1. Introduction
Using the threshold for vitamin D deficiency of a serum 25-hydroxyvitamin D (25(OH)D) concentration <30 nmol/L [
1] acknowledged as a primary risk factor for nutritional rickets in children [
2] and osteomalacia in adults [
3], a recent analysis from the ‘Food-based solutions for optimal vitamin D nutrition and health through the life cycle’ (ODIN) project, representing >55,000 individuals, reported that one in eight European residents has vitamin D deficiency [
4]. Many investigators in the vitamin D field concur with the Endocrine Society’s 25(OH)D threshold of <50 nmol/L to designate vitamin D deficiency for protection of skeletal and non-skeletal health [
5]. Thus, on the basis of the World Health Organization’s criteria [
6] and the 25(OH)D threshold (30 or 50 nmol/L) selected, vitamin D deficiency in Europe can be classified as a mild (5%–19.9%) or severe (>40%) public health problem. Thus, strategies for vitamin D deficiency prevention are urgently required [
7].
Dietary recommendations for vitamin D, hereafter referred to as dietary reference values (DRVs), have a key role in defining the parameters for such prevention strategies. In line with scientific developments, DRVs for vitamin D have been re-evaluated over the last six years [
7,
8]. Most agencies selected musculoskeletal outcomes as the health status indicators with a developed evidence base, particularly from randomized controlled trials (RCTs), on which to nominate serum 25(OH)D thresholds designating deficient/sufficient vitamin D status. Following analysis of vitamin D intake–serum 25(OH)D dose-response data, these thresholds provided the targets for estimating population and/or individual vitamin D intake recommendations [
1,
9,
10,
11,
12]. All of the agencies based their analyses for vitamin D requirements assuming minimal ultraviolet B (UVB) exposure, as sunlight synthesis would interfere with the vitamin D intake–25(OH)D dose-response relationship. However, despite considerable effort by five different agencies over a 6-year period, the published DRVs are diverse. Individual recommendations vary from 10 to 20 µg/day (400–800 IU), depending on the 25(OH)D target, ranging from 25 to 50 nmol/L.
In the UK, the Scientific Advisory Committee on Nutrition (SACN) proposed a ‘population protective’ serum 25(OH)D concentration of 25 nmol/L, defined as the minimum threshold that should be met or exceeded by almost everyone, and recommended a corresponding vitamin D intake of 10 µg/day for all persons >1 year of age [
9]. Four other agencies based their individual recommendations on achieving a serum 25(OH)D threshold ≥50 nmol/L [
1,
10,
11,
12], with some important distinctions. The Institute of Medicine (IOM) [
1] and Nordic Council of Ministers (NORDEN) [
11] recommended vitamin D intakes of 15 and 10 µg/day, respectively, for children and adults, with increased allowances for older adults. These two agencies also proposed an Average Requirement (AR) intake of 10 and 7.5 µg/day, respectively, to meet the needs of 50% of the population. The AR value is not an individual recommendation but used by nutritional epidemiologists to evaluate the adequacy of vitamin D intakes in the population [
13]. The German Nutrition Society [
12] and, more recently, the European Food Safety Authority (EFSA) [
10] opted for Adequate Intake (AI) values for vitamin D of 20 and 15 µg/day, respectively, with the 50 nmol/L threshold in mind. The AI is reserved for use only in cases of much uncertainty in the data.
While the serum 25(OH)D threshold selected will influence dietary requirement estimates for vitamin D, it is only one contributor to the disparity in recently published DRVs. An unintended but serious cause of this disparity is the method used to perform the vitamin D–25(OH)D dose-response analysis: use of a standard meta-regression approach based on aggregate data from several RCTs [
1,
10,
11] versus use of individual data from distinct RCTs [
9,
12]. Apart from the impact of analytical differences in 25(OH)D values in different RCTs [
14], there are pros and cons to both approaches. The meta-regression of aggregate data, which uses information averaged across all individuals within treatment groups in a RCT, can allow for between-study variability, but it cannot incorporate between-participant variability, which is crucial for estimating individual recommendations. A meta-regression based on analysis of individual participant data (IPD), in which the raw data for each RCT are used for synthesis, has many potential advantages, both statistically and clinically, over meta-regression of aggregate RCT data [
15]. Modelling strategies for pooling of individual subject data from cognate dose-response RCTs are likely to provide appropriate estimates of DRVs to meet specified 25(OH)D thresholds, as use of individual data permits estimation of requirement values that cover 97.5% of the population group being considered. A recent Cochrane paper has highlighted that IPD analyses are not only now described as the gold standard, but that their use by guideline developers could lead to improved guidelines, ensuring that routine patient care is based on the most reliable evidence available [
16].
Our hypothesis is that the method used to perform the dose-response vitamin D-25(OH)D analysis, on which DRVs are based, has a profound effect on the recommendation issued, regardless of the serum 25(OH)D target selected or other sources of heterogeneity. Thus, the objectives of this study were firstly to perform an IPD meta-regression using individual subject data (
n = 882) from seven selected winter-based RCTs of the vitamin D intake–serum 25(OH)D dose-response, where raw data were available to the authors [
17,
18,
19,
20,
21,
22,
23], in order to establish recommendations for vitamin D. Secondly, we wished to contrast these IPD-derived results against results from a standard meta-regression based on aggregate data derived from the same RCTs.
3. Results
A collection of seven RCTs, where raw data (
n = 882 individuals) were available to the authors, was included in the present analysis. These RCTs were conducted in: 4–8 year-old children [
17], 11 year-old girls [
18], 14–18 year-old adolescents [
19], adults aged 20–40 years [
20], 50+ years [
21,
22], and 65+ years [
23], and were all implemented using the same study design, analytical platform for serum 25(OH)D, and dietary assessment method. Most of these RCTs were among the 44 used collectively in the IOM, NORDEN, and EFSA exercises for deriving DRVs [
1,
10,
11]. These seven RCTs all fulfill or exceed the previously defined RCT selection criteria established by the IOM as part of their process [
1] (see
Appendix A), and in fact five of the seven [
18,
20,
21,
22,
23] have been used in the recent DRV meta-regression exercises by IOM, NORDEN, and EFSA [
1,
10,
11]; the two most recent RCTs [
17,
19] were published since these. A brief description of each of the studies, including the sex and age of participants, the doses of vitamin D
3, the published DRV estimate to prevent vitamin D deficiency (serum 25(OH)D < 25 nmol/L) in 97.5% of that population, and the trial registry numbers is provided in
Table 1.
3.1. Vitamin D Requirement Estimates Based on the Two-Step IPD Meta-Regression Analyses
Based on the two-step IPD meta-regression model, the vitamin D intake estimates required to maintain serum 25(OH)D concentrations above the four serum 25(OH)D thresholds used by all of the regulatory agencies are shown in
Table 2. Recommended Intakes, defined as the intake estimated to meet the requirement of 97.5% of the population for a specific 25(OH)D threshold, were as follows:
Average Requirements (AR), defined as the intake estimated to meet the requirement of 50% of the population for a specific threshold, were as follows:
Using the IOM 25(OH)D cut-off for the AR of 40 nmol/L [
1], we estimated that the vitamin D intake required to maintain serum 25(OH)D concentrations ≥40 nmol/L in 50% of the subjects was 4.5 µg/day (
Table 2).
The AR meeting the NORDEN [
11] serum 25(OH)D threshold of ≥50 nmol/L was 10.9 µg/day (
Table 2).
Table 2 shows that moving from the 90th percentile through the 95th to the 97.5th percentile of requirement increased the intake estimate dramatically; this is expected to meet the needs of ‘nearly all’ healthy individuals in the population (i.e., 97.5%). The equivalent vitamin D intake requirement estimates based on a regression model unadjusted for age and baseline serum 25(OH)D are also shown in
Table 2, and for the most part were of the order of ~1–3 µg/day higher.
Using a one-stage IPD approach, the Recommended Intake estimates (95% CI) were very close to those from the two-step IPD analyses: 10.8 (9.7, 11.8) and 27.0 (25.6, 28.5) µg/day using serum 25(OH)D thresholds of ≥25 and ≥50 nmol/L, respectively. Likewise, the AR estimates (95% CI) were similar to those from the two-step IPD analyses: 3.6 (2.7, 4.4) and 10.1 (9.5, 10.7) µg/day using serum 25(OH)D thresholds of ≥40 and ≥50 nmol/L, respectively. The similarity in estimates between the one-step and two-step approaches can be explained by the closeness of fitted regression lines for both (see
Figure 1A), even though the two-step approach yielded a slightly narrower prediction band.
3.2. Outcomes of the Sensitivity Analyses for the Two-Step IPD
BMI: Inclusion of BMI as an additional covariate in the adult RCT dataset produced estimates to maintain 97.5% of individuals with serum 25(OH)D >30 and >50 nmol/L of 15.0 and 28.4 µg/day, respectively, compared to 15.0 and 28.7 µg/day, respectively, in the dataset unadjusted for BMI.
Age: The repeated two-step IPD meta-regression analysis in the four RCTs on adults only provided estimates to maintain 97.5% of individuals with serum 25(OH)D >30 and >50 nmol/L of 15.0 and 28.7 µg/day, respectively, compared to 13.1 and 26.1 µg/day, respectively, in the full data set.
3.3. Comparison with Vitamin D Requirement Estimates from Standard Meta-Regression Analyses Based on Aggregate Data
In order to fulfill our secondary objective, which was to perform standard meta-regression analyses based on an aggregate from the same RCTs used in the two-step IPD analysis, we extracted the data on the basis of each RCT arm. This yielded 23 separate RCT summary data points. The relation between achieved serum 25(OH)D concentrations and the total vitamin D intake (diet and supplemental) in the 23 arms of the seven RCTs using the standard meta-regression approach is shown in
Figure 1B. In the standard meta-regression model of aggregate data from the seven RCTs (23 arms), adjusted for age and baseline serum 25(OH)D, the intake estimate (95% CI) required to maintain 97.5% of individuals ≥50 nmol/L serum 25(OH)D was 14.2 (10.1, 18.9) µg/day. The unadjusted model yielded a slightly higher estimate of 15.8 (11.0, 16.9) µg/day. The discrepancy in estimates arising from the IPD-based and standard meta-regression model of aggregate data can be summarized as a more biased regression line with a considerably narrower accompanying prediction band in the latter, whether based on unadjusted or adjusted analyses (
Figure 1B).
While the best fit model in the present analyses was found to be a linear fit, which is in keeping when a plateauing effect at vitamin D intakes above 35 µg/day (1), some of the agencies have used a curvilinear (natural logarithmic (Ln)) model [
1,
10,
11]. For the purposes of comparison, we also applied a curvilinear (Ln) model to the same seven RCT data set as used for the linear model. Using this model, the intake requirement estimates to maintain 97.5% of individuals ≥50 nmol/L serum 25(OH)D were 12.5 μg/day and 13.2 μg/day for adjusted and unadjusted models, respectively.
4. Discussion
DRVs for vitamin D established over the last 6 years are highly variable. This is a major concern in light of their crucial role in providing a framework for the prevention of vitamin D deficiency in the population and evaluating progress towards this goal, as well as providing guidance for clinical practice and for individuals. This study set out firstly to establish a proof of concept that an IPD meta-regression using individual subject data from selected winter-time RCTs of the vitamin D intake–serum 25(OH)D dose-response is a superior approach in the determination of recommendations for vitamin D, both at a population average and individual level, as it avoids some of the limitations intrinsic to standard meta-regression, based on aggregate data, as used in recent DRV estimates [
1,
10,
11]. Secondly, we aimed to provide experimentally-derived DRV estimates that would achieve the range of serum 25(OH)D target values selected by the different international agencies. We conducted our analysis in a pooled sample of 882 participants, ranging in age from 4 years upwards, from seven dose-response vitamin D intervention studies, all implemented using the same design, during wintertime at locations >50° N and with data for the axes for the dose-response analysis (i.e., 25(OH)D concentrations and vitamin D intake produced using certified and validated methods, respectively). Using this data set, the present analysis clearly illustrates that the vitamin D DRV estimates arising from an IPD approach, increasingly recognized as best practice [
15,
16], were considerably higher than those derived from the standard meta-regression approach based on aggregate data. This was strikingly evident in the large gap in estimates for recommended intakes derived by the two-step IPD analyses (~26 µg/day) and by standard meta-regression (~14 µg/day) to maintain serum 25(OH)D ≥50 nmol/L in 97.5% of individuals, favored by most agencies [
1,
10,
11,
12].
The disparity arises, to a large extent, from the inability of the standard meta-regression approach (even in adjusted models) to recover between-individual variability from summary statistics (i.e., mean and standard error per arm), which is a critical short-coming in the context of deriving DRVs. The large inter-individual variation that exists in the response of serum 25(OH)D to any particular intake of vitamin D underlies the wide prediction intervals used to estimate the DRV intake estimates at the 97.5th percentile. At best, between-study variation may be recovered using such an approach [
15], and we have shown in the present analysis that between-study variation does not account for between-individual variation. This is not unexpected, as they arise from two completely different sources of variation. It has been recently suggested that failure to assimilate information from IPD-based approaches may lead to limited recommendations that are inappropriate for population health. On the other hand, the opportunity afforded by their uptake in preference to standard aggregate approaches may better inform guidelines [
16]. This study, by clearly demonstrating the significant advantage of using an IPD approach, adds further evidence to support its application to the process of deriving DRVs.
There are wide differences between our DRV estimates and those from some of the agencies, as summarized in
Table 3. The intakes recommended by IOM, NORDEN, and EFSA, intended to maintain serum 25(OH)D ≥50 nmol/L in 97.5% of individuals during wintertime, used for setting personal targets, or assessing individual intakes, at 10 µg/day (NORDEN) and 15 µg/day (IOM and EFSA) are much lower than our experimentally derived value of 26 µg/day. Due to the inability of the standard meta-regression approaches used by these three agencies to capture between-individual variability [
1,
10,
11], these DRVs for vitamin D will not provide the level of population protection anticipated at the time of their establishment. The UK SACN established a Recommended Intake of 10 µg/day to maintain serum 25(OH)D ≥25 nmol/L in winter for 97.5% of the population [
9], based on three separate analyses of individual subject level data from distinct age-group specific RCTs [
18,
20,
23]. Our analysis, based on the two-step IPD, derived an experimentally derived intake of 9.9 µg/day to maintain serum 25(OH)D concentration ≥25 nmol/L in winter for 97.5% of the population, which in
Table 3 is rounded up to 10 µg/day.
To enable evaluations of the adequacy of vitamin D intakes on a population basis, two agencies proposed estimated ARs [
1,
11]. The IOM’s AR of 10 µg/day is based on maintaining serum 25(OH)D ≥40 nmol/L in 50% of individuals (aged 1 year old and upwards) during winter. Notwithstanding their declared uncertainty in their simulated dose-response relationship, the IOM used it, on the basis that this intake would considerably over-shoot the targeted serum 25(OH)D concentration [
1]. Our IPD analyses, which is free of such uncertainty in the dose-response relationship due to its underpinning individual level analytical data, shows that an intake of 4.5 µg/day during winter will maintain serum 25(OH)D ≥40 nmol/L in 50% of individuals, aged 4–86 years. The AR estimates from the two types of models (IPD analyses v. standard meta-regression with aggregate data) are much closer than these figures might suggest at face value, and this is due to the fact that both are based on an average and do not rely on the ability to use between-individual variation, in contrast to Recommended Intake estimates. To illustrate this point, the IOM standard meta-regression model showed that an intake of 10 µg/day would give a projected mean serum 25(OH)D of ~52 nmol/L [
1], while the present work based on IPD analysis shows that an intake of ~10 µg/day will maintain serum 25(OH)D ≥50 nmol/L in 50% of individuals (aged 4+ years) during winter.
The impact of using our experimentally derived AR estimate of 4.5 µg/day versus the IOM value of 10 µg/day to compare the prevalence of inadequate intakes of vitamin D in the general population is striking. For example, 55% and 89% of participants in the National Adult Nutrition Survey in Ireland would have inadequate intakes of vitamin D using these two benchmarks, respectively [
33]. Furthermore, the median intake of vitamin D in adults (aged 18–84 year) participating in the Irish adult nutrition survey was 3.7 µg/day [
33], and the measured prevalence of serum 25(OH)D <40 nmol/L in winter was 47% [
24]. The apparently contradictory evidence of a much higher prevalence of inadequate intakes compared with vitamin D status is therefore a function of the AR itself. This does not diminish the need to address vitamin D deficiency on a population-wide basis, but it does highlight the need to work towards well-founded DRVs for nutritional surveillance.
It has been stressed that IPD analyses are not without their challenges, including being resource intensive. The issue of limited availability of data for some studies could introduce bias [
15]. In this study, biochemical re-analysis of samples was required to minimize a confounding effect of method-related differences in the outcome measure. In a wider context, this would have a knock-on impact on the availability of samples in addition to data from RCTs identified in the systematic review approach. Another option that has been suggested is to collaborate with other research groups and agree to pool resources to answer specific questions [
16]. This is the approach adopted in this work which allowed us to secure serum samples for re-analysis of serum 25(OH)D and remove some of the method-related confounding that is likely intrinsic in DRV estimates to date.
A strength of the present analyses is the high-quality of the RCTs used, which were designed to address this specific question. The studies used in the present work all fulfill or exceed the previously defined RCT selection criteria established by the IOM [
1], and five of the seven have been used in the recent DRV meta-regression exercises by IOM, NORDEN, and EFSA [
1,
10,
11]. The two most recent RCTs in adolescents and children were published after the most recent reports. That the seven RCTs met with these predefined selection criteria while also minimizing serum 25(OH)D method-related differences ensured the data were of the highest quality for inclusion in an IPD analysis. It could be fairly argued that by striving to achieve a high degree of internal validity in our analysis by our selection of RCT that were all implemented using the same study design, analytical platform for serum 25(OH)D, dietary assessment method, and statistical approach, it may diminish the external validity of our findings and limit their generalizability to other contexts (i.e., other populations beyond those used in the included RCTs). However, the winter-based vitamin D RCT collection used in the present analysis is not substantially different in terms of a number of RCTs of children and adults than that used by IOM (
n = 11; 1) or NORDEN (
n = 6; 11), and all of which used winter-based RCTs performed above a latitude threshold of 49.5° N. EFSA, by choosing less strict inclusion criteria for their winter-based RCTs, and especially a much wider latitude range >40°N, utilized data from 35 RCTs [
10]. We argue that the EFSA’s assumption of minimal UVB-induced synthesis would not be met in several of these RCTs based on the latest UVB availability data for Europe [
34]. It is of note that the IOM, after performing their meta-regression in RCTs in the latitude band of 40–49.5° N, as well as in those conducted at >49.5° N, decided not to use those from the lower latitude band for this reason [
1]. Despite this, it is of note that the unadjusted Ln-models performed by these three agencies [
1,
10,
11] as well as ourselves, and especially in the face of the differences in RCT datasets, all yielded relatively similar mean achieved serum 25(OH)D to a total vitamin D intake of 10 µg/day (as an estimate all four exercises had included and thus allows for comparison), namely, in the range 52–55 nmol/L. This provides added confidence that the generalizability of our estimates is robust. In addition, the fact that the various sensitivity analyses performed in the present work (e.g., removal of RCT data from children, inclusion of BMI, increasing the compliance threshold to ≥95%) had no major impact on the findings further speaks towards the robustness of the estimates. This was also the case in the analyses by EFSA, in which age or BMI did not seem to impact on their modelling and thus were not included in their final model [
10].
The present analyses were based on linear models of vitamin D intake–serum 25(OH)D relationship, not only as the best fit model, but also in line with the suggestion that only at intakes >35 µg/day does the response of serum 25(OH)D to vitamin D begin to plateau [
35]. In terms of vitamin D intakes in the present analyses, only 0.3% of IPD data and none of the meta-regression summary intakes were >35 µg/day. This is also of relevance from a population perspective, where the 95th percentile of intakes of vitamin D in various European [
36] as well as the US population [
37] sit very well with the total vitamin D intake range in the present RCT dataset (1–44 µg/day). Despite these arguments, we recognize the fact that three of the agencies have applied curvilinear meta-regression models to the intake–status RCT data [
1,
10,
11]. Importantly, our analyses show that the estimates to maintain 50 nmol/L from an adjusted Ln-based standard meta-regression model of aggregate data yielded relatively similar estimates as the adjusted linear-based equivalent (13.2 vs. 14.2 µg/day, respectively).
Importantly, we would also stress that even inclusion of aggregate data from as many as 35 RCTs as per EFSA [
10], while enhancing the representativeness and quantifying more precisely between-study variability, will not recover the between-person variability from within the datasets needed for providing an unbiased RDA-like estimate.
Figure 1 in the present analysis demonstrates very clearly what is referred to as an ecological bias in the fitted standard meta-regression lines, reflecting that these models only picked up the relationship between vitamin D intake and serum 25(OH)D at the level of the studies considered, but not at the level of participants (these two relationships need not be the same) [
38]. The IPD analyses do not have such a bias because they rely on participant-level data.
Limitations in the present analysis stem from the type of individual participant data from appropriate RCTs available to us. While our age-range included individuals from 4 to 86 years, <5% of the sample were aged >75 years of age. This lower availability of data for the elderly subset of the population was also noted by NORDEN [
11]. We also note the lack of data for children <5 years, which is an acknowledged research gap [
39]. Access to current and accurate food composition data is a requirement for the estimation of vitamin D intakes, a key input into DRV modelling exercises. More comprehensive coverage of the vitamin D content, including 25(OH)D, of staple foods is required within food composition databases [
40]. It should be noted, however, that the harmonized dietary assessment method used in all seven RCTs in the present analysis is based on food composition data from various sources to minimize such gaps. All data from animal-derived sources included estimates of potency-adjusted 25(OH)D as well as vitamin D [
41].