Clinical trials most often measure efficacy – the ability of an intervention to produce the desired result. They tend not to measure effectiveness, which are the actual results found in clinical practice – a product of efficacy and adherence to therapy. The difference between trial efficacy and clinical effectiveness can be ascribed to the tendency of patients not to adhere to therapy, especially in the longer term. Discontinuation of therapy is most often due to lack of efficacy or unmanageable or intolerable adverse events, or both.

Patient adherence to therapy is known to be low, especially where therapy is prophylactic or does not deliver rapid symptomatic relief. In the USA it is estimated that only about 50% of patients continue statin therapy at 6 months, and 30% to 40% at 1 year [1]. In the UK 50% of patients prescribed low-dose aspirin have discontinued within a year [2]. Where benefit is greater and more tangible, adherence is likely to be higher, even if adverse events are common. Thus, among renal transplant patients, only 15% were found to be nonadherent to immunosuppressants using stringent criteria [3].

Adverse events are a major concern to patients. A Dutch study [4] found that half of 232 chronic prescriptions for long-term medicines were not refilled over 3 months, with adverse events being a major reason. Talking to patients about adverse events can help, but it may well depend on how the adverse event rates are described. In a randomized trial conducted in 120 patients given information about the adverse events associated with the medicine [5], it was found that the patients were more likely to be compliant, and had less fear, when they were presented with information about adverse events in percentage terms rather than in words.

Physician adherence to guidelines is also known to be low. Prescribing of nonsteroidal anti-inflammatory drug (NSAID) plus gastroprotective agent or cyclo-oxygenase-2 inhibitor (coxib) was identified in only 26% of patients with at least one gastrointestinal risk factor in a large systematic review of studies published since 2000 [6]. More or better professional education can improve matters, as with antihypertensive prescribing in Canada [7, 8].

Few patients appear able to take oral NSAIDs prescribed for chronic musculoskeletal problems for a long time. For instance, only 15% to 20% of those started on a study NSAID were still using the same drug at the end of 12 months in an observational study conducted in Seattle [9], whereas in a telephone survey [10], also conducted in the USA, drug continuation beyond 24 months was reported by 33% of patients for paracetamol (acetaminophen), 21% for ibuprofen, 17% for naproxen and 19% for diclofenac. Most studies are too small to measure differences in continuation rates between NSAIDs adequately, although recent larger studies have consistently observed longer continuation rates for coxibs over traditional nonselective NSAIDs [11, 12].

Company clinical trial reports do not suffer from the problems of selective reporting in published papers that results from strict word limitations. Although efficacy in published studies has in the past been poorly presented [13], adverse event information is even more poorly presented in published papers [14]. Both, but particularly adverse events, may be more clearly presented in company clinical trial reports. Company clinical trial reports potentially provide an ideal source of information for systematic review and meta-analysis, particularly on discontinuation rates for different causes [15, 16].

Meta-analysis of discontinuation rates in randomized trials of drug therapy has been undertaken infrequently. In psychiatry, studies in schizophrenia [17] and depression [18, 19] have sought to differentiate drugs by discontinuation rates, as have studies in Helicobacter treatment [20] or hypertension [21]. Discontinuation rates have been included in previous meta-analyses of clinical trials of analgesics in musculoskeletal pain [15, 16].

In this report we examine discontinuation rates in randomized trials of etoricoxib in chronic musculoskeletal conditions. Our prior hypotheses were as follows. First, discontinuation rates would differ between osteoarthritis (OA), rheumatoid arthritis (RA) and other musculoskeletal conditions for lack of efficacy, but probably not because of adverse events. Second, discontinuation rates would differ between shorter and longer trials. Third, discontinuations because of gastrointestinal adverse events would be lower with etoricoxib than with traditional NSAIDs. Finally, higher doses of etoricoxib would produce lower rates of lack of efficacy discontinuations and higher rates of adverse event discontinuations.

The data for this meta-analysis were contained in clinical trial reports of 18 randomized trials of etoricoxib lasting 4 to 12 weeks, and three trials with mean duration of more than 1 year. Merck Research Laboratories (Rahway, NJ, USA) provided clinical trial reports in the form of PDF (19 trials) or Word (two trials) documents. No additional literature searching was conducted.

Information from the clinical trial reports was tabulated, including diagnosis and trial entry criteria, demographic information, drugs, dose and duration of treatment. Patients who were randomized and received at least one dose of drug were considered the intention-to-treat population. Also extracted were the following outcomes: all-cause discontinuations; discontinuations due to lack of efficacy; discontinuations due to a clinical adverse event; discontinuations due to a laboratory adverse event (hepatic, renal, or haematological); and discontinuations for specific reasons, namely gastrointestinal adverse event, oedema, hypertension, or a serious adverse event.

In addition, information about the time course of discontinuations was sought. Where this information was in graphical form, printed versions were used to calculate the number of patients remaining after discontinuation for lack of efficacy or adverse event. Extrapolation was not done where the scale of graphs precluded accurate estimation.

The nomenclature for discontinuation does not appear to be fixed. Although discontinuations for a particular cause might be described as a percentage of the total (100x/n, where x is the number discontinued and n the total number of patients), it is common for discontinuations to be described as a continuation (100 - 100x/n) to produce a description of survival on treatment. This has previously been done for NSAIDs [9], and we chose to follow this example to describe discontinuation over the first 12 weeks of treatment.

Trial reporting quality and validity were examined using two Oxford scales [22, 23]. Relative risk with 95% confidence interval was calculated using the fixed effects model [24], and was considered to be statistically significant when the 95% confidence interval did not include 1. We calculated the number needed to treat to prevent an event (NNTp) or cause an event (NNH), with a 95% confidence interval, from the sum of all events and patients for treatment and placebo groups [25]. Heterogeneity tests were not used because they have previously been shown to be unhelpful, although homogeneity was examined visually [26, 27]. Publication bias was not assessed using funnel plots because these tests have been shown to be unhelpful [28, 29]. QUOROM (Quality Of Reporting Of Meta-analyses) guidelines were followed where appropriate [30].

This analysis demonstrated again that useful and informative information can be obtained from clinical trial reports, which are more detailed in their reporting of the trials than are published papers, if only because they are not constrained by dictates of size. Here, for instance, information from almost 45,000 patients was contained in just over 73,000 pages – close to two pages per patient. This allows for more detail, and the combination of more detail and large numbers of patients offers the possibility of greater insight. This has been seen previously for valdecoxib [15] and celecoxib [16] in arthritis, for efficacy outcomes in acute pain [31, 32] and for adverse events for contraceptives [33].

In this case it was possible to examine discontinuation rates and causes in some detail over the first 12 weeks of treatment. With placebo there were slight differences for lack of efficacy discontinuation between conditions, with slightly higher rates for RA over OA, but no difference between conditions for adverse event withdrawals (Figure 1). As would be expected, for placebo lack of efficacy withdrawal at 12 weeks was at least three times higher than was adverse event withdrawal.

With etoricoxib and NSAIDs, lack of efficacy withdrawal was considerably less than with placebo (Table 2 and Figure 2). Etoricoxib was no different from placebo in terms of clinical adverse event discontinuations, but NSAIDs were associated with higher rates, with a NNH of 23. Etoricoxib was associated with lower all-cause discontinuation rates than was placebo, mainly because of fewer lack of efficacy discontinuations. Etoricoxib produced significantly more discontinuations for oedema and hypertension (Table 2).

Additional file 4 shows a wider comparison of discontinuation rates, where data from this analysis are examined alongside data from similar analyses of valdecoxib [15] and celecoxib [16]; information from the valdecoxib and celecoxib analyses is recalculated where necessary to make the comparisons similar. In total, these three analyses involve 81,664 patients enrolled in randomized controlled trials of high quality and validity in the settings of OA, RA, back pain, and ankylosing spondylitis. In 41,608 patients involved in trials lasting 4 to 12 weeks, 6,581 received placebo, 502 paracetamol 4,000 mg daily, 12,093 traditional nonselective NSAIDs at maximum licensed daily doses, and 21,892 received a coxib within the licensed dose range. In trials lasting 26 weeks or longer, 40,596 patients received NSAID at maximum licensed daily doses or a coxib within the licensed dose range. We could find no similar analyses for rofecoxib or lumiracoxib, although the celecoxib analysis provided data on rofecoxib, and the etoricoxib analysis provided information on celecoxib (presented separately in Additional file 4).

This wider comparison shows a degree of consistency between the different datasets. For instance, with placebo the rate of lack of efficacy discontinuation is consistently about 70% of all-cause discontinuations. All-cause discontinuations were lower with NSAIDs (22% to 31%) and paracetamol (25%) than with placebo (32% to 47%), and tended to be lower still with coxibs in trials of 4 to 12 weeks (14% to 31%). Adverse event discontinuations are similar across all treatments: 4.4% to 6.3% with placebo, 5.4% with paracetamol, 5.2% to 8.5% with coxibs, and 7.4% to 11% with NSAIDs. As expected, discontinuations because of a gastrointestinal adverse event tended to be higher with traditional nonselective NSAIDs than with any other treatment (Figure 5). Comparisons in longer duration trials are hampered by a dearth of trials of longer duration using licensed doses of celecoxib and valdecoxib.

Examining discontinuation data from clinical trials, even when the numbers of patients are very large, does not necessarily predict what will happen in the real world, where it is the clinical effectiveness of therapy rather than efficacy in clinical trials that is paramount. There are few comparisons on which to draw, although it is generally assumed that discontinuation rates are higher in clinical practice than in clinical trials. For coxibs and NSAIDs, both clinical trials and observational studies [11, 12] concur in their conclusion that adherence is higher with coxibs than with NSAIDs.

Discontinuation results examined here in trials of coxibs and NSAIDs over 12 weeks in musculoskeletal conditions can be compared with trials of other interventions in musculoskeletal and other chronic painful conditions. For example, lack of efficacy and adverse event discontinuations with placebo in short-term trials of opioids in nonmalignant pain were 20% and 7%, respectively [34], similar to our findings for etoricoxib (Table 2). Similar discontinuations rates occur with placebo in 12-week trials in neuropathic pain [35]. Although lack of efficacy discontinuation is typically low (≤ 7%) with active therapies in these conditions, adverse event withdrawal over 12 weeks can be higher (>20% with opioids [34] and about 15% in neuropathic pain), with NNH values for adverse event withdrawal of the order of 10 to 15 [35]. Assessment of adherence in randomized trials is not without its complications and methodological difficulties [36], and this potentially complicates comparisons across datasets.

Discontinuation rates in clinical trials represent important aspects of both efficacy and harm, and deserve more attention. Here, with etoricoxib, discontinuations could be examined in detail because large numbers of patients participated in good quality and well recorded clinical trials, using different doses of etoricoxib and comparators in different clinical conditions. Not all new treatments, and few older ones, enjoy this luxury in terms of the data available.

This does not prevent discontinuations from being a useful measure, perhaps as much or more so in clinical practice as in clinical trials, although discontinuations have been mentioned previously in clinical trials in rheumatology, as with meloxicam in ankylosing spondylitis [37]. In clinical practice, discontinuation rates for adverse events and lack of efficacy has been used to distinguish between disease-modifying antirheumatic drugs [38, 39]. Whether discontinuations meet all of the various criteria of truth, discrimination and feasibility that constitute all the possible subquestions in the OMERACT (Outcome Measures in Rheumatology) filter [40] is moot. We know of no formal evaluation, but at face value discontinuation is easy to record, with the cause, and provides useful information.

This review of a large number of randomized trials and patients provided a more detailed account of discontinuation rates in trials of anti-inflammatory drugs in musculoskeletal conditions than has previously been possible. It also demonstrated consistency between three similar meta-analyses of clinical trial reports covering over 80,000 patients with these conditions. Consistent differences between different drugs were seen, principally fewer adverse event discontinuations with coxibs than with NSAIDs.