Reflections on age imprecision and child nutrition in Uganda

We write in reference to a very interesting study published in the American Journal of Physical Anthropology online in July of 2019, titled “Nutritional status of Ugandan school-children: The effect of age imprecision.” In the study, Ornella Comandini and colleagues explore the effects of mis-measurement in child age on estimating the prevalence of various types of malnutrition in Uganda, including stunting, underweight, wasting, and obesity (Comandini et al. 2019). While we applaud the focus on the important issue of age imprecision in nutrition surveys, and find the analysis of the data at hand to be well executed, we are less sanguine than the authors are over the potential for age mis-measurement to drive spurious conclusions about the dynamics of stunting in broader at-risk populations.

We have three major concerns: (a) the age measurements taken by the authors do not mimic any of the methods used by large international demographic surveys; (b) the children under observation experience fundamentally different growth trajectories than the modal child in a low-income country; and (c) the largest concerns with measurement error induced bias in height-for-age z-scores (HAZ) relate to comparisons or treatment effects within populations, not to the population level estimates explored by Comandini et al.

1 NATURE OF MEASUREMENT ERROR

The authors are correct that misreporting child age can result in inaccurate prevalence estimates of stunting and spurious relationships between birth timing and child health (Agarwal et al., 2017; Larsen, Headey, & Masters, 2019). However, it is not clear that the types of measurement error present in this data are similar to those found in nationally representative data sets such as the Demographic and Health Surveys (DHS). By 2012, over 500 academic papers had used DHS child health or nutrition data in published manuscripts, making these some of the most widely used data in all of child development research (Short Fabic, Choi, & Birda, 2012; Assaf et al. 2015). DHS enumerators are trained in specific data collection procedures, including an extensive and complicated script for identifying child age and birthdate. If no official documentation of a birth can be found, or if the birthdate is not easily recalled by survey respondents, survey enumerators begin with a child's age in years and then use a regimented procedure for eliciting the month of birth based on local events that the respondent recalls around the time of the birth (DHS Program 2019; ICF Macro, 2009; Larsen et al., 2019).

We suspect that it is this procedure, combined with the lack of official documentation of birth dates common in low-income countries that leads to the particular age mis-measurement concerns recently espoused in the broader economic and demographic literature. As described in Agarwal et al. (2017), Larsen et al. (2019), and Finaret and Masters (2019), the measurement error in child age is decidedly nonrandom and generates systematic patterns in child HAZ across age in months and month of birth due to heaping on particular ages and birthdates. There are in fact two artifacts that occur in these surveys and cause bias in stunting prevalence, pushing in opposite directions, and it is not clear that either of these effects would be present in the Comandini et al. data.

The first artifact is random error in reported month of birth, which may cause the prevalence of stunting to be estimated as higher than it actually is by increasing the tail-widths of the HAZ distribution. The second artifact is age heaping, which causes age to be underestimated on average due to the tendency to round down versus round up when enumerators or parents are estimating a child's age. This artifact causes a downward bias on stunting prevalence, because HAZ will be measured as higher than it should be since children will be recorded as younger than they are. Perhaps closer to the measurement error in Comandini et al., Larsen et al. (2019) simulate random measurement error where 11% of children have their birth months replaced by a random month, and they find a small (about 1% point) expected impact on the prevalence of stunting, similar to the results in Comandini et al. However, in their Discussion section 4.1, the authors imply that their findings are in contrast to previous work.

It is not precisely clear how Comandini et al. elicit their “age declared” variable, nor is it possible to determine whether any information was based on official birth registration information. It appears that none of the procedures employed by Comandini et al. reflect the well-established methods used by the DHS, or by UNICEF's Multiple Indicator Cluster Survey or in the World Bank's Living Standards Measurement Study, all of which have similar procedures for eliciting child ages. Moreover, the authors themselves remove around 20% of the observations due to invalid age measures, and it is not at all clear that these observations would be excluded under the criteria of any nationally representative survey. While the Comandini et al. data are consistent with age misreporting, it is likely not the same as the age misreporting in the DHS or other nationally representative data sets commonly utilized in nutrition research. It is not a priori clear in any sense whether the systematic errors in DHS data and those in the Comandini et al. data will generate different impacts on stunting rate estimates, but it seems likely that the underlying measurement error generating processes are different.

2 STUDY POPULATION

As the authors note in their introduction, 28.9% of children under five were stunted in Uganda in 2016, but only 11.9–12.7% of children were stunted in the authors' study population (Table 3). This indicates that the population under study is not a representative sample of Ugandan children at risk for malnutrition. Further, given that linear growth faltering can continue to occur through school age and that there is only a limited possibility of catch-up growth, one would expect greater stunting prevalence in the primarily school-aged population the authors study compared to the younger population of children sampled in the DHS (Leroy & Frongillo, 2019). As Comandini et al. rightly note in their introduction, there is a dearth of data on the nutritional status of school-aged children in Sub-Saharan Africa. While this is a helpful contribution to the broader literature, it also limits the external validity of their study for researchers concerned about age mis-measurement in more commonly used data sources.

The difference in the population distributions of HAZ between the nationally representative DHS and the selected sample in Comandini et al. is directly relevant to the effect of age mis-measurement on stunting rate estimates. In a population with a high percentage of stunted children and a reasonably symmetric distribution of HAZ around the median, a large number of children will be “near the threshold” for stunting. For those children, a small age-mismeasurement could easily move them from one side of the threshold to the other (from “normal” to “stunted” or vice-versa). However, in a population with very low stunting levels and relatively high (for a low-income country) mean HAZ, a much larger age mis-measurement problem would be required to substantially affect the population level stunting rate estimates or mean HAZ. This is especially the case for older children, for whom growth has slowed with age and relative differences in HAZ are lower for the same given degree of measurement error in child age.

3 LEVELS v. DIFFERENCES

The extent to which age mis-measurement has affected the estimates of country-level stunting rates remains open, but we agree with the authors on the probable magnitude of such an effect and agree that it is relatively small. However, our biggest concern is not about potential bias in population stunting rate or mean HAZ estimates, but with bias on estimates of the determinants of child health in the form of group-level differences and treatment effect estimates. In a country-level population estimate of stunting rates, a 1 or 2 percentage point error may be close to meaningless to policy makers or demographers who are tracking long term patterns. Stunting rates have fallen from 47.6% in 1988 in Uganda to 28.9% in 2016 (World Bank, 2018), and a 1 percentage point error would only represent about 5% of the total change. However, when estimating the causal effects of policies, programs, or resources provided to poor households, a bias of 0.2 HAZ due to age imprecision would be enough to turn an estimate centered on zero to a statistically significant treatment effect estimate.

The main challenge currently facing policy makers and public health experts is not to obtain as precise an estimate as possible of the prevalence of stunting, but instead to understand the causes of linear growth faltering and the effectiveness of different policies to address the dynamic, which does not always present as stunting (Frongillo, Leroy, & Lapping, 2019). Our biggest concern with age-misreporting in nutrition surveys is the effect on causal inference for studies that tie exposure to some treatment to a child's birth cohort as a source of random or quasi-random variation, or which attempt to measure directly the health impacts of being born in specific times, months, or seasons. In these cases, age imprecision can seriously interfere with understanding the etiology of stunting (Agarwal et al., 2017; Cummins, 2013; Finaret & Masters, 2019).

Thank you for publishing a study of such interest to the broader nutrition and child development community. We believe that the nutrition research community and the physical anthropology community can learn and benefit from each other's expertise on human growth, and we appreciate the opportunity to be part of that conversation.

Sincerely yours,

Joseph Cummins, Department of Economics, University of California, Riverside

Amelia B. Finaret*, Department of Global Health Studies, Allegheny College