appendix - university of chicago · 1/21/2017 · 2. simultaneous processing scale (gestalt...
TRANSCRIPT
Appendix
40
A NFP Randomization Protocol
Pregnant women were enrolled as they visited the Medical Center during pregnancy. The randomization
protocol was sequential, that is to say that each participant was randomized according to the order of
enrollment. Pregnant women who accepted to enroll were classified in strata defined by 5 characteristics:
1. Maternal race (African American vs non-African American);
2. Maternal age (< 17, 17− 18, > 18 years );
3. Gestational age at enrollment (< 20, ≥ 20 weeks);
4. Employment status of the head of household and
5. 4 geographic regions of residence.17
Within strata, randomization was perform following the Soares and Wu (1983) method as follows:
1. If the participant had a sibling already enrolled in the program, the participant was assigned to the
same treatment status of the elder sibling.
2. Else, the participant was randomized according in to control (C) or treatment group (T). However, if the
sample size difference between treatment and control group is larger that a threshold, the participant
is deterministically assigned to the treatment status that has fewer participants.
3. Next the participant was randomized again into her final treatment status. If in step 2 the participant
was assigned to the control, she is next randomized in to:
• Group 1 : control women that received free transportation to and from their prenatal appointments
(sample size: 166).
• Group 2 : control women that received developmental screening and referral services at ages 6, 12
and 24 months in addition to the benefits of Group 1 (sample size: 514).
If she was assigned to the treatment group previously, she is next randomized into
• Group 3 : treated women that had home visits by nurses during pregnancy, one visit in the hospital
and one visit at home after childbirth in addition to the benefits of Group 2 (sample size: 230).
• Group 4 : treated women that received home visits by nurses during pregnancy until the child’s
2nd birthday in addition to the benefits of Group 2 (sample size: 228)
17The regions are: Inner City, Bisson, Cawthon and Hollywood.
41
As in the previous step, if the absolute difference in group size exceed some threshold then the partic-
ipant was deterministically assigned to the group with the lowest number of participants. Otherwise,
the pregnant woman was randomly assigned.
Importantly, the randomization method incorporated a trigger mechanism that deterministically assigned
a treatment status to participants if the sequence of assignments became too imbalanced due to sampling
variation. In this context, imbalance was measured by the difference of persons assigned to T and the persons
assigned to C. In practice less than 1% of the women were assigned according to the trigger mechanism.
Thus, the NFP Memphis trial can be treated as a non-sequential protocol.
B Instruments
B.1 HOME
The Home Observation for Measurement of the Environment (HOME) was first developed in the 1960s by
Caldwell. It measures the quality and quantity of stimulation and support available to a child at home
(Bradley and Caldwell (1984)). The test is based on the child and is carried on during a visit of 45 to 90
minutes. A more in depth explanation of the HOME inventory is in Bradley and Caldwell (1984).
There are several versions of the inventory. The initial version, Infant/Toddler HOME, is for kids aged 0
to 3 years old. It consists of 45 binary-choice items grouped into 6 subscales. The Early Childhood HOME
is for kids aged 3 to 6 years old. It consists of 55 binary-choice items clustered into 8 subscales. Finally, the
Middle Childhood HOME is used for kids aged 6 to 10 years old. It consists of 59 items in 8 subscales. The
NFP uses the first version of the Inventory, the Infant/Toddler HOME.
The 45 items of the HOME inventory contain the six following subscales:
1. Emotional and Verbal Responsiveness of the Mother (11 items): measures the mother’s ability to communicate with the
child.
2. Avoidance of Restriction and Punishment (8 items): measures the mother’s ability to discipline the child.
3. Organization of the Environment (6 items): measures the daily changes in the child’s environments.
4. Provision of Appropriate Play Material (9 items): measures the types of toys and their contributions to the child’s motor
skills.
5. Maternal Involvement with Child (6 items): measures the aspects in which the mother is involved in the child’s daily
life.
6. Opportunities for Variety in Daily Stimulation (5 items): measures the levels of interaction the mother and other family
members have with the child.
42
The NFP measured the HOME Inventory when the child was 12 and 24 months old.
B.2 KABC
The Kafuman Assessment Battery for Children (KABC) was developed by Alan S. Kaufman and Nadeen L.
Kaufman in 1983 with a later revision in 2004. The KABC focuses on processes required to solve problems
compared to intelligence. The KABC contains 16 subtests (10 mental processing and 6 achievement), which
can be grouped into 3 scales. Due to the nature of the subtests, 13 subtests can be taken at once, with the
mandatory age range to be between 7 to 12.5 years old. The NFP used the following 11 subtests:
1. Sequential Processing Scale (Hand Movements, Number Recall, Word Order): measures short-term memory and problem-
solving skills. It emphasizes how children are able to follow ordered sequenced.
2. Simultaneous Processing Scale (Gestalt Closure, Triangles, Matrix Analogies, Spatial Memory, Photo Series): measures
problem-solving skills. It involves several processes at once such as scenes in a partially completed picture.
3. Achievement Scale (Arithmetic, Riddles, Reading/Decoding): measures achievement and focus on applied skills and
facts learned through the home/school environment.
The NFP Program used these three scales when the child was 6 years old.
B.3 PPVT
The Peabody Picture Vocabulary Test (PPVT) is an individual verbal intelligence test that measures re-
ceptive vocabulary, developed by Llyod M. Dunn and Leota M. Dunn in 1959. It is a verbal test that lasts
between 20 and 30 minutes. The child is presented a series of pictures. There are four pictures in a page.
The examiner states a word and asks the child to associate it with a picture. The diffusion of the figures
increases over time. The exam stops when the child answers six out of eight questions incorrectly. After
completion, a raw score is given, normalized to a mean of 100 and standard deviation of 15. The NFP
Program used PPVT when the child was 6 years old.
B.4 WISC-III
The Wechsler Intelligence Scale for Children — Third Edition (WISC-III) was created in 1949. The third
edition was published in 1991 (Wechsler, 1991). WISC is an intelligence test for children between the ages
6 and 16 years old. It can be completed without reading or writing. The exam takes between 65 and 80
minutes. There are two subscales: verbal and performance, which provide a Verbal IQ (VIQ), a Performance
IQ (PIQ), and a Full Scale IQ (FSIQ). The NFP only used the coding, part of the Processing Speed Index.
43
1. Coding: the child marks rows of shapes with different lines to transcribe a digit-symbol code. It measures visual or motor
integration and visual scanning.
The NFP Program used WISC-III when the child was 6 years old.
B.5 CBCL
The Child Behavior Checklist (CBCL) is a parent-report questionnaire developed by Thomas M. Achenbach.
In it, the child is rated on several behavioral and emotional problems. The goal of the inventory is to assess
internalizing and externalizing behaviors. The responses are recorded using a Likert scale: 0 = Not True, 1
= Sometimes True, 2 = Very True. The preschool checklist (18 months to 5 years) contains 100 questions
and the school-age checklist (6 to 18 years) contains 120 questions. The preschool checklist questions can
be broken down into the following subscales: anxious/depressed, withdrawn, sleeping problems, somatic
problems, aggressive behavior, and destructive behavior. The school-age checklist questions can be broken
down into the following subscales: withdrawn, somatic complaints, anxious/depressed, social problems,
thought problems, attention problems, delinquent behavior, aggressive behavior, and other problems. The
NFP Program used the CBCL when the child was 2 and 6 years old.
B.6 C-DISC
The Computerized Diagnostic Interview Schedule for Children (C-DISC) is a comprehensive and structured
interview that studies mental health disorders such as general anxiety, panic, eating, elimination, depression,
ADHD, and conduct. The majority of the questions are “yes and no”. Some questions have additional
options such as “sometimes”. The C-DISC has about 3, 000 questions with 358 core questions that the child
must answer. About 1300 questions are asked depending on the answer to the core questions. There are 732
questions that ask about age of onset and treatment for reported symptoms. Lastly, there are about 700
questions from the optional whole-life module. The NFP Program used C-DISC when the child was 9 years
old.
B.7 MacArthur
The MacArthur Story Stem Battery (MSSB) was created by the MacArthur Narrative Working Group that
included Bretherton, Buchsbaum, and several other collaborators. The story stem method is where the
examiner presents a story to the child that culminates at a high point, at which the child is then asked
to complete the story; this type of method allows insight into the child’s inner workings of the mind. The
44
MSSB uses 15 stories and measures: dysregulated aggression, empathy/warmth, emotional integration, and
performance anxiety.
1. Dysregulated Aggression Dimension: aggression, injury, danger, destruction, dishonesty, escalation of conflict, negative
story endings, inappropriate child power, controlling toward examiner.
2. Empathy/Warmth Dimension: empathy-helping, affiliation, affection, reparation or guilt, parental warmth.
3. Emotional Integration Construct: ability to maintain story coherence with the inclusion of emotional expression. The
affects included are joy, anger, distress, concern, sadness.
4. Avoidance or Withdrawal Dimension: characters leaving the scene repetition of previous story fragments, denial of
central conflict or challenge, family characters leave, avoiding separation from parents, dissociative behaviors.
5. Performance Anxiety Dimension: unwillingness to verbalize, unresponsiveness to examiner, anxious behaviors.
The NFP Program used the MSSB when the child was 6 years old.
C Method for Permutation-based Inference
As mentioned, the standard model of program evaluation describes the observed outcome Yi of participant
i ∈ J by
Yi = DiYi,1 + (1−Di)Yi,0, (15)
where J = 1, . . . , N denotes the sample space indexing set, Di denotes the treatment assignment for
participant i ∈ J, (Di = 1 if treatment occurs, Di = 0 otherwise) and (Yi,0, Yi,1) are potential outcomes for
participant i when treatment is fixed at control and treatment status respectively.
Randomized experiments solve potential problems of selection bias by inducing independence between
counterfactual outcomes (Yi,0, Yi,1) and treatment status Di when conditioned on the pre-program variables
X used in the randomization protocol. All variables are defined in the common probability space (Ω,F ,P).
In our notation, a randomized experiment must satisfy the following assumption:
Assumption A-1. Y (d) ⊥⊥ D | X; d ∈ supp(D),
where variables X = (Xi; i ∈ J), D = (Di; i ∈ J) are N -dimensional vectors of treatment assignments
and pre-program variables, and Y (d) = (Yi,di ; i ∈ J, di ∈ 0, 1) and d ∈ supp(D) = 0, 1|J| denotes the
vector of counterfactual outcomes. In the same fashion, we represent the vector of observed outcomes of
Equation (15) by Y = (Yi; i ∈ I). The no treatment hypothesis is equivalent to the statement that the
conditional counterfactual outcome vectors share the same distribution:
45
Hypothesis H-1. Y (d)d= Y (d′) | X ; d,d′ ∈ supp(D),
Hypothesis H-1 can be restated in more tractable form:
Hypothesis H-1′. Under Assumption A-1 and Hypothesis H-1, we have that Y ⊥⊥ D | X.
Testing Hypothesis H-1′ poses some statistical challenges. First, small sample sizes cast doubt on
inference that rely on the asymptotic behavior of test statistics. We address the problem of small sample size
by generating the exact of a test statistic conditioned on data. Second, the presence of multiple outcomes
allows for the arbitrary selection of statistically significant outcomes. The selectively reporting statistically
significant outcomes is often termed cherry picking and generates a downward biased inference with smaller
p-values. We solve the problem of cherry picking by implementing a multiple-hypothesis testing based on
the stepdown procedure of (Romano and Wolf, 2005a). They explain that the stepdown procedure strongly
controls for Family-wise error rate (FWER), while classical tests do not. Also, Romano and Wolf (2005a)
shows that the strong FWER control can be obtained by ensuring a certain monotonicity condition on the
test statistics. This requirement is weaker than the assumption of subset pivotality, used in various methods
of resampling outcomes presented in Westfall and Young (1993).
In summary, our method is based on three steps. First, we seek to characterise the exact conditional
distribution of D|X. Specifically we characterize the a multiset Dx(d), defined by:
Dx(d) = d′ ∈ 0, 1|J| ; P(D = d|X = x) = P(D = d′|X = x),
such that the distribution of D conditioned on realised data is uniform among elements of Dx(d). Next we
use the assumption of null hypothesis of no treatment effects, i.e. H0 : Y ⊥⊥ D|X, to generate the exact
conditional distribution of a test statistic T (Y,D)|X. Upon it, we can construct an inference that controls
for the probability of falsely rejecting the null hypothesis. We control for this probability in two ways: (1) in
the case of single (joint) null hypothesis, we control for the standard Type-I error; (2) in the case of multiple
hypothesis inference, we control for the Family-wise error rate.
More notation is necessary to describe the method. Let K represent the indexing set for all available
outcomes Yk; k ∈ K. We represent the single (joint) null hypothesis that a set L ⊂ K of outcomes Yk; k ∈ L
are jointly independent of treatment status D conditional on pre-program variables X by
HL : YL ⊥⊥ D|X, where YL = (Yk : k ∈ L). (16)
When L is a singleton, say L = k, then the null hypothesis is given by Hk : Yk ⊥⊥ D|X. In this notation,
we can write the joint Hypothesis HL as HL = ∩k∈LHk.
46
Our goal is to tests a single (joint) null hypothesis controlling for the probability of a Type I error at
level α, that is, P( reject HL|HL is true ) ≤ α. To do so, we rely on the fact that, under Hypothesis (16),
(YL, D)|X d= (YL, gD)|X ∀ g ∈ GX , (17)
where GX comprises all the permutations within strata of X, that is,
GX = g ; g : J → J is a bijection and g(j) = j′ ⇒ (Xj) = (Xj′),
and gD is a vector defined by:
gD = (Di ∈ supp(D); i ∈ J and Di = Dg(i)).
We use Relation (17) to generate a statistical test whose exact distribution the test statistic TL(YL, gD) is
obtained by re-evaluating TL(YL, gD) as g varies in GX . Note that the inference on Hypothesis 16 is depend
on the choice of statistics. That is to say that even though any statistics TL(YL, D) whose value provide
evidence against the null hypothesis can be used, the inference is dependent on this choice of statistic. An
example of such statistic is the maximum of the t-statistic associated with the difference in means between
treated and control groups over outcomes Yk such that k ∈ L. Formally,
TL(YL, D) = maxk∈L
Tk(Yk, D), (18)
where Tk(Yk, D) is the t-statistics for outcome Yk. Relation (17) implies that TL(YL, D)|X d= TL(YL, gD)|X
for any g ∈ GX . Moreover, let d ∈ 0, 1|J| such that P(D = d|X = x) > 0, then the distribution of D
conditioned on X = x is uniform across elements of Dx(d) (see Lehmann and Romano (2005), Chapter 15).
Thus, a critical value cL,x(YL,d, α) such that P(TL(YL, D) > cL,x(YL,d, α)|X = x,HL is true) ≤ α can be
computed as:
cL,x(YL,d, α) = inft∈R
∑d′∈Dx(d)
ITL(YL,d′) ≤ t ≥ (1− α)|Dx|
,
where I· is the indicator function. The following notation is useful to further characterize cL,x(YL,d, α).
Let T(1)L,x, . . . , T
(|Dx(d)|)L,x be the sequence of increasing ordered statistics TL(YL,d
′) as d′ varies in Dx(d). In
this notation we can write the critical value as
cL,x(YL,d, α) = T(d(1−α)|Dx|e)L,x (19)
47
where dae stands for the smallest integer bigger or equal than a.
Under the null hypothesis HL, the probability of a test statistic be bigger or equal than the statistic
TL(YL,d) actually observed, i.e. the p-value, is given by:
pL,x(d) = infα∈[0,1]
cL,x(YL,d, α) ≤ TL(YL,d)
. (20)
Now let rL,x ∈ 1, . . . , |Dx(d)| be the lowest rank that the value of the observed test statistic TL(YL,d)
takes in the sequence T(1)L,x, . . . , T
(|Dx(d)|)L,x , that is to say:
rL,x = 1 +∑
d′∈Dx(d)
ITL(YL,d′) < TL(YL, d).
Thus:
T(rL,x)L,x = TL(YL,d). (21)
Then, by the ordered property of T(r)L,x; r ∈ 1, . . . , |Dx(d)| and the definition of rL,x, we have that:
pL,x(d) = 1− rL,x|Dx(d)|
. (22)
Moreover, p-value pL,x(d) complies with the following property:
P(pL,x(d) ≤ φ|X = x) ≤ φ ∀φ ∈ [0, 1].
We implemented a inference method that tests for the multiple null hypothesis that each outcome Yk; k ∈
L is independent of treatment status D conditional on pre-program variables X. The representation of these
multiple hypothesis in the same fashion as the single (joint) null hypothesis, namely:
HL = ∩k∈LHk ; Hk : Yk ⊥⊥ D|(Z,U).
The multiple hypothesis testing differs from the single (joint) hypothesis testing in the way it controls for
the probability of false rejection. Specifically, let the subset L0 be the set of true Hypothesis Hk such that
k ∈ L0 ⊂ L. Our multiple hypothesis testing controls for the familywise error rate (FWER), that is, the
probability of even one false rejection among the set of true hypothesis L0. Formally, we control for:
P( reject at least one Hk; k ∈ L0|HL0is true ) ≤ α,
48
while single (joint) hypothesis testing controls for P( reject HL|HL is true ) ≤ α.
Bonferroni or Holm are examples of inference methods that test multiple hypothesis controlling for
FWER. These methods rely upon a “least favorable” dependence structure among the p-values. The step-
down procedure of Romano and Wolf (2005a) is less conservative as it accounts for the information about
the dependence structure of p-values. The method is based on the monotonicity assumption, which, in our
case, can be stated as:
cK,x(YK ,d, α) ≥ cL0,x(YL0 ,d, α) for any subset K of L containing L0 i.e. L0 ⊂ K ⊂ L. (23)
Assumption (23) is satisfied by our choice of test statistic (18) and the fact that L0 ⊂ K.
The stepdown procedure given in Romano and Wolf (2005a) is a stepwise method summarized in the
following algorithm:
Algorithm 1.
Step 1: Set L1 = L. If
maxk∈L1
Tk(Yk,d) ≤ cL,x(YL1 ,d, α) , (24)
then stop and reject no null hypotheses; otherwise, reject any Hk with
Tk(Yk,d) > cL,x(YL1,d, α)
and go to Step 2.
...
Step j: Let Lj denote the indices of remaining null hypotheses. If
maxk∈Lj
Tk(Yk,d) ≤ cL,x(YLj ,d, α), (25)
then stop and reject no further null hypotheses; otherwise, reject any Hk with
Tk(Yk,d) > cL,x(YLj ,d, α)
and go to Step j + 1.
...
49
We can compute the multiplicity-adjusted p-values of Equations(24)–(25) in the same fashion described
by Equations (20)–(22).
C.1 Conditioning and Linearity
A typical problem in small sample randomized trials is sampling variation, where pre-program variables differ
across treatment groups by chance. One can increase the power of a statistical inference by conditioning on
those pre-program variables. Let Z be the pre-program variables that were not used in the randomization
protocol and we ought to control for.
Variables Z precede the treatment intervention and therefore Z ⊥⊥ D | X holds due to randomization.
Under the hypothesis of no-treatment, Y ⊥⊥ D | X also holds. These two relations imply that Y ⊥⊥
D | (X,Z). Likewise Section 4.1, we can use this relation to to generate a permutation test that consider
the strata formed by values of covariates X and Z. This way we can generates an inference method that
non-parametrically condition on variables X and Z.
Non-parametric conditioning through block permutation comes at a cost. A fine conditioning set decreases
the share of available data that can be permuted and a sufficiently large conditioning set prohibits the
implementation of a permutation-based test. We solve this problem by evoking linearity. That is to say, we
condition variables through a linear regression instead of a non-parametric block permutation. Anderson and
Legendre (1999) tested a range of permutation methods for linear models. They found that Freedman and
Lane (1983) generated the most consistent and reliable results among the available models in this literature.
We non-parametrically condition on variables used in the randomization protocol to achieve valid ex-
changeable properties (i.e. we use permutations in GX); We linearly condition on additional pre-program
variables Z not used in the randomization protocol. According to Freedman and Lane (1983) method,
our approach can be summarized by the following steps: (1) compute the residuals Y − Zβ such that
β = (Z ′Z)−1Z ′Y ; (2) permute these residuals according to permutations g ∈ GX . (3) add these permuted
residuals to Zβ, call it Y ; (4) regress Y on Z and the treatment statuses D. (5) we then use the t-statistic
associated with covariate D of the last regression as test statistic.
As mentioned, Beaton (1978) and Freedman and Lane (1983) suggested a permutation inference based
on Shuffle Residuals. Buy this I mean regressing Y on X, shuffling the residuals from this regression, and
adding them to the predicted Y, say Y , to form a new variable, say Y , which is then regressed on Z and D.
Formally, let the regression:
Y = Zβ +Dδ + ε,
where Z stands for the pre-program variables we wish to control for and includes a vector of elements ones
50
that play the role of a contant term for the regression. Error term ε is a mean-zero exogenous random
variable independent of Z and D.
Now let Bg; g ∈ GX be a permutation matrix associated with a permutation g in GX . Let also the operator
that projects a vector in the orthogonal space generated by columns of Z be MZ = I−Z(Z ′Z)−1Z ′, where I
denotes the identity matrix. As properties of Matrix MZ , we can say that MZ is symmetric and idempotent,
that is:
MZ = M ′Z = MZMZ = M ′ZMZ . (26)
The estimated residuals of Y generated by the the regression
Y = Zβ + ε
is given by e = MZY. The predicted outcome based on this regression is given by: Y = Z(Z ′Z)−1X ′Y.
We define the new outcome based on the sum of the predicted outcome Y with permuted errors e
according to permutation g ∈ GX as
Y = Y +Bg e. (27)
We then use the newly computed outcome in the following regression:
Y = Zβ +Dδ + ε. (28)
We now examine the δ estimate on Equation (28). This estimate ia actually the same as the one computed
by the following regression:
MZ Y = MZDδ + ε. (29)
Thus. by applying the Ordinary Least Square formula, we obtain:
δg = (D′M ′ZMZD)−1D′M ′ZMZ Y . (30)
51
We now use previous equations to transform Equation (30) into a more general formula:
δg = (D′M ′ZMZD)−1D′M ′ZMZ Y by (30),
= (D′MZDMZD′MZ Y by (26),
= (D′MZDMZD′MZ(Y +Bg e) by (27) ,
= (D′MZDMZD′MZ((I −MZ)Y +Bg e) because MZ = I − Z(Z ′Z)−1Z ′,
= (D′MZDMZD′((MZ −MZ)Y +MZBg e),
= (D′MZDMZD′(MZBg e),
= (D′MZDMZD′(MZBgMZY ) because e = MZY . (31)
Kennedy (1995) pointed out that Freedman and Lane (1983) algorithm is summarized by Equation 31.
Notationally, we can use TZ(Y, gD); g ∈ GX (instead of T (Y, gD); g ∈ GX) to represent the distribution of
the test statistic associated with the t-statsitic of the D covariate in the Freedman and Lane (1983) regression
just described. Using this notation, the analysis of the previous sections holds unaltered.
D Additional Baseline Tables
These tables explore differences in pre-program variables between treatment and control groups at each
follow up wave. These results are presented in order to highlight that there were no differential patterns of
attrition across waves.
52
Tab
leD
.1:
Des
crip
tive
Sta
tist
icof
Base
lin
eC
hara
cter
isti
cs(Y
ear
6)
C M
ean
C SD
T M
ean
T SD
Pval
C M
ean
C SD
T M
ean
T SD
Pval
C M
ean
C SD
T M
ean
T SD
Pval
Back
groun
d Ch
arac
terist
icsM
ater
nal R
ace
(Blac
k)0.
060
0.23
80.
100
0.30
10.
100
0.06
70.
251
0.08
10.
274
0.66
80.
054
0.22
60.
119
0.32
50.
070
Mar
ital S
tatu
s (M
arrie
d)0.
016
0.12
40.
015
0.12
20.
952
0.00
90.
094
0.01
00.
101
0.92
20.
022
0.14
80.
020
0.14
00.
883
Mat
erna
l Age
18.0
603.
220
18.0
603.
294
0.99
918
.219
3.29
918
.152
3.60
70.
874
17.9
023.
138
17.9
702.
971
0.85
0Y
ears
of E
duca
tion
10.2
631.
881
10.1
202.
024
0.39
510
.313
1.84
110
.081
2.06
90.
339
10.2
141.
922
10.1
581.
989
0.81
3M
othe
r in
Scho
ol0.
609
0.48
90.
580
0.49
50.
497
0.57
00.
496
0.61
60.
489
0.43
20.
647
0.47
90.
545
0.50
00.
084
Hea
d of
Hou
seho
ld is
Em
ploy
ed0.
562
0.49
70.
492
0.50
10.
106
0.60
50.
490
0.47
50.
502
0.03
10.
518
0.50
10.
510
0.50
20.
897
% o
f Cen
sus T
ract
Belo
w P
over
ty34
.812
21.3
7135
.518
20.2
210.
687
33.1
9520
.304
36.7
2422
.248
0.17
936
.428
22.3
1634
.336
18.0
490.
371
Hou
seho
ld D
ensit
y0.
940
0.49
71.
027
0.56
90.
064
0.96
10.
499
1.07
00.
669
0.15
10.
920
0.49
50.
986
0.45
10.
236
Tota
l Hou
sehold
Incom
e (Pa
st 6
Mon
ths)
Less
than
$30
000.
283
0.45
10.
365
0.48
30.
044
0.29
00.
455
0.36
40.
483
0.20
20.
277
0.44
80.
366
0.48
40.
116
$300
0 - $
6999
0.23
70.
425
0.22
50.
419
0.74
60.
219
0.41
40.
222
0.41
80.
945
0.25
40.
437
0.22
80.
421
0.60
1$7
000
- $10
999
0.22
80.
420
0.20
50.
405
0.51
50.
219
0.41
40.
222
0.41
80.
945
0.23
70.
426
0.18
80.
393
0.31
7G
reat
er th
an $
1100
00.
161
0.36
80.
125
0.33
20.
222
0.18
80.
391
0.08
10.
274
0.00
50.
134
0.34
10.
168
0.37
60.
434
Inco
me,
No
Resp
onse
0.09
20.
289
0.08
00.
272
0.62
50.
085
0.27
90.
111
0.31
60.
476
0.09
80.
298
0.05
00.
218
0.09
9
Regio
n of
Resid
ence
Inne
r City
0.29
50.
456
0.29
00.
455
0.90
50.
286
0.45
30.
303
0.46
20.
755
0.30
40.
461
0.27
70.
450
0.62
8Bi
sson
0.19
20.
394
0.21
50.
412
0.50
60.
179
0.38
40.
232
0.42
40.
282
0.20
50.
405
0.19
80.
400
0.87
9Ca
wth
on0.
194
0.39
60.
190
0.39
30.
900
0.21
00.
408
0.16
20.
370
0.29
70.
179
0.38
40.
218
0.41
50.
420
Hol
lywoo
d0.
319
0.46
70.
305
0.46
20.
719
0.32
60.
470
0.30
30.
462
0.68
40.
313
0.46
50.
307
0.46
40.
920
Mat
ernal
Men
tal H
ealth M
ater
nal I
Q (S
hipl
ey)
96.2
7010
.287
96.4
4010
.360
0.84
796
.223
10.2
7996
.061
10.6
180.
898
96.3
1710
.317
96.8
1210
.140
0.68
6M
ater
nal B
avol
ek S
core
99.7
947.
657
101.
133
8.50
20.
057
100.
091
7.41
110
1.43
18.
727
0.18
699
.499
7.89
910
0.84
28.
309
0.17
3M
ater
nal M
enta
l Hea
lth10
0.18
49.
979
99.4
4710
.352
0.39
899
.717
9.77
799
.741
10.1
720.
984
100.
649
10.1
7899
.158
10.5
680.
235
Self-
Effi
cacy
100.
083
10.0
1799
.788
9.86
60.
727
100.
862
9.77
810
0.58
39.
253
0.80
699
.307
10.2
1299
.008
10.4
190.
810
Mat
erna
l Mas
tery
100.
065
10.2
1399
.535
9.99
20.
537
99.8
7910
.155
99.1
7310
.246
0.56
810
0.25
010
.290
99.8
919.
774
0.76
4M
ater
nal P
sych
olog
ical R
esou
rces
100.
060
10.0
4599
.533
10.6
490.
554
100.
030
9.71
199
.619
10.6
340.
743
100.
090
10.3
9099
.448
10.7
150.
615
Mat
ernal
Hea
lth C
hara
cteris
tics
Mat
erna
l Heig
ht16
4.55
77.
253
164.
064
6.56
90.
397
164.
331
7.40
416
4.47
26.
546
0.86
516
4.78
17.
108
163.
651
6.60
10.
170
Pre-
Preg
nanc
y W
eight
62.0
9714
.866
62.3
3913
.588
0.83
962
.828
13.7
7561
.394
12.3
750.
355
61.3
6215
.885
63.2
6414
.683
0.29
4G
esta
tiona
l Age
(Int
ake)
16.5
605.
794
16.6
305.
728
0.88
716
.402
5.74
616
.364
5.59
60.
955
16.7
195.
850
16.8
915.
870
0.80
7
Mat
ernal
Socia
l Sup
port
Gra
ndm
othe
r Soc
ial S
uppo
rt10
0.19
79.
474
101.
517
8.56
60.
081
99.3
5710
.486
101.
434
9.10
00.
073
101.
034
8.28
510
1.59
98.
054
0.56
3H
usba
nd/B
oyfr
iend
Socia
l Sup
port
100.
030
9.99
410
0.70
49.
754
0.42
199
.892
10.0
5799
.907
9.52
40.
990
100.
169
9.95
210
1.48
49.
960
0.27
2
Mat
ernal
Risk
y Beh
avior
sA
lcoho
l Con
sum
ptio
n (P
ast 2
wks
)0.
043
0.20
20.
050
0.21
80.
680
0.03
60.
186
0.07
10.
258
0.22
80.
049
0.21
70.
030
0.17
10.
385
Smok
ing
(Pas
t 3 d
ays)
0.08
50.
279
0.11
00.
314
0.33
40.
081
0.27
30.
121
0.32
80.
284
0.08
90.
286
0.09
90.
300
0.78
4U
sed
Miar
juan
a (P
ast 2
wks
)0.
034
0.30
90.
070
0.86
00.
560
0.02
70.
283
0.02
00.
201
0.80
90.
040
0.33
20.
119
1.19
40.
517
Use
d Co
cain
e (P
ast 2
wks
)0.
007
0.14
20.
000
0.00
00.
318
0.00
00.
000
0.00
00.
000
.0.
013
0.20
00.
000
0.00
00.
318
Sexu
ally
Tran
smitt
ed D
iseas
es0.
333
0.47
20.
375
0.48
50.
301
0.33
00.
471
0.35
40.
480
0.68
80.
335
0.47
30.
396
0.49
20.
294
Wh
ole
Sam
ple
Fem
ale
Sam
ple
Mal
e Sa
mp
le
Notes:
This
table
pre
sents
the
stati
stic
al
desc
ripti
on
of
sele
cte
dpre
-pro
gra
mvari
able
saft
er
6years
of
the
pro
gra
m.
The
firs
tcolu
mn
of
the
table
giv
es
the
vari
able
desc
ripti
on.
Vari
able
sare
div
ided
into
gro
ups
that
share
sim
ilar
meanin
gs.
The
rem
ain
der
of
the
table
consi
sts
of
the
desc
ripti
on
of
the
blo
cks
of
vari
able
sass
ocia
ted
wit
hth
ew
hole
sam
ple
,th
efe
male
sam
ple
and
the
male
sam
ple
.E
ach
blo
ckhas
6colu
mns:
(1)
Contr
ol
mean
(CM
ean),
(2)
Contr
ol
standard
devia
tion
(CSD
),(3
)T
reatm
ent
mean
(TM
ean),
(4)
Tre
atm
ent
standard
devia
tion
(TSD
),and
(5)
Asy
mpto
ticp-v
alu
eass
ocia
ted
wit
hth
ediff
ere
nce
inm
eans.
Boldp-v
alu
es
indic
ate
that
the
t-st
ati
stic
betw
een
the
contr
ol
and
the
treatm
ent
means
issi
gnifi
cant
at
the
10%
level.
53
Tab
leD
.2:
Des
crip
tive
Sta
tist
icof
Base
lin
eC
hara
cter
isti
cs(Y
ear
12)
C M
ean
C SD
T M
ean
T SD
Pval
C M
ean
C SD
T M
ean
T SD
Pval
C M
ean
C SD
T M
ean
T SD
Pval
Back
groun
d Ch
arac
terist
icsM
ater
nal R
ace
(Blac
k)0.
057
0.23
20.
084
0.27
80.
244
0.05
60.
231
0.06
50.
248
0.77
00.
057
0.23
30.
101
0.30
30.
208
Mar
ital S
tatu
s (M
arrie
d)0.
014
0.11
90.
010
0.10
20.
690
0.00
50.
069
0.01
10.
104
0.60
30.
024
0.15
30.
010
0.10
10.
346
Mat
erna
l Age
18.0
523.
215
18.0
473.
268
0.98
618
.258
3.32
418
.174
3.58
10.
847
17.8
423.
093
17.9
292.
960
0.81
2Y
ears
of E
duca
tion
10.2
541.
860
10.0
732.
025
0.29
610
.324
1.82
810
.043
2.08
00.
265
10.1
821.
893
10.1
011.
982
0.73
5M
othe
r in
Scho
ol0.
599
0.49
10.
565
0.49
70.
444
0.55
70.
498
0.59
80.
493
0.50
50.
641
0.48
10.
535
0.50
10.
081
Hea
d of
Hou
seho
ld is
Em
ploy
ed0.
556
0.49
70.
495
0.50
10.
163
0.58
50.
494
0.47
80.
502
0.08
90.
526
0.50
10.
510
0.50
20.
793
% o
f Cen
sus T
ract
Belo
w P
over
ty34
.800
21.3
8035
.727
20.1
850.
606
33.6
3220
.150
37.2
0822
.390
0.18
935
.990
22.5
5034
.351
17.9
000.
492
Hou
seho
ld D
ensit
y0.
940
0.48
61.
023
0.55
90.
081
0.96
90.
483
1.04
90.
662
0.29
90.
911
0.48
80.
998
0.44
50.
123
Tota
l Hou
sehold
Incom
e (Pa
st 6
Mon
ths)
Less
than
$30
000.
280
0.44
90.
361
0.48
20.
048
0.27
70.
449
0.33
70.
475
0.30
50.
282
0.45
10.
384
0.48
90.
083
$300
0 - $
6999
0.24
20.
429
0.23
60.
425
0.87
00.
239
0.42
80.
239
0.42
90.
995
0.24
40.
431
0.23
20.
424
0.82
2$7
000
- $10
999
0.23
00.
421
0.18
80.
392
0.23
80.
230
0.42
20.
217
0.41
50.
808
0.23
00.
422
0.16
20.
370
0.15
1G
reat
er th
an $
1100
00.
159
0.36
60.
126
0.33
20.
269
0.17
80.
384
0.08
70.
283
0.02
20.
139
0.34
70.
162
0.37
00.
606
Inco
me,
No
Resp
onse
0.09
00.
287
0.08
90.
285
0.96
70.
075
0.26
40.
120
0.32
60.
251
0.10
50.
308
0.06
10.
240
0.16
6
Regio
n of
Resid
ence
Inne
r City
0.29
10.
455
0.28
30.
452
0.82
50.
282
0.45
10.
293
0.45
80.
836
0.30
10.
460
0.27
30.
448
0.60
3Bi
sson
0.19
40.
396
0.22
50.
419
0.39
10.
169
0.37
60.
239
0.42
90.
176
0.22
00.
415
0.21
20.
411
0.87
4Ca
wth
on0.
204
0.40
30.
188
0.39
20.
657
0.22
10.
416
0.15
20.
361
0.14
80.
187
0.39
10.
222
0.41
80.
477
Hol
lywoo
d0.
310
0.46
30.
304
0.46
10.
867
0.32
90.
471
0.31
50.
467
0.81
90.
292
0.45
60.
293
0.45
70.
985
Mat
ernal
Men
tal H
ealth M
ater
nal I
Q (S
hipl
ey)
96.0
669.
987
96.7
5910
.181
0.43
396
.075
10.0
0296
.011
10.7
890.
961
96.0
579.
997
97.4
559.
585
0.24
0M
ater
nal B
avol
ek S
core
99.9
477.
604
101.
078
8.56
80.
118
100.
190
7.48
910
1.42
78.
718
0.23
899
.701
7.72
910
0.75
48.
459
0.29
6M
ater
nal M
enta
l Hea
lth10
0.10
69.
744
99.5
5010
.612
0.53
899
.766
9.52
999
.909
10.4
290.
911
100.
451
9.96
899
.216
10.8
210.
339
Self-
Effi
cacy
99.8
139.
995
99.6
719.
912
0.87
010
0.64
09.
746
100.
192
9.33
90.
705
98.9
7310
.197
99.1
8610
.440
0.86
6M
ater
nal M
aste
ry10
0.05
910
.236
99.4
4610
.098
0.48
999
.954
10.3
0199
.085
10.5
330.
507
100.
165
10.1
9399
.781
9.71
80.
751
Mat
erna
l Psy
chol
ogica
l Res
ourc
es99
.857
9.65
299
.664
10.9
140.
834
99.9
479.
401
99.5
3710
.896
0.75
499
.765
9.92
399
.782
10.9
840.
990
Mat
ernal
Hea
lth C
hara
cteris
tics
Mat
erna
l Heig
ht16
4.59
57.
349
164.
303
6.68
00.
630
164.
297
7.48
316
4.90
46.
664
0.48
516
4.89
67.
217
163.
732
6.68
00.
170
Pre-
Preg
nanc
y W
eight
62.3
9815
.149
62.7
3513
.786
0.78
663
.078
14.0
7661
.880
12.3
690.
458
61.7
0116
.178
63.5
3015
.003
0.33
2G
esta
tiona
l Age
(Int
ake)
16.4
745.
830
16.6
075.
639
0.78
916
.235
5.79
116
.228
5.48
90.
993
16.7
185.
873
16.9
605.
780
0.73
3
Mat
ernal
Socia
l Sup
port
Gra
ndm
othe
r Soc
ial S
uppo
rt10
0.40
79.
331
101.
623
8.40
60.
110
99.6
2410
.361
101.
370
9.30
60.
148
101.
200
8.10
210
1.85
87.
514
0.48
5H
usba
nd/B
oyfr
iend
Socia
l Sup
port
100.
266
9.95
110
0.29
99.
986
0.96
910
0.02
39.
949
99.8
339.
700
0.87
710
0.51
19.
971
100.
731
10.2
750.
859
Mat
ernal
Risk
y Beh
avior
sA
lcoho
l Con
sum
ptio
n (P
ast 2
wks
)0.
040
0.19
70.
047
0.21
20.
710
0.03
80.
191
0.06
50.
248
0.34
50.
043
0.20
30.
030
0.17
20.
568
Smok
ing
(Pas
t 3 d
ays)
0.08
10.
273
0.10
50.
307
0.35
60.
075
0.26
50.
120
0.32
60.
255
0.08
60.
281
0.09
10.
289
0.89
1U
sed
Miar
juan
a (P
ast 2
wks
)0.
036
0.31
80.
073
0.88
00.
566
0.02
80.
291
0.02
20.
209
0.82
40.
043
0.34
40.
121
1.20
60.
528
Use
d Co
cain
e (P
ast 2
wks
)0.
007
0.14
60.
000
0.00
00.
318
0.00
00.
000
0.00
00.
000
.0.
014
0.20
80.
000
0.00
00.
318
Sexu
ally
Tran
smitt
ed D
iseas
es0.
347
0.47
70.
372
0.48
50.
554
0.33
50.
473
0.33
70.
475
0.97
20.
359
0.48
10.
404
0.49
30.
450
Wh
ole
Sam
ple
Fem
ale
Sam
ple
Mal
e Sa
mp
le
Notes:
This
table
pre
sents
the
stati
stic
al
desc
ripti
on
of
sele
cte
dpre
-pro
gra
mvari
able
saft
er
6years
of
the
pro
gra
m.
The
firs
tcolu
mn
of
the
table
giv
es
the
vari
able
desc
ripti
on.
Vari
able
sare
div
ided
into
gro
ups
that
share
sim
ilar
meanin
gs.
The
rem
ain
der
of
the
table
consi
sts
of
the
desc
ripti
on
of
the
blo
cks
of
vari
able
sass
ocia
ted
wit
hth
ew
hole
sam
ple
,th
efe
male
sam
ple
and
the
male
sam
ple
.E
ach
blo
ckhas
6colu
mns:
(1)
Contr
ol
mean
(CM
ean),
(2)
Contr
ol
standard
devia
tion
(CSD
),(3
)T
reatm
ent
mean
(TM
ean),
(4)
Tre
atm
ent
standard
devia
tion
(TSD
),and
(5)
Asy
mpto
ticp-v
alu
eass
ocia
ted
wit
hth
ediff
ere
nce
inm
eans.
Boldp-v
alu
es
indic
ate
that
the
t-st
ati
stic
betw
een
the
contr
ol
and
the
treatm
ent
means
issi
gnifi
cant
at
the
10%
level.
54
E Additional Inference Results: Addressing Attrition using In-
verse Propensity Weights
One aspect of the NFP that may cause concern is attrition. In order to address this we proceed as follows. We
estimate by OLS a linear regression model which dependent variable is binary and equal to one if the mother
represents an attrition case. We choose the model as to maximize the R2 and incorporate pre-program
variables with very few missing values. This permits us to establish the set of variables that provides the
maximum amount of information when it comes to estimate the probability of attrition. Then, we predict
the probability of attrition through a Logit model and apply an inverse probability weighting scheme to our
estimations. The results do not change much after this correction. Tables E.4–10 show these results. The
tables can be read in the same way as Tables 7–10 in the paper.
55
Tab
leE
.3:
Usi
ng
Logit
toO
bta
inP
rob
ab
ilit
ies
Sam
ple
Psy. Res.Pct. Pov.SmokerDrinkerEduc.Mom SupportHus./BF SupportRaceHH Emp.AgeAge SquaredSTDs# STDsMental HealthMasteryBavolekIncome (5 Cat.)Gest. Wks.HeightPre. Preg. Wt.Schooling
LR C
hi2
Prob
. > C
hi2
AIC
BIC
Trea
tmen
t Fem
ales
Yea
r 12
xx
xx
xx
xx
xx
xx
x25
.79
0.06
45.2
690
.70
Yea
r 9x
xx
xx
xx
xx
x22
.53
0.05
51.3
988
.81
Yea
r 6x
xx
xx
xx
xx
17.6
20.
0230
.361
54.4
17Y
ear 4
.5x
xx
xx
xx
xx
16.6
70.
0521
.291
48.0
19Y
ear 2
xx
xx
xx
xx
xx
24.1
80.
01-2
0.92
8.79
Cont
rol F
emale
s Yea
r 12
xx
xx
xx
xx
20.3
10.
0120
5.54
237.
16Y
ear 9
xx
xx
xx
xx
19.2
90.
0112
2.33
153.
81Y
ear 6
xx
xx
xx
x17
.84
0.01
103.
7213
1.73
Yea
r 4.5
xx
xx
xx
x14
.68
0.02
114.
5413
9.16
Yea
r 2x
xx
xx
xx
10.0
00.
12-1
9.62
5.01
Trea
tmen
t Male
s Yea
r 12
xx
xx
xx
xx
xx
22.9
70.
0410
6.90
144.
96Y
ear 9
xx
xx
9.38
0.05
82.5
596
.14
Yea
r 6x
xx
xx
xx
x17
.66
0.02
62.9
187
.37
Yea
r 4.5
xx
xx
xx
xx
xx
x14
.03
0.23
46.0
478
.66
Yea
r 2x
xx
xx
x15
.73
0.02
-3.5
415
.49
Cont
rol M
ales Y
ear 1
2x
xx
xx
xx
xx
32.0
70.
0015
5.12
200.
37Y
ear 9
xx
xx
xx
x37
.02
0.00
95.7
213
4.01
Yea
r 6x
xx
xx
xx
x34
.24
0.00
68.2
010
9.91
Yea
r 4.5
xx
xx
xx
xx
xx
22.0
30.
0212
1.86
160.
14Y
ear 2
xx
xx
12.3
80.
01-2
4.20
-6.6
3
Notes:
The
table
desc
rib
es
the
pre
-pro
gra
mvari
able
suse
dto
calc
ula
teth
ein
vers
epro
babilit
yw
eig
hts
.T
he
firs
tcolu
mn
pro
vid
es
the
four
div
isio
ngro
ups:
treatm
ent
fem
ale
s,contr
ol
fem
ale
s,tr
eatm
ent
male
s,and
contr
ol
male
s.A
ddit
ionally,
there
isa
corr
esp
ondin
gti
me
peri
od
for
each
row
.T
he
next
23
colu
mns
repre
sents
the
set
of
pre
-pro
gra
mch
ara
cte
rist
ics
that
were
use
dfo
rlo
git
.A
“x”
repre
sents
that
that
vari
able
was
use
dfo
rth
esp
ecifi
csa
mple
and
tim
ep
eri
od.
The
colu
mn
lab
ele
d“L
RC
hi2
”is
the
chi-
square
dcalc
ula
ted
usi
ng
the
logit
regre
ssio
nand
the
next
colu
mn,
“P
rob.>
Chi2
,”pro
vid
es
the
corr
esp
ondin
gp-v
alu
es.
The
last
two
colu
mns,
AIC
and
BIC
,pro
vid
es
the
Akaik
ein
form
ati
on
cri
teri
on
and
Bayesi
an
info
rmati
on
cri
teri
on
resp
ecti
vely
.
56
Tab
leE
.4:
Ch
ild
Hea
lth
Ou
tcom
es
Out
com
e D
escr
iptio
nC
ntr.
Mea
nC
d. D
iff. M
n.C
d. E
ff. S
ize
Asy
P-v
alSi
ngle
P-v
alSt
epdo
wn
Cnt
r. M
ean
Cd.
Diff
. Mn.
Cd.
Eff
. Siz
eA
sy P
-val
Sing
le P
-val
Step
dow
nBi
rth O
utco
mes
for C
hild
Plac
enta
Wei
ght
683.
488
-11.
638
-0.0
730.
707
0.46
70.
717
662.
401
27.9
650.
157
0.11
20.014
0.036
Birt
h W
eigh
t30
50.5
65-1
28.4
56-0
.235
0.96
60.
903
0.90
329
93.7
2620
4.97
70.
292
0.006
0.000
0.001
Hea
d C
ircum
fere
nce
33.2
570.
038
0.02
30.
425
0.20
30.
459
33.5
060.
327
0.14
60.
107
0.060
0.060
Leng
th49
.652
0.23
40.
087
0.23
60.
202
0.51
349
.908
0.71
10.
196
0.042
0.018
0.033
Ges
tatio
nal A
ge a
t Del
iver
y39
.092
-0.5
45-0
.242
0.94
00.
854
0.91
938
.526
0.74
50.
214
0.028
0.001
0.005
Chi
ld H
ealth
Out
com
es (Y
ear 1
2)A
ny I
njur
ies
sinc
e La
st I
nter
view
0.17
5-0
.043
-0.1
220.
171
0.21
60.
386
0.23
2-0
.059
-0.1
490.
132
0.12
00.
474
# H
ospi
taliz
atio
ns fo
r Inj
urie
s si
nce
Last
Int
ervi
ew0.
009
-0.0
11-0
.116
0.13
80.
185
0.45
10.
011
-0.0
13-0
.134
0.13
20.
170
0.58
2To
tal #
Inj
urie
s si
nce
Last
Int
ervi
ew0.
200
-0.0
68-0
.156
0.10
20.074
0.22
40.
278
-0.0
57-0
.110
0.21
20.
268
0.68
5H
ospi
taliz
ed s
ince
Las
t Int
ervi
ew0.
059
-0.0
44-0
.226
0.033
0.035
0.14
00.
040
0.05
40.
299
0.97
50.
890
0.89
0H
ave
Chr
onic
Con
ditio
n/H
ealth
Pro
blem
0.20
3-0
.003
-0.0
090.
473
0.63
90.
639
0.36
00.
077
0.16
30.
885
0.84
90.
965
Stan
dard
ized
Chi
ld B
MI
1.09
0-0
.240
-0.2
770.019
0.012
0.060
0.77
80.
224
0.25
70.
968
0.83
30.
988
Bas
ic S
tatis
tics
Blo
ck P
erm
. FL
Bas
ic S
tatis
tics
Blo
ck P
erm
. FL
Fem
ales
Mal
es
Notes:
The
firs
tcolu
mn
pro
vid
es
the
outc
om
edesc
ripti
on.
Our
resu
lts
are
pre
sente
din
six
colu
mns
for
each
gender.
The
firs
tcolu
mn
(Cntr
.M
ean)
of
each
resu
ltse
tsh
ow
sth
em
ean
for
the
contr
ol
gro
up.
When
facto
rsc
ore
sw
ere
com
pute
d,
we
set
the
mean
inth
econtr
ol
gro
up
tozero
.T
he
second
colu
mn
(Cd.
Diff
.M
n.)
giv
es
the
condit
ional
diff
ere
nce
inm
eans
betw
een
the
treatm
ent
gro
up
and
the
contr
ol
gro
up.
The
thir
dcolu
mn
(Cd.
Eff
.Siz
e)
calc
ula
tes
the
condit
ional
eff
ect
size
for
the
resp
ecti
ve
gro
up.
The
fourt
hcolu
mn
(Asy
.P
-val.)
pro
vid
es
the
asy
mpto
ticp-v
alu
efo
rth
eone-s
ided
single
hyp
oth
esi
ste
stass
ocia
ted
wit
hth
et-
stati
stic
for
the
diff
ere
nce
inm
eans
betw
een
treatm
ent
and
contr
ol
gro
ups.
As
menti
oned
inSecti
on
2,
the
contr
ol
gro
up
stands
for
the
ori
gin
al
treatm
ent
gro
up
2of
the
NF
Pexp
eri
ment
and
the
treatm
ent
gro
up
stands
for
the
ori
gin
al
gro
up
4.
The
fift
hcolu
mn
(Blo
ckP
erm
.F
L/Sin
gle
P-v
al.)
pre
sents
the
one-s
ided
rest
ricte
dp
erm
uta
tionp-v
alu
es
for
the
single
-hyp
oth
esi
ste
stin
gbase
don
thet-
stati
stic
ass
ocia
ted
wit
hth
etr
eatm
ent
indic
ato
rin
the
Fre
edm
an
and
Lane
(1983)
regre
ssio
nas
desc
rib
ed
inSecti
on
4.
By
rest
ricte
dp
erm
uta
tion
we
mean
that
perm
uta
tions
are
done
wit
hin
stra
tadefined
by
the
base
line
vari
able
suse
din
the
random
izati
on
pro
tocol:
mate
rnal
age
and
race,
gest
ati
onal
age
at
enro
llm
ent,
em
plo
ym
ent
statu
sof
the
head
of
the
house
hold
,and
geogra
phic
regio
n.
The
covari
ate
suse
din
the
Fre
edm
an
and
Lane
(1983)
regre
ssio
nare
:m
ate
rnal
heig
ht,
house
hold
incom
e,
gra
ndm
oth
er
supp
ort
,m
ate
rnal
pare
nti
ng
att
itudes
and
moth
er
curr
entl
yin
school.
Fin
ally,
the
last
colu
mn
(Blo
ckP
erm
.F
L/Ste
pdow
n)
pro
vid
esp-v
alu
es
that
account
for
mult
iple
-hyp
oth
esi
ste
stin
gbase
don
the
Ste
pdow
nalg
ori
thm
of
Rom
ano
and
Wolf
(2005a).
Blo
cks
of
outc
om
es
that
are
test
ed
join
tly
are
separa
ted
by
lines.
The
sele
cti
on
of
blo
cks
of
outc
om
es
isdone
on
the
basi
sof
their
meanin
g.
Outc
om
es
that
share
sim
ilar
meanin
gare
gro
up
ed
togeth
er.
Fem
ale
mate
rnal
outc
om
es
all
ude
tom
oth
ers
whose
firs
tch
ild
isa
gir
l.L
ikew
ise,
male
mate
rnal
outc
om
es
alludes
tom
oth
ers
whose
firs
tch
ild
isa
boy.
The
resu
lts
inth
ista
ble
use
an
invers
epro
babilit
yw
eig
hti
ng
schem
eto
addre
ssatt
riti
on.
The
weig
hts
are
base
don
the
pre
dic
ted
pro
babilit
yto
dro
pth
esa
mple
.T
he
pre
dic
tion
isbase
don
aL
ogit
model
that
isdesc
rib
ed
at
the
begin
nin
gof
this
secti
on.
57
Tab
leE
.5:
Fam
ily
Envir
on
men
t
Out
com
e D
escr
iptio
nC
ntr.
Mea
nC
d. D
iff. M
n.C
d. E
ff. S
ize
Asy
P-v
alSi
ngle
P-v
alSt
epdo
wn
Cnt
r. M
ean
Cd.
Diff
. Mn.
Cd.
Eff
. Siz
eA
sy P
-val
Sing
le P
-val
Step
dow
nH
ome E
nviro
nmen
t, Pa
rent
ing (
Year
1) -
Fac
tor S
core
sH
ome
Obs
erva
tion
Mea
sure
men
t of
the
Env
ironm
ent (
HO
ME
)0.
000
0.33
80.
338
0.004
0.004
0.004
-0.0
050.
154
0.15
40.
114
0.079
0.079
Non
-Abu
ssiv
e Pa
rent
ing
Atti
tude
s (B
avol
ek)
0.00
70.
294
0.29
30.010
0.003
0.006
-0.0
020.
364
0.36
40.002
0.001
0.002
Hom
e Env
ironm
ent,
Pare
ntin
g (Ye
ar 2
)- Fa
ctor S
core
sH
ome
Obs
erva
tion
Mea
sure
men
t of
the
Env
ironm
ent (
HO
ME
)0.
001
0.29
80.
297
0.010
0.004
0.007
-0.0
080.
116
0.11
60.
186
0.11
10.
111
Non
-Abu
ssiv
e Pa
rent
ing
Atti
tude
s (B
avol
ek)
0.01
20.
374
0.37
20.003
0.005
0.005
-0.0
050.
481
0.48
10.000
0.001
0.001
Mat
erna
l Men
tal H
ealth
(Yea
r 2)
Anx
iety
-0.0
01-0
.226
-0.2
260.042
0.038
0.086
0.01
2-0
.052
-0.0
520.
340
0.34
80.
633
Dep
ress
ion
0.00
0-0
.115
-0.1
150.
180
0.10
20.
169
0.01
0-0
.011
-0.0
110.
465
0.52
40.
692
Posi
tive
Wel
l-Bei
ng-0
.002
0.09
60.
096
0.22
20.
413
0.41
3-0
.006
-0.2
13-0
.214
0.95
00.
947
0.94
7E
mot
iona
l Sta
bilit
y0.
001
0.18
50.
185
0.076
0.056
0.11
3-0
.012
0.04
20.
042
0.36
70.
427
0.68
9O
vera
ll M
enta
l Hea
lth0.
000
0.19
30.
193
0.066
0.066
0.12
2-0
.011
-0.0
47-0
.047
0.64
40.
666
0.77
2Se
lf-E
stee
m0.
011
0.28
30.
283
0.014
0.003
0.014
-0.0
110.
045
0.04
50.
367
0.46
70.
707
Mas
tery
0.00
90.
251
0.25
00.030
0.018
0.057
-0.0
100.
253
0.25
20.026
0.040
0.13
7
Tota
l Cos
t of
Gov
t. Pr
ogra
ms (
Chi
ld A
ges 1
- 12
Yea
rs)
AFD
C/T
AN
F25
85.2
86-1
77.2
26-0
.070
0.28
00.
627
0.62
726
57.0
84-4
26.4
34-0
.165
0.073
0.087
0.15
6Fo
od S
tam
p29
00.6
13-3
74.6
02-0
.229
0.026
0.24
10.
362
3191
.672
-288
.782
-0.1
870.061
0.11
80.
155
Med
icai
d34
62.0
64-3
67.1
66-0
.221
0.035
0.27
50.
377
3747
.045
-271
.420
-0.1
830.068
0.15
30.
153
Fem
ales
Mal
esB
asic
Sta
tistic
sB
lock
Per
m. F
LB
asic
Sta
tistic
sB
lock
Per
m. F
L
Notes:
The
firs
tcolu
mn
pro
vid
es
the
outc
om
edesc
ripti
on.
Our
resu
lts
are
pre
sente
din
six
colu
mns
for
each
gender.
The
firs
tcolu
mn
(Cntr
.M
ean)
of
each
resu
ltse
tsh
ow
sth
em
ean
for
the
contr
ol
gro
up.
When
facto
rsc
ore
sw
ere
com
pute
d,
we
set
the
mean
inth
econtr
ol
gro
up
tozero
.T
he
second
colu
mn
(Cd.
Diff
.M
n.)
giv
es
the
condit
ional
diff
ere
nce
inm
eans
betw
een
the
treatm
ent
gro
up
and
the
contr
ol
gro
up.
The
thir
dcolu
mn
(Cd.
Eff
.Siz
e)
calc
ula
tes
the
condit
ional
eff
ect
size
for
the
resp
ecti
ve
gro
up.
The
fourt
hcolu
mn
(Asy
.P
-val.)
pro
vid
es
the
asy
mpto
ticp-v
alu
efo
rth
eone-s
ided
single
hyp
oth
esi
ste
stass
ocia
ted
wit
hth
et-
stati
stic
for
the
diff
ere
nce
inm
eans
betw
een
treatm
ent
and
contr
ol
gro
ups.
As
menti
oned
inSecti
on
2,
the
contr
ol
gro
up
stands
for
the
ori
gin
al
treatm
ent
gro
up
2of
the
NF
Pexp
eri
ment
and
the
treatm
ent
gro
up
stands
for
the
ori
gin
al
gro
up
4.
The
fift
hcolu
mn
(Blo
ckP
erm
.F
L/Sin
gle
P-v
al.)
pre
sents
the
one-s
ided
rest
ricte
dp
erm
uta
tionp-v
alu
es
for
the
single
-hyp
oth
esi
ste
stin
gbase
don
thet-
stati
stic
ass
ocia
ted
wit
hth
etr
eatm
ent
indic
ato
rin
the
Fre
edm
an
and
Lane
(1983)
regre
ssio
nas
desc
rib
ed
inSecti
on
4.
By
rest
ricte
dp
erm
uta
tion
we
mean
that
perm
uta
tions
are
done
wit
hin
stra
tadefined
by
the
base
line
vari
able
suse
din
the
random
izati
on
pro
tocol:
mate
rnal
age
and
race,
gest
ati
onal
age
at
enro
llm
ent,
em
plo
ym
ent
statu
sof
the
head
of
the
house
hold
,and
geogra
phic
regio
n.
The
covari
ate
suse
din
the
Fre
edm
an
and
Lane
(1983)
regre
ssio
nare
:m
ate
rnal
heig
ht,
house
hold
incom
e,
gra
ndm
oth
er
supp
ort
,m
ate
rnal
pare
nti
ng
att
itudes
and
moth
er
curr
entl
yin
school.
Fin
ally,
the
last
colu
mn
(Blo
ckP
erm
.F
L/Ste
pdow
n)
pro
vid
esp-v
alu
es
that
account
for
mult
iple
-hyp
oth
esi
ste
stin
gbase
don
the
Ste
pdow
nalg
ori
thm
of
Rom
ano
and
Wolf
(2005a).
Blo
cks
of
outc
om
es
that
are
test
ed
join
tly
are
separa
ted
by
lines.
The
sele
cti
on
of
blo
cks
of
outc
om
es
isdone
on
the
basi
sof
their
meanin
g.
Outc
om
es
that
share
sim
ilar
meanin
gare
gro
up
ed
togeth
er.
Fem
ale
mate
rnal
outc
om
es
all
ude
tom
oth
ers
whose
firs
tch
ild
isa
gir
l.L
ikew
ise,
male
mate
rnal
outc
om
es
alludes
tom
oth
ers
whose
firs
tch
ild
isa
boy.
The
resu
lts
inth
ista
ble
use
an
invers
epro
babilit
yw
eig
hti
ng
schem
eto
addre
ssatt
riti
on.
The
weig
hts
are
base
don
the
pre
dic
ted
pro
babilit
yto
dro
pth
esa
mple
.T
he
pre
dic
tion
isbase
don
aL
ogit
model
that
isdesc
rib
ed
at
the
begin
nin
gof
this
secti
on.
58
Tab
leE
.6:
Cogn
itiv
eA
bil
itie
san
dA
chie
vem
ent
Ou
tcom
es
Out
com
e D
escr
iptio
nC
ntr.
Mea
nC
d. D
iff. M
n.C
d. E
ff. S
ize
Asy
P-v
alSi
ngle
P-v
alSt
epdo
wn
Cnt
r. M
ean
Cd.
Diff
. Mn.
Cd.
Eff
. Siz
eA
sy P
-val
Sing
le P
-val
Step
dow
nK
aufm
an A
ssessm
ent B
atter
y for
Chi
ldre
n (Y
ear 6
)G
esta
lt C
losu
re9.
026
0.24
40.
081
0.26
60.
193
0.48
79.
787
-0.3
88-0
.134
0.83
70.
636
0.63
6H
and
Mov
emen
ts9.
267
0.43
80.
203
0.065
0.025
0.15
09.
319
0.12
70.
060
0.33
20.
398
0.71
1M
atrix
Ana
logi
es8.
636
0.13
60.
080
0.27
30.
300
0.57
98.
480
0.28
50.
180
0.092
0.12
40.
434
Num
ber R
ecal
l9.
390
0.43
70.
152
0.12
00.086
0.32
78.
886
1.00
40.
421
0.002
0.004
0.029
Phot
o Se
ries
7.04
00.
424
0.21
20.055
0.064
0.28
46.
791
0.03
00.
014
0.45
80.
496
0.70
0Sp
atia
l Mem
ory
8.44
10.
204
0.08
40.
264
0.34
10.
516
8.56
80.
218
0.09
00.
258
0.19
40.
531
Tria
ngle
s8.
845
0.43
50.
188
0.070
0.12
90.
402
9.20
10.
094
0.04
10.
382
0.21
30.
522
Wor
d O
rder
9.73
7-0
.079
-0.0
300.
591
0.38
60.
386
9.14
80.
763
0.29
80.016
0.006
0.039
Kau
fman
Asse
ssmen
t Bat
tery f
or C
hild
ren
(Yea
r 6)
Non
verb
al89
.267
2.11
80.
239
0.041
0.051
0.10
489
.466
1.10
40.
118
0.19
60.
187
0.23
5Se
quen
tial P
roce
ssin
g96
.587
1.68
50.
131
0.16
00.071
0.12
494
.353
3.67
60.
312
0.013
0.011
0.023
Sim
ulta
neou
s Pr
oces
sing
88.9
811.
980
0.19
60.072
0.094
0.094
90.1
280.
498
0.05
00.
359
0.23
10.
231
WIS
C-II
I, PP
VT-
III f
or C
hild
ren
(Yea
r 6)
Wec
hsle
r Int
ellig
ence
Sca
le fo
r Chi
ldre
n (W
ISC
-III
)96
.518
0.90
00.
050
0.34
80.
352
0.35
290
.692
1.74
60.
102
0.22
70.
298
0.29
8Pe
abod
y Pi
ctur
e V
ocab
ular
y Te
st (P
PVT-
III)
83.6
821.
685
0.15
40.
119
0.16
40.
286
82.6
952.
325
0.22
10.062
0.013
0.024
Chi
ld C
ogni
tion
(Yea
r 6)
- Fac
tor S
core
sC
ogni
tion
+ a
chie
vmen
t (K
AB
C, P
PVT,
WIS
C)
0.00
50.
118
0.11
80.
188
0.067
0.092
-0.0
100.
182
0.18
20.092
0.063
0.063
Cog
nitiv
e sk
ills
(Men
tal P
roce
ssin
g C
ompo
site
-KA
BC
)0.
000
0.13
70.
137
0.15
00.073
0.073
-0.0
070.
277
0.27
70.023
0.015
0.021
Rea
ding
Ach
ievem
ent f
or th
e Chi
ld (Y
ear 1
2)A
vera
ge R
eadi
ng G
rade
(Gra
des
1 - 5
)2.
694
0.07
60.
101
0.22
80.
107
0.27
12.
348
0.05
80.
078
0.29
60.100
0.16
4T
CA
P %
Lan
guag
e (S
choo
l Yea
rs 1
- 5,
Grd
3+
)51
.600
0.37
20.
016
0.45
60.
180
0.30
737
.918
5.07
60.
233
0.067
0.005
0.021
TC
AP
% R
eadi
ng (S
choo
l Yea
rs 1
- 5,
Grd
3+
)42
.374
0.19
70.
010
0.47
30.
164
0.31
035
.020
1.75
00.
088
0.28
00.043
0.11
7PI
AT
Tot
al R
eadi
ng (D
eriv
ed S
core
)90
.420
0.66
20.
069
0.30
70.
344
0.40
489
.350
1.38
10.
103
0.22
10.063
0.12
9PI
AT
Rea
ding
Com
preh
ensi
on (D
eriv
ed S
core
)88
.458
-0.2
32-0
.026
0.57
60.
546
0.54
687
.641
2.36
90.
203
0.072
0.022
0.070
PIA
T R
eadi
ng R
ecog
nitio
n (D
eriv
ed S
core
)94
.486
2.17
50.
180
0.10
20.
156
0.32
592
.620
0.26
60.
018
0.44
70.
136
0.13
6
Mat
h A
chiev
emen
t for
the C
hild
(Yea
r 12)
Ave
rage
Mat
h G
rade
(Gra
des
1 - 5
)2.
622
0.09
30.
113
0.19
60.
146
0.27
02.
391
0.09
90.
130
0.18
30.072
0.072
TC
AP
% M
ath
(Sch
ool Y
ears
1 -
5, G
rd 3
+)
47.6
101.
896
0.08
00.
291
0.18
80.
279
40.1
763.
086
0.13
90.
185
0.033
0.082
PIA
T M
athe
mat
ics
(Der
ived
Sco
re)
87.4
13-0
.080
-0.0
080.
525
0.72
70.
727
86.5
381.
947
0.19
30.086
0.048
0.085
Fem
ales
Mal
esB
asic
Sta
tistic
sB
lock
Per
m. F
LB
asic
Sta
tistic
sB
lock
Per
m. F
L
Notes:
The
firs
tcolu
mn
pro
vid
es
the
outc
om
edesc
ripti
on.
Our
resu
lts
are
pre
sente
din
six
colu
mns
for
each
gender.
The
firs
tcolu
mn
(Cntr
.M
ean)
of
each
resu
ltse
tsh
ow
sth
em
ean
for
the
contr
ol
gro
up.
When
facto
rsc
ore
sw
ere
com
pute
d,
we
set
the
mean
inth
econtr
ol
gro
up
tozero
.T
he
second
colu
mn
(Cd.
Diff
.M
n.)
giv
es
the
condit
ional
diff
ere
nce
inm
eans
betw
een
the
treatm
ent
gro
up
and
the
contr
ol
gro
up.
The
thir
dcolu
mn
(Cd.
Eff
.Siz
e)
calc
ula
tes
the
condit
ional
eff
ect
size
for
the
resp
ecti
ve
gro
up.
The
fourt
hcolu
mn
(Asy
.P
-val.)
pro
vid
es
the
asy
mpto
ticp-v
alu
efo
rth
eone-s
ided
single
hyp
oth
esi
ste
stass
ocia
ted
wit
hth
et-
stati
stic
for
the
diff
ere
nce
inm
eans
betw
een
treatm
ent
and
contr
ol
gro
ups.
As
menti
oned
inSecti
on
2,
the
contr
ol
gro
up
stands
for
the
ori
gin
al
treatm
ent
gro
up
2of
the
NF
Pexp
eri
ment
and
the
treatm
ent
gro
up
stands
for
the
ori
gin
al
gro
up
4.
The
fift
hcolu
mn
(Blo
ckP
erm
.F
L/Sin
gle
P-v
al.)
pre
sents
the
one-s
ided
rest
ricte
dp
erm
uta
tionp-v
alu
es
for
the
single
-hyp
oth
esi
ste
stin
gbase
don
thet-
stati
stic
ass
ocia
ted
wit
hth
etr
eatm
ent
indic
ato
rin
the
Fre
edm
an
and
Lane
(1983)
regre
ssio
nas
desc
rib
ed
inSecti
on
4.
By
rest
ricte
dp
erm
uta
tion
we
mean
that
perm
uta
tions
are
done
wit
hin
stra
tadefined
by
the
base
line
vari
able
suse
din
the
random
izati
on
pro
tocol:
mate
rnal
age
and
race,
gest
ati
onal
age
at
enro
llm
ent,
em
plo
ym
ent
statu
sof
the
head
of
the
house
hold
,and
geogra
phic
regio
n.
The
covari
ate
suse
din
the
Fre
edm
an
and
Lane
(1983)
regre
ssio
nare
:m
ate
rnal
heig
ht,
house
hold
incom
e,
gra
ndm
oth
er
supp
ort
,m
ate
rnal
pare
nti
ng
att
itudes
and
moth
er
curr
entl
yin
school.
Fin
ally,
the
last
colu
mn
(Blo
ckP
erm
.F
L/Ste
pdow
n)
pro
vid
esp-v
alu
es
that
account
for
mult
iple
-hyp
oth
esi
ste
stin
gbase
don
the
Ste
pdow
nalg
ori
thm
of
Rom
ano
and
Wolf
(2005a).
Blo
cks
of
outc
om
es
that
are
test
ed
join
tly
are
separa
ted
by
lines.
The
sele
cti
on
of
blo
cks
of
outc
om
es
isdone
on
the
basi
sof
their
meanin
g.
Outc
om
es
that
share
sim
ilar
meanin
gare
gro
up
ed
togeth
er.
Fem
ale
mate
rnal
outc
om
es
all
ude
tom
oth
ers
whose
firs
tch
ild
isa
gir
l.L
ikew
ise,
male
mate
rnal
outc
om
es
alludes
tom
oth
ers
whose
firs
tch
ild
isa
boy.
The
resu
lts
inth
ista
ble
use
an
invers
epro
babilit
yw
eig
hti
ng
schem
eto
addre
ssatt
riti
on.
The
weig
hts
are
base
don
the
pre
dic
ted
pro
babilit
yto
dro
pth
esa
mple
.T
he
pre
dic
tion
isbase
don
aL
ogit
model
that
isdesc
rib
ed
at
the
begin
nin
gof
this
secti
on.
59
Tab
leE
.7:
Soci
o-
Em
oti
on
al
Ab
ilit
ies
Out
com
e D
escr
iptio
nC
ntr.
Mea
nC
d. D
iff. M
n.C
d. E
ff. S
ize
Asy
P-v
alSi
ngle
P-v
alSt
epdo
wn
Cnt
r. M
ean
Cd.
Diff
. Mn.
Cd.
Eff
. Siz
eA
sy P
-val
Sing
le P
-val
Step
dow
nC
hild
Beh
avior
Che
cklis
t (Ye
ar 2
) - F
acto
r Sco
res
Aff
ectiv
e Pr
oble
ms
-0.0
01-0
.337
-0.3
370.002
0.003
0.015
0.00
40.
287
0.28
70.
985
0.95
50.
955
Anx
iety
Pro
blem
s-0
.002
-0.1
81-0
.181
0.066
0.24
90.
249
0.00
70.
016
0.01
60.
550
0.63
60.
907
Perv
asio
n D
evel
opm
enta
l Pro
blem
s-0
.005
-0.2
61-0
.261
0.013
0.060
0.100
0.00
50.
185
0.18
50.
925
0.81
70.
950
Atte
ntio
n D
efic
it H
yper
activ
ity D
isor
der
-0.0
01-0
.243
-0.2
420.025
0.019
0.060
0.00
30.
056
0.05
60.
670
0.70
60.
923
Opp
ositi
onal
Def
iant
Pro
blem
s-0
.001
-0.2
17-0
.217
0.040
0.053
0.12
00.
005
0.12
60.
126
0.85
30.
880
0.96
2
Chi
ld B
ehav
ior C
heck
list (
Year
6) -
Fac
tor S
core
sA
ffec
tive
Prob
lem
s-0
.010
-0.0
07-0
.007
0.47
90.
612
0.79
6-0
.004
-0.1
03-0
.103
0.20
30.
151
0.48
1A
nxie
ty P
robl
ems
-0.0
08-0
.061
-0.0
610.
306
0.49
20.
759
0.00
90.
083
0.08
20.
729
0.81
30.
813
Som
atic
Pro
blem
s-0
.003
0.13
00.
130
0.83
20.
884
0.88
40.
007
0.06
30.
063
0.67
80.
442
0.75
7A
ttent
ion
Def
icit
Hyp
erac
tivity
Pro
blem
s-0
.012
-0.2
30-0
.230
0.035
0.096
0.30
7-0
.006
-0.0
40-0
.040
0.37
90.
310
0.71
3O
ppos
ition
al D
efia
nt P
robl
ems
0.00
0-0
.027
-0.0
270.
415
0.28
60.
608
-0.0
13-0
.083
-0.0
830.
270
0.31
70.
672
Con
duct
Pro
blem
s-0
.002
-0.2
67-0
.266
0.013
0.003
0.015
-0.0
09-0
.011
-0.0
110.
467
0.48
50.
665
Mac
Arth
ur S
tory
Stem
Bat
tery (
MSS
B) (Y
ear 6
) - F
acto
r Sco
res
Dys
regu
late
d A
ggre
ssio
n-0
.006
-0.0
27-0
.027
0.41
30.
135
0.26
9-0
.009
-0.1
30-0
.130
0.18
90.
137
0.49
6W
arm
th a
nd E
mpa
thy
-0.0
110.
388
0.38
80.002
0.005
0.019
-0.0
11-0
.099
-0.0
990.
770
0.53
50.
832
Em
otio
nal I
nteg
ratio
n-0
.005
-0.0
28-0
.028
0.58
50.
765
0.76
5-0
.015
0.05
50.
055
0.34
90.
429
0.84
9Pe
rfor
man
ce A
nxie
ty0.
010
-0.0
38-0
.038
0.37
30.093
0.25
9-0
.009
0.07
70.
077
0.70
10.
843
0.84
3A
ggre
ssio
n-0
.005
-0.1
64-0
.164
0.084
0.003
0.012
-0.0
10-0
.095
-0.0
950.
260
0.17
70.
570
Inter
naliz
ing,
Ext
erna
lizin
g, A
bsen
ces (Y
ear 1
2) Inte
rnal
izin
g D
isor
ders
0.24
0-0
.028
-0.0
660.
309
0.45
30.
813
0.39
7-0
.087
-0.1
830.090
0.082
0.15
4E
xter
naliz
ing
Dis
orde
rs0.
183
-0.0
13-0
.032
0.40
20.
641
0.86
60.
183
0.08
90.
239
0.95
10.
859
0.85
9A
vera
ge #
of
Abs
ence
s (S
choo
l Yea
rs 1
- 5)
10.1
440.
263
0.03
50.
605
0.66
60.
666
11.5
48-1
.838
-0.2
460.029
0.027
0.077
Fem
ales
Mal
esB
asic
Sta
tistic
sB
lock
Per
m. F
LB
asic
Sta
tistic
sB
lock
Per
m. F
L
Notes:
The
firs
tcolu
mn
pro
vid
es
the
outc
om
edesc
ripti
on.
Our
resu
lts
are
pre
sente
din
six
colu
mns
for
each
gender.
The
firs
tcolu
mn
(Cntr
.M
ean)
of
each
resu
ltse
tsh
ow
sth
em
ean
for
the
contr
ol
gro
up.
When
facto
rsc
ore
sw
ere
com
pute
d,
we
set
the
mean
inth
econtr
ol
gro
up
tozero
.T
he
second
colu
mn
(Cd.
Diff
.M
n.)
giv
es
the
condit
ional
diff
ere
nce
inm
eans
betw
een
the
treatm
ent
gro
up
and
the
contr
ol
gro
up.
The
thir
dcolu
mn
(Cd.
Eff
.Siz
e)
calc
ula
tes
the
condit
ional
eff
ect
size
for
the
resp
ecti
ve
gro
up.
The
fourt
hcolu
mn
(Asy
.P
-val.)
pro
vid
es
the
asy
mpto
ticp-v
alu
efo
rth
eone-s
ided
single
hyp
oth
esi
ste
stass
ocia
ted
wit
hth
et-
stati
stic
for
the
diff
ere
nce
inm
eans
betw
een
treatm
ent
and
contr
ol
gro
ups.
As
menti
oned
inSecti
on
2,
the
contr
ol
gro
up
stands
for
the
ori
gin
al
treatm
ent
gro
up
2of
the
NF
Pexp
eri
ment
and
the
treatm
ent
gro
up
stands
for
the
ori
gin
al
gro
up
4.
The
fift
hcolu
mn
(Blo
ckP
erm
.F
L/Sin
gle
P-v
al.)
pre
sents
the
one-s
ided
rest
ricte
dp
erm
uta
tionp-v
alu
es
for
the
single
-hyp
oth
esi
ste
stin
gbase
don
thet-
stati
stic
ass
ocia
ted
wit
hth
etr
eatm
ent
indic
ato
rin
the
Fre
edm
an
and
Lane
(1983)
regre
ssio
nas
desc
rib
ed
inSecti
on
4.
By
rest
ricte
dp
erm
uta
tion
we
mean
that
perm
uta
tions
are
done
wit
hin
stra
tadefined
by
the
base
line
vari
able
suse
din
the
random
izati
on
pro
tocol:
mate
rnal
age
and
race,
gest
ati
onal
age
at
enro
llm
ent,
em
plo
ym
ent
statu
sof
the
head
of
the
house
hold
,and
geogra
phic
regio
n.
The
covari
ate
suse
din
the
Fre
edm
an
and
Lane
(1983)
regre
ssio
nare
:m
ate
rnal
heig
ht,
house
hold
incom
e,
gra
ndm
oth
er
supp
ort
,m
ate
rnal
pare
nti
ng
att
itudes
and
moth
er
curr
entl
yin
school.
Fin
ally,
the
last
colu
mn
(Blo
ckP
erm
.F
L/Ste
pdow
n)
pro
vid
esp-v
alu
es
that
account
for
mult
iple
-hyp
oth
esi
ste
stin
gbase
don
the
Ste
pdow
nalg
ori
thm
of
Rom
ano
and
Wolf
(2005a).
Blo
cks
of
outc
om
es
that
are
test
ed
join
tly
are
separa
ted
by
lines.
The
sele
cti
on
of
blo
cks
of
outc
om
es
isdone
on
the
basi
sof
their
meanin
g.
Outc
om
es
that
share
sim
ilar
meanin
gare
gro
up
ed
togeth
er.
Fem
ale
mate
rnal
outc
om
es
all
ude
tom
oth
ers
whose
firs
tch
ild
isa
gir
l.L
ikew
ise,
male
mate
rnal
outc
om
es
alludes
tom
oth
ers
whose
firs
tch
ild
isa
boy.
The
resu
lts
inth
ista
ble
use
an
invers
epro
babilit
yw
eig
hti
ng
schem
eto
addre
ssatt
riti
on.
The
weig
hts
are
base
don
the
pre
dic
ted
pro
babilit
yto
dro
pth
esa
mple
.T
he
pre
dic
tion
isbase
don
aL
ogit
model
that
isdesc
rib
ed
at
the
begin
nin
gof
this
secti
on.
60
F Theoretical Framework for Mediation Analysis
This section develops a theoretical framework that helps to interpret the estimates of a standard mediation
model for early childhood interventions. Our model is based on a technology of skill formation according
to (Cunha and Heckman, 2007b). In it, later skills build on previous ones to generate human capital.
Notationally, let θi,t be the vector of skills during childhood for individual i at period t and t ∈ 0, 1, . . . , T,
where T is the number of periods of childhood. Let Ii,t represent investments at the same period. We use
Xi for family background characteristics and υi,t for an exogenous error term independent of θi,t, Ii,t and
Xi. The structural equation that governs the evolution of skills is given by:
θi,t+1 = qt+1(θi,t, Ii,t+1, Xi, υi,t+1); t ∈ 0, 1, . . . , T − 1. (32)
By structural equation, we mean autonomous functions in the language of Frisch (1938), i.e. deterministic
functions whose functional form do not change as its arguments vary. We also allow for skills to affect
investments, that is:
Ii,t+1 = ht+1(θi,t, Xi, εi,t+1); t ∈ 0, 1, . . . , T − 1, (33)
where εi,t+1 is an exogenous error term independent of θi,t and Xi. Our model is completed by the following
structural outcome equation at period T :
Yi = gT (θi,T , Xi, ξi,T ). (34)
where ξi,T is an exogenous error term independent of θi,T and Xi.
We can use a recursive substitution of investments and skills of Equations (32)–(33) into (34) to generate
the following equation:
Yi = ft′(θi,t′ , Xi, υi,tTt=t′ , εi,tTt=t′ , ξi,T ), (35)
where υi,tTt=t′ = υi,t′ , υi,t′+1, . . . , υi,T and εi,tTt=t′ = εi,t′ , εi,t′+1, . . . , εi,T .
Suppose that an intervention occurs at period t′ where t′ ∈ 1, . . . , T. Let Di ∈ 0, 1 be the treatment
indicator of this intervention which takes value 1 if participant i is treated and 0 otherwise. The intervention
enters our technology of skill formation model as a form of skill investment. Thus we append the investment
Equation (33) at period t′ by:
Ii,t′ = ht′(θi,t′−1, Di, Xi, εi,t′); for some t′ ∈ 0, 1, . . . , T − 1, (36)
61
The counterfactual values investment Ii,t′ are defined by the value Ii,t′ takes when the intervention Di is
fixed at a level d ∈ 0, 1. By fixing, I mean the causal operation defined in Haavelmo (1944) where Di is
set to d ∈ 0, 1 as argument in the structural equation (36). That is:
Ii,t′,d = ht′(θi,t′−1, d,Xi, εi,t′); d ∈ 0, 1 for some t′ ∈ 0, 1, . . . , T − 1. (37)
Let the counterfactual skills be defined in a symmetric fashion by:
θi,t′,d = qt′(θi,t′−1, Ii,t′,d, Xi, υi,t′).
We also define the counterfactual skills and investments for periods t > t′ by:
Ii,t+1,d = ht+1(θi,t,d, Xi, εi,t+1), and
θi,t+1,d = qt+1(θi,t,d, Ii,t+1,d, Xi, υi,t+1); t > t′.
We can also define the counterfactual outcomes by:
Yi,d = ft′(θi,t′,d, Xi, εi,tTt=t′ , εi,tTt=t′ , ξi,T ), (38)
If the intervention assignment uses the method of randomization, then we have that:
(Yi,d,θi,t′,d) ⊥⊥ Di|Xi; d ∈ 0, 1.
We can also write the realized values of skills and outcomes as:
Yi = Yi,1Di + Yi,0(1−Di), and
θi,t = θi,t,1Di + θi,t,0(1−Di); t ≥ t′.
We now cast on Equation (38) to generate a tractable equation to examine mediation effects. Note that
Equation (38) holds ont onlt for t′ but for any t ≥ t′.
Yi,d = ft(θi,t,d, Xi, υi,tTt=t, εi,tTt=t, ξi,T ), for any t ∈ t′, t′ + 1, . . . , T. (39)
Error terms (υi,tTt=t, εi,tTt=t, ξi,T ) are independent of θi,t,d and Xi. For sake of notational simplicity, we
62
can substitute those error terms by ζt without loss of generality. Equation (39) then becomes:
Yi,d = ft(θi,t,d, Xi, ζi,t). (40)
We achieve a linear form of Equation (40) by approximating it through a Maclaurin expansion. This generates
the following equation:
Yi,d = κt +αt,dθi,t,d + βt,dXi + εi,t,d, d ∈ 0, 1. (41)
where εt,d accounts for the approximation error. Equations (40)–(41) are used in our mediation analysis in
Section 5.
G Mediation Methodology
G.1 Three Step Procedure
This part of the appendix explains in detail the three step procedure that we use in order to decompose the
NFP treatment effects. As highlighted in the paper, we perform two sets of analysis. First, we study if the
treatment effects on child skills at age 6 were mediated by program enhancement of birth weight, parenting
attitudes and investments, and maternal socio-emotional skills at age 2. Second, we study if the program
impact on outcomes at age 12 was mediated by the NFP enhancement of skills at age 6. The results from
these analysis shed light on the complementarity of investments and skills in explaining the NFP treatment
effects.
Step One The idea is to develop a measurement system that links the observed items and the latent
skills. In order to do that, we assume that our measurements are dedicated. This means that each observed
measurement is linked to a unique skill. Specifically, let Mj be the index set of measures associated with
trait j, where j ∈ J = P,C, SE. P,C, SE denote, respectively, parenting skills, child cognitive skills, and
child socio-emotional abilities.18 Thus, our linear measurement system looks as follows:19
M jmj ,d = νjmj + ϕjmjθ
jd + ηjmj ,d, (42)
where νjmj is the intercept term and ϕjmj represents the loading factor of trait j. We cannot reject the
null hypothesis that the intercepts and loading factors depend on treatment status. ηmj ,d is a mean zero
18This follows the same notation as Heckman et al. (2013)19We control for pre-program variables X but we keep it implicit to shorten notation.
63
idiosyncratic error term which, by assumption, is independent of θjd ∀j ∈ J . We normalize the loading factor
associated with the first measure of each factor to 1 to set a scale, otherwise the scale is arbitrary.20 Finally,
we allow for factor correlation.
The parameters that identify the measurement system are the factor means, the factor covariances, the
intercepts, the factor loadings, and the variances of the error terms: E[θj(d)] = µjd, V ar[θd] = Σθd , νjmj ,
ϕjmj , V ar[ηjmj ]). Heckman et al. (2013) show that the existence of at least three measures for each latent
skill guarantees identification.21 Broadly, means, variances, and covariances across the measures identify the
parameters of the system.
We estimate the parameters of the measurement system that links skills with measures both at ages 2
and 6. Variables become potential mediators if we estimate an effect of the NFP on it, so that they are
potential meaningful channels. For age 2, non-abusive parenting attitudes are approximated by the Adult-
Adolescent Parenting Inventory (Bavolek), which comprises 32 items, and home investments are measured
by the Bradley and Caldwell Home Observation for measurement of the Environment (HOME) inventory,
which is composed by 45 items.
The maternal skills selected correspond to anxiety, assessed by the Rand Mental Health Inventory, self-
esteem, measured by the Rosemberg scale, and mastery, approximated by the Pearlin scale. Similarly, for age
6, we select as plausible mediators children’s skills influenced by the NFP. Children cognition is measured by
8 subtests from the K-ABC mental processing composite. For children’s socio-emotional skills, we identify
as potential mediators the treatment reduction in conduct, attention and aggression problems, as well as
the enhancement of children’s pro-social skills. Attention and conduct problems are approximated by items
from the Child Behavior Checklist. Pro-social skills (warmth or empathy) and aggression problems are
approximated by items from the MacArthur Story Stem Battery. Section B of the Appendix explains in
more detail these tests, as well as the instruments they use.
We estimate the parameters of the measurement system by maximum likelihood. In order to do this, we
assume that the latent skills and the error terms, , θj , ηjmj , are normal and i.i.d. We use full-information
maximum likelihood to deal with the missing values in the measures for some individuals. Recent work by
citation shows that FIML yields unbiased estimates that are more efficient than ad hoc methods like list-wise
and pair-wise deletion, which work under the implicit assumption of random missing data.22
For the case of the measurement system at age 2, we have 146 items. Although it is ideal to estimate
the complete set of items (skills) jointly, it is not feasible. Thus, we estimate them in two blocks: one for
20Given that the first measure sets the scale, we choose it to be the most correlated with the skill. The results are robust toalterations of this.
21 Carneiro et al. (2003) and Cunha et al. (2010) also discuss identification of factor models.22Missing at random means that the probability of missing data in a variable x can depend on other observed variables but
not on the values of x itself
64
parenting and home investments and other for maternal characteristics. This allows us to account for the
correlation between the skills that are in the same block. For the case of the measurement system at age 6,
the set of items is smaller and we do a joint estimation.
Step Two In the second step we use the parameter estimates from the first step to construct factor
scores for each children. The objective of this is to construct approximations for the latent skills. The two
most common linear scoring methods are the regression method (Bartlett, 1937; Thomson, 1934) and the
Bartlett method, which resembles GLS. We use the Bartlett (1937) method because it estimates unbiased
approximations of the unobserved skills. Actually, this guarantees that the difference in means between the
factor scores for children in the treatment and the control groups equals the difference in means in the true
scores. The derivation of the Bartlett estimator begins with the measurement system summarized as:
Mi︸︷︷︸|M|×1
= ϕ︸︷︷︸|M|×|J |
θi︸︷︷︸|J |×1
+ ηi︸︷︷︸|M|×1
where the dimension of each term is below the braces (recall that J and M are the indexing sets for skills
and measures respectively). Assume that the (θi,ηi), i ∈ 1, . . . , I, are independent across i. For simplicity,
we assume that they are i.i.d.23 Let Cov(Mi,Mi) = Σ, Cov(θi,θi) = Φ and Cov(ηi,ηi) = Ω. The linear
relation between the factor scores and the measures is the following:
θS,i = L′Mi (43)
In order to obtain unbiased estimates, Barlett imposes the restriction that L′ϕ = I|J |. The Bartlett
estimator for the vector of approximated skills (θi) is:
θS,i = (ϕ′Ω−1ϕ)−1ϕ′Ω−1Mi, (44)
where the matrix of loading factors, ϕ, and Ω = Cov(ηi,ηi) are both estimated in the first step. Bartlett’s
estimator is a Generalized Least Squares, GLS, procedure where measures are used as dependent variables
and loading factors are treated as regressors. By the Gauss-Markov theorem, the Bartlett GLS estimator is
optimal and hence leads to the best linear unbiased predictor (BLUE).
There are individuals that have missing data in some of the items that compose the measurement system.
In order to take advantage of the information that they have (instead of list-wise delete them), we predict
factor scores for them. We use the covariance between the measures and the factors from the sample with
23This is not strictly required but simplifies the notation.
65
complete measurement system to predict scores for these people. Additionally, for the cases were individuals
are missing a factor score because they did not have any item in that measurement system, we impute factor
scores with the regression method.24 This procedure recovers around 10% of the randomized sample.
Step 3 In this step, we use factor scores as approximations of the true skills to estimate the models that
link the later outcomes with the intermediate skills. The factor scores are measured with error, which
produces downward biased estimates of the parameters of the outcome equations. This bias corresponds to
the traditional attenuation that results from classical measurement error. In factor scored regressions, Bolck
et al. (2008) prove this. We adopt the bias correction strategy proposed by Croon (2002). In summary, this
approach takes advantage of the fact that we have estimates of all the components of the bias. This strategy,
also used by Heckman et al. (2013), can be summarized as follows:
Consider the model following model. To simplify notation, we use W to denote pre-program variables
X, treatment indicator and the intercept of equation 8:
Yi = αθi + γWi + εi, i = 1, . . . , N. (45)
The covariance matrix of (θi,Wi) is
Cov(θ,θ) Cov(θ,W )
Cov(W ,θ) Cov(W ,W )
.
We measure θi with error. Thus,
θS,i = θi + Vi, i = 1, . . . , N
(Wi,θi) ⊥⊥ Vi, E(Vi) = 0, Cov(V ,V ) = ΣV V
Denote Cov(θS,i,θS,i) = ΣθS,θS . We assume that the (θi,Wi, εi) are i.i,d, but much weaker conditions
suffice. Note that we do not assume that θi ⊥⊥ Wi as in traditional factor analysis. We do assume that
(θi,Wi) ⊥⊥ εi and E(εi) = 0.
If we use θS,i in place of Yi, it follows that:
Yi = αθS,i + γWi + εi − αVi. (46)
24We impute factor scores for individuals that have at least two other factor scores.
66
The estimation of equation 46 using OLS produces estimates that are biased:
plim
α
γ
=
Cov(θS ,θS) Cov(θS ,W )
Cov(W ,θS) Cov(W ,W )
−1 Cov(θ,θ) Cov(θ,W )
Cov(W ,θ) Cov(W ,W )
α
γ
.
Let ΣB,C be Cov(B,C). Observe that Σθ,W = ΣθS ,W as a consequence of our assumptions. In this
notation
plim
α
γ
=
Σθ,θ + ΣV ,V Σθ,W
ΣW ,θ ΣW ,W
−1 Σθ,θ Σθ,W
ΣW ,θ ΣW ,W
︸ ︷︷ ︸
A
α
γ
(47)
which is the usual attenuation formula.
From the estimation of the measurement system, we can identify Σθ,θ, Σθ,W , ΣV ,V , and we have all
the components of A. Hence if we pre-multiply the least squares estimator by A−1, we obtain:
plimA−1
α
γ
=
α
γ
.
This is called “Croon’s method” in psychometrics (Croon, 2002). In our application, there are two groups
corresponding to D = 0 and D = 1 (control and treatment, respectively). We allow θi to vary by treatment
status. Indeed, our method assumes that treatment only operates through shifting the distribution of θ. We
do not normalize the means of θ (or W ) to be zero.
In the third step of our estimation procedure we compute bootstrapped p-values for each decomposition
channel of the treatment effects. We take 100,000 resamples with replacement. The bootstrapped p-value
for the null hypothesis H0 : αj = 0 is calculated as follows:
p-value =1
B
B∑b=1
1(tj,∗b > tj) with tj =αj
σ(αj)and tj,∗b =
(αjb − αj)σ(αjb)
(48)
whereˆαjb is bootstrapped estimated in the bth resample and αj is estimated from the original data. Given
the estimates of the outcome equation and of the factor scores, we construct the bootstrapped p-value for
the contribution of skill k under the null hypothesis H0 : αjE(θj1 − θj0) = 0 as follows:
p-value =1
B
B∑b=1
1(T j,∗b > T j) with T j =αj ∗ E( θj(1)− θj(0))
σ(αj ∗ E( θj(1)− θj(0)))(49)
67
where T j,∗b is the statistic T j computed with the parameters obtained in the bth resample. Notice that the
p-value combines the variation in two population parameters: 1) the coefficient of the outcome equation;
2) the experimentally induced difference in means in the skills. It could be the case that each of these
parameters are, separately, statistically significant. However, the p-value may increase due to a loss in power
when they are combined.
Tables G.8 - G.11 shows the parameters of the outcome equations as wells as the decompositions com-
ponents.
68
Tab
leG
.8:
Fem
ale
Dec
om
posi
tion
(Yea
r6)
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eO
utcom
e Coef
ficien
tsCo
gniti
ve0.
040.
391
0.08
0.08
90.
230.
035
0.06
0.13
90.
210.
087
0.16
0.28
8-0
.08
0.36
330
4A
ttent
ion
Prob
lems
-0.1
50.
083
-0.1
10.
013
-0.1
10.
169
-0.0
70.
042
-0.1
40.
139
0.24
0.20
6-0
.19
0.18
130
4Co
nduc
t Pro
blem
s-0
.15
0.03
6-0
.09
0.01
8-0
.07
0.24
9-0
.03
0.19
2-0
.18
0.03
2-0
.20
0.19
00.
110.
255
304
War
mth
/Em
path
y0.
180.
060
0.05
0.19
20.
290.
003
0.09
0.01
4-0
.01
0.48
1-0
.41
0.10
10.
260.
125
304
Agg
ress
ion
-0.1
30.
103
-0.0
40.
218
-0.1
50.
107
-0.0
10.
416
0.13
0.17
7-0
.13
0.29
9-0
.06
0.38
630
4
Trea
tmen
t Effe
ctCo
gniti
ve0.
040.
391
-0.0
10.
110
0.04
0.03
20.
020.
079
0.03
0.08
10.
030.
246
-0.0
20.
321
304
Atte
ntio
n Pr
oblem
s-0
.15
0.08
30.
010.
099
-0.0
20.
144
-0.0
20.
046
-0.0
20.
115
0.04
0.16
8-0
.05
0.13
430
4Co
nduc
t Pro
blem
s-0
.15
0.03
60.
010.
112
-0.0
10.
223
-0.0
10.
170
-0.0
30.
065
-0.0
40.
153
0.03
0.20
830
4W
arm
th/E
mpa
thy
0.18
0.06
0-0
.01
0.16
90.
050.
007
0.03
0.01
8-0
.00
0.43
4-0
.07
0.07
30.
080.
093
304
Agg
ress
ion
-0.1
30.
103
0.00
0.17
3-0
.03
0.09
0-0
.00
0.40
10.
020.
127
-0.0
20.
263
-0.0
20.
351
304
Trea
tmen
t Effe
ct Fr
actio
nCo
gniti
ve0.
290.
391
-0.0
90.
110
0.35
0.03
20.
140.
079
0.25
0.08
10.
240.
246
-0.1
90.
321
304
Atte
ntio
n Pr
oblem
s0.
730.
083
-0.0
70.
099
0.10
0.14
40.
090.
046
0.10
0.11
5-0
.20
0.16
80.
260.
134
304
Cond
uct P
robl
ems
0.79
0.03
6-0
.06
0.11
20.
060.
223
0.05
0.17
00.
140.
065
0.19
0.15
3-0
.16
0.20
830
4W
arm
th/E
mpa
thy
0.71
0.06
0-0
.03
0.16
90.
210.
007
0.11
0.01
8-0
.01
0.43
4-0
.29
0.07
30.
290.
093
304
Agg
ress
ion
0.74
0.10
3-0
.03
0.17
30.
160.
090
0.02
0.40
1-0
.12
0.12
70.
130.
263
0.10
0.35
130
4
Hom
e y2
Pare
ntin
g y2
Mas
tery
y2
Anx
iety
y2Se
lf-E
stee
m y
2Sa
mpl
e Si
zeTr
eatm
ent
Birth
Weig
ht
Notes:
The
firs
tcolu
mn
pro
vid
es
the
outc
om
edesc
ripti
on
and
the
top
row
pro
vid
es
info
rmati
on
on
the
media
tors
.For
Year
6,
the
media
tors
are
treatm
ent,
bir
thw
eig
ht,
hom
eenvir
onm
ent,
pare
nti
ng,
anxie
ty,
self
-est
eem
and
mast
ery
.T
he
last
colu
mn
pro
vid
es
the
sam
ple
size
for
the
corr
esp
ondin
goutc
om
ein
the
firs
tcolu
mn.
The
row
sare
div
ided
into
3gro
ups:
Outc
om
eC
oeffi
cie
nts
,T
reatm
ent
Eff
ect
and
Tre
atm
ent
Eff
ect
Fra
cti
on.
The
last
of
these
gro
ups
isals
osh
ow
vis
ually
inF
igure
3.
Each
media
tor
has
two
sub
colu
mns
of
info
rmati
on:
the
coeffi
cie
nt
and
thep-v
alu
e.
Boldp-v
alu
es
are
signifi
cant
at
the
10%
level.
We
use
dth
efo
llow
ing
contr
ols
:m
ate
rnal
race,
mate
rnal
age,
mate
rnal
heig
ht,
gest
ati
onal
age,
house
hold
densi
ty,
regio
n,
em
plo
ym
ent
statu
sof
house
hold
head,
gra
ndm
oth
er
supp
ort
,ra
ndom
izati
on
wave,
incom
ecate
gory
,m
oth
er
curr
entl
yin
school,
and
mate
rnal
pare
nti
ng
att
itudes.
69
Tab
leG
.9:
Male
Dec
om
posi
tion
(Yea
r6)
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eO
utcom
e Coef
ficien
tsCo
gniti
ve0.
080.
240
0.08
0.09
30.
350.
004
0.11
0.01
40.
060.
327
-0.0
40.
447
0.02
0.46
430
5A
ggre
ssio
n-0
.08
0.18
6-0
.02
0.33
10.
070.
241
-0.0
50.
091
-0.2
30.
014
0.37
0.05
8-0
.11
0.31
030
5
Trea
tmen
t Effe
ctCo
gniti
ve0.
080.
240
0.02
0.06
40.
040.
054
0.02
0.04
70.
000.
281
-0.0
00.
362
0.00
0.42
230
5A
ggre
ssio
n-0
.08
0.18
6-0
.01
0.29
60.
010.
165
-0.0
10.
091
-0.0
10.
252
0.02
0.17
3-0
.02
0.23
030
5
Trea
tmen
t Effe
ct Fr
actio
nCo
gniti
ve0.
500.
240
0.14
0.06
40.
220.
054
0.11
0.04
70.
020.
281
-0.0
20.
362
0.02
0.42
230
5A
ggre
ssio
n0.
840.
186
0.05
0.29
6-0
.07
0.16
50.
080.
091
0.12
0.25
2-0
.23
0.17
30.
200.
230
305
Mas
tery
y2
Sam
ple
Size
Trea
tmen
tBi
rth W
eight
Hom
e y2
Pare
ntin
g y2
Anx
iety
y2Se
lf-E
stee
m y
2
Notes:
The
firs
tcolu
mn
pro
vid
es
the
outc
om
edesc
ripti
on
and
the
top
row
pro
vid
es
info
rmati
on
on
the
media
tors
.For
Year
6,
the
media
tors
are
treatm
ent,
bir
thw
eig
ht,
hom
eenvir
onm
ent,
pare
nti
ng,
anxie
ty,
self
-est
eem
and
mast
ery
.T
he
last
colu
mn
pro
vid
es
the
sam
ple
size
for
the
corr
esp
ondin
goutc
om
ein
the
firs
tcolu
mn.
The
row
sare
div
ided
into
3gro
ups:
Outc
om
eC
oeffi
cie
nts
,T
reatm
ent
Eff
ect
and
Tre
atm
ent
Eff
ect
Fra
cti
on.
The
last
of
these
gro
ups
isals
osh
ow
vis
ually
inF
igure
4.
Each
media
tor
has
two
sub
colu
mns
of
info
rmati
on:
the
coeffi
cie
nt
and
thep-v
alu
e.
Boldp-v
alu
es
are
signifi
cant
at
the
10%
level.
We
use
dth
efo
llow
ing
contr
ols
:m
ate
rnal
race,
mate
rnal
age,
mate
rnal
heig
ht,
gest
ati
onal
age,
house
hold
densi
ty,
regio
n,
em
plo
ym
ent
statu
sof
house
hold
head,
gra
ndm
oth
er
supp
ort
,ra
ndom
izati
on
wave,
incom
ecate
gory
,m
oth
er
curr
entl
yin
school,
and
mate
rnal
pare
nti
ng
att
itudes.
70
Tab
leG
.10:
Fem
ale
Dec
om
posi
tion
(Yea
r12)
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Out
come C
oeffic
ients
Tota
l # D
ays C
hild
Use
d M
ariju
ana
-0.1
50.
060
-0.1
20.
086
0.03
0.37
30.
000.
515
0.12
0.15
80.
040.
192
271
Child
Use
d A
lcoho
l, M
ari ju
ana,
or T
obac
co in
Las
t 30
Day
s-0
.02
0.19
8-0
.02
0.04
80.
000.
467
0.02
0.33
8-0
.00
0.47
70.
000.
457
268
Stan
dard
ized
Chi
ld B
MI (
Yea
r 12)
-0.3
00.
016
-0.1
10.
100
-0.2
20.
115
0.37
0.03
00.
080.
206
-0.1
10.
197
272
Trea
tmen
t Effe
ctCh
ild E
ver U
sed
Mar
i juan
a-0
.15
0.06
0-0
.01
0.21
8-0
.00
0.31
8-0
.00
0.48
40.
040.
111
-0.0
10.
155
271
Child
Use
d A
lcoho
l, M
ari ju
ana,
or T
obac
co in
Las
t 30
Day
s-0
.02
0.19
8-0
.00
0.21
7-0
.00
0.43
1-0
.00
0.27
3-0
.00
0.46
9-0
.00
0.42
426
8St
anda
rdiz
ed C
hild
BM
I (Y
ear 1
2)-0
.30
0.01
6-0
.01
0.20
90.
030.
109
-0.0
50.
060
0.02
0.16
20.
020.
145
272
Trea
tmen
t Effe
ct Fr
actio
nCh
ild E
ver U
sed
Mar
i juan
a1.
120.
060
0.05
0.21
80.
030.
318
0.00
0.48
4-0
.26
0.11
10.
060.
155
271
Child
Use
d A
lcoho
l, M
ari ju
ana,
or T
obac
co in
Las
t 30
Day
s0.
830.
198
0.05
0.21
70.
020.
431
0.07
0.27
30.
010.
469
0.01
0.42
426
8St
anda
rdiz
ed C
hild
BM
I (Y
ear 1
2)1.
060.
016
0.02
0.20
9-0
.11
0.10
90.
190.
060
-0.0
80.
162
-0.0
80.
145
272
War
mth
/Em
path
yA
ggre
ssio
nSa
mpl
e Si
zeTr
eatm
ent
Cogn
ition
Atte
ntio
n pr
oblem
sCo
nduc
t Pro
blem
s
Notes:
The
firs
tcolu
mn
pro
vid
es
the
outc
om
edesc
ripti
on
and
the
top
row
pro
vid
es
info
rmati
on
on
the
media
tors
.For
Year
12,
the
media
tors
are
treatm
ent,
cognit
ion,
att
enti
on
pro
ble
ms,
Conduct
Pro
ble
ms,
Warm
th/E
mpath
yand
Aggre
ssio
n.
The
last
colu
mn
pro
vid
es
the
sam
ple
size
for
the
corr
esp
ondin
goutc
om
ein
the
firs
tcolu
mn.
The
row
sare
div
ided
into
3gro
ups:
Outc
om
eC
oeffi
cie
nts
,T
reatm
ent
Eff
ect
and
Tre
atm
ent
Eff
ect
Fra
cti
on.
The
last
of
these
gro
ups
isals
osh
ow
vis
ually
inF
igure
7.
Each
media
tor
has
two
sub
colu
mns
of
info
rmati
on:
the
coeffi
cie
nt
and
thep-v
alu
e.
Bold
edp-v
alu
es
are
signifi
cant
at
the
10%
level.
Boldp-v
alu
es
are
signifi
cant
at
the
10%
level.
We
use
dth
efo
llow
ing
contr
ols
:m
ate
rnal
race,
mate
rnal
age,
mate
rnal
heig
ht,
gest
ati
onal
age,
house
hold
densi
ty,
regio
n,
em
plo
ym
ent
statu
sof
house
hold
head,
gra
ndm
oth
er
supp
ort
,ra
ndom
izati
on
wave,
incom
ecate
gory
,m
oth
er
curr
entl
yin
school,
and
mate
rnal
pare
nti
ng
att
itudes.
71
Tab
leG
.11:
Male
Dec
om
posi
tion
(Yea
r12)
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Coef
ficien
tp-
valu
eCo
effic
ient
p-va
lue
Out
come C
oeffic
ients
Ave
rage
TCA
P pe
rcen
tile,
y1-5
: lan
guag
e co
mpo
site
2.88
0.18
811
.93
0.00
02.
210.
379
-0.9
40.
442
2.40
0.17
8-3
.88
0.17
022
2PI
AT
read
ing
com
preh
ensio
n de
rived
scor
e1.
690.
134
7.44
0.00
0-1
.41
0.32
60.
450.
480
0.31
0.39
10.
570.
371
272
Ave
rage
mat
h gr
ade
grad
es 1
-50.
030.
368
0.50
0.00
0-0
.21
0.14
50.
220.
137
0.05
0.28
8-0
.13
0.11
824
3A
vera
ge m
ath
grad
e. Y
ears
1-5
afte
r KG
-0.0
00.
538
0.50
0.00
0-0
.15
0.20
50.
160.
198
0.03
0.32
4-0
.14
0.09
624
6av
erag
e tc
ap p
erce
ntile
y1-
5: m
ath
0.81
0.34
815
.29
0.00
05.
320.
231
-2.8
60.
336
-0.9
80.
380
-3.4
70.
204
223
PIA
T m
ath
deriv
ed sc
ore
1.64
0.10
67.
860.
000
-1.3
60.
324
0.20
0.51
50.
710.
232
1.53
0.14
627
0SC
eve
r trie
d sm
okin
g: 1
=ye
s-0
.05
0.07
90.
020.
254
0.04
0.27
90.
020.
341
-0.0
20.
192
0.01
0.40
427
4SC
use
alc,
mar
, tob
last
30
days
-0.0
50.
04-0
.00
0.40
-0.0
10.
410.
040.
25-0
.02
0.07
0.03
0.20
272
Inte
rnali
zing
diso
rder
s - Y
outh
repo
rt-0
.05
0.21
3-0
.07
0.04
70.
020.
421
0.03
0.40
60.
030.
295
0.10
0.07
227
4A
nxio
us/d
epre
ssed
- cli
nica
l or b
orde
rline
diso
rder
, you
th re
port
-0.0
50.
082
-0.0
50.
016
-0.0
50.
167
0.06
0.10
10.
010.
332
0.07
0.08
327
3A
vera
ge n
umbe
r of a
bsen
ces,
scho
ol y
ears
1-5
-1.0
50.
146
-2.2
50.
001
4.36
0.01
0-3
.87
0.00
90.
220.
372
-1.1
00.
156
267
Trea
tmen
t Effe
ctA
vera
ge T
CAP
perc
entil
e, y1
-5: l
angu
age
com
posit
e2.
880.
188
2.09
0.06
4-0
.02
0.43
20.
030.
361
-0.3
90.
135
0.45
0.16
122
2PI
AT
read
ing
com
preh
ensio
n de
rived
scor
e1.
690.
134
1.44
0.04
10.
110.
255
-0.0
30.
364
-0.0
40.
313
-0.0
70.
270
272
Ave
rage
mat
h gr
ade
grad
es 1
-50.
030.
368
0.08
0.08
0-0
.00
0.36
60.
000.
449
-0.0
10.
200
0.02
0.10
724
3A
vera
ge m
ath
grad
e. Y
ears
1-5
afte
r KG
-0.0
00.
538
0.08
0.06
10.
000.
451
-0.0
00.
335
-0.0
00.
230
0.02
0.12
124
6av
erag
e tc
ap p
erce
ntile
y1-
5: m
ath
0.81
0.34
82.
680.
070
-0.0
90.
367
0.09
0.31
30.
160.
303
0.41
0.20
322
3PI
AT
mat
h de
rived
scor
e1.
640.
106
1.55
0.04
20.
110.
242
-0.0
10.
420
-0.0
80.
169
-0.1
80.
102
270
SC e
ver t
ried
smok
ing:
1=
yes
-0.0
50.
079
0.00
0.19
8-0
.00
0.23
4-0
.00
0.33
40.
000.
194
-0.0
00.
324
274
SC u
se a
lc, m
ar, t
ob la
st 3
0 da
ys-0
.05
0.04
3-0
.00
0.34
20.
000.
328
-0.0
00.
262
0.00
0.14
4-0
.00
0.17
027
2In
tern
alizi
ng d
isord
ers -
You
th re
port
-0.0
50.
213
-0.0
10.
058
-0.0
00.
342
-0.0
00.
303
-0.0
00.
245
-0.0
10.
116
274
Anx
ious
/dep
ress
ed -
clini
cal o
r bor
derli
ne d
isord
er, y
outh
repo
rt-0
.05
0.08
2-0
.01
0.02
80.
000.
213
-0.0
00.
219
-0.0
00.
251
-0.0
10.
085
273
Ave
rage
num
ber o
f abs
ence
s, sc
hool
yea
rs 1
-5-1
.05
0.14
6-0
.36
0.06
3-0
.27
0.25
80.
070.
424
-0.0
30.
272
0.14
0.12
726
7
Trea
tmen
t Effe
ct Fr
actio
nA
vera
ge T
CAP
perc
entil
e, y1
-5: l
angu
age
com
posit
e0.
570.
188
0.41
0.06
4-0
.00
0.43
20.
010.
361
-0.0
80.
135
0.09
0.16
122
2PI
AT
read
ing
com
preh
ensio
n de
rived
scor
e0.
540.
134
0.46
0.04
10.
030.
255
-0.0
10.
364
-0.0
10.
313
-0.0
20.
270
272
Ave
rage
mat
h gr
ade.
Yea
rs 1
-5 a
fter K
G0.
230.
368
0.68
0.08
0-0
.03
0.36
60.
010.
449
-0.0
50.
200
0.16
0.10
724
3av
erag
e tc
ap p
erce
ntile
y1-
5: m
ath
0.20
0.34
80.
660.
070
-0.0
20.
367
0.02
0.31
30.
040.
303
0.10
0.20
322
3PI
AT
mat
h de
rived
scor
e0.
540.
106
0.51
0.04
20.
040.
242
-0.0
00.
420
-0.0
30.
169
-0.0
60.
102
270
SC u
se a
lc, m
ar, t
ob la
st 3
0 da
ys0.
920.
043
0.02
0.34
2-0
.02
0.32
80.
040.
262
-0.0
40.
144
0.08
0.17
027
2In
tern
alizi
ng d
isord
ers -
You
th re
port
0.63
0.21
30.
170.
058
0.01
0.34
20.
020.
303
0.03
0.24
50.
140.
116
274
Anx
ious
/dep
ress
ed -
clini
cal o
r bor
derli
ne d
isord
er, y
outh
repo
rt0.
710.
082
0.14
0.02
8-0
.05
0.21
30.
050.
219
0.02
0.25
10.
120.
085
273
Ave
rage
num
ber o
f abs
ence
s, sc
hool
yea
rs 1
-50.
700.
146
0.24
0.06
30.
180.
258
-0.0
50.
424
0.02
0.27
2-0
.09
0.12
726
7
Trea
tmen
tCo
gniti
onA
ttent
ion
prob
lems
Cond
uct P
robl
ems
War
mth
/Em
path
yA
ggre
ssio
nSa
mpl
e Si
ze
Notes:
The
firs
tcolu
mn
pro
vid
es
the
outc
om
edesc
ripti
on
and
the
top
row
pro
vid
es
info
rmati
on
on
the
media
tors
.For
Year
12,
the
media
tors
are
treatm
ent,
cognit
ion,
att
enti
on
pro
ble
ms,
Conduct
Pro
ble
ms,
Warm
th/E
mpath
yand
Aggre
ssio
n.
The
last
colu
mn
pro
vid
es
the
sam
ple
size
for
the
corr
esp
ondin
goutc
om
ein
the
firs
tcolu
mn.
The
row
sare
div
ided
into
3gro
ups:
Outc
om
eC
oeffi
cie
nts
,T
reatm
ent
Eff
ect
and
Tre
atm
ent
Eff
ect
Fra
cti
on.
The
last
of
these
gro
ups
isals
osh
ow
vis
ually
inF
igure
s5
-6.
Each
media
tor
has
two
sub
colu
mns
of
info
rmati
on:
the
coeffi
cie
nt
and
thep-v
alu
e.
Boldp-v
alu
es
are
signifi
cant
at
the
10%
level.
We
use
dth
efo
llow
ing
contr
ols
:m
ate
rnal
race,
mate
rnal
age,
mate
rnal
heig
ht,
gest
ati
onal
age,
house
hold
densi
ty,
regio
n,
em
plo
ym
ent
statu
sof
house
hold
head,
gra
ndm
oth
er
supp
ort
,ra
ndom
izati
on
wave,
incom
ecate
gory
,m
oth
er
curr
entl
yin
school,
and
mate
rnal
pare
nti
ng
att
itudes.
72
H Mediation Specification Test
In this section we specify how do we empirically test the effect that the mediators have on the final outcomes.
We use J for an indexing set of skills. We use Jp ⊆ J for the subset of measured skills. Our model for the
outcome equation is:
Yd = κd +∑j∈J
αjdθjd + βdX + εd, d ∈ 0, 1,
where κd is an intercept, (αjd; j ∈ J ) are loading factors and βd are |X|-dimensional vectors of parameters.
The error term εd is a zero-mean i.i.d. random variable assumed to be independent of regressors (θjd; j ∈ J )
and X.
The NFP analysts collected a rich array of measures of cognitive and personality skills. However, it
is likely that there are skills that they did not measure. As noted before, we use Jp ⊆ J be the index
set of measured skills. Namely, skills for which we have enough psychological instruments that allows for
estimation. We rewrite the equation for potential outcome Yd as:
Yd = κd +∑j∈J
αjdθjd + βdX + εd
= κd +∑j∈Jp
αjdθjd︸ ︷︷ ︸
skills thatwe measure
+∑
j∈J\Jp
αjθjd︸ ︷︷ ︸skills that we
do not measure
+βdX + εd
= κd +∑
j∈J\Jp
αjd E(θjd)︸ ︷︷ ︸new intercept
+∑j∈Jp
αjdθjd︸ ︷︷ ︸
skills thatwe measure
+∑
j∈J\Jp
αjd(θjd − E(θjd))︸ ︷︷ ︸
skills that wedo not measure
+βdX + εd,
= τd︸︷︷︸new intercept
+∑j∈Jp
αjdθjd︸ ︷︷ ︸
skills thatwe measure
+βdX +∑
j∈J\Jp
αjd(θjd − E(θjd)) + εd︸ ︷︷ ︸
new error term
(50)
where d ∈ 0, 1, τd = κd +∑j∈J\Jp α
jd E(θjd).Any differences in the error terms between treatment and
control groups can be attributed to differences in unmeasured skills. Thus, we assume, without loss of
generality, that ε1d= ε0, where
d= means equality in distribution.
The goal of this section is to examine the statistical assumptions needed to estimate unbiased parameters
(αjd : j ∈ Jp, d ∈ 0, 1). These parameters are used to perform the decomposition of outcome treatment
effects into parts associated with skills enhancement (θj1 − θj0 : j ∈ Jp). Parameters α may suffer from
confounding effects if measured and unmeasured skills are not independent. We can solve this confounding
73
problem by assuming that unmeasured skills are independent of measures skills. Namely,
(θjd; j ∈ J \ Jp) ⊥⊥ (θjd; j ∈ Jp)|X; d ∈ 0, 1,
then the regression:
Yd = τd +∑j∈Jp
αjdθjd + βdX + εd, (51)
produces unbiased estimates of parameter (αjd; j ∈ Jp); d ∈ 0, 1. Indeed error terms εd in equation (51)
are given by
εd = εd +∑
j∈J\Jp
αjd(θjd − E(θjd))
which are independent of (θjd; j ∈ Jp) conditional on X under the assumption that skills are independent.
Now suppose that instead of the skills independence assumption for both groups, we focus only on the
control group, thus,
(θj0; j ∈ J \ Jp) ⊥⊥ (θj0; j ∈ Jp)|X.
Moreover, suppose we also assume that αj1 = αj0; j ∈ J . Equivalently, the outcome loading factors for
both treatment and control groups are the same. In this new setup, the regression
Y0 = τ0 +∑j∈Jp
αjθj0 + β0X + ε0, (52)
also produces unbised estimates of (αj ; j ∈ Jp). Now consider the regression
Y1 = τ1 +∑j∈Jp
αjθj1 + β1X + ε1.
According to our rationale, this regression only produces unbiased estimates of (αj ; j ∈ Jp) if:
(θj1; j ∈ J \ Jp) ⊥⊥ (θj1; j ∈ Jp)|X, (53)
or, alternatively,
(θj1 − θj0; j ∈ J \ Jp) ⊥⊥ (θj1 − θ
j0; j ∈ Jp)|X. (54)
Thus, under this new set of assumptions, testing H0 : α1 = α0 is translated into testing the independence
relations of equations (53)–(54).
While the skill independence assumption in equation (53) may appear strong, the rich settlement of
74
information on NFP surveys makes this assumption more plausible. NFP data has a huge selection of
psychological questionnaires that aims to measure both cognitive and non-cognitive skills though childhood.
We examine all the available data and only a subset of these measures turns out to be statistically relevant
for mediation analysis. We use these measures to estimate factors that are able to explain the majority of
the treatment effects. Thus, it seems unlikely that some unobserved skills overlooked by psychologists could
have a major impact on mediating treatment effects.
H.1 Skills and the Measurement System
The assumption that the loading factors in the measurement system (equation 42) are the same for treatment
and control is not necessary to identify the model. It is useful for clarity in the interpretation because the
treatment operates by the shift of the latent skills and not by the map between measures and skills.
Ultimately, we need the decomposition of the treatment effects, (12), to be invariant to the choice of the
measurement system we used. Thus, for each skill’s contribution to treatment effect on each outcome, we
want to test the null hypothesis that:
H0 : α0(E(θ1 − θ0)) = α1(E(θ1 − θ0)) (55)
where αd = (αjd : j ∈ Jp) and θd = (θjd : j ∈ Jp) such that d ∈ 0, 1 denotes treatment status.
Let θi be the estimated factor score for individual i, assigned to treatment status Di ∈ 0, 1, using the
estimated loading factors from the subsample of individuals with the same treatment status, i.e. for each
individual factor score:
θi = (ϕDi′(ΩDi)
−1ϕDi)−1ϕDi
′(ΩDi)−1Mi.
We would like to test if the contributions to the treatment effects is independent if we use the parameters’
from different measurement system (i.e if we estimate a different set of loading factors for the treatment and
control group).
Hence, an appropriate single hypothesis test statistic for each skill j ∈ Jp becomes:
αj0(θj1 − θj0)− αj1(θj1 − θ
j0)
where we use a hat subscript to denote estimated parameters. α are Croon corrected estimates of α. We can
use a summary statistic to test the joint hypothesis stated in (55).
75
Independence between αd and θd − θd yields:
Var(αd(θ1 − θ0)) = (αd)2 Var(θ1 − θ0) + Var(αd)(θ1 − θ0)2 + Var(α) Var(θ1 − θ0)
Independence between the quantities estimated for each of the d’s yields:
Var(α0(θ1 − θ0)− α1(θ1 − θ0)) = Var(α0(θ1 − θ0)) + Var(α1(¯θ1 − ¯
θ0))
This variance helps us to get the z-statistic:
z =α0(θ1 − θ0)− α1(θ1 − θ0)√
Var(α0(θ1 − θ0)− α1(θ1 − θ0))
A two-sided z-test gives a p-value associated with the skill and outcome null hypothesis of invariance to
the choice of the measurement system.
These (outcome, skill) paired p-values are shown in Tables H.12 and H.13. We find that we can not
reject the null hypothesis for any skill-outcome pair which suggests that our decompositions of the NFP
treatment effects are not driven by the choice of the measurement system.
H.1.1 Additional Specification Test for the Outcome Equation
In order to clearly interpret the channels through which the NFP affects later outcomes, (5) assumes that the
parameters that map skills and pre-program variables with the outcomes are not affected by the programs.
Put another way, the mediated channels operate exclusively on the program effect on the skills. This
assumption is not necessary to identify the model.
For each outcome decomposed, we test the hypothesis that αj1 = αj0,∀j ∈ J and β1 = β0 with a
Wald test. Tables H.14 and H.15 show the results of this test. We cannot reject the null hypothesis of
equality of the coefficients for the treatment and control groups. This evidence strengthens the validity of
our interpretation of the decomposition of the NFP treatment effect into interpretable channels.
I Oaxaca-Blinder Decomposition Results
Oaxaca-Blinder decompositions are often used to examine sources of treatment effects. This method de-
composes the difference in means between two groups (treatment and control) into the part that is due to
the group differences in the channels and into the part that is due to group differences in the parameters
that capture the relationship between the channels and the outcomes. In our context, the Oaxaca-Blinder
76
decomposition is summarized as follows:25
E(Y |D = 1)− E(Y |D = 0)︸ ︷︷ ︸Treatment Effects
= (α1 − α0)θ0︸ ︷︷ ︸Differences unexplained by the skills
+ (θ1 − θ0)α︸ ︷︷ ︸Explained: differences in skills
. (56)
The decomposition that we propose summarizes the unexplained part in the above equation through
the difference in the intercepts between the treatment and the control groups. In order to assess if our
decomposition is a plausible specification, we estimate an Oaxaca-Blinder decomposition. The results in
Tables I.16 - I.20 evidence that the Oaxaca-Blinder unexplained part that accounts for differences in the
mapping of the skills on outcomes is not statistically significant for any outcome. Therefore, the results from
the decomposition of the NFP treatment effects presented in the paper seem to be correctly specified.
25We implicitly control for pre-program variables.
77
Tab
leH
.12:
Sp
ecifi
cati
onT
est
-In
vari
an
ceof
the
Contr
ibu
tion
of
Skil
lsto
the
Ch
oic
eof
the
Mea
sure
men
tS
yst
em(F
emale
s)
Age
6 o
utco
mes
Hom
eP
aren
ting
Mat
erna
l an
xiet
ySe
lf-es
teem
mas
tery
Cog
nitio
n0.
263
0.90
70.
859
0.69
80.
672
Atte
ntio
n pr
oble
ms
0.36
30.
709
0.70
20.
667
0.74
8C
ondu
ct p
robl
ems
0.42
10.
694
0.92
20.
721
0.67
7W
arm
th-e
mpa
thy
(pro
-soc
ial
skill
s)0.
267
0.90
70.
644
0.83
30.
973
Agr
essi
on P
robl
ems
0.69
20.
819
0.86
20.
821
0.78
6
Age
12
outc
omes
Cog
niti
onA
tten
tion
pr
obs
Con
duct
pr
obs
Em
path
yA
gres
sion
SC # days e
ver u
sed marijuana
0.86
70.
878
0.59
20.
885
0.28
0SC use alc, m
ar, tob
last 30 days
0.87
60.
695
0.90
70.
893
0.81
2St
anda
rized
Chi
ld B
MI
0.88
90.
822
0.57
40.
953
0.32
4
Fact
or T
estin
g R
esul
ts -
Fem
ales
Mat
erna
l ski
lls-
age
2
Chi
ldre
n's
skill
s. A
ge 6
Notes:
The
table
show
sp-v
alu
es
for
the
Wald
test
:z
=α
0(¯ θ0 1−
¯ θ0 0)−α
1(¯ θ1 1−
¯ θ1 0)
√ Var(α
0(¯ θ0 1−
¯ θ0 0)−α
1(¯ θ1 1−
¯ θ1 0))
78
Tab
leH
.13:
Sp
ecifi
cati
onT
est
-In
vari
ance
of
the
Contr
ibu
tion
of
Skil
lsto
the
Ch
oic
eof
the
Mea
sure
men
tS
yst
em(M
ale
s)
Age
6 o
utco
mes
Hom
eP
aren
ting
Mat
erna
l an
xiet
ySe
lf-es
teem
mas
tery
Cog
nitio
n0.
349
0.39
40.
971
0.92
70.
946
Agr
essi
on P
robl
ems
0.92
80.
959
0.95
00.
843
0.53
7
Age
12
outc
omes
Cog
niti
onA
tten
tion
pr
obs
cond
uct
prob
sE
mpa
thy
Agr
essi
on
Ave
rage
TC
AP
perc
etile
. Yea
rs 1
-5
afte
r KG
: Lan
guag
e0.
529
0.97
50.
993
0.63
60.
882
PIA
T re
adin
g co
mpr
ehen
sion
de
rived
sco
re0.
420
0.79
40.
941
0.86
70.
812
Ave
rage
mat
h gr
ades
. Yea
rs 1
-5
afte
r KG
0.42
50.
953
0.83
00.
871
0.79
2
Ave
rage
TC
AP
perc
etile
. Yea
rs 1
-5
afte
r KG
: Mat
h0.
571
0.94
00.
951
0.94
00.
875
PIA
T m
athe
mat
ics
deriv
ed s
core
0.43
30.
503
0.81
70.
976
0.65
2SC
use
of
alc,
mar
, tob
. Lat
30
days
0.84
50.
970
0.75
10.
791
0.53
8
Inte
rnal
izin
g di
sord
ers
- you
th
repo
rt0.
582
0.93
40.
911
0.90
50.
509
Clin
ical
or b
orde
rline
an
xiou
s/de
pres
sed
diso
rder
0.
537
0.93
60.
771
0.91
60.
687
Ave
rage
num
ber o
f ab
senc
es,
scho
ol y
ears
1-5
aft
er K
G0.
379
0.90
80.
706
0.83
30.
735
Mat
erna
l ski
lls-
age
2
Chi
ldre
n's
skill
s. A
ge 6
Notes:
The
table
show
sp-v
alu
es
for
the
Wald
test
:z
=α
0(¯ θ0 1−
¯ θ0 0)−α
1(¯ θ1 1−
¯ θ1 0)
√ Var(α
0(¯ θ0 1−
¯ θ0 0)−α
1(¯ θ1 1−
¯ θ1 0))
79
Tab
leH
.14:
Sp
ecifi
cati
on
Tes
t-
Ou
tcom
eE
qu
ati
on
(Fem
ale
s)
Out
com
eTe
st S
tat
P-V
al6
Year
sCo
gniti
on0.
982
0.49
0A
ttent
ion
prob
.1.
753
0.01
8Co
nduc
t Pro
b.0.
846
0.67
5Pr
o-so
cial
1.26
40.
189
Agg
ress
ion
0.55
80.
955
12 Y
ears
SC u
se a
lc, m
ar, t
ob la
st 3
0 da
ys1.
266
0.18
9SC
# d
ays u
se o
f alc,
mar
, tob
last
30
days
1.27
10.
186
Stan
dard
ized
Chi
ld B
MI (
Yea
r 12)
1.17
20.
270
Notes:
The
table
show
sp-v
alu
es
for
Wald
test
sfo
rth
eequality
of
slop
es
betw
een
treatm
ent
and
contr
ol
gro
up
inth
eoutc
om
eequati
on.
80
Tab
leH
.15:
Sp
ecifi
cati
on
Tes
t-
Ou
tcom
eE
qu
ati
on
(Male
s)
Out
com
eTe
st S
tat
P-V
al6
Year
sCo
gniti
on0.
609
0.92
6A
ggre
ssio
n0.
881
0.62
8
12 Y
ears
aver
age
tcap
per
cent
ile, y
1-5:
lang
uage
com
posit
e1.
162
0.28
3PI
AT
read
ing
com
preh
ensio
n de
rived
scor
e1.
286
0.17
5A
vera
ge m
ath
grad
e gr
ades
1-5
1.65
50.
034
Ave
rage
mat
h gr
ade.
Yea
rs 1
-5 a
fter K
G1.
493
0.07
3av
erag
e tc
ap p
erce
ntile
y1-
5: m
ath
1.24
20.
213
PIA
T m
ath
deriv
ed sc
ore
1.10
20.
343
SC u
se a
lc, m
ar, t
ob la
st 3
0 da
ys1.
208
0.23
7In
tern
alizi
ng d
isord
ers -
You
th re
port
0.99
30.
477
Anx
ious
/dep
ress
ed -
clini
cal o
r bor
derli
ne d
isord
er0.
682
0.86
7A
vera
ge n
umbe
r of a
bsen
ces,
scho
ol y
ears
1-5
0.79
80.
738
Notes:
The
table
show
sp-v
alu
es
for
Wald
test
sfo
rth
eequality
of
slop
es
betw
een
treatm
ent
and
contr
ol
gro
up
inth
eoutc
om
eequati
on.
81
Tab
leI.
16:
Oaxaca
-Blin
der
Dec
om
posi
tion
,ou
tcom
esat
age
6(F
emale
s)
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nO
verall
Tota
l Diff
. in
Mea
ns0.
114
0.11
20.
311
--0
.189
0.09
20.
039
--0
.197
0.07
20.
006
-0.
235
0.10
20.
021
--0
.187
0.09
40.
046
-E
xplai
ned
0.08
30.
041
0.04
40.
706
-0.0
800.
037
0.03
30.
271
-0.0
510.
030
0.08
60.
213
0.06
90.
035
0.05
00.
291
-0.0
240.
030
0.42
10.
260
Une
xplai
ned
0.03
10.
112
0.78
40.
294
-0.1
100.
093
0.23
70.
729
-0.1
460.
072
0.04
50.
787
0.16
60.
099
0.09
30.
709
-0.1
630.
090
0.07
10.
740
Exp
laine
d Po
rtion
Hom
e In
dex
0.03
20.
020
0.11
30.
355
-0.0
160.
015
0.27
90.
096
-0.0
100.
013
0.42
00.
063
0.03
40.
019
0.07
10.
214
-0.0
150.
015
0.34
40.
160
Pare
ntin
g In
dex
0.02
80.
026
0.28
40.
137
-0.0
480.
024
0.04
60.
093
-0.0
180.
018
0.31
40.
046
0.04
00.
023
0.07
40.
106
0.00
20.
019
0.93
50.
015
Mat
erna
l Anx
iety
Inde
x0.
024
0.02
00.
217
0.25
4-0
.018
0.01
60.
271
0.10
0-0
.019
0.01
40.
181
0.13
70.
006
0.01
40.
671
-0.0
080.
009
0.01
40.
536
-0.1
16M
ater
nal S
elf-E
stee
m In
dex
0.00
70.
021
0.75
10.
238
0.00
90.
020
0.64
2-0
.204
-0.0
320.
022
0.14
50.
187
-0.0
430.
028
0.12
2-0
.290
-0.0
220.
021
0.28
30.
131
Mat
erna
l Mas
tery
Inde
x0.
002
0.02
10.
913
-0.1
92-0
.019
0.02
30.
409
0.25
60.
017
0.01
90.
375
-0.1
570.
039
0.02
80.
158
0.29
5-0
.003
0.02
30.
893
0.09
6Bi
rthw
eight
-0.0
110.
012
0.39
3-0
.086
0.01
20.
013
0.35
3-0
.070
0.01
10.
012
0.36
1-0
.063
-0.0
060.
009
0.48
1-0
.027
0.00
50.
007
0.48
3-0
.027
Une
xplai
ned
Porti
onH
ome
Inde
x0.
008
0.02
40.
746
0.06
8-0
.004
0.01
80.
842
0.01
90.
000
0.01
40.
994
0.00
10.
028
0.02
30.
222
0.11
8-0
.009
0.01
90.
624
0.05
1Pa
rent
ing
Inde
x0.
020
0.02
80.
461
0.17
90.
033
0.02
20.
130
-0.1
76-0
.001
0.01
60.
965
0.00
40.
007
0.02
10.
741
0.03
00.
010
0.01
90.
615
-0.0
51M
ater
nal A
nxiet
y In
dex
-0.0
170.
021
0.43
4-0
.145
0.03
40.
024
0.16
1-0
.181
0.03
20.
020
0.11
1-0
.161
-0.0
070.
017
0.69
6-0
.029
0.00
50.
016
0.75
7-0
.026
Mat
erna
l Self
-Est
eem
Inde
x-0
.009
0.02
90.
757
-0.0
800.
041
0.02
90.
164
-0.2
160.
022
0.02
20.
306
-0.1
13-0
.018
0.03
50.
599
-0.0
770.
006
0.02
40.
802
-0.0
33M
ater
nal M
aste
ry In
dex
0.01
70.
034
0.62
80.
147
-0.0
460.
036
0.20
00.
244
-0.0
380.
026
0.14
80.
193
0.01
80.
037
0.62
50.
077
0.00
10.
029
0.97
4-0
.005
Birth
weig
ht-0
.002
0.00
70.
719
-0.0
220.
000
0.00
30.
915
-0.0
020.
002
0.00
60.
695
-0.0
11-0
.002
0.00
70.
758
-0.0
090.
002
0.00
50.
749
-0.0
09Re
sidua
l0.
014
0.11
70.
904
0.12
4-0
.169
0.09
40.
074
0.89
1-0
.163
0.07
40.
028
0.82
80.
140
0.09
90.
156
0.59
5-0
.176
0.09
10.
052
0.94
3
Cogn
ition
Atte
ntio
n Pr
oblem
sCo
nduc
t Pro
blem
sW
arm
th/E
mpa
thy
Agg
ress
ion
Notes:
The
indic
es
are
means
of
the
non-m
issi
ng
item
s.T
he
fracti
ons
are
pro
port
ions
of
the
tota
lcondit
ional
diff
ere
nce
inm
eans.
82
Tab
leI.
17:
Oaxaca
-Blin
der
Dec
om
posi
tion
,ou
tcom
esat
age
6(M
ale
s)
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nO
verall
Tota
l Diff
. in
Mea
ns0.
173
0.10
50.
100
--0
.025
0.09
50.
797
--0
.012
0.08
80.
891
--0
.080
0.10
40.
442
--0
.099
0.08
60.
250
-E
xplai
ned
0.09
20.
037
0.01
20.
496
-0.0
640.
033
0.05
1-5
.856
-0.0
560.
035
0.11
1-1
.130
0.02
20.
025
0.38
7-0
.254
-0.0
250.
025
0.32
20.
158
Une
xplai
ned
0.08
10.
101
0.42
20.
504
0.03
90.
092
0.67
06.
856
0.04
40.
090
0.62
82.
130
-0.1
020.
105
0.33
21.
254
-0.0
740.
085
0.38
30.
842
Exp
laine
d Po
rtion
Hom
e In
dex
0.02
10.
022
0.34
40.
220
-0.0
030.
006
0.62
7-0
.303
-0.0
090.
010
0.39
5-0
.395
0.00
60.
009
0.51
3-0
.161
0.00
20.
005
0.72
8-0
.074
Pare
ntin
g In
dex
0.04
40.
022
0.04
20.
112
-0.0
210.
019
0.25
8-0
.479
-0.0
110.
018
0.54
0-0
.041
0.00
80.
017
0.64
0-0
.007
-0.0
190.
013
0.15
00.
085
Mat
erna
l Anx
iety
Inde
x0.
001
0.00
40.
847
0.01
9-0
.007
0.01
20.
561
-0.5
44-0
.008
0.01
40.
546
-0.2
970.
003
0.00
70.
629
-0.0
46-0
.011
0.01
80.
538
0.12
5M
ater
nal S
elf-E
stee
m In
dex
-0.0
050.
009
0.60
7-0
.016
0.00
50.
010
0.59
21.
536
-0.0
110.
015
0.49
0-0
.514
-0.0
030.
009
0.69
00.
034
0.01
20.
014
0.39
9-0
.230
Mat
erna
l Mas
tery
Inde
x0.
009
0.01
80.
598
0.02
1-0
.015
0.01
80.
385
-3.9
78-0
.001
0.01
40.
964
0.55
30.
004
0.01
80.
831
-0.0
20-0
.003
0.01
30.
833
0.19
7Bi
rthw
eight
0.02
20.
017
0.19
20.
140
-0.0
230.
018
0.21
4-2
.089
-0.0
160.
015
0.28
3-0
.435
0.00
40.
014
0.74
7-0
.054
-0.0
070.
010
0.53
40.
055
Une
xplai
ned
Porti
onH
ome
Inde
x0.
002
0.01
10.
834
0.01
3-0
.003
0.00
90.
763
0.11
10.
002
0.00
90.
846
-0.1
520.
008
0.01
20.
529
-0.0
96-0
.002
0.00
90.
836
0.01
9Pa
rent
ing
Inde
x-0
.007
0.02
40.
760
-0.0
430.
012
0.02
40.
627
-0.4
710.
016
0.02
20.
467
-1.3
140.
010
0.02
70.
698
-0.1
290.
034
0.02
10.
110
-0.3
40M
ater
nal A
nxiet
y In
dex
0.00
60.
012
0.62
20.
035
0.00
00.
006
0.94
50.
016
-0.0
050.
011
0.62
50.
451
0.00
30.
009
0.70
8-0
.042
-0.0
040.
008
0.62
70.
040
Mat
erna
l Self
-Est
eem
Inde
x0.
004
0.01
00.
665
0.02
4-0
.002
0.00
80.
771
0.09
10.
004
0.01
00.
688
-0.3
210.
005
0.01
00.
642
-0.0
590.
000
0.00
60.
970
-0.0
02M
ater
nal M
aste
ry In
dex
-0.0
330.
028
0.24
1-0
.192
-0.0
070.
026
0.78
00.
294
0.02
00.
023
0.39
6-1
.627
-0.0
230.
029
0.42
10.
288
0.02
90.
026
0.26
8-0
.289
Birth
weig
ht0.
020
0.02
40.
418
0.11
3-0
.060
0.02
80.
034
2.42
2-0
.065
0.02
80.
022
5.33
5-0
.015
0.02
40.
525
0.19
1-0
.003
0.01
80.
877
0.02
8Re
sidua
l0.
090
0.10
70.
400
0.51
90.
100
0.09
30.
284
-4.0
610.
072
0.09
80.
462
-5.9
70-0
.090
0.11
60.
436
1.12
1-0
.128
0.07
60.
091
1.29
3
Cogn
ition
Atte
ntio
n Pr
oblem
sCo
nduc
t Pro
blem
sW
arm
th/E
mpa
thy
Agg
ress
ion
Notes:
The
indic
es
are
means
of
the
non-m
issi
ng
item
s.T
he
fracti
ons
are
pro
port
ions
of
the
tota
lcondit
ional
diff
ere
nce
inm
eans.
83
Tab
leI.
18:
Oaxaca
-Blin
der
Dec
om
posi
tion
,ou
tcom
esat
age
12
(Fem
ale
s)
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nTo
tal D
iff. i
n M
eans
-0.1
410.
085
0.09
8-
-0.0
210.
020
0.28
1-
-0.1
970.
110
0.07
2-
Exp
laine
d0.
026
0.03
80.
488
-0.1
84-0
.005
0.00
70.
460
0.24
20.
020
0.03
80.
601
-0.1
02U
nexp
laine
d-0
.167
0.11
10.
133
1.18
4-0
.016
0.02
10.
429
0.75
8-0
.217
0.11
40.
056
1.10
2
Exp
laine
dCo
gniti
ve-0
.006
0.01
80.
723
0.04
6-0
.001
0.00
30.
708
0.05
8-0
.004
0.01
30.
777
0.01
9A
ttent
ion
Prob
lems
0.00
00.
013
0.99
10.
001
-0.0
010.
005
0.82
70.
051
0.01
60.
017
0.35
4-0
.081
Cond
uct P
robl
ems
-0.0
030.
010
0.77
50.
019
-0.0
030.
004
0.46
80.
149
-0.0
270.
022
0.21
50.
138
War
mth
/Em
path
y0.
038
0.03
80.
318
-0.2
680.
000
0.00
40.
934
-0.0
150.
023
0.02
30.
331
-0.1
15A
ggre
ssio
n-0
.003
0.00
50.
644
0.01
80.
000
0.00
20.
988
-0.0
010.
013
0.02
10.
554
-0.0
64
Une
xplai
ned
Cogn
itive
0.01
20.
018
0.52
3-0
.084
0.00
30.
004
0.49
6-0
.135
0.00
40.
015
0.77
3-0
.022
Atte
ntio
n Pr
oblem
s0.
002
0.01
50.
889
-0.0
150.
008
0.00
70.
288
-0.3
64-0
.036
0.03
40.
293
0.18
1Co
nduc
t Pro
blem
s0.
013
0.01
80.
482
-0.0
91-0
.008
0.01
20.
522
0.36
40.
095
0.04
70.
043
-0.4
81W
arm
th/E
mpa
thy
-0.0
290.
032
0.37
40.
202
-0.0
050.
004
0.25
40.
228
-0.0
260.
023
0.26
00.
132
Agg
ress
ion
-0.0
020.
009
0.80
30.
016
-0.0
010.
003
0.69
70.
062
0.01
30.
037
0.72
4-0
.066
Resid
ual
-0.1
630.
096
0.08
91.
155
-0.0
130.
027
0.62
90.
603
-0.2
680.
122
0.02
81.
358
SC #
day
s eve
r use
d m
ariju
ana
SC u
se a
lc, m
ar, t
ob la
st 3
0 da
ysSt
anda
rdiz
ed C
hild
BM
I (12
Y)
Notes:
The
indic
es
are
means
of
the
non-m
issi
ng
item
s.T
he
fracti
ons
are
pro
port
ions
of
the
tota
lcondit
ional
diff
ere
nce
inm
eans.
84
Tab
leI.
19:
Oaxaca
-Bli
nd
erou
tcom
esat
age
12,
Dec
om
posi
tion
Part
1(M
ale
s)
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Tota
l Diff
. in
Mea
ns4.
403
3.28
90.
181
-2.
022
1.58
40.
202
-0.
117
0.10
70.
275
-0.
083
0.10
30.
422
-2.
818
3.42
90.
411
-2.
447
1.32
40.
065
-E
xplai
ned
1.56
41.
520
0.30
40.
355
0.94
10.
808
0.24
40.
466
0.07
70.
063
0.21
90.
660
0.06
70.
058
0.24
20.
814
2.22
61.
820
0.22
10.
790
0.94
60.
817
0.24
70.
387
Une
xplai
ned
2.83
93.
105
0.36
10.
645
1.08
01.
454
0.45
70.
534
0.04
00.
090
0.65
80.
340
0.01
50.
087
0.86
00.
186
0.59
23.
111
0.84
90.
210
1.50
11.
103
0.17
40.
613
Exp
laine
dCo
gniti
ve1.
831
1.33
70.
171
0.41
60.
910
0.70
50.
197
0.45
00.
057
0.05
40.
285
0.49
20.
057
0.05
10.
266
0.68
22.
146
1.66
40.
197
0.76
20.
939
0.75
00.
211
0.38
4A
ttent
ion
Prob
lems
0.07
50.
275
0.78
70.
017
0.06
00.
131
0.64
70.
030
0.00
90.
016
0.57
00.
080
0.00
70.
012
0.54
80.
087
-0.0
570.
312
0.85
5-0
.020
0.06
70.
123
0.58
90.
027
Cond
uct P
robl
ems
-0.0
030.
273
0.99
2-0
.001
0.03
70.
122
0.76
50.
018
0.00
00.
010
0.96
4-0
.004
-0.0
010.
008
0.93
3-0
.008
0.02
00.
281
0.94
50.
007
0.03
50.
110
0.75
10.
014
War
mth
/Em
path
y-0
.535
0.46
80.
253
-0.1
22-0
.079
0.15
60.
611
-0.0
39-0
.009
0.01
30.
475
-0.0
79-0
.012
0.01
40.
362
-0.1
49-0
.063
0.38
50.
870
-0.0
22-0
.094
0.12
20.
441
-0.0
38A
ggre
ssio
n0.
197
0.41
90.
639
0.04
50.
014
0.16
10.
930
0.00
70.
020
0.01
80.
267
0.17
10.
017
0.01
70.
328
0.20
30.
181
0.43
90.
681
0.06
4-0
.001
0.14
20.
995
0.00
0
Une
xplai
ned Co
gniti
ve0.
540
0.98
00.
582
0.12
3-0
.015
0.28
80.
958
-0.0
070.
005
0.01
90.
787
0.04
50.
009
0.01
90.
652
0.10
50.
563
0.78
00.
470
0.20
0-0
.016
0.22
80.
943
-0.0
07A
ttent
ion
Prob
lems
0.69
90.
769
0.36
30.
159
0.44
90.
342
0.18
90.
222
0.00
90.
020
0.65
90.
077
0.00
60.
020
0.75
20.
075
0.51
80.
663
0.43
50.
184
-0.1
620.
292
0.57
9-0
.066
Cond
uct P
robl
ems
0.00
30.
427
0.99
50.
001
-0.0
920.
242
0.70
5-0
.045
-0.0
030.
016
0.85
8-0
.024
-0.0
030.
015
0.86
4-0
.031
0.00
20.
402
0.99
50.
001
0.00
40.
183
0.98
30.
002
War
mth
/Em
path
y0.
373
0.69
10.
590
0.08
50.
134
0.27
10.
622
0.06
60.
011
0.02
00.
588
0.09
20.
012
0.02
10.
559
0.15
0-0
.076
0.65
00.
907
-0.0
270.
070
0.22
20.
752
0.02
9A
ggre
ssio
n-0
.078
0.49
50.
875
-0.0
18-0
.071
0.27
00.
793
-0.0
35-0
.006
0.01
70.
740
-0.0
49-0
.001
0.01
30.
933
-0.0
14-0
.266
0.61
40.
664
-0.0
94-0
.168
0.23
10.
468
-0.0
69Re
sidua
l1.
302
3.44
70.
706
0.29
60.
675
1.61
70.
676
0.33
40.
023
0.09
00.
796
0.20
0-0
.008
0.08
90.
927
-0.0
99-0
.150
3.28
90.
964
-0.0
531.
773
1.19
10.
137
0.72
4
PIA
T m
ath
deriv
ed sc
ore
langu
age
com
posit
ede
rived
scor
eA
vera
ge m
ath
grad
e gr
ades
1-5
afte
r KG
aver
age
tcap
per
cent
ile y
1-5:
mat
h
Notes:
The
indic
es
are
means
of
the
non-m
issi
ng
item
s.T
he
fracti
ons
are
pro
port
ions
of
the
tota
lcondit
ional
diff
ere
nce
inm
eans.
85
Tab
leI.
20:
Oaxaca
-Bli
nd
erou
tcom
esat
age
12,
Dec
om
posi
tion
Part
2(M
ale
s)
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nE
ffect
SEP
-Val
Frac
tion
Effe
ctSE
P-V
alFr
actio
nTo
tal D
iff. i
n M
eans
-0.0
630.
034
0.06
5-
-0.0
340.
025
0.17
6-
-0.0
680.
063
0.28
1-
-0.0
530.
030
0.08
1-
-1.0
170.
931
0.27
5-
Exp
laine
d-0
.005
0.01
00.
644
0.07
6-0
.003
0.01
00.
741
0.09
9-0
.030
0.02
20.
163
0.44
5-0
.014
0.01
20.
248
0.26
2-0
.524
0.34
60.
130
0.51
5U
nexp
laine
d-0
.058
0.03
60.
103
0.92
4-0
.031
0.02
50.
231
0.90
1-0
.038
0.06
20.
546
0.55
5-0
.039
0.03
00.
192
0.73
8-0
.493
0.91
80.
591
0.48
5
Exp
laine
dCo
gniti
ve0.
001
0.00
40.
727
-0.0
210.
000
0.00
30.
886
0.01
3-0
.006
0.00
90.
477
0.09
0-0
.006
0.00
60.
330
0.10
8-0
.237
0.22
80.
298
0.23
3A
ttent
ion
Prob
lems
-0.0
040.
005
0.50
50.
057
0.00
10.
003
0.78
9-0
.023
-0.0
040.
008
0.58
80.
062
0.00
00.
002
0.89
00.
006
-0.2
120.
237
0.37
00.
209
Cond
uct P
robl
ems
-0.0
010.
005
0.77
70.
022
-0.0
010.
004
0.74
00.
042
-0.0
020.
007
0.81
10.
024
-0.0
010.
004
0.74
20.
026
0.05
60.
180
0.75
8-0
.055
War
mth
/Em
path
y0.
002
0.00
40.
569
-0.0
340.
004
0.00
40.
296
-0.1
17-0
.002
0.00
70.
762
0.03
00.
001
0.00
40.
811
-0.0
18-0
.125
0.13
10.
339
0.12
3A
ggre
ssio
n-0
.003
0.00
60.
556
0.05
4-0
.006
0.00
80.
459
0.18
4-0
.016
0.01
50.
267
0.23
9-0
.007
0.00
70.
313
0.14
0-0
.004
0.11
90.
970
0.00
4
Une
xplai
ned
Cogn
itive
-0.0
010.
005
0.90
60.
010
-0.0
060.
007
0.38
20.
187
-0.0
030.
012
0.79
90.
046
-0.0
010.
007
0.87
90.
020
0.10
30.
135
0.44
4-0
.102
Atte
ntio
n Pr
oblem
s-0
.004
0.01
00.
650
0.06
90.
000
0.00
50.
980
-0.0
03-0
.007
0.01
60.
657
0.10
7-0
.001
0.00
50.
907
0.01
0-0
.101
0.19
90.
612
0.09
9Co
nduc
t Pro
blem
s0.
002
0.00
60.
783
-0.0
250.
002
0.00
50.
734
-0.0
470.
001
0.01
00.
960
-0.0
070.
002
0.00
60.
765
-0.0
340.
003
0.09
60.
977
-0.0
03W
arm
th/E
mpa
thy
-0.0
070.
007
0.33
50.
105
0.00
00.
004
0.91
30.
012
0.01
40.
014
0.33
8-0
.203
0.00
30.
008
0.67
8-0
.064
-0.0
510.
205
0.80
30.
050
Agg
ress
ion
0.00
10.
006
0.87
2-0
.014
0.00
60.
009
0.52
0-0
.171
0.00
40.
014
0.74
2-0
.066
0.00
10.
008
0.93
2-0
.012
0.05
40.
169
0.74
9-0
.053
Resid
ual
-0.0
490.
042
0.23
90.
780
-0.0
310.
026
0.23
50.
923
-0.0
460.
069
0.50
30.
679
-0.0
440.
035
0.21
10.
818
-0.5
011.
072
0.64
00.
493
SC e
ver t
ried
smok
ing:
1=
yes
SC u
se a
lc, m
ar, t
ob la
st 3
0 da
ysIn
tern
alizi
ng d
isord
ers -
You
th
Anx
ious
/dep
ress
ed -
clini
cal o
r A
vera
ge n
umbe
r of a
bsen
ces,
Notes:
The
indic
es
are
means
of
the
non-m
issi
ng
item
s.T
he
fracti
ons
are
pro
port
ions
of
the
tota
lcondit
ional
diff
ere
nce
inm
eans.
86