Methodologic approach for the Biomarkers Reflecting Inflammation and Nutritional Determinants of Anemia (BRINDA) project

INTRODUCTION

Global and country-level decision makers frequently make policy and program decisions with the use of uncertain data. Improving methods to accurately estimate priority global health indicators facilitates the reduction of such uncertainty and assists policy makers in making better and more-informed decisions. In the nutrition field, there is a need for the accurate measurement of micronutrient biomarkers to determine public-health burdens and micronutrient-intervention priorities as well as to understand the relative contribution of micronutrient deficiencies to anemia. The presence of inflammation and infection pose problems with the interpretability of certain micronutrient biomarkers that can either be substantially overestimated or underestimated in the presence of inflammation and infections (1–3). Challenges exist in accurately assessing micronutrients and their role in anemia and in better understanding the other factors (infection, inflammation, and genetics) that contributed to anemia to formulate a response that matches the causes with the right interventions.

To this end, the Biomarkers Reflecting Inflammation and Nutritional Determinants of Anemia (BRINDA) project consists of a collaborative research group who used population-based nutrition surveys to address 3 primary research aims as follows: 1) to assess factors that are associated with inflammation [defined as elevations in the acute phase proteins α-1-acid glycoprotein (AGP) and C-reactive protein (CRP)]; 2) to assess the associations between AGP, CRP, and malaria infection with iron and vitamin A biomarkers and develop approaches to adjust for the effects of inflammation and malaria infection; and 3) to identify risk factors and their relative contributions to anemia in preschool children (PSC) and women of reproductive age (WRA) across geographic settings.

An overview of the BRINDA project and the rational for our priority research questions have been described elsewhere (4). Briefly, the BRINDA project was formed in 2012 on the basis of an identified need to address programmatic issues related to inflammation, micronutrient biomarkers, and anemia etiology. The project is led by investigators from the Eunice Kennedy Shriver National Institute of Child Health and Human development of the NIH, the Global Alliance for Improved Nutrition, and the US CDC. A governing body was established to ensure project quality, shared ownership by partners, and proper management of project operations. The organizational structure includes an agency-led steering committee and investigator-led working groups. The investigators from the 3 lead agencies and representatives from the countries and agencies that contributed data carried out the research.

The present supplement consists of 9 articles including this methodologic overview article (Table 1). In this article, we lay the foundation for the approaches that were used in the subsequent articles. The first of the subsequent articles explores the demographic and health factors that are associated with inflammation (AGP and CRP) (5). The next 4 articles examine the influence of inflammation and malaria infection on iron and vitamin A biomarkers as well as potential methods for adjusting biomarker data to better account for the effects of inflammation (6–9). The following 2 articles assess the relations in iron deficiency, other risk factors, and anemia (10, 11). The final article highlights the research, policy, and programmatic considerations of the BRINDA project findings (12). To the best of our knowledge, the BRINDA project is the largest and most comprehensive project to date to use individual participant data to systematically address these research questions.

METHODS

Statistics

The analyses plans were developed with input from collaborating biostatisticians on the BRINDA project. Analyses were conducted at both the individual data-set level and the combined data-set level. Statistical survey packages (SAS 9.4; SAS Institute) survey procedures, STATA (version 12; StataCorp) svy command, and SPSS version 23 with the complex sample module (IBM Corp.) were used to account for complex survey designs that were used in each respective data set. In the analyses for the micronutrient biomarker–adjustments (6–8), individual data-set analyses that accounted for the complex survey design were done with the use of the survey package in R 3.2.2 software (R Core Team) (42). The individual data-set estimates were also combined with the use of the metafor package in R 3.2.2 software (R Core Team) (43). Heterogeneity of estimates across the data sets was assessed with the use of Cochran’s heterogeneity test (43, 44). In the analyses for inflammation (5) and anemia (10, 11) articles, heterogeneity of effects, both within and between data sets, was a concern. Thus, to combine the data sets and to examine heterogeneity, within–data-set weighting variables were rescaled so that the sum of the weights was proportional to the target population (PSC or WRA) of that survey.

Biomarker adjustments

The following 3 approaches are presented in this supplement to adjust ferritin, sTfR, and RBP biomarkers for CRP, AGP, malaria infection, or a combination: exclusion, correction factor (CF), and regression correction (RC) (Table 4). Prevalence estimates of micronutrient deficiencies with the use of no adjustment and the 3 adjustment approaches were compared with McNemar’s chi-square statistics. Statistical significance was defined as a P < 0.05 before applying the Bonferroni corrections to correct for multiple comparisons (P = 0.05 ÷ k, where k equals the number of comparisons).

TABLE 4.

Approaches to adjust iron and vitamin A biomarkers for inflammation and malaria infection: the BRINDA project¹

Approach	Method
Unadjusted	• No adjustments for AGP, CRP, or malaria infection.
Exclusion	• Exclude from the data set individuals with a CRP concentration >5 mg/L or AGP concentration >1 g/L or malaria-infection positive.
	• Calculate the estimated prevalence of micronutrient deficiency with the use of the remaining subsample.
CF	• Stratify the data set into groups by inflammation or malaria-infection status depending on the data availability and MB (types of categorizations listed below).
	• Calculate the CF (ratio of the MB value’s GM in the reference group to that of the respective inflammation or malaria-infection group) for each categorization with the use of the equation shown below.
	• Multiply the raw MB values by the appropriate group CF.
	Equation
	where i denotes the data set, ref denotes the reference group, and j denotes the group.
	Category group:
	CRP: 1) no inflammation (CRP concentration ≤5 mg/L) (reference); 2) inflammation (CRP concentration >5 mg/L)
	AGP: 1) no inflammation (AGP concentration ≤1 g/L) (reference); 2) inflammation (AGP concentration >1 g/L)
	Malaria: 1) malaria-infection negative (reference); 2) malaria-infection positive.
	CRP and AGP: 1) no inflammation (CRP concentration ≤5 mg/L and AGP concentration ≤1 g/L) (reference); 2) incubation (CRP concentration >5 mg/L and AGP concentration ≤1 g/L); 3) early convalescence (CRP concentration >5 mg/L and AGP concentration >1 g/L); 4) late convalescence (CRP concentration ≤5 mg/L and AGP concentration >1 g/L).
	AGP and malaria (sTfR only) (9): 1) no inflammation and no malaria infection (AGP concentration ≤1 g/L and negative malaria infection); 2) inflammation and no malaria infection (AGP concentration >1 g/L and malaria-infection negative); 3) no inflammation and malaria infection (AGP concentration ≤1 g/L and positive malaria infection); and 4) inflammation and malaria infection (AGP concentration >1 g/L and positive malaria infection).
RC	• Run linear regression models; outcome variable is ln MB; depending on available data, ln CRP and ln AGP (continuous) and malaria infection (dichotomous) can be included in the model as explanatory variables.
	• Extract slopes from explanatory variables and input into RC equation shown below (slope values multiplied by the CRP, AGP, and malaria infection observations and subtracted from the MB observations).
	• Back-transform adjusted MB values before applying MB cutoffs.
	Equation:
	where β1 is the CRP regression coefficient, β2 is the AGP regression coefficient, β3 is the malaria regression coefficient, obs denotes the observed value, and ref denotes the reference value. MBs CRP, AGP, CRPref, and AGPref are on the ln scale; refs are the maximum values of the lowest CRP and AGP deciles obtained from the combined BRINDA database. The unlogged reference concentrations are as follows—CRP in PSC: 0.10 mg/L; CRP in WRA: 0.16 mg/L; AGP in PSC = 0.59 g/L; and AGP in WRA = 0.53 g/L; only apply adjustments to individuals with either CRP concentrations > CRPref, AGP concentrations > AGPref, or both.

Open in a new tab

¹

sTfR was adjusted for AGP and malaria infection but not for CRP per a biological rationale as described elsewhere (9). AGP, α-1-acid-glycoprotein; BRINDA, Biomarkers Reflecting Inflammation and Nutritional Determinants of Anemia; CF, correction factor; CRP, C-reactive protein; GM, geometric mean; MB, micronutrient biomarker; PSC, preschool children; RC, regression correction; sTfR, soluble transferrin receptor; WRA, women of reproductive age.

Categorical adjustment

We made categorical adjustments to the micronutrient biomarkers for inflammation (CRP concentration >5 mg/L, AGP concentration >1 g/L), malaria (positive infection), or a combination with the use of the exclusion and CF approaches. The exclusion approach entailed estimating the micronutrient-deficiency prevalence with the use of the subsample of the data set in which individuals had nonelevated inflammation or no malaria infection (Table 4). The CF approach involved applying CFs to the raw micronutrient biomarker values with the use of a CF equation as previously described (34, 45) (Table 4). Under this approach, the micronutrient values of individuals who were classified as having inflammation (CRP concentration >5 mg/L or AGP concentration >1 g/L) or malaria (positive infection) were shifted to a reference group (i.e., the segment of the population with no inflammation or malaria infection) (45).

CFs were generated in 3 ways as follows: with internal survey specific data (termed internal correction factor), a previous meta-analysis with the use of distinct data from the BRINDA database (termed Thurnham correction factor) (34, 46), or the meta-analyzed BRINDA database (termed BRINDA correction factor). The BRINDA correction factor was calculated by obtaining coefficient estimates and their variance-covariance matrix from the 2- and 4-group linear regression models from individual data sets and using these values in 2- and 4-group meta-analyses to calculate the combined effect sizes with the use of multivariable random-effects models to allow for variations between data sets. The meta-analyses were conducted on the ln scale, and the synthesized estimates were exponentiated to obtain the multiplicative CFs for the micronutrient biomarker adjustment. Sample sizes were used as weights to combine the summary statistics. The restricted maximum-likelihood approach with an unstructured variance-covariance matrix was used for the meta-analysis.

Continuous adjustment

We made linear adjustments to the micronutrient biomarkers for inflammation (CRP and AGP on a continuous scale) and a categorical adjustment for malaria (positive infection) with the use of the RC approach. Under this approach, we developed a regression-adjustment equation, which consisted of CRP or AGP slopes, or slopes for both (with or without a malaria-infection slope) from a linear regression model and an external reference value (Table 4).

The first step was to build the linear regression model (with the micronutrient biomarker as the outcome and CRP, AGP, malaria infection or a combination, as the explanatory variables). We performed regression diagnostics on the model to examine the appropriateness of the linearity and other modeling assumptions. Box-Cox transformations were used to determine the optimal transformation of the outcome variables (ferritin, sTfR, and RBP) (47). On the basis of this procedure, the ln transformation was deemed most appropriate for ferritin, sTfR, and RBP. Residuals, leverage, and influence were assessed for each data set with the use of the svydiags package in R 3.2.2 software (48, 49). An examination of the residual plot, scatter plots, and Cook’s D resulted in the decision to also use the ln for the explanatory variables CRP and AGP. Collinearity was also assessed in multiple linear regression models that included both CRP and AGP (including or excluding malaria infection) on the basis of tolerance >0.1 and <1 and a variance inflation factor <5.

The second component of the regression-adjustment equation was to include a reference value for CRP and AGP to avoid overadjusting for low levels of inflammation. A common definition of low levels of inflammation across data sets and laboratory methods was obtained with the use of a combined lowest decile (10th percentile) of CRP and AGP from each data set. The deciles were combined with the use of random-effects meta-analysis whereby the metaphor package in R 3.2.2 software was used to obtain a common estimate across data sets [termed external decile (ED)].

The ED was derived with the use of the full BRINDA database with the exception of data from Georgia because the lower concentration of CRP detection was too high (0.5 mg/L). The EDs that were estimated for the regression approach were as follows: CRP = −2.26 ln(mg/L) and AGP = −0.52 ln(g/L) in PSC; and CRP = −1.83 ln(mg/L) and AGP = −0.63 ln(g/L) in WRA. These EDs equate to a CRP concentration of 0.10 mg/L and AGP concentration of 0.59 g/L in PSC and a CRP concentration of 0.16 mg/L and AGP concentration of 0.53 g/L in WRA. A heterogeneity test for the ED was calculated although Mexico 2006 and United States data sets were excluded from the test because these data sets had zero variance in the lowest decile. There was some degree of heterogeneity, but we conducted a sensitivity analysis, and there were no meaningful differences in micronutrient-deficiency prevalences of the regression-adjusted micronutrient biomarkers with the use of internally generated deciles than with EDs (data not shown).

RCs were generated with the use of internal survey-specific data (termed internal regression correction). An SAS macro to adjust micronutrient biomarkers with the use of the internal regression correction approach is shown in the Supplemental Materials. In addition, external RCs were derived with the use of the meta-analyzed BRINDA database (termed BRINDA regression correction). The BRINDA regression correction was generated by running a linear regression to calculate the coefficients for each individual data set and running a meta-analysis to compute a mean coefficient.

Factors associated with inflammation and risk factors for anemia

To quantify the potential factors associated with inflammation, bivariable and multivariable logistic regression models were estimated with inflammation as the outcome with the use of clinically relevant cutoffs. The inflammation article (5) used an exploratory approach to quantify the potential factors associated with inflammation. Variables were selected for inclusion in a multivariable model if they were significant in the bivariable models by calculating Wald’s chi-square test. Data were analyzed at the individual data-set level. In addition, a pooled analyses restricted to countries with malaria data was conducted to further investigate the role of malaria in inflammation when controlling for sociodemographics and anthropometric variables.

The anemia articles (10, 11) used a conceptual approach to quantify the potential factors associated with anemia because the evidence base is better established than that for inflammation. Bivariable and multivariable models were estimated with anemia and severe anemia as the outcomes with the use of clinically relevant cutoffs both separately for each data set as well as pooled by countries with low, medium, and high burdens of infectious disease. The modeling strategy adopted was the “block stepwise regression approach” (50). Variables that have been considered in the literature to have high theoretical importance were retained in the model irrespective of significance. The other variables were examined in blocks starting with the proximal variables in the causal pathway and ending with distal variables. Theoretically, plausible 2-way interactions that were based on theory were tested and remained in the model at the P ≤ 0.10 level of significance.

The articles on potential factors associated with inflammation (5) and risk factors for anemia (10, 11) used statistical approaches to adjust for the sampling weight, strata, and cluster (as applicable) to account for the complex survey design for an individual data-set analysis. For pooled analyses, population-based weights were converted to standardized weights by dividing the individual weights by the mean population-based weight. The sum of the study-level weights was rescaled to represent the size of the population represented by the data set. In the multivariable analysis, variables that were not measured in ≥3 data sets were omitted.

RESULTS

DISCUSSION

The first important contribution of the BRINDA project was to refine approaches to estimate the prevalence of micronutrient deficiencies (6–8). The BRINDA group compared different approaches to account for the effects of inflammation and malaria infection on iron and vitamin A biomarkers with the use of a predefined set of characteristics to summarize the strengths and weaknesses of each approach (Table 7). The BRINDA group was only able to consider 3 of 5 characteristics as follows: 1) precision, 2) variability reflecting relation between micronutrient biomarkers and severity of inflammation, and 3) feasibility to implement across countries. Two other important considerations (validity and the ability to monitor trends) could not be assessed with the use of the current BRINDA database because gold-standard micronutrient biomarkers and time-series data were not available. These characteristics should be assessed in future studies.

All adjustment approaches resulted in substantial changes in the estimated prevalence of micronutrient deficiencies compared with unadjusted values with even greater differences when the RC approach was used. The CF approach resulted in similar prevalence estimates as with the exclusion approach but had the advantage of retaining the full sample. The exclusion and CF approaches rely on cutoffs for elevated CRP and AGP that were not derived for the purposes of adjusting micronutrient biomarkers; thus, the relevance of these cutoffs when adjusting for micronutrient biomarker concentrations is unclear. The BRINDA group concluded that the RC approach was more promising than the exclusion or CF approach was because it accounted for the influence of the entire spectrum of inflammation. With the use of the BRINDA database, we illustrated that there was no clear threshold value below which the micronutrient biomarkers (ferritin, sTfR, and RBP) were unaffected by CRP and AGP; rather, the relation appeared to follow a linear pattern throughout the range of CRP and AGP, thereby lending support for the use of the RC approach.

Other authors have used the RC approach (18, 54), but the reference value that was selected by the BRINDA group to avoid overadjustment at the lower concentrations of AGP and CRP was a major divergence from previous studies. For instance, Engle-Stone et al. (18) used a similar RC approach but selected higher reference values of 3.75 mg/L for the CRP concentration and 0.75 g/L for the AGP concentration. In addition, the authors only applied the RC adjustments to those individuals with CRP concentrations >5 mg/L or AGP concentration >1 g/L. The BRINDA group showed that even at low values of CRP and AGP there was an effect on micronutrient biomarkers, and therefore we have proposed a lower reference value (10th percentile) to use when applying RC adjustments.

The development of a new-regression adjustment approach is a major step forward because no universally accepted methods for accounting for inflammation when assessing micronutrient status currently exist to our knowledge. Although the WHO currently recommends the omission of individuals with elevated acute-phase protein concentrations (AGP or CRP) when calculating depleted iron stores using ferritin, this approach often results in a substantial portion of the data being lost from analyses in areas with high levels of inflammation. To our knowledge, there are also currently no WHO recommendations for adjusting other iron biomarkers (e.g., sTfR) or any vitamin A biomarkers (e.g., RBP or retinol). The lack of a consensus on how to adjust micronutrient biomarkers results in various approaches being used, which hinders the interpretability and comparability of micronutrient-deficiency data. It also continues to impede inflammatory biomarkers becoming a standard measurement in nutrition surveys.

The second major contribution of the BRINDA project is the characterization of potential factors associated with inflammation (5) and risk factors for anemia (10, 11) to better understand the drivers of these widespread phenomena. Inflammation is increasingly recognized as playing an important role in disease (55) and nutritional deficiencies, but little is known, to our knowledge, about the prevalence and demographic differences of inflammation in low- and middle-income countries. In contrast, anemia has long been recognized as a problem with recent global estimates in children (43%) and women (33%) revealing little to no reduction in anemia between 1995 and 2011 (56). Although poorly designed and implemented interventions have largely contributed to this lack of progress, an additional failure has been a poor understanding of anemia risk factors. Recently, a cause-specific attribution to anemia has been modeled with the use of estimation techniques that required a high degree of extrapolation (57). Although this is an important step forward, it underscores the need to assess risk factors directly. The body of work in the current supplement highlights the importance of the use of real data to assess risk factors for anemia and provides a conceptual framework for dealing with such analyses.

There are several strengths of the BRINDA project including its multipartner and collaborative nature, comparability of laboratory methods, and the large sample size from diverse countries. There are also important caveats about the data available. Data were aggregated from household-based surveys. The surveys were not designed a priori to answer the BRINDA project research questions, and as such, there were some differences in what variables were available and, in some cases, how variables were defined. In addition, although all WHO regions were represented in the BRINDA database, the data sets were selected with the use of a convenience sample and did not collectively represent the global population. Last, the findings cannot necessarily be applied to clinical aspects of inflammation and anemia in individual patients; thus, this investigation was focused at the population level (3).

Because the data sets were derived from cross-sectional surveys, we were unable to establish a temporal relation between the exposures and outcomes. In regard to the biomarker papers, this inability meant that a differentiation between true micronutrient deficiency and the body’s biological response to inflammation was not possible. In addition, we were unable to perform sensitivity analyses that compared adjustment approaches in the absence of gold-standard measurements of micronutrient biomarkers [iron (bone marrow aspirates) and vitamin A (liver reserves)]. However, as previously mentioned, the RC approach appears to better reflect the underlying linear relation between the micronutrient biomarker and inflammation.

Similar to the biomarker papers, the use of cross-sectional data was also a weakness for the inflammation and anemia papers because of the sequential nature of the relation between these outcomes and their risk factors. The BRINDA group was unable to fully capture the complex biological interplay between anemia, infections, inflammation, and micronutrient deficiencies. In addition, it was not possible to distinguish between different types of anemia in the present investigation. The measurement of the mean corpuscular volume, reticulocyte count, and haptoglobin may better elucidate the interrelations between the risk factors for anemia (58). There were also important risk factors for anemia and inflammation that were either not assessed or were not assessed in a sufficient number to be included in our models such as helminth infections, genetic disorders, and folate, vitamin B-12, vitamin D, and zinc deficiency (59–63).

The BRINDA project has made advances toward improving biomarker interpretation and identifying the major causes of anemia. It has also highlighted the value of the use of nutrition surveys to address important research questions at the same time as showing the need to conduct micronutrient surveys in a harmonized fashion and undertake longitudinal studies. Further research is needed to identify and compare different adjustment approaches, understand their biological implications, examine whether different types of infections modify the relation between inflammation and the micronutrient biomarkers, and explore adjustments to other micronutrient biomarkers. In addition, measuring more risk factors for anemia and performing structural equation modeling would lead to a better understanding of the factors that contribute to anemia, which, in turn, would result in more-effective programs.

In conclusion, this research has important policy and programmatic implications. Better prevalence estimates of iron deficiency and vitamin A deficiency would improve our ability to identify populations who are in greatest need of nutrition programs or populations who could potentially be at risk of excessive micronutrient intake. At the individual level, of particular importance is moving toward screening individuals for iron deficiency before supplementation. To mitigate the potential adverse effects of unnecessary iron supplementation, this screening requires the ability to assess iron status at the individual level, particularly in settings with high infection burden. Although it is unclear whether the results of BRINDA project can be applied to individuals, this project serves as a catalyst for future longitudinal and intervention studies on individual biomarker assessment (64). Above all, understanding the relative contribution of nutritional and other risk factors for anemia will improve the design of interventions that address the multifactorial causes of anemia. The insights that are gained from this project will help guide global and country investments to prevent and control micronutrient deficiencies and anemia globally.