Bone Quality Project

For both the Bioimaging and BTM Projects, the Project Team will request the following types of data for individual patients: age; height; weight; BMI; fracture history (vertebral and nonvertebral); time in study; fractures during study; treatment assignment; adherence to study medication; use of other osteoporosis medications. For BMD the Project Team will request thefollowing: site of BMD; date; and density values (adjusted, unadjusted, raw values as well as tscores, manufacturer, etc.). During the data acquisition phase of the study, the data management programmer will create a set of standard variables and work with each data set to yield variables that can be pooled in the analysis across all studies. For both projects, there will be two types of analyses: individual-level analyses which will be based on individual patients pooled across all studies and study-level analyses in which the unit of analysis will be the study. The individual-level analysis will be the primary analysis.During the first 6 months of the project, a detailed analysis plan will be developed by the Project Team. This plan will include criteria for inclusion/exclusion of specific data based on relevance, availability of fracture data and a formal study quality grading analysis. Otherquestions will be addressed such as analysis methods for trials of differing length. In this phase, a number of important questions about analysis of BTM's such as developing methods for pooling different markers and adjustments of differing marker assays. This plan will alsodelineate specific publications to be developed to address the primary as well as secondary objectives.Analysis for Bioimaging ProjectThe Bioimaging Project will focus on examining the relationship of change in QCT-derived parameters, including FEA, to reduction in risk. The overall goal is to examine whether QCT can qualify as a biomarker for bone strength. This project will include testing of the following hypotheses:H1: There will be a significant relationship between change in QCT-derived density and structural parameters and reduction in fracture risk for non-vertebral, hip and morphometric vertebral fracture. Change in QCT density and structure will explain a significant portion of the variation in fracture reduction seen with treatment. Using analytic methods described below, this analysis will be performed using pooled data from individual patients. We will look at changes in overall volumetric bone density in the hip, as well as specific structural parameters including bone size and cortical volume derived from QCT scans, and their relationship to reduction in fracture risk. As described below, the primary summary statistic will be Proportion of Treatment Explained (PTE) which will be examined forstatistical significance for each parameter. As a secondary analysis, a study-level analysis will be performed in which the overall change in QCT-derived parameters versus reduction in fracture risk will be plotted and used in a meta-regression as has previously been done for DXA BMD and fracture risk (Cummings et al. 2002; Hochberg et al. 2002). This will provide more descriptive information about the relationship between change in QCT-derived parameters and reductions in fracture risk.H2: There will be a significant relationship between change in bone strength, as estimated by FEA and observed reduction in fracture risk. FEA-estimated bone strength will explain a significant portion of the variation in fracture reduction seen with all agents. This analysis will be performed using parallel methods to that described for H1. PTE will be theprimary outcome measure.H3: Change in FEA strength will explain a higher proportion of the variation in fracture reduction than either QCT density/bone structure or DXA-derived areal BMD. Using the methods described below, we will perform this analysis to compare changes in FEA estimated bone strength to more traditional QCT parameters and also to DXA-derived areal BMD in the magnitude of their relationship to change in fracture risk.H4: A standardized analysis program applied to these QCT scans will reduce the variation in QCT response to treatment compared to the varying types of analysis programs used in the original analyses. We will compare changes in QCT density and bone structure as derived from the original analyses of the QCT scans to parallel changes derived from the our new analyses which will be performed under more carefully controlled conditions and in a standardized way across all studies. We will look the correlation of changes from the original analyses and the new analyses but also focus on the variation of changes as a measure of precision. These analyses will be performed in the overall pooled data set but will also examine the influence of factors such as acquisition parameters on the improvement in precision and PTE.Analysis for Bone Turnover MarkersH1: There will be a significant relationship between greater short-term changes in BTMs and reduction in fracture risk. Change in BTMs will explain a significant portion of the variation in fracture reduction seen with all agents. The primary analysis for this hypothesis will be an individual-level meta-analysis pooling multiple studies in which the short term percent change in the BTM variable (e.g., serum PINP, urine CTX) is modeled as the independent variable and risk of fracture is modeled as the dependent variable. These analyses will be performed separately for three fracture types: spine fractures (morphometrically defined); non-vertebral fractures; and hip fractures. Recognizing the important limitation of the difference between the nature of BTM changes fororal bisphosphonates compared to other antiresorptive treatments with non-continuous marker changes (e.g., denosumab and zoledronic acid), the primary analyses will be limited to oral bisphosphonate studies using BTM change/level 6 to 12 weeks after the start of therapy. Secondary analyses will include all available antiresorptive trials and additional analyses will attempt to account for different follow-up marker intervals and patterns. Additional analyses using other methods to model the extent and duration of BTM reduction, such as “area under the curve” and time dependent effects, will be assessed. We also plan to assess the utility of combinations of BTMs, such as the standardized ratio of levels of formation to resorption and other linear combinations of BTMs.Separate models will be constructed for specific BTMs (such as urine CTX or serum PINP) when feasible, and secondary analyses will pool categories of BTMs (such as serum formation markers or urine resorption markers) after transformation to facilitate standardization when identical BTMs across studies are not available. These pooled analyses will be performedamong both study participants with BTM measurements, and as discussed below, among all trial participants using multiple imputation.As with the Bioimaging aims, the primary summary statistic will be Proportion of Treatment Explained, or PTE, which will be examined for statistical significance for each BTM parameter. A study-level secondary analysis will be performed to provide more descriptive informationabout the relationship between change in BTM parameters and reductions in fracture risk.H2: The relationships between treatment-related changes in BTMs and fracture risk will persist after accounting for the effects of bone mass and other clinical characteristics. To determine if the relationships between short term treatment-related changes in BTM and fracture outcomes are independent of other potential predictors (such as pre-treatment BMD, change in BMD and clinical risk factors for fracture), we will determine the effect of stepwise addition of each to the primary pooled models constructed above. If after further adjustment the relationships between change in BTMs and fracture are not substantially attenuated, we will infer that they are independent. H3: Among bisphosphonates, the relationships between treatment-related changes in BTMs and fracture risk will not differ by dosing interval, route of administration, or specific BTM. In stratified secondary analyses we will search for evidence that the relationships between change in BTM and fracture differ among dosing regimens, routes of administration, andspecific BTM. Apparent differences will be formally tested with models that contain interaction terms for the strata of interest.H4: The fracture prediction ability of specific treatment-related change in BTM cutpoints (such as a 30% or 50% reduction in bone resorption marker) will differ, informing selection of a preferred cutpoint for clinical use.Using the pooled subject-level data from bisphoshonate trials, we will test the utility of specific change in BTM cutpoints for the optimal prediction of fracture outcomes among treated individuals.H5: Patterns of changes in formation vs. resorption BTMs and duration of those changes will be predictive of efficacy for fracture risk, BMD or bone strength increase for anabolic or nontraditional antiresorptive therapies and for therapies with novel mechanisms of action. In this exploratory aim, which will be limited by the number of fracture trials among nonbisphosphonate treatments, we will use therapy-specific pooled subject-level data from anabolic and other osteoporosis treatments to assess the magnitude and patterns of shortterm change BTMs, and where possible relate them to fractures and other skeletal outcomes of interest. We recognize the possibility that the relationships between treatment-related changesin BTMs and subsequent skeletal outcomes may differ substantially among various antifracture treatments. These treatment specific differences will limit our ability to use BTM data derived from one treatment category to predict relevant outcomes with another, but will provide important descriptive data and provide some insight into mechanisms of action.Statistical methods for analysis of individual-level results: (note that in this section, “biomarker” refers to both imaging and bone turnover markers). The basis for many of the analyses in both Project 1 and Project 2 is to fit an individual patient-level meta-analysis (Riley et al. 2010),incorporating random treatment effects for each trial. This will allow us to calculate PTE explained by various biomarkers (including BMD, imaging biomarkers or BTMs) and also to assess the association of biomarkers with fracture risk at the individual-level.Fracture outcomes (other than morphometric vertebral fracture) will be analyzed by a Coxregression analysis as in the original studies. The basic model (without a biomarker) for participant j in trial i for the log hazard rate will be (with the subscript O for outcome):logHR=x(ij)Bo+b(0,i)+y(0)T(j)+d(0,i)T(j), (1)where x(ij) represents covariates for adjustment (e.g., adjusting for bone mass), b(O,i) is a random intercept, is the overall treatment effect (with T(j) =1 indicating treatment and 0 otherwise), and dO,i is the trial specific treatment effect (also a random effect). To assess the percent oftreatment explained by the biomarker we will additionally fit the model below (with the subscript M for biomarker) and with M(ij) denoting a measured biomarker. This biomarker can be imaging based (e.g., QCT or FEA) or based on BTMs. For imaging markers, the focus will be on changes from the start to end of study. For BTMs, and for changes in the early phase of thestudy as well as the current values of the biomarkers.logHR=x(i,j)B(m)+b(M,i)+y(m)T(j)+d(m,i)T(j)+v(M)M(ij), (2)Our primary quantity of interest will be the Proportion of Treatment Explained, defined as PTE=1-(Ym/Y0). The above models can be fit either using a frailty approach (Rondeau et al. 2012) or using robuststandard errors (Lin et al. 1989). Based on the above model fits and bias-corrected bootstrapping, we will derive confidence intervals for the PTE.Details and variations on the basic analysis1. We wish to use all the fracture data for each trial, but will have biomarkers measured on only a subset of the data. To address this issue we will first use multiple imputation to fill in missing biomarker values (Schafer 1997). Multiple imputation will allow us to use the full outcome and biomarker data but will fairly represent the amount of information actually available to measure the association.2. A meta-analysis at the study-level corresponds to using multiple imputation with a simplistic and single imputation analysis, followed by averaging up to the study-level. So multiple imputation will give us more flexibility to correct situations in which the biomarkers are measured in a selected subset. For example, some studies measure biomarkers only inparticipants who are adherent to study medication, whereas other studies measure the biomarkers in random subsets. By modeling the relationship between outcome and adherence, we can use multiple imputation to correct for selection bias.3. Via bootstrapping we can also compare the PTE across different biomarkers or cutpoints of biomarkers (by deriving a bootstrap CI for the difference in PTE).4. We can include bone mass and other clinical characteristics as covariates in models (1) and (2) to check for persistence of PTE. 5. The coefficient, for the biomarker in (2) measures the influence of the biomarker on fractureoutcome after adjusting for treatment group. As such it describes the individual-level effect of the biomarker on outcome.6. We can allow the coefficient, for the biomarker in (2) to be dependent on biomarker type (after standardization of the biomarkers) to test for equivalence of biomarker effects, either directly in the coefficient or its effect on PTE. This will be achieved by “stacking” all the standardized biomarker values into a single predictor and creating a subsidiary predictor to keep track of biomarker type. A test of the interaction between the standardized biomarker and biomarker type measures is a test of equivalence of biomarker effects on fracture.7. We can allow the coefficients, and in (1) and (2) to be dependent on route of administration or resorption class to test for their effects on PTE, by including interaction terms.8. Model (2) can be modified to allow linear combinations of more than a single biomarker to check the combined effects of multiple biomarkers on PTE. Using 10 times 10-fold crossvalidation (Witte et al. 2011) we will assess which combination of biomarkers explains the maximal PTE. Cross-validation allows unbiased assessment of the PTE without danger of overfitting (as opposed to model selection based on statistical significance, which is susceptible to over-fitting). Using cross-validation we can assess whether BTMs add over and above imaging based biomarkers (as well as vice-versa) and whether a combination of the two does betterthan either alone. The total number of biomarkers is small enough that we can exhaustively search for inclusion/exclusion of all possible biomarkers in model (2).Study-level analyses: In these secondary analyses, there will be one data point in the regression representing each study as has been done in several previous meta-analyses (Cummings et al. 2002; Hochberg et al. 2002). The reduction in fracture risk will be derived based on the entire study sample while the percent change in the biomarker will be based on the sample in which data are available. For DXA BMD in most studies, the sample will be of similar size to the entire study since DXA is generally collected on most participants. However, for the biochemical markers and for the QCT-derived variables the number contributing to the percent change will represent only a small proportion of the overall study population. We will summarize therelationship between change in imaging and reduction in fracture risk and present this regression visually.Power/sample size: The sample size will be determined by the number of studies and participants per study for which FNIH can acquire data on behalf of the Project Team. So at issue is the accurate assessment of the power or minimal detectable effect sizes. For a metaanalysis this depends on both the within-study accuracy of the estimated parameters (primarily the PTE as noted above) as well as study-to-study variation in those parameters. While there are isolated previous publications that might provide single-study preliminary data for one or another of our novel biomarkers (e.g., FEA or some of the BTMs) there are not a series studies that can be assembled to assess the study to study variation in treatment effects and, likely, PTE, as modeled with the random treatment effects in equations (1) and (2) and which form a crucial part of the analysis. That said, previous meta-analyses of BMD have found highly statistically significant associations. We expect that our study, will offer even higher precision due both to the increased number of studies and our individual patient-level meta-analysis.We calculated power for the bioimaging analysis using a simulation study. Because power calculation algorithms are not available for individual patient data meta-analyses and also not readily available for percent treatment explained (PTE), we wrote a simulation program to estimate the power for individual patient data meta-analyses for the bioimaging project. We examined power under two scenarios. First, we assumed that we would have 24 studies with sample sizes ranging from 1,000 to 7,000 (uniformly distributed) with an average of 4,000.This should closely approximate the meta-analysis examining change in DXA with fracture risk reduction assuming we are able to obtain most of the studies listed in appendix 1. Under a more conservative scenario, we also simulated with only 12 studies with an average of 1250 patients. This more closely approximates the situation where we have QCT on a more limitedset of patients (and we are able to gain some power by using multiple imputation for missing QCT values) or where we are only able to obtain DXA data from about half of the studies in Appendix 1. In both scenarios, half of the participants were assumed to be on treatment. The values for parameters in the simulation were derived from the raw data files for three RCT's (FIT vertebral fracture arm (Black et al. 1996), FIT clinical fracture arm (Cummings et al. 1998) and HORIZON PFT (Black et al. 2007)). Each of these studies includes fracture incidence (clinical fractures for this simulation) and percent change in femoral neck BMD (via DXA). The model weused for the risk of fracture in study i as a function of both treatment and change in DXA was the following:"logit" (p_ij )=?_0i+?_1i ?"trt" ?_ij "+" ?_2i ?"ddexa" ?_ij,where " " p_ij is the probability of fracture for participant j in study i, trtij identifies if the participant is on treatment or not, and ddexaij is the change in DXA value. Note that the parameters (and hence associations) are allowed to vary across studies. Furthermore, thechange in DXA was assumed to respond to treatment:?"ddexa" ?_ij~N(?_0i+?_1i ?"trt" ?_ij,s_d^2).From this it is possible to work out theoretically a close approximation to the PTE as ?PTE ?_i "= " (?_1i+?_2i (?_1i-?_0i )-?_1i ?_i)/(?_1i+?_2i (?_1i-?_0i ) ),where ?_i=v(1+(3?_2i^2 s_di^2)/p^2 ).Except for the PTE, the values for each of above parameters (e.g., the ?_ki) were calculated from the three studies. Then, to be conservative, the simulation used a range of values about 25% wider than those encountered in the preliminary data. The PTE was varied systematically to assess change in power.To each simulated dataset we calculated both a study level meta-analysis and a patient data level meta-analysis. Since PTE is calculated from two separate fits to the same dataset (one adjusting and one not adjusting for the mediator), it can be challenging to derive standard errors. To derive standard errors for PTE for the study-specific summaries in the study-levelmeta-analysis, which is required for combining the estimates, we used a “seemingly unrelated regression” routine to calculate the standard errors. For the patient data level meta-analysis, we created a stacked dataset (Freedman et al. 1992) that allows direct estimation of the PTE andcalculated the standard errors by jackknifing and the delta method. In this balanced data situation without other patient level factors or missing data, the power for the two procedures was comparable. All calculations were performed using Stata (Version 12, StataCorp, College Station, TX).The power curves are given in the figure and show that there is excellent power over a range of small PTE's. For the 24 study scenario, power is >80% even when the PTE is as small as 15% (blue line). For the 12 study scenario, power is over 80% for PTE as small as 18% and over 90% if PTE is > 22% (red line). Since we expect that the true PTE is much larger, we believe thatpower is high for any plausibly interesting PTE.