The Relationship Between Vitamin D Intake and Serum 25-hydroxyvitamin D in Young Children: A Meta-Regression to Inform WHO/FAO Vitamin D Intake Recommendations

Abstract

Background

This work was commissioned by the World Health Organization and Food and Agriculture Organization to inform their update on the vitamin D requirements for children aged <4 y.

Objectives

The objective of this work was to undertake multilevel and multivariable dose-response modeling of serum 25-hydroxyvitamin D (25OHD) to total vitamin D intake in children aged <4 y with the goal of deriving updated vitamin D requirements for young children.

Methods

Systematically identified randomized controlled trials among healthy children from 2 wk up to 3.9 y of age provided with daily vitamin D supplements or vitamin D-fortified foods were included. Linear and nonlinear random effects multilevel meta-regression models with and without covariates were fitted and compared. Interindividual variability was included by simulating the individual serum 25OHD responses. The percentage of individuals reaching set minimal and maximal serum 25OHD thresholds was calculated and used to derive vitamin D requirements.

Results

A total of 31 trials with 186 data points from North America, Europe, Asia, and Australasia/Oceania, with latitudes ranging from 61°N to 38°S, and with participants of likely mostly light or medium skin pigmentation, were included. In 29 studies the children received vitamin D supplements and in 2 studies the children received vitamin D-fortified milk with or without supplements. The dose-response relationship between vitamin D intake and serum 25OHD was best fitted with the unadjusted quadratic model. Adding additional covariates, such as age, did not significantly improve the model. At a vitamin D intake of 10 μg/d, 97.3% of the individuals were predicted to achieve a minimal serum 25OHD threshold of 28 nmol/L. At a vitamin D intake of 35 μg/d, 1.4% of the individuals predicted to reach a maximal serum 25OHD threshold of 200 nmol/L.

Conclusions

In conclusion, this paper details the methodological steps taken to derive vitamin D requirements in children aged <4 y, including the addition of an interindividual variability component.

Introduction

In 2004, the FAO and WHO published global nutrient intake requirements [1]. Many countries adopt these estimates as part of their national dietary allowances and/or food standards, as well as a foundation to develop food-based dietary guidelines [2]. Dietary recommendations to meet nutrient requirements are, by their nature, intended to be iterative, and their revision is usually based on an extended body of evidence [3]. In keeping with this ethos, in 2019, the FAO-WHO decided to update their nutrient intake recommendations for infants and young children (0–3.9 y) [4] and prioritize, among other nutrients, vitamin D, in light of the new evidence that has emerged since 2004.

The availability of a large body of new data on vitamin D has also been the stimulus for a number of other authorities to update their vitamin D recommendations in recent years [[5], [6], [7], [8], [9], [10], [11]]. The approach followed by many of these authorities consisted of the undertaking of a sequence of independent systematic evidence-based reviews, followed by an appraisal of the evidence around the relationship of serum 25-hydroxyvitamin D (25OHD) and the determination of reference level of the critical indicator health outcome of nutrient adequacy, so as to derive population serum 25OHD targets, and these, in turn, are used to establish the recommended vitamin D intake [3]. To date, these vitamin D requirement exercises have had a regional focus, either North America [5], Europe-wide [7], the United Kingdom [9], or the Nordic region [6], and in conditions of minimal UV-B sunlight.

In an effort to provide global vitamin D requirements, the FAO-WHO has decided to update its vitamin D intake requirements with evidence from all regions of the world, irrespective of sunlight exposure. Accounting for sunlight exposure when setting vitamin D intake requirements is very challenging for a number of reasons, not least because often it is not possible to quantify the contribution sunlight exposure makes to serum 25OHD concentrations within the general population [9]. A number of systematic reviews were commissioned to enable this update. One review highlighted serum 25OHD as a useful biomarker for vitamin D status in young children [12]. Two reviews on breast milk vitamin D content [13] and breast milk intake volume [14] provided new intake exposure data. Another review proposed a definition of serum 25OHD threshold for the minimization of nutritional rickets in young children [15]. An additional review provided a summary of the evidence around vitamin D intake, 25OHD status, and health outcomes in young children but included only a cursory analysis of the vitamin D dose-response relationship [16].

The objective of this work was to undertake detailed, multilevel, and multivariable modeling of the response of serum 25OHD to total vitamin D intake in children aged <4 y, including interindividual variability, in order to derive updated vitamin D requirements for young children.

Methods

Eligibility criteria

The study inclusion and exclusion criteria are summarized in Table 1. Studies with healthy children aged <4 y were included, whereas those of children with diseases (e.g., rickets) and certain conditions (such as prematurity and low birth weight) were excluded. Study arms with daily vitamin D supplementation or vitamin D-fortified foods were included, whereas those with weekly, monthly, or single (bolus) vitamin D dose(s) were excluded. Study arms in which lactating mothers received ≤12.5 μg/d vitamin D supplements were included. The control groups within eligible RCTs could consist of a placebo, no vitamin D addition, or low-dose vitmain D supplementation (compared with a higher dose). Study arms with supplementation of other nutrients (e.g., calcium) concomitantly were included as long as the effect of vitamin D could be isolated. Only the data points from 2 wk of age onward were included, as vitamin D status during the first 2 wk of life was considered more reflective of the mother’s vitamin D intake rather than of the infants’. Studies with a minimum follow-up of 2 wk were included, because this was considered to be the minimum duration required for the vitamin D intervention to have an effect on serum 25OHD.

TABLE 1.

Eligibility criteria for randomized controlled trials to contribute data for vitamin D intake requirement modeling

	Inclusion	Exclusion
Participants	Healthy children (including children with vitamin D deficiency) from 2 wk of age <4 y (extended to 9 y to make sure sufficient data was available, as a sensitivity analysis)	Children with diseases (e.g., rickets) and conditions (very preterm and low birth weight)
Intervention	Daily vitamin D supplements or fortified foods, with a follow-up of a minimum 2 wk	Weekly, monthly, single-dose vitamin D supplements or injections
Comparator	Low or zero vitamin D comparator	Invalid comparator (e.g., meat) or unable to isolate the effect of vitamin D
Other	Maternal vitamin D supplementation ≤12.5 μg/d	Maternal vitamin D supplementation >12.5 μg/d

Study selection

Studies from a previous systematic review [16] were used as the starting point for the collection of the vitamin D intake-status modeling in the present work. In brief, the latter review (PROSPERO registration number: CRD42020198843) searched online databases (Medline, Embase, and Cochrane Central) from inception up to June 2020, and a total of 51 vitamin D randomized controlled trials (RCTs) were identified [16]. These RCTs were screened again according to a refined set of eligibility criteria (see Table 1), which had been more tailored to this vitamin D requirement modeling exercise.

In addition, vitamin D guidelines and reviews from other authoritative bodies [5,7,9,[17], [18], [19]] were reviewed, and studies not previously identified were screened against the same eligibility criteria and, if eligible, included.

Data extraction

Only aggregated data were available, with the achieved serum 25OHD (nmol/L) and the vitamin intake (μg/d) both expressed as continuous variables (study arm group means). The data from the eligible RCTs (n = 26) in the original review [16], which were extracted by one investigator and spot-checked by a second investigator, were additionally verified, edited, and extra information added (MRL) (e.g., adding missing data points, intermediate timepoints, and values reported only in figures using PlotDigitizer (plotdigitizer.sourceforge.net)). Data from newly identified studies (n = 5) were also extracted by one reviewer (MRL), and all of this newly extracted data was checked by another reviewer (KDC). The age of the infants was defined as their age at the corresponding time of the measurement. Data from baseline, intermediate, and final timepoints, at ages <4 y, were extracted (Note: Within-study correlation among timepoints and dose groups was accounted for in the model). Standard errors (SE) that could not be derived from standard deviations (SD), confidence intervals (CI), interquartile ranges (IQR), or ranges were imputed with the weighted mean SE of all the included studies [20].

Total vitamin D intake was calculated as the sum of vitamin D intake from the background diet and the vitamin D intake from the supplements or fortified food interventions. Vitamin D intake from the background diet was extracted from the papers whenever reported or requested of the study authors by e-mail where not presented; in cases where the data was not presented nor provided by authors, it was imputed using a single imputation method with data from nationally representative samples from the same country, cognate studies (i.e., same country, age, year, and feeding type), or the Global Dietary Database (www.globaldietarydatabase.org), under the assumption that the surrogate data are sufficiently similar to the study population in terms of characteristics relevant for the dietary consumption (e.g., age and feeding type). For exclusively breastfed infants aged 0–6 mo, vitamin D intakes provided by breast milk were estimated using the FAO-WHO-commissioned systematic reviews [13,14].

Risk of bias assessment

The quality of the included studies was assessed with the Cochrane Risk of Bias Tool 2.0 [21]. The overall strength of the evidence was assessed with a grading of recommendation, assessment, development, and evaluation approach [22].

Data modeling

Under the assumption that the causal relationship between vitamin D intake and risk of rickets or other adverse effects was exclusively mediated by 25OHD with no other direct relationships, a model was established integrating a number of components (illustrated in Figure 1). Based on this model, the daily total vitamin D intake that will maintain serum 25OHD concentrations above or below target 25OHD thresholds in a stated percentage of individuals, as is the convention with nutrient requirement recommendations [3], was estimated. Data analyses were conducted with R (v 3.6.3) and RAnalyticFlow (v 3.1.8) using the package metafor.

FIGURE 1.

Illustration of the steps used to model total vitamin D intake and serum 25OHD and to derive vitamin D requirements. (A) Included RCTs provided aggregated data with total vitamin D intake and achieved study mean serum 25OHD. (B) The relationship between vitamin D intake and serum 25OHD was modeled with random effects multilevel meta-regression dose-response model. The impact of the inclusion of potential covariates that could play the role of modifiers of this vitamin D intake-status relationship was tested with adjusted models. (C) A target serum 25OHD threshold of 28 nmol/L was used as the basis of derivation of the INL98 for children aged 0–3.9 y, whereas an upper threshold of 200 nmol/L was used as the basis of derivation of the UL for children aged 0–3.9 y. (D) Interindividual variability of the response of serum 25OHD at different vitamin D intake concentrations was simulated. (E) The modeling approach was used to estimate the vitamin D intake needed to maintain 98% of individuals over a stated serum 25OHD threshold concentration (INL98) and the vitamin D intake, which is judged to be unlikely to lead to serum 25OHD concentrations associated with adverse health (UL). Abbreviations: INL98, individual nutrient level 98; RCT, randomized controlled trial; UL, upper level; 25OHD, 25-hydroxyvitamin D.

Unadjusted random effects multilevel meta-regression models

The relationship between doses of total vitamin D intake and study mean concentrations of serum 25OHD was fit using the collection of RCT data from included studies (Figure 1A). Random effects multilevel meta-regression models with study, study arm, and time of measurements included as nested random factors to reflect the hierarchical structure in the data were used (Figure 1B). A continuous-time autoregressive structure was assumed for the variance under the assumption that measurements closer in time have a stronger correlation and data are not equally spaced in time. Different shapes were tested, including linear, quadratic, cubic, logarithmic, and 3 knots-restricted cubic spline [23]. The best-fitting model was selected based on the Akaike Information Criterion (AIC), the significance of the parameters for the dose, and additional considerations related to the biological explicability of the model. The model was used to predict the mean of the serum 25OHD at different concentrations of vitamin intake, its 95% CI, and prediction interval (95% PI) [24].

Adjusted random effects, multilevel meta-regression models

The impact of the inclusion of potential covariates in the vitamin D intake-status relationship was tested with adjusted models, using backward and forward stepwise selection approaches. The infants’ age, baseline 25OHD, region, country income category (according to 2020 United Nations classification), 25OHD assay, season, skin pigmentation, and latitude were considered as possible modifiers. In the absence of more appropriate data on exposure to UV-B, and in order to crudely cluster participants based on differences in the potential for the synthesis of vitamin D in the skin, a cutaneous synthesis score with latitude and season was computed (high: ≤40°N/S or >40°N/S in summer or autumn; or low: >40° N/S in winter or spring). Sensitivity analyses were conducted excluding where study authors categorized their participants as “vitamin D deficient” at baseline (only one study mentioned including children with vitamin D deficiency, which was defined in that study as having a serum 25OHD concentration <50 nmol/L [25]), imputed background vitamin D intake, no external standards for vitamin D assay, use of vitamin D₂ supplements, nonexclusively breastfed infants aged 0–6 mo, infants below 6 mo, and follow-up of only 2 wk.

Target serum 25OHD thresholds for the derivation of individual nutrient level 98 and upper-level intake recommendations

Target serum 25OHD thresholds, as they relate to FAO-WHO’s individual nutrient level 98 (INL98) and upper level (UL) reference intakes [26], were used for the present modeling (Figure 1C). The INL98 is intended to estimate the total vitamin D intake needed to maintain 97.5% of individuals over a stated serum 25OHD threshold concentration [26] and was derived from the serum 25OHD threshold of 28 nmol/L for children aged 0–3.9 y [15]. The UL is intended to estimate the total vitamin D intake which is judged to be unlikely to lead to serum 25OHD concentrations associated with adverse health effects in young children and was derived from the serum 25OHD threshold of 200 nmol/L. This threshold was identified (as a no observed adverse effect level) by the FAO/WHO expert group on nutrient requirements based on a systematic review of studies investigating the association between serum 25OHD and vitamin D supplementation and adverse effects, especially hypercalcemia and hypercalciuria, full details of which will be available in the FAO/WHO report (J. Montez, WHO, personal communication, 2023).

Integration of interindividual variability in the modeling

As the model was only able to predict the mean group-level serum 25OHD response, the individual-level response was simulated by adding another layer to the model. The meta-regressive dose-response models on aggregate data described in the previous section provide predicted group-mean, 95% CI, and 95% PI values of serum 25OHD. None of these estimators are indicative of the serum 25OHD concentration achieved by an individual, as befitting the INL98 and UL definitions. In order to simulate individual-level responses, the interindividual variability was built into the model based on a method developed by the European Food Safety Authority (EFSA) (see Section 3.5.2.4 of the EFSA vitamin D UL opinion [27]).

The interindividual variability distribution of serum 25OHD was simulated based on the studies included in the meta-regression model as well as based on individual data collected on standardized serum 25OHD concentrations in young children [28]. The interindividual distribution was considered to be left-truncated normal (minimum 0) [28], with a coefficient of variation (CV) of 0.34 (weighted mean CV of the studies included) and 0.10 right-skewness [28] (Figure 1D) under the assumption that the shape and skewness observed in the individual data study was representative of the studies in the meta-regression. For each concentration of vitamin D intake between 1 and 60 μg/d, 100,000 random samples of individual serum 25OHD responses were generated using the Markov Chain Monte Carlo algorithm [29], with the first 10,000 simulations discarded (burn-in step). The mean of each interindividual distribution at each vitamin D intake level was set as the predicted mean 25OHD from the best-fitting meta-regression dose-response model for deriving the INL98 and the upper bound of the 95% CI for deriving the UL. Based on the simulated individual values, the percentage of individuals reaching a serum 25OHD of 28 nmol/L (for INL98) and 200 nmol/L (for UL) was calculated. Sensitivity analyses were performed assuming a non-skewed distribution, a CV of 0.40, as well as thresholds other than 28 and 200 nmol/L (i.e., 20, 25, 30, 35, 50, and 150, 180, 190, 210, 220, 250) and using as the mean of the interindividual distribution, the 95% CI and 95% PI bounds in addition to the predicted study mean.

Results

Characteristics of the studies

A total of 31 studies of children aged 2 wk to 3.9 y were included in the present modeling work [25,[30], [31], [32], [33], [34], [35], [36], [37], [38], [39], [40], [41], [42], [43], [44], [45], [46], [47], [48], [49], [50], [51], [52], [53], [54], [55], [56], [57], [58]], of which 26 were already identified in the original systematic review [16] and 5 from other guidelines or reviews (see Figure 2). The characteristics of the included studies are shown in Table 2. The mean age of the children within the studies was 6 mo, with most of the studies (N = 28) initiated in children aged <6 mo. The total duration of the trials ranged from 4 wk to 24 mo, with some trials providing intermediate measurements after a minimum of 2 wk since the start of the intervention. Total vitamin D intakes ranged from 0.6 to 57 μg/d (median of 11 μg/d; mean of 15 μg/d), including supplemental vitamin D ranging from 0 to 50 μg/d. The supplemental vitamin D forms used were vitamin D₃ in 20 studies, vitamin D₂ in 3 studies, both vitamers (i.e., vitamin D₂ than vitamin D₃) in 2 studies, and unspecified in 6 studies. Of the 31 RCTs, 29 used vitamin D supplements, and only 2 used either breast milk from mothers who were supplemented or vitamin D supplements plus or minus vitamin D-fortified infant formula. The included studies were conducted in North America (N = 11), Europe (N = 9), Asia (N = 8), and Australasia/Oceania (N = 3), with latitudes ranging from 61°N to 38°S. No studies were identified from Africa or South America. The measures in the studies were taken across multiple seasons for 47% of the data points, in winter for 8%, in spring for 8%, in summer for 6%, in autumn for 4%, and season was not reported for 26% of the data points. In terms of skin pigmentation, 45% of the studies were conducted in mixed skin types and 13% in light skin types, whereas data on skin pigmentation was unspecified in 42% (but were probably a majority of light or medium skin pigmentation). The methods of serum 25OHD measurement were enzyme immunoassay/chemiluminescence (10 studies), competitive protein-binding assays (9 studies), radioimmunoassays (6 studies), liquid chromatography with tandem mass spectrometry (3 studies), high-performance liquid chromatography (2 studies), and unspecified in one study. Only 6 of the 31 included trials participated in an external quality assessment scheme for serum 25OHD measurement.

FIGURE 2.

Study selection flowchart.

TABLE 2.

Overview of included randomized controlled trials contributing data for vitamin D intake requirement modeling

First author, y (reference)	Country	Mean age (range) at baseline	Intervention description (dose)1	Follow-up	Adherence/compliance2	Vitamin D intake estimation	25OHD assay	Sun exposure	Skin pigmentation
Aglipay et al., 2017 [30]	Canada	2.7 y (1–5 y)	Daily child vitamin D3 supplementation (10 or 50 μg/d)	4 mo	100% and 98%	Estimated from study in similar population [59]	Protein-binding assay	Low (43°N; trial during winter; 35–60 min unstructured free play outdoors per week at baseline)	Multiple (Fitzpatrick skin type 13% I, 31% II, 33% III, 11% IV, 6% V, 4% VI)
Ala-Houhala, 1985 [31]	Finland	Birth	Daily child vitamin D supplementation (10 or 25 μg/d)	8, 20 wk	NR	Estimated from breast milk concentration [60] and intake [14]	Competitive protein-binding assay	Variable (61°N; groups in winter and summer)	Light (largely fair skin color)
Ala-Houhala et al., 1986 [32]	Finland	Birth	Daily child vitamin D3 supplementation (0 or 10 μg/d) for 15 wk	8, 15 wk	NR	Estimated from breast milk concentration [60] and intake [14]	Competitive protein-binding assay	Low (61°N; recruited in January)	Light (largely fair skin color)
Alonso et al., 2011 [33]	Spain	1 mo	Daily child vitamin D3 supplementation (0 or 10 μg/d)	3, 6, 12 mo	Excluded noncompliant	Estimated from study in similar population [61]	EIA/Chemiluminescence	Variable (43°N; recruited over 1 y; excluded children with sunlight exclusion)	Light/medium (excluded dark skin pigmentation)
Atas et al., 2013 [34]	Turkey	15 d	Daily child vitamin D3 supplementation (5 or 10 μg/d)	4 mo	Excluded infants lost to follow-up and improper vitamin D supplementation	Estimated from breast milk concentration [13] and intake [14]	HPLC	Variable (40.6°N; recruited over 1 y)	Probably medium (Middle East)
Chan, 1982 [35]	United States	2 wk	Human milk with daily maternal supplementation (vitamin D 10 μg/d and calcium 250 mg/d) or human milk with daily child vitamin D supplementation (10 μg/d) or vitamin D-fortified formula (vitamin D 10 μg/L, calcium 0.51 mg/dL and phosphorus 0.39 mg/dL))	2, 4, 6 mo	NR	Reported in the study and combined with estimates from breast milk concentration [13] and intake [14]	Competitive protein-binding assay	Probably low (40.8°N; no seasonal variation was found in the study)	Light (Caucasian)
Chandy et al., 2016 [36]	India	2–4 d	Daily child vitamin D3 supplementation (0 or 10 μg/d)	3.5 mo	94%	Estimated from breast milk concentration [13] and intake [14]	RIA kits	Probably significant (26°N; mothers were instructed to give infant massage under the sun 15 min/d)	Probably medium (India)
Enlund-Cerullo et al., 2019 [37]	Finland	2 wk	Daily child vitamin D3 supplementation (10 or 30 μg/d)	12, 24 mo	89% and 87%	Estimated from study in similar population [62]	EIA/Chemiluminescence3	Probably variable (60.1°N; recruited over several times of the year)	Probably light (mothers of Northern European origin)
Gallo et al., 2013a [38]	Canada	1 mo	Daily child vitamin D2 or D3 supplementation (10 μg/d)	3 mo	89%	Estimated from study in similar population [39]	LC-MS/MS3,4	Probably variable (45.5°N; recruited over >1 y; 58% infants born during vitamin D-synthesizing period April-October)	Multiple (67% self-identified as White, skin color based on individual topical angle: 10% very fair, 46% fair, 35% medium, 6% olive, 4% dark)
Gallo et al., 2013b [39]	Canada	1 mo	Daily child vitamin D3 supplementation (10, 20, 30, or 40 μg/d) for 12 mo	1, 2, 5, 8, 11 mo	84–93%	Reported in the study	LC-MS/MS3,4	Probably variable (45.5°N; recruited over >1 y; 60% infants born during vitamin D-synthesizing period April-October; sun exposure did not differ between groups, but infant sun index increased from 7 at 1 mo to 71 at 9 mo old)	Multiple (84% White)
Gordon et al., 2008 [25]	United States	10 mo (8–24 mo)	Daily child vitamin D2 or D3 supplementation (vitamin D2 50 μg/d or vitamin D3 50 μg/d, both groups received calcium 50 mg/kg/d)	6 wk	NR	Estimated from study in similar population [63]	EIA/Chemiluminescence	Probably variable (42°N; recruited over the year)	Multiple (skin pigmentation 1 (heaviest) 62%, 2 27%, 3 4%, 4 (lightest) 8%)
Grant et al., 2014 [40]	New Zealand	Birth	Daily child vitamin D3 supplementation (placebo 0 μg/d) for 6 mo	2, 4, 6 mo	78–90%	Estimated from study in similar population [39]	LC-MS/MS3	Probably variable (36°S; recruited at all times of the year; mean time spent outdoors 0.21 h/d at 2 mo, 0.25 h/d at 4 mo, and 0.40 h/d at 6 mo)	Multiple (Mother 38% European, 24% Maori, 46% Pacific, 25% Other)
Greer et al., 1982 [41]	United States	3 wk	Daily child vitamin D supplementation (0 or 10 μg/d)	3, 9, 23 wk	80%	Estimated from breast milk concentration [13] and intake [14]	Competitive protein-binding assay	Probably variable (43°N; unclear season; sunshine exposure 35 min/d)	Light/medium (94% Caucasian, 6% Asian-Indian)
Hollis et al., 2015 [42]	United States	5 wk (4–6 wk)	Daily child and maternal vitamin D3 supplementation (10/10 μg/d)	3, 6 mo	NR	Estimated from breast milk concentration [13] and intake [14]	RIA kits	Probably variable (38°N; recruited over different times of the year)	Multiple (59% White, 22% Hispanic, 19% Black/African American
Holst-Gemeiner et al., 1978 [43]	Austria	1 wk (2–10 d)	Daily child vitamin D3 supplementation (30 μg/d)	2, 4–6 wk	NR	Estimated from breast milk concentration [13] and intake [14]	RIA	Probably low (48°N; newborns)	Probably light (Western Europe)
Huynh et al., 2017 [44]	Australia	Birth	Daily child vitamin D3 supplementation (10 μg/d)	3–4 mo	69%	Estimated from breast milk concentration [13] and intake [14]	EIA/Chemiluminescence	Probably low (38°S; considered minimal by authors)	Multiple (Maternal skin pigmentation 50% light-olive, 50% dark)
Kunz et al., 1982 [45]	Germany	Birth	Daily child vitamin D3 supplementation (12.5 or 25 μg/d)	6 wk	NR	Estimated from breast milk concentration [13] and intake [14]	Protein-binding assay	Probably low (48°N; season NR; newborns)	Probably light (Western Europe)
Madar et al., 2009 [46]	Norway	6 wk	Daily child vitamin D2 supplementation (0/usual care or 10 μg/d)	7 wk	91%	Estimated from breast milk concentration [13] and intake [14]	HPLC-APCI-MS3	Probably low (60°N; all seasons, no differences in 25OHD found between seasons)	Medium/dark (Pakistani, Turkish, or Somali)
Pehlivan et al., 2003 [47]	Turkey	2 wk	Daily child vitamin D supplementation (10 or 20 μg/d)	4 mo	NR	Estimated from the Global Dietary Database	EIA/Chemiluminescence	Probably low (40.8°N; according to authors, sunlight exposure is low due to dressing habits, low vitamin D dietary intake, and air pollution; time of year not mentioned, except for the control group; maternal vitamin D intake and dressing habits were correlated with 25OHD, correlations for infant 25OHD were NR)	Probably medium (Middle East)
Pittard et al., 1991 [48]	United States	Birth	Daily child vitamin D supplementation (10 or 20 μg/d)	2, 4, 6, 8, 10, 14, 16 wk	NR	Reported in the study	Competitive protein-binding assay	Probably variable (32.8°N; time of year not mentioned)	Multiple (20% White, 80% Black)
Ponnapakkam et al., 2010 [49]	United States	Birth	Daily child vitamin D3 supplementation (0 or 5 μg/d from birth or starting at 2 mo)	2, 4, 6 mo	82%	Estimated from study in similar population [63]	EIA/Chemiluminescence	Variable (30°N; across several times of the year; differences in skin color and clothing were equally distributed between groups at randomization)	Multiple (dark skin color was distributed evenly between groups at randomization)
Rueter et al., 2019 [50]	Australia	<28 d	Daily child vitamin D3 supplementation (0 or 10 μg/d)	3, 6 mo	NR	Estimated from studies in similar populations [39]	EIA/Chemiluminescence	Variable (32°S; recruitment across multiple seasons, no differences between seasons found; UV light exposure was measured in 42% of infants, was 1204 J/m2 in the vitamin D group and 815 J/m2 in the control group, was not correlated with 25OHD or season of birth)	NR
Shakiba et al., 2010 [51]	Iran	Birth	Daily child vitamin D3 supplementation (5 or 10 μg/d)	6 mo	NR	Estimated from the Global Dietary Database	EIA/Chemiluminescence	Probably variable (32°N; January-September)	Probably medium (Middle East)
Siafarikas et al., 2011 [52]	Germany	4–5 d	Daily child vitamin D3 supplementation (6.25 or 12.5 μg/d)	6 wk	NR	Estimated from breast milk concentration [13] and intake [14]	RIA kits3	Low (52.5°N; recruited during summer and winter equally; absolute UV-B exposure measured 2.5–20 J/m2)	Light (included only photo-types I and II according to Fitzpatrick and Bolognia)
Singh et al., 2018 [53]	India	Birth	Daily child vitamin D3 supplementation (0 or 10 μg/d)	6 mo	NR	Estimated from breast milk concentration [13] and intake [14]	EIA/Chemiluminescence	Probably variable (29°N; January-September)	Probably medium (Southeast Asia)
Specker et al., 1992 [54]	China	Birth	Daily child vitamin D supplementation (2.5, 5, or 10 μg/d)	6 mo	96–131%	Estimated from the Global Dietary Database	EIA/Chemiluminescence	Variable (22, 30, 40, 47°N; enrolled during fall and spring)	Probably medium (North and South China)
Vervel et al., 1997 (Study 1) [55]	France	1 mo	Daily child vitamin D2 supplementation (25 μg/d) with vitamin D-fortified or nonfortified formula	1.5–2, 2.5–4 mo	NR	Reported in the study and combined with estimates from breast milk concentration [13] and intake [14]	Competitive protein-binding assay	Probably variable (49°N; measures at different times of the year)	NR
Vervel et al., 1997 (Study 2) [55]	France	Birth	Daily child vitamin D2 supplementation (12.5 or 25 μg/d) from mothers supplemented during pregnancy (0 or 12.5 μg/d)	3 mo	NR	Reported in the study and combined with estimates from breast milk concentration [13] and intake [14]	Competitive protein-binding assay	Probably variable (49°N; recruited April-July)	NR
Wagner et al., 2006 [56]	United States	1 mo	Daily child and maternal vitamin D3 supplementation (7.5/10 μg/d)	4, 7 mo	≥61% and ≥80%	Estimated from breast milk concentration [13] and intake [14]	RIA	Probably low (33°N; mothers were instructed to avoid direct sunlight exposure of their infants during the first 6 mo)	Multiple (maternal ethnicity 11% African American, 74% White, 15% Hispanic)
Zhou et al., 2018 [57]	China	7.8 mo (3–12 mo)	Daily child vitamin D3 supplementation (10 or 30 μg/d)	2, 4 mo	Excluded noncompliant	Estimated from the Global Dietary Database	NR	Variable (29°N; recruited over multiple seasons)	Probably medium (China)
Ziegler et al., 2014 [58]	United States	24–32 d	Daily child vitamin D3 supplementation (5, 10, 15, or 20 μg/d)	1, 3, 4.5, 6.5, 8, 11 mo	103.40%	Estimated from breast milk concentration [13] and intake [14] and food intake from study in similar population [63]	RIA kits	Low (41°N; main assessment during winter, minimal sun exposure)	Multiple (90% White, 4% Hispanic, 3% African American, 2% American Indian, 1% Asian)

Abbreviations: EIA, enzyme immunoassay; HPLC, high-performance liquid chromatography; LC-MS/MS, liquid chromatography with tandem mass spectrometry; NR, not reported; RIA, radioimmunoassay; UV, ultraviolet; 25OHD, 25-hydroxy-vitamin D; HPLC-APCI-MS, high-performance liquid chromatography atmospheric pressure chemical ionization mass spectrometry; UV-B, ultraviolet B.

¹The study arms that did not correspond to the inclusion criteria were excluded from the data analyses (i.e., weekly, monthly, single-dose vitamin D supplementation, impossible to isolate effect of vitamin D or maternal vitamin D supplementation <12.5 μg/d).

²Expressed as a percentage of the dose taken, unless stated otherwise.

³Participated in an external quality assessment scheme for serum 25OHD measurement.

⁴Also measured serum 25OHD with immunoassay.

The detailed risk of bias assessments and strength of evidence assessed by grades of recommendation, assessment, development, and evaluation are shown in Supplementary Figure 1 and Supplementary Tables 1 and 2. The overall risk of bias was low in 6 studies, some concerns in 12 studies, and high in 13 studies. The risk of bias was most often high due to deviations from the intended intervention, e.g., participants and/or personnel were aware of the intervention received (15 studies) or because of inappropriate analysis used to estimate the effect of assignment to intervention (15 studies). The overall strength of evidence was considered low due to the risk of bias in the included studies, the paucity of standardized 25OHD measurements and the scarcity of studies in darker skin individuals, as well as the high heterogeneity between the studies, which covariates (such as latitude, season, and skin pigmentation) could not explain significantly.

Unadjusted multilevel meta-regression modeling: the total vitamin D intake – serum 25OHD dose-response relationship

The best fit (i.e., the lowest AIC) was obtained with the cubic model. However, the cubic term was not significant. The second best-fitting model was the quadratic model, which was selected for further analyses also because of its biological plausibility (the increase of 25OHD by vitamin D unit dose is larger at low intakes of vitamin D and lower at higher intake levels – see Figure 3). The log model showed the highest AIC. An overview of the models tested and their results are shown in Supplementary Table 3.

FIGURE 3.

Relationship between total vitamin D intake (μg/d) and serum 25OHD (nmol/L) on 0–3.9 y old children fitted with unadjusted quadratic multilevel meta-regression. Black round dots represent the observed study arm means (N = 186 data points). The blue line represents the mean response, the light blue fill represents the 95% confidence interval, and the light grey fill represents the 95% prediction interval. Abbreviation: 25OHD, 25-hydroxyvitamin D.

The possibility of fitting different models separately from the overall age group of 0–3.9 y into infants and young children’s age was also explored. However, age was not significant when included as a continuous variable in the model. In addition, when meta-regressions were conducted for different age categories individually (i.e., <6 mo, 6–11.9 mo, 0–11.9 mo, and 0.5–3.9 y), model parameter estimates were very similar within the vitamin D intake range of 10–45 μg/d. At the lowest and highest intakes, the age categories models diverged due to the lack of data points (see Supplementary Figure 2).

The main quadratic model on 0–3.9 y old children was also analyzed with data from 10 additional studies of children aged 4–9 y [59,[64], [65], [66], [67], [68], [69], [70], [71], [72]] to investigate whether adding supplementary evidence could improve the model at the potential expense of increasing the uncertainty in terms of reflecting the true relationship in the target population (0–3.9 y children). However, the inclusion of these additional studies did not significantly change the shape of the model. The final model selected was the quadratic unadjusted model for children aged 0–3.9 y, shown in Figure 3.

The inclusion of different covariates and their combinations (infant age, baseline 25OHD, region, country income category, 25OHD assay, season, skin pigmentation, and latitude) did not improve the model fit significantly or explain a significant part of the heterogeneity.

Interindividual variability component: the full integrated model for INL98 and UL

The predicted percentage of young children reaching the serum 25OHD thresholds of 28 nmol/L and 200 nmol/L, associated with INL98 and UL, respectively, at selected vitamin D intakes are shown in Table 3 and Figure 4. The predicted percentage of individuals achieving the INL98-associated serum 25OHD threshold of 28 nmol/L ranged from 97.3% at 10 μg/d vitamin D intake to 99.1% at 60 μg/d. The predicted percentage of individuals exceeding the UL-associated serum 25OHD threshold of 200 nmol/L ranged from 0% at 10 μg/d vitamin D intake to 3.7% at 60 μg/d.

TABLE 3.

Predicted percentage of individuals (%) reaching the serum 25-hydroxyvitamin D (25OHD) thresholds 28 and 200 nmol/L respectively (used to derive individual nutrient level 98 (INL98) and upper level (UL) respectively). Modeling with left-truncated normal distribution, right-skewed (0.10), and CV of 0.34, and predicted mean response and predicted upper bound of 95% confidence interval mean response as the mean value of the interindividual distribution respectively for INL98 and UL.

Vitamin D intake (μg/d)	Percentage individuals reaching serum 25OHD threshold of 28 nmol/L (used to set the INL98)	Percentage individuals reaching serum 25OHD threshold of 200 nmol/L (used to set the UL)
10	97.30	0.00
15	97.88	0.02
20	98.35	0.09
25	98.57	0.35
30	98.71	0.79
35	98.84	1.41
40	98.98	1.96
45	98.96	2.61
50	99.08	3.19
55	99.05	3.43
60	99.07	3.65

Abbreviations: CV, coefficient of variation; INL98, individual nutrient level 98; UL, upper limit; 25OHD, 25-hydroxyvitamin D.

FIGURE 4.

Interindividual variability distribution at vitamin D intake of 10 and 35 μg/d. Both interindividual distributions were simulated with a left-truncated normal, CV of 0.34, and 0.10 right-skewness. The interindividual distribution at vitamin D intake of 10 μg/d, simulated using the mean predicted response from the unadjusted quadratic multilevel meta-regression (blue distribution), illustrates that nearly 98% of the individuals would achieve a serum 25OHD of 28 nmol/L, providing a basis for setting the INL98. The interindividual distribution at vitamin D intake of 35 μg/d, simulated with the upper bound 95% confidence interval of the predicted response from the unadjusted quadratic multilevel meta-regression (red distribution), illustrates that <2% of the individuals would achieve a serum 25OHD of ≥200 nmol/L, providing a basis for setting the UL. The dotted blue line represents the lower threshold of 28 nmol/L used to derive the INL98. The dotted red line represents the upper threshold of 200 nmol/L used to derive the UL. Black round dots represent the observed study arm means. The solid black line represents the predicted mean response, the dark grey fill represents the 95% confidence interval, and the light grey fill represents the 95% prediction interval. Abbreviations: CV, coefficient of variation; INL98, individual nutrient level 98; UL, upper level; 25OHD, 25-hydroxyvitamin D.

The findings of sensitivity analyses are shown in Supplementary Tables 4–6. The differences with the main analysis are limited except for the cases when using the 95% PI bounds as a mean of the interindividual distribution. However, this approach was considered over-conservative and not reliable for the setting vitamin D INL98 and UL, and the model results overall were considered robust.

Discussion

This paper presents the results from a novel multilevel and multivariable modeling of the response of serum 25OHD to total vitamin D intake in children aged <4 y, including an interindividual variability component. Our findings suggest that a vitamin D intake of 10 μg/d is required to maintain serum 25OHD concentrations in the vast majority (97.3%) of the young children >28 nmol/L (i.e., a threshold associated with minimized risk of rickets), corresponding to an INL98. From a safety perspective, the present analyses suggest that vitamin D intakes <35 μg/d would keep serum 25OHD concentrations in almost all young children (98.6%) <200 nmol/L as the upper threshold associated with the UL for this age group.

The vitamin D requirement estimates arising from this work differ partly from previous recommendations (see Supplementary Table 7) due to differences in the body of evidence used, the thresholds selected, the analyses conducted, and the type of recommendations derived. Compared with the 2004 FAO-WHO vitamin D recommendation (5 μg/d) for infants and young children [1], these new estimates, arising from the current modeling for the FAO-WHO update exercise, represent a more data-driven derivation of the vitamin D dietary requirement. Although the serum 25OHD targets in the present analyses (of 28 nmol/L) and that of the 2004 recommendations (27 nmol/L) were extremely close, the former stem from a systematic review and individual participant data meta-analysis [15], whereas the latter was based on the prevailing view of the level necessary to ensure normal bone health as well as being the lower limit of the normal range [1]. The present analyses used meta-regressive modeling to relate vitamin D intake to serum 25OHD, which also included an interindividual variability component allowing for the estimation of the intake required to maintain serum 25OHD concentration of >28 nmol/L in 97.3% of young children. In contrast, the 2004 WHO recommendations relied on a more simplified approach that involved the estimation of the mean group dietary intake of vitamin D required to maintain the plasma 25OHD concentrations of >27 nmol/L [1]. In this method, the dietary intake of vitamin D for each population group was rounded to the nearest 50 IU (1.25 μg) and then doubled to cover the needs of all individuals within that group, irrespective of sunlight exposure. Notably, in the case of infants and young children, the mean intakes were based on only a few studies overall. The 2004 FAO-WHO report on nutrient requirements did not establish a UL for vitamin D but noted that the adverse effects of high vitamin D intakes - hypercalciuria and hypercalcemia - did not occur at the recommended intake levels proposed in the report [1].

The present INL98 vitamin D estimate cannot be directly compared with the international vitamin D reference values from Institute of Medicine (IOM) for North America [5] or EFSA for Europe [7]. The modeling approach underpinning all 3 sets of vitamin D requirement estimates differed in various aspects, especially in relation to serum 25OHD thresholds (28 compared with 40 and/or 50 nmol/L), use or nonuse of covariates, and with respect to eligible RCT data - use or nonuse of restrictions in relation to latitude (>40 or 49.5^oN), winter-time only RCTs, ethnicity of RCT participants, among other differences [5,7]. The present INL98 vitamin D estimates for children aged 0–3.9 y stemmed from the multilevel and multivariable modeling, which included interindividual variability simulations, whereas the estimates from IOM and EFSA’s modeling were restricted to children aged ≥1 y; for infants, both agencies set their vitamin D recommendations based on two vitamin D supplementation trials in breastfed babies [5,7]. In addition, although the derivation of the UL for vitamin D was not based on meta-regression modeling in the case of the IOM reference values, it was in the case of EFSA’s UL for infants aged ≤1 y. EFSA fitted a meta-regression dose-response model and adjusted for baseline 25OHD, integrating an interindividual variability component, to predict the percentage of infants with serum 25OHD concentration of >200 nmol/L at different vitamin D intakes to establish a UL [27]. The method used by EFSA to add this interindividual variability component in the model was taken and adapted for the present analysis, not only to derive a UL but also an INL98.

The present work had a number of weaknesses. Firstly, many of the included studies had evidence of high bias, and the certainty of the evidence was considered low. Evidence of high bias among the collection of RCTs used in vitamin D requirement derivation was also noted by EFSA in their exercise [7]. Second, vitamin D intake from the general diet, which was added to vitamin D provided by the supplements or fortified foods to calculate the total vitamin D intake, had to be imputed from other sources for several studies. Although this need to impute data on dietary intake was a limitation, it was outweighed by the benefit of accounting for dietary supply of vitamin D from background diet to the estimate of total vitamin D intake. Third, the analysis did not have estimates of vitamin D cutaneous synthesis and relied instead on indirect measures of potential UV-B availability, such as latitude and season. However, even using these 2 proxies of cutaneous vitamin D synthesis did not provide major additional insight into the role of sunlight exposure when setting vitamin D intake requirements, similar to the experience of EFSA [7]. Although skin pigmentation is also an important factor that can affect cutaneous vitamin D synthesis, this was explored but could not be informatively included in the score calculations because 42% of studies did not report (but were probably a majority of light or medium skin pigmentation), 45% were reported as mixed, 13% reported as light skin type, and none reported dark skin type only. The majority of the studies were conducted in countries where the predominant racial group is White. This is an inherent limitation of the data rather than of the analysis. Nevertheless, this limitation should be a consideration as agencies make local context adjustments to these new estimates. In this regard, one cautious interpretation of the present vitamin D intake estimates is that they are most protective of those young children not synthesizing vitamin D in the skin. In addition, this analysis did not include premature and low birth weight infants, which can represent a significant portion of the infant population and which should be considered when interpreting our findings. Moreover, method-related differences in the measurements of serum 25OHD [73] were likely to have contributed additional variability to the modeling of the vitamin D intake and serum 25OHD dose-response in the present work, as it has for other vitamin D recommendations from competent authorities [[5], [6], [7], [8], [9]]. Standardization protocols exist to harmonize existing serum 25OHD data, but these are for observational-type studies [74], and for RCTs, it would mean reanalysis of serum 25OHD samples using certified liquid chromatography with tandem mass spectrometry method [75], which was beyond the scope of the present exercise. However, the data used to inform the interindividual variability within the modeling was based on standardized 25OHD data. These data, nonetheless, was still a surrogate for empirical variability data from the 31 included RCTs, which were not available. In the absence of the availability of individual data from the RCTs, the interindividual variability distribution shape and skewness were derived from one study and applied to the entire range of 25OHD predicted mean and corresponding vitamin intake. Although this approach represents a limitation because not based on real distributions observed in the studies, to our knowledge, no better methods are available at the moment, and frequently, the issue of interindividual variability is ignored. Another weakness is that the data were extracted by a single reviewer and not two independent reviewers; however, the risk of errors was minimized by the thorough verification by a second reviewer. Lastly, although the literature search covered the period from inception to June 2020, it will have missed additional studies, which would likely be eligible for inclusion in the modeling [76,77]. This was outside the control of the present authors, as allocated resources within the exercise were such that the present work began some after the original review [16] was completed. Furthermore, the collection of eligible RCTs in the present work is a major advancement over that collected in previous vitamin D intake requirement exercises.

The present work also has some important strengths, which includes the use of data coming from various steps within the overall risk assessment framework, which either framed or facilitated the present modeling. This included, for example, evidence around the robustness of using 25OHD, the definition of serum 25OHD threshold for the minimization of nutritional rickets in young children [15], and key new exposure data from two systematic reviews on breast milk vitamin D content and breast milk intake volume [13,14]. The vitamin D RCT data was also identified to a large extent from an independently commissioned systematic review, which was further refined in terms of use for the present modeling. The modeling used in the present work was comprehensive and goes beyond that of previous vitamin D requirement exercises, especially by including evidence from the entire globe and by its incorporation of an interindividual variability component.

This review also highlights a number of key research gaps that should be addressed going forward, more precisely, the lack of published data from Africa and South America, the limited data available for children aged 1–3.9 y compared with ≤1 y of age, and further investigation of the role of ethnicity, sun exposure, as well as prematurity and low birth weight on dietary vitamin D intake estimates.

In conclusion, the present analysis provided new global estimates of vitamin D intake requirements (INL98 and UL) for children aged <4 y. These new estimates can be used by countries across the globe once appropriate, local context adjustments (such as contribution to vitamin D status from sun exposure) are made to suit the intended population.

Acknowledgments

We thank the FAO/WHO expert group on nutrient requirements for children aged 0–3 y for their valuable inputs, the team from Tufts University, led by Mei Chung, who conducted the original systematic review, for sharing their data, and Jason Montez, Scientist at WHO, and Maria Xipsiti, Nutrition Officer at FAO, for supporting this work.

Author contributions

The authors’ responsibilities were as follows – KDC, MRL: screened the studies; MRL: collected and extracted the data, and KDC verified them; LM, MRL: conducted the analyses; KDC, LM, MRL: wrote the manuscript; and all authors: read and approved the final manuscript.

Conflict of interest

MRL received financial support from the WHO to conduct this work as an independent consultant. LM was an independent consultant on statistical issues with no financial support from WHO. The consultancy was performed in a personal capacity. LM is employed with the European Food Safety Authority (EFSA). However, the positions and opinions presented in this paper are those of the author alone and do not represent the views of EFSA. KDC was part of the FAO/WHO expert group. The individuals in the FAO/WHO expert group were required to declare a lack of conflict of interest.

Funding

MRL received financial support from the WHO to conduct this work as an independent consultant. The European Food Safety Authority covered the open-access publication fee.

Data availability

The data underlying this article will be shared on reasonable request to the corresponding author.

Appendix A. Supplementary data

The following is the Supplementary data to this article: