The Reverse J-Shaped Association Between Serum Total 25-Hydroxyvitamin D Concentration and All-Cause Mortality: The Impact of Assay Standardization

Abstract

We evaluated the impact of standardizing the originally measured serum total 25-hydroxyvitamin D (25(OH)D) values from Third National Health and Nutrition Examination Survey (NHANES III, 1988–1994) on the association between 25(OH)D and rate of all-cause mortality. Values were standardized to the gold-standard laboratory method. Follow-up from 1990–2006 consisted of 15,099 participants aged at least 20 years at baseline, among whom there were 3,784 deaths. Relative risk of death was adjusted for age, sex, race/ethnicity, and season using Poisson regression. Results were obtained for eight 25(OH)D (nmol/L) categories: <20 nmol/L, 20–29 nmol/L, 30–39 nmol/L, 40–49 nmol/L, 50–59 nmol/L, 60–74 nmol/L, 75–99 nmol/L (reference), and ≥100 nmol/L. Assay standardization dramatically shifted original 25(OH)D values toward zero. Accordingly, risk ≥120 nmol/L could not be evaluated (i.e., n = 7 and n_deaths = 2). Relative risk (95% confidence interval (CI)) <40 nmol/L remained significant (30–39 nmol/L: relative risk (RR) = 1.4 (95% CI: 1.1, 1.6); 20–29 nmol/L: RR = 1.6 (95% CI: 1.3, 1.9), and <20 nmol/L: RR = 2.1 (95% CI: 1.6, 2.7). However, adjusted relative risk estimates for 25(OH)D levels ≥40 nmol/L were no longer significant (40–49 nmol/L: RR = 1.2 (95% CI: 0.99, 1.4); 50–59 nmol/L: RR = 1.2 (95% CI: 1.04, 1.4); 60–74 nmol/L: RR = 1.1 (95% CI: 0.94, 1.2); 75–99 nmol/L: RR = 1.0 (referent), and ≥100 nmol/L: RR = 1.1 (95% CI: 0.6, 2.1). In summary, after standardization, risk of death from all causes increased with decreasing 25(OH)D <40 nmol/L, while there was no association with values in categories between 40 nmol/L and 120 nmol/L.

In this study, we returned to the Third National Health and Nutrition Examination Survey (NHANES III, 1988–1994) with 15 years of follow-up data in order to evaluate the impact that assay standardization would have on the association between serum total 25-hydroxyvitamin D (25(OH)D) and the rate or risk of all-cause mortality. Previously, we found a J-shaped association between 25(OH)D level and death that persisted with increasing length of follow-up and after adjustment for confounding (1). In addition, a J-shaped association was reported in a meta-analysis and more recently in an analysis of the Longitudinal Aging Study Amsterdam (2, 3). Unfortunately, there is a tremendous amount of variability among the 25(OH)D laboratory assays used, which has the potential to bias the interpretation of those published research results (4) and, therefore, the development of vitamin D clinical and public health guidelines (5). Accordingly, we address 3 important questions: 1) Using standardization methodology developed by the Vitamin D Standardization Program (VDSP), how is the 25(OH)D distribution affected by standardizing the original radioimmunoassay (RIA) (DiaSorin, Stillwater, Minnesota) concentration data to the gold standard reference-measurement procedures of National Institute of Standards and Technology (NIST), Ghent University, and Centers for Disease Control and Prevention (CDC) (6–8)? 2) How does assay standardization affect the association between 25(OH)D and the risk of death, as adjusted for age, sex, race/ethnicity, and season? 3) What is the importance of assay standardization to the development of consensus cutpoints to define clinically meaningful states of vitamin D status (e.g., deficient, sufficient, and toxic)?

METHODS

Measurements

25(OH)D was measured using an RIA kit (DiaSorin) at the CDC under the direction of Dr. Rosemary Schleicher (1). 25(OH)D values recalibrated to the kit used in 2004 were used in the previous study (9, 10). More recently, a calibration study was conducted to standardize the original data using the reference measurement procedures of the NIST and Ghent University. In that study, RIA data from NHANES III were ranked by quartile and then sorted by date of analysis within each quartile; approximately 125 specimens per quartile were selected from NHANES III to represent the full range of assay dates and concentrations (n = 505). Based on a plot of the calibration study data (Figure 1), a piecewise regression model was used to standardize the original RIA (on the x-axis) values to the NIST, Ghent University, and CDC reference-measurement-procedure values (on the y-axis). Going from the minimum (8.7 nmol/L) to maximum (239.1 nmol/L), a grid search procedure was used to fit a series of 120 pairs of regression models to the left and to the right of a selected point. The starting point was 8.7 nmol/L, with constant increments equal to a delta of 1.92 nmol/L, where delta = (239.1 − 8.7) ÷ 120. The best piecewise model was the one with the highest overall R² (0.86) at the selected point of 102 nmol/L.

$$\begin{array}{c}{\mathrm{If\; RIA}}_{\mathrm{original}}\le {102,\text{then VDSP Standardized}}_{\mathrm{equivalent}}\hfill \\ =\text{1.57548}+\text{0}{\text{.8429}\times \text{RIA}}_{\mathrm{original}}.\hfill \end{array}$$

$$\begin{array}{c}{\mathrm{If\; RIA}}_{\mathrm{original}}{>102,\text{then VDSP Standardized}}_{\mathrm{equivalent}}\hfill \\ =\text{59.2296}+\text{0}{\text{.2788}\times \text{RIA}}_{\mathrm{original}}.\hfill \end{array}$$

Figure 1.

Comparison of possible models to standardize 25-hydroxyvitamin D (25(OH)D) measurements with the reference measurement procedures of the National Institute of Standards and Technology and Ghent University, using the bridging-study data set of the Third National Health and Nutrition Examination Survey. Overall, the piecewise model indicated that the impact of assay standardization was to decrease the observed concentrations of serum 25(OH)D. Moreover, it indicated that the performance of the original radioimmunoassay (RIA) changed dramatically at a serum 25(OH)D concentration of approximately 102 nmol/L, and concentrations above that level were greatly reduced as a result of VDSP standardization. Overestimates above 102 nmol/L may have been due to cross-reactivity of the antibody in the assay with 24,25-dihydroxyvitamin D3. LC-MS/MS: liquid chromatography–dual mass spectrometry.

All 25(OH)D values are reported in units of nmol/L (ng/mL = nmol/L ÷ 2.5).

Assessment of vital status

Vital status was assessed based on a probabilistic match between personal identifiers from NHANES III and the death certificate records from the National Death Index. All NHANES III participants who were aged 17 years or older at the time of the survey were eligible for mortality follow-up. However, for this study, the sample is limited to those aged 20 years or older. Vital status through 2006 was used in order to be consistent with our previous publication (1).

Analytical sample

A total of 23,258 participants aged 20 years or older were selected to participate in NHANES III, 18,825 were interviewed, and 16,573 visited the mobile examination center (MEC) for assessment, including a blood draw. Excluded from that sample were those missing information on vital status (n = 11), women who were pregnant at baseline (n = 288), participants with zero length of follow-up time from the date of examination (n = 11), and those who were missing data for serum total 25(OH)D (n = 762), serum creatinine (n = 344), body mass index (n = 34), or systolic blood pressure (n = 24). The total analytical sample size was 15,099 participants (1).

Statistical analysis

All analyses were conducted using SAS, version 9.2 (SAS Institute, Inc., Cary, North Carolina), or STATA, version 14 (StataCorp LP, College Station, Texas). “Survey” procedures from both computer packages were used in order to incorporate the sampling weights and design effects into the data analyses. Poisson regression (11) was used to model the association between serum total 25(OH)D and risk of death using a traditional categorical variable approach (12). Because we previously found that the results from minimal-adjustment models were similar to maximal-adjustment models, all models in this study were adjusted only for age, sex, race/ethnicity, and season (1). The traditional approach was used to evaluate risk in a categorical variable for serum total 25(OH)D with 9 ranges of 25(OH)D: <20 nmol/L, 20–29 nmol/L, 30–39 nmol/L, 40–59 nmol/L, 50–59 nmol/L, 60–74 nmol/L, 75–99 nmol/L, 100–119 nmol/L, and ≥120 nmol/L, as were used in the previous study (1). The reference group was 75–99 nmol/L unless stated otherwise. Estimates of the mortality rates and mortality rate ratios, also referred to as relative risk, were calculated for each of the 9 intervals, defined above in the traditional approach at the median value of each interval. The complex survey standard error and 95% confidence intervals were estimated using the delta method.

RESULTS

25(OH)D values for original, recalibrated, and VDSP-standardized results ranged from 9–400 nmol/L, 10–340 nmol/L, and 9–171 nmol/L, respectively, indicating that with each “recalibration” step, the tail at the high end of the distribution was pushed in toward the left side of the distribution (Figures 1 and 2). The piecewise linear model indicated that the RIA performance characteristics (i.e., accuracy and precision) changed at approximately the intersection of the 2 lines, or 102 nmol/L. Moreover, it indicated that the performance of the original RIA assay changed dramatically at a serum 25(OH)D concentration of approximately 102 nmol/L. At the low end of the 25(OH)D distribution, there was a tendency for concentrations to decrease with each successive “recalibration.” However, because of the changes in assay performance in the high concentration range, possibly due to cross-reactivity of the antibody in the assay with 24,25-dihydroxyvitamin D3 (13), the pressure to decrease 25(OH)D was progressively much greater at the upper tail with the move from the recalibrated data set to the VDSP-standardized data set, such that there were only 7 participants and 2 decedents in the VDSP-standardized data set with a concentration of ≥120 nmol/L. That made it impossible to model the risk of death at 25(OH)D levels ≥120 nmol/L in the VDSP-standardized data set.

Figure 2.

Density (weighted relative frequency) distribution of 25-hydroxyvitamin D measurements for the original, recalibrated, and VDSP-standardized values, Third National Health and Nutrition Examination Survey, 1988–1994.

For all the categories within 0–100 nmol/L, the recalibrated and VDSP-standardized models yielded very similar results (Figure 2 and Table 1). Up to a concentration of approximately 80 nmol/L, the distributions for the recalibrated and VDSP-standardized concentrations were almost superimposable. However, with the piecewise VDSP-standardized model, concentrations ≤102 nmol/L were decreased at a much slower rate than were concentrations >102 nmol/L, such that at approximately 85–90 nmol/L the 2 waves caught up with each other to produce a second peak.

Table 1.

Effect of 2004 Radioimmunoassay Calibration and Vitamin D Standardization Program Standardization on Original Serum 25-Hydroxyvitamin D Radioimmunoassay Values From the Third National Health and Nutrition Examination Survey, 1988–1994^a

Total 25(OH)D, nmol/L
Original Value	2004 RIA Calibrationb	VDSP Standardizationc
10	11	10
15	15	14
20	19	18
30	28	27
40	36	35
50	45	44
60	53	52
75	66	65
80	70	69
90	78	77
100	87	86
125	108	94
150	129	101
200	171	115

Abbreviations: 25(OH)D, 25-hydroxyvitamin D; NHANES III, Third National Health and Nutrition Examination Survey; RIA, radioimmunoassay; VDSP, Vitamin D Standardization Program.

^a Sources: Centers for Disease Control and Prevention, National Center for Health Statistics, NHANES III, 1988–1994, and the public-use version of the NHANES III Linked Mortality File (24).

^b Original serum 25(OH)D values (nmol/L) calibrated to the radioimmunoassay kit available in 2004 using the following equation (9): NHANES III 25(OH)D_{Corrected to 2004 RIA} = 0.8429 × NHANES III 25(OH)D_{1988–1994 RIA} + 2.5762 (nmol/L).

^c Original serum 25(OH)D values (nmol/L) standardized to the reference measurement procedures of the National Institute of Standards and Technology, and Ghent University using the following equations (10):

$$\begin{array}{c}{\mathrm{If\; RIA}}_{\mathrm{original}}\le {\mathrm{102\; then\; VDSP\; Standardized}}_{\mathrm{equivalent}}\hfill \\ =\text{1h.57548}+\text{0}{\text{.8429}\times \text{RIA}}_{\mathrm{original}}\hfill \end{array}$$

$$\begin{array}{c}{\mathrm{If\; RIA}}_{\mathrm{original}}{>\text{102,then VDSP Standardized}}_{\mathrm{equivalent}}\hfill \\ =\text{59.2296}+\text{0}{\text{.2788}\times \text{RIA}}_{\mathrm{original}}\hfill \end{array}$$

The risk curves based on the original, recalibrated, and VDSP-standardized data were very similar (Table 2 and Figure 3). However, as you move from the original to the recalibrated to the VDSP-standardized Poisson risk models, there was a tendency for the lines to flatten, so that within the range of 40–100 nmol/L, there appears to be a trough where risk remains relatively constant. In addition, the rate at which risk increases at 25(OH)D levels <40 nmol/L appears to be similar in all 3 sets of 25(OH)D data. Moreover, it must be emphasized that, after standardization, the risk of death continues to increase significantly with decreasing concentration of 25(OH)D.

Table 2.

Adjusted Relative Risk of Death From All Causes by Source and Category of Serum Total 25-Hydroxyvitamin D, Third National Health and Nutrition Examination Survey, 1988–2006^a

Source and Category of Serum 25(OH)D, nmol/L	Median 25(OH)D Concentration, nmol/L	No. of Deaths	Total Sample Size	Model-Based RRb	95% CI
Original RIA
<20	17.6	68	215	2.0	1.5, 2.7
20–29	26.1	220	910	1.6	1.3, 2.0
30–39	35.3	459	1,814	1.5	1.2, 1.8
40–49	45.4	565	2,263	1.2	0.97, 1.4
50–59	55.0	575	2,261	1.1	0.9, 1.3
60–74	67.4	777	2,999	1.1	0.9, 1.2
75–99	85.0	796	3,127	1.0	Referent
110–119	108.3	205	941	0.9	0.7, 1.1
≥120	133.5	119	569	1.2	0.9, 1.5
Recalibrated 2004 RIA
<20	17.7	79	251	2.2	1.6, 2.9
20–29	26.2	297	1,270	1.6	1.4, 2.0
30–39	35.2	592	2,340	1.5	1.2, 1.7
40–49	45.1	694	2,790	1.2	1.02, 1.4
50–59	54.8	668	2,526	1.2	1.1, 1.4
60–74	67.0	775	3,046	1.1	1.0, 1.3
75–99	83.6	533	2,156	1.0	Referent
110–119	106.9	110	518	1.1	0.8, 1.4
≥120	133.4	36	202	1.5	1.02, 2.3
VDSP-standardizedc
<20	17.5	105	334	2.1	1.6, 2.7
20–29	26.3	314	1,366	1.6	1.3, 1.9
30–39	35.5	640	2,467	1.4	1.1, 1.6
40–49	44.8	663	2,714	1.2	0.99, 1.4
50–59	54.8	666	2,502	1.2	1.04, 1.4
60–74	66.9	746	2,970	1.1	0.94, 1.2
75–99	87.7	631	2,603	1.0	Referent
110–119	104.1	17	136	1.1	0.6, 2.1
≥120		2	7

Abbreviations: 25(OH)D, 25-hydroxyvitamin D; RIA, radioimmunoassay; VDSP, Vitamin D Standardization Program.

^a Sources: Centers for Disease Control and Prevention, National Center for Health Statistics, NHANES III, 1988–1994, and the public-use version of the NHANES III Linked Mortality File (24).

^b Minimal-adjustment model: adjustment for age, sex, race/ethnicity, and season.

^c Data pooled with the concentration group of 100–119 nmol/L in mortality analyses.

Figure 3.

Association between 25-hydroxyvitamin D (25(OH)D) and the risk of death, Third National Health and Nutrition Examination Survey, 1988–2006. Original 25(OH)D values (A); recalibrated to 2004 radioimmunoassay (B); Vitamin D Standardization Program–standardized (C).

Relative risk estimates from the models based on the original, recalibrated, and VDSP-standardized data were qualitatively and quantitatively similar for the concentration groups of <40 nmol/L and 75–99 nmol/L (Table 2). Nor were the relative risk estimates—except for the 25(OH)D category <20 nmol/L (after multivariable adjustment, relative risk (RR) = 1.7; 95% confidence interval (CI): 1.3, 2.3)—affected by adjusting for self-reported preexisting disease, blood pressure, body mass index, glomerular filtration rate, and physical activity (data not shown).

However, for the 25(OH)D concentration ranges of 40–49 nmol/L, 50–59 nmol/L, and 60–74 nmol/L, the relative risk estimates based on the VDSP-standardized data were quantitatively but not qualitatively similar to the results previously published using the recalibrated data (Table 2). Using the recalibrated 25(OH)D data, it appeared as though there might be a significant negative association for the concentration groups in the range of 40–59 nmol/L, and because the lower bound of 95% confidence interval for the 60–69 nmol/L group was almost exactly 1, the risk relationship in that region was uncertain (Table 2, Figure 3B). Depending on the modeling approach (Poisson categorical or Cox proportional hazards models (data not shown)), the confidence intervals for the 40–49 nmol/L, 50–59 nmol/L, and 60–74 nmol/L VDSP-standardized concentration groups indicated that the relative risk estimates were either slightly above 1 (i.e., 50–59 nmol/L) or below 1 (i.e., 40–49 and 60–69 nmol/L). Similarly, using the Cox proportional hazards model with the VDSP-standardized 25(OH)D data, the relative risk (and 95% CI) for those 25(OH)D groups, after adjustment for age, sex, race/ethnicity, and season, were: for 40–49 nmol/L, RR = 1.15 (95% CI: 0.967, 1.4); for 50–59 nmol/L, RR = 1.2 (95% CI: 1.03, 1.4); and for 60–69 nmol/L, RR = 1.05 (95% CI: 0.91, 1.2). On the other hand, because the 95% confidence intervals are so close to 1 and because one effect of standardization was to flatten the curve in the range of 40–70 nmol/L, it was our interpretation that the risk of death did not appear to vary significantly given a total 25(OH)D concentration within the range of 40–100 nmol/L using the VDSP-standardized data.

Finally, as a result of standardization, there were too few values ≥120 nmol/L to evaluate whether risk increases at very high levels of serum total 25(OH)D.

DISCUSSION

Standardization of serum total 25(OH)D values from NHANES III had a significant effect on the interpretation of how the risk of death changes with increasing concentrations of 25(OH)D. For concentrations <40 nmol/L, the risk of death is clearly greater compared with the reference group of 75–99 nmol/L. Within the range of 40–100 nmol/L, however, the risk does not appear to change significantly with increasing 25(OH)D level. That is, there is a large trough where the concentration of 25(OH)D within the range of 40–100 nmol/L does not appear to be related to mortality risk. Finally, standardization led to such a dramatic lowering of levels in the upper tail that it became impossible to evaluate whether risk of death was related to increasing concentration of 25(OH)D at levels >100 nmol/L.

A number of other research groups have reported finding a reverse J-shaped association between 25(OH)D concentration and risk of death (2, 3, 14–18). Given our results and those of others that standardization tends to lead to a much narrower range of 25(OH)D values, it would be useful to know the effect of standardization on the results of those other studies (19, 20). A partial answer to that question may come from the currently ongoing European Commission–funded integrated project, Food-Based Solutions for Optimal Vitamin D Nutrition and Health Through the Life Cycle (ODIN) (21). A component of ODIN is to evaluate the impact of standardization on the risk relationship between 25(OH)D and death in meta-analyses. Results from ODIN will help to clarify the risk curve.

The relationship of 25(OH)D to risk of death is, however, only one issue in vitamin D science. While having standardized measurements to evaluate the risk relationship is important, the general lack of standardized research results for 25(OH)D is of larger concern. The lack of standardized total 25(OH)D research data is the fundamental limitation to the development of consensus cutpoints to define clinically meaningful states of vitamin D status (e.g., deficient, sufficient, and toxic) (22). While standardization of 25(OH)D measurements in future studies will undoubtedly be useful, we cannot afford to discard all the previous studies with unstandardized 25(OH)D measurements, nor is it realistic to think that we can standardize all of the measurements in those past studies. Standardizing serum total 25(OH)D results from previous studies is possible (23) and critical for the development of clinical and public health guidelines for vitamin D.

In summary, after standardization of the serum total 25(OH)D measurements from NHANES III using the reference measurement procedures of the NIST and Ghent University, it was clear that the risk of death increased significantly with decreasing 25(OH)D concentration for values <40 nmol/L. A note of caution is that within the concentration range of 40–59 nmol/L, there remained some ambiguity over the significance of the relative risk estimates; however, for the concentrations of 40–100 nmol/L, our results showed a trough where risk of death was unrelated to 25(OH)D concentration. Moreover, standardization can have profound effects on the entire distribution of laboratory values; in this study there were too few participants to evaluate the 25(OH)D-mortality association at high end of the distribution of 25(OH)D concentrations. Finally, the results of this study help to emphasize that only assay standardization can provide the ability to form conclusions based on the true values of serum total 25(OH)D. Without it the chaos surrounding the development of consensus cutpoints will continue.