1 | P a g e
The effect of breastfeeding on children’s early cognitive
outcomes: Evidence from an Irish Cohort Study
-Master Thesis Economics- July 2017
University of Amsterdam
David Mooney
MSc. Economics (Public Policy)
Prof. Dr. Hessel Oosterbeek Supervisor
2 | P a g e
Statement of Originality
This document is written by student David Mooney who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
3 | P a g e
Abstract
This thesis uses an alternative approach to identify the cognitive outcomes of breastfed children at age five. The theory behind the alternative approach is to restrict the comparison group of those who did not breastfeed to those who were not able to breastfeed. By restricting the comparison group, the self-selection bias reduces as individuals could not choose to breastfeed due to either mother’s illness, baby’s illness, or an inability to produce milk. A more restricted control group is created for “Inability to produce milk” in order to limit endogeneity concerns. Using this approach, the cognitive outcomes for math ability are shown to drop from a large and significant coefficient to an estimate close to zero. Results on reading ability are shown as not statistically significant using the alternative approach once confounders such as income and maternal education are accounted for. Propensity score matching results are also provided estimating the average treatment effect and the average treatment effect on treated for those breastfed.
4 | P a g e
Contents
Section Pages
I. Introduction 5-7
II. Literature Review 7-10
III. Data 11-13 IV. Methodology 13-23 V. Empirical Results 24-34 VI. Discussion 34-36 References 36-37 Appendix 38-40
5 | P a g e
I. Introduction
To breastfeed is an inevitable decision every mother must make when raising a child from infancy. Not only do mothers decide on whether or not to breastfeed their child, but also on the number of days of feeding. Immediate advantages are evident cost-wise, as breastfeeding remains the cheapest option for parents despite the increased time and effort. In terms of health benefits, the dosage in breastmilk is biologically engineered to provide the child with the exact nutrients and antibodies needed in optimal quantities (Allen & Hector, 2005). Policy makers’ recommended time for exclusive feeding is six months, with complimentary foods and breastfeeding introduced thereafter (World Health Organization, 2001). Nevertheless, a natural variation occurs in the length of breastfeeding because mothers are either unable to or choose not to breastfeed for the recommended duration. This variation is also seen on a global scale, where countries who offer the most in terms of paid maternity leave, showing a greater percentage of new-borns being breastfed for the recommended duration1 (Save the Children, 2012).
In terms of cognitive advantages, this question still remains relatively unclear and contradictory in the existing literature. As explored later on, a number of studies show that breastfed children have higher life and cognitive outcomes than those who were not breastfed (Belfort, et al., 2016) (Anderson, Johnstone, & Remley, 1999) (Horwood & Fergusson, 1998). However, the causal relationship between breastfeeding and cognitive outcomes is difficult to establish due to the ethical concerns surrounding randomized control trials of infant feeding methods. Since breastfeeding is not a random event, but a choice made by parents, comparisons are usually made between two groups of individuals; the children who were breastfed and the children who were not. This comparison is seen as the “traditional approach”. Nonetheless, issues surrounding sample selection bias occur without proper randomization, as those who breastfeed may be inherently different from those who do not. While analysis is restricted to observational studies, different methods exist to limit the degree of sample selection bias. This thesis aims to provide an additional alternative method to reduce bias in estimating the cognitive benefits of breastfeeding.
The methodology of the alternative approach is to reduce the comparison group of the treatment (breastfed) group to those who did not breastfeed based on their response to the question “Why
1 Figure A1 in the Appendix shows breastfeeding and paid maternity leave rates based on the “State of the
6 | P a g e
not breastfed”? If the choice to breastfeed was out of the mother’s control as the mother was ill, the baby was ill, or they were unable to produce milk, then respondents were included in the “alternative” control group. Respondents who did not breastfeed for different reasons than mentioned above were excluded from the alternative control group. A sub-set of this alternative control group is also created using the “not enough milk” reason based on potential limitations of the “illness” reasons. This is referred to as the “milk” control group. Using this methodology, the aim is to limit the sample selection bias as respondents were not able to self-select breastfeeding and therefore, should be more comparable to the treatment group. This methodology is also compared to a more conventional and common technique in reducing sample selection bias using propensity score matching (PSM).
Comparisons using the traditional, alternative, and matching approaches are made using longitudinal survey data obtained from the Growing Up in Ireland (GUI) data set. This data set follows the lives of over 11,000 infants born in the Republic of Ireland with primary data being recorded at 9 months of age. The approaches are tested using three distinct models based on previous literature. Under the most stringent of the models, the results show a difference in approaches. For children who were breastfed, a significant 2.98 percentage point increase in the probability of being above average in math is shown using the traditional approach. The alternative approach does not show a statistically significant effect for breastfeeding on math ability, with estimates being close to zero for both the alternative and milk control groups. These results are compared against the more conventional approach of PSM, which shows a significant average treatment effect of 2.85%. When specifications are run for reading ability, estimates are shown to be higher using the alternative approach, than the traditional approach when controlling for maternal education and household income. These results are significant at the 10% level and do not yield as statistically significant once the comparison group is restricted to the milk control group. PSM results for reading ability displays a significant average treatment effect of 4.78%.
Using the traditional and alternative methods, specifications are also re-run for both methods with the addition of the duration of breastfeeding variable, which is shown to correct for upward bias present in the breastfeeding coefficient. The gap between intensive margins, defined as the amount of days spent breastfeeding, and extensive margins, defined as breastfed or not, are also shown to reduce when the sample is restricted for math ability, while no real change is shown in reading estimates.
7 | P a g e
The remainder of this thesis is organized as follows; the next section explores the previous literature surrounding the estimates of breastfeeding as well as the duration of breastfeeding. Section III and Section IV presents the data and methodology used in this thesis as well as descriptive statistics on the comparison groups. Section V provides the empirical results of the traditional, alternative and Propensity Score Matching (PSM) methods. Section VI gives a brief discussion and conclusion on the results found in this thesis.
II. Literature Review
Hoefer and Hardy (1929) were the first to examine any post-infancy difference in physical and mental development among breastfed and artificially fed children. This paper was influential as many papers up to this point had only looked at breastfeeding on health outcomes for infants in their first two years of life. It was first to use what is referred to in this thesis as the “traditional approach”, where post-infancy outcomes were compared on the margin between those who breastfed and those who did not.
This section identifies the concerns surrounding the traditional approach such as sample-selection bias and examines some of the more recent studies that try to mitigate this problem.
Section 2.1 Randomization
While randomization of breastfeeding against artificial feeding is argued as unethical, randomization into a breastfeeding promotion programme is seen as a more feasible option. To date, only one randomized study has taken place on breastfeeding techniques and childhood health outcomes. A cluster-randomization trial on breastfeeding promotion interventions took place in Belarus modelled off the World Health Organization/UNICEF baby friendly hospital initiative (Kramer, et al., 2001). Over thirty one Belarusian maternity hospitals and 17,046 breastfeeding infants were enrolled into the programme. The initiative promoted exclusive breastfeeding and extended duration among mothers in maternity hospitals, which were randomly selected for the initiative. The experimental intervention lead to a large increase in exclusive feeding methods, with at 3 months old, 43.3% of the experimental group reporting exclusive feeding, compared to 6.4% in the control group. The study also showed a significantly higher prevalence of any breastfeeding up to and including 12 months.
The randomization lead to a follow up study to test the potential for bias when randomization is not possible by analyzing data ignoring treatment (Kramer, et al., 2002). The two randomized groups were combined to compare infants who weaned after the first 30 days and those who
8 | P a g e
lasted the full duration of 12 months. The initial results – based on randomization – showed that prolonged and exclusive breastfeeding inhibited faster weight and length gain. However, the observational data reported faster weight and length gain with early weaning and slower gains through prolonged exclusive feeding. The authors attribute these differences to unmeasured confounding or an actual biological effect of formula feeding. In terms of cognitive development, this same cohort was investigated further at 6.5 years of life to see whether breastfeeding had any impact on IQ scores, as well as reading, writing, and mathematical ability, which was assessed via teacher’s evaluation (Kramer, et al., 2008). Of the initial sample only 13,889 (81.5%) were assessed in this follow up study2. The results showed that from the initial intervention, the experimental group had higher cluster adjusted mean differences of + 7.5 for verbal IQ, + 2.9 for performance IQ, and + 5.9 for overall IQ. Teacher’s evaluation of infants showed academic cognitive ability to be higher among the experimental group for both reading and writing.
This randomization poses a unique setting and is free from the self-selection bias discussed in the next section. However, the external validity and applicability of these results is hard to determine as this experiment has not been replicated in further settings.
Section 2.2 Sample Selection Bias
Sample selection bias occurs when there is an absence of randomization, or when randomization is not conducted properly. In the context of breastfeeding, the sample selection bias appears when individuals self-select whether to breastfeed or not. This self-selection produces difficulty in distinguishing whether the treatment (e.g. breastfeeding), or the non-random factors that affect being assigned to the treatment, had an effect on the outcome. Confounding issues also appear where unobservable characteristics influence an individual’s choice to enrol in the treatment, as well as the outcome variable. Failing to account for these biases can lead to an under or overestimation in the parameter of observation, which in this case is the effect of breastfeeding.
While both of these issues remain a key problem in observational studies, a number of papers have tried to limit these biases through various methods and models. Evenhouse and Reily (2005) use data on over 2,734 sibling pairs from a national longitudinal study to limit the bias
9 | P a g e
in response to cognitive, physical health, and emotional health outcomes. For each outcome variable the authors create a between-family model as well as a within-family model. The between-family model follows more in line with the traditional approach where indicators between children of families who did and did not breastfeed are compared. The within-family model investigates whether differences in sibling breastfeeding histories are associated with differences in cognitive and health outcomes. The idea behind the within family model is that individual specific characteristics, such as maternal education and household income, are controlled for. Thus, if self-selection occurs between siblings, it is less likely due to these unobservable characteristics. The findings showed that nearly all of the correlations found significant in the between-family model, showed as insignificant in the within family model. The only exception to this finding was cognitive ability, which showed a persistent positive correlation with breastfeeding, whether breastfeeding was measured in terms of duration, or as a yes/no binary variable.
A similar study by Der, Batty, and Deary (2006) assesses the impact of maternal intelligence and other important confounders in determining the link between breastfeeding and children’s academic ability. The study regresses important confounders, such as maternal age, race, and household income simultaneously, inclusive and exclusive of maternal IQ. Results showed that pre-adjustment for maternal IQ, breastfeeding was associated with an increase of around 4 percentage points in a child’s academic ability. Adjusting for maternal IQ showed to account for most of the effect. The authors also perform a meta-analysis that explicitly controls for maternal IQ, including one report that directly controls for paternal intelligence via sibling comparison. The meta-analysis showed that the studies with the highest breastfeeding coefficients - adjusted for maternal IQ - were those that controlled for fewer important confounders, such as household income and socio-economic status.
Nonetheless, using sibling comparison prompts a different set of econometric issues. Davanzo, Starbird, and Leibowitz (1990) examine data on women with at least two children to see if experiences in breastfeeding with the first child affects whether subsequent children are breastfed.The results indicate that women often repeat the decision made with their first child with later children. Mothers who did not breastfeed their first child were more likely to switch for successive children the higher their education levels were beyond secondary school. Conversely, women who breastfed their first child are less likely to breastfeed later-born children if the first breastfeeding duration was shorter than six months, or if the woman had not gone beyond high school. Therefore, without properly controlling for birthing order or first
10 | P a g e
feeding duration, the results can be biased for within sibling outcomes as this influences future decisions to self-select into breastfeeding.
Section 2.3 Duration & Propensity Score Matching
Within the group that chooses to breastfeed, we also see variation among individuals in the duration of breastfeeding. In a 2011 paper by McCrory & Layte - using a later cohort of the “Growing Up in Ireland” data set - the traditional approach is used to see if the length of breastfeeding impacts standardized test scores at age 9. This allows for an intensity margin to be established where differences in length of feeding can potentially influence test scores. However, even restricting the sample to individuals who breastfed, self-selection is also present as mother’s age and education is highly correlated with breastfeeding duration (Quarles, Williams, Hoyle, Brimeyer, & Williams, 1994). The authors combat this problem by creating different models and introducing covariates at new stages to account for changes in the breastfeeding point estimate. This thesis follows a similar methodology, whereby models are based around the literature to account for significant confounders as they are added into the model. The results indeed showed that adjusting for a range of child, maternal, and socio-environmental characteristics, breastfed children enjoy a significant test score advantage of 3.24 and 2.23 percentage points in math and reading ability. Any amount of breastfeeding was associated with a significantly higher test score advantage, however evidence of a dose-response relationship - based on the length of feeding - was weak3.
The study also uses a technique of propensity score matching in order to test the robustness in test scores advantages among breastfed children found in the initial results. This technique involves trying to negate the sample selection bias, by restricting the comparison group to those who share similar pre-treatment characteristics to the treatment group. If in theory, the effects of breastfeeding are being overestimated by important confounders - such as maternal education and household income - then when groups are balanced so that they share these similar pre-treatment4 characteristics the effects of breastfeeding should diminish in
magnitude. The results however, showed that the test score advantage of breastfed children was indeed robust, and that the magnitude effects varies across groups, being largest in the most socially disadvantaged and falling to near zero in the most advantaged group.
3 The authors do associate the weak dose response results to the inability to disentangle between children who
were exclusively breastfed and those that are complementary fed. i.e. children who are exclusively fed should receive more of a “dose” of breastmilk than those that are fed in addition to supplements.
11 | P a g e
III Data
3.1 Growing up in Ireland data setThe micro data used to analyze the cognitive outcomes of breastfeeding was acquired from the Irish Social Science Data Archive, 2017 (ISSDA). The data set, “Growing up in Ireland” (GUI), uses longitudinal survey data to follow two cohorts of individuals from infancy and childhood. The information collected is used as a representative sample of the children and infants growing up in the Republic of Ireland on a national level. The study itself is conducted from a conglomerate of researchers from the Economic and Social Research Institute (ESRI) and Trinity College Dublin (TCD). For the purpose of this research, the infancy cohort is solely used to analyze the effects of breastfeeding on higher cognitive ability. This is done in order to limit distortion by parental recall of breastfeeding length.
The infancy cohort provides data on the same child starting at nine months of age in 2009, age three in 2011, and age five in 2013. Children were randomly selected from the Child’s Benefit Registrar (provided by the Department of Social and Family Affairs) so as to be 9 months of age at time of interview. The initial sample consisted of over 11,134 infants at wave one, with this dropping to 9,793 at wave two, and 9,001 for wave three. The main reason for dropouts in sample size were attributed to either non-response, moving outside the jurisdiction of study, or death of the infant (ESRI, 2011). Since there is no reason to believe that attrition is based on the decision to breastfeed, the 9,001 observations at wave three should still be a representative sample. Figure 1 displays the number of children breastfed in wave one and wave three of the data set5. Currently, data regarding wave four at age 7 has not been made publically available and fieldwork for the pilot phase of wave five of the infant cohort is underway6.
In the first two waves, the surveys are only filled out by the child’s Primary Care Giver (PCG), which is defaulted to the mother, and provides information at the household level for income, parental education levels, birth characteristics, and infant characteristics. At the third wave, the child has already entered their first year of primary school and their respective teachers and principals fill out questionnaires about the type of school, learning environment, and child’s cognitive ability. The PCG also fills out information on behalf of the child in wave three similar to the previous waves.
5 Graph only includes children where information on breastfeeding was available.
12 | P a g e
Figure 1: Number of children breastfed in wave one and wave three of the GUI data set.
3.2 Key Variables & Design
The initial design of this study is broken down into three parts, differing only in the comparison group used to analyze the effects of breastfeeding on higher cognitive ability. In the first part, the distinction between the two groups is made between individuals who did breastfeed and those who did not, using wave one of the GUI data set. This is referred to as the traditional approach which I use to create the “traditional control group”. The second part involves narrowing down the comparison group from the established non-breastfed group to those where breastfeeding was outside the mother’s control. This methodology follows a similar approach to Leuven and Oosterbeek (2008), wherein the comparison groups are refined to individuals who partook in a singular training, and a control group who missed training due to a random event. This “alternative” methodology used in the second part is further expanded in section
4.2 of this thesis. Thirdly, propensity score matching is used to provide a final control group,
who are similar to the treated participants in all relevant pre-treatment characteristics.
After splitting individuals into breastfed and non-breastfed groups (treatment and control respectively), individuals are assessed on their academic ability using wave three of the GUI data set. The data provided to assess the child’s ability in math and reading is initially split into five responses: “well above average”, “above average”, “average”, “below average”, and “well below average”. The above average response is used as the cut-off point for establishing higher cognitive ability. With the cut off established, “well above average” and “above average” are
13 | P a g e
grouped to create the variable of interest of “higher cognitive ability”. All other responses are used to represent average or lower cognitive ability.
3.3 Missing Cases and Data Limitations
While the full sample size population started at 9,001 individuals, a number of participants had to be excluded from the study. Firstly, if information pertaining to breastfeeding was not available for a particular child they were not included in the sample.7 Secondly, if any child’s
ability could not be assessed in either math or reading, they were subsequently dropped from further analysis in that particular subject. This meant that any teacher who answered as “don’t know”, “not applicable”, or failed to answer entirely, excluded the child from consideration. Lastly, information on household income could also not be attained for every child in the data. In order to obtain the consistency of observations across all three models, observations were dropped where information on household income could not be attained. After these missing cases had been excluded, this left the entire sample size at 8,567 individuals, with 7,897 and 7,083 participants for analysis on math and reading ability respectively8.
As of the 15th of July 2017, information is only available for the infant cohort at age five. Information on the infant cohort for later ages has either not been collected or made publically available by the ISSDA. Therefore, analysis is restricted to the time infants enter their first year of academia. Follow-up research is needed to examine whether the effects of breastfeeding (if any) persist or diminish into later academic periods.
IV Methodology
The empirical strategy for this section proceeds as follows. First, the Ordinary Least Squares methodology and models are explained to estimate the effects of breastfeeding on higher cognitive ability, for the traditional and alternative control groups. Second, the alternative approach is described in detail, which is used to create the “alternative” and “milk” control groups. Third, Propensity Score Matching (PSM) is tested to see if matching is eligible, and another comparison group can be created. Fourth, descriptive statistics are provided for the respective control groups.
7 Information on breastfeeding could only not obtained for a total of 3 children in the entire study.
14 | P a g e
4.1 Models
Table 1 shows the three models I use to regress breastfeeding on higher cognitive ability. Model 1 follows a simple bivariate regression model where breastfeeding is solely regressed on higher cognitive ability. The bivariate regression is used to establish a benchmark before any control variables are entered. This enables analysis on whether the inclusion of new variables lead to a change in the point estimate. OLS regressions with robust standard errors are run under each of these models to assess if breastfed individuals are more likely to have higher cognitive ability outcomes than non-breastfed individuals.
𝑌𝑖 = 𝛼1+ 𝛽1𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑𝑖 + 𝜀𝑖 (1)
Equation 1 illustrates this regression mathematically. The dependant variable “Yi” is a binary
variable used to represent whether child “i”, had higher cognitive ability in math or reading. The “α1”coefficient represents the constant term in the model. The “Breastfed”coefficient is
used as a dummy variable to signify whether child “i” had been breastfed and takes on a value of 1 if breastfed, 0 otherwise. The 𝜀𝑖𝑡 variable is the error term and includes any omitted variable bias and some random error.
𝑌𝑖 = 𝛼2+ 𝛽2𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑𝑖 + 𝛿2𝑋𝑖 + 𝜀𝑖 (2)
Model 2 introduces birth characteristic of the child into the OLS regression, including variables for gender of the child, weight at birth, native background, gestation period, family size, and Neonatal Intensive Care Unit (NICU) admittance. All of these variables are determined from wave one of the GUI data set, i.e. at 9 months old. Mathematically, this is estimated similarly to equation 1, apart from the inclusion of “Xi” which is a vector of control variables of birth
characteristics demonstrated in equation 2. Here gender, native background, family size, birth weight, gestation, and NICU admittance are all dummy variables. These values take on a value of 1 if: the child is male, comes from a native background, is not first born, has a birth weight over 2,500grams, or admitted to the NICU at birth (0 otherwise). The gestation period variable is separated into three responses. These responses are: “gestation low” for babies who were
Model 1: Bivariate regression
Model 2: Model 1 + Birth characteristics (e.g. Birth Weight, Gender etc.) Model 3: Model 2 + HH income and PCG Education
15 | P a g e
born under 37 weeks of gestation, “gestation healthy” for any babies born between 38 and 40 weeks of gestation, and “gestation high” for any babies born after 40 weeks of gestation. The inclusion of these variables into the regression limits any potential bias that may influence the real effect of the breastfeeding coefficient. If the breastfeeding coefficient is significantly higher or lower than estimated in equation 1 then we can conclude that there was a bias in the first estimation, which is now accounted for.
𝑌𝑖 = 𝛼3+ 𝛽3𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑𝑖+ 𝛿3𝑋𝑖+ 𝛾3𝐻𝐻𝐼𝑛𝑐𝑜𝑚𝑒𝑖+ 𝜔3𝑃𝐶𝐺𝐸𝑑𝑢𝑐𝑖 + 𝜀𝑖 (3) Model 3 contains, what is perceived in the literature, as the most important confounders (Der, Batty, & Deary, 2006 ). It is estimated similarly to model 2, except with the inclusion of two new vectors of “HHIncome” and “PCGEduc”. This is shown in equation 3. I use Mother’s education as a proxy for mother’s IQ since this information was not available in the GUI data set.
The vector “HHincome” includes the respective dummy variables for the income quintile of each household, with “Quintile 1” being the lowest and “Quintile 5” being the highest. Similarly, “PCGEduc” is used as a vector to include the primary care giver education level. This variable is used as a proxy for maternal education level and is separated into three distinct levels. “PCGEducHigh” represents at least a third-level qualification or higher, “PCGEducMid” represents at least a secondary-level qualification, and “PCGEducLow” represents anything lower than a secondary level qualification. All education levels are reported as dummy variables. Information on “PCGEduc” and “HHIncome” are obtained from wave one of the GUI data set. The inclusion of household income and maternal education allow us to test - by analyzing the change in the point estimate of “𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑𝑖” - if these confounders are important determinants of higher cognitive ability.
After initial estimates are run for the traditional control group, the variable of “length” is included into the regression so as to establish an intensity margin for breastfeeding. In other words, if the duration of breastfeeding leads to a higher probability of a child being above average in math or reading. The GUI data set is restricted in providing the exact day parents stopped breastfeeding their child. Instead, the data gives the day parents stopped breastfeeding rounded to the nearest 10 days. This data is taken from wave one of the GUI data set when the child is 9 months old. The variable of length is continuous up to “190 days” or more and does not provide information after this day.
16 | P a g e
𝑌𝑖 = 𝛼4+ 𝛽4𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑𝑖+ 𝜑4𝐿𝑒𝑛𝑔𝑡ℎ𝑖 + 𝛿4𝑋𝑖+ 𝛾4𝐻𝐻𝐼𝑛𝑐𝑜𝑚𝑒𝑖+ 𝜔4𝑃𝐶𝐺𝐸𝑑𝑢𝑐𝑖
+ 𝜀𝑖 (4) Empirically this does not change the equations in models 1 to 3 by much, as only one new variable is added in each equation. Model 1 changes from a bivariate regression to a multivariate regression as we are now concerned with 2 variables of interest in “length” and “breastfed”. Equation 4 shows the change to model 3 once “lengthit” is added to the regression.
Table 5 and Figure 6 in section 4.4 shows a breakdown of the frequency in days of feeding.
4.2 Alternative Approach
As referenced earlier, this study is broken up into three parts. After examining the traditional approach in the first part, the second part involves narrowing down the comparison group who did not breastfeed and re-running the same models as were done previously for the traditional control group. As participants self-sampled themselves into “breastfed” and “non-breastfed” groups, we use information provided in wave one of the GUI data set on why participants did not initiate breastfeeding. Table A1 accompanied in the appendix of this paper lists all responses to the above question. The selection criteria for the “alternative” control group were reasons where the choice to breastfeed was out of the mother’s control – and potentially exogenous. For example if the baby was ill, this was counted as a potential exogenous reason and participants were assigned to the “alternative” control group. Conversely, if mother decided to not breastfeed their child as they “wanted to drink alcohol” or due to “Embarrassment/Social Stigma” then individuals were not assigned to the alternative control group. Table 2 summarizes the reasons for assignment to the alternative control group.
The mechanism behind each of these reasons is as follows:
Reason 1 “Not enough milk”: If mothers were not able to produce enough milk for a baby to begin breastfeeding, then this is not a choice but a biological barrier to breastfeeding.
Reason 2 “Mother’s illness”: If the mothers are physically unwell or at risk of passing an illness to their child, then they cannot choose to breastfeed as it would put the child at risk
Reason 1 Not enough Milk N= 171
Reason 2 Mother's illness N= 182
Reason 3 Baby's illness N= 31
Table 2 Reasons for Alternative Approach
respondent included baby's illness and milk. Leaving 380 individual responses *Note of the 384 responses, 3 responses chose Mother's illness & milk and 1
17 | P a g e
Reason 3 “Baby’s illness”: If the baby is unwell enough, or has a biological condition that they cannot breastfeed, such as lactose intolerance, then mothers cannot choose to breastfeed. A sub-sample of the “alternative control group” was also created using the “Not enough milk” reason. This was created as lactation failure was seen as the most exogenous out of the three reasons, whereby mothers biologically could not choose to breastfeed their child. If the mother could not produce enough prolactin from the pituitary gland due to endocrine problems, insufficient glandular tissue, breast hypoplasia, or breast cancer, these were perceived as random reasons where the mother could physically not produce milk. As the scope of illness is much wider with “mother’s illness” and “baby’s illness”, I exclude these reasons from the “milk control group” since “not enough milk” has less variation in diagnosable problems and has a higher potential for randomness.
Unfortunately the data does not go into the exact detail of why mothers could not produce enough milk, and therefore endogenous reasons still exist. Reasons such as substance abuse of certain drugs (amphetamines and bromocriptine), obesity, breast surgery, and smoking are judged to be endogenous as pre-existing selection factors may be associated with these reasons (Ferry, McLean, & Nikitovitch-Winer, 1974) (Bennett & Jensen, 1996). The other sickness reasons suffer from the same disentanglement problems, where the data does not provide enough detail on the severity of illness to rule out possible endogeneity introduced by the mother’s choice. For example if one infant is diagnosed with galactosemia, and the other a fever, one illness completely rules out the possibility of breastfeeding beyond the mother’s control, while the other is still feasible.
This criteria reduced the traditional control group from 3,323 to 380 participants for the alternative control group, and from 380 to 170 for the milk control group. Once participants were assigned to the alternative control groups, OLS estimates with robust standard errors are run identical to part one of this study. This allows for comparisons in the point estimates in breastfeeding and length to be made between the “traditional” group, and the “alternative” and “milk” control groups.
4.3 Propensity Score Matching
Propensity Score Matching is used as the last empirical technique in this thesis to create the final comparison group. Matching is a popular technique to estimate causal treatment effects and is applicable to all situations where one has a treatment (breastfeeding), a set of treated individuals (breastfed) and untreated individuals (non-breastfed).
18 | P a g e
When trying to estimate the treatment effect of breastfeeding for a particular individual, there exist an inherent problem where we are unable to observe the effect of being treated and untreated for the same individual on a given set of outcomes, i.e. one cannot be both breastfed and not breastfed, as well as simultaneously being below average and above average in math ability.
This is shown mathematically in equation 5, illustrating the Average Treatment Effect on the Treated (ATET) for outcome Yi (above average in math) for all breastfed individuals, where:
𝐴𝑇𝐸𝑇 = 𝐸(𝑌1 | 𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑 = 1) − 𝐸(𝑌0 | 𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑 = 1) (5)
Here we allow the second term “𝐸(𝑌0 | 𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑 = 1)”, to be unobserved for all breastfed
students. For individuals who were not breastfed, i.e. the untreated, we allow 𝐸(𝑌0 | 𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑 = 0) to be observed. Thus we are able to estimate the Average Treatment Effect (ATE) as follows:
𝐴𝑇𝐸 = 𝐸(𝑌1 | 𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑 = 1) − 𝐸(𝑌0 | 𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑 = 0) = 𝐴𝑇𝐸𝑇 + 𝑆𝐵𝑖𝑎𝑠 (6) The “SBias” term in equation 6 stands for selection bias. In order to successfully estimate the average treatment effect, the selection bias must equal to zero. As discussed previously, the selection bias occurs in breastfeeding with important confounders - such as household income and maternal education - which may make a child more likely to be breastfed. Propensity score matching attempts to bypass this problem by finding a comparable group of non-treated individuals who share similar pre-treatment characteristics to the treatment group, thus removing the selection bias (Rubin & Rosenbaum, 1983).
However before implementing propensity score matching, a number of conditions must be satisfied and two important choices must be made about model choice and the variables to be included within the model. The model choice (between probit and logit) is less important when dealing with the binary treatment case9 compared to the multiple treatment case (Caliendo & Kopeining, 2008). As our case is binary, a probit model is chosen and is estimated by:
𝑃(𝑌𝑖 = 1 | 𝐵𝑟𝑒𝑎𝑠𝑡𝑓𝑒𝑑𝑖) = 𝜎(𝑍𝑖′𝛽) (7)
19 | P a g e
The left hand side of equation is the probability of outcome Yi (above average math ability)
conditional on the breastfeeding treatment variable. 𝜎 is used as the cumulative distribution function of the standard normal distribution, and Zi is a vector of all possible regressors.
With the model choice chosen, the next important step is to consider which variables for inclusion. This is reliant on the Conditional Independence Assumption (CIA) which states “that given a set of covariates X which are not affected by treatment, potential outcomes are independent of treatment assignment” (Caliendo & Kopeining, 2008). This implies that selection is solely based on observable characteristics and that any variables that influence treatment assignment and potential outcomes simultaneously are observed by the agent. This assumption is essential for propensity score matching to take place. The variables used in propensity score matching will be the same as used for model 3 described in section 4.1. Models 1 and 2 will not be estimated as the conditional independence assumption fails if propensity scores are not conditioned on all possible variables10.
After model and variable choice has taken place, the choice of matching algorithm is the next important step. The most straightforward matching estimator is nearest neighbour matching. This is where individuals from a comparison group is chosen as a matching partner for a treated individual that is closest in terms of propensity score. Nearest neighbour matching of (2) is chosen in this particular case based on the ratio of treated to untreated individuals11. While the ratio for both math and reading is just under [2:1], oversampling is used in matching. This form of matching subsequently involves a trade-off between variance and bias. It trades reduced variance, giving more information to construct a counterfactual for each participant with increased bias, resulting from on average, poorer quality matches (Abadie, Drukker, Herr, & Imbens, 2004).
Finally the common support assumption must hold which states that, “any person with the same values of X12 have positive probability of being both participants and non-participants” (Heckman, LaLonde, & Smith, 1999) . In other words there needs to be enough overlap, or “common support” in characteristics, between both the treated and untreated individuals. This assumption can be inspected graphically as argued by (Lechner, 2000b). Figure 2 shows the propensity score distribution among breastfed and non-breastfed students for math ability, and
10 Length will also be omitted as it is not possible to match to the treatment group since length is = 0 for
everyone in the control group.
11 [4,847:3,050] for Math & [4,298:2,785] for reading – treated to untreated.
20 | P a g e
figure 3 illustrates this information in a histogram. The same is shown for reading ability in figures 4 and 5 respectively. As shown in the figures below there is enough overlap in the two groups for the common support assumption to be satisfied, and for suitable matches to be made. (Math)
Figure 2 & 3: Distribution of propensity score among breastfed and non-breastfed students for math ability
(Reading)
Figure 4 & 5: Distribution of propensity score among breastfed and non-breastfed students for math ability
4.4 Descriptive Statistics
Table 3 illustrates descriptive statistics between individuals in the treatment group (breastfed) as well as the three control groups (“traditional”, “alternative”, and “milk”). When comparing
21 | P a g e
the means between approaches, the control groups created using the alternative approach show to be more comparable to the treatment group. The only variables where the traditional control group is shown as a better comparison is gestation and NICU stay. For the alternative and milk control group, family size, gender, birth weight, background, household income, and primary caregiver education are shown to be more alike to the treatment. This distinction is important as in theory we are reducing the sample-selection bias - present in the traditional control group - by making individuals more comparable with either the alternative or milk control group. When examining the means for household income and primary caregiver education, there is a noticeable difference between the treatment and control groups. Over 50% of the treatment group hold at least a third-level qualification. In all control groups, the majority of individuals hold a second-level qualification with third-level much rarer. Similarly for household income, the table shows a much higher percentage of individuals present in the top income quintile for the treatment group compared to all other control groups. The number of individuals present in each quintile for the treatment group decrease as quintile levels fall. Conversely, in the traditional control group, this number increases for each drop in income quintile. The distribution among the “alternative” and “milk” control group is much smoother.
22 | P a g e
Table 4 reports the test statistics on the differences in means between the treatment and control groups. All p-values here under the 5% are used to represent groups where the means are statistically different from each other. When comparing the traditional control group to the treatment, the groups are shown to be statistically different from each other with the exception of NICU stay and the third income quintile. Comparing the alternative group to the treatment group, only gender and two of the income quintiles are shown to be statistically similar. While the alternative control group is shown to be more comparable to the treatment than the traditional control group, it does not show as considerably different.
The group that is seen as the most statistically similar is the treatment and milk control group. A total of 8 out of the 14 variables are seen as statistically similar to the treatment group with p-values over the 5% level. In these results, the significance is that the milk control group is only statistically different to the treatment group mainly in family size and the maternal education composition. Unfortunately no control group shows a mean statistically similar to the treatment for maternal education. However, the “milk” comparison group does show the most similarity to the treatment for the other important confounder introduced in model 3, household income.
23 | P a g e
4.4.1 Descriptive Statistics – Duration
Table 5 and Figure 6 illustrates the frequency of distribution in the days of breastfeeding. The control group is shown in the “0 days” frequency, which changes based on whether the traditional, alternative, or milk control group is used for analysis. Of the initial group who breastfed (5,244), only 4,321 individuals filled out information on the duration of days spent breastfeeding. Thus, for analysis when the “length” variable was included, these missing values are dropped. The table also shows hints of rounding errors to “month” values as the distribution is much more concentrated around intervals of 30 days. After 60 days of breastfeeding, the rounding effect becomes a lot clearer as shown in the intervals between “120 days to 150 days”, and “150 days to 180 days”. This may indicate a measurement error in the length variable as people were recalling months of breastfeeding instead of days.
24 | P a g e
Figure 6: Number of children breastfed in wave one and wave three of the GUI data set.13
V Empirical Results
5.1 Traditional Approach
Table 6 examines the initial results using the traditional approach. The bivariate regression in model 1 shows the breastfeeding coefficient as large and statistically significant at the 1% level for both math and reading. As the covariates gender, birthweight, NICU stay, family size, gestation period, and background are added in model 2, the breastfeeding coefficient increase in size, significant at the 1% level. The inclusion of the covariates in model 2, have corrected for the downward bias present in the “breastfeeding” coefficient as it gets closer to the true value.The bias appears to be larger for reading than for math ability as the coefficient increased by 0.0127 for reading compared to 0.0060 for math.
13 Figure 2 shows the frequency distribution for the “Alternative” control group. The distribution is the same for
25 | P a g e
Under the third model, correcting for both household income and primary caregiver education, the breastfeeding coefficient drops by more than 50% for both math and reading ability. This confirms that estimates in model 2 were upward biased. The breastfeeding coefficient on above average math ability decreases substantially under model 3 and is significant at the 5% level. Therefore at the 95% confidence level, breastfeeding your child is shown to have a significant 2.98 percentage point increase in the probability of the child being above average in math. Similarly for reading, at the 99% confidence level, breastfeeding your child is shown to have a significant 4.61 percentage point increase in the probability of the child being above average in reading. Hence, when we have not restricted any sample populations, breastfeeding is shown to have a positive effect on math and reading ability. These coefficients are used as the main comparison to the alternative and milk control group estimates in the next section as they control for the most variables and still show as large and significant.
Table 7 shows the coefficient of breastfeeding once the intensity – or duration – of breastfeeding is included into the regression. With length included, the coefficient for breastfeeding decreases for both math and reading ability under all models. Therefore, it appears that by adding length of breastfeeding to the model, it reduces the explanatory power of the breastfeeding coefficient, i.e. the length of feeding appears to reduce the importance derived from breastfeeding on a yes/no margin. For math ability, the breastfeeding coefficient is significant at the 1% level for both model 1 and model 2, and not statistically significant once household income and primary caregiver education is controlled for in model 3. This further suggest the idea that with the inclusion of length, household income and maternal
26 | P a g e
education are stronger determinants of ability than breastfeeding, as all explanatory power in the breastfeeding and length coefficient is lost for model 3.
The reading estimate for breastfeeding is shown to be similar to math, except for the third model, where the coefficient is statistically significant at the 10% level. Similarly, when breastfeeding is examined separately to length, model 3 shows the most change in the breastfeeding coefficient with estimates dropping by over half.
When examining the length variable, we see that it is significant for both math and reading estimates at the 1% level for model 2. Model 3 is not shown to be statistically significant for the length variable for both math and reading. The coefficient for length moves in a similar direction to breastfeeding between models, whereby we see an initial increase between model 1 and model 2, and a substantial decrease between model 2 and model 3. Model 3 shows to correct for upward bias as both duration and the breastfeeding coefficient decrease by over half the value in model 2 and are no longer statistically significant. Thus, we can conclude that the length of breastfeeding has no statistically significant effect on cognitive ability once household income and primary care giver education are controlled for.
27 | P a g e
5.2. Alternative margin results
Table 8 compares the results of breastfeeding from the traditional approach against the results of the alternative approach (i.e. once control groups have been restricted). For math ability, comparing the traditional control group and the alternative control group, estimates are lower across all models and are not statistically significant with the exception of model 1. Similarly for the milk group, we see no coefficient being statistically significant across all models with the point estimate being close to zero (0.0005) for model 3. Based on these results we can conclude that once we restrict the sample size of the control group to the aforementioned reasons in the alternative approach, breastfeeding does not appear to have a statistically significant effect on higher cognitive math ability.
For reading estimates, we find a significant breastfeeding effect across most models when looking at the alternative and milk control groups. In model 1 (without any controls) initial estimates of breastfeeding appear to be higher in the alternative and milk control group compared to the traditional control group. However, after covariates are added into model 2, estimates appear as lower for breastfeeding in both the alternative and milk control groups. The last model shows the breastfeeding coefficient as statistically significant for the alternative control group at the 10% level. However, this is due to the large standard errors present. When the sample gets restricted even further to the milk control group, there appears to be no statistically significant effect of breastfeeding on reading ability once primary care giver education and household income are accounted for.
28 | P a g e
5.2.1 Duration comparison
Table 9 compares the breastfeeding and length coefficients for all control groups using the traditional and alternative approach. The length coefficient remains roughly similar for both math and reading ability across all approaches. In both the alternative and milk control group it is shown to reduce the explanatory power of the breastfeeding coefficient and can be interpreted similar to the traditional approach. The explanation for the lack of variation in length estimates across control groups is that the only frequency changing in the length variable is the no group, i.e. the “0 days”. All other frequencies (30, 40, 50 days etc.) remain constant as they are based on the treatment group.
Analyzing the breastfeeding coefficient for math ability, neither the alternative nor the milk control group show any statistically significant effects for breastfeeding. Similar to when breastfeeding is examined excluding length, the results appear to show that when we restrict the sample of the control group, the breastfeeding coefficient does not produce a statistically significant effect on higher cognitive math ability.
In terms of reading ability, the coefficient for breastfeeding only appears statistically significant at the 5% level for the alternative control group in model 1 and model 2. The estimates for model 2 in the alternative control group show as 1.02 percentage points lower than the
29 | P a g e
traditional control group. When household income and primary caregiver education are controlled for in model 3, the results are no longer statistically significant, yet this may be attributed to large standard errors. This again suggests that these confounders are driving higher cognitive ability and not the act of breastfeeding. The milk control group is only significant in model 1 and decreases steadily with the addition of covariates in models 2 and 3. The estimates also highlight a lack of variation in the magnitude of the breastfeeding coefficient across approaches in reading ability, as for model 3, the breastfeeding coefficient is shown at 0.0304, 0.0337, 0.0335 for the traditional, alternative, and milk control groups respectively.
5.3 – Intensive and Extensive Margins
Figure 7 *Author’s calculation based on mean probability of being above average of math from Wave 3 of the GUI data set.
Figures 7 and 8 show the intensive and extensive margins for breastfeeding in math and reading ability respectively. The extensive margin is used to show the probability of the child being above average in cognitive ability. This is expressed as a flat line since there is no variation with the length of feeding. The intensive margin shows the probability of the child being above average in cognitive ability the longer the duration of breastfeeding. Results of the intensive margin is fitted with a linear trend-line. The aim is to graphically compare the extensive margins of each control group, to the intensive margins to see if the gap in estimates have reduced once the sample has been restricted to limit endogeneity concerns. This approach is similar to that of Caetano (2005) examining the effects of number of cigarettes smoked per day (intensive margin) on birth weights.
30 | P a g e
The above hypothesis shows to be true in figure 7 as the gap between the treatment and the traditional group is the largest among the control groups. The 0.06 gap in the groups is reduced on the y-axis once we restrict our sample to those who did not self-select in the “alternative” or “milk” control group, as the difference between these and the treatment is only 0.03. The milk control group shows a similar gap to the alternative control group between margins as they approach closer to the intensive margin14. The vertical comparison between approaches would suggest that when the sample is restricted using the alternative approach, the difference in the effects of breastfeeding is reduced.
Figure 8 – *Author’s calculation based on mean probability of being above average of reading from Wave 3 of the GUI data set.
The results for reading ability do not support the same theory as math ability. The extensive margins for the traditional, alternative, and milk control group all appear as roughly similar with means at .3199, .3133, and 0.3055 respectively. Analyzing the vertical gap between the treatment group and the control groups, the traditional group is still seen as the closest comparison. Therefore, it appears that even when the sample size is restricted, there still exists a significant gap in the probability of being above average in math between those who breastfed and those who did not breastfeed. This is further supported by estimates in the previous section, as once samples were restricted to either the alternative or milk control groups, the point
31 | P a g e
estimates showed to be similar to the traditional group and demonstrated less of a variation than the math variable.
5.4. Matching Results
Table 10 compares the OLS and Probit results for the traditional control group for both math and reading. As stated earlier, the common independence assumption needed for propensity score matching is violated if propensity scores are not conditioned on all possible values. Since the probit model is only needed for propensity score matching, singular results are reported for model 315.The probit model is estimated via marginal effects to act as a comparison to the OLS results.
The estimates between the OLS model and the probit model show as extremely similar which is to be expected16. The probit model predicts that, holding all else constant, breastfeeding increases the probability of above average math ability by 3.01 percentage points. Similarly, the probit model also predicts that breastfeeding increases the probability of above average reading ability by 4.73 percentage points, holding all else constant. Both math and reading estimates show as slightly higher in magnitude for the probit model, compared to the OLS (or Linear Probability Model), with math and reading estimates showing the same level of significance between estimators. While the results for probit are not as important as the models use in propensity score matching, the difference in results may be attributed to the underlying assumptions of each model17.
15 Model 3 includes all variables for model 2, as well as Household Income and PCG Education.
16 As there is enough subjects in both the breastfed and non-breastfed groups, estimates are expected to be
similar for both the OLS/Linear Probability Model and Probit Model.
17 Since all variables on the RHS of equation (3) are dummy variables, OLS estimates will not suffer from the
32 | P a g e
Table 11 PSM Balancing Table
MATH READING Unmatched Mean P-Value Mean P-Value Variable Matched Treatment Control Treatment Control
Male U 0.495 0.521 0.023 0.493 0.518 0.037 M 0.495 0.512 0.094 0.493 0.507 0.185 BirthWeight U 0.951 0.942 0.072 0.951 0.942 0.090 M 0.951 0.964 0.002 0.951 0.964 0.003 Background U 0.825 0.964 0.000 0.817 0.964 0.000 M 0.825 0.834 0.213 0.817 0.823 0.415 GestHigh U 0.314 0.293 0.042 0.315 0.297 0.109 M 0.314 0.329 0.142 0.315 0.330 0.117 Gest Low U 0.101 0.121 0.004 0.103 0.124 0.005 M 0.101 0.095 0.357 0.103 0.097 0.341 FamilySize U 0.562 0.677 0.000 0.560 0.677 0.000 M 0.562 0.566 0.705 0.560 0.566 0.587 NICU Stay U 0.136 0.146 0.191 0.136 0.148 0.150 M 0.136 0.131 0.571 0.136 0.122 0.060 PeducHigh U 0.503 0.219 0.000 0.495 0.212 0.000 M 0.503 0.511 0.428 0.495 0.499 0.666 Peduc Mid U 0.469 0.649 0.000 0.475 0.651 0.000 M 0.469 0.464 0.691 0.475 0.474 0.948 Quin1 U 0.145 0.237 0.000 0.148 0.244 0.000 M 0.145 0.136 0.237 0.148 0.140 0.304 Quin2 U 0.168 0.221 0.000 0.170 0.225 0.000 M 0.168 0.161 0.381 0.170 0.163 0.394 Quin3 U 0.194 0.205 0.256 0.198 0.203 0.574 M 0.194 0.195 0.969 0.198 0.198 0.999 Quin4 U 0.233 0.183 0.000 0.226 0.178 0.000 M 0.233 0.228 0.552 0.226 0.24 0.120
*Note - Figures in bold show where variables have been successfully matched
Table 11 presents the PSM balancing table for both math and reading variables18 to see if the nearest neighbour matching algorithm has taken place successfully. Here the means are reported before and after matching for the control and treatment groups, as well as p-values of a t–test between the means of both groups. The table would suggest matching has indeed taken place successfully for most variables in both math and reading. The P-values in the third and sixth column show if the means are statistically different from each other after matching. If the p-value is over the 10% level, then we can safely reject the null hypothesis that the means are
18 Variables: Gender, Birthweight, Background, Gestation Period, Family Size, NICU stay, PCG education and
33 | P a g e
statistically different from each other. As the table shows, “gender” and “birth weight” variables were not able to be matched successfully for math ability. For reading ability only the “birth weight” and “NICU stay” variables still show as statistically different. What is interesting to note - in comparison to the traditional and alternative method used earlier in this thesis - is that the education variables now do not show as statistically different, which was not the case in Table 4 for any of the control groups. The means for math ability for high primary care giver education levels showed as 50.3% for the treatment group and 21.9% for the traditional control group in the original sample. In the matched sample, this mean has changed for the traditional control group to 51.1%, which is much closer to the original treatment mean of 50.3%.
As not all variables could be matched, the matched sample cannot be considered as a perfect comparison group. However since most of the variables are matched it can be considered a more suitable comparison group than the initial traditional control group. This statement is further supported by examining the PSM kernel density graphs provided in figures A2 and A3 of the appendix.
Table 12 shows both the Average Treatment Effect (ATE) and Average Treatment Effect on Treated (ATET) for both math and reading ability. These results include all variables specified in the PSM balancing table. Focusing on the ATET first, the table shows a statistically significant, and positive impact of breastfeeding on higher math and reading ability at the 5% and 1% level respectively. The results can be interpreted as “the average breastfed student is 3.32% (4.98%) more likely to be above average in math (reading) ability had they never been breastfed”.
34 | P a g e
While these results are significant, the ATE results are of greater interest as they display the effects of breastfeeding across the entire sample. The ATE shows as positive, and has a significant impact for both math and reading ability at the 5% and 1% significance level. These results can be interpreted as “the average student in the sample would be 2.85% (4.78%) more likely to be above average in math (reading) ability had they been breastfed”. These matching results are contradictory to the results found in the alternative and milk control groups, and more in line with the results of the traditional control group found in table 819. The results also
show much higher in magnitude and significance for reading ability over math ability. Based on this method, matching of propensity scores does seem to show a significant difference in cognitive outcomes between being breastfed and not breastfed.
VI Discussion
The results in this thesis highlight the complicated issue and often conflicting results, in trying to establish the causal effect of breastfeeding. When comparing results between individuals who did and did not breastfeed, it is easy to see that there is a positive and significant effect of breastfeeding on cognitive ability, either through the traditional method or through PSM. However, if we look at an alternative methodology that restricts the comparison group based on mother’s reasons for not breastfeeding, this effect disappears and suggests that breastfeeding may not be the perfect guarantee for better cognitive outcomes that physicians and policy maker’s try to assert.
The results in this thesis stresses the importance of controlling for the confounders of maternal education and household income when estimating breastfeeding effects. Using the traditional and alternative approaches, we see a drop in the breastfeeding coefficient by over 50% when changing between model 2 and model 3. If we restrict analysis to the results of the alternative method, it appears that the cognitive gains are coming from these channels instead of the nutritional value of breastmilk augmenting neurological development. This is in line with recent research that control for such confounders, as while breastfeeding does present many advantages for the mother and child, enhancement of cognitive ability in either math or reading is unlikely to be one of them (Der, Batty, & Deary, 2006 ).
These conclusions are based, however, on an alternative approach which is not completely free from limitations. The theory behind this method is to present a more realistic way of estimating
19 Table 8 is used over Table 9 as it is impossible to match the length of breastfeeding variable for the control
35 | P a g e
the cognitive returns to breastfeeding, by attempting to balance the characteristics of both the initial treatment and control groups. Indeed this is shown by the similarity of the characteristics in the treatment group to the milk control group when testing for statistical difference. However, it does not completely eliminate endogeneity issues associated with observational studies and self-selection in breastfeeding. Information on the exact nature of lactation failure or illness could negate the sample section bias if disclosed in a more detailed data set. In theory, this should lead to a more refined estimate as endogeneity can truly be ruled out if the nature of lactation failure is indeed random.
If the focus is shifted to PSM, matching methods give a significant ATE of breastfeeding for both math and reading cognitive ability. While a positive ATE is shown, it is worth noting that while PSM balances observe baseline characteristics, it still does nothing to balance unmeasured characteristics and confounders. Therefore as with all observational studies, and unlike randomized control trials, PSM analyses have limitations in that unmeasured confounding may still be present (Wolfgang & Kurth, 2004). PSM also showed to increase imbalances between groups relative to the original data, for variables such as birth weight and NICU stay for reading ability20. As such these variables were not able to be balanced21. One advantage of PSM over the alternative method proposed in this thesis is that large numbers of observations are not lost in analysis, as is the case when restricting control groups. As both methods have pros and cons, it is difficult to say which method is better at reducing sample selection bias in breastfeeding.
Both the alternative and matching approaches did however exhibit a difference in outcomes between the cognitive abilities of math and reading. Breastfeeding estimates for reading - within models - did not show much variation in the approach used and extensive margins showed as similar regardless of sample size. This is in contrast to math ability, where the alternative approach dropped the coefficient by at least two percentage points in most models, and the narrowing of sample size clearly reduced the gap between extensive and intensive margins. One potential reason for this difference may be the measurement of the ability itself as they are based on teacher evaluation. Reading is a harder ability to evaluate and variation on which factors to measure, such as comprehension, speed, as well as reading out-loud or silently, are debated amongst academics (Rubin G. , 2013). Scores on standardized tests for math and
20 Changing the matching algorithm to “nearest neighbour 1” still leaves variables unmatched for both math and
reading ability.
21 This can be shown in figure either the PSM Balancing table or the kernel density graphs before and after
