Literature search. A thorough literature search was performed together with a medical librarian (JP), to reduce bias by increasing the likelihood of retrieving all relevant studies (^{Vochteloo et al. 2010}). The following bibliographies were searched up to 2009: PubMed, Embase, Web-of-Science, and the Cochrane Library. Relevant articles were screened for additional references. Additionally, a separate search was conducted in 9 leading orthopedic and biomechanical journals (Acta Orthop, Clin Orthop Relat Res, J Arthroplasty, J Bone Joint Surg (Am and Br), Knee Surg Sports Traumatol Arthrosc, J Orthop Res, J Biomech, and Clin Biomech). Finally, Google Scholar was used. Articles in English, French, Italian, Spanish, Dutch, and German were considered. The search strategy consisted of the following components, each defined by a combination of controlled vocabulary and free text terms: (1) RSA, and (2) joint replacement. See Appendix (Supplementary data) for more details on the strategy and terms.

Inclusion and exclusion analysis. Initial screening on the basis of title and abstract of RSA studies was performed by BP to identify studies on patients treated with TKPs for OA or RA. When the information in the abstract did not suffice or where there was any doubt, the studies remained eligible. The full text of eligible studies was independently evaluated in duplicate by 2 reviewers (BP and EV). The inclusion criteria for RSA studies were: (1) primary TKP, and (2) minimal RSA follow-up of 1 year, measuring tibial component migration. Non-clinical studies (animal, phantom) were excluded.

Data extraction. BP and KN independently extracted migration data in duplicate from the RSA studies. Migration data comprised translations, rotations, and maximal total point motion (MTPM) of the tibial component in the first postoperative year. MTPM is the unit of measurement for the largest 3D migration of any point on the prosthesis surface (^{Ryd et al. 1995}). Data concerning patient demographics and regional influences were also extracted to allow for confounder correction.

Quality assessment. The quality of the RSA studies was independently appraised in duplicate by BP and KN at the level of outcome using the AQUILA methodological score (^{Pijls et al. 2011}). For the RSA studies, we modified the AQUILA by removing items not considered relevant for early migration: long-term follow-up and the revision assessment.

Literature search. The search strategy and bibliographies were the same as those in the RSA review, with the exception of the components of the search strategy. The search strategy for the survival studies consisted of the following components, each defined by a combination of controlled vocabulary and free text terms: (1) joint replacement, (2) implant failure, and (3) survival analysis. See Appendix for more details of the strategy and terms. In the search strategy, no distinction was made between total knee prostheses and total hip prostheses (THPs), because some studies reported on both TKPs and THPs (^{Ryd 1992}).

Inclusion and exclusion analysis. The procedure of screening the survival studies for eligibility, and subsequent inclusion and exclusion analysis, was identical to the procedures for the RSA studies, with the exception of inclusion and exclusion criteria. The inclusion criteria for survival studies were (1) primary TKP; (2) follow-up of 5, 10, 15, 20, or 25 years; (3) endpoint revision surgery for aseptic loosening of the tibial component, or indication for revision surgery in patients with poor general health or decline; and (4) survival or percentage revised to be available for specific follow-up (see point 2). Studies with less than 75 TKPs at baseline were excluded.

Data extraction. From the survival studies, BP and KN independently determined the revision rates for aseptic loosening of the tibial component at 5-year intervals. Data concerning patient demographics and regional influences were extracted to allow for confounder correction.

Quality assessment. The quality of the survival studies was independently appraised in duplicate by BP and KN at the level of outcome using the AQUILA methodological score (^{Pijls et al. 2011}).

A detailed description of the analysis, methodology, and a worked example are available in the online Appendix (see Supplementary data). To determine the association between early migration and late revision, we matched the results from the RSA review to the results of the survival review on type of prosthesis, fixation method (e.g. cement or bone ingrowth), and articulating insert (e.g. modular or non-modular). The combination was termed PFI. Since PFI involves technical factors known to be associated with both migration and the likelihood of revision for aseptic loosening, matching on PFI prevents confounding by PFI (^{DKAR 2009}, ^{NJR 2009}, ^{AJR 2010}, ^{SKAR 2010}). Depending on the studies available, it was possible that there would be more than 1 combination of matching RSA and survival studies for a particular PFI. For instance, if there are 3 RSA studies and 2 survival studies for the same PFI, then there are 6 possible combinations (3 times 2). All combinations were considered in the analysis. A meta-analysis for the revision rate at 5 years was performed. A model for the censoring mechanism was employed to reconstruct the data and then a generalized linear mixed model with study as a random effect was applied to estimate the survival at 5 years and its 95% confidence interval (CI) (^{Fiocco et al. 2009}, ²⁰¹¹, ^{Putter et al. 2009}). Regarding the RSA studies, pooling of migration results at the level of PFI was based on weights according to study size (N).

The 10-year results for TKPs with high revision rates may not be published once the 5-year results have been published. Since 10-year revision rates in the registries are on average 1.7 times higher than 5-year revision rates, any missing 10-year results were estimated on 5-year results by applying a factor of 1.7. This method was validated by comparing the estimated 10-year results with the known 10-year results, for the complete cases (^{DKAR 2009}, ^{NJR 2009}, ^{AJR 2010}, ^{SKAR 2010}).

Since migration data and revision rate data were extracted from different studies, differences between study populations might confound the observed association. In order to address this issue, we determined the degree of similarity of the population from RSA and survival study combinations, expressed by a match score, for age, sex, diagnosis, hospital type, and continent. The match score was constructed according to the results of a recent Delphi survey among an international group of 37 independent experts, and can vary between 5 (excellent) and 0 (poor) (^{Pijls et al. 2011}). The RSA study and survival study combination scored 1 point for each of the following 5 criteria (up to a maximum of 5 points): (1) the difference in mean age between the patients from RSA study and those from the survival study was 5 years or less; (2) the difference in percentage of females between the RSA study and survival study was 10% or less; (3) the difference in percentage of patients diagnosed with osteoarthritis between the RSA study and the survival study was 10% or less; (4) the RSA study and the survival study were performed in similar types of hospital (e.g. both university medical centers); (5) the RSA study and the survival study were performed on the same continent. All other cases scored zero points.

We used a weighted regression model to assess the association between early migration and late aseptic revision corrected for match score, RSA study quality, survival study quality, number of TKP in the RSA studies, and number of TKP in the survival studies.

According to the principle of “primum non nocere“ (first do no harm), new implant designs should perform at least as well as the revision standard of national registries: < 3% revision at 5 years and < 5% revision at 10 years (^{DKAR 2009}, ^{NJR 2009}, ^{AJR 2010}, ^{SKAR 2010}). Based on this revision standard, the following 3 categories were constructed for the phased introduction of new TKPs: “acceptable”, “at risk”, and “unacceptable”. The “acceptable” category was defined as the level of migration up to which all survival studies have lower revision rates than the standard. The “unacceptable” category was defined as the level of migration from which all revision rates are higher than the standard. The category “at risk” was defined as the migration interval between the “acceptable” and “unacceptable” thresholds, in which studies with revision rates lower and higher than the standard were observed.

We assessed the potential effect of publication bias by comparing the results from the meta-analysis to the results from national joint registries, since they do not suffer from publication bias (^{DKAR 2009}, ^{NJR 2009}, ^{AJR 2010}, ^{SKAR 2010}). Accordingly, the PFI combinations that perform better than average in the meta-analysis should also perform better than average in the national joint registries. The same principle also applies to PFI combinations that perform worse than average. For this purpose, the migration pooled according to the specific combination of prosthesis type, fixation method and articulating insert (PFI) and visualized in a dot chart (^{Jacoby 2006}).

The literature search yielded 629 hits for the RSA review and 50 studies were included, covering a total of 847 patients (^{Ryd et al. 1986}, ¹⁹⁸⁷, ¹⁹⁸⁸, ¹⁹⁹⁰, ¹⁹⁹³, ^{Albrektsson et al. 1990}, ¹⁹⁹², ^{Nilsson et al. 1991}, ¹⁹⁹³, ¹⁹⁹⁸, ²⁰⁰⁶, ^{Hilding et al. 1995}, ²⁰⁰⁶, ²⁰⁰⁷, ^{Nielsen et al. 1995}, ^{Nelissen et al. 1998}, ^{Onsten et al. 1998}, ^{Toksvig-Larsen et al. 1998}, ^{Adalberth et al. 1999}, ²⁰⁰⁰, ²⁰⁰¹, ²⁰⁰², ^{Østgaard et al. 1999}, ^{Li et al. 2000}, ^{Toksvig-Larsen et al. 2000}, ^{Uvehammer et al. 2001}, ^{Hildebrand et al. 2003}, ^{Catani et al. 2004}, ^{Kienapfel et al. 2004}, ^{Norgren et al. 2004}, ^{Dalen et al. 2005}, ^{Hyldahl et al. 2005a}, ^b, ^{Henricson et al. 2006}, ^{Muller et al. 2006}, ^{van der Linde et al. 2006a},^b, ^{Uvehammer et al. 2007}, ^{Wilson et al. 2007}, ^{Hansson et al. 2008}, ^{Henricson et al. 2008}, ^{Therbo et al. 2008}, von ^{Schewelov et al. 2008}). Details of the study selection and flow of the review are given in Figure 1. The mean quality score of the RSA studies was 3.8 (SD 1.7) on a 7-point scale. MTPM at 1 year was the most frequently and most consistently reported migration value: 44 out of 50 RSA studies reported it. Translations and rotations of the tibial component were reported infrequently and inconsistently, and did not allow a meaningful analysis. Thus, all analyses focused on MTPM at 1 year.

The literature search resulted in 5,290 hits for the survival review and 56 studies were included, with a total of 20.599 patients; see Figure 1 (^{Goldberg et al. 1988}, ^{Laskin 1990}, ^{Samuelson et al. 1990}, ^{Wright et al. 1990}, ^{Moran et al. 1991}, ^{Grewal et al. 1992}, ^{Ranawat et al. 1993}, ¹⁹⁹⁴, ^{Rinonapoli et al. 1994}, ^{Weir et al. 1996}, ^{Knight et al. 1997}, ^{Scott 1997}, ^{Ansari et al. 1998}, ^{Hsu et al. 1998}, ^{Ewald et al. 1999}, ^{Mont et al. 1999}, ^{Buehler et al. 2000}, ^{Emerson et al. 2000}, ^{Robertsson et al. 2000}, ^{Stukenborg-Colsman et al. 2000}, ^{Berger et al. 2001a},^{b, Faraj et al. 2001}, ^{Gill et al. 2001}, ^{Khaw et al. 2001}, ^{Schroder et al. 2001}, ^{Sextro et al. 2001}, ^{Fetzer et al. 2002}, ^{Forster et al. 2002}, ^{Khaw et al. 2002}, ^{Worland et al. 2002}, ^{Mayman et al. 2003}, ^{Goldberg et al. 2004}, ^{Arora et al. 2005}, ^{Bozic et al. 2005}, ^{Campbell et al. 2006}, ^{Clayton et al. 2006}, ^{Gioe et al. 2006}, ^{Lachiewicz et al. 2006}, ^{Vessely et al. 2006}, ^{Bertin 2007}, ^{Kim et al. 2007}, ^{Rodricks et al. 2007}, ^{Zaki et al. 2007}, ^{Anderson et al. 2008}, ^{Chana et al. 2008}, ^{Dalury et al. 2008}, ^{Parsch et al. 2008}, ^{Ritter et al. 2008}, ^{Santini et al. 2008}, ^{Smith et al. 2008}). The mean quality score of the survival studies was 6.0 (SD 1.8) on an 11-point scale.

The matching procedure resulted in 28 different PFI combinations and 89 combinations of RSA and survival studies (Table 1). There was a clear association between early migration, expressed as MTPM at 1 year, and the 5-year revision rate as expressed as prosthesis survival (Figure 2). For every mm of increase in migration, 7.6% (CI: 5.7–9.5) was added to the 5-year revision rate (p < 0.001). The influence of RSA study quality, survival study quality, number of TKPs in the RSA study, number of TKPs in the survival study, and match score were small relative to the overall effect of migration on revision rate (Table 2).

For TKPs that rely on primary fixation (cemented and uncemented with screws), 7.1% (CI: 4.7–9.5) was added to the 5-year revision rate for every 1 mm increase in MTPM (p <0.001). For TKPs that rely on secondary fixation (uncemented without screws), 10% (CI: 2.7–17) was added to the 5-year revision rate for every 1 mm increase in MTPM (p = 0.018).

Figure 3 shows the 3 categories for the migration of TKP. For MTPM at 1 year of between 0 mm and 0.54 mm, there was no tibial component with more than 3% revision for aseptic loosening at 5 years. With 1-year MTPM of more than 1.6 mm, no tibial components had less than 3% revision for aseptic loosening at 5 years. This indicates that acceptance of 3% revision at 5 years resulted in a threshold of 0.54 mm or acceptable MTPM at 1 year, and a threshold of 1.6 mm for unacceptable MTPM at 1 year. For the 10-year revision rates, the thresholds for acceptable and unacceptable migration were 0.45 mm and 1.6 mm, respectively (Figure 4).

The mean difference between the estimated 10-year revision rate and the known 10-year revision rate was 0.17% (SD 2.1), indicating that there was no systematic error. The 5-year revision rates for the studies with missing 10-year revision rates were already higher than the 10-year revision rate of 5% that is considered to be acceptable. Thus, the 10-year thresholds were not influenced by any missing values.

The pooled MTPM ranked by the pooled revision rate for each PFI combination is presented in Figure 5. The PFI combinations that migrated statistically significantly less than the acceptable threshold—classified as acceptable—have had excellent track records and low revision rates in several national joint registries (^{DKAR 2009}, ^{NJR 2009}, ^{AJR 2010}, ^{SKAR 2010}). Conversely, the PFI combinations that were classified as unacceptable on basis of their pooled migration have been abandoned and are no longer used. The possible influence of publication bias on the results was therefore small.