See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/317403934

A methodologic framework for modeling and assessing biomarkers of

environmental enteropathy as predictors of growth in infants: An example
from a Peruvian birth cohort

Article in American Journal of Clinical Nutrition · June 2017

DOI: 10.3945/ajcn.116.151886

CITATIONS READS

22 123

9 authors, including:

Josh M. Colston Gwenyth O'Neill Lee

University of Virginia Rutgers, The State University of New Jersey
54 PUBLICATIONS 654 CITATIONS 105 PUBLICATIONS 3,920 CITATIONS

SEE PROFILE SEE PROFILE

Margaret Kosek
University of Virginia
241 PUBLICATIONS 11,517 CITATIONS

SEE PROFILE

All content following this page was uploaded by Josh M. Colston on 08 October 2018.

The user has requested enhancement of the downloaded file.

 AJCN. First published ahead of print June 7, 2017 as doi: 10.3945/ajcn.116.151886.

A methodologic framework for modeling and assessing biomarkers of

environmental enteropathy as predictors of growth in infants: an
example from a Peruvian birth cohort
Josh M Colston,1 Pablo Pe~
nataro Yori,1 Elizabeth Colantuoni,2 Lawrence H Moulton,1,2 Ramya Ambikapathi,1
Gwenyth Lee,3 Dixner Rengifo Trigoso,4 Mery Siguas Salas,4 and Margaret N Kosek1

Departments of 1International Health and 2Biostatistics, Johns Hopkins School of Public Health, Baltimore, MD; 3Department of Epidemiology, University of
Michigan, Ann Arbor, MI; and 4Biomedical Research Unit, Asociación Benéfica PRISMA (PRISMA Charitable Association), Iquitos, Peru

ABSTRACT INTRODUCTION
Background: Environmental enteropathy (EE) impairs the gut’s Environmental enteropathy (EE) is increasingly understood
absorptive capacity and immune function and causes decelerations as a critical pathway that mediates the interaction between nu-
in statural growth that manifest gradually over time. tritional status and immune response to infections (1). For
Objective: To illustrate an approach for assessing emerging bio- children raised in conditions of poor sanitation, environmental
markers of EE, we separately assessed the associations between 3 exposure to enteric pathogens starts early in infancy and occurs
such markers and subsequent nutritional status. continually thereafter, putting the gut in a perpetual state of low-
Design: Stool samples were routinely collected between January level T cell–mediated inflammatory immune stimulation (1–3).
2010 and November 2014 from a cohort of 303 Peruvian infants Over time, this process disrupts normal intestinal function, al-
and analyzed for concentrations of the biomarkers a-1-antitrypsin tering the structure of the mucosal lumen surface and compro-
(AAT), myeloperoxidase, and neopterin. For each marker, a mising the capacity of the gut to absorb available nutrients (2, 4,
mixed-effects linear regression model was fitted for length-for- 5). As the intestinal epithelium is rendered permeable to luminal
age z scores (LAZs) obtained from anthropometric assessments contents, already depleted nutrient reserves are diverted away
that incorporated covariate predictors, polynomial terms for age, from investment in growth toward the mounting of systemic
and product interaction terms to test associations over varying responses to translocated microbes (1–4). Subclinical self-limiting
lag lengths. The biomarkers’ contribution to the models was enteric infections may establish themselves more readily in
assessed with the use of the likelihood ratio test and partial R2 outwardly asymptomatic EE-affected infants, with a cumulative
statistics. impact that manifests only later as a deceleration in statural
Results: Test statistics for the combined inclusion of the 4-model growth (2, 3, 5–7).
terms that involved the biomarker were highly statistically signifi- Because growth faltering starts early in life, there is a limited
cant for AAT (28.71; P , 0.0001) and myeloperoxidase (62.79; window of opportunity to intervene to prevent undernutrition and
P , 0.0001) over a 3-mo lag and moderately so for neopterin mitigate its long-term health impacts (8–10). Sensitive prog-
(13.97; P = 0.0074). AAT and myeloperoxidase seemed to interact nostic markers are needed that can detect the subclinical pre-
strongly with age, with the magnitude and direction of the effect cursors to later deviations from healthy growth while being
varying considerably over the first 3 y of life. The largest proportion
of the variance explained by any biomarker (2.8%) and the largest
difference in LAZ predicted between the 5th and 95th percentile Supported by the Bill & Melinda Gates Foundation, the NIH, and Johns
Hopkins School of Medicine. This is a free access article, distributed under
(0.25) was by myeloperoxidase over a 2-mo lag.
terms (http://www.nutrition.org/publications/guidelines-and-policies/license)
Conclusions: Of the 3 fecal biomarkers studied, 2 that related to that permit unrestricted noncommercial use, distribution, and reproduction in
intestinal function—AAT and myeloperoxidase—were associated any medium, provided the original work is properly cited.
with small but highly statistically significant differences in future The funders played no role in the design and implementation of the study
statural growth trajectories in infants in this cohort, lending further or the analysis and interpretation of the data.
evidence to the EE hypothesis that increased gut permeability and Supplemental Figure 1 and Supplemental Material are available from the
inflammation adversely affects subsequent nutritional status. This “Online Supporting Material” link in the online posting of the article and
from the same link in the online table of contents at http://ajcn.nutrition.org.
association exhibited a complex interaction with age. This trial was
Address correspondence to MNK (e-mail: [email protected]).
registered at clinicaltrials.gov as NCT02441426. Am J Abbreviations used: AAT, a-1-antitrypsin; EE, environmental enteropa-
Clin Nutr doi: https://doi.org/10.3945/ajcn.116.151886. thy; LAZ, length-for-age Z score; MAL-ED, Interactions of Malnutrition and
Enteric Infections: Consequences for Child Health and Development Project.
Keywords: environmental enteropathy, biomarkers, nutritional Received February 3, 2017. Accepted for publication May 9, 2017.
status, mixed-effects model, Peru, MAL-ED doi: https://doi.org/10.3945/ajcn.116.151886.

Am J Clin Nutr doi: https://doi.org/10.3945/ajcn.116.151886. Printed in USA. Ó 2017 American Society for Nutrition 1 of 11

Copyright (C) 2017 by the American Society for Nutrition

2 of 11 COLSTON ET AL.

minimally invasive and permitting frequent monitoring (6, 8, 11, potential (12, 18, 19). LAZs were calculated with the use of
12). Because the luminal histopathology characteristic of EE can WHO Anthro version 3.2.2 and represent the number of SDs the
only be diagnosed by intestinal biopsy, an invasive, technically child’s stature lies from the mean in the distribution of a healthy
demanding procedure that is ethically and logistically unsuitable reference population, accounting for age, sex, edema status, and
for routine screening, attention has turned to surrogate measures stature measure (20). Infants’ lengths were measured on marked
(4, 6, 13). Numerous analytes detectable in blood, stool, and platforms with a sliding footboard. Rigorous quality control
urine have been proposed as potential biomarkers of the pro- measures were implemented and have been described elsewhere
cesses underlying EE; however, to our knowledge, their pre- (19). LAZ was treated as a continuous variable rather than di-
dictive potential has yet to be systematically explored because of chotomized to minimize the loss of information, maximize the
methodologic challenges (4, 14–16). statistical power to detect an association, preserve the variation
We aimed to present a theoretical framework and methodo- in the outcome, and avoid residual confounding (21).
logic approach for assessing the prognostic potential of bio-
markers of EE as predictors of subsequent statural growth
trajectories during infancy. To explore effects over varying lag Exposure variables
lengths, we used data on the fecal concentrations of 3 candidate The primary exposure variables were age and the 3 time-
biomarkers for EE: myeloperoxidase and neopterin (indicators of varying candidate fecal markers of EE: myeloperoxidase, neo-
inflammation) and a-1-antitrypsin (AAT) (indicator of intestinal pterin, and AAT. These biomarkers were measured with the use of
permeability). We hope this approach can be used by those ELISA tests on stool samples collected from the infants at
monitoring biomarkers and infant growth and applied to novel monthly intervals from birth to 12 mo of age and quarterly from
and emerging biomarkers as they become more feasible and 12 to 36 mo of age. In addition to routine monthly collections,
affordable to track. some stool samples were collected between scheduled assess-
ments on the days on which the child’s caregiver reported that the
child had experienced a diarrheal episode (5). The samples were
METHODS collected by fieldworkers who stored them for processing at
2708C without fixative (12). The assays were run on the sam-
Study population
ples locally at the MAL-ED site laboratory per the instructions
We analyzed data from the Peruvian site of the MAL-ED on the package inserts at initial dilutions of 1:500 ng/mL for
(Interactions of Malnutrition and Enteric Infections: Conse- both myeloperoxidase (ALPCO) and AAT (BioVendor) and
quences for Child Health and Development) project, a network of 1:1000 nmol/L for neopterin (GenWay Biotech). If the calcu-
institutions that has recruited and monitored birth cohorts at 8 lated values were found to be out of the range of the standard
locations in different low- and middle-income countries. MAL-ED curve for any of the assays, the samples were run again at a 2-
uses standardized surveillance protocols and assays to collect fold higher or lower concentration (as appropriate) (12). Stool
data on infectious disease incidence, nutritional and anthropo- samples were excluded from the analysis if they were collected
metric outcomes, cognitive development, and biological markers 1) during or #2 d after a diarrheal episode had been reported
(5). The Peruvian cohort consisted of 303 infants enrolled be- (defined from daily household surveillance visits as 3 semi-
tween December 2009 and February 2012 from a peri-urban liquid stools in a 24-h period separated by $2 diarrhea-free
community located in the country’s forested interior 15 km days) or 2) alongside urine samples obtained for a routine lac-
from the city of Iquitos and on a tributary of the Amazon tulose:mannitol ratio test of intestinal permeability (5, 8). This
River, a study setting that has been described in detail elsewhere was to ensure that the analysis was restricted to markers ob-
(17). Selection and enrollment of each child took place within tained from samples of suitable consistency to avoid overly di-
17 d of birth according to eligibility and exclusion criteria that luting the biomarker concentration. They were also excluded
have also been outlined elsewhere (5) (Supplemental Figure 1). if they were recollections of an initial sample that was in-
Daily data on morbidity, care seeking, and infant feeding were adequate for testing but were collected .24 h after the initial
ascertained by caregiver report from twice-weekly household sample. Because the distributions of the 3 biomarker measures
visits, whereas anthropometric and clinical data were collected were highly skewed with a small number of very high outlying
during monthly assessments (5). Prior written informed consent values, the log2-transformed biomarker values were used instead
from a parent or guardian was a prerequisite for all participants’ of the crude concentrations so that the exposure measures were
inclusion in the study. Study staff read the consent form to the more normally distributed and the effect measures derived could
caregiver, and then a copy was left in their possession. Ethical be interpreted as those incurred by 2-fold increments in the marker
approval was given by the Johns Hopkins Institutional Review concentration.
Board as well as from the partner organization, PRISMA Benefit The child’s age in months at the time of outcome ascertain-
Association, in Iquitos, Peru. ment (anthropometric assessment) was treated as the time
metric for the analysis and included in the model as a predictor.
The relation between age and statural growth during infancy
Outcome variable follows a nonlinear curved trajectory (Figure 1). There is a large
The outcome of interest in this study was the child’s length- body of literature relating to methods for characterizing growth
for-age z score (LAZ), a widely used measure of nutritional curves with the use of parametric models. One of the most
status and attained statural growth and a stable sentinel indicator flexible approaches and one that allows for the estimation of
of physiologic insults that has well-documented associations random effects is to incorporate polynomial functions of the
with overall health, adult economic performance, and cognitive age metric as additional terms in the model (19, 22). For this
 FECAL BIOMARKERS OF ENVIRONMENTAL ENTEROPATHY 3 of 11

FIGURE 1 Mean length-for-age z score by age with 95% CIs in a cohort of 303 Peruvian infants.

analysis, quadratic and cubic terms for age (age2 and age3) were formed stools as the reference category and soft and
included as second- and third-degree polynomials. liquid or watery stools as comparison categories.
 Infant feeding: Breastfeeding and the timing of the introduc-
tion of complementary foods and weaning are related to both
nutritional and immune status as measured by the 3 bio-
Potential confounders
markers (23). Daily breastfeeding status and the introduction
The analysis was adjusted for the following covariates: of complementary food to the infants’ diets were ascertained
 Calendar time: To adjust for potential secular trends and ef- by a caregiver report during home visits (5) and categorized
fects that might have influenced the values for both the ex- in accordance with standard definitions described elsewhere
posures and outcome over the course of follow-up, calendar in the literature (26). Subjects’ feeding status on the day
time in continuous months was added to the model and cen- before the biomarker measure was included in the analysis
tered on the midpoint between the earliest and latest obser- as an ordinal time-varying predictor with partially breastfed
vation in the dataset. (the category with the largest number of observations)
 Seasonality: Biomarker concentrations vary seasonally (23), treated as the reference category and exclusively breastfed,
and nutritional status may also be subject to fluctuations over predominantly breastfed, and not breastfed as comparison
an annual cycle. To adjust for the potentially confounding categories.
effects of these variables, Fourier-series sine and cosine
functions were added as terms in the model (see Equation Statistics
1 in the Supplemental Material online) (24, 25).
 Sex: Although LAZ was calculated separately for males and For each candidate fecal biomarker of EE, we constructed a
females, the infants’ sexes were included to adjust for po- linear mixed-effects model to estimate the mean LAZ as a
tential residual confounding. function of age and the biomarker after adjusting for calendar
 Birth weight: The child’s weight in kilograms at birth was time, seasonality, and the a priori–selected infant and mother
included in the model as a continuous time-fixed variable confounders. The relation between mean LAZ and age was
and centered at the sample mean. modeled as a cubic function (main terms of age, age2, and age3)
 Maternal height: The mother’s height in centimeters was and was allowed to vary by the concentration of the biomarker
included as a continuous time-fixed variable centered on the by including the main effect of the biomarker and interactions of
sample mean to adjust for inherited variation in stature. the age terms and the biomarker. Product interaction terms for
 Stool consistency: The concentrations of analytes may be the differential effect of age in the presence of increased bio-
diluted in samples from stools that are of a soft or watery marker concentrations were calculated by multiplying the log2
consistency (8). Numerous samples were excluded from biomarker values by each of the 3 age variables (linear, qua-
the data set because they were collected soon after a diar- dratic, and cubic) and were included in the model to capture
rheal episode or lactulose:mannitol ratio test. Among samples potential-effect measure modification. The covariates entered
that did not meet this exclusion criterion, the consistency of the model as main terms with the exception of seasonality,
each stool sample, as classified by the fieldworkers, was in- which was modeled with the use of Fourier-series sine and co-
cluded in the model as an ordinal time-varying variable with sine functions (24, 25). To account for the correlation of LAZs
4 of 11 COLSTON ET AL.

within a child over time, the model included a random intercept characterize the pattern of dropout. We explored patterns of
and random slopes for the linear and quadratic age terms; these dropout and missing data by comparing the LAZ trajectories
random effects were allowed to be correlated with one another. among infants that dropped out of the study with those that did
The within-child residual LAZs were assumed to be indepen- not and estimating the difference in the mean LAZ among those
dent. Because we interested in predicting future deceleration in with missing outcomes at the next monthly assessment and the
statural growth, the biomarker values were initially lagged by rest of the sample at various time points. We assumed that the
3 mo so that the analysis assessed the association between the data were missing at random, and our analyses were valid under
children’s LAZ at month of age j and the biomarker measured this assumption because the use of a mixed model (a likelihood-
at age j-3. This lagging was to establish the temporal order of based approach) can adequately account for such missingness
the association, an important methodologic consideration in patterns (22, 29). Analyses were conducted with the use of Stata
exploring the complex hypothesized relation between infec- version 13.1 (StataCorp) (30).
tion, immune response, and nutritional status, which is both
bidirectional and acts over a causal pathway that is not in-
stantaneous but cumulative over the course of many weeks and RESULTS
months of life (12). A lag time of 3 mo was considered to be of Table 1 presents the demographic, nutritional, anthropomet-
clinical significance because it allows a sufficient window for ric, and clinical characteristics of the study population. Although
nutritional intervention. The models were also rerun with in- anthropometric data were collected from children aged #5 y,
creasing lag lengths from 0 to 6 mo to compare the associations the mean age at censoring (either administrative or because of
with growth over increasing age windows. A mathematical attrition) was 37 mo and varied considerably because of the
representation of the model is presented in the Supplemental staggered enrollment. The crude (untransformed) biomarker
Material. concentrations exhibited a high level of both within- and
Likelihood ratio tests were used to determine 1) whether the between-subject variability and had overall SDs that were larger
biomarkers modified the association between the mean LAZ and than the mean values, a result of their distribution being heavily
age by comparing our model to a simpler model that dropped the skewed toward higher values. This variability and skewing was
interaction terms between age and the biomarker and 2) whether greatly reduced when they were log2-transformed. A larger
the biomarkers contributed any information to describe the mean proportion of stool samples was of soft consistency rather than
LAZ by comparing our model to a simpler model that dropped formed (67.0% compared with 26.4%), a fact that may have
all biomarker terms (main effect and interactions). R2 statistics introduced measurement error into the calculation of the bio-
were calculated to estimate the proportion of the total variation marker concentrations. Liquid and watery stools made up just
in LAZ that the model explained (27). Partial R2 statistics were 6.6% of the samples, probably because samples collected soon
calculated to estimate the proportion of the total variation in after a diarrheal episode were excluded from the data set. Ap-
LAZ that is explained solely by the biomarker and its interaction proximately half of the stool samples were taken from infants
with age. This partial R2 calculation used the sum of squares who had been partially breastfed the previous day, whereas the
error of our full model and that of the simpler null model that remaining half was divided roughly equally between the feeding
excluded all biomarker terms (main effect and terms for in- categories. Data on child’s birth weight and mother’s height
teractions) (27). The partial R2 may be an optimistic estimate were fairly complete (,5% missing). The slightly higher pro-
because the model was fitted and the R2 computed from the portion of males in this birth cohort is consistent with the natural
same data. human sex ratio at birth.
Standard regression diagnostics, including added variable Figure 1 plots the mean LAZ by month of age for the sample,
plots and observed compared with expected plots, were used to demonstrating that mean nutritional status in this study site
assess the fit of our models to the observed data. To confirm that declined sharply over the first 19 mo of life and then steadily
the polynomial approach accurately characterized the shape of increased to 5 y but without regaining its initial value. This is a
the growth curves, results were compared with similar models fairly typical pattern for a Latin American context and many
that used restricted cubic splines with 5 knot locations determined developing country settings (31). Figure 2 shows the same
by percentiles. Changes in the direction of the association over trends for each of the 3 biomarkers. The overall pattern for all 3
age were verified by running cross-sectional regressions at each was of decreasing biomarker concentration over the first 3 y of
month of age and comparing the sign of the main coefficient. life. For AAT, this downward trend was roughly linear, whereas
Finally, the model was rerun with all 3 biomarkers, and their for myeloperoxidase and neopterin it was curved convexly, ris-
interaction terms were included together to determine whether ing in early infancy and then declining steadily after a peak
any of their significance levels changed in the presence of the between 6 and 9 mo of age.
others. Table 2 presents the parameter estimates from the linear
At 36 mo of age, anthropometric data were missing for 43.6% mixed-effects models that were fit separately for each of the 3
of the initial sample, whereas 37.0% of the subjects had some biomarkers with a 3-mo lag. The data strongly support an age-
intermittent outcome data. LAZ was unavailable for $1 mo of dependent association between the biomarker and subsequent
age but available for a later assessment. Attrition from the study LAZ for AAT (likelihood ratio chi-square test: 28.71; P , 0.0001),
was rarely because of mortality and is instead explained by the myeloperoxidase (likelihood ratio chi-square test: 62.79;
temporary or intermittent outmigration of families from the P , 0.0001), and, to a lesser extent, neopterin (likelihood
study site (numerous participants returned to the study after ratio chi-square test: 13.97; P = 0.0074). Figure 3 plots the
having been lost to follow-up for many months) (28). Attrition growth trajectory predicted by the models for 3 concentration
and intermittent missingness in the outcome were explored to levels of each biomarker (5th, 50th, and 95th percentiles of the
 FECAL BIOMARKERS OF ENVIRONMENTAL ENTEROPATHY 5 of 11
TABLE 1
Characteristics of the study population of 303 Peruvian infants1
SD Observations

Mean Maximum Minimum Between- Within- Per

Characteristic value value value Overall subject2 subject Total infant

LAZ 21.6 2.6 26.0 0.9 0.9 0.5 10,964 36.2

Age at censoring, mo 37 60 0 20 — — 303 —
Biomarker
AAT, ng/mL 4169 13.8
Concentration 598.3 6337.8 0.4 651.0 365.3 598.7
Log2 8.4 12.6 21.5 1.9 0.8 1.7
MPO, ng/mL 4064 13.6
Concentration 12,482.8 111,145.7 0.0 12,910.7 5627.3 11,909.1
Log2 12.6 16.8 1.9 2.0 0.8 1.8
NEO, nmol/L 3892 13.0
Concentration 2902.2 73,505.9 4.1 2856.5 1047.2 2703.9
Log2 10.9 16.2 2.0 1.5 0.5 1.5
Consistency of
stool sample,3 %
Formed 26.4 — — — — — 1029 —
Soft 67.0 — — — — — 2606 —
Liquid/watery 6.6 — — — — — 256 —
Total 100.0 — — — — — 3891 —
Infant feeding,3 %
Exclusively 16.7 — — — — — 650 —
breastfed
Predominantly 15.2 — — — — — 592 —
breastfed
Partially breastfed 49.6 — — — — — 1926 —
Not breastfed 18.4 — — — — — 715
Total 100.0 3883
Birth weight, kg 3.1 4.9 1.7 0.4 — — 299 —
Mother’s height, cm 150.0 167.5 133.5 5.5 — — 290 —
Sex
Males 52.8 — — — — — 160 —
Females 47.2 — — — — — 143 —
1
AAT, a-1-antitrypsin; LAZ, length-for-age z score; MPO, myeloperoxidase; NEO, neopterin.
2
Computed from the number of subjects and mean number of months for which a subject was observed.
3
Values are those for the samples from which the NEO was detected. The distributions for the other 2 biomarkers
differed only negligibly.

overall sample distribution), with 95% CIs from 3 to 39 mo of age age points (data not shown), their results broadly confirmed the
accounting for interaction with age but holding all other covariates direction of association, although individually, few of these
constant. This figure illustrates that both AAT and myeloperox- models attained any degree of statistical significance.
idase interacted strongly with age and in a similar way, such that There was no statistically significant association between
the magnitude, and possibly even the direction of the effect, varied calendar time and the outcome in any of the models, suggesting
considerably over the first 3 y of life. Both markers predicted wide the absence of any secular trend in infant nutritional status in the
differences in LAZs between the 2 extremes of their distributions study site over the follow-up period. In all 3 models, the co-
at 10 and 27–29 mo of age; the myeloperoxidase model had the efficient estimate for the sine function attained some degree of
largest difference during the former period, and AAT had the statistical significance, evidence that trends in nutritional status
largest difference during the latter period. The neopterin model fluctuated seasonally, peaking in the first half of the year. This
did not predict the same differences with age. Instead, the esti- may be because of changes in the availability or market price of
mated trajectories across the distribution of the biomarker were certain foods over annual cycles or to seasonally varying climatic
similar, only converging for a brief period between w24 and 30 mo drivers of diarrheal disease risk (such as fluctuations in the river
of age. These patterns were broadly similar to those predicted by levels that affect sanitation conditions). A likelihood ratio test for
the cubic spline models that were run as a sensitivity analysis the inclusion of the seasonality terms in the model showed that
(data not shown). The CIs for the curves predicted by the 2 ex- they significantly improved the model for all 3 biomarkers. In all
treme values overlapped throughout the 3 y for all 3 biomarkers, 3 models, females had an estimated LAZ that was statistically
indicating a lack of evidence that these predictions reflected a true significantly higher than males over the follow-up period by a
difference in growth for these values of the markers. When cross- quantity that varied from 0.27 (myeloperoxidase) to 0.31 (AAT),
sectional linear regression models were run at various different indicating that the adjustment by sex of the LAZs failed to
6 of 11 COLSTON ET AL.

FIGURE 2 Mean log2 concentration values of AAT, MPO, and NEO by age with 95% CIs in a cohort of 303 Peruvian infants. AAT, a-1-antitrypsin;
MPO, myeloperoxidase; NEO, neopterin.

eliminate residual confounding by sex. Each 1-kg increment been partially breastfed, but there was no equivalent statistically
in birth weight above the mean was statistically significantly significant difference either for those who were exclusively
associated with a 0.90–0.95 increase in LAZ. The equivalent breastfed or those who received no breast milk at all.
estimate for a 1-cm increase in mother’s height was a highly Approximately 90% of the observed variations in LAZs was
statistically significant 0.03 increase. Children whose stool explained by this combination of explanatory variables in all 3
samples were soft experienced a small (0.05) additional decline cases. Although this proportion might seem high for such a
in LAZ over the subsequent 3 mo relative to those who con- complex and variable biological process as growth, it is worth
tributed formed stool samples at the same age. No equivalent bearing in mind that the effects of age on LAZ were modeled
difference was detected for liquid or watery stool samples. This quite closely with the use of the polynomial terms and subject-
may be because they made up too small a proportion of the data specific random effects, which together capture a large proportion
to detect a difference or because higher biomarker concentra- of the within-subject variance. The inclusion of basic baseline
tions in diarrheal stools (which are typically more liquid) offset characteristics of the child such as birth weight, sex, and mother’s
their greater dilution (23). A child having been predominantly height may have ensured that a high proportion of between-
breastfed on the day before the stool sample was taken was subject variation was also captured. However, the partial R2
associated with a small but statistically significant decrease in for the combined contribution of the biomarkers to the explan-
LAZ over the following 3 mo compared with those who had atory capability of the models was small for all biomarkers
 FECAL BIOMARKERS OF ENVIRONMENTAL ENTEROPATHY 7 of 11
TABLE 2
Parameter estimates from multilevel mixed-effects linear regression models of each of the 3 fecal biomarkers on LAZ scores over a 3-mo lag in a cohort of
303 Peruvian infants1
Characteristic AAT MPO NEO

Fixed intercept 20.81 (21.20, 20.42)*** 20.70 (21.21, 20.19)** 21.11 (21.78, 20.44)**
Log2 biomarker 20.07 (20.11, 20.03)*** 20.05 (20.09, 20.01)** 20.03 (20.09, 0.03)
Age
Linear term 20.13 (20.20, 20.06)*** 20.23 (20.32, 20.13) 20.10 (20.23, 0.02)
Interaction with biomarker 0.02 (0.01, 0.03)*** 0.02 (0.01, 0.03)*** 0.01 (0.00, 0.02)*
Quadratic term 0.01 (0.00, 0.01)*** 0.01 (0.01, 0.02)*** 0.00 (20.00, 0.01)
Interaction with biomarker 20.00 (20.00, 20.00)*** 20.00 (20.00, 20.00)*** 20.00 (20.00, 20.00)*
Cubic term 20.00 (20.00, 20.00)*** 20.00 (20.00, 20.00)*** 20.00 (20.00, 0.00)
Interaction with biomarker 0.00 (0.00, 0.00)*** 0.00 (0.00, 0.00)* 0.00 (0.00, 0.00)*
Calendar time 20.01 (20.02, 0.00) 20.01 (20.02, 0.00) 20.01 (20.02, 0.00)
Sine function 0.02 (0.00, 0.04)* 0.03 (0.01, 0.04)** 0.03 (0.01, 0.05)***
Cosine function 0.01 (20.00, 0.03) 0.02 (20.00, 0.03) 0.02 (0.00, 0.03)*
Sex
Males Reference Reference Reference
Females 0.31 (0.15, 0.46)*** 0.27 (0.11, 0.43)*** 0.29 (0.13, 0.45)***
Birth weight 0.95 (0.76, 1.13)*** 0.90 (0.71, 1.10)*** 0.93 (0.74, 1.12)***
Mother’s height 0.03 (0.02, 0.05)*** 0.03 (0.02, 0.05)*** 0.03 (0.02, 0.05)***
Stool consistency
Formed Reference Reference Reference
Soft 20.05 (20.08, 20.02)** 20.05 (20.08, 20.02)*** 20.05 (20.08, 20.01)**
Liquid/watery 0.02 (20.04, 0.07) 20.02 (20.08, 0.04) 20.02 (20.07, 0.04)
Feeding status
Exclusively breastfed 0.02 (20.03, 0.08)*** 0.01 (20.04, 0.07) 0.02 (20.03, 0.07)
Predominantly breastfed 0.12 (0.07, 0.16) 0.12 (0.07, 0.16)*** 0.12 (0.07, 0.17)***
Partially breastfed Reference Reference Reference
Not breastfed 0.00 (20.06, 0.07) 20.02 (20.09, 0.04) 20.00 (20.07, 0.06)
Random-effect parameters
Variance of intercept 0.76 (0.62, 0.94) 0.82 (0.66, 1.00) 0.82 (0.66, 1.01)
Variance of age slope 0.01 (0.01, 0.01) 0.01 (0.01, 0.01) 0.01 (0.01, 0.01)
Variance of age-squared curve 0.00 (0.00, 0.00) 0.00 (0.00, 0.00) 0.00 (0.00, 0.00)
Age/intercept covariance 20.05 (20.07, 20.04) 20.05 (20.07, 20.04) 20.06 (20.08, 20.04)
Age-squared/intercept covariance 0.00 (0.00, 0.00) 0.00 (0.00, 0.00) 0.00 (0.00, 0.00)
Age/age-squared covariance 20.00 (20.00, 20.00) 20.00 (20.00, 20.00) 20.00 (20.00, 20.00)
Variance of residuals 0.12 (0.12, 0.13) 0.11 (0.10, 0.11) 0.11 (0.10, 0.12)
Model statistics
LR test for interaction (df 3) 26.37*** 42.40*** 6.46
LR test for biomarker (df 4) 28.71*** 62.79*** 13.97**
R2 of model, % 87.2 88.7 88.7
Partial R2 for biomarker, % 0.7 2.4 0.2
1
Values in parentheses are 95% CIs. ***P , 0.001, **P = 0.001–0.01, and *P = .0.01–0.05. AAT, a-1-antitrypsin; LAZ, length-for-age z score; LR,
likelihood ratio; MPO, myeloperoxidase; NEO, neopterin.

(Table 2). Myeloperoxidase had the highest partial R2, ex- by myeloperoxidase over a 2- and a 4-mo lag; the former model
plaining 2.4% of the variability in LAZ 3 mo later. The equiv- also predicted the largest difference in LAZ between the 5th and
alent statistic for neopterin was just 0.2%, and for AAT it was 95th percentile (0.25 at 8 mo of age). AAT with a 1-mo lag
0.7%, despite the fact that this latter biomarker was also highly predicted a clinically significant 0.24 difference in LAZ at
significantly associated with the outcome. 27 mo of age; however, the partial R2 for the biomarker in that
Table 3 compares the likelihood ratio test and partial R2 model was much smaller than for myeloperoxidase. Even without
statistics for the 3 biomarkers over increasing lag lengths from applying a lag (the 0-mo model), myeloperoxidase was still highly
0 to 6 mo. It also gives the largest difference in LAZ that the significantly associated with the outcome, a fact that may be ex-
model predicted between the 5th and 95th percentile of the log2 plained by the high within-subject correlation of LAZ from one
biomarker from 3–30 mo of age and the month of age at which age to the next.
this largest difference occurred (an indicator of the clinical When all 3 biomarkers and the terms for their interaction with
significance of the association). For all 3 markers, the statistical age were included in a single multilevel mixed-effects linear
significance of the likelihood ratio test and the R2 statistic was regression model along with the other covariates (a total of 12
highest for a 1- or 2-mo lag and generally decreased with each biomarker terms), all terms for AAT and neopterin lost all sta-
increase in lag length (except for a 6-mo lag). The largest pro- tistical significance, whereas the terms for myeloperoxidase
portion of the variance explained by any biomarker (2.8%) was remained highly statistically significant. The likelihood ratio test
8 of 11 COLSTON ET AL.

FIGURE 3 Growth trajectories with 95% CIs predicted by multilevel mixed-effects linear regression models of AAT (4169 samples), MPO (4064
samples), and NEO (3892 samples) at 3 concentrations of their log2-transformed values on LAZ adjusting for all other covariates over a 3-mo lag in a cohort of
303 Peruvian infants. AAT, a-1-antitrypsin; LAZ, length-for-age z score; MPO, myeloperoxidase; NEO, neopterin.

statistics for each biomarker (compared with a null model in The statistical significance of a detected association is no
which that biomarker was excluded but the terms for the other guarantee of clinical significance because any single biomarker is
2 were kept) were just 6.52 (P = 0.163) for AAT and 4.22 likely to explain a very small proportion of the overall variance in
(P = 0.377) for neopterin but 62.34 (P , 0.001) for myelo- such a multifarious outcome. Furthermore, as shown herein, the
peroxidase. This indicates that myeloperoxidase remained relation between immune response and longer-term enteropathy
highly significantly associated with the outcome in the pres- is not linear. Some degree of self-limiting inflammatory response
ence of the other biomarkers, whereas the AAT terms lost all is protective against enteric pathogens, and the duration over
significance and the lack of association between neopterin and which prolonged inflammation becomes pathologic and the
LAZ remained. balance between injury and insult may vary with age and other
factors (11).
In this study, we piloted a methodologic framework for
DISCUSSION modeling infant growth with the aim of assessing the prognostic
There are numerous methodologic challenges in assessing performance of EE biomarkers. By including polynomial terms
biomarkers of EE as predictors of subsequent statural growth. for age, random effects, baseline characteristics that can be
 FECAL BIOMARKERS OF ENVIRONMENTAL ENTEROPATHY 9 of 11
TABLE 3
LR test and R2 statistics and the largest difference in LAZ in the first 30 mo between the 5th and 95th percentiles of
biomarkers predicted by multilevel mixed-effects linear regression models of each biomarker on LAZs over increasing lag
lengths in a cohort of 303 Peruvian infants1
Lag length

0 mo 1 mo 2 mo 3 mo 4 mo 5 mo 6 mo

AAT
LR test for biomarker (df 4) 11.80* 29.14*** 29.29*** 28.71*** 19.76*** 8.98 13.70**
Partial R2 for biomarker, % 0.4 0.6 0.5 0.7 0.4 0.2 0.3
Largest difference in LAZ 0.17 0.24 0.23 0.20 0.16 0.11 0.12
Age at largest difference, mo 26 27 27 28 28 27 25
MPO
LR test for biomarker (df 4) 33.02*** 46.63*** 77.54*** 62.79*** 52.81*** 24.34*** 37.48***
Partial R2 for biomarker, % 1.8 1.9 2.8 2.4 2.8 1.1 1.6
Largest difference in LAZ 0.18 0.20 0.25 0.19 0.18 0.11 0.14
Age at largest difference, mo 10 10 8 10 28 30 12
NEO
LR test for biomarker (df 4) 9.89* 24.15*** 17.86** 13.97** 8.18 7.68 11.11*
Partial R2 for biomarker, % 0.3 0.6 0.4 0.2 0.0 0.0 0.1
Largest difference in LAZ 0.11 0.17 0.13 0.12 0.11 0.10 0.12
Age at largest difference, mo 3 8 11 11 4 5 6
1
***P , 0.001, **P = 0.001–0.01, and *P = .0.01–0.05. AAT, a-1-antitrypsin; LAZ, length-for-age z score; LR,
likelihood ratio; MPO, myeloperoxidase; NEO, neopterin.

measured affordably, caregiver-reported data on feeding status, growth trajectories shown in Figure 3 are abstractions to some
and coefficients for calendar time and seasonality, we were able extent because it would be rare for any given infant to remain
to explain .85% of the variance in LAZs. The biomarkers at the same percentile of the biomarker distribution from one
could then be assessed based on the additional proportion of assessment to the next, much less over the entire course of the
the variance that they were able to explain as well as on the follow-up.
statistical significance of their effect. All 3 biomarkers were These considerations have implications for those monitoring
highly significantly associated with the outcome for $1 lag both existing and future novel biomarkers for interventions. The
length—results that are consistent with findings from a study explanatory power of any given analyte is likely to be small
in Northeast Brazil for myeloperoxidase and AAT (32) and within a model that includes much more strongly predictive
in Gambia for neopterin (33). In addition, within each in- covariates such as age and sex. As existing biomarkers become
dividual biomarker, this statistical significance seemed to more routinely monitored and new ones become available, they
correlate with the partial R 2 . This builds on the work of will need to be judged by multiple standards: their additive ex-
Kosek et al. (8), who analyzed an earlier subset of stools planatory power relative to other covariates as well as their clinical
from all 8 MAL-ED sites and demonstrated a 1-U increase in and statistical significance. By these standards, myeloperoxidase
an aggregated index of these 3 markers to be statistically performed fairly promisingly as a prognostic marker of under-
significantly predictive of a 0.47 decrease in LAZ, an ap- nutrition because a partial R2 of 2.8% is large for a single enzyme,
proach that has since been applied to similar data from and this measure remained high across increasing lag lengths.
Bangladesh (34). However, there was considerable differ- Furthermore, myeloperoxidase’s statistical association with nu-
ence in the partial R2 between biomarkers, and myeloperoxidase tritional status retained its strong significance even when adjusting
and its interactions consistently explained a proportion of the for the other 2 biomarkers. Future research should focus on this
variance in LAZ that, although small compared with the overall marker’s ability to predict nutritional outcomes over a 2-mo
model, was an order of magnitude higher than that of the other 2 window during the first year of life with the use of data from
biomarkers. Furthermore, the results presented herein serve as a other populations. AAT may yet show some limited promise
reminder that the statistical significance of a biomarker and its as a clinical marker, particularly late in the second and early in
ability to explain variance in the outcome does not necessarily the third year of life, during which it seems to predict its widest
translate into clinical significance of the association. One of the difference in growth trajectories.
largest differences in LAZ between the 5th and 95th percentile of We have also demonstrated that the associations identified
the biomarker distribution predicted by any model was for AAT herein had a complex interaction with age. This age dependence
over 1 mo (0.24). However, the partial R2 and likelihood ratio test was broadly intelligible within the EE paradigm that emphasizes
statistic for this biomarker and its interactions was small compared the dynamic nature of the relation between infection, immunity,
with myeloperoxidase, which predicted a similar difference over a and nutrition during infancy. We speculate that, although some
2-mo lag. Furthermore, because of the high within-subject vari- degree of intestinal inflammation may be beneficial for fighting
ability exhibited by all 3 of these biomarkers and the small effect off enteric infections that might otherwise hinder growth, chronic
sizes, their clinical potential as prognostic markers of future nu- inflammation persisting into later infancy may eventually become
tritional status is likely to be somewhat limited. The predicted pernicious, a delicate balance that may shift as a child ages. The
10 of 11 COLSTON ET AL.

finding that the direction of the relation seems to reverse at least 7. George CM, van Geen A, Slavkovich V, Singha A, Levy D, Islam T,
once over the first 3 y of life, although not without precedent (35), Ahmed KM, Moon-Howard J, Tarozzi A, Liu X, et al. A cluster-based
randomized controlled trial promoting community participation in ar-
should be interpreted with caution and in light of the largely senic mitigation efforts in Bangladesh. Environ Health 2012;11:41.
overlapping CIs for the predicted trajectories. Further studies 8. Kosek M, Haque R, Lima A, Babji S, Shrestha S, Qureshi S, Amidou S,
with larger sample sizes will be needed before this can be Mduma E, Lee G, Yori PP, et al. Fecal markers of intestinal in-
confirmed. The finding that the association between AAT and flammation and permeability associated with the subsequent acquisi-
tion of linear growth deficits in infants. Am J Trop Med Hyg 2013;88:
LAZ loses significance when myeloperoxidase is present in the 390–6.
model also merits further attention. The permeability of the 9. Bhutta ZA. Early nutrition and adult outcomes: pieces of the puzzle.
intestine may be a more distal determinant on the pathway be- Lancet 2013;382:486–7.
tween pathogen exposure and undernutrition, mediated by the 10. Horton R. Maternal and child undernutrition: an urgent opportunity.
proximal factor of inflammation. If this is the case, it would Lancet 2008;371:179.
11. Petri WA, Naylor C, Haque R. Environmental enteropathy and mal-
suggest that nutrient loss because of the extravasion of serum nutrition: do we know enough to intervene? BMC Med 2014;12:187.
proteins may play a secondary role in EE compared with the 12. Kosek M, Guerrant RL, Kang G, Bhutta Z, Yori PP, Gratz J,
primary factor of the diversion of nutrients away from growth Gottlieb M, Lang D, Lee G, Haque R, et al. Assessment of environ-
toward the mounting of inflammatory immune responses to in- mental enteropathy in the MAL-ED cohort study: theoretical and an-
fections. However, the relative roles of these 2 pathways may alytic framework. Clin Infect Dis 2014;59 Suppl 4:S239–47.
13. Trehan I, Kelly P, Shaikh N, Manary MJ. New insights into environ-
themselves vary with age. mental enteric dysfunction. Arch Dis Child 2016;101:741–4.
There are numerous ways in which this methodology could be 14. Kosek MN, Mduma E, Kosek PS, Lee GO, Svensen E, Pan WK,
adapted to address other exposures or different outcome defini- Olortegui MP, Bream JH, Patil C, Asayag CR, et al. Plasma tryptophan
tions. Change in nutritional status could be defined instead as the and the kynurenine-tryptophan ratio are associated with the acquisition
change in LAZ over a given window (8, 14), as a binary indicator of statural growth deficits and oral vaccine underperformance in pop-
ulations with environmental enteropathy. Am J Trop Med Hyg 2016;
of stunting, or by exposure to wasting—either as a time-varying 95:928–37.
weight-for-age z score or by total exposure calculated as the AUC. 15. Campbell DI, Elia M, Lunn PG. Growth faltering in rural Gambian
Other covariates that could be considered for inclusion in similar infants is associated with impaired small intestinal barrier function,
models include diarrheal episodes (which herein were adjusted for leading to endotoxemia and systemic inflammation. J Nutr 2003;133:
1332–8.
by excluding diarrheal stools and adjusting for the consistency of 16. Lu M, Zhou J, Naylor C, Kirkpatrick BD, Haque R, Petri WA, Ma JZ.
the remaining samples), presence of infection with enteric path- Application of penalized linear regression methods to the selection of
ogens, or dietary intake. Studies that involve large panels of an- environmental enteropathy biomarkers. Biomark Res 2017;5:9.
alytes could proceed by applying machine-learning methods to 17. Yori PP, Lee G, Olortegui MP, Chavez CB, Flores JT, Vasquez AO,
Burga R, Pinedo SR, Asayag CR, Black RE, et al. Santa Clara de
identify the optimal subset, which can then be treated as expo- Nanay: the MAL-ED cohort in Peru. Clin Infect Dis 2014;59:S310–6.
sures in a model such as the one presented herein. 18. WHO. WHO child growth standards: methods and development.
Geneva (Switzerland): WHO; 2006.
We thank Benjamin JJ McCormick for his valuable comments and 19. Richard SA, McCormick BJ, Miller MA, Caulfield LE, Checkley W.
suggestions. Modeling environmental influences on child growth in the MAL-ED co-
The authors’ responsibilities were as follows—PPY and MNK: designed hort study: opportunities and challenges. Clin Infect Dis 2014;59:S255–60.
the research; DRT and MSS: conducted the research; JMC: analyzed the 20. WHO. WHO Anthro version 3.2.2. Geneva (Switzerland): WHO; 2011.
data; JMC, EC, and MNK: wrote the paper; MNK: had primary responsi- 21. Altman DG, Royston P. The cost of dichotomising continuous vari-
bility for the final content; EC, LHM, RA, GL, and MNK: provided consul- ables. BMJ 2006;332:1080.
tation on the statistical analysis and interpretation of the results; and all 22. Rabe-Hesketh S, Skrondal A. Multilevel and longitudinal modeling
using Stata. 3rd ed. College Station (TX): Stata Press; 2012.
authors: read and approved the final manuscript. None of the authors
23. McCormick BJ, Lee GO, Seidman JC, Haque R, Mondal D, Quetz J,
reported a conflict of interest related to the study. Lima AA, Babji S, Kang G, Shrestha SK, et al. Dynamics and trends in
fecal biomarkers of gut function in children from 1–24 months in the
MAL-ED study. Am J Trop Med Hyg 2017;96:465–72.
24. Stolwijk AM, Straatman H, Zielhuis GA. Studying seasonality by using
REFERENCES sine and cosine functions in regression analysis. J Epidemiol Com-
1. Ngure FM, Reid BM, Humphrey JH, Mbuya MN, Pelto G, munity Health 1999;53:235–8.
Stoltzfus RJ. Water, sanitation, and hygiene (WASH), environmental 25. Christiansen CF, Pedersen L, Sørensen HT, Rothman KJ. Methods to
enteropathy, nutrition, and early child development: making the links. assess seasonal effects in epidemiological studies of infectious dis-
Ann N Y Acad Sci 2014;1308:118–28. eases—exemplified by application to the occurrence of meningococcal
2. Korpe PS, Petri WA. Environmental enteropathy: critical implica- disease. Clin Microbiol Infect 2012;18:963–9.
tions of a poorly understood condition. Trends Mol Med 2012;18: 26. Labbok M, Krasovec K. Toward consistency in breastfeeding defini-
328–36. tions. Stud Fam Plann 1990;21:226–30.
3. Prendergast A, Kelly P. Enteropathies in the developing world: ne- 27. Nakagawa S, Schielzeth H. A general and simple method for obtaining
glected effects on global health. Am J Trop Med Hyg 2012;86:756–63. R 2 from generalized linear mixed-effects models. Methods Ecol Evol
4. Ali A, Iqbal NT, Sadiq K. Environmental enteropathy. Curr Opin 2013;4:133–42.
Gastroenterol 2016;32:12–7. 28. Kosek M, Yori PP, Pan WK, Olortegui MP, Gilman RH, Perez J,
5. MAL-ED Network Investigators. The MAL-ED study: a multinational Chavez CB, Sanchez GM, Burga R, Hall E. Epidemiology of highly
and multidisciplinary approach to understand the relationship between endemic multiply antibiotic-resistant shigellosis in children in the
enteric pathogens, malnutrition, gut physiology, physical growth, Peruvian Amazon. Pediatrics 2008;122:e541–9.
cognitive development, and immune responses in infants and children 29. Hogan JW, Roy J, Korkontzelou C. Handling drop-out in longitudinal
up to 2 years of age in resource-poor environments. Clin Infect Dis studies. Stat Med 2004;23:1455–97.
2014;59(Suppl 4):S193–206. 30. StataCorp. Stata statistical software: release 13. College Station (TX):
6. Gilmartin AA, Petri WA Jr. Exploring the role of environmental StataCorp; 2013.
enteropathy in malnutrition, infant development and oral vaccine 31. Shrimpton R, Victora CG, de Onis M, Lima RC, Blössner M,
response. Philos Trans R Soc Lond B Biol Sci 2015;370:pii: Clugston G. Worldwide timing of growth faltering: implications for
20140143. nutritional interventions. Pediatrics 2001;107:E75.
 FECAL BIOMARKERS OF ENVIRONMENTAL ENTEROPATHY 11 of 11
32. Guerrant RL, Leite AM, Pinkerton R, Medeiros PH, Cavalcante PA, 34. Arndt MB, Richardson BA, Ahmed T, Mahfuz M, Haque R, John-
DeBoer M, Kosek M, Duggan C, Gewirtz A, Kagan JC, et al. Biomarkers Stewart GC, Denno DM, Petri WA, Kosek M, Walson JL. Fecal markers
of environmental enteropathy, inflammation, stunting, and impaired of environmental enteropathy and subsequent growth in Bangladeshi
growth in children in Northeast Brazil. PLoS One 2016;11:e0158772. children. Am J Trop Med Hyg 2016;95:694–701.
33. Campbell DI, McPhail G, Lunn PG, Elia M, Jeffries DJ. Intestinal 35. Prendergast AJ, Rukobo S, Chasekwa B, Mutasa K, Ntozini R,
inflammation measured by fecal neopterin in Gambian children with Mbuya MNN, Jones A, Moulton LH, Stoltzfus RJ, Humphrey JH, et al.
enteropathy: association with growth failure, Giardia lamblia, and in- Stunting is characterized by chronic inflammation in Zimbabwean in-
testinal permeability. J Pediatr Gastroenterol Nutr 2004;39:153–7. fants. PLoS One 2014;9:e86928.

View publication stats