Multiple sclerosis and fracture risk : traditional meta-analysis versus mega-analysis of individual patient data

Introduction The aim of this systematic review was to evaluate the difference between a traditional meta-analysis and a mega-analysis of individual patient data when combining observational studies. Materials and methods We used data from two studies that evaluated the risk of fracture in patients with multiple sclerosis using the British General Practice Research Database and the Danish National Health Registries. The published results were pooled together in an inverse-variance fixed effect metaanalysis. Using patient level data, we made the study populations as comparable as possible regarding the index date, calendar time, selection of incident/prevalent patient and follow-up. The individual patient data of these populations were combined in a mega-analysis. Cox proportional hazards models were used to estimate hazard ratios (HRs) of fracture, adjusted for shared confounders. Results A traditional meta-analysis of the original studies resulted in pooled adjusted hazard ratios of 1.13 [95%CI 1.03–1.23] for any fracture, hazard ratio 1.22 [95%CI 1.07–1.41] for osteoporotic fracture, and hazard ratio 2.47 [95%CI 1.72–3.53] for hip fracture. The mega-analysis of individual patient data showed an adjusted hazard ratio of 1.20 [95%CI 1.12–1.28] for any fracture, hazard ratio 1.36 [95%CI 1.24–1.50] for osteoporotic fracture, and hazard ratio 3.27 [95%CI 2.65–4.04] for hip fracture. The traditional meta-analysis of the original studies showed significant heterogeneity, which disappeared in a meta-analysis that pooled the two more comparable studies together. This meta-analysis yielded similar results as the mega-analysis with individual patient data. Conclusion A crucial step in performing a multicountry study is to reduce the level of heterogeneity between studies as much as possible before combining the data. Introduction Over the past years, an increasing number of multi-country studies have been performed in Europe. Different methods can be used to combine data from separate registries. The vast majority of published metaanalyses are based on aggregate patient data from completed studies1. Therefore, this type of meta-analysis can be done relatively quickly and easily. However, a ‘mega-analysis’ that combines individual patient data from various data sources has many theoretical advantages compared to a traditional meta-analysis, such as the ability to use common definitions, to explore heterogeneity at a patient level and to perform subgroup analyses of patient level data1,2. To date, studies that compare the two methods with actual patient data are scarce. The objective of the present study was to combine the results of two observational studies in (1) a traditional meta-analysis, by pooling the estimated risks from the published papers together, and (2) a megaanalysis of individual patient data, by using common definitions, selecting study populations with comparable patient characteristics and adjusting for the same set of confounders. Materials and methods Data sources We used data from two published articles that evaluated the risk of fracture in patients with multiple sclerosis (MS) using the British General Practice Research Database (GPRD), currently known as the Clinical Practice Research Datalink, and the Danish National Health Registries3,4. The GPRD comprises computerised medical records for over 10 million patients under the care of general practitioners (GPs) in the UK. The data recorded in the GPRD since 1987 include demographic information, prescription details, clinical events, hospital admissions and their major outcomes. A recent review of all validation studies found that medical data in the GPRD were generally * Corresponding author Email: f.devries@uu.nl 1 Utrecht Institute for Pharmaceutical Sciences, Utrecht University, The Netherlands 2 MRC Lifecourse Epidemiology Unit, University of Southampton, Southampton, UK 3 Clinical Practice Research Datalink, Medicines and Healthcare products Regulatory Agency, London, UK 4 National Institute of Public Health, University of Southern Denmark, Copenhagen, Denmark 5 Aalborg University, Aalborg, Denmark 6 MS Center Amsterdam, VU University Medical Center, Amsterdam, The Netherlands 7 The Danish Multiple Sclerosis Registry, Rigshospitalet, Copenhagen, Denmark 8 MS Clinic of Southern Jutland, Vejle, Sønderborg and Esbjerg, Denmark 9 Institute of Regional Health Service Research, University of Southern Denmark, Denmark 10 Department of Clinical Pharmacy and Toxicology, Maastricht University Medical Centre, Maastricht, The Netherlands


Introduction
Over the past years, an increasing number of multi-country studies have been performed in Europe.Different methods can be used to combine data from separate registries.The vast majority of published metaanalyses are based on aggregate patient data from completed studies 1 .Therefore, this type of meta-analysis can be done relatively quickly and easily.However, a 'mega-analysis' that combines individual patient data from various data sources has many theoretical advantages compared to a traditional meta-analysis, such as the ability to use common definitions, to explore heterogeneity at a patient level and to perform subgroup analyses of patient level data 1,2 .To date, studies that compare the two methods with actual patient data are scarce.
The objective of the present study was to combine the results of two observational studies in (1) a traditional meta-analysis, by pooling the estimated risks from the published papers together, and (2) a megaanalysis of individual patient data, by using common definitions, selecting study populations with comparable patient characteristics and adjusting for the same set of confounders.

Data sources
We used data from two published articles that evaluated the risk of fracture in patients with multiple sclerosis (MS) using the British General Practice Research Database (GPRD), currently known as the Clinical Practice Research Datalink, and the Danish National Health Registries 3,4 .
The GPRD comprises computerised medical records for over 10 million patients under the care of general practitioners (GPs) in the UK.The data recorded in the GPRD since 1987 include demographic information, prescription details, clinical events, hospital admissions and their major outcomes.A recent review of all validation studies found that medical data in the GPRD were generally Licensee OA Publishing London 2013.Creative Commons Attribution License (CC-BY) For citation purposes: Bazelier MT, Van Staa TP, Bentzen J, Vestergaard P, Uitdehaag BMJ, Leufkens HGM, et al.Multiple sclerosis and fracture risk: traditional meta-analysis versus mega-analysis of individual patient data.OA Epidemiology 2013 Jul 22;1(1):9.
Competing interests: none declared.Conflict of interests: declared in the article.
All authors contributed to conception and design, manuscript preparation, read and approved the final manuscript.
All authors abide by the Association for Medical Ethics (AME) ethical rules of disclosure.
to the time of diagnosis than to the time of the first symptom.We defined the index date as 31st December of the year of diagnosis in Denmark.
• Calendar time: we defined the start of the follow-up as 1 January 1996 in both the databases.In the GPRD study, this meant restricting the study population to those patients with a first MS record after 1 January 1996.In the Danish study, this meant selecting those patients with an MS diagnosis after 1 January 1996.• Selection of incident/prevalent MS patients: because the GPRD population captured a case-mix between incident and prevalent patients (reflected in a mean age of 44.9 years at index date), we similarly created a case-mix of incident and prevalent patients in the Danish population.This was done by adding prevalent patients to the study who were diagnosed from 1 January 1972 onwards; all patients were followed from 1 January 1996.The incident and prevalent patients together had a mean age at index date of 44.9 years (the same as in GPRD).The distribution of ages at the index date was also similar between these case-mix populations: the 10th percentiles of ages in GPRD were [18, 28, 33,  37, 40, 44, 47, 51, 56, 62, 95] and in Denmark they were [18, 29, 34, 38,  41, 45, 48, 51, 55, 61, 86].• Duration of follow-up: because the mean duration of follow-up was longer in the Danish registries than in GPRD, we cut off the follow-up time for patients in Denmark to a maximum time period of 5.7 years.This resulted in a mean duration of follow-up of 4.7 years in both groups of patients.

Outcome
In both the registries, the study outcome was the first fracture after the index date.The types of fracture were classified according to the In-the first record of MS after starting the valid data collection.Due to the dynamic character of the GPRD, the study population captured a casemix of incident and prevalent MS patients, reflected in a mean age of 44.9 years at index date 3 .
In Denmark, we had information about all patients (aged 18+) with an accepted diagnosis of MS in the DMSR between 1 January 1949 and 31 December 2007.To capture information on both fractures that were treated in an inpatient setting (recorded from 1977) as well as fractures that were treated in an outpatient setting (recorded from 1995), the beginning of the follow-up was from 1 January 1996.In the original study, incident MS patients were studied from the disease onset: the index date was defined as the 31st December of the year of the first symptom of the disease 4 .
All MS patients (from the GPRD and the Danish registries) were matched to six control individuals.For the GPRD, patients were matched by year of birth, sex, and practice; in Denmark they were matched by year of birth, sex, and region.In both registries, control persons were assigned the same index date as their matched MS patient.MS patients and controls were followed from their index date to the end of data collection, emigration, the date of transfer of the patient out of the practice area, or the patient's death.

Adjusted study populations
Using the individual patient data, we made the study populations from the GPRD and from Denmark as comparable as possible in terms of patient characteristics and definitions, as follows: • Index date: instead of following patients in the Danish population from the year of the first symptom of the disease, we now followed patients from the year of diagnosis because the first record of MS in the GPRD was probably closer of high quality 5 .The GPRD comprises a dynamic study population, where both patients and practices can enter and leave the database over calendar time.From the GPRD, we selected all patients with a record of MS between January 1987 and August 2009.GPRD data were linked to the national Hospital Episode Statistics (HES) for approximately 45% of all practices.We had access to HES data between April 1997 and March 2008.
In Denmark, separate registries of computerised medical records on all contacts to hospitals and on the use of drugs can be linked for the entire population and patients are followed from birth until death.Information on hospital admissions comes from the National Hospital Discharge Register 6 , which covers all inpatient contacts from 1977 onwards and from 1995 also all outpatient visits to hospitals, outpatient clinics, and accident and emergency rooms.In general, the validity of registrations is high 7 .The Danish Medicines Agency keeps a nation-wide register of all prescription drugs sold at pharmacies throughout the country from 1996 onwards 8 .For our data set, the end of data collection for all hospital and pharmacy records was December 2007.These data were linked to the Danish Multiple Sclerosis Registry (DMSR), a nation-wide database covering approximately 90% of all patients with MS, which started operating in 1949 9 .Patients' records are reviewed by a neurologist and classified according to their expertise 10 .The DMSR provided us with data until December 2007.

Original studies
The GPRD study included all patients aged 18 years or older with at least one recorded diagnosis of MS during the period of GPRD or HES data collection (1987-2009) 3 .Patients with a history of MS before the start of data collection were excluded.The index date was defined as the date of

Licensee OA Publishing London 2013. Creative Commons Attribution License (CC-BY)
For citation purposes: Bazelier MT, Van  Competing interests: none declared.Conflict of interests: declared in the article.
All authors contributed to conception and design, manuscript preparation, read and approved the final manuscript.
All authors abide by the Association for Medical Ethics (AME) ethical rules of disclosure.
tients and 15,436 controls.Their mean age at index date was 36.9 years, and patients were on average followed longer than in GPRD with a mean duration of 7.2 years.
were estimated using Cox models.We adjusted the HRs for confounders that were present in both registries, and for a binary variable indicating the data source (i.e.GPRD or the Danish registries).In addition, we conducted an inverse-variance fixed effect meta-analysis of the two more comparable populations.All statistical analyses were conducted using SAS ® 9.1/9.2 and Review Manager ® 5.1 software.

Results
Baseline characteristics of the original study populations from the GPRD and the Danish registries are presented in Table 1.The study population from the GPRD consisted of 5565 MS patients and 33,360 population-based controls.For these patients, the mean age at index date was 44.8 years and the mean duration of follow-up was 5.7 years.In Denmark, there were 2963 MS pa-ternational Classification of Diseases (ICD-10) categories.For any fracture we included S02, S12, S22, S32, S42, S52, S62, S72, S82, S92, T02, T08, T10, and T12.An osteoporotic fracture was defined as a fracture of the radius/ulna, humerus, rib, femur, hip, pelvis, or vertebrae.In both databases, the total followup period was divided into 30-day intervals.The presence of risk factors was assessed by reviewing the computerised medical records for any evidence of risk factors before the start of an interval.Potential confounders have been previously listed 3,4 .

Statistical analysis
We provided baseline characteristics of MS patients from the original study populations.In the original studies, Cox proportional hazards models had been used to provide an estimate of the relative risk (hazard ratio [HR]) of fracture among MS patients, adjusted for any potential confounders that changed the HR > 1% in an age-/sex-adjusted analysis.We conducted an inverse-variance fixed effect meta-analysis of these two original studies.In this analysis, the overall log HR is simply a weighted average of the individual log HRs, with the weights inversely proportional to the variance of the log HR of each study 11 .The crude and adjusted HRs were pooled for any, osteoporotic and hip fracture.
We then provided baseline characteristics of MS patients and controls from the more comparable study populations.Cox proportional hazards models were used to estimate HRs of fracture among these groups of patients (first, separately for GPRD and Denmark).A minimal confounder set was used for every fracture type, including risk factors that were present both in the GPRD as well as in the Danish registries.We combined the individual patient data of these more comparable populations in a mega-analysis.Fracture risks Table 6 shows that the difference in estimates between a traditional meta-analysis and an individual patient data mega-analysis varied between 6% (any fracture) and 32% (hip fracture).The megaanalysis showed an adjusted HR of 1.20 [1.12-1.28]for any fracture, HR 1.36 [1.24-1.50]for osteoporotic fracture, and HR 3.27 [2.65-4.04]for hip fracture.A meta-analysis that pooled the two more comparable studies together yielded similar results as a mega-analysis with individual patient data.When the fully adjusted HRs from Table 5 were pooled in a meta-analysis, the HRs were 1.20 [1.12-1.28]for any fracture, 1.36 [1.24-1.50]for osteoporotic fracture, and 3.31 [2.67-4.09]for hip fracture.In this meta-analysis, there was no significant heterogeneity between the results for any of the fracture types.

Discussion
The authors have referenced some of their own studies in this review.These referenced studies have been conducted in accordance with the Declaration of Helsinki (1964) and the protocols of these studies have been approved by the relevant ethics committees related to the institution in which they were performed.All human subjects, in these referenced studies, gave  4 shows the baseline characteristics of the adjusted study populations.In both the GPRD and the Danish population, MS patients were on average 44.9 years old at the index date and they had a mean duration of follow-up of 4.7 years.Patient characteristics were comparable between the original and the adjusted GPRD population.In the adjusted Danish population, MS patients were on average older than in the original study (44.9 years vs 36.9 years), and they had used more antidepressants, hypnotics/ anxiolytics and anticonvulsants at baseline (12.2%, 21.9% and 7.5% vs 7.0%, 9.5% and 3.1%, respectively).
The changes that were made to make the two populations more comparable resulted in comparable fracture risks for the GPRD study and in greater fracture risks for the Danish study (Table 5).For GPRD, the fully adjusted HR for hip fracture changed from 2.79 [1.83-4.26informed consent to participate in these studies.
The difference in estimates between a traditional meta-analysis of the original studies and an individual patient data mega-analysis, in which the differences between the study characteristics were reduced as much as possible, was 32% for hip fracture.A traditional meta-analysis of the original studies showed significant heterogeneity, which disappeared in a meta-analysis that pooled the two more comparable studies together.
The selection of a different patient group, including not only incident but also prevalent patients, changed the findings in the Danish study.This may be explained by increasing fracture risks with increasing age and worsening of the disease.In a previous study, we found that the risk of osteoporotic fracture increased with age and was most elevated for patients aged 50-59 12 .In another study, we found that the risk of osteoporotic fracture increased with increasing disability 4 .
There have been other examples where different study designs have led to different results, even with the same database.When two case-control studies, both performed in the GPRD, found different risks of fracture for patients using statins, a third study examined this discrepancy in results 13 .It was found that the ageband for matching cases and controls, the selection of potential confounders, the exclusion of high-risk patients, and different definitions for exposure time-windows explained different results in the two studies.Two studies on the use of oral bisphosphonates and risk of oesophageal cancer, both using data from the GPRD, also reached different conclusions 14,15 .They could, however, be explained by a difference in follow-up time and thus the potential to include people with longer durations of bisphosphonate use 14 .Other examples include the risk of venous thromboembolism with third generation Adjusted for age, sex, the use of oral/ iv glucocorticoids, antidepressants, hypnotics/anxiolytics, anticonvulsants, opioids in the previous six months, history of falling at index date, history of fracture at index date, history of fatigue in the previous six months, history of smoking and body mass index.d Adjusted for age, sex, the use of oral/ iv glucocorticoids, antidepressants, hypnotics/anxiolytics, anticonvulsants, opioids in the previous six months, history of cerebrovascular disease and epilepsy.
e Adjusted for age, sex, the use of antidepressants, antipsychotics, hypnotics/anxiolytics, anticonvulsants, opioids in the previous six months, history of cerebrovascular disease and epilepsy.f Adjusted for age, sex, the use of antidepressants, anticonvulsants in the previous six months and history of fracture at index date.All authors contributed to conception and design, manuscript preparation, read and approved the final manuscript.All authors abide by the Association for Medical Ethics (AME) ethical rules of disclosure.
contraceptives 16 , and the association between the use of proton pump inhibitors and fracture 17 .
There is a strong need for comparison between methods to conduct multi-country studies with electronic healthcare databases.The vast majority of published meta-analyses are based on aggregate patient data from completed studies 1 .A meta-analysis that uses individual patient data instead of aggregate patient data has many theoretical advantages 1,2,18,19 , such as the ability to use common definitions, to adjust for the same variables across studies, to explore heterogeneity at a patient level and to perform subgroup analyses of patient level data.In the analysis of clinical trial data, there have been a few examples where an aggregate meta-analysis failed to detect a result that was found with individual patient data 20 .To our knowledge, in observational electronic healthcare database research, peer-reviewed comparisons between the two methods are lacking.Our study is the first one to compare a traditional metaanalysis with an individual patient data mega-analysis of observational studies, using real life patient data.
Different patient characteristics between the GPRD and the Danish study caused significant heterogeneity in a meta-analysis that simply pooled the estimates from the original studies together.The most important difference between the studies was the start of follow-up: the Danish study followed incident MS patients from the first symptom of the disease, while the GPRD study comprised a case-mix of incident and prevalent MS patients.Heterogeneity disappeared in a meta-analysis that pooled results from more comparable patient groups together and this method showed similar results as an individual patient data mega-analysis.Because the case-mix between incident and prevalent patients elevated fracture risks compared to the incident Danish MS population, the meta-analysis, a small degree of heterogeneity may be accounted for using a random-effect meta-analysis, but this reduces the precision of the estimate and more importantly, cannot explore heterogeneity at a patient level.

Conclusion
A crucial step in performing a multicountry study is to reduce the level of heterogeneity between studies in terms of their design as much as possible, so that only a small amount of unexplained heterogeneity remains, before combining the data.Individual patient data are therefore a major advantage for a multi-country study.
Marloes T.  combined analysis of the GPRD and the Danish study obtained higher fracture risks than the combination of the original studies.Other differences that we synchronised between the two studies were calendar time, duration of follow-up and the selection of confounders.
A certain level of heterogeneity will always persist between studies that use different healthcare registries, even after making the best effort in reducing the differences.For example, patients with MS from the UK may be genetically different than MS patients from Denmark; guidelines for treatment of patients may slightly differ between countries, environmental factors may differently alter risks of fracture for MS patients, and residual confounding may alter estimated risks.In a traditional Staa TP, Bentzen J, Vestergaard P, Uitdehaag BMJ, Leufkens HGM, et al.Multiple sclerosis and fracture risk: traditional meta-analysis versus mega-analysis of individual patient data.OA Epidemiology 2013 Jul 22;1(1):9.

Table 1 Baseline characteristics of MS patients and controls from the original studies GPRD Denmark MS patients Controls MS patients Controls
risks of any, osteoporotic and hip fracture were not significantly increased, while for the adjusted population they were.We found an adjusted HR of 1.19[1.11-1.28]foranyfracture,an adjusted HR of 1.34 [1.20-1.49]forosteoporoticfracture, and an adjusted HR of 3.44 [2.73-4.34]forhipfracture.Table5further shows that fracture risks were comparable between the adjusted study populations.For both cohorts, there was a 1.2-fold increased risk of any fracture, a 1.3-to 1.5-fold increased risk of osteoporotic fracture and an approximately tripled risk of hip fracture.
Licensee OA Publishing London 2013.Creative Commons Attribution License (CC-BY)For citation purposes: Bazelier MT,VanStaa TP, Bentzen J, Vestergaard P, Uitdehaag BMJ, Leufkens HGM, et al.Multiple sclerosis and fracture risk: traditional meta-analysis versus mega-analysis of individual patient data.OA Epidemiology 2013 Jul 22;1(1):9.Competing interests: none declared.Conflict of interests: declared in the article.All authors contributed to conception and design, manuscript preparation, read and approved the final manuscript.All authors abide by the Association for Medical Ethics (AME) ethical rules of disclosure.the

Table 1 (Continued) Comorbidity ever before
] to 2.70[1.60-4.56].In the original Danish study, Licensee OA Publishing London 2013.Creative Commons Attribution License (CC-BY) For citation purposes: Bazelier MT,VanStaa TP, Bentzen J, Vestergaard P, Uitdehaag BMJ, Leufkens HGM, et al.Multiple sclerosis and fracture risk: traditional meta-analysis versus mega-analysis of individual patient data.OA Epidemiology 2013 Jul 22;1(1):9.Competing interests: none declared.Conflict of interests: declared in the article.All authors contributed to conception and design, manuscript preparation, read and approved the final manuscript.All authors abide by the Association for Medical Ethics (AME) ethical rules of disclosure.

Table 2 Risk of fracture in MS patients versus controls: original studies GPRD
a Adjusted for age, sex, the use of oral/iv glucocorticoids, antidepressants, hypnotics/anxiolytics, anticonvulsants, opioids in the previous six months, history of falling at index date, history of fracture at index date, history of cerebrovascular disease, epilepsy and history of smoking.b Adjusted for age, sex, the use of oral/iv glucocorticoids, antidepressants, hypnotics/anxiolytics, anticonvulsants, opioids in the previous six months, history of falling at index date, history of fracture at index date, history of cerebrovascular disease, epilepsy, history of smoking and body mass index.c

Table 3 Risk of fracture in MS patients versus controls: traditional meta-analysis of the original studies
a Significant heterogeneity between the two studies (p < 0.05).Licensee OA Publishing London 2013.Creative Commons Attribution License (CC-BY)For citation purposes: Bazelier MT,VanStaa TP, Bentzen J, Vestergaard P, Uitdehaag BMJ, Leufkens HGM, et al.Multiple sclerosis and fracture risk: traditional meta-analysis versus mega-analysis of individual patient data.OA Epidemiology 2013 Jul 22;1(1):9.Competing interests: none declared.Conflict of interests: declared in the article.

Table 4 Baseline characteristics of MS patients and controls from the adjusted populations GPRD Denmark MS patients Controls MS patients Controls
Licensee OA Publishing London 2013.Creative Commons Attribution License (CC-BY)For citation purposes: Bazelier MT,VanStaa TP, Bentzen J, Vestergaard P, Uitdehaag BMJ, Leufkens HGM, et al.Multiple sclerosis and fracture risk: traditional meta-analysis versus mega-analysis of individual patient data.OA Epidemiology 2013 Jul 22;1(1):9.Competing interests: none declared.Conflict of interests: declared in the article.All authors contributed to conception and design, manuscript preparation, read and approved the final manuscript.All authors abide by the Association for Medical Ethics (AME) ethical rules of disclosure.
Staa also works for the Clinical Practice Research Datalink (CPRD), UK.CPRD is owned by the UK Department of Health and operates within the Medicines and Healthcare prod-

Table 5 Risk of fracture in MS patients vs controls: adjusted populations
Adjusted for age, sex, the use of oral/iv glucocorticoids, antidepressants, hypnotics/anxiolytics, anticonvulsants, opioids in the previous six months, history of fracture at index date, history of cerebrovascular disease and epilepsy.b Adjusted for age, sex, the use of oral/iv glucocorticoids, antidepressants, hypnotics/anxiolytics, anticonvulsants, opioids in the previous six months and history of fracture at index date. a