Dr Eisman holds a patent in relation to VDR and osteoporosis. All other authors have no conflict of interest.
With the rise of molecular and genetic epidemiology, molecular association studies are increasingly common; however, meta-analysis of these studies has been a neglected area. This study performed a meta-analysis of the association of the vitamin D receptor (VDR) gene polymorphisms and BMD. We also highlight methodological issues that need to be resolved.
Introduction: With the rise of molecular and genetic epidemiology, molecular association studies are increasingly common; however, meta-analysis of these studies has been a neglected area. This study performed a meta-analysis of the association of vitamin D receptor (VDR) gene polymorphisms and BMD/osteoporosis and highlights methodological issues.
Materials and Methods: Studies published from 1994 to 2001 were identified through Medline using PubMed software. The reference lists of the articles retrieved were also reviewed. Where eligible papers had insufficient information, we contacted authors by mail (up to three mailings) for additional information. Any observational study, which tested the association between VDR BsmI genotypes and either BMD or osteoporosis at the femoral neck or spine in adult women, was included in the review. Data were extracted independently by two reviewers (AT and JA) using a standardized data extraction form.
Results: The B allele was significantly associated with BMD at the spine; it seemed to follow a recessive model, with the BB genotype having lower BMD than Bb/bb genotypes at baseline, which led to greater bone mineral loss over time. Highlighted methodological lessons included the need to check Hardy-Weinberg equilibrium and the importance of exploring heterogeneity, pooling data in a manner that is sensitive to genetic models, and avoiding multiple comparisons.
Conclusion: With the proliferation of molecular association studies, there will be an increased need to quantify the magnitude of the risk associated with genetic polymorphisms. This will likely entail meta-analytic methods, and this meta-analysis highlights some of the methodological issues that will need to be resolved.
OSTEOPOROSIS IS A DIMINUTION of skeletal mass in which bone is normally mineralized, but the amount of bone tissue in a given volume of bone is reduced, causing mechanical weakness and leading to fractures, especially of the hip and spine, which can occur either spontaneously or with minimal trauma.(1) One of the primary predictors of this fracture risk is BMD (g/cm2). Although many environmental factors impact BMD, a large component of variation in BMD seems to be genetic.(2, 3) Twin and familial studies show heritabilities of ∼60–80% for BMD.(4, 5) Dissecting the genes responsible for this contribution can be achieved by two broad lines of inquiry: linkage analysis, using family pedigrees and genome-wide markers (which will not be discussed here), and molecular association studies, using candidate genes in a population-based study design.
A number of polymorphisms in multiple candidate genes have been investigated in this regard,(6) the vitamin D receptor (VDR) gene being the first(7) and most intensively studied.(3, 8) The VDR plays a role in regulating calcium homoeostasis through binding and nuclear translocating of 1α,25(OH)2D3, affecting bone resorption, and increasing calcium absorption. Although numerous association studies relating polymorphisms in this gene to BMD have been published, results are conflicting,(9) possibly because of variations in study design, small sample sizes, and heterogeneous populations, among other issues.
Meta-analysis may be able to overcome the shortcomings of individual studies; by systematically combining results from individual studies, this method increases the power to detect an association, increases the precision of the magnitude of effect, and sheds light on reasons for discrepant results by exploring heterogeneity. Although methods for meta-analysis of traditional association studies are well established,(10, 11) applying this method to molecular association studies raises unique issues. Some of these relate to sources of error at the individual study level and have been enumerated previously(12–14) (e.g., population stratification, genotyping error, linkage dysequilibrium, and gene-environment interaction). Others relate to pooling the data in a way that reflects the biology of gene effects and handling at least three separate genotype groups while controlling for multiple comparisons.(15)
We performed a meta-analysis of the VDR BsmI polymorphism in relation to BMD and change in BMD. We also viewed this as a case study, highlighting methodological issues in the meta-analysis of molecular association studies.
MATERIALS AND METHODS
We searched for all observational studies published from January 1994 (when the first VDR association study was published) to May 2001 using PubMed software to search Medline. The search terms were as follows.
1. vitamin D receptor or VDR (MeSH)
2. genotype(s) or allele(s) or polymorphism(s) (MeSH)
3. bone mineral density or BMD or bone density (MeSH)
4. low BMD or low density (textword)
5. osteoporosis (MeSH)
6. fracture (MeSH)
7. (1 and 2) and (3 or 4 or 5 or 6)
The reference lists of the articles retrieved were also reviewed to identify publications on the same topic. The most complete and recent results were used when there were multiple publications from the same study group.
Any observational study (cohort, case-control, and cross-sectional study), regardless of sample size, which tested the association between VDR BsmI genotypes and either BMD or osteoporosis at the femoral neck or spine and fulfilled the following criteria, was included.
•BMD measurements at lumbar spine or femoral neck by DXA or dual-photon absorptiometry (DPA) method
•Results described in sufficient detail for extraction of data, that is, mean and SD of BMD and number of subjects for each VDR genotype for continuous outcomes. Where eligible papers had insufficient information, we contacted authors by mail (up to three mailings) for additional information.
•Participants were pre- or postmenopausal adult women
•VDR polymorphism was determined by the BsmI restriction site. The possible genotypes were BB, Bb, or bb, where B and b indicate absence and presence of the restriction site, respectively.
•Outcomes were mean BMD or percent change in BMD per year.
Data were extracted independently by two reviewers (AT and JA) using a standardized data extraction form. Any disagreement was resolved by discussion and consensus. Co-variables such as mean age, mean body mass index (BMI), study frame (e.g., population-based versus hospital-based), ethnicity, and menopausal status were also extracted for each study.
Quality score assessment
Quality of studies was also independently assessed by the same two reviewers. Quality scoring criteria were modified from previous meta-analyses of observational studies(16–19) (Appendix 1).
Studies were pooled separately according to site of BMD measurement. Data analysis followed the methods described in a separate paper.(20) Briefly, this method follows five steps.
1. Checking each study for Hardy-Weinberg equilibrium (HWE) and doing a sensitivity analysis including and excluding studies not in HWE
2. Checking for heterogeneity, and if present, trying to ascertain reasons for this, rather than pooling
3. Using ANOVA methods to test for an overall gene effect
4. If ANOVA is significant, looking at multiple pairwise comparisons to determine the genetic model, for example, dominant, recessive, etc.
5. Using the genetic model to collapse the three genotype groups into two groups and using random or fixed effects models to pool the data
HWE was checked in the entire cohort using a χ2 goodness of fit test.(21) A Q-test of heterogeneity based on the standardized mean difference (SMD) was performed separately for three differences of means (BB versus bb [D1], Bb versus bb [D2], and BB versus Bb [D3]). The SMD, calculated using Cohen's method,(11, 22) was chosen because the BMDs were measured using different methods or scanners. If there was heterogeneity on at least one comparison, we refrained from pooling and instead explored the cause of heterogeneity(23) by fitting the co-variables described above in a meta-regression model.(24, 25)
ANOVA was used to determine whether the VDR genotypes could significantly explain BMD. The outcome variable in the analysis was the mean BMD in each genotype group, and the unit of analysis was study. The weighted least-squares method was used to determine the main difference in the mean level of the BMD between genotypes, with weights proportional to the inverse of the variance of the mean of each group in each study. If there was an overall gene effect, the mode of inheritance was further determined using linear regression. The genotypes and study were fitted in the model as indicator variables. The double positive was treated as a reference group (bb), and the three pairwise differences were tested (BB versus bb [D1], Bb versus bb [D2], and BB versus Bb [D3]). These differences were used to indicate the most appropriate genetic model, as outlined below.
1. If D1 = D3 ≠ 0 and D2 = 0, then a recessive model is suggested.
2. If D1 = D2 ≠ 0 and D3 = 0, then a dominant model is suggested.
3. If D2 = −D3 ≠ 0 and D1 = 0, then a complete overdominant model is suggested.
4. If D1 > D2 ≠ 0 and D1 > D3 > 0, then a codominant model is suggested.
Once the best genetic model was identified, this model was used to collapse the three genotypes into two groups (except in the case of a codominant model) and pool the results using traditional meta-analysis. Again, heterogeneity was checked. If heterogeneity was absent, pooling using the fixed-effects model was used; if present, the random-effects model was used. Publication bias was checked using Egger's test.(11, 26, 27)
Sensitivity analyses were performed by including or excluding studies not in HWE and by excluding the largest studies from analysis. All analyses were performed using STATA version 6.0.(28) A p value less than 0.05 was considered statistically significant, except for tests of heterogeneity, where a level of 0.10 was used.
Characteristics of studies
Sixty-one studies were identified by the specified search terms. Twenty-two studies were ineligible for the following reasons: three studies were conducted in men only, one study reported results for men and women together, three were twin studies, two studies were in prepubertal subjects, three studies reported BMD only as a Z-score, one study reported BMD for the whole body only, one study used a CT scanner, and eight studies did not address the BsmI polymorphism. Therefore, 39 studies were considered in the analysis.(29–67) The characteristics of the studies are given in Table 1.
Table Table 1.. Characteristics of Studies Determining Association Between BsmT Polymorphism and BMD
Among the 22 studies in HWE,(29, 31–33, 35-37, 39-41, 45, 50, 52, 54-56, 58, 59, 63, 65-67) heterogeneity was assessed for the three pairwise comparisons of D1, D2, and D3 and was found to be present in all three (χ2 = 35.7, df = 21, p = 0.024 for D1; χ2 = 30.4, df = 21, p = 0.085 for D2; χ2 = 31.4, df = 21, p = 0.068 for D3). Meta-regression indicated that menopausal status was significantly associated with the SMD (coefficient = −0.32, p = 0.046); therefore, we performed a subgroup analysis according to menopausal status.
There were 13 studies on postmenopausal women that were in HWE.(29, 35–37, 39, 41, 50, 52, 54, 55, 59, 66, 67) The average age and BMI ranged from 51 to 75 years and 23 to 29 kg/m2, respectively; one(36) and four studies(29, 36, 41, 67) did not provide mean age and BMI, respectively. The frequency of allele B ranged from 0.29 to 0.53. Total sample sizes were 454 (range, 2-107), 1345 (range, 14-306), and 901 (range, 7-196) for BB, Bb, and bb groups, respectively (Table 2).
Table Table 2.. Studies Determined Association Between BsmI Polymorphism and Spine BMD in Postmenopausal Women
There was no evidence of heterogeneity in D1, D2, and D3 (χ2 = 16.1, df = 12, p = 0.185 for D1; χ2 = 15.1, df = 12, p = 0.235 for D2; χ2 = 13.8, df = 12, p = 0.315 for D3). ANOVA was used to determine the overall gene effect; we found an association between the BsmI genotype and spine BMD (F = 4.16, df = 2/24, p = 0.028). The estimated effect sizes were D1 = −0.027 (95% CI: −0.046, −0.008), D2, = −0.007 (95% CI: −0.021, 0.006), and D3 = −0.019 (95% CI: −0.037, −0.001). D1 and D3 were statistically significant, whereas D2 was not, indicating that a recessive effect was most likely. The mean spine BMDs of the Bb and bb groups were combined. The SMD among BB versus Bb/bb was estimated, and heterogeneity was again checked. The estimated pooled SMD was −0.131 (95% CI: −0.232, −0.029), with no heterogeneity (χ2 = 15.41, df = 12, p = 0.220; Fig. 1). The estimated effect size was −0.022 (95% CI: −0.036, −0.007), that is, those with the BB genotype had a lower spine BMD than those with the Bb/bb genotypes by about 0.022 g/cm2. There was no evidence of publication bias for this result (coefficient = −0.52, SE = 0.74, p = 0.498 by Egger's test).
This result was also quite robust. After adjusting for age, D1, D2, and D3 remained similar; the values were −0.028 (95% CI: −0.048, −0.008), −0.008 (95% CI: −0.022, 0.006), and −0.019 (95% CI: −0.038, −0.001), respectively. Too few studies reported BMI to adjust for this variable.
Sensitivity analysis was also performed by taking out the two largest studies,(41, 52) which contributed 24% and 18% of the total weight; the recessive effects persisted.
Sensitivity analysis was also performed by including the two studies(38, 49) that did not observe HWE. Among the 15 studies,(29, 35–39, 41, 49, 50, 52, 54, 55, 59, 66, 67) there was heterogeneity in D3 (χ2 = 25.14, df = 14, p = 0.033) but not in D1 (χ2 = 20.14, df = 14, p = 0.126) and D2 (χ2 = 18.25, df = 14, p = 0.196). Neither meta-regression nor subgroup analysis could detect the causes of heterogeneity (data not shown). Applying ANOVA methods despite the heterogeneity indicated that the gene effect was no longer statistically significant (F = 2.62, df = 2/28, p = 0.092).
Nine studies in premenopausal women were in HWE.(31–33, 36, 41, 45, 58, 63, 65) Mean age and BMI ranged from 28 to 40 years and 22 to 25 kg/m2, respectively. Total sample size of BB, Bb, and bb groups were 275 (range, 8-114), 798 (range, 12-323), and 553 (range, 9-240), respectively (Table 3). D1 and D2 were homogeneous but D3 was not (χ2 = 9.21, df = 8, p = 0.325; χ2 = 3.92, df = 8, p = 0.864; χ2 = 13.37, df = 8, p = 0.100; respectively). Meta-regression indicated that study frame, that is, population-based versus non-population-based, was associated with the SMD, and this might be the cause of the heterogeneity (coefficient = 0.598, SE = 0.223, p = 0.007).
Table Table 3.. Studies Determined Association Between BsmI Polymorphism and Spine BMD in Premenopausal Women
A subgroup analysis was performed using only the six population-based studies.(33, 36, 45, 58, 63, 65) Sample sizes of genotype groups BB, Bb, and bb were 242 (range, 8–114), 691 (range, 26-323), and 465 (range, 9-240), respectively. There was no evidence of heterogeneity for D1 (χ2 = 0.65, df = 5, p = 0.986), D2 (χ2 = 2.17, df = 5, p = 0.825), or D3 (χ2 = 1.87, df = 5, p = 0.867). ANOVA indicated that there was no main effect of gene (F = 2.78, df = 2/10, p = 0.110), and we concluded that there was no association between the VDR gene and spine BMD in premenopausal women (D1, D2, and D3 were 0.013 [95% CI: 0.001, 0.026], 0.006 [95% CI: −0.004, 0.016], and 0.007 [95% CI: −0.005, 0.019], respectively). However, to explore if the recessive model indicated in the postmenopausal group might be applicable to the premenopausal group, we imposed a recessive model in these six studies; the estimated SMD was 0.07 (95% CI: −0.07, 0.21), and this was not statistically significant (p = 0.309).
BsmI polymorphisms and femoral neck BMD
Five studies were ineligible for the following reasons: one study(64) seemed to be a duplicate of another,(52) and four studies(34, 43, 51, 61) did not provide femoral neck BMD for the BB or Bb genotype groups.
Postmenopausal white women:
Among 21 white studies, 16 studies(29, 35–37, 39, 41, 46, 47, 50, 52, 53, 55, 57, 59, 66, 68) provided mean femoral neck BMD of each genotype in postmenopausal white women. Total sample sizes for BB, Bb, and bb were 604 (range, 2-107), 1700 (range, 14-306), and 1223 (range, 7-196), respectively. These studies were pooled with no heterogeneity (χ2 = 16.3, df = 15, p = 0.361 for D1; χ2 = 16.1, df = 15, p = 0.378 for D2; χ2 = 20.7, df = 15, p = 0.146 for D3). ANOVA indicated that there was no overall gene effect (F = 0.15, df = 2/30, p = 0.863). The estimated D1, D2, and D3 were 0.002 (95% CI: −0.010, 0.014), −0.001 (95% CI: −0.010, 0.008), and 0.003 (95% CI: −0.009, 0.015), respectively.
This result was relatively robust. Among these 16 studies, 2 contributed the most to sample size.(44, 52) Sensitivity analysis after removing these two largest studies one by one did not change the results. In addition, imposing the recessive effect found for BsmI genotypes at the spine did not change the results (SMD = 0.01; 95% CI: −0.08, 0.10). Sensitivity analysis including the one study not in HWE(49) also did not change the results (F = 0.37, df = 2/32, p = 0.697).
Premenopausal white women:
Six studies(33, 36, 41, 45, 63, 65) determined association between femoral neck BMD and BsmI polymorphism in premenopausal women. The sample sizes for BB, Bb, and bb groups were 181 (range, 8–114), 533 (range, 25-323), and 361 (range, 9-240), respectively. All pairwise comparisons were homogeneous (χ2 = 4.2, df = 5, p = 0.523 for D1; χ2 = 4.2, df = 5, p = 0.520 for D2; χ2 = 3.3, df = 5, p = 0.653 for D3). ANOVA found no association between the VDR gene and femoral neck BMD in this group (F = 2.44, df = 2/10, p = 0.137). The estimated D1, D2, and D3 were 0.022 (95% CI: −0.0002, 0.043), 0.006 (95% CI: −0.009, 0.022), and 0.015 (95% CI: −0.005, 0.036), respectively.
BsmI polymorphism and percent change in spine BMD
There were 13 cohort studies(30, 35, 37, 39, 41, 50, 55, 58–60, 65, 69, 70)investigating the association between mean percent BMD change over time and BsmI polymorphisms. Four studies were ineligible for the following reasons: one did not provide SD,(58) one included only men,(70) one did not provide data separated by gender,(69) and one did not provide data for the BB genotype.(60) Of the nine remaining studies, all observed HWE. Total sample sizes for BB, Bb, and bb groups were 158 (range, 2-46), 501 (range, 14-134), and 399 (range, 7-96), respectively.
There was evidence of heterogeneity in D1 (χ2 = 16.88, df = 8, p = 0.031), D2 (χ2 = 19.78, df = 8, p = 0.011), and D3 (χ2 = 19.49, df = 8, p = 0.012). Meta-regression and subgroup analyses did not reveal the cause of heterogeneity.
Determination for gene effect despite this heterogeneity indicated a statistically significant gene effect (F = 5.28, df = 2/16, p = 0.017). We also found significant differences in D1 (−0.589; 95% CI: −1.105, −0.074), and D2 (−0.444; 95% CI: −0.768, −0.119) but not in D3 (−0.145; 95% CI: −0.643, 0.352), that is, the BB and Bb genotypes had greater loss in BMD per year than the bb genotype, consistent with a dominant mode of effect.
Mean change in BMD for genotype groups BB and Bb were therefore collapsed. With the random effect model, the estimated SMD was −0.28 (95% CI: −0.49, −0.06), and this was statistically significant (Z = 2.55, p = 0.011). The estimated difference was −0.43, that is, those with BB and Bb genotypes had a mean percent BMD loss per year of 0.43 more than those with genotype bb (see Fig. 2). Egger's test indicated no publication bias (coefficient = 0.48, SE = 1.648937, p = 0.778)
BsmI polymorphism and percent change in femoral neck BMD
There were 11 studies(35, 37, 39, 41, 47, 50, 55, 58, 59, 65, 70) that determined the association between mean percent BMD change at the femoral neck and the BsmI polymorphism. Two studies were excluded: one did not provide SDs,(58) and one included only men.(70) The total sample sizes for BB, Bb, and bb groups were 169 (range, 2–46), 501 (range, 14-134), and 364 (range, 7-96), respectively.
Heterogeneity was present (χ2 = 75.85, df = 8, p = 0.008 for D1; χ2 = 84.41, df = 8, p < 0.001 for D2; χ2 = 26.70, df = 8, p = 0.001 for D3). Neither meta-regression nor subgroup analyses could identify the causes of heterogeneity. Determination of gene effect despite heterogeneity indicated no association between genotypes and mean percent BMD change (F = 2.82, df = 2/20, p = 0.089). The estimated D1, D2, and D3 were −0.627 (95% CI: −1.44, 0.185), −0.619 (95% CI: −1.199, −0.039), and −0.008 (95% CI: −0.773, 0.758), respectively. Imposing a dominant model, as found for change in spine BMD, did not change the results; the estimated SMD was −0.27 (95% CI: −0.81 to 0.27), and this was not significant (p = 0.329).
We used a new process of meta-analysis to pool molecular association studies addressing the relationship between the most common VDR gene polymorphism and various measures of bone mass. Our main result was the presence of an association between the BsmI polymorphism and spinal BMD in postmenopausal, but not premenopausal, women. This association was modest and seemed to follow a recessive mode of action; those with the BB genotype had lower BMD than those with the Bb/bb genotype by ∼0.022 g/cm2. This result was very robust; sensitivity analyses that removed the largest studies, those not in HWE, and those adjusted for age did not significantly change the magnitude of the gene effect or the genetic model. This effect is also consistent with results from previous meta-analyses; Cooper et al.(71) found an effect size of 0.03 g/cm2 or 2.5% in the same direction, although this did not reach significance (p = 0.062). Gong et al.(72) also concluded that there was a significant association, although they could not state the magnitude.
The magnitude of the decrease in spinal BMD with the BB genotype is very modest. To put this in context, the gene effect is similar to “aging the bones” by 1 year or decreasing BMD by 0.3 SD. Given that the average BMD for the Bb/bb genotype groups was 0.933, this gene effect represents a 2.4% decrease in spinal BMD. If each SD decrease in BMD causes a 50% increase in the risk of fractures and we assume that one SD = 10% as a minimum, the gene effect would translate into a ∼12% increase in the risk of fractures in those with the BB genotype. Given that the pooled prevalence of the BB genotype was 16.8%, this results in an estimated population attributable risk of spine fracture of 1.98%, that is, almost 2% of the spine fractures in the general population can be attributed to the BB genotype.
One caveat with our analysis is that it is based on regression analysis. Other approaches to dealing with multiple comparisons (e.g., Tukey's, Scheffe's, or Bonferroni's tests) can also be applied after ANOVA, but they are more conservative. For example, applying Tukey's test on BMD in postmenopausal women indicated a significant difference in D1 (95% CI: −0.049, −0.005) but not in D2 (95% CI: −0.024, 0.010) or D3 (95% CI: −0.019, 0.009). Thus, although the overall gene effect persists, the recessive model is undetectable with this approach.
This gene effect was not seen in premenopausal women. This seems to be consistent with results from calcium supplementation trials; calcium supplementation seems to decrease BMD loss at the spine in late, but not early, postmenopausal women(73) and has no effect in premenopausal women.(74)
Pooled results relating BsmI genotype to change in spinal BMD per year were heterogeneous, and we were unable to identify the source. Pooling despite this heterogeneity indicated a dominant model, such that BB and Bb genotypes lost more BMD per year than bb genotypes. Thus, it seems that the B allele carries a double deleterious effect, leading to lower baseline BMD and greater losses in BMD at the spine, although we are cautious about this latter conclusion because of heterogeneity.
We did not find any association between BsmI polymorphisms and femoral neck BMD. This lack of association was relatively robust in sensitivity analysis. This result is at variance with previous meta-analyses,(71, 72) which found a significantly lower BMD in the BB group by ∼0.02 g/cm2,(71) as well as biological data, indicating that femoral neck BMD does respond to calcium supplementation.(73) The reason for this discrepancy is unclear; it may be because of the small number of studies included in the previous meta-analyses, that is, more negative studies have appeared since then, or it may be because of the fact that previous meta-analyses pooled despite the presence of heterogeneity.
With the explosion in molecular epidemiological methods and microarray technology, the hope is that genetic “risk profiles” for various diseases can be developed. The current paradigm in the field suggests that polymorphisms in multiple genes, each with a small effect, will act, or interact, together to determine overall risk. If this proves true, very large individual studies, or meta-analyses of multiple smaller studies, will be needed to detect these marginal to modest genetic effects. Although there are many attempts underway to create large scale DNA and information databases to allow large scale association studies, these will likely take many years to set up, and it is likely that meta-analysis will remain the method of choice in the near future. To date, few methods have been developed for meta-analysis of molecular association studies.(15) This meta-analysis is instructive in identifying a number of methodologic problems and issues particular to molecular studies.
1. Heterogeneity. Although it is good practice to explore heterogeneity in meta-analysis of traditional studies, for example, RCTs, it seems that this is particularly important in molecular association studies. The possible sources of heterogeneity are numerous and include some unique to genetics, such as population stratification, admixture, linkage disequilibrium, HWE, and varying allele frequencies in different ethnic groups. Pooling results despite heterogeneity has the potential to generate meaningless or even misleading results,(23, 25, 75) perhaps more so with molecular association studies than traditional studies. In our example, pooling despite heterogeneity gave different results than pooling homogenous subgroups. In addition, meta-regression was not always able to uncover the source of heterogeneity, and as in traditional meta-analysis, one needs to be guided by clinical judgment and biological evidence.
2. Pooling data from more than two groups. At minimum, di-allelic polymorphisms will generate three genotype groups (as in our example); however, there are potentially many more genotype groups, and meta-analytic methods need to be developed that handle multiple groups and do so in a way that reflects potential genetic models of action (e.g., dominant, recessive, co-dominant, etc.).
3. Putting genetic data in context. In our example, too few studies measured and included potential confounders and effect modifiers such as menopausal status, calcium intake, smoking and alcohol history, and BMI. To use genetic polymorphisms clinically, one would need to account for these co-variates to estimate the incremental information provided by genotyping, above and beyond that obtained from “traditional” variables.
4. Gene-environment interaction. Gene × environment interactions may also be significant; for example, a BB genotype may have one effect if calcium intake is low and another if intake is high. Gene-gene interactions may also occur (e.g., VDR polymorphisms may interact with polymorphisms in the estrogen receptor or collagen A1 genes to influence BMD).
5. The need for biological data. In our example, it was disappointing that after 8 years of research, so little was known about the functional effects of the BsmI polymorphism. There is a strong need to have biological data to help formulate the hypotheses regarding molecular associations.
In summary, with the proliferation of molecular association studies, the ease of genotyping, and the prospect of developing genetic risk profiles for complex diseases, there will be an increased need to quantify the magnitude of the risk associated with genetic polymorphisms. This will likely entail meta-analytic methods, and this meta-analysis highlights some of the methodological issues that will need to be resolved.
We thank Prof Gerard Lucotte and Drs Bente L Langdahl, G Sigurdsson, HL Jorgensen, B Lawrence Riggs, J Marc, Patrick Garnero, Joseph M Zmuda, Omar M Hauache, and Maria Luisa Brandi for generously providing us with additional information on their studies.
Table Appendix 1. Criteria of Methodlogic Quality Assessment for Cross-Sectional or Cohort Study 11