MASTERMIND: Stratification of glycaemic response the ADOPT and RECORD studies

1.   Clinical predictors of glycaemic response:i.    Models of glycaemic response to thiazolidinedione therapy as a continuous variable: We will examine clinical predictors of response (HbA1c change) within the first 6 months of therapy as a continuous measure using linear regression analysis, with baseline adjusted change in HbA1c as the outcome and clinical characteristics as the independent variables. In addition to baseline HbA1c results will be adjusted for dose of the medication of interest and for co-treatment (metformin and sulphonylurea) status. This work will be extended further using more complex analysis such as mixed models. We will only include participants who have had received sufficient medication to have taken >80% of expected treatment (this may be altered on the basis or a separate analysis of the relationship between measures of adherence and treatment response below) and have not changed other glucose lowering medications. ii.   Covariates of interest: We will assess the relationships between glycaemic response and the following baseline characteristics, where available. Characteristics may be grouped to create composite variables (e.g. does response differ in individuals exhibiting multiple characteristics associated with insulin resistance?)A. Markers of beta cell failure: Diabetes duration, age of diagnosis, C-peptide (and/or insulin), insulogenic index, GAD autoantibodies, proinsulin insulin ratio B. Markers of insulin resistance: BMI, fasting triglycerides, SHBG (if available), HOMA2IRC. Race, genderiii.   Specificity of characteristic to Thiazolidinedione response: To explore whether a characteristic is specifically associated with response to thiazolidinediones (rather than being associated with response to any treatment) we will assess the relationship between characteristics associated with thiazolidinedione response and response to sulphonylurea and metformin therapy, using the same methods described in i above.iv.   Exploration of confounding: Potential confounders will include concordance, timing of outcome measurement and lifestyle change. We will test whether proportion of medication taken (depending on available concordance data e.g. pill counts) is a predictor of poor response in the analysis above. We will also test whether weight change, as a surrogate for lifestyle change, is associated with response. We expect from published data that near maximum HbA1c reduction will be achieved by 3-4 months treatment with stable glycaemia between 4 and 9 months post therapy, this may be affected by the dose titration regime of the study therefore we will explore whether different lengths of follow up confound our results and adjust minimum length of time on treatment if necessary. The distribution of baseline characteristics will depend on the study of origin (for example RECORD recruited slightly older participants with longer duration diabetes), which could confound results if variation in characteristics potentially predictive of response are not sufficiently represented in those treated with a particular agent. This is unlikely as the smallest study/treatment group (sulphonylureas in RECORD) represents over 1000 participants, however to ensure this is not confounding results we will explore the relationship between characteristics associated with response in the whole group (pooled results) and response within the individual studies.v.   Precision estimate: Assuming at least 3000 participants allocated to rosiglitazone (of 3676 randomised to this therapy) are eligible for inclusion in the analysis we will have 90% power to detect a covariate that explains 0.4% of variation in treatment effect with an alpha of <0.05. For sulphonylureas and metformin, assuming at least 2000 (out of >2500 randomised) participants for each agent meet inclusion criteria, we will have a 90% power to detect a covariate that explains 0.5% of variation in treatment effect with an alpha of <0.05. vi.   Validation of findings: It will be important to validate findings in other datasets. We have access to the GoDARTS, a Tayside-based population dataset of treatment response, and observational primary care response data for 30000 sulphonylurea/thiazolidinedione treated patients from the UK clinical practice research datalink (CPRD). We will use these observational datasets to validate our findings.2.    Side effectsi.   Analysis: We will assess the relationship between any baseline characteristics associated with thiazolidinedione response as a continuous variable and occurrence of the specific side effects listed in A.5 using logistic regression with adjustment for (depending on outcome of interest) age, gender, duration of diabetes, baseline glycaemia or liver function, and study allocation (RECORD/ADOPT). We will explore the use of more complex modelling based on fractional polynomials taking into account occurrence of these events in the comparison (sulphonylurea/metformin) groups (Roystan, Stat Med 2004, PMID 15287081). ii. We will also explore the relationship between the following covariates and side effects as listed below: a.   Hypoglycaemia: Baseline creatinine, liver function, age, diabetes duration, concurrent glucose lowering medicationsb.   Oedema: prior evidence of vascular disease (as defined in study protocols: prior myocardial infarction, prior stroke, coronary, carotid or peripheral artery revascularization, previous documented myocardial ischemia on either an exercise stress test or on any cardiac imaging, or previous unstable angina with ECG changes or cardiac enzyme elevation), microalbuminuria or proteinuria, renal impairment (at EGFR <60 and <30), documented history of hypertension.c.   ALT >3 times upper limit of normal: Baseline creatinine, liver function (ALT, AST, Billiribin, GGT), fasting triglycerides, age, BMI, diabetes duration, gender, concurrent medication. iii.    Confounding: the covariates of interest above may simply be prognostic factors of occurrence of these factors in the population, and unrelated to treatment allocation, rather than predictors of occurrence with Thiazolidinedione treatment. The analysis in ii above is therefore exploratory and methods adjusting for occurrence of these in a comparison group (such as that described above) will therefore be required to validate any findings from logistic regression.iv.   Precision estimate: Assuming a 10% prevalence of the side effects listed above (n=350) we will have 90% power to detect a 0.25 standard deviation difference in a potential predictor between those with and without a side effect, with an alpha of <0.05. While hypoglycaemia and oedema have are likely to have higher prevalence's (10% and 14% respectively in the ADOPT study), ALT >3 times the upper limit of normal was reported at 1% therefore power to detect associations for this outcome will be low.V.    Validation of findings: as outlined in 1 above findings will be validated in the additional observational datasets available to the MASTERMIND consortium.3.   Diabetes progressioni.   Defining glycaemic progression: We will define progression based on annualised rate ofHbA1c change on stable therapy in those participants who have at least 3 consecutive HbA1c measurements spanning >9 months on stable glucose lowering therapy between 6 months and up to 5 years after commencing allocated therapy. Mean rate of change will be calculated based on all eligible periods in an individual, weighted for duration (i.e. while some individuals may have several years without treatment change, others may have more than one eligible period, such as one year on stable dual therapy and 2 years on stable triple therapy, a time weighted mean would be used in the latter case).ii. Analysis: We will assess the relationship between the baseline characteristics associated with initial glycaemic response and progression on thiazolidinedione therapy using linear regression with adjustment for baseline HbA1c c and co-treatment (sulphonylurea, metformin).iii. Confounding: the purpose of this analyses is to examine whether the relationship between baseline factors and glycaemic response is confounded by diabetes progression. Therefore we will not adjust for additional covariates in this analysis, so that the response and progression analysis is comparable. T0 maintain comparability we will exclude HbA1c measurements when measures of treatment concordance are low (as in 1 (glycaemic response) above). If our exploration of confounders in 1 above leads to adjustment for additional confounders we will include those as covariates in this analysis.iv. Precision estimate: It is likely the majority of study participants will be eligible for inclusion given the controlled nature and long duration of these studies (ADOPT median duration of participation 4 years, RECORD 5 years). Assuming a conservative estimate of approximately 50% of those randomised (1800 participants) being eligible for this analysis we would have 90% power to detect a covariate that explains 0.6% of variation in glycaemic progression with an alpha of <0.05.