Study Design

The research was based on a single center randomized controlled trial conducted at Kent & Canterbury Hospital (K&C). Research Ethics Committee approval was obtained before the commencement of the study. All study patients were recruited from the routine orthopaedic clinics, which they attended after referral either from their general practitioner (GP) or from the accident and emergency department. The aim was for trial patients to be representative of the range of knee problems seen in orthopedic outpatient clinics at a DGH. Eligibility for inclusion in the trial was assessed for all patients with a persisting knee problem who attended the clinics run by participating general orthopedic surgeons.

Patients were defined as suitable for the trial if:

• diagnostic or therapeutic arthroscopy was being considered (in the absence of MRI);
• there had been no previous major surgery in the injured knee, such as knee replacement (previous arthroscopy and partial meniscectomy did not exclude patients from the trial);
• there was no pre-existing chronic knee pathology;
• there was no serious condition requiring immediate attention, for example, a serious knee infection;
• there was no history or current experience of recurrent locking of the knee;
• patients were between sixteen and fifty-five years old;
• anterior knee pain was not the main clinical indication.

The sample size calculation was based on rate of surgery and assumed that without MRI the vast majority (that is, 90%) of patients presenting with a knee injury would be investigated (or directly treated) using arthroscopy. On the basis of published observational data (4), there was reason to believe that arthroscopy might be avoided in up to 28% of cases by the use of a preliminary MRI investigation. For such a difference between groups (that is, 90% without MRI against 64.8% with MRI) to be established as statistically significant (power=80%; p value <.05), the trial required a total sample of approximately 100 patients allocated evenly between arms.

Once written consent was obtained, study patients were allocated to one of two trial arms:

• investigation using an MRI scan (MRI trial arm);
• investigation using arthroscopy (No-MRI trial arm).

The on-site study researcher, by using randomly ordered opaque sealed envelopes, undertook the patient allocation process. Patients allocated to the MRI trial arm were booked for an MRI scan (median wait for scan: 29 days) and placed immediately on the arthroscopy waiting list, even though they may not have required surgical treatment. This management pathway was adopted on the advice of the Hospital Research Ethics Committee to ensure that patients were not “disadvantaged” as a result of participating in the trial. Patients were reviewed in an outpatient clinic after their scan and a decision on appropriate management was made. Routine clinical follow-up then continued until their knee recovered. Patients allocated to the No-MRI trial arm were immediately listed for arthroscopy, reviewed in clinic, both before and after surgery, and followed-up according to routine clinical practice until the knee problem resolved. It was clearly neither feasible nor sensible to blind the study patients, researchers, or those involved in providing care to the outcome of the allocation process and so an “open-label” policy was adopted.

Assessment of Benefits

By using the categorization of Fineberg et al.; this study investigated the benefits of knee MRI at two levels: diagnostic/therapeutic impact and patient outcome (13). The use of arthroscopy as a form of diagnosis as well as a therapeutic intervention complicates the picture for knee investigations and prevents a complete separation in this study of diagnostic and therapeutic impacts. The observed data on whether or not surgery was undertaken, over the twelve-month follow-up period, allow an assessment of the diagnostic and therapeutic impact of MRI. From a patient's perspective, diagnostic or therapeutic arthroscopies are virtually identical: both are typically day-case procedures involving similar recovery periods.

Benefits in terms of patient outcome were measured by using two generic health status measurement instruments: the SF-36 (5;15;21;30) and the EuroQol EQ-5D (7;8;19). Assessments of quality of life were taken at baseline (at the point of recruitment) and at six and twelve months after recruitment into the trial. The baseline questionnaire and the two follow-up questionnaires (at six and twelve months) were all distributed by post- with pre-paid return envelopes. Nonresponders were sent two reminder letters with a further copy of the questionnaire.

Cost Data

The perspective for the cost analysis reported here is the Health Service and patients, such that the main objective was to identify all important National Health Service (NHS) and patient resource use and cost differences between trial arms. The first step in the costing process was to collect data on the NHS resources devoted to trial patients. Data were collected on the following resource parameters: the MRI scan; surgery on the knee; other relevant diagnostic procedures; other relevant therapeutic procedures; relevant drugs prescribed; associated outpatient attendances; associated inpatient episodes; and relevant contacts with other health-care professionals, such as physiotherapists and GPs. Most of these data were extracted from the patient medical records up to twelve months after recruitment. Data on the use of community health-care resources were collected by using the patient follow-up questionnaires distributed at six and twelve months. The health-care resource use items were costed using unit cost information from three sources (see Table 1).

The costs incurred by patients in attending for an MRI scan, outpatient visit, or inpatient/day-case procedure were also investigated through a survey of a subsample of all trial patients. Total patient costs were calculated by summing costs associated with travel, time, and other expenses. Unit costs were obtained from a variety of published sources.

Data Analyses

All analyses were conducted on the basis of “intention to treat.” For the between-group comparisons of baseline patient characteristics and baseline quality of life, for continuous data, the analyses involved either t-tests or Mann-Whitney Wilcoxon tests, and for the comparisons of proportions, a π² test was used (1).

At each follow-up time point (that is, at both six and twelve months), between-group comparisons of quality of life scores were made using the same methods as adopted in the baseline comparison. However, because this approach takes no account of the longitudinal nature of the data, mean changes in quality of life scores were estimated separately for the zero- to six-month change, the zero- to twelve-month change, and the six- to twelve-month change, using a “complete case analysis.” To allow for potential differences between groups at baseline, quality of life scores were analyzed using regression analyses. A two-limit tobit model with random effects and sample selectivity was developed (technical details of this new technique are available from the authors). Models were estimated for the six SF-36 dimensions for which data can be treated as being continuous (that is, physical functioning, bodily pain, general health, vitality, social functioning, mental health) and the EQ-5D tariff and VAS scores (2). The independent variables included in the models were trial arm, hospital site, referral source, patient sex, and patient age at baseline.

The analysis of cost data initially involved exploring the distributional form of the data separately for the two arms of the trial, to identify the extent of any skew that might present problems for standard parametric statistical tests (6). Second, because skewed distributions for cost data are commonly found, and were expected a priori in this study, a nonparametric approach to the cost analysis was likely to be required. However, estimates of the mean cost are appropriate for use in economic evaluation, even where the data are skewed, because we are interested in both the average per patient cost of a particular treatment and (because there exists a budget constraint) the total cost of care for a patient group. Thus, the nonparametric approach of bootstrapping was used to estimate confidence limits around the estimates of mean cost (6;11).