A randomised controlled trial of the clinical effectiveness and cost-effectiveness of different knee prostheses: the Knee Arthroplasty Trial (KAT)

Health Technology Assessment programme-commissioned call

In the late 1990s, new developments in knee replacement were identified as a priority for research within the NHS. Although they held the promise of better results, the newer forms of arthroplasty were more expensive and information was needed on their safety and cost-effectiveness. This report describes the Knee Arthroplasty Trial (KAT), which was commissioned by the Health Technology Assessment (HTA) programme to address this need. When the trial was funded, it was recognised that any differential performance of the alternative prostheses that were to be compared would become apparent only after a long period of follow-up. Therefore, a plan to report findings after 10 years was built into the study from the start, and results up to a median of 10 years after surgery are described in this report.

Aims

The trial was designed to address questions about four developments in knee replacement surgery:

1. Is it better to resurface the patella as part of a knee replacement or not?

2. Are polyethylene moving components (‘mobile bearing’) between the tibia and femur better than standard designs with a fixed bearing?

3. Are tibial components made out of just high-density polyethylene (‘all polyethylene’) better than those with a metal backing plate and stem, and polyethylene bearing (‘metal-backed’)?

4. Is it better to replace a single compartment of the knee (unicompartmental replacement) or to replace the whole knee joint [total knee replacement (TKR)]?

Clinical background including updated review of evidence base

Total knee arthroplasty is now a common and established surgical procedure. Long-term observational studies have shown that there is variability in the failure rates of different designs of knee replacement, with the best having a 20-year survivorship of about 90%. 1–3 Overall, up to 20% of patients are not satisfied with the result of their knee replacement. Continued developments in design have aimed at further improving function and quality of life and increasing the duration of prosthetic survival.

A substantial proportion of patients have a poor functional result and persistent knee pain after knee replacement. 4,5 Many of these poor results are attributed to problems arising from the patellofemoral joint, and there is considerable debate regarding whether the patella should be resurfaced at the time of the primary total knee arthroplasty. Theoretically, patellar resurfacing should decrease the incidence of pain related to the patellofemoral joint, although the resurfacing can fail. Previous evidence in the form of non-randomised cohort studies, small randomised controlled trials (RCTs) and systematic reviews has not resolved the uncertainty regarding the benefits of patellar resurfacing. 6–21

Many previous RCTs had insufficient sample sizes to detect clinically worthwhile differences in outcomes. There is great variability in the use of patellar resurfacing. Some surgeons routinely do not resurface, whereas others routinely do resurface. There is also a subgroup that sometimes resurfaces the patella, but there is no clear consensus as to which patients should undergo resurfacing.

Theoretically, by using mobile rather than fixed bearings between the tibial and femoral components the performance and longevity of knee replacement could be improved. 22,23 Mobile bearings are usually designed to be more congruent than fixed bearings and, therefore, have larger contact areas, which should reduce wear. However, as there are two bearing surfaces, the decrease in wear may be limited. They also tend to have less constraint than fixed bearings, which should limit the shear stress at the bone–implant interfaces. Strong interfacial forces and osteolysis resulting from polyethylene wear debris are the most important causes of loosening, which is the most common cause of knee replacement failure. By allowing an appropriate amount of mobility, a mobile bearing design could optimise kinematics and thus improve function. However, mobile bearings can dislocate or be associated with instability. Previous non-randomised cohort studies, RCTs and systemic reviews did not resolve the benefits of using a mobile versus a fixed bearing. 24–29 Again, the evidence base suffers from limited evidence from RCTs with sufficient sample sizes to detect worthwhile differences.

Another common variation is the design of the tibial component. Use of a metal-backed base plate has theoretical advantages in that it distributes load more evenly across the implant–bone interface than an all-polyethylene tibia, and thus should decrease the risk of loosening. In addition, as the bearing is modular, the surgeon can select the thickness and constraints of the bearing after the components are fixed. However, metal backing reduces the thickness of the polyethylene that can be implanted in the available space, thus increasing the internal stresses within the polyethylene and increasing the risk of wear. Furthermore, metal backing is more expensive, and good medium- and long-term results have been reported for the use of non-metal-backed components. 30,31 Limited comparisons between non-metal-backed and metal-backed components have been performed, and to our knowledge no definitive difference has been determined. 4,32 As there is no apparent clinical advantage of metal-backed tibias and as they are more expensive, it is generally recommended that all-polyethylene tibias should be used in the elderly to save money. 33,34

Overall, in the NHS about 7% of knee replacements are unicompartmental. There is, however, great variability: in different institutions unicompartmental knee replacements are used for between 0% and 70% of knee replacements. The results from specialist centres that implant large numbers of unicompartmental knee replacements have demonstrated that unicompartmental replacements give a faster recovery, lower morbidity, lower cost, better function and better pain relief than total replacements. However, in national registers, even though the risk of serious complications, such as death and infection, is lower with unicompartmental than with total replacement, the revision rate is about three times higher. There is clearly a need to establish the relative advantages and disadvantages of unicompartmental and TKRs so as to determine whether the use of unicompartmental replacement should or should not be encouraged. Although the only randomised study we are aware of did suggest that unicompartmental replacement is better than total, it was too small to form the basis of any strong recommendation. 13

Study question for economic evaluation

We aimed to assess the cost-effectiveness of the following types of knee prosthesis, in line with the comparisons addressed by the trial:

• patellar resurfacing versus no patellar resurfacing

• mobile versus fixed bearing

• all-polyethylene versus metal-backed tibial component

• unicompartmental versus TKR.

Following UK guidelines for economic evaluation,47 the KAT economic evaluation was conducted from the costing perspective of the UK NHS and from the health perspective of the patients undergoing knee replacement. The analysis, therefore, excluded costs incurred by other sectors of the economy, such as lost productivity, informal care, participants’ out-of-pocket expenses or home/residential/rehabilitation care.

As KAT was a pragmatic trial involving long-term follow-up of elderly participants, data collection was streamlined further to reduce the questionnaire burden on participants and maximise response rates to annual questionnaires. In particular, the amount of costing data collected from annual questionnaires was limited to the main NHS costs directly attributable to knee replacement.

Framework for economic evaluation

Cost–utility analysis (CUA) was used to evaluate cost-effectiveness in order to capture any differences in quality of life between randomised comparisons. The primary end point therefore comprised the cost per quality-adjusted life-year (QALY) gained. CUA was used regardless of whether functional status or quality of life differed between randomised comparisons, as our primary interest is in the joint distribution of cost and effect differences, and also because KAT was not powered to demonstrate equivalence, and thus assuming no difference in outcomes could give misleading conclusions and bias estimates of decision uncertainty. 48,49

The base-case analysis comprised a within-trial economic evaluation using only data from KAT. Results were not extrapolated beyond the end of the trial, as the within-trial period was already substantial (up to 12 years). Attrition and administrative censoring (i.e. data from time points not yet reached as a result of staggered recruitment times) were dealt with using multiple imputation and inverse probability weighting (IPW), as described below. The economic evaluation included all participants formally included in the trial with the exception of 34 participants who died before surgery and 66 participants who were randomised to the total versus unicompartmental knee replacement comparison or withdrew from the trial prior to surgery, giving a total sample of 2252 participants. In line with the clinical analysis, the economic evaluation includes all annual questionnaires received by 8 June 2012. Additional data on hospital readmissions were obtained from HES up to 31 March 2011 (for participants living in England) and from ISD up to 31 December 2010 (for participants living in Scotland).

The base-case economic evaluation took a 10-year time horizon and, therefore, included all data on quality of life and costs and readmissions that occurred within 10 years of primary TKR or before administrative censoring (if sooner). However, alternative time horizons are presented in the sensitivity analyses. Costs and QALYs accrued beyond year 1 were discounted to present values at 3.5% per annum. 50

Collection of resource-use data and unit costs

Quantities of knee-related resources used in the primary admission and subsequent follow-up were collected prospectively for each trial participant. Data collection focused on resources associated with the study knee and excluded resource use from unrelated causes to simplify data collection and reduce the length of participant questionnaires. Focusing on resource use directly related to the study knee and the complications of knee surgery also avoids the risk of catastrophic episodes involving high health-care resource use unrelated to treatment (e.g. renal failure, cancer or extended psychiatric admissions) swamping the main effect of treatment on costs and, therefore, reduces uncertainty around incremental costs and cost-effectiveness. 51

Data on quantities of NHS resources related to the study knee were identified from the surgeon’s form (see Appendix 1), the hospital care form (see Appendix 1) and annual questionnaires (see Appendix 1) completed by participants, with additional data on hospitalisations being collected from HES and ISD (Table 1). Questionnaires focused on the main cost drivers (see Table 1), which were identified following discussions with clinicians and from evidence from previous publications. In line with clinical analyses, readmissions to hospital that were related to the study knee or that were considered (by clinical adjudication) relevant to the primary TKR procedure [e.g. readmissions for myocardial infarction, DVT or pulmonary emboli (PE) within 3 months of TKR] were included in the analysis, but all other readmissions (including readmissions for thromboemboli that occurred after revision procedures on the study knee) were excluded from the analysis. The assumptions used to estimate resource use and costs from questionnaire responses are listed in Box 1.

Mobility aids and medications used to manage pain and arthritis in non-hospitalised participants were excluded from the analysis to simplify questionnaires and reduce the burden on participants. Mobility aids are unlikely to comprise a large proportion of 5- or 10-year total costs; furthermore, they will often be funded from outside the NHS perspective taken in this analysis, and the need for mobility aids will also be affected by comorbid conditions. Similarly, pharmaceuticals are likely to account for only a small proportion of total costs: recent evidence suggests that analgesics and arthritis medication accounts for only around 2–3% of total health-care spending in participants who have undergone joint replacement. 62

Additionally, the following resources were assumed to be included in other unit costs and not considered separately to avoid double counting:

• Drugs (e.g. antibiotics, anaesthetics, anticoagulants) and non-surgical treatments (e.g. urinary catheters) administered during knee-related hospital admission: assumed to be covered by the cost of a bed-day.

• Cement, time in recovery room, intraoperative complications and blood products transfused while in theatre: assumed to be included in the cost of theatre time and not accounted for separately to avoid double counting.

• Surgical instruments used in primary knee replacement: assumed to be supplied at no cost by prosthesis manufacturers.

• Diagnostic tests or postoperative complications other than imaging for DVT/PE, transfusion and admission to high-dependency unit: assumed to be covered by the cost of a bed-day/outpatient consultation and any associated increases in length of stay.

Unit cost data were collected from routinely available sources (see Table 1). 52,56,58,60 The index year for pricing was 2011/12 and prices from earlier (or later) years were inflated (or deflated) to 2011/12 values using the hospital and community hospital services (HCHS) pay and prices index. 53,55

Stickers giving the product codes for the knee prosthesis components used in primary knee replacement and revisions were attached to the surgeon’s form to indicate which brand and model was used in that operation. For all readmissions identified from participant questionnaires or HES/ISD that were deemed potentially related to knee surgery, the hospital was asked to provide a copy of participants’ notes and/or complete the readmission form (see Appendix 2) including codes for all components used. Each component used was then valued using list prices obtained from manufacturers. List prices from 2008 were used for primary procedures when available and inflated to 2011/12 values. However, when such prices were not available (e.g. for components that have since been discontinued), list prices payable at the time of the KAT operations (1999–2002) were inflated to 2011/12 values using the HCHS pay and prices inflation index;53,55 when the price of discontinued components at the time of operation was unavailable, prices were based on comparable components available today. Prices for revision components were obtained between 2008 and 2012 and inflated/deflated based on HCHS when necessary. The pay and prices index appears to be appropriate for knee replacement components, as those components for which list prices were available for 2001 and 2008 increased in price by 38.9% over that period (cf. 36.2% HCHS inflation). 53

In practice, the prices hospitals pay for devices are substantially lower than list prices, as each hospital negotiates discounts with their supplier(s). Although actual discounts vary between hospitals and are commercially sensitive, we applied a fixed discount of 30% on all components in the base-case analysis; this discount was varied in sensitivity analyses.

Calculation of quality-adjusted life-years

Participants completed the three-level EQ-5D and SF-12 questionnaires preoperatively, 3 months after the primary TKR and annually thereafter as part of the participant annual form (see Appendix 1). In the base-case analysis, QALY calculations were based on EQ-5D utilities as prespecified in the trial protocol. EQ-5D is recommended47 and widely used,63 enabling the cost-effectiveness of trial interventions to be compared with other economic evaluations. EQ-5D health-state preference values were calculated using the UK N3 tariff, which is based on time trade-off valuations from 3395 members of the UK general public. 64 The number of QALYs that each participant accrued following TKR was calculated as the area under the utility curve, with linear interpolation between utility measurements (see Box 1).

Methods for imputing missing data

As the calculation of costs and QALYs requires summation of cost and utility data across numerous resource-use items and time points, missing data are a greater problem for economic evaluation than for clinical end points. Specifically within KAT, around 7% (12,102/173,404) of data points prior to administrative censoring were missing across the 77 variables included in the analysis, with 63% (1414/2252) of participants having missing data for at least one resource-use item or quality-of-life measurement, in addition to 884 participants (38%) who were administratively censored before the 10-year follow-up. A complete case analysis excluding all such participants would therefore have been highly inefficient. Furthermore, complete case analysis would also have been prone to bias,65 as some missing data were not missing completely at random: for example, length of stay for revision procedures can be missing only for participants who underwent revisions and utility cannot be missing for participants known to have died in previous years. [The assumption that data are missing completely at random means that the probability of data being missing does not depend on either the values of observed data or the values of missing or unobserved data. Whereas complete case analysis assumes that data are missing completely at random, multiple imputation techniques assume that data are missing at random, i.e. that the probability that data are missing may depend on observed covariates included in the imputation model, but not on unobserved data.] Mean imputation and multiple imputation were used to ensure complete data for all participants up to their date of administrative censoring. 66 IPW was then used to adjust for administrative censoring.

Statistical analysis, calculation of cost-effectiveness ratios and allowance for uncertainty

Although KAT is partially factorial, each randomised comparison was analysed separately, in line with clinical analyses. As a result, each comparison was assumed to address independent questions about separable aspects of knee replacement. In line with the clinical analysis (see Statistical analyses of clinical end points) and the study protocol (see Appendix 1), it was also assumed that there was no interaction between the different comparisons and that treatment effects were additive. The analysis, therefore, assumed that the cost-effectiveness of patellar resurfacing compared with no resurfacing was not affected by whether the tibial component was mobile versus fixed, or all polyethylene versus metal-backed, and that the decision about whether or not to resurface the patella was not affected by decisions about tibial component design. Interactions were not assessed in the base-case analysis, as the study is not powered to detect interactions and 85% of participants were randomised to only one comparison. However, a secondary regression analysis investigated interactions between randomised comparisons for the two subsets of participants randomised to two comparisons.

Following imputation of missing data on resource use, component prices and utilities, QALYs were calculated as the area under the utility profile and quantities of each type of resource were multiplied by unit costs to give the total QALYs and costs accrued by each trial participant in each of the 100 imputed data sets. Costs and QALYs accrued beyond year 1 were discounted at a rate of 3.5% per annum. 50

Imbalances in baseline utility have been shown to introduce substantial bias into unadjusted economic evaluations, as baseline utility is directly included in QALY calculations and normally strongly predicts on-treatment utility. 75 We therefore adjusted for differences in baseline utility by estimating the number of QALYs accrued in each year using ordinary least squares (OLS) regression, controlling for baseline utility and treatment allocation and generating predictions for each study arm based on the mean baseline utility across both arms.

Total costs, resource use and QALYs across the 10-year time horizon were calculated using IPW73,74 to allow for administrative censoring and differences in follow-up time among participants. To implement IPW, OLS regression estimates of the average costs and QALYs accrued in each study arm in each year across uncensored participants were multiplied by the number of participants who were alive and not administratively censored at that time point. Dates of censoring for each participant were identified as described above. Total costs and QALYs were divided by the Kaplan–Meier estimate of the probability of being administratively censored by that time point and then divided by the number of participants included in the analysis. Weighted costs and QALYs were summed across all time periods to give total costs.

Uncertainty around resource use, costs, QALYs and incremental cost-effectiveness ratios (ICERs) was quantified using bootstrapping. Bootstrapping was conducted separately for each randomised comparison (excluding participants not randomised to that comparison) and bootstrapping on each comparison was repeated for each of the 100 imputed data sets. For imputed data set m, bootstrapping76 was used to sample participants with replacement from the trial data set; OLS regression was used to predict resource use, costs and QALYs in each year on that bootstrap replicate and the results of such regression analyses were combined in IPW with Kaplan–Meier estimates of the probability of being censored at each year in that bootstrap replicate. This procedure was repeated on 100 bootstrap replicates drawn (with replacement) from each of the 100 imputed data sets. The standard errors (SEs) around IPW estimates of total costs and QALYs for each imputed data set were calculated as the standard deviation across the 100 bootstrap replicates; results were combined across all imputed data sets using Rubin’s rule to calculate SEs, CIs and p-values around resource use, costs and QALYs. 66

Results are presented as mean ± SE for each group and mean ± 95% CI for between-group differences, with costs in pounds sterling at 2011 prices. The empirical distributions of 10,000 bootstrap estimates of incremental costs and QALYs for each comparison were plotted as scatter graphs on the cost-effectiveness plane. The net benefit method77,78 was also used to estimate cost-effectiveness acceptability curves,79,80 which plotted the probability that each type of prosthesis refinement was cost-effective compared with its comparator against the ceiling ratio. The ceiling ratio represents the amount the NHS is assumed to be willing or able to pay per QALY gained. As the distributions of costs and QALYs for all three comparisons span all four quadrants of the cost-effectiveness plane, the 95% CIs around the ICERs are not defined and range from dominant to dominated. As a result, no 95% CIs are shown for ICERs.

Conclusions were based on the assumption that the NHS is willing to pay £20,000 per QALY gained, based on published social value judgements. 81 Prosthesis refinements generating more QALYs than their comparator were inferred to be cost-effective compared with their comparator if they cost < £20,000 per QALY gained, whereas less effective refinements were inferred to be cost-effective if they saved > £20,000 per QALY forgone compared with their comparator. We assumed that the NHS has symmetrical preferences with respect to losses and gains, as savings from one intervention will fund the cost of others within the largely fixed NHS budget and to ensure that conclusions are not affected by which treatment is designated the comparator.

Sensitivity and subgroup analyses

In total, 18 sensitivity analyses were conducted to relax the assumptions used in the base-case analysis and explore the effect of using alternative methodologies:

• Complete case analysis: excluding all participants with missing data on any component of QALYs or resource use before death or administrative censoring. To ensure complete data on costs and QALYs in participants’ last year of life, EQ-5D utility at time of death was assumed to be equal to participants’ last observed EQ-5D utility measurement and the number of ambulatory consultations attended in participants’ last year of life was assumed to be equal to the number of consultations attended in the previous year, multiplied by the fraction of the year for which the participant was alive.

• Per-protocol analysis: excluding all participants who did not receive the procedure to which they were randomly allocated, or for whom it was unclear whether they had received the allocated procedure.

• Length of stay reduced by 46% for all primary KAT procedures to reflect the fact that mean length of stay for TKR is now 5.3 days,82 compared with 10.0 days among KAT participants.

• 0% price discount applied to component prices (such that all component prices are based on list prices).

• 50% price discount applied to component prices.

• Varying cost per bed-day by ± 50% to reflect uncertainty around costs.

• Varying cost per minute of theatre time by ± 50% to reflect uncertainty around costs and the possibility of double-counting of knee components, which were (to a small degree) included in the cost of theatre time by aggregation of costs over all orthopaedic procedures.

• Alternative discount rates allowing for time preference [0% and 5% for both costs and QALYs and using differential discounting (0% for QALYs and 3.5% for costs)].

• No adjustment for baseline utilities.

• Within-trial time horizon: including all data collected before administrative censoring and making no adjustment for administrative censoring.

• 8-year time horizon.

• 9-year time horizon.

• 11-year time horizon.

• Using multiple imputation to adjust for administrative censoring, rather than IPW. This analysis used the multiple imputation estimates of utilities and resource use after administrative censoring that were excluded from the base-case analysis. For simplicity, this analysis assumed that no participants were readmitted after administrative censoring. Mean imputation was used to impute dates of death for participants who were administratively censored before year 10, assuming that the interval between censoring and death equalled their remaining life expectancy at the time of censoring (obtained from Government Actuary’s Department estimates for age, sex and country of residence). 83

• Participants receiving a femoral component with a patellofemoral joint that is shaped to accommodate an anatomic patella (conducted only for the patella comparison).

• Participants receiving a femoral component with a patellofemoral joint that is shaped to accommodate a domed patella (conducted only for the patella comparison).

All sensitivity analyses were based on 10,000 bootstrap replicates. Complete case analysis and per-protocol analyses were based on separate runs of bootstrapping, whereas all other sensitivity analyses were estimated from the raw data generated in the base-case bootstrapping analysis.

The partial factorial design permitted participants recruited to the patellar resurfacing versus no resurfacing comparison to also be randomised to either mobile versus fixed bearing or all-polyethylene versus metal-backed tibial components. In total, 338 participants were randomised in more than one comparison, 193 of whom were in the mobile versus fixed bearing comparison and 145 in the all-polyethylene versus metal-backed comparison. Although the base-case analysis assumed no interactions between randomised comparisons, interactions were evaluated in sensitivity analyses run on the subset of participants who were randomised to more than one comparison to evaluate whether interactions exist and (if so) what impact this has on the study conclusions. Linear regression was used to estimate the magnitude and statistical significance of interactions between patellar resurfacing and mobile versus fixed bearing and (separately) between patellar resurfacing and all-polyethylene versus metal-backed tibial components. OLS regression was used to predict the total annual cost and total annual QALYs as a function of the main effects for each comparison and an interaction term. For example, the explanatory variables within the analysis of mobile versus fixed bearing comprised a dummy indicating whether or not participants were randomised to patellar resurfacing, a dummy indicating whether or not participants were randomised to mobile bearing and an interaction term (the product of the other two dummies). Baseline EQ-5D utility was included as an additional predictor of annual QALYs. Regression analyses were repeated on 100 bootstrap replicates on each of the 100 imputed data sets, and predicted annual QALYs and annual costs for each treatment arm were adjusted for censoring using IPW and summed to give total costs and total QALYs over the first 10 years after TKR. As a secondary evaluation of interactions, subgroup analyses were conducted to estimate the costs and QALYs for each comparison on the subset of participants randomised to patellar resurfacing and on the subset randomised no patellar resurfacing. Both sets of analyses should be interpreted with caution, as they are based on small participant numbers.

Six post-hoc subgroup analyses were also conducted to explore whether or not differences between randomised comparisons varied between participant subgroups:

• participants < 70 years of age

• participants aged ≥ 70 years

• participants also randomised to patellar resurfacing (conducted only for the metal-backed and mobile bearing comparisons)

• participants also randomised to have no patellar resurfacing (conducted only for the metal-backed and mobile bearing comparisons).

Although no subgroups other than randomised patella allocation were prespecified in the protocol, there are strong a priori reasons to expect costs and/or outcomes of the KAT comparisons to differ by age and patellofemoral joint shape. In particular, age is currently an important factor taken into account in surgeons’ decisions about which component type to use, as it predicts whether or not participants are likely to out-live their prosthesis. In particular, all-polyethylene tibial components are often used on cost grounds in older participants, who are unlikely to out-live the prosthesis. In the absence of prior data to inform the cut-off, the sample was divided into two approximately equal halves. Similarly, the shape of the patella button (anatomical or domed) may affect the cost and/or efficacy of patellar resurfacing. As patella buttons are usually matched to the patellofemoral groove in the femoral component (which is used regardless of whether the participant had patellar resurfacing), we therefore subgrouped by groove shape in the patella comparison.

Funding for the trial

The trial was funded by the National Institute for Health Research (NIHR) HTA programme (project number 95/10/01). Additional industry funding for research support in clinical centres was provided by the following: Howmedica Osteonics (Newbury, UK); Zimmer (Swindon, UK); J&J DePuy (Leeds, UK); Corin Medical (Cirencester, UK); Smith & Nephew Healthcare Ltd (Cambridge, UK); Biomet Merck Ltd (Bridgend, UK); and Wright Cremascoli (Woking, UK).

The project management group

The trial was overseen by a project management group (see Appendix 4 for membership). This group met at variable intervals, typically 4-monthly, usually by teleconference, through the course of the study.

The data monitoring committee

Accumulating trial data were periodically reviewed by a data monitoring committee, independent of the trial organisers (see Appendix 4 for membership). The committee had three members: a statistician with experience of monitoring accumulating RCT data (who also acted as chairperson); an orthopaedic surgeon who was not involved in the trial; and a clinician with experience of RCTs.

The committee met four times between February 2002 and March 2005 at approximately yearly intervals, as decided by the committee. The committee’s terms of reference were guided by the Peto approach to data monitoring in RCTs (see protocol in Appendix 1). Using this to recommend a change in the protocol (such as stopping recruitment in one or all elements) requires both (1) proof beyond reasonable doubt that for all or some types of participants one particular type of prosthesis is clearly indicated or contraindicated (often taken as three SEs difference in the primary outcome) and (2) evidence that might reasonably be expected to influence materially the care of people who require knee replacement by clinicians who know the results of this and comparable trials. On each occasion, the committee recommended continuation of the trial with no change of protocol. All other people, including the project management group, clinical collaborators and trial staff (except those who supplied the confidential analyses), remained ignorant of the interim results considered by the committee. The committee stood down once the initial trial results up to 2 years after surgery were analysed.

Participating centres

In total, 116 surgeons in 34 centres in the UK participated in KAT (Table 2).

Numbers recruited to each comparison

From July 1999 to January 2003, 4070 potentially eligible participants were identified and 2374 (58%) gave their consent and were randomised (Figure 2). The main reasons for non-randomisation were the participant’s refusal to take part in the trial (546; 32%); the surgeon not wanting the participant to be randomised (462; 27%); a missed opportunity to recruit a scheduled participant (351; 21%); cancellation or deferral of the surgery or non-attendance on the part of the participant (84; 5%); the surgeon undertaking the procedure not being registered to participate in the trial (38; 2%); unavailability of necessary equipment (24; 1%); and unknown reasons (45; 3%). Subsequently, 22 participants were found to have been randomised in error: 14 were randomised twice, 3 were not eligible, 3 were treated by surgeons who were not registered to participate in the comparison and 2 were excluded for other reasons. This left 2352 participants formally in the trial: 1715 were included in the comparison assessing the patellar resurfacing; 539 in the comparison assessing the mobile bearing; 409 in the comparison assessing the metal backing; and 34 in the comparison assessing total versus unicompartmental knee replacement. There were 345 participants randomised in more than one comparison (see Figure 1). Separate CONSORT diagrams are presented for each comparison in the individual comparison chapters (see Chapters 3–6).

FIGURE 2.

Recruitment graph.

Oxford Knee Score

There was no evidence of a between-group difference in OKS at baseline or any stage thereafter (Table 6). The mean OKS in both patellar resurfacing and non-resurfacing groups was 18 preoperatively. It increased to 35 at 1 year and thereafter remained about the same, although it did decrease slightly in the long term (Figure 4). The estimated adjusted difference in OKS between the two groups was 0.45 (95% CI –0.66 to 1.56) at 10 years (see Table 6). The marginal estimate over the whole 10-year follow-up was 0.61 (95% CI –0.23 to 1.44; p = 0.153), a difference in favour of the patellar resurfacing intervention, which reflects slightly, but consistently, higher values in favour of patellar resurfacing over the whole 10-year follow-up (Figure 5). Sensitivity analysis imputing the last value before revision gave practically identical results; the marginal estimate was 0.73 (95% CI –0.12 to 1.58; p = 0.09).

FIGURE 4.

Mean (SD) OKS by group at each follow-up time point for patella comparison. PR, patellar resurfacing.

FIGURE 5.

Estimated treatment effect on OKS (95% CI) at each follow-up time point for patella comparison. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours patellar resurfacing.

The distribution of the OKS at 10 years in the two groups is shown in Figure 6. The distributions are very similar, and in particular there are few participants with a poor outcome in either group. Question 12 of the OKS enquires about symptoms relating to stair descent. This item showed a similar pattern to the OKS, that is a small but consistent difference in favour of patellar resurfacing over the whole 10-year follow-up (Figure 7).

FIGURE 6.

Histogram of OKS at 10 years by treatment group for patella comparison. PR, patellar resurfacing.

FIGURE 7.

Estimated treatment effect on OKS question 12, stairs descent (95% CI), at each follow-up time point for patella comparison. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours patellar resurfacing.

Subgroup analysis

There were two post-hoc subgroup analyses proposed for the primary outcome in this comparison: the shape of the groove in the femoral component (anatomical vs. domed) and age (< 70 vs. ≥ 70 years). The shape of the groove is described as either anatomical or domed, depending on whether it is designed to articulate with an anatomically shaped or domed-shaped patella button.

All prostheses were classified as anatomical or domed. Of the 1614 participants included in the analysis of the primary outcome, OKS, 48 (3%) could not be classified given the information recorded. The subgroup analysis was run three times: first excluding the unknowns, and then again reclassifying these 48 as anatomical, and then domed. The breakdown of shape was the same in each arm of the trial: 29% anatomical, 68% domed and 3% unclassified. Figure 8 is a plot of the interaction term (or the difference in differences) at each time point for the anatomical shape by patellar resurfacing interaction. The estimates are about zero with 95% CIs fairly wide throughout, reflecting no evidence of a shape effect modification on the primary outcome, where a positive difference would indicate that resurfacing was more favourable if the prostheses were anatomical.

FIGURE 8.

Interaction term (solid line) at each follow-up time point for OKS anatomical shape by patellar resurfacing. Dotted lines represent 95% CIs.

Figure 9 plots the difference in favourable effect for patellar resurfacing for those aged < 70 years, compared with those ≥ 70 years. A positive difference in differences suggests a higher relative benefit for resurfacing in the younger age group. After an early peak favouring patellar resurfacing in the younger subgroup, the difference in differences settles around zero and indicates no evidence for a treatment modification on the primary outcome OKS by age.

FIGURE 9.

Interaction term (solid line) at each follow-up time point for OKS age < 70 years old by patellar resurfacing. Dotted lines represent 95% CIs.

EuroQol 5D

There was no evidence of a between-group difference in EQ-5D at baseline or at any stage thereafter (Table 7). The mean EQ-5D utility was about 0.40 preoperatively. It increased to about 0.74 at 1 year and thereafter steadily decreased to about 0.66 at 10 years (Figure 10). At 10 years, the difference in EQ-5D was 0.012 (95% CI –0.018 to 0.042) (see Table 7). The marginal estimate over the whole 10-year follow-up was 0.011 (95% CI –0.008 to 0.030; p = 0.27) in favour of the patellar resurfacing intervention (Figure 11).

FIGURE 10.

Mean (SD) EQ-5D utility by group at each follow-up time point for patella comparison. PR, patellar resurfacing.

FIGURE 11.

Estimated treatment effect on EQ-5D utility (95% CI) at each follow-up time point for patella comparison. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours patellar resurfacing.

Short Form 12

There was no evidence of a between-group difference in SF-12 at baseline or at any stage thereafter. SF-12 physical component score (PCS) was 31 for both groups preoperatively (Table 8). It increased to 41 at 1 year and thereafter slowly decreased to 37 for both groups at 10 years (Figure 12). The marginal estimate over the whole 10-year follow-up was 0.40 (95% CI –0.78 to 1.57; p = 0.51) in favour of the patellar resurfacing intervention (Figure 13).

FIGURE 12.

Mean (SD) SF-12 PCS by group at each follow-up time point for patella comparison. PR, patellar resurfacing.

FIGURE 13.

Estimated treatment effect on SF-12 PCS (95% CI) at each follow-up time point for patella comparison. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours patellar resurfacing.

The mean SF-12 mental component score (MCS) was about 50 for both groups preoperatively. It increased to about 52 at 1 year and then decreased slowly to 49 at 10 years (Figure 14 and Table 9). The marginal estimate over the whole 10-year follow-up was 0.56 (95% CI –0.16 to 1.23; p = 0.13) in favour of the patellar resurfacing intervention (Figure 15).

FIGURE 14.

Mean (SD) SF-12 MCS by group at each follow-up time point for patella comparison. PR, patellar resurfacing.

FIGURE 15.

Estimated treatment effect on SF-12 MCS (95% CI) at each follow-up time point for patella comparison. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours patellar resurfacing.

Clinical outcomes

During the first 10 postoperative years, 15% (122/841) of the resurfaced group and 15% (128/830) of the non-resurfaced group required readmission and/or further intervention (Table 10; odds ratio 0.93, 95% CI 0.71 to 1.23, p = 0.63); 7% (58/841) of the resurfaced group and 8% (67/830) of the non-resurfaced group required further minor or intermediate operations (odds ratio 0.85, 95% CI 0.53 to 1.32, p = 0.39); 2% (15/841) of the resurfaced group and 2% (16/830) of the non-resurfaced group required patellar-related operations (odds ratio 0.93, 95% CI 0.46 to 1.90, p = 0.85); and 3% (26/841) of the resurfaced group and 5% (39/830) of the non-resurfaced group required other further major operations (odds ratio 0.65, 95% CI 0.39 to 1.10, p = 0.11). The reasons for further surgery included infection, pain, stiffness, loosening and instability. There was no statistically significant difference in the proportion of participants requiring further surgery in the resurfaced or non-resurfaced groups for any of the different levels of secondary intervention. The majority of the readmissions and reoperations were in the first 5 years (81%). Late patellar resurfacing was carried out on 1.9% (16/830) of the non-resurfaced group and 1.1% (9/841) of the resurfaced group (odds ratio 0.55, 95% CI 0.24 to 1.26, p = 0.16). The reason why some resurfaced group participants had a late resurfacing was that, although they were allocated to resurfacing, they did not have resurfacing at the original operation. For reasons stated previously, 129 of the 861 (15.0%) participants allocated to resurfacing crossed over clinically and did not have the patellar resurfaced at the primary procedure. Of these 129 participants, 9 (7%) had subsequent late resurfacing of the patella. Conversely, 761 of the 854 allocated to the non-resurfaced group did not have a resurfacing at the primary operation; 16 of these 761 participants (2.1%) had late patellar resurfacing. All the ‘late’ patellar resurfacing procedures took place in the first 5 years. The reasons for these, when recorded, were either ‘pain’ or ‘pain and/or stiffness’. During the second 5 years, there were six patella-related reoperations. They were all in the patella resurfaced group and they were all the result of complications of the patellar resurfacing: two were for patella fracture, two were patella revisions, one was the removal of a patella button and one was patella realignment for patella dislocation. Time-to-event analyses showed that there was no evidence of a difference between the randomised groups on time to any major reoperation or patella-related operation (hazard ratio 0.75; 95% CI 0.50 to 1.14; p = 0.18; Figure 16); time to any reoperation (hazard ratio 0.87; 95% CI 0.65 to 1.17; p = 0.35; Figure 17); or time to any reoperation or OKS dropping to below baseline levels beyond 1 year (hazard ratio 0.94; 95% CI 0.78 to 1.12; p = 0.47; Figure 18).

FIGURE 16.

Kaplan–Meier failure curves for time to first major reoperation or patella-related operation for patella comparison. PR, patellar resurfacing.

FIGURE 17.

Kaplan–Meier failure curves for time to any reoperation for patella comparison. PR, patellar resurfacing.

FIGURE 18.

Kaplan–Meier failure curves for time to any reoperation or OKS dropping below baseline level at 1 year or later for patella comparison. PR, patellar resurfacing.

The OKS for those participants who had late resurfacing is shown in Figure 19. Prior to the resurfacing, the OKS was found to deteriorate. In the year before the late resurfacing, the mean OKS was 15.9 (standard deviation 8.9). After the procedure, OKS improved again, averaging 21.3 (standard deviation 8.9) during the second postoperative year. This was higher than before the late resurfacing, but the OKS for these patients remained considerably lower than the mean OKS for the whole trial group.

FIGURE 19.

Mean (95% CI) OKS for the years pre and post late resurfacing for patella comparison.

Base-case analysis

The total cost accrued in each year of the trial generally decreased over time, reflecting mortality, the decreasing probability of readmission and the falling intensity of outpatient follow-up, although some oscillations were observed in later years due to variations in the small number of readmissions each year (Table 12, see Figure 20). In the first year, participants randomised to patellar resurfacing accrued non-significantly higher costs as a result of the added cost of patella components (p = 0.67). However, costs were markedly, but not significantly, lower in the patellar resurfacing group in years 2 (p = 0.42), 3 (p = 0.10) and 4 (p = 0.26) and similar thereafter.

As discussed above, participants had a very poor quality of life at baseline, with a mean baseline utility of around 0.4. Following TKR, utility rose in both groups to around 0.69 at 3 months and 0.74 at 1 year. Quality of life was higher in the patellar resurfacing group at all time points, although differences did not reach statistical significance (all p > 0.05).

From EQ-5D utilities, QALYs were calculated by taking the area under the EQ-5D curve, with multiple imputation of missing data and adjustments for mortality, time preference censoring and the small imbalance in baseline EQ-5D utility. This suggested that the average participant experienced around 0.7 QALYs during each of the first 3 years after TKR (see Table 12). However, as the QALY metric also allows for mortality (assigning participants a utility of zero after death) and the QALYs in Table 12 are discounted to allow for the fact that society places a lower value on benefits accrued in the future, the numbers of QALYs observed beyond year 2 decreased substantially more quickly than the mean EQ-5D utility.

The patellar resurfacing group had higher EQ-5D utility (see Table 7) and accrued more QALYs (see Table 12) than the no resurfacing group in every year. In many cases, the difference in QALYs accrued in a given year was larger than the difference in EQ-5D utility, and QALY differences (unlike those for EQ-5D utility) reached statistical significance in years 3 and 4. The larger QALY differences appear to be partly the result of multiple imputation of missing utility values and partly the result of a non-significant difference in mortality, as mean survival was around 21 days longer for participants randomised to patellar resurfacing (p = 0.60). Across the first 10 years after TKR, the average participant randomised to patellar resurfacing accrued 5.30 QALYs: 0.19 more than the no patellar resurfacing group (p = 0.08).

At a 10-year time horizon, the patellar resurfacing group, therefore, accrued more QALYs and lower costs than the group allocated to no patellar resurfacing. Patellar resurfacing can, therefore, be said to dominate no patellar resurfacing, being more effective and less costly.

However, differences in neither costs (p = 0.70) nor QALYs (p = 0.08) reached the conventional level of statistical significance. Plotting incremental costs and incremental QALYs on the cost-effectiveness plane (Figure 21) demonstrates that there is a 63% probability that patellar resurfacing dominates no resurfacing, a 34% probability that it is more costly and more effective and only a 4% probability that it is less effective.

FIGURE 21.

Stochastic cost-effectiveness results for patellar resurfacing vs. no resurfacing: scatter graph on cost-effectiveness plane.

Although there is no significant difference in either costs or QALYs individually, the joint distribution of costs and QALYs shows that we can be reasonably confident that patellar resurfacing is either dominant or produces health gains that are large compared with its incremental cost. In most NHS decision-making, treatments are considered cost-effective if they cost no more than a ‘ceiling ratio’ of around £20,000 per QALY gained. 81 The distribution of costs and QALYs observed here shows that we can be > 95% confident that patellar resurfacing is good value for money if the NHS is willing and able to pay at least £7250 per QALY gained (Figure 22), suggesting that patellar resurfacing is very good value for money.

FIGURE 22.

Cost-effectiveness acceptability curve for patella comparison.

Sensitivity analysis

However, as with all economic evaluations, our analysis required a number of assumptions and choices among a number of alternative methodologies. We, therefore, conducted sensitivity analyses to assess the impact of using different methods or assumptions.

These analyses demonstrate that the results are extremely robust, with patellar resurfacing dominating no resurfacing and having a > 95% probability of being cost-effective at a ceiling ratio of £20,000 per QALY gained in every analysis except for the complete case analysis (Table 13). In particular, the analysis is robust to changes in the time horizon and the discount rates used to adjust for time preference, as the majority of differences in costs and QALYs are in the first few years after primary TKR (which are given higher weight). Changes in costing methodology also have relatively little impact because the conclusion is driven by the comparatively large difference in QALYs and readmissions. It is particularly notable that changing the discount applied to component list prices has no effect on the conclusions and that the 46% reduction in length of stay since KAT operations were completed has little effect on incremental costs, as length of stay is similar in the two groups. Changing the methods used to deal with censoring and imbalance in baseline utility also had minimal impact.

By contrast, the complete case analysis found patellar resurfacing to be substantially more costly and only slightly more effective than no resurfacing, costing £49,160 per QALY gained. As treatments that cost < £20,000 per QALY gained are normally considered cost-effective in NHS decision-making,81 the complete case analysis would suggest that patellar resurfacing is poor value for money. However, by excluding all participants who had missing data on any variable collected before the participant was administratively censored (including data from up to 14 questionnaires completed over 10 years), the complete case analysis excludes more than half of the sample, substantially reducing statistical power. Furthermore, complete case analyses are prone to bias, as they do not include all randomised participants and assume that data are missing completely at random. 65 Bias is particularly likely here, as participants cannot have missing data on costs or QALYs after they have died or been administratively censored. Furthermore only participants who receive a patella can have missing data on the cost of a patella and only participants who are readmitted can have missing data on component costs or length of stay during subsequent hospitalisations. Consequently, the results of this analysis should be interpreted with caution.

Subgroup analysis

As surgeons often consider participants’ age and their likelihood of out-living their knee prosthesis in decisions around component design, a post-hoc subgroup analysis estimated outcomes for subgroups divided by age (see Table 13). Younger participants (< 70 years old at the time of TKR) accrued higher total costs and more QALYs within 10 years of TKR than those aged ≥ 70 years, presumably because of their longer life expectancy. For younger participants, the estimated cost savings and QALY gains from patellar resurfacing were greater than those of the base case, although the finding that patellar resurfacing dominated no resurfacing was the same. By contrast, older participants randomised to patellar resurfacing accrued higher costs than those randomised to no resurfacing and the QALY gains from resurfacing were also lower than in the base case; as a result, patellar resurfacing cost £1629 per QALY gained in participants aged ≥ 70 years, but remained cost-effective.

As discussed above, patella components may be either a domed shape or an anatomical shape, with corresponding changes to the shape of the trochlear groove within the femoral component. It is generally assumed that a non-resurfaced patella would perform better with an anatomical trochlea, rather than one designed for a domed patella button. Subgrouping participants by the shape of the groove within the femoral component suggested that the effectiveness of patellar resurfacing is slightly better among participants with a femoral groove shaped for an anatomical patella than among those with grooves shaped for domed patellas, while patellar resurfacing was also less costly than no resurfacing in the group with anatomical patellas, but more costly in those with domed patellas. However, patellar resurfacing remained very good value for money in both groups. Costs were also higher among participants with anatomical patellofemoral grooves than among those with dome-shaped grooves.

Oxford Knee Score

There was no evidence of a between-group difference in OKS at baseline or at any stage thereafter (Table 17). The mean OKS in both mobile bearing and fixed bearing groups was approximately 17 at baseline. It increased to approximately 33 at 1 year and thereafter remained about the same, although it did decrease slightly in the long term (Figure 24). The difference in OKS between the two groups was small, 0.28 (95% CI –1.86 to 2.43) at 10 years (see Table 17). The marginal estimate over the whole 10-year follow-up was 0.29 (95% CI –1.17 to 1.75; p = 0.70) in favour of the mobile bearing intervention (Figure 25). Sensitivity analysis imputing the last value before revision gave practically identical results; the marginal estimate was 0.34 (95% CI –0.16 to 1.85; p = 0.65).

FIGURE 24.

Mean (SD) OKS by group at each follow-up time point for mobile vs. fixed bearing.

FIGURE 25.

Estimated treatment effect on OKS (95% CI) at each follow-up time point for mobile vs. fixed bearing. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours mobile bearing.

Figure 26 explores the potential for interaction in those allocated to two interventions by plotting difference in OKS between mobile and fixed bearings among those allocated to patellar resurfacing compared with those allocated to no patellar resurfacing. A positive difference in differences suggests a higher relative benefit for mobile bearings in the patellar resurfacing group. The graph indicates that there may be a potential interaction, suggesting patellar resurfacing when using a mobile bearing may be beneficial. However, there is considerable uncertainty around these estimates as a result of the reduced sample size in the partial factorial aspect of the trial.

FIGURE 26.

Interaction term (solid line) at each follow-up time point for OKS for patellar resurfacing by mobile bearing. Dotted lines represent 95% CIs.

EuroQol 5D

There was no evidence of a between-group difference in EQ-5D at baseline or at any stage thereafter (Table 18). The mean EQ-5D utility was approximately 0.32 at baseline. It increased to approximately 0.70 at 1 year and thereafter steadily decreased to about 0.67 at 10 years (Figure 27). At 10 years the difference in EQ-5D was 0.041 (95% CI –0.021 to 0.104) (see Table 18). The marginal estimate over the whole 10-year follow-up was 0.018 (95% CI –0.020 to 0.056; p = 0.36) in favour of the mobile bearing intervention (Figure 28).

FIGURE 27.

Mean (SD) EQ-5D utility by group at each follow-up time point for mobile vs. fixed bearing.

FIGURE 28.

Estimated treatment effect on EQ-5D utility (95% CI) at each follow-up time point for mobile vs. fixed bearing. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours mobile bearings.

Short Form 12

There was no evidence of a between-group difference in SF-12 measured either at baseline or at any stage thereafter. Mean SF-12 PCS was approximately 31 for both groups at baseline (Table 19). It increased to approximately 39 at 1 year and thereafter slowly decreased to approximately 36 for both groups at 10 years (Figure 29). The difference in score at 10 years was –0.15 (95% CI –2.37 to 2.07) (see Table 19). The marginal estimate over the whole 10-year follow-up was 0.19 (95% CI –1.26 to 1.64; p = 0.79) (Figure 30).

FIGURE 29.

Mean (SD) SF-12 PCS by group at each follow-up time point for mobile vs. fixed bearing.

FIGURE 30.

Estimated treatment effect on SF-12 PCS (95% CI) at each follow-up time point for mobile vs. fixed bearing. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours patellar resurfacing.

The mean SF-12 MCS was 48 for both groups preoperatively (Table 20). It increased to about 50 at 1 year and then decreased slowly to 48 at 10 years (Figure 31). The difference in score at 10 years was –0.91 (95 %CI –3.31 to 1.48). The marginal estimate over the whole 10-year follow-up was –0.18 (95% CI –1.41 to 1.26; p = 0.91) (Figure 32).

FIGURE 31.

Mean (SD) SF-12 MCS by group at each follow-up time point for mobile vs. fixed bearing.

FIGURE 32.

Estimated treatment effect on SF-12 MCS (95% CI) at each follow-up time point for mobile vs. fixed bearing. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours patellar resurfacing.

Clinical outcomes

During the first 10 postoperative years, 16% (41/262) of the mobile bearing group and 18% (45/255) of the fixed bearing group required readmission and/or further intervention (odds ratio 0.84; 95% CI 0.52 to 1.34; p = 0.47; Table 21); 8% (22/262) of the mobile bearing group and 6% (16/255) of the fixed bearing group required further minor or intermediate operations (odds ratio 1.36; 95% CI 0.69 to 2.68; p = 0.37); and 3% (9/262) of the mobile bearing group and 3% (8/255) of the fixed bearing group required other further major operations (odds ratio 1.07; 95% CI 0.36 to 3.25; p = 1.00). There were six reoperations for bearing dislocation or instability in the mobile bearing group in five participants (2%) compared with none in the fixed bearing group (p = 0.062). Time-to-event analyses showed that there was no evidence of a difference between the randomised groups on time to any major reoperation or reoperation for instability (hazard ratio 1.47; 95% CI 0.60 to 3.61; p = 0.39; Figure 33); time to any reoperation (hazard ratio 1.39; 95% CI 0.81 to 2.37; p = 0.23; Figure 34); or time to any reoperation or OKS dropping to below baseline levels beyond 1 year (hazard ratio 0.93; 95% CI 0.67 to 1.29; p = 0.66; Figure 35).

FIGURE 33.

Kaplan–Meier failure curves for time to first major reoperation for mobile vs. fixed bearing.

FIGURE 34.

Kaplan–Meier failure curves for time to any reoperation for mobile vs. fixed bearing.

FIGURE 35.

Kaplan–Meier failure curves for time to any reoperation or OKS dropping below baseline score for mobile vs. fixed bearing.

Base-case analysis

Following the quality-of-life trends described above, the mobile bearing arm accrued non-significantly more QALYs during the first 4 years after TKR (minimum p = 0.14). However, the mobile bearing arm accrued fewer QALYs in most subsequent years (minimum p = 0.59; Table 23), despite life expectancy being 9.29 years in the two groups (p = 0.98). Over the 10-year time horizon, the larger quality of life increases observed in earlier years outweighed the quality of life decreases seen in later years and the mobile bearing group therefore accrued 0.051 (95% CI –0.333 to 0.435) more QALYs than the fixed bearing group (p = 0.79).

Mobile bearings were therefore associated with non-significantly higher costs (mean difference £85; p = 0.87) and marginally more QALYs (mean difference 0.051; p = 0.79) over the 10-year time horizon. The ICER for mobile bearings is therefore £1666 per QALY gained compared with fixed bearings, although there is substantial uncertainty around this point estimate. In NHS decision-making, treatments that increase health and NHS costs are generally considered to be good value for money if they have an ICER below £20,000 per QALY gained,81 making mobile bearings highly cost-effective based on their ICER point estimate, although there remains substantial uncertainty around this figure.

However, there was substantial uncertainty around both incremental costs and incremental QALYs, with a joint distribution spread across the four quadrants of the cost-effectiveness plane (Figure 37). There was a 32% probability that mobile bearings were more costly and more effective, a 29% probability that mobile bearings dominated fixed bearings (being less costly and more effective), a 27% probability that mobile bearings were dominated and a 13% probability that they were less costly and less effective (south-west quadrant). In particular, the uncertainty meant that the cost-effectiveness acceptability curve was very flat, with the probability of mobile bearings being cost-effective varying between 42% and 60% (Figure 38). At a £20,000/QALY ceiling ratio, the probability of mobile bearings being cost-effective was 59%.

FIGURE 37.

Stochastic cost-effectiveness results for mobile vs. fixed bearing: scatter graph on cost-effectiveness plane.

FIGURE 38.

Cost-effectiveness acceptability curve for mobile vs. fixed bearing.

Sensitivity analyses

Sensitivity analyses suggested that the base-case conclusions were sensitive to the methods used to deal with missing data and protocol violations (Table 24). The complete case analysis, based on 96 and 97 participants (excludes all participants with missing data on any resource-use variable or quality-of-life measurement prior to death or administrative censoring), found mobile bearings to be substantially more costly and marginally less effective than fixed bearings. The per-protocol analysis also found mobile bearings to be dominated by fixed bearings (being more costly and less effective).

However, all other sensitivity analyses were consistent with the base-case finding that participants randomised to mobile bearings accrued marginally higher QALYs than those randomised to fixed bearings and that there is a 51–65% probability that mobile bearings are cost-effective at a £20,000/QALY ceiling ratio, although two analyses (increasing the discount of component prices and increasing the cost per bed-day) found mobile bearings to dominate fixed bearings. Time horizon had a marked effect on both incremental costs and incremental QALYs because of the tendency for participants assigned to mobile bearing to accrue fewer QALYs and higher costs in years 5–11. At an 8-year time horizon, the estimated incremental QALYs were greater than those in the base-case analysis, whereas both lower incremental costs and lower incremental QALYs were observed at an 11-year time horizon. This may suggest that a longer follow-up could reverse the direction of differences in both costs and QALYs, making mobile bearings less costly and less effective than fixed bearings.

Subgroup analyses

Subgroup analyses suggested that both the costs and benefits of mobile bearings differ with age (see Table 24). In particular, both the incremental costs and incremental health benefits of mobile bearings were markedly larger for participants aged < 70 years; although the cost-effectiveness ratio for this group was similar to that of the total population, the probability that mobile bearings were cost-effective compared with fixed bearings rose to 86%, compared with 59% in the base-case analysis, despite the smaller sample size. By contrast, mobile bearings were less costly but produced a smaller QALY gain than fixed bearings in older participants, saving £618 per QALY lost, which would not be considered good value for money.

Potential for interactions between mobile bearings and patellar resurfacing

Analysing the subset of 240 participants randomised to both the mobile bearing and patellar resurfacing comparisons as a factorial trial, it was found that there were interactions that had a marked effect on estimated incremental costs and QALYs. In particular, non-significant qualitative interactions between the two treatment allocation factors were observed for costs (p = 0.12), QALYs (p = 0.08) and net monetary benefits (p = 0.06), which means that the incremental effect of mobile bearings changes sign depending on whether participants were allocated to patellar resurfacing or no resurfacing. [Net monetary benefits are a linear measure of cost-effectiveness that facilitates statistical analysis and comparison of multiple groups. The total net monetary benefit was calculated by multiplying the total number of QALYs accrued in each treatment arm by the £20,000/QALY ceiling ratio and subtracting total costs.] In particular, participants randomised to both mobile bearing and no patellar resurfacing accrued substantially higher costs and substantially fewer QALYs than those allocated to the other three combinations of treatment allocation. As a result of the interactions for QALYs and costs, mobile bearings dominated fixed bearings in participants who were also randomised to patellar resurfacing and had a 93% chance of being cost-effective at the £20,000/QALY ceiling ratio, but were dominated by fixed bearings, with a 17% chance of being cost-effective in participants randomised to no resurfacing.

As there is evidence that the incremental costs and benefits of mobile bearings may depend on whether or not participants are also randomised to patellar resurfacing, it is useful to also examine whether making a joint decision about these two aspects of TKR (rather than independent decisions) would change the conclusions. Treating the four combinations of treatment allocation as mutually exclusive strategies for TKR suggests that fixed bearings without patellar resurfacing dominate fixed bearings with patellar resurfacing, and mobile bearings without resurfacing are less costly and more effective than both of these alternatives. However, the strategy with highest clinical effectiveness and cost-effectiveness comprises mobile bearing with patellar resurfacing, which costs £1109 per QALY gained compared with fixed bearing and no patellar resurfacing. Taking account of interactions and evaluating the cost-effectiveness of mobile bearings and patellar resurfacing simultaneously is, therefore, consistent with the conclusion of the base-case analysis that both mobile bearings and patellar resurfacing are expected to be cost-effective.

Oxford Knee Score

There was no evidence of a between-group difference in OKS at baseline or at any stage thereafter (Table 28). The mean OKS in both groups was about 17.5 preoperatively. It increased to about 33 at 1 year and thereafter remained about the same. The difference in OKS between the two groups was small and was –1.19 (95% CI –3.48 to 1.11; p = 0.311) at 10 years (Figure 40, see Table 28). The marginal estimate over the whole 10-year follow-up was –1.36 (95% CI –2.98 to 0.26; p = 0.10) in favour of the metal-backed intervention (Figure 41). Sensitivity analysis imputing the last value before revision gave practically identical results; the marginal estimate was –1.41 (95% CI –3.06 to 0.25; p = 0.10).

FIGURE 40.

Mean (SD) OKS by group at each follow-up time point for all-polyethylene vs. metal-backed tibial components.

FIGURE 41.

Estimated treatment effect on OKS (95% CI) at each follow-up time point for all-polyethylene vs. metal-backed tibial components. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours all-polyethylene components.

Figure 42 explores the potential for interaction in those allocated to two interventions by plotting the difference in effect for all-polyethylene components for those allocated to patellar resurfacing compared with those allocated to no patellar resurfacing. A positive difference in differences suggests higher relative benefit for all-polyethylene backing in the patellar resurfacing group. The graph indicates that there may be a potential interaction, suggesting that not resurfacing the patella, if using all polyethylene components, might be beneficial. However, there is considerable uncertainty around these estimates owing to the reduced sample size in the partial factorial aspect of the trial and estimating CIs around interaction terms.

FIGURE 42.

Interaction term (solid line) at each follow-up time-point for OKS for patellar resurfacing by all-polyethylene backing. Dotted lines represent 95% CIs.

EuroQol 5D

There was a trend towards the metal-backed group having better EQ-5D scores than the all polyethylene group (Figure 43). In 3 years (years 4, 5 and 9), the p-value was < 0.05 (Table 29, Figure 44). The marginal estimate over the whole 10-year follow-up was –0.042 (95% CI –0.081 to –0.003; p = 0.033) in favour of the metal-backed intervention (see Figure 44).

FIGURE 43.

Mean (SD) EQ-5D utility by group at each follow-up time point for all-polyethylene vs. metal-backed tibial components.

FIGURE 44.

Estimated treatment effect on EQ-5D utility (95% CI) at each follow-up time point for all-polyethylene vs. metal-backed tibial components. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours all-polyethylene components.

Short Form 12

There was a trend towards the metal-backed group having a significantly better SF-12 PCS than the all polyethylene group (Figure 45). In 3 years (3, 4 and 9) the p-value was < 0.05 (Table 30, Figure 46). The marginal estimate over the whole 10-year follow-up was –1.63 (95% CI –3.19 to –0.069; p = 0.041) in favour of the metal-backed intervention.

FIGURE 45.

Mean (SD) SF-12 PCS by group at each follow-up time point for all-polyethylene vs. metal-backed tibial components.

FIGURE 46.

Estimated treatment effect on SF-12 PCS (95% CI) at each follow-up time point for all-polyethylene vs. metal-backed tibial components. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours all-polyethylene components.

The SF-12 MCS was similar between the two groups at most time points (Table 31, Figure 47). The marginal estimate over the whole 10-year follow-up was –0.22 (95% CI –1.73 to 1.29; p = 0.77), a minimal difference between groups (Figure 48).

FIGURE 47.

Mean (SD) SF-12 MCS by group at each follow-up time point for all-polyethylene vs. metal-backed tibial components.

FIGURE 48.

Estimated treatment effect on SF-12 MCS (95% CI) at each follow-up time point for all-polyethylene vs. metal-backed tibial components. Treatment effect is the estimated difference between treatments at that time point adjusted for minimisation covariates and surgeon; a positive difference favours all-polyethylene components.

Clinical outcomes

There were 21/203 (10%) and 28/199 (14%) participants who were readmitted over the 10-year follow-up period (Table 32) (odds ratio 0.72; 95% CI 0.39 to 1.31; p = 0.30). The reoperation rate for minor and intermediate operations was very similar between the two groups (odds ratio 0.85; 95% CI 0.37 to 1.91; p = 0.81), and the majority of these reoperations took place within the first 5 years. Major reoperations were slightly more common in the all polyethylene group, but were rare overall (odds ratio 2.32; 95% CI 0.52 to 14.16; p = 0.35). The majority of the major operations took place in the first 5 years. Time-to-failure analysis of time until the first major reoperation estimated the hazard ratio as 2.30 (95% CI 0.60 to 8.90; p = 0.23; Figure 49), reflecting the higher number of reoperations in the all polyethylene group, but the CI is wide owing to the low number of reoperations. For time to any reoperation, the estimated hazard ratio was 0.85 (95% CI 0.44 to 1.68; p = 0.65; Figure 50). Broadening the definition of failure, to include any participant with an OKS after 1 year dropping below the baseline reported OKS, resulted in a hazard ratio of 1.45 (95% CI 0.99 to 2.18; p = 0.056; Figure 51); participants in the all polyethylene group were more likely to fail by this definition.

FIGURE 49.

Kaplan–Meier failure curves for time to first major reoperation for all-polyethylene vs. metal-backed tibial components.

FIGURE 50.

Kaplan–Meier failure curves for time to any reoperation for all-polyethylene vs. metal-backed tibial components.

FIGURE 51.

Kaplan–Meier failure curves for time to any reoperation or OKS dropping below baseline level at 1 year or later for all-polyethylene vs. metal-backed tibial components.

Base-case analysis

As was the case for the mobile bearing comparison, total cost in years 3–12 was primarily driven by readmissions and was low in the years in which no participants were admitted (Figure 52). Nonetheless, incremental cost was higher in the all polyethylene group in all years other than year 1 (in which costs were driven by the cost of the primary TKR procedure) and years 5 and 9 (in which the metal-backed group had more readmissions and more outpatient consultations; Table 34).

FIGURE 52.

Illustration of cost breakdown by year after discharge from hospital for all-polyethylene vs. metal-backed tibial components. Error bars show SEs around total cost.

However, QALYs showed a consistent trend over the first 10 years after primary TKR, being consistently (but not statistically significantly) lower in the all polyethylene group than in those randomised to metal-backed tibial components at every time point (minimum p = 0.06). The difference in QALYs tended to increase over time, suggesting that the long-term benefits of metal-backed components may be greater than is observed with a 10-year time horizon. This trend appears to be a result of increasing differences in quality of life, as life expectancy was actually around 36 days longer in the all polyethylene group (p = 0.66).

Over the 10-year time horizon, the incremental cost of all-polyethylene tibial components compared with metal-backed components was –£10 (95% CI –£872 to £851; p = 0.98), whereas the incremental QALY gain was –0.293 (95% CI –0.706 to 0.119; p = 0.16). The evidence from KAT therefore suggests that all-polyethylene tibial components are less costly and less effective than metal-backed components, with the point estimate lying in the ‘south-west’ quadrant of the cost-effectiveness plane (Figure 53). In principle, treatments that are cost-saving and less effective than their comparators could increase the amount of health generated by NHS treatments, by freeing up resources that can be invested in other treatments that generate greater health gains than those lost by using the less effective treatment. Based on the £20,000 ceiling ratio typically used in NHS decision-making and assuming that the NHS has symmetrical preferences for losses and gains, treatments that are less effective and less costly would be considered good value for money if they saved at least £20,000 per QALY lost.

FIGURE 53.

Stochastic cost-effectiveness results for all-polyethylene vs. metal-backed tibial components: scatter graph on cost-effectiveness plane.

The base-case KAT results suggest that the NHS would save £10 and lose 0.293 QALYs for every participant treated with all-polyethylene tibial components rather than metal-backed components, which equates to an ICER of just £35 per QALY lost. This is substantially below the £20,000/QALY threshold, suggesting that all-polyethylene components are poor value for money and should not be used in place of metal-backed components, which cost just £35 per QALY gained compared with all polyethylene.

However, there is a modest amount of uncertainty around both incremental costs and incremental QALYs (see Figure 53). In particular, there is an 8% chance that all-polyethylene tibial components are more effective than metal-backed components and a 47% chance that they are more costly.

Taking account of the joint density of incremental costs and QALYs demonstrates that there is a 91% probability that metal-backed tibial components are good value for money compared with all-polyethylene components at a £20,000/QALY ceiling ratio (Figure 54).

FIGURE 54.

Cost-effectiveness acceptability curve for all-polyethylene vs. metal-backed tibial components.

Sensitivity analyses

Sensitivity analyses suggested that the all polyethylene group accrued higher costs than the metal-backed group in four scenarios (Table 35). First, as expected, increasing the discount on knee components, such that hospitals pay only 50% (not 70%) of the list price, reduces costs in the metal-backed group more than in the all polyethylene group. Second, increasing the cost per bed-day to £448/day (50% higher than the national average excess bed-day cost in England and Wales) increases the cost in the all polyethylene group more than in the metal-backed group because of the longer primary hospital stay and additional readmissions. Third, if future costs and benefits are not discounted to current values, costs increase proportionately more in the all polyethylene group, as the additional costs accrued beyond year 1 are given greater weight. Fourth, reducing the time horizon and excluding costs accrued in years 9 and 10 changes the conclusions owing to the readmission and outpatient/physiotherapy consultations that occurred in the all metal-backed group in year 9. In other analyses, the magnitude of the cost savings varied from £1 to £156.