Systematic review of methods for individual patient data meta analysis with binary outcomes
 Doneal Thomas^{1},
 Sanyath Radji^{4} and
 Andrea Benedetti^{1, 2, 3, 5}Email author
DOI: 10.1186/147122881479
© Thomas et al.; licensee BioMed Central Ltd. 2014
Received: 5 November 2013
Accepted: 11 June 2014
Published: 19 June 2014
Abstract
Background
Metaanalyses (MA) based on individual patient data (IPD) are regarded as the gold standard for metaanalyses and are becoming increasingly common, having several advantages over metaanalyses of summary statistics. These analyses are being undertaken in an increasing diversity of settings, often having a binary outcome. In a previous systematic review of articles published between 1999–2001, the statistical approach was seldom reported in sufficient detail, and the outcome was binary in 32% of the studies considered. Here, we explore statistical methods used for IPDMA of binary outcomes only, a decade later.
Methods
We selected 56 articles, published in 2011 that presented results from an individual patient data metaanalysis. Of these, 26 considered a binary outcome. Here, we review 26 IPDMA published during 2011 to consider: the goal of the study and reason for conducting an IPDMA, whether they obtained all the data they sought, the approach used in their analysis, for instance, a twostage or a one stage model, and the assumption of fixed or random effects. We also investigated how heterogeneity across studies was described and how studies investigated the effects of covariates.
Results
19 of the 26 IPDMA used a onestage approach. 9 IPDMA used a onestage random treatmenteffect logistic regression model, allowing the treatment effect to vary across studies. Twelve IPDMA presented some form of statistic to measure heterogeneity across studies, though these were usually calculated using twostage approach. Subgroup analyses were undertaken in all IPDMA that aimed to estimate a treatment effect or safety of a treatment,. Sixteen metaanalyses obtained 90% or more of the patients sought.
Conclusion
Evidence from this systematic review shows that the use of binary outcomes in assessing the effects of health care problems has increased, with random effects logistic regression the most common method of analysis. Methods are still often not reported in enough detail. Results also show that heterogeneity of treatment effects is discussed in most applications.
Keywords
Individual patient data Metaanalysis Random effects Systematic review Heterogeneity OnestageBackground
A metaanalysis (MA) attempts to synthesize the results from various distinct studies. The goal is to summarize the evidence for a particular statistical measure of interest, such as a risk difference or odds ratio. It is an especially important tool in clinical practice and medical research, where evidencebased information is preferred [1].
Individual patient data (IPD) MA are the gold standard of metaanalysis. In an IPDMA linebyline patient data are collected from the relevant studies, rather than just the measure of effect as in a standard aggregate data (AD) MA. This permits researchers to define exposures and outcomes consistently across studies, and to analyze them more similarly (e.g. adjusting for the same confounders), which may minimize heterogeneity [2, 3].
For IPDMA, two broad analytic strategies (one and twostep approaches) are possible; both preserve the clustering of subjects within studies, comparability of study arms, and both may be either fixed or random. A fixed effects analysis assumes that the estimated effect is the same across all studies, while a random effects analysis assumes that the estimated effect varies across studies due to differences in patient populations, study procedures, etc [1, 4].
A twostep approach first analyzes each study separately and as identically as possible, and then uses standard metaanalytic techniques to pool the measure of interest. The wellknown random effects method of Der Simonian and Laird is frequently used in the second step of a twostep IPDMA approach [1].
One step approaches use one statistical model while accounting for the clustering among patients in the same study, to estimate an overall effect. A one step model also takes advantage of the ability to standardize elements of the analysis across studies, but offers more flexibility to explore the differences that may exist between patients in the same study as well as across studies [2, 3, 5]. In particular, a onestep approach allows better control of confounding by patient and study level covariates, improves power for detecting interactions and subgroup analyses, as well as avoids and reduces the potential for ecological bias that may occur if group level information is included in the analysis [6, 7].
In conventional ADMA, it is difficult to estimate the effects of patientlevel covariates on the treatment effect [8, 9]. In the context of an ADMA, this is known as metaregression and may use study level covariates or aggregated patient level information. Metaregressions are prone to ecological bias, and to confounding from variables not included in the model [5, 6, 9] and may have limited power. IPDMA have higher power than metaregression to detect the effect of an interaction between covariates and treatment, and are preferable when the interest is in estimating interactions with patientlevel covariates [9–11].
Importantly, IPDMA are not prone to ecological bias if inferences about individuals are not based on aggregated data and model misspecification is evaded [6]. For these reasons, and others, IPDMA are considered the gold standard of metaanalysis, despite the complexity and cost of collecting the data, and are published with increasing frequency [2].
Despite the many advantages, the wide range of methods used for analysis of IPDMA and the lack of a standardized data analysis plan is a serious drawback [12, 13]. A previous review of methods used in practice for IPDMA, reviewed 44 articles published during 1999–2001, of which 14 considered a binary outcome [13]. That review found that the twostep approach was used about twothirds of the time [13].
The aim of this systematic review is to update that report, nearly a decade later when random effects models have been well integrated into other areas of health research, are readily available in many software packages and computing power is also up to the challenge. Our objective was to investigate the statistical approach taken to analyze IPDMA with binary outcomes. In particular, we were interested in (i) whether twostage or onestage approaches were more common; (ii) how heterogeneity was investigated and reported; and (iii) if a one step approach was used, were intercepts permitted to vary across primary studies considered as random.
Methods
Eligibility criteria for included studies were articles published in 2011 that reported results of an individual patient data metaanalysis for a binary outcome and were indexed in PUBMED or Medline. We believed that this would provide a good overview of the methods currently used for analysis of IPDMA. We performed the search in June 2012.
We searched in PUBMED and MEDLINE for articles published between January 1, 2011 and December 30, 2011. The search terms used were “meta analysis” and (“individual patient data” or “ipd” or “patient level” or “individual participant” or “integrated analysis”). The titles and abstracts of these articles were reviewed to ensure that they reported results of an IPDMA.
For the full text review, a standardized form was filled independently by two reviewers (SR, DT). Discordant entries were resolved by a third reviewer (AB). The data we collected from each article included: the reason for performing an IPDMA, the goal of the IPDMA, the types of studies collected, the number of studies sought and retrieved; the number of patients sought and retrieved; the type of outcome (e.g., binary, timetoevent or continuous); the method of analysis for the primary outcome and whether the analytic approach was onestage or twostage; whether intercept and/or the treatment effect were allowed to vary across studies (fixed or random effects); how heterogeneity was quantified, addressed and reported; the method of analysis of covariates: whether by one or twostage methods; methods for study or patientlevel covariates; and, whether subgroup analyses were performed (See Additional file 1: Table S1). For this review, we have considered only those articles which used a binary outcome.
We present descriptive analyses only.
Results
Twentyseven articles presented timetoevent outcome data, 2 presented continuous outcome data and only one article had a count outcome. We focus on the 26 articles that presented results using a binary outcome.
Goal of study, overall and stratified according to whether the IPDMA included only randomized controlled trials, or included both randomized controlled trials and observational studies ^{ 1 }
Reason  Included only randomized controlled trials (n = 15) N (%)  Included observational studies (n = 11) N (%)  Overall N (%) 

To estimate a treatment effect  10 (67%)  3 (27%)  13 (50%) 
To investigate safety of a treatment  2 (13%)  1 (9%)  3 (12%) 
To estimate diagnostic accuracy  1 (7%)  4 (36%)  5 (19%) 
To identify predictors  1 (7%)  3 (27%)  4 (15%) 
Other/Unclear  2 (13%)  1 (9%)  3 (12%) 
Over half of IPDMA (15/26) included only randomized control trials while the other IPDMA included only observational studies. IPDMA that included observational studies had a different profile in terms of goal with a greater proportion of studies that aimed to estimated diagnostic accuracy, and fewer IPDMA that aimed to estimate the effect or safety of a treatment (See Table 1).
Why IPD?
Reasons provided to support conducting an IPD ^{ 1 }
Reason  N (%) 

To perform subgroup analyses  13 (50%) 
To improve consistency across studies (in terms of inclusion criteria, outcome definition, etc.)  4 (15%) 
To consider other outcomes  4 (15%) 
To adjust for confounding variables  1 (4%) 
To estimate diagnostic accuracy  5 (19%) 
To identify predictors of an outcome  2 (8%) 
Unclear  6 (23%) 
Numbers of studies and patients
Statistical methods
Statistical analysis method categorized by overall strategy among 26 IPD metaanalyses of binary outcomes
Analytic approach^{1}  n/N (%)  

One stage approach (n = 19)  
Ignored clustering by study  Logistic regression  5/19 (26%) 
Fixed effects  Logistic regression  4/19 (21%) 
Random effects  Logistic regression  10/19 (52%) 
Fixed study effect with random treatment effect^{2}  1/10 (10%)  
Random study effect with fixed treatment effect^{2}  2/10 (20%)  
Random study effect with random treatment effect^{2}  2/10 (20%)  
Unclear^{1}  5/10 (50%)  
Two stage approach (n = 6)  
Fixed effects  Unspecified  2/6 (33%) 
CochraneMantelHaenszel  1/6 (17%)  
Random effects  Der Simonian Laird  2/6 (33%) 
Unspecified  1/6 (17%) 
Among the 19 onestage analyses, logistic regression was the most frequent technique employed. Ten of these IPDMA used a random effects analysis. However, in 5 of these it was not clear whether intercepts, treatment effects or both were allowed to vary across studies. In the remaining 5 IPDMA, 2 allowed both intercepts and treatment effects to vary, 1 allowed only the treatment effect to vary, and 2 allowed only the intercepts to vary. In general, little justification was offered for these choices. None specified the estimation method (e.g. penalized quasilikelihood (PQL) [40] or adaptive Gaussian Hermite quadrature [41], etc.) used.
A fixed effects onestage approach was used in 9 IPDMA. Of these, 5 IPDMA seemed to ignore clustering of subjects by study completely, and pooled all subjects together.
Twostage methods were used in 6 of 26 studies reviewed. Of these, three studies used random effects for the treatment. One study initially used a Der Simonian Laird approach, but due to very low estimated heterogeneity, used a fixed treatment effect. The CochraneMantelHaenszel twostage approach was used in one study, where no indication of heterogeneity across studies was found.
Heterogeneity
Statistic used to measure heterogeneity among studies in the 26 IPD metaanalyses stratified by analytic approaches
Statistics  

Q Statistics  I^{2}  Multiple statistics  Other measures  Unclear  
(N = 6)  (N = 6)  (N = 2)  (N = 6)  (N = 6)  
n (%)  n (%)  n (%)  n (%)  n (%)  
Onestep  3 (50)  4 (67)  0 (0)  6 (100)  6 (100) 
Twostep  2 (33)  2 (33)  2 (100)  0 (0)  0 (0) 
Covariates
Covariates were used in three ways: (i) to assess subgroup effects; (ii) to adjust a treatment effect for possible confounders; and (iii) to identify predictors of an outcome.
Among the 16 studies where the goal of the IPDMA was to estimate a treatment effect or the safety of a treatment, all considered subgroup analyses. Among studies that reported the number of subgroups considered, the median number of subgroups investigated was 2.5, with a range from 1–15. In all but one case, subgroups were formed by using categorical variables or categorizing a continuous variable. In one study, an interaction between the treatment and a continuous or ordinal risk score was evaluated. The subgroups investigated were based on patientlevel characteristics in 13 IPDMA, and on both patient and studylevel characteristics in 3 IPDMA.
Among the studies that used a onestage approach, 9/10 included interaction terms in the model, and presented stratum specific estimates as well as a pvalue for the interaction. Among studies that used a two stage approach, 5/6 presented the stratum specific effect estimates, and 5/6 presented a pvalue for the interaction. In two cases this pvalue was calculated as described in [43].
Among the 3 IPDMA that included observational studies and aimed to estimate a treatment effect or safety, all three adjusted for potential patientlevel confounders. One of these studies used a twostep approach first adjusting for confounders in each study separately then pooling the adjusted effect estimates. Among the IPDMA that only included randomized trials, and aimed to estimate a treatment effect or safety (n = 13), only 2 adjusted for patient level confounders. They did so by including them in a one stage model.
Finally, of the four IPDMA that aimed to identify predictors of an outcome, three included observational studies.
Missing data
While there are a number of approaches that could be taken to deal with missing data, 16/26 IPDMA did not report how missing data were handled. Three studies used multiple imputation and two studies used single imputation. The remaining studies used a variety of other approaches to dealing with missing data including excluding subjects with missing data, or excluding variables with too much missing data, or it was unclear what approach was taken.
Discussion
In this paper, we reviewed a sample of published individual patient data metaanalyses where the primary outcome was dichotomous, focusing on the statistical approach taken and results reported. To identify relevant articles in our review, we used a thorough search strategy and assessed 26 IPD MA articles published in the year 2011 that presented results for a binary outcome. It is possible that some relevant papers that reported the results of IPD MA with binary outcomes and were published in 2011 have been missed or excluded unintentionally, but these would be unlikely to differ substantially methodologically than those included. Two reviewers extracted all information independently and a third reviewer resolved conflicts. It might also be possible due to the lack of sufficient details to distinguish the methods used, that methods were incorrectly classified since the precise method used was sometimes inferred.
This review also highlighted the strengths and weaknesses of individual patient data metaanalyses (IPDMA) where the outcome was binary. IPDMA are clearly the gold standard of metaanalytic methods and publications featuring results from IPDMA are growing steadily in recent years. However, there are considerable variations in the methodology employed, for instance, the use of fixed or random effects for the estimated effect measures, measures of heterogeneity and strategies used to estimate treatment effects. In many studies, the statistical aspects were not clearly reported, with insufficient details provided to distinguish the methods used. Most times, little justification was given for the approaches taken in the studies, perhaps due to the lack of specific guidelines available for the IPD metaanalysis of binary outcomes. While guidelines exist for the reporting of systematic reviews and metaanalyses, these guidelines are not specific to IPDMA. For example, the PRISMA guideline #14 suggests that the methods of handling data and combining results, including measures of heterogeneity be described [44]. Extending those guidelines to encompass issues specific to IPD MA, such as stating if a one or twostage approach was used, would likely improve the reporting of IPD metaanalyses of binary outcomes.
In a previous systematic review of articles published in 1999–2001 [13], 14 (32%) of the IPD MA dealt with a binary outcome. While the proportion was similar, we found nearly twice the number of IPDMA of a binary outcome in just one year in 2011.
This review of 26 IPD metaanalyses of binary outcome encouragingly shows that practitioners often obtain a large proportion of the IPD required. IPD from 90% or more of the total number of studies were obtained in 62% of IPD studies, an important improvement to the 41% found in the previous review [13].
We found that more than half (73%) of studies did not use a twostep approach (i.e. analyzing each study separately and as identically as possible and pooling via standard meta analytic methods) but instead used the more flexible onestage method. This finding was contrary to the previous review [13], in which most analyses were performed using a twostage approach (82%) with little consideration of the onestep approach. This finding likely reflects the greater comfort with randomeffects models for binary outcomes in health research, as these models are used much more frequently now and are readily available in most mainstream statistical packages.
Heterogeneity was considered in some manner by 81% of included reviews, whether by known quantitative measures or other assessments. The most frequently used measure of heterogeneity was the I^{2} statistic. Alternative measures included the Q Statistic (Chisquare statistic), and BreslowDay test. In a few instances, heterogeneity was estimated and reported from a twostage approach; even when a onestage approach was used for the main analysis.
Investigating subgroup effects was one of the primary reasons for conducting an IPDMA, and among IPDMA that aimed to estimate a treatment effect or treatment safety all investigated subgroup effects. On the other hand, IPDMA were unlikely to adjust for potential confounders unless observational studies were included.
Within the realm of IPDMA with binary outcomes, our review shows that a variety of methods were used to estimate a pooled treatment effect. Many of the articles reviewed contained insufficient details on the approach used and the rationale for that approach. We next provide some recommendations and emphasize the use of the PRISMA statement to help authors ensure transparent and complete reporting of systematic reviews and metaanalyses [3, 44, 45]. First, if individual raw data is available for all studies and irrespective of the final approach, most statisticians and methodologists prefer the onestage rather than a twostage approach [2]. In some cases, the one and twostage approaches will give similar results [46]. However, it is currently unknown under what conditions this may be expected. Moreover, one stage methods may be preferred for evaluating treatmentcovariate interactions of continuous covariates, incorporating nonlinear relationships, when studies are small, and there is heterogeneity across studies, and particularly for pooling of non randomized trials that may need to be adjusted for several confounders [46].
Moreover, methods have been developed to incorporate both individual patient data with summary level data when necessary, so that having partial IPD should not be an impediment to using a onestage approach [5, 11].
However, when random effects logistic regression is used, several details should be reported including: whether study and/or treatment were considered as random, and the statistical method used to estimate the GLMM (e.g. PQL or adaptive Gaussian Hermite quadrature). On the other hand, if a twostage approach is used, we suggest that the metaanalytic technique used to pool results should be stated explicitly. Moreover, simply pooling subjects from various studies together is not appropriate.
Assessment and exploration of heterogeneity should always be performed in any MA, or IPDMA. Nonetheless, how best to quantify heterogeneity remains unclear. While some advocate using the estimated variance of the random treatment effect, difficulties with its interpretation may imply that I^{2} as estimated from a twostage approach is the optimal choice for quantifying heterogeneity. Of course, whether heterogeneity estimated from a twostage approach is relevant to a onestage model is an open question.There are some limitations to the work presented here. First, we have focused on binary outcomes, while survival outcomes were reported in about half of the studies retrieved (See Figure 1). Second, we limited our study retrieval to articles published in 2011. This choice was made because this gave us a sufficient sample of studies to work with that were recently completed. Moreover, we believe that there are unlikely to be major differences in the methods used, or in how they were reported between e.g. 2010 and 2011. Finally, we have focused only on the statistical approach used in these studies; whereas some may be interested more generally in how well IPDMA are reported.
Conclusion
As found previously, we have demonstrated that a diversity of methods are employed when dealing with IPD metaanalyses for binary outcomes. Evidence from this systematic review shows that the use IPDMA of binary outcomes has increased, with random effects logistic regression the most common method of analysis. The statistical approach taken, along with justification for that approach, is still often not reported in sufficient detail. Standardized guidelines both for the best approach to use, as well as what details to report may be needed in this area.
Abbreviations
 IPD:

Individual patient data
 MA:

Metaanalysis
 IPDMA:

Individual patient data metaanalysis
 IQR:

Inter quartile range
 PQL:

Penalized Quasilikelihood.
Declarations
Acknowledgements
We thank our two reviewers Johannes Reitsma and David Fisher whose thoughtful reviews enabled us to improve our work substantially.
Authors’ Affiliations
References
 DerSimonian R, Laird N: Metaanalysis in clinical trials. Control Clin Trials. 1986, 7: 177188.View ArticlePubMedGoogle Scholar
 Riley RD: Commentary: like it and lump it? Metaanalysis using individual participant data. Int J Epidemiol. 2010, 39: 13591361.View ArticlePubMedGoogle Scholar
 Riley RD, Lambert PC, AboZaid G: Metaanalysis of individual participant data: rationale, conduct, and reporting. BMJ. 2010, 340: c221View ArticlePubMedGoogle Scholar
 Ades AE, Lu G, Higgins JP: The interpretation of randomeffects metaanalysis in decision models. Med Decis Making. 2005, 25: 646654.View ArticlePubMedGoogle Scholar
 Riley RD, Lambert PC, Staessen JA, Wang J, Gueyffier F, Thijs L, Boutitie F: Metaanalysis of continuous outcomes combining individual patient data and aggregate data. Stat Med. 2008, 27: 18701893.View ArticlePubMedGoogle Scholar
 Berlin JA, Santanna J, Schmid CH, Szczech LA, Feldman HI: Individual patient versus grouplevel data metaregressions for the investigation of treatment effect modifiers: ecological bias rears its ugly head. Stat Med. 2002, 21: 371387.View ArticlePubMedGoogle Scholar
 Fisher DJ, Copas AJ, Tierney JF, Parmar MK: A critical review of methods for the assessment of patientlevel interactions in individual participant data metaanalysis of randomized trials, and guidance for practitioners. J Clin Epidemiol. 2011, 64: 949967.View ArticlePubMedGoogle Scholar
 Higgins JP, Thompson SG, Spiegelhalter DJ: A reevaluation of randomeffects metaanalysis. J R Stat Soc Ser A Stat Soc. 2009, 172: 137159.View ArticlePubMedPubMed CentralGoogle Scholar
 Schmid CH, Stark PC, Berlin JA, Landais P, Lau J: Metaregression detected associations between heterogeneous treatment effects and studylevel, but not patientlevel, factors. J Clin Epidemiol. 2004, 57: 683697.View ArticlePubMedGoogle Scholar
 Lambert PC, Sutton AJ, Abrams KR, Jones DR: A comparison of summary patientlevel covariates in metaregression with individual patient data metaanalysis. J Clin Epidemiol. 2002, 55: 8694.View ArticlePubMedGoogle Scholar
 Riley RD, Simmonds MC, Look MP: Evidence synthesis combining individual patient data and aggregate data: a systematic review identified current practice and possible methods. J Clin Epidemiol. 2007, 60: 431439.PubMedGoogle Scholar
 Simmonds MC, Higgins JP: Covariate heterogeneity in metaanalysis: criteria for deciding between metaregression and individual patient data. Stat Med. 2007, 26: 29822999.View ArticlePubMedGoogle Scholar
 Simmonds MC, Higgins JP, Stewart LA, Tierney JF, Clarke MJ, Thompson SG: Metaanalysis of individual patient data from randomized trials: a review of methods used in practice. Clin Trials. 2005, 2: 209217.View ArticlePubMedGoogle Scholar
 Rodseth RN, Lurati Buse GA, Bolliger D, Burkhart CS, Cuthbertson BH, Gibson SC, Mahla E, Leibowitz DW, Biccard BM: The predictive ability of preoperative Btype natriuretic peptide in vascular patients for major adverse cardiac events: an individual patient data metaanalysis. J Am Coll Cardiol. 2011, 58: 522529.View ArticlePubMedGoogle Scholar
 van der Pas MH, Meijer S, Hoekstra OS, Riphagen II, de Vet HC, Knol DL, van Grieken NC, Meijerink WJ: Sentinellymphnode procedure in colon and rectal cancer: a systematic review and metaanalysis. Lancet Oncol. 2011, 12: 540550.View ArticlePubMedGoogle Scholar
 Broeze KA, Opmeer BC, Coppus SF, van GN, Alves MF, Anestad G, Bhattacharya S, Allan J, GuerraInfante MF, Den Hartog JE, Land JA, Idahl A, Van der Linden PJ, Mouton JW, Ng EH, Van der Steeg JW, Steures P, Svenstrup HF, Tiitinen A, Toye B, Van d V, Mol BW: Chlamydia antibody testing and diagnosing tubal pathology in subfertile women: an individual patient data metaanalysis. Hum Reprod Update. 2011, 17: 301310.View ArticlePubMedGoogle Scholar
 Broeze KA, Opmeer BC, Van GN, Coppus SF, Collins JA, Den Hartog JE, Van der Linden PJ, Marianowski P, Ng EH, Van der Steeg JW, Steures P, Strandell A, Van d V, Mol BW: Are patient characteristics associated with the accuracy of hysterosalpingography in diagnosing tubal pathology? An individual patient data metaanalysis. Hum Reprod Update. 2011, 17: 293300.View ArticlePubMedGoogle Scholar
 Leroy S, Romanello C, GalettoLacour A, Bouissou F, FernandezLopez A, Smolkin V, Gurgoze MK, Bressan S, Karavanaki K, Tuerlinckx D, Leblond P, Pecile P, Coulais Y, Cubells C, Halevy R, Aygun AD, Da DL, Stefanidis CJ, Vander BT, Bigot S, Dubos F, Gervaix A, Chalumeau M: Procalcitonin is a predictor for highgrade vesicoureteral reflux in children: metaanalysis of individual patient data. J Pediatr. 2011, 159: 644651.View ArticlePubMedGoogle Scholar
 Askie LM, Ballard RA, Cutter GR, Dani C, Elbourne D, Field D, Hascoet JM, Hibbs AM, Kinsella JP, Mercier JC, Rich W, Schreiber MD, Wongsiridej PS, Subhedar NV, Van Meurs KP, Voysey M, Barrington K, Ehrenkranz RA, Finer NN: Inhaled nitric oxide in preterm infants: an individualpatient data metaanalysis of randomized trials. Pediatrics. 2011, 128: 729739.View ArticlePubMedPubMed CentralGoogle Scholar
 Bonati LH, Fraedrich G: Age modifies the relative risk of stenting versus endarterectomy for symptomatic carotid stenosis–a pooled analysis of EVA3S, SPACE and ICSS. Eur J Vasc Endovasc Surg. 2011, 41: 153158.View ArticlePubMedGoogle Scholar
 von MG, Untch M, Nuesch E, Loibl S, Kaufmann M, Kummel S, Fasching PA, Eiermann W, Blohmer JU, Costa SD, Mehta K, Hilfrich J, Jackisch C, Gerber B, du BA, Huober J, Hanusch C, Konecny G, Fett W, Stickeler E, Harbeck N, Muller V, Juni P: Impact of treatment characteristics on response of different breast cancer phenotypes: pooled analysis of the German neoadjuvant chemotherapy trials. Breast Cancer Res Treat. 2011, 125: 145156.View ArticleGoogle Scholar
 Houben RM, Crampin AC, Ndhlovu R, Sonnenberg P, GodfreyFaussett P, Haas WH, Engelmann G, Lombard CJ, Wilkinson D, Bruchfeld J, Lockman S, Tappero J, Glynn JR: Human immunodeficiency virus associated tuberculosis more often due to recent infection than reactivation of latent infection. Int J Tuberc Lung Dis. 2011, 15: 2431.View ArticlePubMedGoogle Scholar
 Lejoyeux M, Lehert P: Alcoholuse disorders and depression: results from individual patient data metaanalysis of the acamprosatecontrolled studies. Alcohol Alcohol. 2011, 46: 6167.View ArticlePubMedGoogle Scholar
 Cardwell CR, Stene LC, Joner G, Bulsara MK, Cinek O, Rosenbauer J, Ludvigsson J, Svensson J, Goldacre MJ, Waldhoer T, JaroszChobot P, Gimeno SG, Chuang LM, Roberts CL, Parslow RC, Wadsworth EJ, Chetwynd A, Brigis G, Urbonaite B, Sipetic S, Schober E, Devoti G, IonescuTirgoviste C, de Beaufort CE, Stoyanov D, Buschard K, Radon K, Glatthaar C, Patterson CC: Birth order and childhood type 1 diabetes risk: a pooled analysis of 31 observational studies. Int J Epidemiol. 2011, 40: 363374.View ArticlePubMedGoogle Scholar
 Patti G, Cannon CP, Murphy SA, Mega S, Pasceri V, Briguori C, Colombo A, Yun KH, Jeong MH, Kim JS, Choi D, Bozbas H, Kinoshita M, Fukuda K, Jia XW, Hara H, Cay S, Di SG: Clinical benefit of statin pretreatment in patients undergoing percutaneous coronary intervention: a collaborative patientlevel metaanalysis of 13 randomized studies. Circulation. 2011, 123: 16221632.View ArticlePubMedGoogle Scholar
 Porter CK, Riddle MS, Tribble DR, Louis BA, McKenzie R, Isidean SD, Sebeny P, Savarino SJ: A systematic review of experimental infections with enterotoxigenic Escherichia coli (ETEC). Vaccine. 2011, 29: 58695885.View ArticlePubMedGoogle Scholar
 Groeneveld E, Broeze KA, Lambers MJ, Haapsamo M, Dirckx K, Schoot BC, Salle B, Duvan CI, Schats R, Mol BW, Hompes PG: Is aspirin effective in women undergoing in vitro fertilization (IVF)? Results from an individual patient data metaanalysis (IPD MA). Hum Reprod Update. 2011, 17: 501509.View ArticlePubMedGoogle Scholar
 Berghella V, Rafael TJ, Szychowski JM, Rust OA, Owen J: Cerclage for short cervix on ultrasonography in women with singleton gestations and previous preterm birth: a metaanalysis. Obstet Gynecol. 2011, 117: 663671.View ArticlePubMedGoogle Scholar
 de Boer SP, Barnes EH, Westerhout CM, Simes RJ, Granger CB, Kastrati A, Widimsky P, de Boer MJ, Zijlstra F, Boersma E: Highrisk patients with STelevation myocardial infarction derive greatest absolute benefit from primary percutaneous coronary intervention: results from the Primary Coronary Angioplasty Trialist versus thrombolysis (PCAT)2 collaboration. Am Heart J. 2011, 161: 500507.View ArticlePubMedGoogle Scholar
 Laporte S, Liotier J, Bertoletti L, Kleber FX, Pineo GF, Chapelle C, Moulin N, Mismetti P: Individual patient data metaanalysis of enoxaparin vs. unfractionated heparin for venous thromboembolism prevention in medical patients. J Thromb Haemost. 2011, 9: 464472.View ArticlePubMedGoogle Scholar
 Jefferis J, Perera R, Everitt H, Van WH, Rietveld R, Glasziou P, Rose P: Acute infective conjunctivitis in primary care: who needs antibiotics? An individual patient data metaanalysis. Br J Gen Pract. 2011, 61: e542e548.View ArticlePubMedPubMed CentralGoogle Scholar
 Zinkstok SM, Vergouwen MD, Engelter ST, Lyrer PA, Bonati LH, Arnold M, Mattle HP, Fischer U, Sarikaya H, Baumgartner RW, Georgiadis D, Odier C, Michel P, Putaala J, Griebe M, Wahlgren N, Ahmed N, Van GN, de Haan RJ, Nederkoorn PJ: Safety and functional outcome of thrombolysis in dissectionrelated ischemic stroke: a metaanalysis of individual patient data. Stroke. 2011, 42: 25152520.View ArticlePubMedGoogle Scholar
 Saber W, Moua T, Williams EC, Verso M, Agnelli G, Couban S, Young A, De CM, Biffi R, van Rooden CJ, Huisman MV, Fagnani D, Cimminiello C, Moia M, Magagnoli M, Povoski SP, Malak SF, Lee AY: Risk factors for catheterrelated thrombosis (CRT) in cancer patients: a patientlevel data (IPD) metaanalysis of clinical trials and prospective studies. J Thromb Haemost. 2011, 9: 312319.View ArticlePubMedPubMed CentralGoogle Scholar
 Kelder JC, Cowie MR, McDonagh TA, Hardman SM, Grobbee DE, Cost B, Hoes AW: Quantifying the added value of BNP in suspected heart failure in general practice: an individual patient data metaanalysis. Heart. 2011, 97: 959963.View ArticlePubMedGoogle Scholar
 Waldman AT, Stull LB, Galetta SL, Balcer LJ, Liu GT: Pediatric optic neuritis and risk of multiple sclerosis: metaanalysis of observational studies. J AAPOS. 2011, 15: 441446.View ArticlePubMedGoogle Scholar
 Baman TS, Meier P, Romero J, Gakenheimer L, Kirkpatrick JN, Sovitch P, Oral H, Eagle KA: Safety of pacemaker reuse: a metaanalysis with implications for underserved nations. Circ Arrhythm Electrophysiol. 2011, 4: 318323.View ArticlePubMedGoogle Scholar
 Lanas A, McCarthy D, Voelker M, Brueckner A, Senn S, Baron JA: Shortterm acetylsalicylic acid (aspirin) use for pain, fever, or colds  gastrointestinal adverse effects: a metaanalysis of randomized clinical trials. Drugs R D. 2011, 11: 277288.View ArticlePubMedGoogle Scholar
 He W, Gandhi CD, Quinn J, Karimi R, Prestigiacomo CJ: True aneurysms of the posterior communicating artery: a systematic review and metaanalysis of individual patient data. World Neurosurg. 2011, 75: 6472.View ArticlePubMedGoogle Scholar
 Black JA, Herbison GP, Lyons RA, Polinder S, Derrett S: Recovery after injury: an individual patient data metaanalysis of general health status using the EQ5D. J Trauma. 2011, 71: 10031010.View ArticlePubMedGoogle Scholar
 Breslow N, Clayton D: Approximate inference in generalized linear mixed models. J Am Stat Assoc. 1993, 88: 925.Google Scholar
 Pinheiro J, Bates D: Approximations to the loglikelihood function in the nonlinear mixedeffects model. J Comput Graph Stat. 1995, 4: 1235.Google Scholar
 Higgins JP, Thompson SG: Quantifying heterogeneity in a metaanalysis. Stat Med. 2002, 21: 15391558.View ArticlePubMedGoogle Scholar
 Altman DG, Bland JM: Interaction revisited: the difference between two estimates. BMJ. 2003, 326: 219View ArticlePubMedPubMed CentralGoogle Scholar
 Moher D, Liberati A, Tetzlaff J, Altman DG: Preferred reporting items for systematic reviews and metaanalyses: the PRISMA statement. J Clin Epidemiol. 2009, 62: 10061012.View ArticlePubMedGoogle Scholar
 Liberati A, Altman DG, Tetzlaff J, Mulrow C, Gotzsche PC, Ioannidis JP, Clarke M, Devereaux PJ, Kleijnen J, Moher D: The PRISMA statement for reporting systematic reviews and metaanalyses of studies that evaluate health care interventions: explanation and elaboration. J Clin Epidemiol. 2009, 62: e1e34.View ArticlePubMedGoogle Scholar
 Stewart GB, Altman DG, Askie LM, Duley L, Simmonds MC, Stewart LA: Statistical analysis of individual participant data metaanalyses: a comparison of methods and recommendations for practice. PLoS One. 2012, 7: e46042View ArticlePubMedPubMed CentralGoogle Scholar
 The prepublication history for this paper can be accessed here:http://www.biomedcentral.com/14712288/14/79/prepub
Prepublication history
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.