Child development at 5 years of age predicted mathematics ability and schooling outcomes in Malawian adolescents

Authors

  • Mihir Gandhi,

    1. Centre for Quantitative Medicine, Office of Clinical Sciences, Duke-NUS Graduate Medical School, Singapore
    2. Biostatistics, Singapore Clinical Research Institute, Singapore
    Search for more papers by this author
  • Tiina Teivaanmaki,

    1. Department of International Health, University of Tampere Medical School, Tampere, Finland
    Search for more papers by this author
  • Kenneth Maleta,

    1. College of Medicine, University of Malawi, Malawi
    Search for more papers by this author
  • Xiaolian Duan,

    1. Department of International Health, University of Tampere Medical School, Tampere, Finland
    Search for more papers by this author
  • Per Ashorn,

    1. Department of International Health, University of Tampere Medical School, Tampere, Finland
    2. Department of Paediatrics, Tampere University Hospital, Tampere, Finland
    Search for more papers by this author
  • Yin Bun Cheung

    1. Centre for Quantitative Medicine, Office of Clinical Sciences, Duke-NUS Graduate Medical School, Singapore
    2. Biostatistics, Singapore Clinical Research Institute, Singapore
    3. Department of International Health, University of Tampere Medical School, Tampere, Finland
    Search for more papers by this author

Yin Bun Cheung, Singapore Clinical Research Institute, 31 Biopolis Way, Nanos #02-01, Singapore City 138669, Singapore. Tel: +65 6508 8310 | Fax: +65 6508 8317 |
Email: yinbun.cheung@scri.edu.sg

Abstract

Aim:  This study aimed to examine the association between child development at 5 years of age and mathematics ability and schooling outcomes at 12 years of age in Malawian children.

Methods:  A prospective cohort study looking at 609 rural Malawian children. Outcome measures were percentage of correctly answered mathematics questions, highest school grade completed and number of times repeating school grades at 12 years of age. A child development summary score obtained at 5 years of age was the main exposure variable. Regression analyses were used to estimate the association and adjust for confounders. Sensitivity analysis was performed by handling losses to follow-up with multiple imputation (MI) method.

Results:  The summary score was positively associated with percentage of correctly answered mathematics questions (p = 0.057; p = 0.031 MI) and with highest school grade completed (p = 0.096; p = 0.070 MI), and negatively associated with number of times repeating school grades (p = 0.834; p = 0.339 MI). Fine motor score at 5 years was independently associated with the mathematic score (p = 0.032; p = 0.011 MI). The association between child development and mathematics ability did not depend on school attendance.

Conclusion:  Child development at 5 years of age showed signs of positive association with mathematics ability and possibly with highest school grade completed at 12 years of age.

Key notes

  •  Child development during early childhood is likely to be positively related to mathematics ability and school grade completed at 12 years of age.

Introduction

Early child development and education may play an important role in preventing intergenerational transmission of poverty and promote adult well-being (1,2). Blair (3) proposed a developmental neurobiological model in which not only cognitive ability but also self-regulatory skills that relate to emotionality influence children’s school readiness. A review of studies have found that a range of abilities, including physical and cognitive abilities, measured during kindergarten were predictive of early elementary school outcomes (4,5). The ‘lag hypothesis’ assumes that children who have poor developmental status in early life are slower in development but they would eventually catch-up and become proficient. In contrast, the ‘deficit hypothesis’ assumes that these children will suffer a long-term failure in functions and skills (6). A deficit is more of a public health concern than a lag. Without sufficient long-term follow-up data, it is impossible to know which hypothesis is more accurate and whether promotion of early child development may have long-term impact on improving lives.

Many of the earlier studies about early child development and later academic performance were oriented toward empirical prediction instead of studying plausible risk or protective factors (4,5,7). Hence, they did not adjust for potential confounders. Growth stunting and developmental delays are related phenomena (8,9). Stunting is a known risk factor of inadequate child development and school achievement (10,11). Without controlling for growth stunting, among others, it is not clear whether previous studies were showing the impact of poor early development or the impact of early growth stunting.

There is limited amount of information about early developmental status of children and its influence on later cognitive ability and school performance in developing countries. A Brazilian study showed that child development assessed with Griffith Mental Development Scales at 4 years of age was associated with grades attained at 18 years of age (N = 152), having adjusted for mother’s education and a wealth index (1). A Guatemalan study suggested not only that early biological risk factors and cognitive ability predicted psycho-educational test scores in adolescence but also that education could buffer the negative impact of early risk factors (12). However, this study also showed that preschool cognition explained only a small amount of variation in the psycho-educational test scores despite statistical significance. Furthermore, only 222 of the original over 2300 cohort members were analyzed. In another analysis of the same study involving 333 of the original cohort members, the risk factors were also found associated with number of years of schooling (13). Useful although they are, the common problems of loss to follow-up in long-term studies, the limited sample size, and the limited amount of evidence from a low-resource setting, where education quality may be different, suggests the needs for further studies.

One of the United Nations Millennium Development Goals is to ensure by the year 2015 that all children complete primary schooling (14). In 1999, Mozambique and Malawi, Southern Africa, had the world’s highest (72%) and second highest (63%), respectively, rate of primary school dropouts (15). In 2006/2007, Mozambique improved to 56% while Malawi worsened to 64% (15). Hence, development and education for Malawian children are major international concerns.

Using data from the Lungwena Child Survival Study (LCSS), rural Malawi, we aimed to assess association between child development at 5 years of age and mathematics ability and schooling outcomes at 12 years of age. In addition to having a relatively large proportion of cohort members being successfully followed, we also used a multiple imputation (MI) method to assess the impact of loss to follow-up. Height-for-age at the age of 5 years and other covariates were adjusted for as potential confounders as the purpose of this study was to examine the relationship between the outcomes and poor early development, not the predictors of poor early development.

Method

Study overview

The LCSS is an ongoing prospective cohort study of 813 children looking at the health and development of rural Malawian infants and children (16,17). Lungwena is an area in southern Malawi where a government health center serves about 100 km2 rural area with some 20 000 inhabitants. The original cohort for the LCSS was enrolled between June 1995 and August 1996. All pregnant women presenting for antenatal care were eligible for the LCSS and 97% of the pregnant women in the area, at that time, were enrolled. Anthropometric measurements (height/length and weight) were collected regularly from birth. At the age of 5 years, each child was invited to a developmental assessment, carried out by a trained research assistant. Training in the use of the developmental assessment tool consisted of four 2-day sessions over a series of 2 months run by a pediatrician who was an accredited trainer of the Griffiths Scales, and had experience in the use of other developmental screening tools such as the Denver II. The training consisted of lectures regarding normal child development and practical training with normal children in Lungwena. The research assistant was then assessed over a 2-month period regularly by the pediatrician. At age 12 years, each child was assessed for health status, anthropometry, schooling performance and given a mathematics test. Ethical approval for the LCSS was obtained from the National Health Science Research Committee in Malawi (HSRC 93/94) while the additional data collection at 12 years of age visit was approved by the College of Medicine Research and Ethics Committee.

Variables

A developmental assessment inventory was developed based on Denver II, Denver Developmental Screening Test and Griffiths Mental Developmental Scales, with modifications based on focus groups with local research workers to make the milestones culturally appropriate in Malawi (17). The inventory originally had 138 items with 34 gross motor, 34 fine motor, 35 language and 35 social items. Feedback on their face validity and content validity were collected from local pediatricians, research assistants, medical students, and a language expert. After revision/addition/deletion based on the interviews, the set of items were administered to 1130 children. Logistic regression was used to assess whether there was an expected association between each item and age and unexpected association with gender. Reliability for each item was tested using two subsamples of 46 (inter-observer) and 25 (intra-observer) randomly selected children who were seen at 7 and 14 days after initial assessment. Inappropriate items were removed after a consensus meeting, resulting in a final set of 110 items. Association between the developmental scores based on the 110 items and height-for-age Z score, a known predictor of child development (1), was demonstrated in a sample of Malawian children age 3–6 years (9). Items were administered until the child failed seven consecutive items in the same domain. After this, it was assumed that the child would not attain any further milestones in that domain. Details of the items in the four domains are listed in the previous report (17). Percent scores and the summary score derived from 5-years development assessments results were used as independent variables. The percent score was derived for each development domain (gross motor, fine motor, language and social). It was calculated by number of items passed divided by number of items administered (i.e. not ‘don’t know’), multiplied by 100. To prevent inflated type I error due to multiple testing and to study the combined predictive power of all four development domains, a factor score was derived from the percent scores of the four domains using factor analysis with the Bartlett method (18,19). The Bartlett factor score minimizes the sum of the squared components for the error factors. This method produces unbiased estimates of the true underlying factor scores for the participants. The score was then standardized, that is (score − mean)/SD, so that it has mean 0 and standard deviation 1. We call this standardized factor score the summary score of child development.

The outcomes in the present analysis were based on age 12-years assessments: percentage of correctly answered mathematics test questions, highest school grade completed, and number of times a school grade was repeated. In the analysis, highest school grade completed was defined as 0 for children who never attended school, and only children who had ever attended school were included in the analysis of number of times a school grade was repeated. The mathematics test had eight questions, each with four answer choices, and was designed with reference to the mathematics test for children aged 7–14 years used in the Indonesian Family Life Survey (20,21). The test has previously been shown to associate with changes in height-for-age (22), which is an established predictor of cognitive ability in children in low-income countries (1). The association supported the criterion validity of the test. Potential confounders that were adjusted by regression analysis were age and height-for-age Z score at 5-years development assessment, weight-for-age Z score near birth, gender, gestational duration (preterm/term), father’s occupation, father’s literacy, mother’s literacy, and a wealth index. Height-for-age Z score and weight-for-age Z score were derived using WHO 2006 Child Growth Standards (23). Gestational duration was derived using fundal height measurements (24). It was dichotomized as preterm (<37 weeks) and term (≥37 weeks). Wealth index summarized household ownership of radio, bicycle/tricycles, mattress, number of family supporters, ownership of land per person, and number of cattle (cow, goats/sheep, and chickens) (25). The wealth index was assessed perinatally and categorized into three levels: poor (below 40 percentiles), middle (40–80 percentiles), and rich (top 20 percentiles).

Analyses

Cohort members with no missing values in any exposure variables at 5 years of age, outcome variables at 12 years of age, and covariates were included in the analyses. To address the potential bias due to missing values, missing values were replaced by MIs using all the aforementioned variables measured near birth and at 5 and 12 years as predictors (26,27). The imputation by chained regression is an iterative procedure to get one set of imputed values (28). It updates the prediction for the missing values for one variable by (a) regressing the variable upon the observed and tentatively imputed values on other predictors and then randomly samples from the conditional distribution, (b) iterating the process through the other predictors, and (c) updating the imputed values by repeating the process for multiple cycles so that the conditional distribution stabilizes. Then multiple sets of imputed values were obtained separately, and pooled analysis was performed taking into account the uncertainty in the imputations (26). The MI procedure was performed using ICE package by Stata software (27). Considering the high proportion of missing values in the data set, 50 sets of imputation were used (28). The children who died by 12 years assessment were not included in the imputed analysis.

The association between the developmental scores at 5 years of age and percentage of correctly answered mathematics questions at 12 years of age was analyzed by multiple linear regression. The associations between the developmental scores and highest school grade completed till 12 years of age (0, 1, 2, 3, ≥4 grades) were analyzed by ordinal logistic regression. Similar analysis was conducted with number of times school grades repeated till year 12 (0, 1, ≥2 times) as the dependent variable. The above associations were assessed in three sets of models. Model I fitted the regression model using the summary score of child development. Model II fitted the regression models separately with each domain’s percent score. Model III simultaneously included all four domain percent scores. Model I would allow studying overall child development. Model II would enable us to study the association between outcome variable and each domain separately. Model III would evaluate association between outcome variable and each domain after adjusting the effect of remaining domains. The model I based on the summary score of child development was used as the main analysis to prevent inflated type I error due to multiple testing for four developmental domains. To assess the interaction between early development and schooling attendance, an additional multiple linear regression model was performed for mathematics percent score including interaction between summary score of child development and ever attended school. The potential confounders were force-entered into each of the above regression models. All the models were performed using observed data as well as imputed data sets. All analyses were performed in Stata/SE 11.2 for Windows (StataCorp, College Station, TX, USA).

Results

Of 813 children, 489 (60%) children were evaluated at 12-years follow-up, 204 (25%) children died by the age 12 year visits, and majority of the remaining 120 (15%) children were not contactable as their family had relocated to other places. From the 489 evaluated children, 74 children were excluded from the analysis due to missing 5-years development assessments or baseline demographic data required in the analysis. That is, 415 (51%) of 813 children were included in the analysis.

Table S1 describes the background characteristics and development scores at 5 years of age of the cohort members included in the analysis. The proportion of preterm birth was 18.3% and the mean weight-for-age near birth and height-for-age at 5 years were −0.58 and −0.205, respectively. The 415 cohort members included in the analysis and who were excluded were comparable in most aspects, except for slightly more proportion of preterm birth (difference = 6%; p = 0.041), and lower mean weight-for-age near birth (difference = −0.27 Z score; p = 0.003) and height-for-age at 5-years assessment (difference = −0.34 Z score; p = 0.001) in the group excluded from the analysis. Language domain score was lower in the excluded group, with an effect size (mean difference divided by SD) of about 0.22 SD (p = 0.036).

Table 1 tabulates the summary statistics of the outcome variables related to schooling at 12 years of age. Majority of the children had attended school for at least 1 year (83.4%) but only 4.1% of the children completed grade 4 and above. Close to half of the children who had ever attended school had repeated a grade at least once (49.7%).

Table 1. Summary of school attendance, highest school grade completed, number of school grades repeated and percentage of correct mathematics questions at year 12 (N = 415)
VariablesSummary statistics
  1. *Percentages are based on ‘Ever attended school’ (n = 346).

School attendance, n (%)
 Ever attended school346 (83.4)
 Never attended school69 (16.6)
Highest school grade completed, n (%)
 0100 (24.1)
 1122 (29.4)
 2113 (27.2)
 352 (12.5)
 ≥417 (4.1)
 Not reported11 (2.7)
Number of school grades repeated n (%)*
 0171 (49.4)
 1124 (35.8)
 ≥248 (13.9)
 Not reported3 (0.9)
Percentage of correct mathematics questions, mean (SD)36.4 (19.51)

Table 2 shows multiple linear regression coefficients of percentage of correctly answered mathematics questions at 12 years of age controlling for potential confounders. Model I showed a positive association between the summary score of child development and mathematics ability in observed data (coefficient = 1.774, p = 0.057). MI analysis gave similar results (coefficient = 1.841, p = 0.035). Model II revealed that the fine motor score was positively associated with mathematics ability in both the observed (coefficient = 0.412, p = 0.032) and imputed data (coefficient = 0.445, p = 0.011). However, models II and III indicated positive but statistically non-significant association (each p > 0.05) between each domain’s percent score and mathematics ability in most of the cases.

Table 2. Summary of multiple linear regression for percentage of correct mathematics questions at year 12*
RegressorModel IModel IIModel III
Regression coefficient (p) [95% CI]Regression coefficient (p) [95% CI]Regression coefficient (p) [95% CI]
Observed (N = 415)Imputed (N = 609)Observed (N = 415)Imputed (N = 609)Observed (N = 415)Imputed (N = 609)
  1. *All models were adjusted for weight-for-age Z score near birth, gender, gestational duration, father’s occupation, father’s literacy, mother’s literacy, socio-economic level, age at 5-years assessment, height-for-age Z score at 5-year assessment.

  2. Model II: regression models fitted separately each domain score as the independent variable.

  3. Model III: regression model fitted simultaneously all four domain scores as the independent variables.

Summary score1.774 (0.057) [−0.055, 3.602]1.892 (0.031) [0.178, 3.606]
Gross motor percent score0.206 (0.176) [−0.092, 0.503]0.184 (0.216) [−0.108, 0.475]−0.034 (0.874) [−0.461, 0.392]−0.066 (0.742) [−0.461, 0.329]
Fine motor percent score0.412 (0.032) [0.036, 0.789]0.445 (0.011) [0.104, 0.787]0.395 (0.140) [−0.130, 0.921]0.432 (0.068) [−0.032, 0.895]
Language percent score0.160 (0.158) [−0.063, 0.383]0.180 (0.072) [−0.016, 0.376]0.052 (0.705) [−0.219, 0.324]0.080 (0.507) [−0.156, 0.316]
Social percent score0.122 (0.357) [−0.139, 0.383]0.074 (0.570) [−0.183, 0.331]−0.001 (0.997) [−0.307, 0.306]−0.032 (0.834) [−0.330, 0.266]

Table 3 and 4 show the ordinal logistic regression analysis results for highest school grade completed and number of school grades repeated by 12 years, controlling for potential confounders. As indicated by the summary score (model I), the higher the child development status at 5 years, the higher the school grade completed years (OR = 1.218; p = 0.096), and the lower the odds of repeating school grades by 12 years (OR = 0.834; p = 0.320). However, all the analyses concerning highest school grades and repeating did not reach conventional level of statistical significant (each p > 0.05). Analysis of MI data gave similar results.

Table 3. Summary of ordinal logistic regression for highest school grade completed by year 12*
RegressorModel IModel IIModel III
Odds ratio (p) [95% CI]Odds ratio (p) [95% CI]Odds ratio (p) [95% CI]
Observed (N = 404)Imputed (N = 609)Observed (N = 404)Imputed (N = 609)Observed (N = 404)Imputed (N = 609)
  1. *All models were adjusted for weight-for-age Z score near birth, gender, gestational duration, father’s occupation, father’s literacy, mother’s literacy, socio-economic level, age at 5-years assessment, height-for-age Z score at 5-year assessment.

  2. Model II: regression models fitted separately each domain score as the independent variable.

  3. Model III: regression model fitted simultaneously all four domain scores as the independent variables.

Summary score1.218 (0.096) [0.966, 1.536]1.246 (0.070) [0.982, 1.580]
Gross motor percent score1.028 (0.138) [0.991, 1.067]1.021 (0.380) [0.975, 1.069]1.017 (0.451) [0.973, 1.063]1.032 (0.120) [0.992, 1.074]
Fine motor percent score1.029 (0.153) [0.990, 1.070]1.029 (0.241) [0.981, 1.078]1.015 (0.549) [0.966, 1.067]1.036 (0.072) [0.997, 1.076]
Language percent score1.011 (0.319) [0.989, 1.033]0.997 (0.786) [0.972, 1.022]0.999 (0.932) [0.973, 1.025]1.010 (0.367) [0.989, 1.031]
Social percent score1.020 (0.130) [0.994, 1.047]1.01 (0.560) [0.980, 1.039]1.012 (0.418) [0.983, 1.043]1.017 (0.207) [0.991, 1.043]
Table 4. Summary of ordinal logistic regression for number of school grades repeated by year 12*
RegressorModel IModel IIModel III
Odds ratio (p) [95% CI]Odds ratio (p) [95% CI]Odds ratio (p) [95% CI]
Observed (N = 343)Imputed (N = 493)Observed (N = 343)Imputed (N = 493)Observed (N = 343)Imputed (N = 493)
  1. *All models were adjusted for weight-for-age Z score near birth, gender, gestational duration, father’s occupation, father’s literacy, mother’s literacy, socio-economic level, age at 5-years assessment, height-for-age Z score at 5-year assessment.

  2. Model II: regression models fitted separately each domain score as the independent variable.

  3. Model III: regression model fitted simultaneously all four domain scores as the independent variables.

Summary score0.834 (0.320) [0.584, 1.192]0.852 (0.339) [0.614, 1.183]
Gross motor percent score0.991 (0.677) [0.948, 1.035]0.995 (0.828) [0.953, 1.039]0.998 (0.933) [0.953, 1.045]1.005 (0.833) [0.960, 1.052]
Fine motor percent score0.957 (0.152) [0.901, 1.016]0.965 (0.217) [0.913, 1.021]0.961 (0.211) [0.902, 1.023]0.973 (0.361) [0.916, 1.032]
Language percent score0.988 (0.445) [0.959, 1.018]0.983 (0.217) [0.959, 1.010]0.994 (0.681) [0.963, 1.025]0.988 (0.377) [0.960, 1.015]
Social percent score1.012 (0.560) [0.972, 1.053]1.017 (0.391) [0.978, 1.057]1.012 (0.558) [0.972, 1.054]1.016 (0.423) [0.977, 1.056]

The linear regression models for mathematics percent score with interaction between summary score of child development and ever attended school showed some positive but statistically non-significant interactions (p = 0.941). Among those who never attended school, every 1 SD increase in the summary score of child development was associated with 1.6% (p = 0.131) higher score in the mathematics test, whereas among those who attended school the increase was 1.8% (p = 0.339) in the observed data. Similar results were observed with the imputed values (details not shown).

Discussion

We aimed to assess relation between child development at 5 years of age and mathematics ability and schooling achievements at 12 years of age. We demonstrated positive associations between the summary score of child developments at 5 years of age and mathematics ability at 12 years of age in a rural Malawian cohort, with adjustment for a range of covariates including growth stunting in early life. The analysis of the observed data did not reach conventional level of statistical significance (p = 0.057), but the multiply-imputed data gave similar regression coefficient and a stronger level of statistical significance (p = 0.031) allowing for losses to follow-up. The results appeared to be stable despite the use of different methods and sample data, suggesting the reliability of the findings. It was estimated that for one SD increase in the summary score of child development, the mathematics score increased by about 0.1 SD. Taking plus and minus one (or two) SD in the summary score as indicators of slightly (or moderately) high and low levels of child development, the difference in mathematics score would be about 0.2 (or 0.4) SD. This would mean a small-to-moderate effect as per Cohen’s suggestion (29). Similarly, the odds of having higher school grades completed differed by about 50% (100%) for the slightly (moderately) high and low levels of child development. This again appears to be practically significant.

We also assessed associations between individual domain of child development at 5 years of age and the outcomes at 12 years of age. We found that individual domain had a weak association with outcome measures. Only fine motor development had significant association with mathematics score when assessed individually. But there was no clear evidence of one domain being more important than the others in models for highest grade completed and number of times repeating grades at 12 years of age. This suggests that a wide spectrum, instead of a single aspect, of developmental impairment is associated with long-term outcomes. This supports the previous findings about the harmful effects of multiple risk factors on schooling (13). In contrast to the Guatemalan study, school attendance has not shown significant interaction with summary score for child development on mathematics ability.

Our findings are similar to those in the previous studies conducted in developed countries and in Brazil and Guatemala, where early development predicted later academic abilities (1,12). However, the present study analysis adjusted for potential confounders including growth stunting, eliminating the effect of known risk factors and hence provides more accurate estimates of the associations between early development and long-term outcomes. In the present study, there was indication of association between development at 5 years of age and mathematical skills. We did not see a strong enough association between early child development and highest school grade completed and number of times repeating school grades although the observed associations were in line with the hypothesis of poor child development leading to a deficit in schooling outcomes. This finding might be because schooling outcomes could be affected by environmental issues like poverty, illnesses or other practical issues. The prevalence of ever repeating a school grade was high in this cohort. We would hypothesize that being in the rural area is itself a risk factor but the precise factors are unknown. Further research into its determinants will be useful. The data suggest that the plausibility of a long-term deficit associated with poor development in early life. The ‘lag hypothesis’ was a potential reason for people to disregard poor early child development. But the findings appeared to support the ‘deficit hypothesis’ although the p-values did not clearly confirm it. The United Nations’ Millennium Development Goals included not only completion of a full course of primary schooling for all children but also elimination of poverty. In Malawi, completion of primary schooling was uncommon. Education and functional skills such as mathematics may affect income and the risk of poverty (1). Early child development can be promoted, and programs to promote it are within reach (30). The results here suggest that early childhood development is an area that may have implications to not only short-term behaviors but also long-term development and the achievement of the Millennium Development Goals.

Sample attrition is a common problem in long-term follow-up studies. The Guatemalan study, for example, had only about 10% to 15% of the cohort members included in the analyses of long-term outcomes (12,13), whereas the Brazilian study was able to include only about 40% of the cohort members in the long-term analysis. In the present study, over 50% of the cohort members were included in the assessment and analysis. The main reason of exclusion was death prior to age 12, not insufficient follow-up efforts. To make a valid ‘prediction’ about their academic outcomes if they had survived is a question different from to make a valid assessment of the actual association among the survivors. This study only aimed at assessing the actual association. Furthermore, we have also included sensitivity analysis based on MI, an advanced statistical method for handling missing data, and found that the results are robust. There is no evidence of bias despite the attrition rate. Nevertheless, further research with planned long-term follow-up is warranted to confirm the present findings.

There are several limitations to our study. Firstly, the sample size was not powered for the present analysis, and some practically significant associations were statistically non-significant. Nevertheless, the use of MI method somewhat alleviate the issue of sample size and also showed that the results were robust in relation to loss to follow-up. Secondly, test results on a developmental inventory may be presented as Z score or developmental quotient if there are well-established norms from a reference population. The Malawian inventory is a newly adapted, culturally appropriate measure that has not yet developed such a norm. A simple percentage score was calculated for each domain to represent the results. However, a recent study has established that in terms of estimation of association, the choice of scoring approach has little influence on the results (9). Thirdly, the mathematics test used at 12 years of age was not a formally validated instrument. However, it was modeled on an instrument used in a large-scale Indonesian study that included assessment of children at a similar age, the instrument possesses face validity, and it had been shown previously to be associated with height gain in children (22), which is a known predictor of cognitive ability (1). They provide evidence about its validity.

In summary, there were signs that child development at age 5 years was positively associated with mathematics ability and schooling outcomes at age 12 years, even after adjusting for growth stunting and other potential confounders. The strength of association was not negligible. Early child development programs preceding the school admission age may have an impact on long-term cognitive and educational outcomes. Further studies with stronger statistical power and higher follow-up rate are needed to confirm the findings.

Acknowledgements

We are thankful to Dr. Melissa Gladstone for conducting the developmental assessment training. The last author (YBC) was supported by Singapore Ministry of Health’s National Medical Research Council under its Clinician Scientist Award.

Ancillary