þ ðDiff.Þ ðDiff.Þ
Race:
Black .16 2.00 2.00 2.00
White .56 .02 .00 .01
Hispanic .20 2.01 .00 2.01
Asian .04 2.00 2.01 .00
Gender:
Female .49 2.01 2.01 2.01
Mom’s highest grade completed <12 .14 2.00 .01 .02
Mom’s highest degree received 5
high school diploma .64 .00 2.00 2.01
Mom’s highest degree received ≥ BA .22 2.00 2.01 2.00
Number of books at home 71.82 1.94 .89 2.26
Birth weight ðpoundsÞ 7.35 .02 .05 .05
SES 5 1 .20 2.01 .01 .01
SES 5 2 .20 .01 2.01 2.00
SES 5 3 .20 2.00 .00 .00
SES 5 4 .20 2.00 2.01 .00
SES 5 5 .20 .01 2.00 2.01
characteristics such as the state poverty rate, male labor force participa-
tion rate, and high school graduation rate are small and statistically
insignificant.
E. Estimation Results
Before estimating the EPF in equation ð17Þ, I first obtain estimates of
math and reading ability by estimating the fall of kindergarten EPF in
equation ð15Þ. I assume that there are four inputs into the production
of achievement in the fall of kindergarten: age, mother’s educational
attainment, household income, and prenatal inputs as measured by birth
weight. Assuming that race is independent of ability but correlated with
the fall test scores through the family inputs, I use race indicators as in-
struments for family inputs. The specification imposes that the error term
from the 2SLS estimation of equation ð15Þ is composed only of ability
and measurement error.
The ^hij1 estimates from equation ð15Þ are shown in figures 15, 16, 17,
and 18. Overall, the ability estimates appear not only to have the same
mean as imposed by the IV estimation ðMA4*Þ but also to have similar

TABLE 11
Seasonality of Characteristics: Means and Differences from Reference Group by Month
All Children
Jan. Feb. March April May June July Aug. Sept. Oct. Nov. Dec.
Characteristic ðMeanÞ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ
Race:
Black .17 .00 2.03 2.00 2.01 2.03 2.01 2.01 2.01 2.02 2.01 2.00
White .56 2.01 .03 .01 .03 .03 .01 .01 .01 .02 .02 .01
Hispanic .19 .01 2.00 2.00 2.02 .01 .00 .01 .01 .00 2.00 2.01
Asian .04 2.00 2.00 .00 2.00 2.01 2.01 2.00 2.00 2.00 .00 .00
Gender:
Female .49 .01 2.01 .01 2.02 2.02 .01 2.01 2.02 2.03 2.00 .01
Mom’s highest grade completed <12 .15 2.00 2.01 2.01 2.01 .00 .01 .00 .01 .01 .02 .01
high school diploma .64 2.01 2.01 .01 2.00 2.02 2.02 2.00 2.01 2.02 2.03 2.02
Mom’s highest degree received ≥ BA .21 .02 .02 2.01 .01 .02 .01 .00 2.00 .01 .01 .00
Number of books at home 69.07 1.97 5.96 4.92 5.30 3.90 3.95 4.35 2.58 2.73 3.11 1.67
Live with father .76 .02 .01 .02 .03 .02 .02 .02 .02 .03 .01 .01
Parents married at birth .67 .03 .02 .04 .03 .05 .04 .02 .05 .03 .02 .03
Birth weight ðpoundsÞ 7.45 2.17 2.12 2.04 2.07 2.11 2.09 2.06 .01 2.03 2.07 2.05
SES 5 1 .21 2.01 2.03 2.01 2.04 2.02 2.00 2.01 .00 2.01 .00 2.00
SES 5 2 .20 2.01 .01 .00 .01 .01 2.01 .00 2.01 2.01 .01 2.01
SES 5 3 .19 .02 .01 .00 .02 .01 2.00 .02 .02 .02 .01 .02
SES 5 4 .20 2.01 .01 .00 2.00 .00 .00 2.01 2.01 .01 .00 .00
SES 5 5 .20 .00 2.00 .01 .01 .01 .01 2.01 2.01 2.00 2.02 2.01

TABLE 12
Seasonality of Characteristics: Means and Differences from
Reference Group by Quarter
Sample
Quarter 1 Quarter 2 Quarter 3 Quarter 4
Characteristic ðMeanÞ ðDiff.Þ ðDiff.Þ ðDiff.Þ
Race:
Black .17 2.07 2.13 .00
White .58 2.03 2.01 .05
Hispanic .19 .10 .03 2.04
Asian .03 .02 2.00 2.01
Gender:
Female .49 2.01 .07 .00
Mom’s highest grade completed <12 .13 .01 2.01 .01
high school diploma .66 2.03 2.03 2.03
Mom’s highest degree received ≥ BA .22 .02 .04 .02
Number of books at home 72.56 .66 2.57 2.38
Birth weight 7.29 .15 .38 .05
SES 5 1 .17 .01 2.10 .01
SES 5 2 .20 .00 .03 .01
SES 5 3 .21 2.03 2.06 2.02
SES 5 4 .23 2.02 2.03 2.01
SES 5 5 .20 .03 .17 2.00
overall distributions as originally assumed in MA4. The deviations from

identical distributions could be explained by immigrants speaking a lan-
guage in the home other than English. When asked whether the home
language of the child is English or non-English, only 1 percent of whites
and blacks say non-English. This number jumps to 46 percent for Hispan-
ics and 63 percent for Asians.
Using the ^hij1 as ability estimates allows us to estimate the EPFs in
equation ð17Þ. The family inputs included in estimation are household
income, mother’s education, birth weight, whether a foreign language
is spoken at home, whether enrolled in a prekindergarten program,
mother’s employment between birth and kindergarten, whether the fa-
ther was living in the household, and whether the child ever received
WIC benefits. School inputs include observed entrance age, observed
relative age, the interaction of the two ages, school and state fixed effects,
gender, and race.
Estimates of the entrance age and relative age parameters in equation
ð17Þ are displayed in tables 14 and 15. The first impression we might take
away from the estimates in these tables is that entrance age parameters are
positive between kindergarten and third grade in both math and reading.

TABLE 13
Seasonality of Characteristics: Means and Differences from Reference Group by Month
Sample
Jan. Feb. March April May June July Aug. Sept. Oct. Nov. Dec.
Characteristic ðMeanÞ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ ðDiff.Þ
Race:
Black .17 .03 2.02 2.05 2.08 2.10 2.12 .00 2.17 2.01 2.01 .01
White .58 2.02 .01 .06 2.14 2.12 .02 .00 2.58 .02 .06 .05
Hispanic .19 2.01 .00 2.01 .19 .23 2.01 .00 .61 2.02 2.04 2.06
Asian .04 2.01 2.01 .00 .02 .01 2.02 .00 .16 2.01 2.01 2.02
Gender:
Female .48 .01 .00 .01 .01 2.05 .05 .00 .52 2.02 .01 .02
Mom’s highest grade completed <12 .15 2.04 2.03 2.04 .02 2.02 2.07 .00 .65 2.02 2.00 2.02
514
high school diploma .67 2.00 2.02 2.01 2.07 2.06 2.02 .00 2.47 2.04 2.04 2.03
Mom’s highest degree received ≥ BA .18 .05 .06 .04 .05 .08 .09 .00 2.18 .05 .05 .06
Number of books at home 68.51 5.27 6.55 8.06 22.34 4.16 5.52 .00 234.91 1.06 4.79 4.24
Live with father .79 .01 2.02 .04 .06 .08 .15 .00 .21 .03 2.00 2.01
Parents married at birth .67 .05 .03 .05 .03 2.01 .15 .00 2.47 .01 .08 .03
Birth weight 7.41 2.19 2.16 .05 2.07 .10 .35 .00 21.41 2.15 2.00 2.08
SES 5 1 .21 2.06 2.05 2.03 .00 2.05 2.13 .00 2.21 2.05 2.02 2.03
SES 5 2 .20 2.02 .01 .03 2.05 2.03 2.02 .00 .80 2.00 .03 2.00
SES 5 3 .21 .02 2.02 2.02 2.05 2.01 2.06 .00 2.21 2.03 2.03 2.00
SES 5 4 .21 .01 .03 2.04 .07 .01 2.01 .00 2.21 .04 2.01 .00
SES 5 5 .17 .05 .04 .06 .02 .08 .22 .00 2.17 .04 .03 .03

Figure 11.—State poverty and 1998 kindergarten entrance cutoff dates. State-level data are
from the US Census/NHGIS, and entrance cutoff dates are taken from Elder and Lubotsky
ð2009Þ.
Figure 12.—State male labor force participation and 1998 kindergarten entrance cutoff
dates. State-level data are from the US Census/NHGIS, and entrance cutoff dates are taken
from Elder and Lubotsky ð2009Þ.
515

Figure 13.—State female high school graduation rate and 1998 kindergarten entrance cut-
off dates. State-level data are from the US Census/NHGIS, and entrance cutoff dates are
taken from Elder and Lubotsky ð2009Þ.
Figure 14.—State male high school graduation rate and 1998 kindergarten entrance cutoff
dates. State-level data are from the US Census/NHGIS, and entrance cutoff dates are taken
from Elder and Lubotsky ð2009Þ.
516

Figure 15.—Math ^
h estimates
Figure 16.—Math ^
h estimates by race
517

Figure 17.—Reading ^
h estimates
Figure 18.—Reading ^
h estimates by race

TABLE 14
Math Test Score Parameter Estimates for All Children
Spring
Kindergarten Spring 1st Grade Spring 3rd Grade
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb 2.10 .18 5.99 .33 3.91 .16
ð.5Þ** ð.78Þ** ð1.36Þ**
âðtÞ;2 Þ
Relative age ðb 2.50 2.04 24.77 2.26 .58 .02
ð.38Þ ð.83Þ** ð1.14Þ
âðtÞ;3 Þ
Interaction ðb 2.05 2.00 2.01 2.00 2.26 2.01
ð.05Þ ð.1Þ ð.14Þ
Observations 2,800 2,730 2,350
Note.— Outcomes are IRT test scores. Home inputs included in estimation are income,
mother’s education, whether a foreign language is spoken at home, whether enrolled in a
prekindergarten program, mother’s employment between birth and kindergarten, whether
the father was living in the household, and whether the child ever received WIC benefits.
School inputs include entrance age, relative age, the interaction of the two ages, school and
state fixed effects, gender, and race. Ability estimates as described in Secs. III.C and III.E are
also included in estimation.
** p < .01.
All else equal, in this time period, 1 month of extra entrance age increases
math IRT scores between 0.16 and 0.33 standard deviations and reading
IRT scores between 0.19 and 0.28 standard deviations. In the spring of
third grade the entrance age coefficient is under 0:2j for both math and
reading test scores.
A second impression from tables 14 and 15 is that the parameters
pertaining to being relatively older than one’s classmates are quite
TABLE 15
Reading Test Score Parameter Estimates for All Children
Spring
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb 3.85 .28 5.05 .22 5.06 .19
ð.56Þ** ð1.18Þ** ð1.94Þ**
âðtÞ;2 Þ
Relative age ðb 23.10 2.23 22.89 2.12 22.43 2.09
ð.4Þ** ð1.15Þ* ð1.55Þ
âðtÞ;3 Þ
Interaction ðb 2.007 2.01 2.06 2.00 2.05 2.00
ð.06Þ ð.12Þ ð.19Þ
Note.—Outcomes are IRT test scores. Home inputs included in estimation are income,
mother’s education, whether a foreign language is spoken at home, whether enrolled in a
prekindergarten program, mother’s employment between birth and kindergarten, whether
the father was living in the household, and whether the child ever received WIC benefits.
School inputs include entrance age, relative age, the interaction of the two ages, school and
state fixed effects, gender, and race. Ability estimates as described in Secs. III.C and III.E are
also included in estimation.
* p < .05.
** p < .01.

different from the parameters for entering at an older age. All else equal,
being 1 month relatively older tends to decrease IRT scores. For reading
these effects are large but fade over time, although for math these effects
are sometimes large and sometimes near zero. A final impression from
tables 14 and 15 is that the interaction effects are very small.
In light of the evidence of essential heterogeneity provided in Sec-
tion II, equation ð17Þ is also estimated by subgroup. Tables 16, 17, 18, and
19 show separate effect estimates for boys and girls. Tables 20, 21, 22, and
23 show similar estimates for children with different home environments
as measured by the number of children’s books they have at home.
The estimates indicate that entrance age effects are extremely different
for boys and girls. By the spring of third grade the entrance age parameter
for math scores is approximately 0:1j for girls, but for boys it is 0:5j!
Although the trends in earlier grades are different for reading, we see an
almost identical pattern in the entrance age effect by the spring of third
grade: increasing entrance age has very large and beneficial effects on
boys but relatively small impacts on girls.
Relative age effects tend to be negative or zero for both boys and girls.
Although many of these effects are not statistically different from zero, by
the spring of third grade, relative age effects are negative and large for
boys.
Finally, I investigate heterogeneity in effects by one measure of the
home environment, the number of children’s books the child has at home.
I define children to have few books if they have fewer than 35 books,
which is the 33rd percentile of the sample distribution, and I define chil-
dren to have many books at home if they have more than 100 books, the
66th percentile of the sample distribution.
Entrance age and relative age have huge effects on children with few
books at home. All else constant, for every month older they are when they
enter school, these children score 0.28 and 0.47 standard deviations worse
on their third grade math and reading tests, respectively. Relative age
effects have the opposite sign and in the spring of third grade have an
TABLE 16
Math Test Score Parameter Estimates for Boys
Spring
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb 1.24 .10 4.10 .23 12.27 .50
ð.76Þ ð1.32Þ** ð2.39Þ**
âðtÞ;2 Þ
Relative age ðb 1.65 .14 21.30 2.07 25.03 2.21
ð.74Þ* ð1.97Þ ð2.31Þ*
âðtÞ;3 Þ
Interaction ðb 2.14 2.01 2.13 2.01 2.41 2.02
ð.08Þ ð.18Þ ð.26Þ
* p < .05.
** p < .01.

TABLE 17
Math Test Score Parameter Estimates for Girls
Spring Spring 3rd
Kindergarten Spring 1st Grade Grade
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb .18 .02 1.28 .07 3.13 .13
ð1.00Þ ð1.46Þ ð2.54Þ
âðtÞ;2 Þ
Relative age ðb 1.55 .13 1.05 .06 .84 .03
ð.76Þ* ð1.32Þ ð1.71Þ
âðtÞ;3 Þ
Interaction ðb 2.05 2.00 2.08 2.00 2.24 2.01
ð.09Þ ð.15Þ ð.21Þ
* p < .05.
TABLE 18
Reading Test Score Parameter Estimates for Boys
Spring
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb 1.74 .13 21.03 2.04 12.94 .48
ð.95Þ ð1.93Þ ð2.87Þ**
âðtÞ;2 Þ
Relative age ðb .04 .00 5.54 .24 26.24 2.23
ð.75Þ ð2.35Þ* ð3.56Þ
âðtÞ;3 Þ
Interaction ðb 2.07 2.01 2.19 2.01 2.29 2.01
ð.09Þ ð.20Þ ð.37Þ
* p < .05.
** p < .01.
TABLE 19
Reading Test Score Parameter Estimates for Girls
Spring Spring 3rd
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb 3.00 .22 5.16 .22 4.08 .15
ð1.10Þ** ð2.00Þ* ð3.51Þ
âðtÞ;2 Þ
Relative age ðb 23.52 2.26 25.18 2.22 22.44 2.09
ð.97Þ** ð1.91Þ** ð2.77Þ
âðtÞ;3 Þ
Interaction ðb .07 .01 .06 .00 2.02 2.00
ð.10Þ ð.21Þ ð.31Þ
* p < .05.
** p < .01.

TABLE 20
Math Parameter Estimates with Few Books ð≤35Þ at Home
Spring
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb 22.03 2.17 27.48 2.41 26.92 2.28
ð1.02Þ* ð2.09Þ** ð4.21Þ
âðtÞ;2 Þ
Relative age ðb 2.23 .19 6.38 .35 8.33 .34
ð.84Þ** ð1.49Þ** ð2.75Þ**
âðtÞ;3 Þ
Interaction ðb .03 .00 .14 .01 2.04 2.00
ð.10Þ ð.21Þ ð.33Þ
Observations 810 800 670
* p < .05.
** p < .01.
TABLE 21
Math Parameter Estimates with Many Books ð≥100Þ at Home
Spring
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb 2.34 .20 8.47 .47 6.51 .27
ð1.09Þ* ð1.59Þ** ð2.47Þ**
âðtÞ;2 Þ
Relative age ðb 21.60 2.13 25.35 2.29 22.90 2.12
ð.75Þ* ð1.37Þ** ð2.14Þ
âðtÞ;3 Þ
Interaction ðb .009 .00 2.17 2.01 2.21 2.01
ð.10Þ ð.18Þ ð.28Þ
Observations 1,070 1,040 910
* p < .05.
** p < .01.
TABLE 22
Reading Parameter Estimates with Few Books ð≤35Þ at Home
Spring Spring 1st
Kindergarten Grade Spring 3rd Grade
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb 213.00 2.96 21.22 2.05 212.70 2.47
ð2.15Þ** ð3.53Þ ð5.68Þ*
âðtÞ;2 Þ
Relative age ðb 16.06 1.18 3.68 .16 16.34 .60
ð1.44Þ** ð2.74Þ ð4.54Þ**
âðtÞ;3 Þ
Interaction ðb 2.14 2.01 2.10 2.00 2.05 2.00
ð.15Þ ð.36Þ ð.50Þ
Observations 730 760 660
* p < .05.
** p < .01.

TABLE 23
Reading Parameter Estimates with Many Books ð≥100Þ at Home
Spring Spring 3rd
Raw j Raw j Raw j
âðtÞ;1 Þ
Entrance age ðb .65 .05 212.45 2.53 4.05 .15
ð.95Þ ð3.58Þ** ð3.70Þ
âðtÞ;2 Þ
Relative age ðb 2.75 2.06 12.24 .52 21.39 2.05
ð.79Þ ð2.47Þ** ð3.72Þ
âðtÞ;3 Þ
Interaction ðb .03 .00 .06 .00 2.09 2.00
ð.10Þ ð.27Þ ð.36Þ
Observations 1,070 1,040 910
** p < .01.
even larger magnitude for these children, at 0.34 and 0.60 standard
deviations for math and reading test scores.
Entrance age and relative age have smaller parameters for children with
many books at home, and these parameters tend to have the opposite sign
of those for children with few books at home. Entering 1 month older
increases math test scores in the spring of third grade, on average, by 0.27
standard deviations for children with many books, but entering relatively
older decreases test scores at this time by 0.12 standard deviations. The
effects on reading test scores are similar but of smaller magnitudes; the
entrance age parameter in spring of third grade is 0:15j and the relative
age parameter is 20:05j.
IV. Effects of Changing Entrance Cutoff Dates

One issue of interest for policy makers is understanding how states making
changes to their entrance cutoff dates would affect children. That is, we
might be interested in understanding how an earlier entrance cutoff date
would affect a child’s achievement. We can use the estimated EPF param-
eters to investigate the effects of this counterfactual entrance age policy
change.
It is important to note that changing the entrance cutoff date does not
change all children’s experiences in the same way. For those children
whose eligibility would not change, their entrance age would stay the same
ðthis assumes they are not redshirtedÞ. For those children whose eligibility
would change, their entrance age would change by 12 months. And
although this would change every child’s relative age, it would do so dif-
ferently for those who remain eligible versus those whose eligibility
changes.
Our EPF parameter estimates allow us to make inferences only about
those children who are in our sample, and these children’s eligibility
would not be changed by a 1-month change in their state’s entrance cutoff
date. It is important to remember that our sample is composed of only the

relatively oldest children in a given cohort. Furthermore, I cannot speak to

the overall effects of such an entrance age policy change because the EPF
parameters are not informative about the large, discontinuous changes
in entrance and relative ages experienced by children whose eligibility
would be altered. The counterfactual identified by a state-level identifi-
cation strategy using state-level data would speak to these effects, which
would be a combination of the effects identified in this section and the
effects of redshirting, as experienced by different subpopulations.
I assume that changing the entrance cutoff date will not change any
family inputs and will change only one school input, age relative to
classmates, RA, for children in the sample. Returning to the EPF in
equation ð17Þ, asking how an earlier entrance cutoff date would affect an
eligible child’s achievement is the same as asking how decreasing a child’s
relative age affects his achievement:
yYijt
2 5 2baðtÞ;2 2 baðtÞ;3 EA: ð18Þ
yRA
The relative age parameter baðtÞ;2 from the EPF determines the level of the
effect from changing the entrance cutoff date, and the interaction param-
eter bt3 determines how much this varies at different entrance ages.
When interpreting these results, it is useful to recall the mechanisms
discussed in Section III.C.1 through which relative age functions as a
school input. One such mechanism is that as a child becomes older
relative to other children in his classroom, the child may receive additional
inputs at school, perhaps in the form of additional attention or encour-
agement from his teacher or from tracking with a higher-achieving peer
group. Stimulated from the extra input, the child’s achievement goes up.
This mechanism could also work in reverse for children receiving fewer
inputs as a result of being relatively younger than their classmates. If this
mechanism were to dominate, we would expect relatively older children to
perform better on achievement tests or for there to be a negative effect
from decreasing a child’s relative age.
The second mechanism through which relative age might serve as a
school input is also related to peer effects and social interactions. If
teachers teach to the average level of achievement in their classroom,
school inputs will be tailored toward the average age in a classroom.
Inputs aimed at higher-achieving peers could cause children to accelerate
their learning to keep pace with their peers, raising their test scores.15 In
this case, we would expect the effect of decreasing relative age to have a
positive sign.
Table 24 shows the implied effects in the spring of third grade of an
earlier entrance cutoff date based on the EPF parameter estimates at
15
Alternatively, it might be the case that if children cannot keep pace, such inputs might
also leave children behind, decreasing their test scores.

TABLE 24
Counterfactual 1: Effects from Changing to an Earlier Entrance
Cutoff Date Spring of Third Grade
Math Reading
Sample Raw j Raw j
All children 1.21 .05 2.76 .11
ð.70Þ ð.94Þ
Gender:
Boys 7.94 .33 8.26 .34
ð1.47Þ ð1.74Þ
Girls .85 .03 2.61 .11
ð1.40Þ ð2.11Þ
Home environment:
Few books 28.08 2.33 216.00 2.66
ð2.54Þ ð3.61Þ
Many books 4.36 .18 2.05 .08
ð.90Þ ð2.62Þ
âðtÞ;2 2 b
Note.— 2yYijt =yRA 5 2b âðtÞ;3 EAj
EA57 .
5 years and 7.00 months. The effects on all children in the sample of
changing the entrance cutoff date are large, especially considering that
this date typically varies across a 4-month interval. The estimates imply
that the average child would experience effects of 0.20 and 0.44 stan-
dard deviations on math and reading test scores if the state moved from
the latest to the earliest cutoff date in the sample. However, the effects
of 1-month changes are muted and are no larger than 0:11j for both
math and reading.
It is perhaps surprising that the effects of an earlier cutoff date are
positive because this indicates that children perform better when relatively
younger than their classmates. This positive effect indicates that the sec-
ond mechanism through which relatively younger children speed up their
learning to match the level of their peers dominates the harm done by
the first mechanism. And while it might be surprising that children have
higher achievement when relatively younger, similar results have also been
reported in the literature. Elder and Lubotsky ð2009Þ document that rel-
atively younger children have higher achievement, and Dobkin and Fer-
reira ð2010Þ find that younger children have higher academic attainment.
The heterogeneity in effects by gender and home environment is
remarkably strong, and this heterogeneity is likely to be the most impor-
tant finding for policy makers. The effects on boys are much larger than
the effects on girls. If it is surprising that children perform better when
relatively younger than their classmates, it might be even more surprising
to see this result for boys. Together with the well-documented practice of
redshirting boys more often than girls ðDeming and Dynarski 2008; Ali-
prantis 2012Þ, this heterogeneity suggests that a topic for future research
to investigate is whether early childhood trends might be contributing to
the gender gap in attainment ðGoldin, Katz, and Kuziemko 2006Þ.

Children with few books at home perform much worse when relatively
younger than their classmates. Thinking about the two mechanisms dis-
cussed earlier, one hypothesis is that such children’s family inputs work in
favor of the first mechanism dominating the second mechanism. Children
with low family inputs who are also relatively young may simply be too far
behind to catch up with their peers. Policy makers moving their state’s
entrance cutoff date to an earlier point hurts children with few home
inputs as measured by books at home.
V. Effects of Changing Birth Dates

An issue of interest for parents is understanding how their children being
born at a different point in time would affect their achievement. This
might be posed by parents as the question, How would an earlier birth
ðand conceptionÞ date affect an eligible child’s achievement? In contrast
to the effect in Section IV, this question is more relevant for individual-
level choices than for entrance age policy choices.
Thus in addition to the relative age mechanisms discussed in Section IV,
two mechanisms related to entrance age are relevant for understanding
the effect of changing a child’s birth date. A first entrance age mechanism
one might imagine is that increasing the age at which children enter
school increases the time they remain in the home or preschool environ-
ment. This mechanism can be seen as a change to the allocation of
children’s time ðFiorini and Keane 2011Þ. A second mechanism would be
that children have more time for their biological development process to
unravel. A large literature in child psychology has focused on understand-
ing the relationship between these two mechanisms, that is, how children’s
environments and activities interact with the process of their biological
development ðTudge, Freitas, and Doucet 2009Þ.
We can also use the estimated EPF parameters to investigate the effects
of this counterfactual scenario. Inserting EA and R A from equations
ð8Þ and ð9Þ into the production function in equation ð17Þ, we can see
that, all else constant, moving a child’s birthday to an earlier date is the
same as increasing both the child’s entrance and relative ages. Thus an
earlier birth date results in a combination of entrance and relative age
effects:
yYt
2 5 baðtÞ;1 1 baðtÞ;2 2 baðtÞ;3 ðEA 1 RAÞ:
yBDay
Table 25 shows estimates of these effects evaluated at the midpoint of

the support of EA 1 RA in our sample, 15. As a result of the imprecision of
these effect estimates, it is difficult to make strong statements from the
point estimates. However, the effects in table 25 are all estimated to be
positive, even if they are not statistically significant. This helps to illustrate

TABLE 25
Counterfactual 2: Effects from an Earlier Birth Date Spring of Third Grade
Math Reading
Sample Raw j Raw j
All children 8.33 .34 3.34 .14
ð4.17Þ ð5.89Þ
Gender:
Boys 13.46 .55 11.05 .45
ð7.73Þ ð11.27Þ
Girls 7.59 .31 2.01 .08
ð6.50Þ ð9.38Þ
Home environment:
Few books 1.97 .08 4.38 .18
ð10.01Þ ð15.04Þ
Many books 6.72 .28 4.08 .17
ð8.55Þ ð10.80Þ
âðtÞ;1 1 b
Note.— 2yYt =yBDay 5 b âðtÞ;2 2 b
âðtÞ;3 ðEA 1 RAÞj
EA1RA515 .
that entrance age parameters typically have the opposite sign and a larger
magnitude than their corresponding relative age parameters.
VI. Policy Discussion

The analysis conducted in this paper is ultimately aimed at helping to
determine the optimal age for children to start school. There are at least
three hurdles to assessing when states should set their entrance cutoff
dates.
The first hurdle is that deciding when children should start school
requires that we understand how changing the entrance cutoff date would
affect the achievement of an entire cohort of children. As discussed in
Section IV, the changes to children’s entrance and relative ages induced by
changing the entrance cutoff date do not satisfy monotonicity assump-
tions ði.e., the changes in these ages are in different directions for differ-
ent childrenÞ. The evidence presented here pertains to the monotonic
impacts for only the oldest children in a cohort.
Although it should be possible to estimate the effects from different
types of changes in entrance and relative ages, it must be noted that such
effects, like the EPF parameters estimated in this paper, will typically not
be structural in the sense of being policy invariant. The parameters in this
analysis describe the technology producing educational achievement for
a specific subset of children in the United States entering kindergarten in
the fall of 1998. It is entirely possible that the parameters of the produc-
tion function have changed since then as the organization and behaviors
of families, schools, and communities have themselves changed. In fact,
one key reason for doing research such as that conducted in this paper is
to inform education policy with the goal of improving the technology of

the production function. It seems unlikely that any of the estimated

parameters represent inherent or unaltering relationships between age
and achievement.
Finally, even if we did fully understand the achievement production
technology at all ages and under all policies, we would still need to use
some objective function in order to weigh the effects of changing entrance
cutoff dates on achievement against the concurrent effects to society from
changing child care costs ðCascio 2009; Fitzpatrick 2010Þ, attainment
ðAngrist and Krueger 1991; Dobkin and Ferreira 2010Þ, and foregone
labor force participation and government revenue ðDeming and Dynarski
2008Þ. Such an objective function, from the perspective of policy makers,
would most likely be concerned with the achievement of children at a
given age. However, most studies in the literature are not informative
about achievement at a given age under counterfactual entrance age
policies. Rather, most of the literature takes the perspective of children
evaluated relative to their peers without conditioning on age. This makes
sense from the individual’s perspective because this is how children are
evaluated in schools and in society. However, optimal policy would likely
be designed around maximizing children’s development at a given age.
These hurdles make it difficult to imagine a single answer to the
question, When should children start school? Furthermore, the evidence
of heterogeneity presented in this analysis indicates that from a broader
policy perspective, answering this question may be of less importance than
answering the question, How can schools best meet the diverse needs of
new students? These questions need not be independent of each other,
but children will continue to enter school with large age differences no
matter when the entrance cutoff date is set.
VII. Conclusion
This paper studied causal effects informative for determining the age
when children should start kindergarten. The first results were a set of
hypothesis tests rejecting the homogeneity of treatment effects from
delaying kindergarten entry. Because choices to delay entry create viola-
tions of the instrument monotonicity condition, this evidence suggests
that standard IV strategies do not identify the effect of choosing to delay a
child’s entry for any subpopulation of interest ðBarua and Lang 2009;
Aliprantis 2012Þ.
I presented an identification strategy combining features of Angrist and
Krueger ð1991Þ and Todd and Wolpin ð2003Þ modified to identify educa-
tion production function parameters in spite of these methodological
difficulties. The identified EPF parameters allowed for the construction
of outcomes under counterfactual policies and parental choices. The
central result of the paper was that increasing the relative age of students’
classmates would actually increase their achievement in the spring of third
grade, at least for the older children in a class. That is, relative age has

negative returns for older children. The finding that initial age advantages
need not translate into superior development trajectories is evidence that
relative advantages at early ages need not determine subsequent develop-
ment trajectories.
The heterogeneity of effects documented in this study points early
childhood education policy toward early childhood environments and
schools that can accommodate the relatively large differences in ages and
skills at initial enrollment. Effects identified in the analysis were remark-
ably heterogeneous by gender and by home environment as measured by
the number of books at home. Relative age effects on boys were much
larger than the effects on girls. Children with few books at home per-
formed much worse when entering at an older entrance age. If a goal of
policy is to foster skills more efficiently begetting skills without advantages
necessarily begetting advantages, future research could contribute to this
goal by further investigating the trade-offs alternative early childhood
education policies represent for various subpopulations.
Appendix A
Policy Background
One reason for increasing interest in policy determining kindergarten entrance

age is research showing the importance of the childhood environment. There is a
line of research showing that investments in children and their parents before the
age of 5 can have profound impacts on adult outcomes ðAlmond and Currie 2010Þ.
Another line of research shows that outcomes determined by age 15 or 16 can
explain a great deal of subsequent education and labor market outcomes ðNeal
and Johnson 1996; Keane and Wolpin 1997, 2000; Eckstein and Wolpin 1999;
Cameron and Heckman 2001Þ.
This research has coincided with an increase during recent decades in the
average age of children entering kindergarten in the United States ðElder and
Lubotsky 2009Þ. This increase can be seen as the result of state-level policy changes
moving entrance cutoff dates to an earlier point in the year. Since an older entrance
age is empirically associated with better performance in school, these changes in
policy can first be seen as a response to legislation such as 1994’s Goals 2000:
Educate America Act and 2002’s No Child Left Behind Act of 2001, which give
schools and state governments increasing incentives to raise children’s test
scores. Increasing age at enrollment could be seen as a simple way to achieve this
objective without making more fundamental changes to instruction or the distri-
bution of student characteristics across schools.
However, high-stakes testing and policy changes are not the only cause of the
increasing kindergarten entrance age, as the average kindergarten entrance age
began to increase before the introduction of high-stakes testing ðDeming and
Dynarski 2008Þ. Recent changes in entrance age can also be attributed to increas-
ing concerns about the “readiness” of children for school, a vague concept that
might be best interpreted as relative development at a given calendar date ðGraue
1993; Graue and DiPierna 2000Þ. Evidence of these concerns is seen in parents and
schools delaying the enrollment of many children until the year after they are first

eligible to enroll or else having children repeat kindergarten. The prevalence of

this practice, known as redshirting, is 8 percent for children in the ECLS-K, with
increasing frequency for the relatively youngest in any cohort.16
Overall, the choices made by parents, teachers, and schools appear to be more
important in explaining changes in the increasing average age at enrollment than
changes in entrance cutoff dates. Deming and Dynarski ð2008Þ find that only a
quarter of the decrease in 6-year-olds enrolled in first grade in recent decades has
come from changes in state entry laws, with the rest of the decrease coming from
choices made by parents, teachers, and schools. Of those chosen to be delayed by
third grade, Bedard and Dhuey ð2006Þ find that 41 percent are behind because of
delayed enrollment and 59 percent are behind because of retention.
Appendix B
Monotonicity in the Redshirting Model Used to Test Treatment

Effect Homogeneity
Recall assumption EH1 from Section II.

EH1. gi 5 g for all i and g ≠ 0.
Agents select into treatment D on the basis of the latent index D * :

← 1 if D * ≥ 0
D5 ðB1Þ
0 otherwise;
with
←
D* 5 mD ðX ; Z Þ 2 V
←
ðB2Þ
5 mðX Þ 1 gZ 2 V :
Assumption EH1 holds in this model because of the restricted subsample
selected in Section II.D. Consider the full sample of all children indexed by their
fifth birthday. Specifically, consider the children born on dates e, f, and g in fig-
ure B1. Suppose that a child whose fifth birthday was on date g was counterfactu-
ally moved to date f. The child faces two possibilities. If he enrolls when first eli-
gible, then this counterfactual change in birth date would result in his entering at
a younger age. If he were redshirted, or delayed entry, he would in contrast enter
later as a result of this counterfactual change in birth date. This is the violation of
monotonicity discussed in Barua and Lang ð2009Þ and Aliprantis ð2012Þ.
This violation of monotonicity is not of concern for the model in Section II
because of the restricted nature of the treatment considered and because of the
restricted subsample chosen in Section II.D. In terms of treatment, children enter
either within a 3-month window or else 1 year later. The model abstracts from
variation in entrance or relative ages within either of these 3-month windows and
instead attributes all age effects to the variation in entrance and relative ages
generated by moving between these 3-month windows, or by redshirting.
16
Other estimates of the prevalence of redshirting range between 4 and 5.5 percent ðfor
first-time kindergartners; Bassock and Reardon 2013Þ to 16 percent ðDeming and Dynarski
2008Þ in the United States and 15 percent in Australia ðEdwards, Fiorini, and Taylor 2011Þ.
The data used to construct this estimate are discussed in depth in Secs. II.C and III.

Figure B1.—Sample selection for the effect of redshirting model: A, all children facing a
September 1, 1998, cutoff date; B, the entering cohort ðonlyÞ facing a September 1, 1998,
cutoff date. Date e: a child turning 5 on July 1, 1998. Date f : a child turning 5 on August 1,
1998. Date g: a child turning 5 on October 1, 1998.
In terms of the sample, the only children used in estimation of the model were
children who turned 5 within 3 months before the entrance cutoff date. Thus the
only possible counterfactual manipulations to fifth birth dates is between dates
such as e and f, not between e and g or f and g. Owing to the binary nature of
treatment, counterfactual manipulations of fifth birth dates between dates e and f
leave treatment unaffected at D 5 0 as long as the child enters when first eligible.
The birth date instrument affects outcomes only through its affect on redshirting
by making children more likely to receive the redshirting treatment and therefore
to switch from D 5 0 to D 5 1. Had the sample been less restricted and included
children turning 5 on date g, this simple model of treatment would not apply, and
monotonicity would be violated.
Appendix C
Comparison with Value-Added Strategies

Since assumptions MA1–MA4 impose strong restrictions on the data-generating
process, we might be interested in alternative specifications allowing for the
identification of parameters of interest in equation ð10Þ:
←
Yijt 5 YAðtÞ fF ij ½Ai ðtÞ; S ij ½Ai ðtÞ; mij0 ; eijt g: ð10Þ
Define the vector of family and school inputs to be
X ij ½Ai ðtÞ ; ½F ij ½Ai ðtÞ; S ij ½Ai ðtÞ;

and suppose that we were to estimate the following value-added specification:

←
Yijt 5 X ij ½Ai ðtÞa 1 gYijt21 1 nijt : ðC1Þ
Under what assumptions would the OLS estimation of equation ðC1Þ identify
the parameters of equation ð10Þ? Recall that according to Todd and Wolpin
ð2003Þ, the empirical implementation of equation ð10Þ has floundered on two
basic problems.
Problem i. Ability mij0 is unobserved.

Problem ii. Data sets typically do not include the entire vectors of Fij½AiðtÞ
and Sij½AiðtÞ.
Suppose that the true technology in equation ðC1Þ were as follows:

←
Yijt 5 X ij ½Ai ðtÞa1;aðtÞ 1 X ij ½Ai ðt 2 1Þa2;aðtÞ 1
ðC2Þ
1 X ij ½Ai ð1Þaa;aðtÞ 1 daðtÞ mij0 1 eijt :
Under the following assumption:
MAðiÞ. aaðtÞ 5 a for all a.
Then it is possible to write
←
Yijt 2 gYijt21 5 X ij ½Ai ðtÞa1 1 X ij ½Ai ðt 2 1Þða2 2 ga1 Þ 1
ðC3Þ
1 X ij ½Ai ð0Þðaa 2 gaa21 Þ 1 ½daðtÞ 2 gdaðt21Þ mij0 1 eijt 2 geijt21 :
Todd and Wolpin ð2003Þ derive the assumptions under which equation ðC3Þ re-
duces to equation ðC1Þ so that input coefficients can be consistently estimated via
OLS estimation of ðC1Þ. They show that in addition to MAðiÞ, the following as-
sumptions must also be made.
MAðiiÞ. aj 5 gaj21 for j ∈ f2, . . . , ag.

MAðiiiÞ. daðtÞ 5 gdaðt21Þ .
MAðivÞ. E½eijt 2 geijt21 5 0.
Assumptions MA1–MA4 are similar to assumptions MAðiÞ–MAðivÞ. For example,

relaxing assumption MAðiÞ or MAðiiÞ would require the adoption of an assumption
analogous to MA2 in order to address problem ii. Similarly, relaxing assumption
MAðiiiÞ would require the identification of ability, therefore requiring some as-
sumptions analogous to MA3–MA4 in order to address problem i.
There are also reasons to believe that assumptions MA1–MA4 compare favor-
ably to assumptions MAðiÞ–MAðivÞ, which is the judgment adopted in this analysis.
There is some theoretical reasoning to believe that assumptions MA1–MA4 may
be violated. However, as discussed in the text, similar reasoning might lead us to
conclude that this violation is relatively minor and that assumptions MA1–MA4
approximate reality. At the very least, there is reason to believe that assumptions
MA1–MA4 approximate reality better than assumptions MAðiÞ–MAðivÞ. The restric-
tions on the technology implied by assumptions MAðiÞ–MAðivÞ are not motivated
by theory, but rather by the fact that they allow us to empirically implement

equation ð10Þ. Empirically, Rothstein ð2010Þ and Andrabi et al. ð2011Þ find that
value-added specifications do not overcome problem i because of the correlation
of ability with inputs, and Todd and Wolpin ð2007Þ find that in the NLSY79
children sample, there is “some support for the pattern of declining coefficients
with age; however, statistical tests reject the strict value-added formulation” ð100Þ.
Although value-added specifications may perform well in certain contexts ðTodd
and Wolpin 2007; Kinsler 2012Þ, there is no reason to believe that assumptions
MAðiÞ–MAðivÞ impose weaker restrictions on the underlying technology in equa-
tion ð10Þ than the assumptions adopted in this analysis, MA1–MA4.
References
Aliprantis, D. 2012. “Redshirting, Compulsory Schooling Laws, and Educational
Attainment.” J. Educ. and Behavioral Statis. 37 ð2Þ: 316–38.
———. 2014a. “Assessing the Evidence on Neighborhood Effects from Moving to
Opportunity.” Working Paper no. 12-33r, Fed. Reserve Bank Cleveland.
———. 2014b. “Covariates and Causal Effects: The Problem of Context.” Working
Paper no. 13 -10r, Fed. Reserve Bank Cleveland.
Aliprantis, D., and D. Carroll. 2013. “Neighborhood Dynamics and the Distribu-
tion of Opportunity.” Working Paper no. 12-12r, Fed. Reserve Bank Cleveland.
Almond, D., and J. Currie. 2010. “Human Capital Development before Age Five.”
Working Paper no. 15827, NBER, Cambridge, MA.
Andrabi, T., J. Das, A. I. Khwaja, and T. Zajonc. 2011. “Do Value-Added Estimates
Add Value? Accounting for Learning Dynamics.” American Econ. J.: Appl. Econ. 3
ð3Þ: 29–54.
Angrist, J. D., and A. B. Krueger. 1991. “Does Compulsory School Attendance
Affect Schooling and Earnings?” Q.J.E. 106 ð4Þ: 979 –1014.
Barua, R., and K. Lang. 2009. “School Entry, Educational Attainment and Quarter
of Birth: A Cautionary Tale of LATE.” Working Paper no. 15236, NBER, Cam-
bridge, MA.
Bassock, D., and S. F. Reardon. 2013. “ ‘Academic Redshirting’ in Kindergarten:
Prevalence, Patterns, and Implications.” Educ. Evaluation and Policy Analysis 35
ð3Þ: 283 –97.
Bedard, K., and E. Dhuey. 2006. “The Persistence of Early Childhood Maturity:
International Evidence of Long-Run Age Effects.” Q.J.E. 121 ð4Þ: 1437–72.
Black, S. E., P. J. Devereux, and K. G. Salvanes. 2011. “Too Young to Leave the Nest?
The Effects of School Starting Age.” Rev. Econ. and Statis. 93 ð2Þ: 455– 67.
Blackmon, D. A. 2008. Slavery by Another Name: The Re-enslavement of Black People in
America from the Civil War to World War II. New York: Doubleday.
Bobak, M., and A. Gjonca. 2001. “The Seasonality of Live Birth Is Strongly Influ-
enced by Socio-demographic Factors.” Human Reproduction 16 ð7Þ: 1512–17.
Bogard, K., and R. Takanishi. 2005. “PK-3: An Aligned and Coordinated Approach
to Education for Children 3 to 8 Years Old.” Soc. Policy Report 19 ð3Þ: 3–23.
Bound, J., and D. A. Jaeger. 2000. “Do Compulsory School Attendance Laws Alone
Explain the Association between Quarter of Birth and Earnings?” In Worker Well-
Being, edited by S. W. Polachek, 83–108. Research in Labor Economics, vol. 19.
New York: JAI.
Buckles, K., and D. M. Hungerman. 2008. “Season of Birth and Later Outcomes:
Old Questions, New Answers.” Working Paper no. 14573, NBER, Cambridge,
MA.
Cameron, S. V., and J. J. Heckman. 2001. “The Dynamics of Educational Attain-
ment for Black, Hispanic, and White Males.” J.P.E. 109 ð3Þ: 455–99.
Cascio, E. U. 2009. “Maternal Labor Supply and the Introduction of Kindergartens
into American Public Schools.” J. Human Resources 44 ð1Þ: 140 –70.

Cascio, E. U., and E. G. Lewis. 2006. “Schooling and the Armed Forces Qualifying
Test.” J. Human Resources 41 ð2Þ: 294–318.
Cascio, E. U., and D. W. Schanzenbach. 2007. “First in the Class? Age and the Ed-
ucation Production Function.” Working Paper no. 13663, NBER, Cambridge, MA.
Chen, B., and J. Pearl. 2012. “Regression and Causation: A Critical Examination of
Econometrics Textbooks.” Manuscript, Cognitive Systems Lab., Univ. California,
Los Angeles.
Cook, P. J., and S. Kang. 2013. “Birthdays, Schooling, and Crime: New Evidence
on the Dropout-Crime Nexus.” Working Paper no. 18791, NBER, Cambridge,
MA.
Cunha, F., and J. J. Heckman. 2007. “The Technology of Skill Formation.” A.E.R.
97 ð2Þ: 31– 47.
Cunha, F., J. J. Heckman, L. Lochner, and D. V. Masterov. 2005. “Interpreting the
Evidence on Life Cycle Skill Formation.” IZA Discussion Paper no. 1675, Inst.
Study Labor, Bonn.
Currie, J. 2001. “Early Childhood Education Programs.” J. Econ. Perspectives 15 ð2Þ:
213–38.
Datar, A. 2006. “Does Delaying Kindergarten Entrance Give Children a Head
Start?” Econ. Educ. Rev. 25:43–62.
Dawid, H., and G. Muehlheusser. 2012. “Repeated Selection with Heterogeneous
Individuals and Relative Age Effects.” IZA Discussion Paper no. 6478, Inst. Study
Labor, Bonn.
Deming, D., and S. Dynarski. 2008. “The Lengthening of Childhood.” J. Econ.
Perspectives 22 ð3Þ: 71–92.
Dhuey, E., and S. Lipscomb. 2008. “What Makes a Leader? Relative Age and High
School Leadership.” Econ. Educ. Rev. 27 ð2Þ: 173 –83.
Dickert-Conlin, S., and A. Chandra. 1999. “Taxes and the Timing of Births.” J.P.E.
107 ð1Þ: 161–77.
Dickert-Conlin, S., and T. Elder. 2009. “Suburban Legend: School Cutoff Dates
and the Timing of Births.” Manuscript, Michigan State Univ.
Dobkin, C., and F. Ferreira. 2010. “Do School Entry Laws Affect Educational
Attainment and Labor Market Outcomes?” Econ. Educ. Rev. 29 ð1Þ: 40–54.
Eckstein, Z., and K. I. Wolpin. 1999. “Why Youths Drop Out of High School: The
Impact of Preferences, Opportunities, and Abilities.” Econometrica 67 ð6Þ: 1295–
1339.
Edwards, B., M. Fiorini, and M. Taylor. 2011. “Who Gets the ‘Gift of Time’ in
Australia? Exploring Delayed Primary School Entry.” Australian Rev. Public Affairs
10 ð1Þ: 41– 60.
Elder, T. E., and D. H. Lubotsky. 2009. “Kindergarten Entrance Age and Children’s
Achievement: Impacts of State Policies, Family Background, and Peers.” J.
Human Resources 44 ð3Þ: 562–98.
Fiorini, M., and M. Keane. 2011. “How the Allocation of Children’s Time Affects
Cognitive and Non- cognitive Development.” Manuscript, Univ. Technology
Sydney.
Fitzpatrick, M. D. 2010. “Preschoolers Enrolled and Mothers at Work? The Effects
of Universal Prekindergarten.” J. Labor Econ. 28 ð1Þ: 51–85.
Fredriksson, P., and B. Öckert. 2005. “Is Early Learning Really More Productive?
The Effect of School Starting Age on School and Labor Market Performance.”
IZA Discussion Paper no. 1659, Inst. Study Labor, Bonn.
Goldberger, A. S., and C. F. Manski. 1995. Review of The Bell Curve by Herrnstein
and Murray. J. Econ. Literature 33 ð2Þ: 762–76.
Goldin, C., L. F. Katz, and I. Kuziemko. 2006. “The Homecoming of American
College Women: The Reversal of the College Gender Gap.” J. Econ. Perspectives
20 ð4Þ: 133 –56.

Graue, M. E. 1993. Ready for What? Constructing Meanings of Readiness for Kinder-
garten. Albany: State Univ. New York Press.
Graue, M. E., and J. DiPierna. 2000. “Redshirting and Early Retention: Who Gets
the ‘Gift of Time’ and What Are Its Outcomes?” American Educ. Res. J. 37 ð2Þ:
509–34.
Hahn, J., P. Todd, and W. van der Klaauw. 2001. “Identification and Estimation of
Treatment Effects with a Regression-Discontinuity Design.” Econometrica 69 ð1Þ:
201– 9.
Heckman, J. J. 2005. “The Scientific Model of Causality.” Sociological Methodology 35
ð1Þ: 1–97.
Heckman, J. J., H. Ichimura, J. Smith, and P. Todd. 1998. “Characterizing Selection
Bias Using Experimental Data.” Econometrica 66 ð5Þ: 1017–98.
Heckman, J. J., S. H. Moon, R. Pinto, P. Savelyev, and A. Yavitz. 2010a. “Analyzing
Social Experiments as Implemented: A Reexamination of the Evidence from the
HighScope Perry Preschool Program.” Quantitative Econ. 1 ð1Þ: 1–46.
———. 2010b. “The Rate of Return to the HighScope Perry Preschool Program.”
J. Public Econ. 94 ð1–2Þ: 114–28.
Heckman, J. J., D. Schmierer, and S. Urzúa. 2010. “Testing the Correlated Random
Coefficient Model.” J. Econometrics 158 ð2Þ: 177–203.
Heckman, J. J., S. Urzúa, and E. Vytlacil. 2006. “Understanding Instrumental
Variables in Models with Essential Heterogeneity.” Rev. Econ. and Statis. 88 ð3Þ:
389–432.
Heckman, J. J., and E. Vytlacil. 2005. “Structural Equations, Treatment Effects, and
Econometric Policy Evaluation.” Econometrica 73 ð3Þ: 669 –738.
Holm, S. 1979. “A Simple Sequentially Rejective Multiple Test Procedure.” Scandi-
navian J. Statis. 6 ð2Þ: 65–70.
Imbens, G. W., and J. D. Angrist. 1994. “Identification and Estimation of Local
Average Treatment Effects.” Econometrica 62 ð2Þ: 467–75.
Keane, M. P., and K. I. Wolpin. 1997. “The Career Decisions of Young Men.” J.P.E.
105 ð3Þ: 473–522.
———. 2000. “Eliminating Race Differences in School Attainment and Labor
Market Success.” J. Labor Econ. 18 ð4Þ: 614–52.
Khan, S. 2012. The One World Schoolhouse: Education Reimagined. New York: Twelve.
Kinsler, J. 2012. “Assessing Rothstein’s Critique of Teacher Value-Added Models.”
Quantitative Econ. 3 ð2Þ: 333–62.
Kirp, D. L. 2007. The Sandbox Investment: The Preschool Movement and Kids-First Politics.
Cambridge, MA: Harvard Univ. Press.
Koedel, C., R. Leatherman, and E. Parsons. 2012. “Test Measurement Error and
Inference from Value-Added Models.” B.E. J. Econ. Analysis and Policy 12 ð1Þ: art.
52.
Lam, D. A., and J. A. Miron. 1996. “The Effects of Temperature on Human
Fertility.” Demography 33 ð3Þ: 291–305.
Lam, D. A., J. A. Miron, and A. Riley. 1994. “Modeling Seasonality in Fecundability,
Conceptions, and Births.” Demography 31 ð2Þ: 321–46.
Lillard, A. S. 2008. Montessori: The Science behind the Genius. Oxford: Oxford Univ.
Press.
Manski, C. F. 2011. “Genes, Eyeglasses, and Social Policy.” J. Econ. Perspectives 25 ð4Þ:
83–94.
McCrary, J., and H. N. Royer. 2011. “The Effect of Maternal Education on Fertility
and Infant Health: Evidence from School Entry Policies Using Exact Date of
Birth.” A.E.R. 101 ð1Þ: 158–95.
McEwan, P. J., and J. S. Shapiro. 2008. “The Benefits of Delayed Primary School
Enrollment: Discontinuity Estimates Using Exact Birth Dates.” J. Human Re-
sources 43 ð1Þ: 1–29.

Minnesota Population Center. 2004. National Historical Geographic Information Sys-

tem. Prerelease version 0.1 ed. Minneapolis: Univ. Minnesota. http://www.nhgis
.org.
Muscha, J., and S. Grondin. 2001. “Unequal Competition as an Impediment to
Personal Development: A Review of the Relative Age Effect in Sport.” Develop-
mental Rev. 21 ð2Þ: 147–67.
Neal, D. A., and W. R. Johnson. 1996. “The Role of Premarket Factors in Black-
White Wage Differences.” J.P.E. 104 ð5Þ: 869–95.
Puhani, P. A., and A. M. Weber. 2005. “Does the Early Bird Catch the Worm?
Instrumental Variable Estimates of Educational Effects of Age of School Entry
in Germany.” IZA Discussion Paper no. 1827, Inst. Study Labor, Bonn.
Rizzi, E. L., and G. Dalla-Zuanna. 2007. “The Seasonality of Conception.” Demog-
raphy 44 ð4Þ: 705–28.
Rolnick, A., and R. Grunewald. 2003. Early Childhood Development: Economic Develop-
ment with a High Public Return. Minneapolis: Fed. Reserve Bank.
Rothstein, J. 2010. “Teacher Quality in Educational Production: Tracking, Decay,
and Student Achievement.” Q.J.E. 125 ð1Þ: 175–214.
Shonkoff, J., et al. 2000. From Neurons to Neighborhoods: The Science of Early Childhood
Development. Washington, DC: Nat. Acad. Press ðfor Nat. Res. Council and Inst.
MedicineÞ.
Stipek, D. 2002. “At What Age Should Children Enter Kindergarten? A Question
for Policy Makers and Parents.” Soc. Policy Report 16 ð2Þ: 1–16.
Todd, P., and K. I. Wolpin. 2003. “On the Specification and Estimation of the
Production Function for Cognitive Achievement.” Econ. J. 113 ð485Þ: F3–F33.
———. 2007. “The Production of Cognitive Achievement in Children: Home,
School, and Racial Test Score Gaps.” J. Human Capital 1 ð1Þ: 91–136.
Tudge, J. R. H., L. B. L. Freitas, and F. Doucet. 2009. “The Transition to School:
Reflections from a Contextualist Perspective.” In Educational Theories, Cultures,
and Learning: A Critical Perspective, edited by H. Daniels, H. Lauder, and J. Porter,
117–33. London: Routledge.
Washington, H. A. 2006. Medical Apartheid: The Dark History of Medical Experimenta-
tion on Black Americans from Colonial Times to the Present. New York: Harlem Moon
Broadway.
Weil, E. 2007. “When Should a Kid Start Kindergarten?” New York Times Mag.,
June 3.
White House. 2013. “Fact Sheet: President Obama’s Plan for Early Education for
All Americans.” February 13. http://www.whitehouse.gov.
Willett, J. B. 1988. “Questions and Answers in the Measurement of Change.” In
Review of Research in Education, vol. 15, edited by E. Z. Rothkopf, 345– 422.
Washington, DC: American Educ. Res. Assoc.
Wilson, W. J. 1987. The Truly Disadvantaged: The Inner City, the Underclass, and Public
Policy. Chicago: Univ. Chicago Press.
Zuberi, T. 2001. Thicker than Blood: How Racial Statistics Lie. Minneapolis: Univ.
Minnesota Press.


Aliprantis - When Should Children Start School

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Aliprantis - When Should Children Start School

Загружено:

Авторское право:

Доступные форматы

When Should Children Start School?

[ Journal of Human Capital, 2014, vol. 8, no. 4]

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

avoid this problem at the cost of identifying effects pertaining to a more

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

II. Effects of Delaying Entry

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

In this scenario of essential heterogeneity, to be formally defined in

B. Model and Test of Redshirting Effect Homogeneity

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

which monotonicity must hold is taken as evidence about essential het-

Here it is worth discussing the precise combination of changes to en-

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

Note that nowhere in EH1–EH5 is any restriction imposed regarding

If we define bi ; Yi ð1Þ 2 Yi ð0Þ, then note that essential heterogeneity

H0IV : bIV;½p ;p 1 5 bIV;½p ;p 2 ; ð6Þ

where ½p ; p 1 and ½p ; p 2 represent two disjoint intervals of the support of

for j 5 1, 2. The Wald test statistic of H0IV is asymptotically distributed as

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

RA ; T ðCutoff Date; Year of EntryÞ 2 T ð5th BDayÞ

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

age of 12 months. Redshirted children will have relative ages between 12

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

D. Empirical Specification and Test Results

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

Figure 5.—Predicted probability of redshirting ðconditional on mðX Þ and percentiles of Z Þ

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

^ ðx; zÞ of Redshirted Subsample

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

^ ðx; zÞ of Redshirted Subsample

against effect homogeneity is a rather negative result as it suggests that

III. Identifying Education Production Function Parameters

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

shirting. However, this evidence does not preclude the possibility of

B. Identification Strategy: Sample Selection

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

C. Identification Strategy: Education Production Function Specification

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

Figure 8.—The distributions of reading IRT test scores

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

of student i in school j at age A iðtÞ is a function of the entire history of

Equation ð10Þ is specified to allow for certain age-related features

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

we would expect relatively younger children to perform better on achieve-

2. Empirical Specification and Identifying Assumptions

Problem i. mij0 is unobserved.

This analysis imposes four key additional assumptions in order to over-

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

Furthermore, it is assumed that in the fall of kindergarten ðt 5 1Þ, school

Assumptions MA1 and MA2 are maintained to overcome problem ii.

A natural question related to MA2 is whether input measures at the fall

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

Owing to a limited number of instruments, the t 5 1 vector Fij½Aið1Þ is as-

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

Assumption MA3 is clearly true; the first-stage regressions of inputs

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM

D. Identification Strategy: Identifying Variation

This content downloaded from 160.036.178.025 on September 15, 2016 08:35:26 AM