Several studies have addressed the effect of the Sp1 polymorphism of the collagen Iα 1 (COLIA1) gene on the prevalence of fractures. The results are not in full agreement on whether this polymorphism is associated with fracture risk. To clarify this uncertainty, we performed a meta-analysis including 13 eligible studies with 3641 subjects. The COLIA1 Sp1 polymorphism showed a dose-response relationship with the prevalence of fractures. The risk was 1.25-fold (95% CI, 1.09–1.45) in Ss heterozygotes versus SS homozygotes, 1.68-fold (95% CI, 1.35–2.10) in ss homozygotes versus SS> homozygotes, and 1.35 (95% CI, 1.04–1.75) for ss homozygotes versus Ss heterozygotes by random effects calculations. There was modest heterogeneity for these three effect estimates (p value for heterogeneity, 0.17, 0.16, and 0.08, respectively). The Sp1 polymorphism effects possibly were larger when the analysis was limited to studies considering only vertebral fractures (pooled risk ratios [RR], 1.30, 2.07, and 1.46, respectively). Conversely, the Sp1 polymorphism effects tended to be smaller in studies with mean patient age ≥65 years than in studies with younger patients on average, but the differences were not formally significant. We estimated the total average attributable fraction (AF) of fractures due to the s allele in European/U.S. populations as 9.4%. The meta-analysis suggests an important role for the Sp1 polymorphism in the regulation of fracture risk; however, potential heterogeneity across ethnic groups, age groups, and skeletal sites may be important to clarify in future studies. Very large studies or meta-analyses are required to document subtle genetic differences in fracture risk.
A POLYMORPHISM at an Sp1 binding site in the collagen Iα 1 (COLIA1) gene has been proposed as a regulator of bone mineral density (BMD).(1) A large number of studies have been performed since the original discovery of this association. However, the unfavorable effect of the s allele has not been seen consistently across different studies and substantial controversy has arisen. More importantly, it has been questioned whether this polymorphism has a meaningful clinical effect on the risk of fractures(2, 3) and data have been most controversial for vertebral fractures in particular. Isolated studies have reached different conclusions and have been too small to document the magnitude of a plausible effect with accuracy or to detect dose-response relationships for ss homozygotes versus Ss heterozygotes. The attributable fraction (AF) of fractures that may be caused by the Sp1 polymorphism is unknown.
To address these issues, we performed a meta-analysis of all available studies relating the Sp1 polymorphism to the prevalence of fractures. The meta-analysis incorporated data from all identified studies in the field. We aimed at summarizing the results of these investigations, understanding whether significant heterogeneity may exist among them and clarifying whether study design, population characteristics, and other parameters may underlie this heterogeneity.
MATERIALS AND METHODS
Identification and eligibility of relevant studies
A MEDLINE and EMBASE search (last updated 8/2000) was complemented with perusal of abstracts from major meetings and communication with experts to avoid publication bias. The search strategy was based on several combinations of pertinent terms including “osteoporosis,” “collagen,” “polymorphisms,” “fracture,” and “genetics.” The references of retrieved articles and of review articles also were screened to identify potentially eligible studies.
All studies of Sp1 genotyping qualified for inclusion into the meta-analysis, provided that Sp1 genotyping had been performed with molecular methods (polymerase chain reaction of fragments containing the Sp1 polymorphism)(1) and both subjects with and without fractures had been studied. Only prevalent fractures were considered in the meta-analysis and quantitative data synthesis. The more sparse data on incident fractures also were collected for comparison. We excluded ecological analyses (studies relating the prevalence of Sp1 polymorphism with the prevalence/incidence of fractures across different populations without individual patient data correlations), studies of populations in which the Sp1 polymorphism was absent, and studies of populations without any reported fractures (studies of children or young adults and those not quoting fracture data).
Data were extracted independently by two investigators on prespecified data forms. A few disagreements were discussed and consensus was reached on all data. Extracted information focused on demographics, country of origin, definition of fractures, type of study design and definition of the control group, and numbers of subjects with and without fractures for each genotype of the Sp1 polymorphism (SS homozygotes, Ss heterozygotes, and ss homozygotes).
Three different comparisons were performed: (a) Ss heterozygotes versus SS homozygotes; (b) ss homozygotes versus SS homozygotes; and (c) Ss heterozygotes versus ss homozygotes. This approach allowed assessment of whether there was a gradation of risk between Ss heterozygotes and ss homozygotes. In an alternative approach, we performed logistic regressions using the allele dose as the dependent variable. The results were qualitatively similar (not shown), but this approach is inferior in that it makes the assumption a priori that there is a dose effect, while the separate comparisons allow for dose effects to be evaluated and shown if they are present.
The main analysis used the risk ratio (RR) as the metric of effect. In a two by two table format, if a is the number of subjects with fractures who have the polymorphism, b is the number of subjects without fractures who have the polymorphism, c is the number of subjects with fractures who do not have the polymorphism, and d is the number of subjects without fractures who do not have the polymorphism, then the RR is [a(c + d)]/[c(a + b)]. Results were similar when using the odds ratio (OR) instead of the RR as the metric of choice, although in some comparisons there was more heterogeneity. This was anticipated, because diversity in the populations and study designs caused the proportion of subjects with fractures to differ across different study populations. The OR is given by ad/bc and thus it diverges more from the RR when the rate of fractures in the study population becomes larger.(4) Thus, pooled ORs are shown for the main analyses for comparison with the RRs.
The synthesis of the data was performed both with fixed effects and random effects models weighting each study by a measure of its precision, the inverse of the estimated variance.(4) The Mantel-Haenszel fixed effects model(5) assumes that there is a common effect of the Sp1 polymorphism across different studies and results differ only by chance. Heterogeneity was assessed by the χ2-based Q statistic and was deemed significant when p < 0.10.(4) The DerSimonian and Laird random effects model incorporates also the between-study heterogeneity to the within-study heterogeneity and does not assume that the polymorphism effect is common across different studies.(6) Generally, random effects estimates have wider confidence intervals (CIs) and thus are more conservative. When there is no detectable heterogeneity, the two models coincide. In the results section, random effect estimates are reported as the primary analysis. Fixed effects estimates also are provided for the main comparisons of interest.
When heterogeneity is suspected in a meta-analysis, it is important to assess whether parameters of clinical and/or biological importance or bias may be responsible for the diversity.(7) In this regard, we performed (i) a sensitivity analysis excluding two studies in which a few patients had been included in the fracture group, although they only had very low BMD but no fractures and (ii) subgroup analyses evaluating the strength of the Sp1 polymorphism effect on fracture risk separately for studies with different study designs (case-control, cross-sectional, cohort with population controls), different skeletal sites (vertebral fractures vs. other sites or mixed sites), and different mean ages in the fracture group (≥65 vs. <65 years).
Furthermore, we evaluated the potential for bias. Publication bias was probed by inspecting inverse funnel plots.(4) This method examines whether small studies with large variance give more favorable estimates for the postulated associations than larger studies with more precision in their results. Finally, recursive cumulative meta-analysis(8) examined whether the strength of the postulated associations has been changing markedly over time, as new studies have been published on this topic. In the presence of bias, an early study may show a large effect, but it may not be confirmed in subsequent publications.(8)
AF for fractures
The AF for fractures due to the Sp1 polymorphism for the population is estimated as
where PR1 and RR1 are the prevalence of and RR associated with Ss heterozygosity, while PR2 and RR2 are the prevalence of and RR associated with ss homozygosity. We used the prevalence and RR estimates for the overall meta-analysis population to obtain a value for the AF. We also calculated separate estimates for AF for cohort studies and for case-control studies.
Meta-analyses were performed in SPSS 10.0 (SPSS, Inc., Chicago, IL, USA) and in Meta-Analyst (Joseph Lau, Boston, MA, USA). All p values are two-tailed.
Characteristics of eligible studies
Twenty-two studies(1-3, 9-27) addressing the relationship between COLIA1 and BMD and/or fractures were retrieved. Eight studies were excluded from all analyses(9-16) for the following reasons: ecological analysis (n = 1),(9) Asian populations where the s allele was not detected in any patients (n = 2),(10, 11) premenopausal women without fracture data (apparently none were present; n = 3),(12-14) prepubertal children without fractures (n = 1),(15) and no mention of fracture data (n = 1).(16) Finally, one study(1) was an early publication of 55 patients with fractures and 55 age- and sex-matched controls; 47 of the fracture cases and 12 of these controls had been included in a subsequent publication with 93 patients with fractures and 88 age- and sex-matched controls. Only the latter study was included in the main analysis to avoid duplication of the majority of the fracture cases in the calculations. However, in a sensitivity analysis we evaluated the effect of adding also the results of the early publication.
The meta-analysis therefore included 13 studies (Table 1) in which both subjects with and without prevalent fractures had been included and COLIA1 genotyping had been performed.(2, 3, 17-27) Only one of these studies provided data on incident fractures as well,(17) showing a detrimental effect of the s allele(17); the same study did not show such a detrimental effect for prevalent fractures. One other study(24) had been based on a preexisting cohort established in 1977; selected subjects had been evaluated by spine radiography for vertebral fractures 18 years after the inception of the cohort. Although these also may be considered to be incident fractures, we included the study in the prevalent fracture calculations because the fracture determination had been performed in a cross-sectional fashion.
Table Table 1.. Studies Included in the Meta-Analysis
In total, 3641 subjects (3542 women) were included in the quantitative synthesis (2418 SS homozygotes [67.7%], 1086 Ss heterozygotes [29.8%], and 137 ss homozygotes [2.5%]). Subjects were enrolled in France (two studies), the United Kingdom (two studies), Denmark (two studies), the United States, Spain, Netherlands, Sweden, Greece, Belgium, and the Czech Republic. There were seven case-control studies,(2, 18, 21, 22, 25-27) four cross-sectional studies,(17, 20, 23, 24) and two comparisons against controls from the general local population.(3, 19) In the two studies with population controls, the fracture and nonfracture groups were not of similar age range. Age matching was successful in all the case-control studies and the mean age of the two groups was comparable in all the cross-sectional cohorts. In two studies,(19, 21) the group of women with fractures also included some women with severe osteoporosis without fractures; thus, sensitivity analyses were performed excluding these two studies to evaluate whether the pooled estimates and their heterogeneity would change.
Using DerSimonian and Laird random effects calculations, the risk for fractures was 1.25-fold (95% CI, 1.09-1.45) higher in the Ss group versus SS group (p = 0.17 test for heterogeneity by the Q statistic), 1.68-fold (95% CI, 1.35-2.10) in the ss group versus SS group (p = 0.16 for heterogeneity), and 1.35 (95% CI, 1.04-1.75) in the ss group versus Ss group (p = 0.08 for heterogeneity; Fig. 1). The fixed effects RRs were 1.21 (95% CI, 1.09-1.35), 1.41 (95% CI, 1.15-1.73), and 1.21 (95% CI, 0.99-1.50), respectively, and they are also shown in Fig. 1 for comparison. Generally, fixed effects estimates were smaller, but the difference was not prominent. Moreover, both fixed and random effects calculations suggested the presence of a dose effect.
The random effects ORs estimates were 1.43, 1.84, and 1.37, respectively, for the three comparisons and there was significant between-study heterogeneity for all three of them (p value for heterogeneity testing, 0.04, 0.07, and 0.10, respectively). The Mantel-Haenszel fixed effects ORs estimates were 1.37, 1.85, and 1.45, respectively.
The results were similar when we also included one study,(1) most of whose patients with fractures had been included in a later publication. The random effects RRs were 1.28 (95% CI, 1.12-1.47) for Ss heterozygotes versus SS homozygotes, 1.76 (95% CI, 1.44-2.14) for ss homozygotes versus SS homozygotes, and 1.38 (95% CI, 1.10-1.72) for ss homozygotes versus Ss heterozygotes. The inclusion of this study did not reduce the observed between-study heterogeneity (p value for heterogeneity testing, 0.07, 0.2, and 0.11, respectively).
The results also were similar when we excluded studies in which not all patients in the fracture group actually had fractures. The random effects RRs were 1.20 (95% CI, 1.03-1.39) for Ss heterozygotes versus SS homozygotes, 1.67 (95% CI, 1.30-2.13) for ss homozygotes versus SS homozygotes, and 1.42 (95% CI, 1.07-1.87) for ss homozygotes versus Ss heterozygotes. The exclusion of these two studies did not reduce the observed between-study heterogeneity among the remaining 11 studies (p value for heterogeneity testing, 0.09, 0.09, and 0.07, respectively).
As shown in Table 2, cross-sectional studies showed a clear effect only for ss homozygotes and studies with population controls showed a clear effect only for Ss heterozygotes, while case-control studies showed a more typical dose-response relationship of fracture risk with the number of s alleles. However, the CIs for specific study design subgroups were very wide, and these differences could have been caused by chance.
Table Table 2.. Subgroup Analyses for the Risk of Prevalent Fractures
Analyses limited to the eight studies focusing on vertebral fractures yielded more prominent effects for the Sp1 polymorphism, as compared with the five studies in which other skeletal sites were studied or both the spine and the other sites qualified. The difference between the two subgroups of studies was seen in all three genotype comparisons (Table 2), and it was formally statistically significant in the case of ss homozygotes versus SS homozygotes (p = 0.017).
Finally, there were trends for larger effect estimates in studies in which the mean age of the fracture group was <65 years than in studies of older subjects. However, the differences were not formally statistically significant and CIs overlapped substantially (Table 2).
In inverted funnel plots there was no asymmetry suggesting publication bias for any of the genotype comparisons (not shown). Similarly, no strong bias was discerned with recursive cumulative meta-analysis, although the first study published in the field(1) showed a larger protective effect than most of the subsequent studies. Specifically, the RRs observed in the first study were 1.61 for Ss versus SS, 2.21 for ss versus SS, and 1.39 for ss versus SS.(1) The estimated pooled effect of the Sp1 polymorphism has diminished slightly over time with the emergence of additional data because the most recent studies show less prominent Sp1 effects. However, the overall change has been small so far for all three genotype comparisons.
In the total meta-analysis cohort, the frequency of Ss was 29.8% and of ss was 2.5%. Using these frequency estimates and the random effects RRs (1.25 and 1.68), we estimate that the average AF of fractures caused by the COLIA1 s allele in the European and U.S. populations included in the meta-analysis would be approximately 9.4% (6.9% from Ss heterozygosity and 2.5% from ss homozygosity). When considering RR estimates from different study designs, the attributed fraction varied from 3.6% in cross-sectional studies to 8.9% in case-control studies and 9.7% in studies with population controls.
The meta-analysis documents that, overall, the s allele probably has an important association with the prevalence of fractures, in particular vertebral fractures. Although AFs need to be interpreted with caution, on average this allele may account for approximately 9% of vertebral fractures in the European/U.S. populations. We should caution that it is conceivable that the estimate of the AF analysis may be skewed by publication bias and it differed modestly according to study design. However, we did not observe strong evidence for publication bias in inverted funnel plot diagnostics, and the different estimates in case-control and cohort studies could have been caused by chance. The meta-analysis complements the previous finding that the incidence of nonvertebral fractures during follow-up seems to be increased among subjects with this allele.(17) Importantly, the effect of the Sp1 polymorphism on fracture risk seems to have a clear dose-response character: the risk increases stepwise from SS homozygotes to Ss heterozygotes and from Ss heterozygotes to ss homozygotes.
The clinical importance of the Sp1 polymorphism obviously would depend on its prevalence in different populations. Thus, the results of this meta-analysis should not be extrapolated outside Europe and the United States. The allele is absent, for example, in Asian populations.(10, 11) Moreover, the meta-analysis shows that there is probably substantial heterogeneity in the strength of the effect across different studies. This may reflect differences in study designs and eligibility criteria for the compared groups, while ethnic genetic differences also are possible.
The study design may influence the magnitude of the observed protective effects. Studies that do not match for age may provide biased estimates if the effects of Sp1 or any other polymorphism are age dependent. Age dependence also may create between-study heterogeneity in the estimates of matched studies targeting different age groups. In the meta-analysis we observed trends for larger detrimental effects of Ss and ss genotypes in patient populations with mean age less than 65 years than in older groups, although there were exceptions to this observation.(25) The observed difference may be caused by chance or may be susceptible to ecological bias given the nature of the data (group average rather than individual patient data). Future studies using individual patient data and adequately large sample size should try to dissect whether the detrimental effect of the s allele is more prominent in early postmenopausal women. One hypothesis may be that the s allele results in earlier occurrence of fractures, but the difference is gradually lost in very old populations in which fractures may accumulate even in subjects with more favorable genetic profiles.
We also observed that on average the detrimental effect of the s allele was more prominent in studies focusing on vertebral fractures. Detrimental effects also have been seen in fractures involving the wrists(27) but not in fractures involving the hips.(26) Although these differences also may be caused by chance, the meta-analysis suggests that this is another field in which differences are worthwhile dissecting in future studies.
In the meta-analysis we did not attempt to synthesize the available data on BMD from these studies. This is because such data were not measured and reported in the same way across studies and because we deemed that the effect on fractures is more important clinically than the effect on BMD. Moreover, it has been postulated that the Sp1 effect may be mediated in part by pathways other than its influence on BMD.(17)
Finally, despite the observed heterogeneity, the meta-analysis shows a clear dose-response effect dependent on the number of s alleles carried by an individual. This relationship would be hard to discern with clarity with the limited number of patients in any single study. Further studies also would need to address the independence and potential synergism/epistasis of the Sp1 effect with the postulated effects of other polymorphisms involving other genes that may regulate the osteoporosis process.(28-31) Given the subtlety and increasing complexity of genetic effects, large numbers of patients are required to achieve reasonable power to test complex hypotheses. Small studies may reach seemingly discordant results by chance alone and this may create confusion in the field. In this regard, meta-analysis of the evidence may offer a prime means of documentation of subtle, yet clinically potentially important, genetic associations in the field of osteoporosis.