the determinants of human capital formation during the early years of life

The Determinants of Human Capital Formation Duringthe Early Years of Life

Theory, Measurement, Empirics and Policies.

Orazio P. Attanasio

UCL, IFS, NBER & [email protected]

GSE Lecture - Barcelona November 27th 2014

Co-authors and research projects

I will use liberally material from studies with many co-authors.

The Colombia study:Attanasio, Fitszimons, Fernandez, Grantham- McGregor, Meghir, Rubio(impacts of a stimulation intervention) - BMJ 2014.Attanasio, Cattan, Fitzsimons, Meghir, Rubio (2014) (a model of theproduction function and investment)

Production function in observational dataAttanasio, Meghir, Nix (2014) (Young Lives Data - India)

Beliefs and distorted production functionsAttanasio, Cattan (in progress) - Data from ColombiaAndrew, Attanasio, Malde, Vera-Hernandez, Wilhelm - Data fromMalawiAttanasio, Cunha, Jervis (in progress) - Data from Colombia

Measuring child developmentAraujo, Attanasio, Rubio (2014), Bogota survey, in progressHamadami, Grantham McGregor, Attanasio et al, Bangladesh data,Pediatrics, 2014Attanasio et al. (2014), Bogota Survey, forthcoming JHR.

Orazio P. Attanasio ( UCL and IFS) Determinants of HK formation GSE - Barcelona 27.11.14 2 / 76

Co-authors and research projects

I will use liberally material from studies with many co-authors.The Colombia study:

Attanasio, Fitszimons, Fernandez, Grantham- McGregor, Meghir, Rubio(impacts of a stimulation intervention) - BMJ 2014.Attanasio, Cattan, Fitzsimons, Meghir, Rubio (2014) (a model of theproduction function and investment)

Production function in observational dataAttanasio, Meghir, Nix (2014) (Young Lives Data - India)

Beliefs and distorted production functionsAttanasio, Cattan (in progress) - Data from ColombiaAndrew, Attanasio, Malde, Vera-Hernandez, Wilhelm - Data fromMalawiAttanasio, Cunha, Jervis (in progress) - Data from Colombia

Measuring child developmentAraujo, Attanasio, Rubio (2014), Bogota survey, in progressHamadami, Grantham McGregor, Attanasio et al, Bangladesh data,Pediatrics, 2014Attanasio et al. (2014), Bogota Survey, forthcoming JHR.

Antecedents and Debts

Sally Grantham McGregor

The Jamaica Study (Lancet, 1991, Science, 2014)The Bangladesh study (Journal of Nutrition, 2006)The Lancet Series (2007, 2011)

Jim Heckman and co-authors:

Models of Human Capital accumulation.Latent Factor models

Chuck Manski, Tom Juster

Measurement: beyond revealed preferencesExpectations, beliefs, attitudes

Outline

1 IntroductionThe importance of early yearsSome Successful Interventions

2 Known and Unknown

3 A theoretical framework

4 Using the modelA Colombian intervention and its impactsEstimating the model and interpreting the impacts

5 Two DigressionsStructural Models and RCTsMeasurement

6 Beliefs

7 Conclusions

Introduction

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Introduction The importance of early years

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Introduction: Human Capital Development

A considerable amount of attention devoted to human capitalaccumulation

In the process of growth and developmentTo reduce inequality and increase equality of opportunities

Human capital accumulation is a very complex process:

MultidimensionalComplex interactionsIt starts very early in life

The early years are important and have long run consequences.

Vulnerability / MalleabilitySalient for policy... it does not mean every thing is fixed by age 3!

Vulnerability / MalleabilitySalient for policy

... it does not mean every thing is fixed by age 3!

Human development in developing countries.

Children vulnerability and policy interventions can be particularlyrelevant in developing countries where risk factors are particularlysalient.

According to the Lancet (2007, 2011) series, 200m children are at riskof not developing their full potential.

Risk factors in developing countries.

Bio factors and nutrition – undernutrition in perinatal period andearly childhood, iron deficiency, iodine deficiency, othermicro-nutrients deficits;

Infectious diseases;

Environmental factors – clean water, hygiene, health;

Psycho-social factors – stimulation, parenting and responsiveness ofcare-givers, violence, maternal depression.

Developmental lags are associated with poverty:some evidence from Ecuador.

The delay accumulated by age 6 by the poorest children is 3 s.d. ofstardardised scores.This is equivalent to a 2.5 years delay.

(Paxson and Schady, JHR 2007).Orazio P. Attanasio ( UCL and IFS) Determinants of HK formation GSE - Barcelona 27.11.14 10 / 76

Developmental lags are associated with poverty:some evidence from Bangladesh.

Developmental lags are significant already at 7 months of ageBy age 5 they are very large

(Hamadami, Grantham-McGregor, Attanasio et al, forthcoming inPediatrics).

Developmental lags are associated with poverty:some evidence from Bogota.

Socieconomic gradient significant at 12 months in Bogota.Delays evident in several domains.

(Attanasio et al., forthcoming in Journal of Human Resources).

Introduction Some Successful Interventions

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Successful interventions

The fact that the early years are malleable and important implies thatwell designed and well targeted policies can have spectacular impacts.

This has been shown both in developed and developing countries:

US programs that have shown substantive impacts :

The Perry Pre-School ProgramThe Abecedarian ProgramThe Nurse Family PartnershipIHDP: Infant Health and Development Program

Interventions in developing countries:

The Jamaica StudyBangladesh

The Jamaica study.

129 stunted children, aged between 9 and 24 months at baseline wererandomly divided into 4 groups:

Stimulation group;Nutrition group;Nutrition + Stimulation group;Control group.

The intervention lasted 2 years and the children were observed:

at the end of the intervention;at age 7-8;at age 11-12;at age 17-18;at age 22-23.

The Jamaica study.

The results were stunning:(Grantham Mc Gregor et al., Lancet 1991)

The Jamaica study.

The results were stunning: (Walker et al., Lancet 2005)

Known and Unknown

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Known and Unknown

What we know and what we don’t.

Over the last few decades we have learned much:

Human capital is a multidimensional object and its different dimensionsare all important;Early years are key;Nutrition is important;Stimulation is key to child development and might be more effectivethan nutrition;The household environment is crucial, especially in the early years;There are important dynamic interactions.

Known and Unknown

Over the last few decades we have learned much:

Human capital is a multidimensional object and its different dimensionsare all important;Early years are key;Nutrition is important;Stimulation is key to child development and might be more effectivethan nutrition;The household environment is crucial, especially in the early years;There are important dynamic interactions.

Known and Unknown

... and yet there is much we do not know!

We still do not understand the role of genes and epigenetics and howthey affect multiple aspects of development;We still do not fully understand how nutrition affect the developmentof cognitive and non-cognitive skills;

Example: Iron and Zinc; Iron and malaria.

We still do not fully understand the nature of the interactions ofdifferent aspects of human capital;We still do not fully understand the role played by different inputs andtheir interactions.

Known and Unknown

We still do not understand what determines investment in humancapital.

What do parents do?What constraints do they face?How are resources allocated within the family?

We still do not know how to build cost effective interventions: fromefficacy to effectiveness.

Known and Unknown

We still do not understand what determines investment in humancapital.

What do parents do?What constraints do they face?How are resources allocated within the family?

We still do not know how to build cost effective interventions: fromefficacy to effectiveness.

A theoretical framework

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

The Production Function of Human Capital

from Cunha, Heckman and Schennach (2010)

Ht+1 = gt(Ht ,Xt ,Zt , et+1)

Ht is Human Capital (including cognition c , socio-emotional dev. sand health h capital)..

Zt are background variables (either fixed or time varying) (includingmother m, father f and other r).

Xt are Investments in HK (including materials M and time T ). .

et+1 are shocks.

All variables are multidimensional:

Ht = {θct , θst , θht }Zt = {θmt , θft , θrt}Xt = {θMt , θTt }

The Investment Problem

Parents choose investment to maximise utility that depends onHuman Capital and Consumption

subject to a budget constraintsubject to (their perception of) the production function

Max{Ct ,Xt}U(Ct ,Ht+1)

s.t. Ct + PxtXt = Yt

Investment could be in commodities or time.

Could easily add labour supply choices

Could easily add inter-temporal dimension, independently of theprocess of HK

The Investment Problem

Parents choose investment to maximise utility that depends onHuman Capital and Consumption

subject to a budget constraintsubject to (their perception of) the production function

Max{Ct ,Xt}U(Ct ,Ht+1)

s.t. Ct + PxtXt = Yt

Investment could be in commodities or time.

Could easily add labour supply choices

Could easily add inter-temporal dimension, independently of theprocess of HK

Issues: what we do observe?

Many of the variables in this model are not directly observed.

Cunha et al. (2010) propose a useful latent factor approach.

The structural model is complemented with a measurement system

mkjt = αjk

t θjt + εkjt , j = {c , s, h,m, f , r ,M,T}, k = {1, 2, ...}

We need at least two measurements to identify the distribution ofunobservable latent factors.

Some exclusion restrictions are required.

Some independence of measurement error is required.

mkjt = αjk

Issues: Endogeneity of Investment

Parents choose investment in reaction to shocks.

The structural framework we sketched suggests a solution.

One can derive an Investment Function:.

Xt = ft(Yt ,Pt ,Zt, et)

One can estimate the investment equation together with theproduction function

... or approximate it and use a control function approach

Notice the role played by the variables Pt ,Yt in the investmentfunction: they are excluded from the production function.

One can use a control function approach to estimate the roleinvestment plays in the production function.

The parameters of the investment function will depend on:

The parameters of the utility functionThe parameters of the production function as perceived by the parents

Notice the role played by the variables Pt ,Yt in the investmentfunction: they are excluded from the production function.

One can use a control function approach to estimate the roleinvestment plays in the production function.

The parameters of the investment function will depend on:

The parameters of the utility functionThe parameters of the production function as perceived by the parents

Estimation

Estimation can be performed in two steps (Attanasio, Meghir andNix, 2014):

Estimation of the joint distribution of all latent factorsDraws from such a distribution can be used to estimate the structuralparameters on simulated data by NLLS or Control Function.This applies to both production function and investment function.Endogeneity of the latter is taken into account by a control functionapproach.Inference can be done by using appropriate bootstrapping.

Using the model

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Using the model

We now use the model to analyse the data from a RCT designed toevaluate a stimulation intervention.

I will briefly describe the intervention.

I will report the impacts.

I will show how we use the data from the intervention to estimate theproduction function of human capital at early years.

I will discuss why this is useful.

Using the model A Colombian intervention and its impacts

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

An early years stimulation intervention in Colombia

Results from Attanasio, Fernandez, Fitzsimons,Grantham-McGregor, Meghir and Rubio, (2014)

The intervention consists of home visits targeted at children aged 12to 24 months at baseline.

It adapted Grantham-McGregor Jamaican curriculum to theColombian context.

One hour weekly visits by a visitor who interacts with mother andchild and develops an integrated and well structured curriculum

Focus on cognition, language, fine motor skills.

Innovations relative to the Jamaica study:

The intervention was delivered by local women (rather than healthworkers), identified through a pre-existing welfare program.The intervention was implemented on a large scale, attempting todesign a scalable program.It lasted 18 months.We collected rich data to be used to explain the impacts of theintervention.

The intervention was evaluated through a Cluster RCT.

Randomization was across 96 small towns in 3 regions.

The resulting data is one of the largest RCT in this field: 1400children in 96 clusters;

Effective sample size, taking into account ICC, 880

Innovations relative to the Jamaica study:

The intervention was delivered by local women (rather than healthworkers), identified through a pre-existing welfare program.The intervention was implemented on a large scale, attempting todesign a scalable program.It lasted 18 months.We collected rich data to be used to explain the impacts of theintervention.

The intervention was evaluated through a Cluster RCT.

Randomization was across 96 small towns in 3 regions.

The resulting data is one of the largest RCT in this field: 1400children in 96 clusters;

Effective sample size, taking into account ICC, 880

The intervention’s impacts

The intervention had some significant impacts on cognitivedevelopment and receptive language.

source: Attanasio et al. BMJ 2014;349:g5785 doi: 10.1136/bmj.g5785 (Published29 September 2014)

The intervention’s impacts

The cognitive impact fills about a third of the gap between top andbottom wealth quintiles in Bogota.

ley co

gnitive

ardise

28 30 32 34 36 38 40 42Age in months

Bogota -‐ poorest quartile Bogota -‐ richest quartileC ontrol G roup S timulation G roup

Impacts along the distribution - cognition

!!!!! !

-4 -2 0 2 4x

kdensity z_cog kdensity z_cog

-4 -2 0 2 4x

kdensity z_cog kdensity z_cog

Some other impacts

First Hint at Mechanisms: Increased Parental Investment in Children

•  Suggestive evidence of “crowding-in” of resources

!"#$%&'($%)"*+

,"-./0%)"*+

1'23$03$+%"4%56'*%

&'0$23'6+56'*%78039303$+

:;2$93"-+%<%('*+=,"">+%

4"2%7(-60+

?03#-6'03"@ !"#$%&& !"'(%& !"))*&& !")*%&& !"!$((+!"$(!, +!"$-%, +!"$'(, +!"$)', +!"!($,

?03#%A%&382"@-02 !".$#&& !"$*. !"%)'&& !".-$&& !"$%!+!"$(#, +!"$--, +!"$-., +!"$)-, +!"!(.,

&382"@-023$@0+ !"!((* !"--.& !"'$- !"'$. !"$!%+!"$(., +!"$)$, +!"$*., +!"$)-, +!"!(.,

/01$-'#20&345/46478/90890):20&&345/46478/90890$:

Using the model Estimating the model and interpreting the impacts

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Incorporating the role of the intervention in the framework

We now use the theoretical framework to interpret these results.

from Attanasio, Cattan, Fitzsimons, Meghir and Rubio (2014)

The purpose is to establish how the intervention obtained theobserved impacts.

Within the production function approach, the intervention could haveaffected the formation of skills:

By shifting the distribution of investmentsBy shifting the productivity of the inputs

We consider the impact of stimulation (κ = 1) vs. no stimulation(κ = 0) and let the investment functions and the productionfunctions to depend on the intervention.

The investment functions are specified as:

ln(I τi ,t) = λτκ,0 + λτκ,1 ln(θci ,t) + λτκ,2 ln(θsi ,t) + λτκ,3 ln(θm,c) +

λτκ,4 ln(θm,s) + λτκ,5 ln(Zi ,t) + uτi ,t , τ = M,T

where Z are variables that determine investment but have no directimpact in the production function:

Prices, wagesFamily resourcesFamily composition (but we do allow them in the prod. fun.)

The production function we estimate is:

θji ,t+1 = Ajκ[γj1,κθ

ci ,tρj + γj2,κθ

si ,tρj + γk3,κθ

m,ci ,t

ρj + γj4,κθm,si ,t

γj5,κIMi ,t+1ρj + γj6,κITi ,t+1

ρj ]1ρj eη

ji,t j = c , s

The joint distribution of factors (θt+1, θt , It+1,P) is allowed to differbetween the treatment and control samples.

Latent Factors distributions: child development

Figure 1: Kernel densities of latent factors

−2 −1 0 1 2

Density

TreatedControl

(a) Children’s cognitive skill at follow-up

−1.5 −1.0 −0.5 0.0 0.5 1.00.0

Density

TreatedControl

(b) Children’s non-cognitive skill at follow-up

−2 −1 0 1

Density

TreatedControl

(c) Children’s cognitive skill at baseline

−1.0 −0.5 0.0 0.5 1.0

Density

TreatedControl

(d) Children’s non-cognitive skill at baseline

Latent Factors distributions: investments

−3 −2 −1 0 1 2 3

Density

TreatedControl

(e) Material investment

−2 0 2 40.0

Density

TreatedControl

(f) Time investment

−2 −1 0 1 2 3

Density

TreatedControl

(g) Mother’s cognitive skille

−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

Density

TreatedControl

(h) Mother’s non-cognitive skill

The Investment Functions

Specification Basic With interactionsConstant 0.000 -0.001

[-0.026,0.03] [-0.033,0.044]Treat 0.249 0.254

[0.104,0.364] [0.108,0.386]Log child's cognitive skill (t) 0.094 0.035

[-0.076,0.279] [-0.286,0.275]Log child's cognitive skill (t) * Treat 0.029

[-0.269,0.555]Log child's non-cognitive skill (t) -0.224 -0.161

[-0.548,0.055] [-0.611,0.285]Log child's non-cognitive skill (t) *Treat -0.202

[-0.865,0.439]Log mother's cognitive skill 0.647 0.886

[0.522,0.876] [0.504,1.223]Log mother's cognitive skill * Treat -0.437

[-0.797,0.158]Log mother's non-cognitive skill -0.034 0.083

[-0.199,0.173] [-0.244,0.378]Log mother's non-cognitive skill * Treat -0.242

[-0.541,0.274]Log wealth (factor) (t) 0.955 0.339

[0.153,1.089] [-0.646,1.562]Log wealth (factor) * Treat 1.269

[-0.836,1.946]Married (t+1) 0.215 0.080

[0.049,0.292] [-0.067,0.384]Married * Treat 0.278

[-0.201,0.421]Log number of children (t+1) -0.213 -0.181

[-0.278,-0.121] [-0.296,-0.088]Log number of children * Treat -0.134

[-0.292,0.156]Log avg males wages (t+1) 0.013 -0.008

[-0.04,0.104] [-0.156,0.113]Log avg males wages (t+1) * Treat 0.031

[-0.12,0.28]Log avg females wages (t+1) 0.021 0.010

[-0.045,0.084] [-0.115,0.136]Log avg females wages (t+1) * Treat 0.022

[-0.16,0.167]

TABLE 8aEstimates of the material investment functions

Xujhnknhfynts Gfxnh \nym�nsyjwfhyntsx

Htsxyfsy QOQQQ NQOQQR

|NQOQSWMQOQT~ |NQOQTTMQOQUU~

Ywjfy QOSUZ QOSVU

|QORQUMQOTWU~ |QORQYMQOTYW~

Qtl�hmnqi,x�htlsnyn{j�xpnqq�-y.� QOQZU QOQTV

|NQOQXWMQOSXZ~ |NQOSYWMQOSXV~

Qtl�hmnqi,x�htlsnyn{j�xpnqq�-y.�/�Ywjfy QOQSZ

|NQOSWZMQOVVV~

Qtl�hmnqi,x�sts2htlsnyn{j�xpnqq�-y.� NQOSSU NQORWR

|NQOVUYMQOQVV~ |NQOWRRMQOSYV~

Qtl�hmnqi,x�sts2htlsnyn{j�xpnqq�-y.�/Ywjfy NQOSQS

|NQOYWVMQOUTZ~

Jxynrfyjx�tk�ymj�rfyjwnfq�ns{jxyrjsy�kzshyntsx

Specification Basic With interactionsConstant -0.005 -0.003

[-0.027,0.036] [-0.034,0.045]Treat 0.330 0.304

[0.193,0.428] [0.149,0.407]Log child's cognitive skill (t) 0.023 0.083

[-0.157,0.215] [-0.165,0.266]Log child's cognitive skill (t) * Treat -0.174

[-0.571,0.238]Log child's non-cognitive skill (t) -0.103 0.015

[-0.368,0.142] [-0.548,0.453]Log child's non-cognitive skill (t) *Treat -0.239

[-0.873,0.472]Log mother's cognitive skill 0.381 0.543

[0.263,0.611] [0.202,0.869]Log mother's cognitive skill * Treat -0.262

[-0.594,0.296]Log mother's non-cognitive skill 0.077 0.033

[-0.061,0.301] [-0.255,0.311]Log mother's non-cognitive skill * Treat 0.010

[-0.297,0.488]Log wealth (factor) (t) 0.644 -0.189

[-0.142,0.831] [-1.012,1.008]Log wealth (factor) * Treat 1.278

[-0.907,2.177]Married (t+1) 0.163 -0.014

[0.001,0.235] [-0.187,0.25]Married * Treat 0.305

[-0.159,0.51]Log number of children (t+1) -0.163 -0.179

[-0.211,-0.077] [-0.304,-0.101]Log number of children * Treat -0.011

[-0.141,0.268]Log avg males wages (t+1) -0.087 0.065

[-0.156,-0.001] [-0.075,0.214]Log avg males wages (t+1) * Treat -0.239

[-0.466,-0.055]Log avg females wages (t+1) 0.048 -0.031

[-0.014,0.111] [-0.168,0.067]Log avg females wages (t+1) * Treat 0.129

[0.003,0.341]

TABLE 8bEstimates of the time investment functions

Xujhnknhfynts Gfxnh \nym�nsyjwfhyntsx

Htsxyfsy NQOQQV NQOQQT

|NQOQSXMQOQTW~ |NQOQTUMQOQUV~

Ywjfy QOTTQ QOTQU

|QORZTMQOUSY~ |QORUZMQOUQX~

Qtl�hmnqi,x�htlsnyn{j�xpnqq�-y.� QOQST QOQYT

|NQORVXMQOSRV~ |NQORWVMQOSWW~

Qtl�hmnqi,x�htlsnyn{j�xpnqq�-y.�/�Ywjfy NQORXU

|NQOVXRMQOSTY~

Qtl�hmnqi,x�sts2htlsnyn{j�xpnqq�-y.� NQORQT QOQRV

|NQOTWYMQORUS~ |NQOVUYMQOUVT~

Qtl�hmnqi,x�sts2htlsnyn{j�xpnqq�-y.�/Ywjfy NQOSTZ

|NQOYXTMQOUXS~

Jxynrfyjx�tk�ymj�ynrj�ns{jxyrjsy�kzshyntsx

QOTTQ QOTQU

QOQST QOQYT

The Production Functions

Without control function

With control function

Baseline cognitive skills 0.707* 0.646*[0.664,0.778] [0.606,0.761]

Baseline non-cognitive skills 0.028 0.045[-0.056,0.138] [-0.042,0.153]

Mother's cognitive skills 0.103* -0.123[0.038,0.182] [-0.174,0.126]

Mother's non-cognitive skills 0.119* 0.055[0.026,0.184] [-0.018,0.152]

Material investments 0.056* 0.277*[0.02,0.084] [0.005,0.315]

Time investments -0.021 0.065[-0.056,0.012] [-0.09,0.178]

Number of children in household 0.007 0.034[-0.012,0.022] [-0.003,0.042]

Control function for material investments - -0.24[-0.295,0.049]

Control function for time investment - -0.097[-0.229,0.07]

Complementarity parameter 0.027 0.094[-0.156,0.263] [-0.053,0.243]

Elasticity of substitution 1.027* 1.104*[0.865,1.356] [0.949,1.321]

Productivity parameter (A) 0.996* 0.986*[0.986,1.008] [0.978,1.004]

Productivity parameter interacted with treatment 0.064* -0.022[0.006,0.128] [-0.055,0.102]

Note: 90% confidence intervals in brackets based on 200 bootstraps. * significant at the 10% level

TABLE 6aEstimates of the CES production function for cognitive skill

The Production Functions

Without control function

With control function

Baseline cognitive skills 0.156* 0.148*[0.102,0.277] [0.03,0.291]

Baseline non-cognitive skills 0.611* 0.536*[0.424,0.705] [0.371,0.678]

Mother's cognitive skills -0.047 0.012[-0.09,0.034] [-0.27,0.21]

Mother's non-cognitive skills 0.134* 0.034[0.02,0.27] [-0.072,0.184]

Material investments 0.073* -0.319[0.035,0.105] [-0.418,0.09]

Time investments 0.048* 0.578*[0.014,0.085] [0.198,0.724]

Number of children in household 0.025 0.012[-0.008,0.077] [-0.014,0.087]

Control function for material investments - 0.41[-0.008,0.509]

Control function for time investment - -0.564[-0.738,-0.16]

Complementarity parameter -0.107 0.013[-0.29,0.15] [-0.072,0.056]

Elasticity of substitution 0.904* 1.013*[0.775,1.176] [0.933,1.059]

Productivity parameter 1.005* 1.000*[0.989,1.023] [0.99,1.009]

Productivity parameter interacted with treatment -0.009 -0.09[-0.04,0.027] [-0.176,0.023]

Note: 90% confidence intervals in brackets based on 200 bootstraps. * significant at the 10% level

TABLE 6bEstimates of the CES production function for non-cognitive skill

Two Digressions

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Two Digressions Structural Models and RCTs

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Why a structural model?

We have shown the impacts of a cluster RCT.

As such they are simple to obtain and easy to explain.

What is the use of the structural model?

It helps us to understand the mechanisms through which the programoperates.

These are key for the design of policy.

The structure we have built is also a useful way to synthetise muchinformation.

Notice that we do not use the RCT to identify the productionfunction or the investment rule.

This makes necessary the use of an ’instrumental variable’ approachto take into account the role of parental investment.

We therefore need exogenous sources of variation in the determinantsof investment.

Two Digressions Measurement

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Measurement issues

Measurement is central to the approach we have taken.

Appropriate measurement can be used to solve identification problemsand estimate richer and more realistic models.

Much has to be learned in terms of developing appropriate measures:1 Much work needs to be done in developing measures in some crucial

areas.2 New technologies can and should play a central role in developing new

measures.3 Theory should inform the development of suitable measures.4 It might be worthwhile to go beyond the standard ‘revealed preference’

approach.

I will give some examples of these issues.

Measurement issues

approach.

Measurement issues

areas.

2 New technologies can and should play a central role in developing newmeasures.

3 Theory should inform the development of suitable measures.4 It might be worthwhile to go beyond the standard ‘revealed preference’

approach.

Measurement issues

measures.

3 Theory should inform the development of suitable measures.4 It might be worthwhile to go beyond the standard ‘revealed preference’

approach.

Measurement issues

measures.3 Theory should inform the development of suitable measures.

4 It might be worthwhile to go beyond the standard ‘revealed preference’approach.

Measurement issues

approach.

Measurement issues

approach.

Measuring child development is hard

Measuring young children development accurately is very hard.

Some of the measures that are considered the ‘gold standard’ are veryexpensive

The Bayley scales of infant development (BSID) take about 1.5 hoursto administerThey need to be administered by a specially trained psychologistThey cannot be administered in the child’s home but in standardisedsettings.

Unfortunately alternative ‘cheap’ measures can be very noisy.

How good are cheap measures?

Results from Araujo, Attanasio, and Rubio-Codina (2014, in progress)

0.368 0.392

6 - 18 meses 19 - 30 meses 31 - 42 meses

Correlations with Bayley Cognition

ASQ3 Res. Problemas (0.160)

ASQ3 Comunicación (0.218)

Battelle Cognitivo (0.302)

Battelle Lenguaje (0.308)

Denver Mot. Fina (0.316)

Denver Lenguaje (0.274)

How good are cheap measures?

Results from Araujo, Attanasio, and Rubio-Codina (2014, in progress)

6 - 18 meses 19 - 30 meses 31 - 42 meses

Correlations with Bayley Expressive Language

ASQ3 Comunicación (0.406)

Denver Lenguaje (0.470)

Battelle Lenguaje (0.472)

MacArthur Leng. Exp.* (0.475)

* 8 - 30 meses

New technologies

Much is happening and new tools are being developed.

Some examples:

Neil Marlowe’s PARCA-R, originally developed for premature babies,claims correlation with the Bayley of 0.8. (Martin et al. Arch.Child.Dis.2012)

Anne Fernald’s eye tracking test seems to be very predictive of lateracademic outcomes. (Fernald et al. Dev. Psych. 2006)Functional near infrared spectroscopy (fNIRS) with portable devices:Fox-Lloyd et al. (Nature, 2014) implemented it in the Gambia.

There might be some low hanging fruits to be reaped.

New technologies

Some examples:

Neil Marlowe’s PARCA-R, originally developed for premature babies,claims correlation with the Bayley of 0.8. (Martin et al. Arch.Child.Dis.2012)Anne Fernald’s eye tracking test seems to be very predictive of lateracademic outcomes. (Fernald et al. Dev. Psych. 2006)

Functional near infrared spectroscopy (fNIRS) with portable devices:Fox-Lloyd et al. (Nature, 2014) implemented it in the Gambia.

New technologies

Some examples:

Neil Marlowe’s PARCA-R, originally developed for premature babies,claims correlation with the Bayley of 0.8. (Martin et al. Arch.Child.Dis.2012)Anne Fernald’s eye tracking test seems to be very predictive of lateracademic outcomes. (Fernald et al. Dev. Psych. 2006)Functional near infrared spectroscopy (fNIRS) with portable devices:Fox-Lloyd et al. (Nature, 2014) implemented it in the Gambia.

New technologies

Some examples:

Neil Marlowe’s PARCA-R, originally developed for premature babies,claims correlation with the Bayley of 0.8. (Martin et al. Arch.Child.Dis.2012)Anne Fernald’s eye tracking test seems to be very predictive of lateracademic outcomes. (Fernald et al. Dev. Psych. 2006)Functional near infrared spectroscopy (fNIRS) with portable devices:Fox-Lloyd et al. (Nature, 2014) implemented it in the Gambia.

Theory should inform the choice of measurement tools

Appropriate measurement tools can solve identification problems.

This is obvious when some specific ‘unobservables’ become (partially)observable.

There is a more subtle sense in which appropriate design can helpidentification.

In the factor models we have used, non-parametric identificationrequires multiple measures with uncorrelated measurement errors.

This has been discussed in the literature:

Schennach (ECA, 2004)Browning and Crossley (2009) ”Are Two Cheap, Noisy MeasuresBetter Than One Expensive, Accurate One?,” AEA P&P

The assumption of uncorrelated measurement errors in differentmeasures can be made realistic by specific design:

In the case of the Colombia study we have different measures ofcognitive development (Bayleys’ scales, MacArthur Language tests,ASQ, etc.)Some of them are administered by different individuals in different days.Some of them are based on the observation of the child , other arebased on maternal reports.

Additional examples can be constructed with respect to the loading ofdifferent factors.

Beyond revealed preferences: observing unobservables

Another dimension in which measurement tools can be developed isby eliciting individual responses on:

tastes, preferences, attitudesexpectationsbeliefs

This approach involves often going beyond ‘revealed preferences’.

In addition to choices, questionnaires elicit other dimensions oftenbased on hypotheticals.

This approach has not been used in the profession....

... but it has been pioneered by a number of scholars:

Tom Juster and researchers in MichiganChuck Manski on subjective expectations.Willingness to pay.

Flavio Cunha has been asking questions about the return to parentinginvestment in Philadelphia (Cunha et al., 2014)

In collaboration with Flavio Cunha we have been eliciting beliefs onthe returns to parenting in Colombia.

Flavio Cunha has been asking questions about the return to parentinginvestment in Philadelphia (Cunha et al., 2014)In collaboration with Flavio Cunha we have been eliciting beliefs onthe returns to parenting in Colombia.

Beliefs

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Beliefs

The importance of beliefs and attitudes

Some interventions have an impact without providing targetedindividuals any resources, beside information.

The intervention in Colombia, according to our results, did notchange the production function, did not make parents more effective.

Instead, it increased investment.

But then the question is: why did parents not invest before?

One possibility is that they were not aware of the productivity ofinvestment.

In the model we sketched before, parents’ investment depend on theirtastes and their perception of the production function.

Beliefs

A model of distorted beliefs: a parametric example

from Attanasio and Cattan (in progress).

Suppose parents maximise utility subject to a budget constraint andtheir perceived production function.

Suppose that utility is the sum of two power functions.

The production function of HK is Cobb Douglas .

Distorted beliefs are about the parameters of the Cobb Douglas.

U(C ,H) =C 1−θ

1− θ+

H1−φ

1− φC = Y − pI

H = Hα0 I β

H = H α0 I β

Beliefs

U(C ,H) =C 1−θ

1− θ+

H1−φ

1− φC = Y − pI

H = Hα0 I β

H = H α0 I β

Beliefs

U(C ,H) =C 1−θ

1− θ+

H1−φ

1− φC = Y − pI

H = Hα0 I β

H = H α0 I β

Beliefs

Substituting the budget constraint solved for C in the utility functionand the production function:

(Y − pI )θH αφo I βφ

F.O.C.βH

α(1−φ)0 I β(1−φ)−1 = p(Y − PI )−θ

Taking logs and re-arranging this yields:[β(1− φ)− 1

]lnI = lnp − θln(Y − pI )

− lnβ − α(1− φ)lnH0

These choices are not necessarily optimal, unless α = α and β = β.

Beliefs

F.O.C.βH

α(1−φ)0 I β(1−φ)−1 = p(Y − PI )−θ

Beliefs

F.O.C.βH

α(1−φ)0 I β(1−φ)−1 = p(Y − PI )−θ

Beliefs

F.O.C.βH

α(1−φ)0 I β(1−φ)−1 = p(Y − PI )−θ

Beliefs

Identification of distorted beliefs

Several strategies are possible to identify the distorted belief.

The most obvious is the elicitation of preferences

This is what we do in Attanasio, Cunha and Jervis (in progress) orwhat Cunha et al. (2014) do.

Alternatively, we can use interventions where information provision issuccessful to identify distorted preferences.

The idea is to use:

outcomes on HK and investment choices of ‘informed parents’ toidentify the production function and taste parameters.investment choices of ‘uninformed’ parents to identify the parametersof the distorted production function.

A RCT is very useful because allows us to assume that the tasteparameters of ‘treated’ and ‘control’ parents are the same.

Beliefs

The idea is to use:

Beliefs

The idea is to use:

Beliefs

The idea is to use:

Beliefs

The idea is to use:

Beliefs

Estimation of distorted beliefs

We can rewrite the log investment function:

ln I = bκ0 + bκ1 ln p + bκ2 ln(Y − pI ) + bκ3 ln H0, κ = T ,C

where:

bκ0 =lnβκ

1− βκ(1− φ)

bκ1 = − 1

1− βκ(1− φ)

bκ2 =θ

1− βκ(1− φ)

bκ3 =ακ(1− φ)

1− βκ(1− φ)

We can estimate this equation by OLS.Orazio P. Attanasio ( UCL and IFS) Determinants of HK formation GSE - Barcelona 27.11.14 70 / 76

Beliefs

This can be done both for treatment and control sample, so to estimatebκ’s for κ = T ,C ’s.

By doing so, if we assume that βT = β we identify:

1/bC1 + 1

1/bT1 + 1

bC3 /bC

bT3 /bT

Beliefs

The ‘true’ values for the parameters of the production function can beestimated directly relating outcomes to investment.

Notice that for such a purpose we can use both treatment and controlsamples.

Investment choices are treated as endogenous, so that we will have tointerment them.

To get consistent estimates, we need the first stage regressions tocontain full interactions with treatment.

If we are not willing to assume that the treatment does not affect theproduction function, we will need additional instruments to identifythe production function parameters.

The model indicates household resources or, even more convincingly,prices as a source of identification.

Beliefs

Preliminary results for Colombia

Using the estimates above, we can recover the ratios of productionfunction parameters.

1/0.941 + 1

1/0.684 + 1= 0.84

0.170/0.941

0.0824/0.684= 1.50

This is consistent with anecdotal evidence:

poor parents tend to overestimate the importance of initial condition.poor parents tend to underestimate the importance of investment,especial for ‘normal’ children.Annette Lareau: ‘Unequal Childhoods’: Concerted Cultivation vNatural Growth.

Beliefs

Preliminary results for Colombia

Using the estimates above, we can recover the ratios of productionfunction parameters.

1/0.941 + 1

1/0.684 + 1= 0.84

0.170/0.941

0.0824/0.684= 1.50

This is consistent with anecdotal evidence:

poor parents tend to overestimate the importance of initial condition.poor parents tend to underestimate the importance of investment,especial for ‘normal’ children.Annette Lareau: ‘Unequal Childhoods’: Concerted Cultivation vNatural Growth.

Beliefs

Very preliminary results for Colombia

... and it is consistent with some very preliminary evidence on elicitedreturns from our new data.

!! !! !! !! !! !!Maternal!Investment!

!! !! Mean! St.!Dv.! Min! Max!

Ra3os!(low/high!inv.)!for!LESS!developed!

Easy! 1.21! 0.32! 0.31! 4.00!

More!difficut! 1.21! 0.28! 0.30! 3.17!

Even!more!difficult! 1.22! 0.28! 0.34! 4.00!

Ra3os!(low/high!inv.)!for!MORE!developed!

Easy! 1.10! 0.24! 0.25! 2.50!

More!difficut! 1.11! 0.22! 0.33! 3.27!

Even!more!difficult! 1.13! 0.22! 0.34! 2.45!

!! !! !! !! !! !!*!1,252!observa3ons.! !! !! !! !!**!On!average,!a!86.90%!answered!in!the!correct!way!(ra3o!is!higher!than!1)!for!LESS!developed.!***!On!average,!a!83.5%!answered!in!the!correct!way!(ra3o!is!higher!than!1)!for!MORE!developed.!!! !! !! !! !! !!

Conclusions

Outline

2 Known and Unknown

6 Beliefs

7 Conclusions

Conclusions

The study of human capital accumulation is important, especially inthe context of developing countries.

We know much:Early years are important.Early years are malleable.Targeted and well-designed interventions can be effective.

There is much we do not know.The process of HK formation is complex: dynamics, the role ofnutrition, genetics and epigenetics.What do parents do and why?How to build scalable interventions?

A research agenda:Better measurement tools.Use of structural models to identify the production function andparental behaviour.Instrumental to the design of effective policies.

Conclusions

We know much:Early years are important.

Early years are malleable.Targeted and well-designed interventions can be effective.

Conclusions

We know much:Early years are important.Early years are malleable.

Targeted and well-designed interventions can be effective.

Conclusions

There is much we do not know.

The process of HK formation is complex: dynamics, the role ofnutrition, genetics and epigenetics.What do parents do and why?How to build scalable interventions?

Conclusions

There is much we do not know.The process of HK formation is complex: dynamics, the role ofnutrition, genetics and epigenetics.

What do parents do and why?How to build scalable interventions?

Conclusions

There is much we do not know.The process of HK formation is complex: dynamics, the role ofnutrition, genetics and epigenetics.What do parents do and why?

How to build scalable interventions?

Conclusions

the determinants of human capital formation during the early years of life

Economy & Finance