Child stunting is associated with weaker human capital among native Amazonians

1 INTRODUCTION

Child stunting, the chronic restriction of a child's potential growth, is widespread in low-income nations owing to maternal and child malnutrition (Black et al., 2017, 2008; de Onis, Blössner, & Borghi, 2011; Lu, Black, & Richter, 2016; Walker, Chang, Wright, Osmond, & Grantham-McGregor, 2015). Several factors contribute to malnutrition, including deprivations from poverty, inadequate nutrients and protein, poor water and sanitation, and high disease burden, resulting in short stature or slow growth in stature (Black et al., 2008, 2013, 2017; Lu et al., 2016; Victora et al., 2008b). Recent estimates suggest that globally as many as 171 million children (∼25%) < 5 years of age suffer from stunting (de Onis et al., 2011; United Nations, 2016). While stature is partly determined by genetic potential, nutrition and disease loads during childhood play major roles in shaping world-wide variation in human growth (Martorell & Habicht, 1986). Thus, stunting is amenable to public health interventions (Bhutta et al., 2013; Engle et al., 2007; Hoddinott, Alderman, Behrman, Haddad, & Horton, 2013a).

Predictors of persistent stunting affect brain development and function, risk of illness, energy levels, motor development, and exploratory behaviors and—through some of these paths—erode cognitive and academic skills and educational attainment during early life (Berkman, Lescano, Gilman, Lopez, & Black, 2002; Casale & Desmond, 2016; Crookston et al., 2013; Fink & Rockers, 2014; Georgiadis et al., 2016; Ghosh, Chowdhury, Chandra, & Ghosh, 2015; Sudfeld et al., 2015). These effects may last into adulthood (Alderman, Hoddinott, & Kinsey, 2006; Hoddinott et al., 2013b; Hoddinott, Maluccio, Behrman, Flores, & Martorell, 2008; Prendergast & Humphrey, 2014; Victora et al., 2008b) and even across generations (Behrman, Calderón, Preston, Hoddinott, & Martorell, 2009; Victora et al., 2008a; Walker et al., 2015).

Some researchers have defined a window for interventions among children aged <2 years to redress stunting, suggesting that the factors that stunted growth might have irreversible effects after this window (Victora et al., 2008a; Victora, de Onis, Hallal, Blössner, & Shrimpton, 2010). However, recent longitudinal evidence suggests that stunting is reversible after the age of 2 years (Georgiadis et al., 2016; Lundeen et al., 2014; Mani, 2012; Prentice et al., 2013; Schott, Crookston, Lundeen, Stein, & Behrman, 2013). In some instances this reversal has been observed even without public health interventions, probably from general improvements in socio-economic conditions (Crookston, Forste, McClellan, Georgiadis, & Heaton, 2014). Some recent evidence also suggests that children who recover from stunting have better cognitive skills and complete more schooling grades than children who do not recover (Crookston, et al., 2013; Fink & Rockers, 2014), but supporting evidence was not found in at least one other context (Casale & Desmond, 2016).

We contribute to research on child stunting and child human capital by estimating these associations in a unique setting: a remote, economically self-sufficient rural society of native Amazonian horticulturalists-foragers in Bolivia, the Tsimane' (Figure 1). Such a setting adds insights to childhood stunting studies in at least three ways. First, with a few possible exceptions (e.g., Hoddinott et al., 2013b), most estimates of the links between child stunting and child human capital are associative, and may be biased by effects of unobserved omitted socioeconomic confounders such as ethnic and racial heterogeneity, access to health care, residential segregation, or differential exposure to pollution (Deaton & Lubotsky, 2003; Prendergast & Humphrey, 2014; Shaw et al., 1999). While Tsimane' households vary in their access to resources (Undurraga, Nica, Zhang, Mensah, & Godoy, 2016b), including food and monetary income, many of these other confounding factors vary much less among the Tsimane' than among many other populations in which associations between stunting and human capital have been studied. Tsimane' villages are racially and ethnically homogeneous, with non-Tsimane' accounting for ∼5% of the population in their villages, and most Tsimane' have limited access to Western health care (Byron, 2003; Gurven, Kaplan, & Supa, 2007). Second, the setting allowed us to extend the work linking childhood stunting with human capital by examining the association between child stunting and child knowledge of local plants, a culturally-appropriate measure of human capital (Reyes-García et al., 2008). Last, we used longitudinal data to examine whether growth faltering or recovery were associated with changes in human capital over time in a traditional rural society that is increasingly integrated into the market economy.

The main form of human capital through history has not been schooling or academic skills, but traditional ecological knowledge, the cumulative body of knowledge of living things, natural resources, and ecosystems that is socially transmitted between generations (Berkes, Colding, & Folke, 2000). This type of human capital can be critical to the subsistence, health, and nutritional status of people in small-scale rural societies (Johns, 1996; McDade et al., 2007; Reyes-García et al., 2008; Thomas, Vandebroek, Sanca, & Van Damme, 2009). Traditional ecological knowledge may improve the use of local natural resources, thus generating benefits such as securing diets of better quality and adequate quantity, using pharmacological properties of plants, or improving farm productivity (Reyes-García, 2015). We used local plant knowledge as a proxy for traditional ecological knowledge. Prior research has found that parental local plant knowledge in rural societies is positively associated with both parental and child health and may overshadow formal educational attainment and academic skills in practical importance (McDade et al., 2006; Quave & Pieroni, 2015; Reyes-García et al, 2016b; Reyes-Garcia et al., 2007).

For the analysis, we used cross-sectional and longitudinal data from the Tsimane', a native Amazonian society in the department of Beni, Bolivia (Supporting Information Figure S1). At the time of research, the Tsimane' had low levels of monetary income, weak links to the market economy, and mostly primary schools that operated sporadically, with Tsimane' as the primary language of instruction (Godoy et al., 2007b; Undurraga et al, 2013). During 2010, the last year of our nine-year annual longitudinal survey (2002–2010), the mean monetary daily earnings per person reached only US $0.90, about half the global poverty line (US $1.90 Purchasing Power Parity) used by the World Bank (2015). In a worldwide comparative study of 15 small-scale rural societies, the Tsimane' ranked next-to-last in market interactions, with only 7% of total household calories bought in the market (Henrich, Ensminger, McElreath, & Barrat, 2010). Nevertheless, Tsimane' villages vary in the depth and frequency of contact with outsiders. We control for some of these effects in the analysis by using walking distance (in hours) to the nearest town. Last, most Tsimane' self-reported as being monolingual in Tsimane'. For instance, the 2008 survey from the longitudinal study showed that of the 396 children and adults for whom we had data on language skills, 57% self-reported being monolingual in Tsimane', 35% reported having some fluency in Spanish, and 8% reported being fluent in Spanish.

Previous work among the Tsimane' suggests that the share of stunted children is high but declining (Godoy et al., 2010; Zhang et al., 2016). In 2000, 52% of all boys and 43% of all girls aged < 9 years were stunted (Foster et al., 2005a); by 2010 stunting rates for children aged < 9 years had fallen to 34% for all boys and to 30% for all girls (Zhang et al., 2016). In 2008, the year we used for much of the analyses, 34% of children aged 6 to 16 years were stunted (height-for-age Z score, HAZ<–2; girls =31%; boys 36%), 41% were moderately stunted (–2 < HAZ<–1; girls = 44%; boys = 38%), and 26% were not stunted (HAZ>–1; girls =25%; boys = 26%). Based on a sample of 409 children aged <9 years in 58 villages, Foster et al. (2005b) examined whether village attributes predicted stunting and found no significant results. McDade et al. (2007) found no significant predictors of stunting among 330 children between 2 and 10 years of age, but the probability of severe stunting (HAZ<–3; ∼12% of the sample) declined with maternal ethnobotanical knowledge, village population size, and father's body mass index. Nyberg et al. (2012) found that acculturation into national society and infection predicted stunting in a sample of 53 Tsimane' children >1.6 years of age, but Tanner (2014), based on a larger sample of children between 2 and 14 years of age, found weak associations between child stunting and helminth infections or market interactions. Using a sample of 238 children aged <2 years, Gurven (2012) found that the probability of being stunted decreased with maternal weight and increased with child birth order. Last, our prior work shows some evidence of catch-up growth (Godoy et al., 2010), but height deficits persist despite modest year-to-year changes between height categories (Zhang et al., 2016). None of the prior work has focused on the links between child stunting and human capital.

We focused on children in primary school (aged 6 to 16 years) to estimate two sets of associations. First, using cross-sectional data from 2008, we estimated the contemporaneous association between (i) children's height categories—stunted (HAZ <–2), moderately stunted (–2≤ HAZ ≤–1), nonstunted (HAZ >–1)—and (ii) human capital, measured as completed grades of schooling, test-based academic skill (math, reading, writing), and local plant knowledge. We increased our sample size by combining the 2008 survey of the longitudinal study with the baseline survey of a randomized controlled trial (RCT). Second, using a nine-year annual (2002–2010) longitudinal dataset from all children in 13 villages along the Maniqui River, we estimated the association between (a) changes in height category (e.g., from stunted to nonstunted, or vice versa) from the first to the last year in which the child was measured and (b) the schooling and academic skills during the last year of measurement. Supporting Information Figure S1 shows visually the overlap between the two datasets, and how each dataset relates to the associations we want to estimate.

2 STUDY PARTICIPANTS AND METHODS

2.4 Analysis

2.4.1 Height category and human capital (2008 cross-sectional data)

We used the following Ordinary-Least Squares (OLS) regression with household fixed effects and robust standard errors adjusted for clustering at the village level to analyze the 2008 cross-sectional data:

()

where the subscripts stand for individual child (i), household (h), and village (v), with (j) indexing the type of human capital (HK) used as an outcome. HK_ihv stands for human capital, and captures one of the following: grades of completed schooling, score in the tests of math, reading, writing, or local plant knowledge. S_ihv and MS_ihv are binary variables for stunted (S) and for moderately stunted (MS), with nonstunted children serving as a reference category.

Control variables (C) included the child's age in 2008 and the child's sex, a binary variable for the study (TAPS longitudinal study [2002–2010] = 1; 2008 cross-sectional sample from RCT =0), and the walking distance (in hours) from the village to the nearest town. We controlled for the study type because repeated exposure by study participants to the same surveyors in the longitudinal study might influence responses (Zwane et al., 2011). Because most households (59%) had only one child who was tested for local plant knowledge, we did not use household fixed effects when using local plant knowledge as an outcome, but we adjusted standard errors for clustering at the village level. For ease of interpretation, we used OLS even for binary outcomes, but later used Tobit, Poisson, and Logistic regressions as robustness checks for our main results. ɛ_ihv is an error term.

2.4.2 Change in height category and human capital (2002–2010 longitudinal data)

To estimate what happens to schooling and academic skills when a stunted child became nonstunted (recovery) or when a nonstunted child became stunted (faltering) we used annual data collected during 2002–2010 from ∼850 children aged 6 to 16 years from the 13 villages of the longitudinal study. Following Crookston et al. (2010), we created a variable that captured changes in height categories (ΔHAZ), defined as HAZ during the last year a child was measured minus HAZ during the first year the child was measured. The child had to be in the range of 6 to 16 years of age for both measurements to be included in the analysis. Thus, the first and the last measure did not necessarily refer to the first (2002) and the last (2010) year of the longitudinal study. We decided to consider the first and the last measure to avoid reducing the sample size. Our definition of change in HAZ has the advantage of capturing any change in height category during the maximum length of time we measured a child who was between 6 and 16 years of age; more transitory changes between height categories were thus swept away in the sample we considered (Cole, 1997; Hermanussen, Grasedyck, Kromeyer-Hauschild, Prokopecs, & Chrzastek-Spruch, 2002). The mean and median duration between the first and the last measure was four years (SD = 2.2; min = 1; max = 8).

Using ΔHAZ, we created three binary variables for the regression analysis: ΔHAZ⁺, ΔHAZ⁻, and ΔHAZ⁰. These variables stand for children whose height improved (recovered) from stunted to nonstunted (ΔHAZ⁺=1 recovered, 0 otherwise), children whose height worsened (faltered) from nonstunted to stunted (ΔHAZ⁻=1 faltered, 0 otherwise), and children who remained stunted in both measurement years (ΔHAZ⁰=1 stunted, 0 otherwise). Children who remained nonstunted in both measurement years served as the reference category in the regressions. We also estimated ΔHAZ between the first and last measurements for all children without considering the stunted and nonstunted distinction, and then estimated the association between the total change in HAZ and human capital outcomes.

Because policy questions have focused on what happens to a child's human capital when a nonstunted child becomes stunted or vice versa, we examined the two-way change from stunted to nonstunted, without considering the category of moderately stunted. The category of moderately stunted was useful in the cross-sectional analysis because it permitted a finer-grained description of how children's human capital varied in relation to height categories. The trade-off was a smaller sample size in each height category, which may have resulted in less precise estimates. Focusing on the two-way change from stunted to nonstunted had the limitation of defining the boundary between stunted and nonstunted by a single number (i.e., HAZ = −2). A child whose height changed from −2.01 HAZ to −1.99 HAZ would be technically regarded as someone who had transitioned from stunted to nonstunted even though the size of the height improvement was trivial. To address this concern, we used the standard definition of stunting, but then considered a child to have changed height category only if the child passed the threshold by an additional 0.2 HAZ units. For example, a child with a HAZ of −2.1 would have to improve to at least −1.9 HAZ to have been considered as having transitioned from stunted to nonstunted. Adding 0.2 HAZ units to the threshold reduced the likelihood of counting as meaningful very small height changes near the boundary.

To estimate the association between changes in height categories and human capital outcomes we used the following OLS regressions, with one observation per child, and with household fixed effects and robust standard errors adjusted for clustering at the village level:

()

In Equation 2, the outcomes remained the same as in Equation 1, but referred to the last year in which the child was measured. Control variables included the child's age and HAZ during the first year of measure, the child's sex, baseline measure of the outcome, walking distance (in hours) from the village to the nearest town, and a variable for the number of years between the first and the last height measurement.

With both the cross-sectional and longitudinal data we used household fixed effects to control for well-known confounds linking deprivation, child stunting, and child human capital at the household level. This was important to ensure we observed the association between child stunting and child human capital, and not the association between stunting and household socioeconomic status. By adding household fixed effects, we controlled for all observed and unobserved, unmeasured fixed household variables. Examples of these confounds include the height and education of the parents, abilities and preferences of the parents, household socioeconomic status (e.g., asset wealth), and household demographics. Elsewhere we showed little change in household socioeconomic status over time among the Tsimane' (Undurraga et al., 2010).

3 RESULTS

3.1 Height category and human capital (2008 cross-sectional data)

Table 1 show that, compared with their nonstunted peers, stunted children differed significantly in completed grades of schooling and in math skills. Table 1 and Figure 2 show that stunted children had significantly lower mean (1.5) and median (1) grades of completed schooling than nonstunted children (mean = 1.8; median = 2) (P = .01). Almost two-thirds (59.2%) of stunted children scored zero on the math test, compared with 56.1% and 50.8% among moderately stunted and nonstunted children. The mean math score of stunted children (0.9) was significantly lower than the mean math score of nonstunted children (1.1) (P = .03). We found no statistically significant difference in reading, writing, or local plant knowledge between children of different height categories. In sum, despite low levels of schooling and few opportunities to practice academic skills, descriptive statistics suggested that nonstunted children had more years of completed schooling and more math skills than their age-sex matched stunted peers, but they did not have better reading or writing skills, and they did not know more about local plants than other children.

Table 2 contains the results of the regression analyses (Equation 1; OLS) and shows three findings. (i) Stunting bore a negative association with schooling and with the probability of having any math or writing skills. Compared with their nonstunted age-sex peers, stunted children had 0.38 fewer completed grades of schooling (P = .05) and were 15% less likely to have any math skills (P = .04) and 13% less likely to have any writing skills (P = .07). (ii) Moderate stunting bore no statistically significant association with schooling, or with math and reading ability, but children who were moderately stunted were 9% less likely to have any writing skills compared to nonstunted children (P = .08). (iii) Children with moderate stunting scored 1.86 fewer points (or ∼19% less) in the test of local plant knowledge than nonstunted children (P = .04), but we found no difference between stunted and nonstunted children in their score of local plant knowledge.

Table 2. Association between (i) height categories and (ii) schooling and academic skills among Tsimane' children 6 years ≤ age ≤ 16 years, 2008

		Academic skills
Explanatory variables:	Schooling	Math	Reading	Writing	Local plant knowledge
Stunteda	−0.38**	−0.15**	−0.03	−0.13*	−0.52
	(0.19)	(0.07)	(0.06)	(0.07)	(1.00)
Moderately stunteda	−0.18	−0.09	−0.03	−0.09*	−1.86**
	(0.16)	(0.06)	(0.04)	(0.05)	(0.80)
Male	0.31***	0.13***	0.06**	0.10***	0.44
	(0.11)	(0.04)	(0.03)	(0.04)	(0.56)
Age	0.35***	0.09***	0.07***	0.08***	0.76***
	(0.03)	(0.01)	(0.01)	(0.009)	(0.16)
Distanceb	−0.11***	−0.02***	0.003	−0.03***	−0.04*
	(0.01)	(0.004)	(0.003)	(0.002)	(0.02)
TAPSc	0.15	0.23***	0.74***	0.44***	Not applicable
	(0.11)	(0.03)	(0.04)	(0.04)
Constant	−1.16***	−0.41***	−0.50***	−0.29***	1.15
	(0.29)	(0.08)	(0.08)	(0.08)	(2.50)
R2	0.76	0.72	0.76	0.76	0.16
N	1246	1045	940	1030	129
Household fixed effects	Yes				No

• Notes: *, **, and *** significant at ≤10%, ≤5%, and ≤1%. OLS regressions; robust standard errors (in parentheses) were adjusted for clustering at the village level. See Table 1 for a definition and description of the variables.
• ^a Nonstunted children are the excluded category.
• ^b Walking distance (hours) of the village to the nearest town.
• ^c TAPS = Tsimane' Amazonian Panel Study; longitudinal study with 13 villages. The variable TAPS = 1 if the village is in the TAPS sample, and 0 if the village was part of the baseline survey for the randomized controlled trial.

We used OLS regressions with household fixed effects and robust standard errors in our main results for ease of interpretation. But alternative regression specifications may be more adequate in some cases. For example, a large proportion of children in the sample never attended school, which may result in a violation of the assumption that error terms have to be normally distributed. We addressed this by using a lowered-censored Tobit regression for schooling. Table 3 and Supporting Information Table S1 contain robustness analyses, using different regressions specifications than those used in Table 2. Changes included the use of a lower-censored Tobit regression for schooling because schooling was censored at zero, with 23.6% of children having no schooling; the use of a Poisson regression for raw scores of academic skills and local plan knowledge; logit regressions for binary academic skills; and the use of OLS regressions with random effects for all outcomes.

Table 3. Robustness analysis of Table 2, using different regression specifications and household fixed effects. Summary of coefficients for stunted and moderately stunted children

		Academic skills			Local plant
Height category:	Schooling	Math	Reading	Writing	knowledge
[A] Tobit (schooling) and censored Poisson regressions (skills and plant knowledge)a
Stunted	−0.55***	−0.45***	−0.28	−0.33**	−0.05
	(0.04)	(0.13)	(0.24)	(0.15)	(0.06)
Moderately stunted	−0.27***	−0.31***	−0.28	−0.26**	−0.20***
	(0.03)	(0.11)	(0.20)	(0.13)	(0.07)
[B] Eliminating moderate stunting as a category; nonstunted is used as the reference
Stunted	−0.25*	−0.09*	−0.01	−0.07	0.49
	(0.13)	(0.05)	(0.05)	(0.05)	(0.72)
[C] OLS using HAZ as a continuous variableb
HAZ	0.19**	0.06*	0.01	0.04*	0.22
	(0.08)	(0.03)	(0.03)	(0.02)	(0.32)
[D] Including severe stuntingc
Severely stunted	−0.42	−0.06	0.07	−0.04	1.83
	(0.29)	(0.13)	(0.10)	(0.09)	(1.04)
Stunted	−0.37*	−0.17**	−0.05	−0.15**	−0.83
	(0.20)	(0.07)	(0.05)	(0.07)	(1.06)
Moderately stunted	−0.18	−0.09	−0.04	−0.10*	−1.86**
	(0.16)	(0.06)	(0.06)	(0.05)	(0.84)
[E] Logit regressions for academic skills with odds ratio and 95% CI also reported
Stunted	NA	−0.86***	−0.63**	−0.74**
		(0.32)	(0.29)	(0.32)
Odds ratio		0.42***	0.53**	0.48**
95% CI		(0.23–0.79)	(0.30–0.94)	(0.25–0.89)	NA
Moderately stunted		−0.56**	−0.22	−0.38
		(0.26)	(0.22)	(0.27)
Odds ratio		0.57**	0.80	0.68
95% CI		(0.34–0.95)	(0.52–1.24)	(0.40–1.15)
[F] Using household random-effects
Stunted	−0.38***	−0.15***	−0.06*	−0.13***	−0.47
	(0.10)	(0.03)	(0.04)	(0.03)	(0.81)
Moderately stunted	−0.18**	−0.09**	−0.03	−0.09***	−1.76**
	(0.09)	(0.03)	(0.03)	(0.03)	(0.74)
Hausman test [χ2, (p)]	6.39 (0.17)	3.30 (0.51)	1.70 (0.79)	2.70 (0.61)	0.80 (0.94)

• Notes: Same notes as in Table 1 and same regressions as in Table 2 (main manuscript), except where noted. ^aLower-censored Tobit for schooling and lower-censored Poisson regressions for math, reading, writing, and local plant knowledge with raw scores of academic skills used as outcome variables (see Table 1 for raw scores). CI = confidence interval. ^bHAZ was used as a continuous variable. ^cSevere stunting was defined as HAZ< −3. NA = not applicable.

The results from Table 3 buttress the results from Table 2. The Tobit regression (Table 3, Panel A) showed that stunting and moderate stunting were associated with 0.6 and 0.3 fewer completed grades of schooling (P < .001). The Poisson regressions (panel A), showed that stunting and moderate stunting bore a statistically significant negative association with math (P < .001) and writing skills (P = .03; P = .04), and that moderate stunting was associated with a lower score in local plant knowledge (P = .004). Eliminating moderate stunting as a height category (Table 3, panel B), reduced the magnitude of the coefficients for the stunting variable, but still showed that stunting was associated with 0.3 fewer completed grades of schooling (P = .06) and with a 9% lower probability of having any math skills (P = .06), compared with nonstunted children. Using HAZ as a continuous variable (Table 3, panel C) showed that a unit increase in HAZ was associated with 0.2 more completed grades of schooling (P = .02), a 6% higher probability of knowing any math (P = .08) and a 4% higher probability of having any writing skills (P = .09). This is consistent with previous results, indicating that higher HAZ was associated with an increase in human capital outcomes. We also included an indicator variable for severe stunting (HAZ < −3; Table 3, panel D), which showed no significant results, possibly due to a small sample of children in the severely stunted category (n = 73). The coefficients for stunted and moderately stunted remained largely unchanged. In Table 3, panel E, we used a Logit regression for academic skills, and found statistically stronger results than when using OLS, with previously statistically insignificant coefficients becoming significant. For example, moderate stunting was negatively associated with math skills (P = .03) and stunting was negatively associated with reading skills (P = .03).

Last, we examined the robustness of our results using household random effect models, which are also robust to serial correlation in the error terms. Much of the bias from important confounds, such as household socioeconomic status and parental attributes, is eliminated using household fixed-effects. While regressions with random effects result in smaller standard errors, the coefficients may be biased if the error terms were correlated with the explanatory variables, which was likely in our data. A Hausman specification test, however, suggested that we could not reject the random-effect specification in favor of the fixed-effect specification (Table 3, Section F). The use of a household random-effect model (Section D) showed slightly larger negative coefficients for the stunting variable and that, compared with nonstunted children, stunted children were six percentage points less likely to have any reading skills (P = .09). Moderately stunted children had less schooling (P = .04), and lower scores in tests of math (P = .01), writing (P = .004), and local plant knowledge (P = .02), but not in reading (P = .24).

Supporting Information Table S1 shows that stunting and moderately stunted generally bore no significant interaction effects (P < 10%) with the following: (i) child's sex, (ii) child's age, (iii) village road access, and (iv) household socioeconomic status. Stunting and moderate stunting also bore no significant interaction effects with schooling in the regressions with academic skills as outcomes.

4 DISCUSSION AND CONCLUSIONS

The study yielded two main findings and two puzzles. First, the ubiquitous cross-sectional association between child stunting and child human capital found in previous studies extended to remote societies without well-established educational systems. The magnitude of these associations was meaningful. For example, Table 2 suggested that stunted children had 0.4 fewer completed grades of schooling than their nonstunted age-sex peers. Since the average number of completed grades of schooling was 1.7, a 0.4 deficit amounted to ∼24% fewer completed grades of schooling in the area studied.

The second main finding related to the longitudinal relation between changes in stunting and human capital over time. We found that persistent stunting was associated with 0.4 fewer completed grades of schooling, but that it was not consistently associated with weaker academic skills. Faltering was associated with a 15–21% lower probability of having any math or writing skills. Because the datasets we used were part of larger studies that examined several different aspects of the transition of the Tsimane' from a largely autarkic to a market economy (Leonard et al., 2015b; Undurraga et al., 2016a), we had to limit the length of the education module to lower respondent burden. The ability to sign one's name, to read simple sentences, or to solve basic math operations resemble the type of questions used to measure child human capital in the Young Lives Study, the largest international comparative longitudinal study of child growth (n∼12,000, in four nations: India, Ethiopia, Peru, and Vietnam) (Barnett et al., 2013; Crookston et al., 2013). On the other hand, we found recovery from earlier stunting leads to estimates for schooling attainment that are no different from those for children who were never stunted. However, we could not fully discard the possibility that our measures of human capital were not sensitive enough to capture small differences between children who recovered and those who were never stunted.

Our results in some respects agreed with—but also differed in other respects from—the results of the Young Lives Study. Crookston et al. (2013) measured changes in HAZ categories of children between about 1 and 8 year of age and, depending on the nation, found that 18–32% of children changed their height status from stunting to nonstunting or vice versa. Children in the four nations who remained stunted were more likely to be overage for their grade, a result consistent with our finding that persistent stunting was associated with less completed grades. As in this study, Crookston and colleagues found no consistent association between schooling and recovery; in Peru and Vietnam recovery was associated with an increase in the age of the child for their school grade, but in Ethiopia and India recovery bore no significant association with appropriate school age. In academic skills, we found only one consistent association: an inverse association between faltering and math and writing skills. In contrast, Crookston et al. only found a consistent negative association between persistent stunting and math scores, reading comprehension, and receptive vocabulary. Children aged 5 to 8 years who recovered improved their math, reading, and receptive vocabulary compared with children who remained stunted, but did not fully catch up with children who were never stunted (Crookston et al., 2013). However, in a study restricted to Peru, children aged ∼5 years who recovered scored as well in verbal vocabulary and in quantitative tests as children who were never stunted (Crookston et al., 2010). This finding was consistent with our study; we found that children who recovered were statistically indistinguishable from those who were never stunted, suggesting an improvement in human capital outcomes. These and other studies (Berkman et al., 2002; Casale & Desmond, 2016; Fink & Rockers, 2014) suggested that, compared with children who remained nonstunted, children who faltered, or who remained stunted, did not tend to fully catch up in academic or cognitive skills.

We found two puzzling results. First, it is not clear why moderate stunting, but not stunting, was associated with lower scores in local plant knowledge. The statistical significance of this finding could be spurious and attributed to a small sample size of children in these categories. Nevertheless, the negative association between moderate stunting and local plant knowledge was consistent with our other findings showing negative associations between stunting and other indicators of human capital. Taken together, the evidence suggested that the factors that stunted growth, such as malnutrition or disease burden, may also have decreased the efficiency of cultural transmission of traditional ecological knowledge.

The second puzzling result was that stunting or moderate stunting almost never bore a significant association with reading skills. Our measure of reading skills may have been hampered by the fact that the sentences were written in Spanish, the second language of instruction in schools, which only some Tsimane' can speak. One possible explanation for the lack of results in reading, compared with other academic skills, is that the tests of math and writing did not hinge as much on understanding Spanish.

The study had at least three limitations. First, measurement errors with reported age, height, and perhaps years of schooling may have occurred in our data (Godoy et al., 2008; Gurven et al., 2007). This was largely due to parental report of a child's age and grades of schooling instead of relying on vital and school records, which were many times nonexistent.

Second, we could not control for unobserved abilities, such as drive, attention, or patience. If these traits were fixed and covaried positively with child human capital and negatively with stunting, then our estimated negative coefficients of stunting would be more pronounced than the true values. Likewise, if stunting covaries negatively with factors such as task comfort, that may increase performance (Gibson, Piantadosi, Jara-Ettinger, & Levy, submitted). Unfortunately, it was difficult to quantify these effects with the available data.

Third, we did not have information to capture the pathways that would allow stunting to influence human capital outcomes. For example, the early life stressors that led to stunting may have impaired attention, motivation, memory, or a combination of these factors (Aburto, Ramirez-Zea, Neufeld, & Flores-Ayala, 2009; Baker-Henningham, Hamadani, Huda, & Grantham-McGregor, 2009; Chang, Walker, Grantham-McGregor, & Powell, 2002; Fernald & Grantham-McGregor, 1998; Gardner, Grantham-McGregor, & Himes, 1999). We also lacked information on vision and hearing loss in our cohorts, which would have hampered children's ability to accumulate human capital (Chua & Mitchell, 2004; Davis & Hind, 1999; Emmett et al., 2014; Lim et al., 2010; Sharma et al., 2010; Steward-Brown, Haslum, & Butler, 1985; Teasdale & Sorensen, 2007). This did not bias our estimates of the associations between stunting and the outcomes that we considered, but understanding the pathways would be useful for directly addressing the underlying human capital deficits.

We did not measure the social and household context in which stunting occurred. In an earlier study, we found that adult Tsimane' attributed more positive traits to taller children, but not to taller adults (Undurraga et al., 2012). Some of the adverse associations between stunting and child human capital might have taken place because parents invested less in stunted than in nonstunted children, for example, reinforcing a child's endowments. Parental investment has been shown to affect the number of years of schooling of a child, and family environment is a key determinant of cognitive abilities (Alderman & King, 1998; Heckman, 2006; Knudsen, Heckman, Cameron, & Shonkoff, 2006; Luby et al., 2013; Wilder, 2014). Stunting could also have elicited compensatory behaviors from parents through extra stimulation to redress the adverse association with stunting or faltering (Engle & Fernández, 2010; Pollitt et al., 1993; Prado & Dewey, 2014; Wachs, 2009; Wachs et al., 1992). We did not have data on parental attitudes or behaviors toward stunted children or on caretakers' behavior, though we controlled for some of these attitudes and behaviors by using regressions with household fixed effects.

Further, several studies have documented the impact of school quality, and particularly teacher quality (Rivkin, Hanushek, & Kain, 2005; Rockoff, 2004), in children's academic skills and years of schooling. It is theoretically possible that nonstunted children attended better quality schools because more motivated parents were willing to migrate to villages with better schools.

One final potential limitation deserves mention, related to our reference sample. We used the WHO (2006) standards in our analysis because they are currently the most widely accepted and preferred standards for evaluating child growth and nutritional status. However, it is possible that they do not closely represent the growth of healthy local populations, including native Amazonian children, even though there are no significant differences among the six diverse populations that WHO used to produce its standards (Blackwell et al., 2016; Urlacher et al., 2016). Further, the results may not necessarily be affected or biased even if local growth standards, such as those for Amazonian populations, differed from WHO standards. This would be true if, for example, local and international standards were correlated except for random differences.

Prior studies have documented the ubiquity of child stunting in remote native Amazonian societies (Benefice, Monroy, Jiménez, & López, 2006; Blackwell, Pryor, Pozo, Tiwia, & Sugiyama, 2009; Cobayashi et al., 2014; Godoy, Reyes-García, Byron, Leonard, & Vadez, 2005; Pedraza, Sales, de Queiroz, & Al, 2014; Roche, Creed-Kanashiro, Tuesta, & Kuhnlein, 2011). Our results suggested that some of these remote societies might not be able to offer their children enough protection to foster normal growth, leaving open the question of how best to redress faltering and avoid losses in the human capital of these children.

CONFLICT OF INTEREST

The authors declare they have no conflicts of interest.