I'm more interested in the responses to the 'study' which are at the end of the report. Did Mr Mendosa read them do you think ?
Re: Risk of cardiovascular disease and all cause mortality among patients with type 2 diabetes prescribed oral antidiabetes drugs: retrospective cohort study using UK general practice research database. A reply to Tzoulaki et al
Pioglitazone was launched in 2000 . It seems unlikely to have been widely used till 2002-03. Yet this study ran from 1990 till 2005. During this period mortality from vascular disease fell sharply by about a third.
This study is confounded by this effect of time , the pioglitazione treatment group if a different temporally cohort from the other treatment cohorts. Has this been considered in the analysis ? Could a reanalysis of all the data from only 2000 onwards be done ? Also, this might account for the negative outcome of Sulphonylureas which during the 1990s were the most widely used medications but then fell from favour. Observation is not always causation.
Competing interests: None declared
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
In the article by Tzoulaki et al1, a very large cohort of UK diabetic patients are analysed according to previous oral antidiabetic prescriptions and associated risk of developing myocardial infarction or congestive heart failure, or death, during the period from 1990 to 2005.
We acknowledge the very large cohort of more than 90 000 patients analysed. Sulphonylureas are reported to be associated with a considerably increased risk of events in comparison with metformin and glitazones; however, major methodological biases are to be listed:
- 1. Important heterogeneity of the different treatment groups, with the baseline characteristics of the sulphonylurea group showing a much higher risk of cardiovascular disease (average age 5 years older, higher systolic blood pressure, less coprescription with hypolipemic agents, antihypertensives, or aspirin).
- 2. Mixing data for an older-generation sulphonylurea with data for newer ones like glimepiride or gliclazide might be very confusing, as these drugs have already been shown to have very different prognostic value with regard to cardiovascular events. Moreover, some of the analysed data for pioglitazone only concern combination therapy, which might make the comparisons unbalanced.
- 3. The fact that glitazones were introduced later in the study (after the year 2000) might have led to discrimination in the patient profiles, such as a priori exclusion of patients at risk of heart failure in the glitazone group because of specific contraindications.
- 4. Adjusting for potential confounders as proposed in the three models cannot fundamentally correct for all the confounding variables, as some relationships are not linear.
Moreover, the discrepancy should be noted between these results and those of solid randomised, controlled studies (UKPDS2, ADVANCE3, ADOPT4), which introduces some basic questions regarding the solidity of this retrospective analysis.
References
1- Tzoulaki I., BMJ 2009;339-b4731
2- UKPDS 33, Lancet 1998;352(9131):837-53.
3- ADVANCE NEJM 2008 ;358 :2560-72
4- ADOPT NEJM 2006 ;355 :2427-43
Competing interests: I have received fees for lectures and consultations from Astra-Zeneca, BMS, Eli Lilly, GlaxoSmithKline, Novartis, Roche, Servier, and Takeda
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
We have misgivings about the validity of the above study.
The most important problem with the study design would appear to be the definition of each interval of drug treatment; the parameter that the investigators used to define treatment exposure. We are at a loss to understand why a prescription for aspirin or any other non-diabetes drug would be used to define an end date for a period of diabetes treatment exposure! This resulted in an bizarrely short duration of diabetes treatment of only 24 days; far less than even the typical interval between repeat prescriptions issued by GPs. Patients with type 2 diabetes will usually be exposed to a given treatment regimen for an extended period; often years. For example, our most recent studies using similar data resulted in an estimated median follow-up time for metformin monotherapy of more than 2¼ years [2, 3]; a figure itself reduced by right censoring. As such, we would question the use of Cox proportional hazards models to infer that several days' exposure to a given treatment was associated with outcomes that are known to be influenced over months or years. This may explain why no survival curves feature in the report.
Another major concern is that the findings from the various statistical models in this study were inconsistent. For example, in Model 3 (their table 2), none of the drug regimens showed any difference in mortality. Model three—albeit possibly over-parameterised—was the most plausible Cox model. We accept that the use of the biochemical markers in this model resulted in a large amount of data being excluded because of missing data; thus there are fewer events. Substitution of historic clinical diagnosis here, instead of the related biochemistry would have reinstated this statistical power. Thus the overall conclusions of the study are—we would argue—overstated.
There is a further analytical concern in that the results of the study inferred an independent association between drug treatment classes and adverse outcomes, yet no attempt was made to account for the inter-dependence of the treatment intervals for each subject. For example, a notable proportion of second generation sulfonylurea intervals (so called) were preceded by TZD or metformin intervals, whereas TZD intervals where much less likely to have been prescribed alternative agents. Clearly treatment failure with TZDs or metformin would have been associated with adverse clinical profiles, themselves associated with poorer outcomes. As the authors' colleagues Pocock & Smeeth recently commented in The Lancet, even after adjustment for confounders, residual selection bias might distort any true (lack of) differences between treatments [4].
The truncated interval design devised by Tzoulaki and colleagues may have introduced a number of other potential flaws that could also invalidate the findings from this study. Patients on a given combination therapy are likely to have received repeat prescriptions for each individual component of their combination regimen. For instance, taking the most common combination oral diabetes therapy of metformin plus a sulfonylurea; we have observed that patients do not necessarily have both drugs prescribed simultaneously; they are commonly staggered over a period of weeks (the length of which would be largely determined by the quantity issued and dose last prescribed). The design used in by Tzoulaki and colleagues would have spuriously defined these patients as being exposed variously to either metformin or sulfonylurea monotherapy.
The original focus of their study was the safety and efficacy of pioglitazone. The most likely alternative to this regimen was metformin plus sulfonylurea combination therapy. This regimen was never specifically defined for some reason, although as illustrated, the use of the exposure parameter would have made this unreliable anyway.
Finally, there was no dose-response data in this study. Interestingly, the use of the spurious “interval of drug treatment” would have lent itself to the easy definition of treatment dose. There are other issues that raise concern about the reliability of this study. They did not define baseline morbidity in any meaningful way. Although the investigators have gone to considerable length to define some potentially confounding effects, they have seemingly ignored or missed other ones that would be commonly regarded as being important with regard to all-cause mortality; for example cancer. Cancer is now known to be related to diabetes therapy [2]. Thus, there are various important issues that have been ignored in terms of elimination of important confounders that would have been related to the hazard of death. We are also surprised that smoking did not feature in models 1 and 2.
In this study, the investigators may have failed to appreciate the basic patterns of clinical glycaemic management and the complexity of the way that this is recorded in routine prescribing data, and as such their work may represent a missed opportunity.
Chris D. Poole
Epidemiologist
Pharmatelligence, Cardiff
Craig J. Currie
Reader in Diabetes Pharmacoepidmiology
School of Medicine, Cardiff University
Contact:
[email protected]
References
1. Tzoulaki I, Molokhia M, Curcin V, Little MP, Millett CJ, Ng A, Hughes RI, Khunti K, Wilkins MR, Majeed A, Elliott P. Risk of cardiovascular disease and all cause mortality among patients with type 2 diabetes prescribed oral antidiabetes drugs: retrospective cohort study using UK general practice research database. BMJ 2009 doi: 10.1136/bmj.b4731
2. Currie CJ, Poole CD, Gale EA. The influence of glucose-lowering therapies on cancer risk in type 2 diabetes. Diabetologia 2009;52:1766-77
3. Currie CJ, Peters JR, Tynan A, Evans M, Heine R, Bracco AL, Zagar T and Poole CD. Survival in people with type 2 diabetes as a function of HbA1c: a retrospective cohort study. The Lancet (in press)
4. Pocock SJ, Smeeth L. Insulin glargine and malignancy: an unwarranted alarm. The Lancet 2009;374:511-513
Competing interests: None declared
)