Hip fracture risk in relation to vitamin D supplementation and serum 25-hydroxyvitamin D levels: a systematic review and meta-analysis of randomised controlled trials and observational studies

We applied the inclusion and exclusion criteria below to studies identified through searches of the PubMed and Web of Science databases, additional articles identified from the reference lists of sourced papers, hand searching of relevant journals and forward citations searches of key papers, to include publications up to April 2009.

Studies were included if they were published, English language, original research articles and were one of the following.

Exclusion criteria included randomised controlled studies that used vitamin D treatment combined with other therapies such that the individual effect of vitamin D could not be established [8‐11]; studies that included hip fractures as a component of all fractures but without individual counts; and studies published as abstracts only.

Although only studies from English-language publications were included, PubMed encompasses abstracts from non-English publications. This revealed only one relevant non-English language study [19] and this abstract provided similar results to the other studies that were included.

Data were independently extracted by three reviewers (JL, EB, RL) and discrepant results reconciled through arbitration. Two of the reviewers (EB and RL) are researchers and specialist public health physicians and one (JL) has qualifications in statistics. Adjusted results were used where possible. JL conducted the quality assessment.

Randomised controlled trials were assessed for quality based upon allocation concealment, blinding of interventions, and the total loss to follow up. Lower quality studies were those considered to be inadequate in any of the above areas (no allocation concealment, no blinding of interventions or >25% loss to follow up) including studies where the description presented was not clear. We investigated heterogeneity between lower and higher quality studies.

For observational studies, the quality assessment was based upon adjustment for confounding (none = 1 point, age and sex = 2 points, age and sex plus other possible confounders = 3 points), the selection of cases and controls over a comparable time period (1 point for same period or season), and the adequate description of patient characteristics (1 point for inclusion of characteristics such as residential status, sunshine exposure, calcium intake and medical history). These broad criteria were taken from the list of most commonly used measures in a review of quality assessments of observational studies [20]. Adjustment for confounding was weighted more heavily as it was considered the best proxy for study quality. Based upon the number of points scored out of a maximum of 5, each study was given a quality rating from 1 (lowest) to 5 (highest). We investigated heterogeneity between lower quality studies (1 and 2) compared to higher quality studies (3, 4 and 5).

For the randomised controlled trials we tabulated the measure of association for hip fracture comparing vitamin D treatment and control groups. Some studies provided the hazard ratio from a Cox Proportional Hazards model. For studies where this was not available, we calculated the relative risk from raw data [21, 22]. To check the validity of directly comparing these measures of association, we calculated the relative risk (RR) (from the raw data) for all studies and performed a sensitivity analysis using a standard t-test with unequal variances to compare the results obtained by this method to those using the adjusted hazard ratio. This revealed that the treatment effects using both methods were statistically equal (p = 0.99) and as such the adjusted results were used where possible. Outcomes were analysed on an intention-to-treat basis. We examined heterogeneity of results according to the type and dose of vitamin D used, mode of administration, use of placebo, reported levels of compliance, and according to whether or not participants were community dwelling or institutionalized, using the Cochrane Q statistic [23, 24] with the p-value obtained from a chi-squared distribution with n-1 degrees of freedom. Overall heterogeneity between individual studies was investigated using the Cochrane Q statistic in addition to the H and I² statistics with the p-value obtained from a standard normal distribution. The I² statistic is a transformation of the H statistic and represents 'the proportion of variance attributable to heterogeneity'. Further details on its calculation are outlined in Higgins et al [25]. Cut-offs for heterogeneity were taken at p = 0.05. We tested for publication bias using funnel plots and asymmetry determined using Egger's linear regression approach with the p-value obtained from a t-distribution with n-2 degrees of freedom [26].

Most case-control studies reported the mean and standard deviation of serum 25(OH)D levels directly. However, in studies where participants were stratified according to gender or season, we obtained the combined standard deviation from the usual ANOVA sum-of-squares breakdown.

Within case-control studies, 25(OH)D and PTH measurements of hip fracture patients and controls were compared using a ratio estimator of the respective means. To account for the positively skewed distribution common in serological measurements a logarithmic transformation of the data was used to better approximate normality. As most studies reported only the arithmetic mean and standard error, it was necessary to estimate the geometric mean and standard error on the log transformed scale using asymptotic Taylor series approximations. Details of the methodology used are outlined in Higgins et al [27].

Results were summarised across studies using the weighted average of the study-specific log ratios, with each individual weight inversely proportional to its estimated variance. Graphically log ratios are represented as black squares with areas proportional to their weights indicating the amount of statistical information for each particular study. The corresponding confidence interval (CI) is drawn as a line extending from the estimated log ratio.

All analyses were undertaken with the R computing package (version 2.5.1; 2007, available at: http://​www.​cran.​r-project.​org) [28].

Major results from cohort studies were presented as reported in the original studies. A meta-analysis of estimates was not possible due to the different cut points for serum 25(OH)D from each study.

A total of seven eligible randomised controlled trials recording hip fractures was identified (see Figure 1) [21, 22, 29‐33]. Six studies were vitamin D versus placebo or "no treatment" trials [22, 29‐33], and one was a factorial design investigating vitamin D versus placebo and vitamin D and calcium versus calcium alone [21]. The same study was a secondary fracture prevention study that selected participants based upon a previous fracture history. Table 1 summarises the studies including study populations, vitamin D dose and type, mode of administration, change in 25(OH)D levels and the relative risk or hazard ratio reported in the study paper.

The pooled analysis of these studies (Figure 2) included 424 and 377 hip fractures in the vitamin D and control groups respectively. The weighted RR was 1.13 (95%CI, 0.98-1.29). There was no evidence of heterogeneity (χ² ₇ = 4.44, p = 0.73 and H = 1, I² = 0%, p = 1), including in relation to geographical location. All studies reported adequate allocation concealment; one study did not use placebo and did not adequately blind intervention [32] and two studies reported a loss to follow up over the course of the study of >25% [21, 22]. Further, one study did not report loss to follow up figures [33]. Performing a sub-group analysis comparing these studies with the higher quality studies showed no evidence of heterogeneity (χ² ₁ = 0, p = 1). No significant variations were found between results of studies randomising participants to: <800 IU/day 1.14 (95%CI, 0.86-1.49) or ≥800 IU/day 1.12 (95%CI, 0.96-1.32); vitamin D₂ 1.21 (95%CI, 0.99-1.48) or vitamin D₃ 1.06 (95%CI, 0.88-1.28); oral 1.08 (95%CI, 0.94-1.25) or intramuscular injections 1.49 (95%CI, 1.02-2.18); placebo 1.11 (95%CI, 0.97-1.28) or no placebo 1.36 (95%CI, 0.80-2.34); reported levels of non-compliance (≤15% 1.26 [95%CI, 1.03-1.55] vs >15% 1.03 [95%CI, 0.86-1.24]); or between nursing home 1.11 (95%CI, 0.91-1.36) and community residents 1.14 (95%CI, 0.95-1.37). There was no evidence of publication bias.

Table 3 shows the three identified cohort studies investigating 25(OH)D levels and hip fracture [51‐53]. Two studies used nested case-control designs. One of these studies did not find a significant association between serum 25(OH)D and hip fracture using a cut-off point of 47.5 nmol/L [51]. Another study containing 400 hip fractures found a dose-related increase in hip fracture risk for lower serum 25(OH)D levels (OR = 1.33 [95%CI, 1.06-1.68] for each 25 nmol/l decrease) [52], while the third study concluded that there was a significantly reduced risk of hip fracture in those with 25(OH)D levels greater than or equal to 62.5 nmol/L compared to levels below this (RR = 0.64 [95%CI, 0.48-0.89]) [53]. For comparability with the previous studies, the reciprocal of this value is RR = 1.56 (95%CI, 1.12-2.08). Despite the differences in the cutoff points in serum 25(OH)D levels used in these analyses, the results of the cohort studies are essentially consistent with one another and qualitatively similar to those of the case-control studies, with overlapping confidence intervals and generally elevated risks of hip fracture in those with lower 25(OH)D levels. It should be noted that in the one study that permitted the comparison of mean 25(OH)D levels in cases and controls, the difference appeared less extreme than that seen in the case-control studies, although 25(OH)D levels were still statistically different (cases 55.95 nmol/L [SD 20.28] vs. controls 59.60 nmol/L [SD 18.20], p = 0.007).

A strength of this meta-analysis is the presentation of a comprehensive collection of randomised controlled trials, case-control studies and cohort studies, allowing a broader comparison of published results than previously considered. The effect of vitamin D supplementation was separated from that of other medications in the RCT data. Clearly defined methods were used in the selection of the studies and analysis of the results to ensure validity and consistency. Furthermore, this represents the first summary of serological evidence on both vitamin D and PTH in relation to hip fracture, and the first meta-analysis of case-control studies on this topic. By placing the evidence of the case-control studies alongside the results of the randomised controlled trials we highlight the complexity of the problem and the difficulty in drawing adequate conclusions.

This study is constrained by the detail and quality of published data of the respective studies included. Many of the case-control studies were relatively small, contained few hip fractures and reported highly variable serum measurements. Individual participant data were rarely published especially in larger studies, limiting a more detailed analysis of patient characteristics. The substantial heterogeneity in the results of the population-based case control studies of 25(OH)D and hip fracture, and the studies of PTH levels mean that the summary results should be considered as providing a broad indication of the overall direction of study findings, rather than a precise estimate of the combined study results. Despite some studies being of lower quality, investigations revealed no heterogeneity between results of higher and lower quality studies, both for randomised controlled trials and case-control studies.

Other potential limitations may include measurement error in 25(OH)D levels from poor assay standardisation [54], misclassification bias in hip fracture status or publication bias from the inclusion of smaller studies.

There have been several meta-analyses of randomised controlled trials published on this topic. A Cochrane review showed similar results for hip fracture risk to the present study with a summary RR of 1.15 (95%CI, 0.99-1.33), in individuals randomised to receive vitamin D compared to placebo or control [5]. A pooled analysis of individual level randomised controlled trial data showed a borderline statistically non-significant decrease in hip fracture risk from vitamin D + calcium supplementation and no reduction in risk from vitamin D supplementation alone [55]. However, these reviews did not consider results from observational studies or the role of PTH as we have done here. Shrier et al suggest that the advantages of including observational studies with RCTs in meta-analyses may outweigh the disadvantages [18]. By considering the evidence in its entirety we are able to offer plausible hypotheses for the apparent null effect seen in RCTs.

Several explanations have been previously proposed for the null Randomised Controlled Trial results. A previous meta-analysis of the relationship between vitamin D supplementation and overall fracture risk stratified by dosage concluded that supplementation of 700-800 IU/d is necessary to reduce non-vertebral fractures [6]. This was followed up most recently by the same authors with another meta-analysis supporting the same conclusion of a dose-dependence [7]. However, these studies did not assess the impact of vitamin D alone but rather included combined treatments with calcium, and calcium alone has been shown to have a significant impact on fracture risk [12]. The latter meta-analysis of calcium treatment reported that the addition of trials of vitamin D + calcium to calcium-only trial results did not change the treatment effect significantly. The RR (95%CI) for fractures at all sites for calcium-only supplementation compared to placebo was 0.90 (95%CI, 0.80-1.00), which became 0.87 (95%CI, 0.77-0.97) when calcium + vitamin D versus placebo trial results were added; these RR did not differ significantly from one another (p = 0.63). This highlights the critical question of whether vitamin D supplementation is itself effective, either with or without calcium supplementation.

Where randomised and observational evidence are inconsistent, it is generally most appropriate to place more weight on the randomised evidence for clinical decision making. Although the reason for the apparent difference between the randomised and observational evidence on vitamin D and hip fracture is not known, there are a number of possible explanations, the most obvious being that the results of the observational studies may be affected by uncontrolled confounding. The risk of hip fracture is increased by many factors such as advancing age, lack of physical activity, low body mass index, smoking, co-morbidities and frailty. Many of these factors are also likely to reduce exposure to sunlight and therefore vitamin D levels. Since only six of the case-control studies summarised here accounted for age and none adjusted for current illness or disability, it seems likely that the results of the observational studies may be subject to residual confounding. This is supported by our finding of a greater difference in 25(OH)D levels in case-control studies using population rather than hospital controls. Previous studies suggest hospital patients and nursing home residents may have frailty patterns more comparable to hip fracture patients and have lower 25(OH)D levels compared to independent elderly participants [56] who are likely more ambulant and healthier. One of the cohort studies did attempt to adjust for frailty and physical functioning [52] and these factors were found to attenuate the association between low serum vitamin D status and hip fracture risk, although it remained marginally statistically significant. Further similar investigations are necessary to better quantify the likely impact of such confounding and may help explain these conflicting results.

A second explanation is that vitamin D is indeed beneficial and the randomised controlled trials may not have been able to detect this effect. There may have been inadequate power in the trials because of limited follow up and the relatively small number of recorded hip fractures. However the results of a Cochrane review of all fracture sites (summary RR = 1.01 [95%CI, 0.93-1.09]), including more events with greater statistical power, and vertebral fractures only (RR = 0.90 [95%CI, 0.42-1.92]) were similar to those reported here, in individuals randomised to receive vitamin D compared to placebo or control [5]. Performing the analysis on an intention-to-treat basis is consistent with recommendations [57] although it may obscure the effect of poor adherence to treatment or follow up that may reduce the observed treatment effect. However, not all studies reported detailed information on adherence rates and of those that did, four estimated the proportion of adherent participants greater than 75% [22, 29‐31], and only one study reported lower adherence to treatment of around 50% [21]. This latter study reported a similar treatment effect compared to other studies where there was better adherence.

Another proposed explanation for the different findings of RCTs and observational studies, in particular the lack of a significant treatment effect in the RCTs, is that the doses of vitamin D used in many of the studies may have been too low to achieve hip fracture reduction. Some studies suggest that a minimum supplementation of 700-800 IU/d is necessary to reduce non-vertebral fractures [6]. However as already noted, the treatment effect of trials within this dosage range were not found to differ significantly from those in lower dose trials in the present meta-analysis. Furthermore, the doses of vitamin D supplementation used in both high and low dose randomised controlled trials resulted in increases in serum 25(OH)D levels that were similar to, or in excess of, the differences between cases and controls in the observational data [21, 29, 30, 32, 33]. This suggests that the apparent discrepancies between the observational and randomised data are unlikely to be explained by dose alone; it remains possible, though currently unsupported, that very high doses of vitamin D could prevent fracture.

Finally, it is possible that some other factor, such as exposure to sunlight or genotypic factors, independently reduces hip fracture risk and is related to increased serum 25(OH)D levels, but is not simulated by vitamin D supplementation.

There was no significant difference between PTH levels in hip fracture cases and controls. Low 25(OH)D levels lead to a small decrease in serum 1,25(OH)₂D and calcium absorption that in turn stimulates the secretion of PTH to maintain adequate 1,25(OH)₂D production and calcium homeostasis [56]. This secondary hyperparathyroidism leads to increased bone turnover, bone loss and possibly increased hip fracture risk. It might be expected then that hip fracture patients compared to controls would have elevated PTH measurements consistent with lower 25(OH)D levels. This was not evident although there was significant heterogeneity between studies.

The heterogeneity was not attributable to the timing of sera collection or stratification by population-based or hospital-based controls. However, PTH can vary significantly over a short period and the impact on serum measurements from fracture and subsequent trauma is not well understood. One study showed elevated PTH levels immediately after hip fracture that fell significantly two weeks later [58], while others have reported serum levels increasing [59] or remaining stable [60] during hip fracture recovery. Comparatively, 25(OH)D has a much longer half-life in circulation and studies have consistently showed no changes in 25(OH)D levels from post-fracture through to recovery [58, 61].

Despite the null findings from the randomised controlled trials, vitamin D is widely recommended and used with the aim of preventing fracture. Current guidelines for bone health support both vitamin D supplementation only, starting from as little as 400 IU/day for individuals who may be sunlight deprived [1], and combined vitamin D and calcium supplementation, as part of broader osteoporosis treatment [3]. However despite current practice, the question still remains of whether vitamin D itself is actually effective. There is insufficient evidence at present to support widespread vitamin D supplementation for fracture prevention. In particular, lack of consistency within the epidemiological evidence highlights the possible impact of bias and confounding, as well as uncertainties on dosage and therapy requirements.

The issue of confounding between vitamin D, physical activity and comorbidity has implications beyond the investigation of vitamin D and fracture. Physical activity, and related body mass index, influences the risk of a wide range of conditions and frailty and co-morbidity have profound impacts on survival. Recently, low vitamin D has been suggested as a risk factor for a number of conditions, including prostate cancer [62], breast cancer [63, 64], type 1 and 2 diabetes [65], hypertension [66], multiple sclerosis [67, 68] and is speculated to reduce survival from melanoma [69]. These types of observations are potentially affected by similar issues to those outlined here and high quality evidence, ensuring that potential confounding factors have been properly taken into account, is required before definitive conclusions can be reached regarding the effects of vitamin D. It should also be borne in mind that a number of previous observational studies showed reduced rates of cancer and other diseases in individuals taking supplements such as beta-carotene, vitamin A and vitamin E [70]. Subsequent randomised controlled trials revealed that these supplements did not significantly prevent disease [71] and in the case of beta-carotene led to an increased risk of cancer and cardiovascular disease in smokers [72, 73].

The way forward for fracture should focus upon strengthening the evidence, informed by the results of previous studies. Specifically, observational studies with tighter control for confounding are likely to be informative. Randomised controlled trials using larger vitamin D doses and combined vitamin D and calcium therapies (versus calcium alone), with appropriate control groups may also contribute to our understanding. Studies should account appropriately for factors such as age, physical activity, functional capacity and appropriate lifestyle factors when assessing the relationship between vitamin D status and hip fracture. Genetic studies examining factors relating to 25(OH)D levels may also provide useful insights for future research [74‐76].