2 cancers dating
Studies restricted to populations with additional comorbidities or diseases were excluded.
During an initial title and abstract screen, reviews, meta analyses and editorial pieces that looked at metformin and cancer were retained so that reference lists could be checked.
A systematic data extraction and bias assessment was conducted, in which risk of eight bias domains (outcome, exposure, control selection, baseline confounding, time-dependent confounding, immortal time, missing data, censoring methods) were assessed against pre-defined criteria, and rated as unlikely, low, medium or high.
Results: Of 46 studies identified, 21 assessed the effect of metformin on all cancer.
When the time-dependent confounder is also affected by previous treatment, as depicted in the causal diagram below, standard statistical methods cannot provide unbiased estimates of the total causal effect of time-varying treatment. The search involved using Me SH headings as well as keyword searches in the title and abstract. Conference abstracts and unpublished studies were excluded.
Articles were included in the review if they were of a standard epidemiological design and presented original observational research. Studies were required to present a measure of effect of metformin on risk of cancer incidence (either all cancer or site-specific) in patients with T2DM, with age adjustment as a minimum.
However, 28/46 studies were at risk from bias through exposure definition, 22 through insufficient baseline adjustment and 35 from possible time-dependent confounding.
A 10% random sample of the extracted studies were screened by an additional researcher (H. An example extraction table is supplied in Although none of the investigators were blinded to the aims of the review, detailed criteria for assessment of bias were produced in order to consider risks of bias for each study.
F.) to test the reliability of the inclusion criteria. The data extraction table was piloted on five studies (by R. The eight domains assessed for bias were: (i) outcome definition; (ii) exposure definition (including choice of comparator): (iii) control selection (case control studies only); (iv) consideration of Hb A1c, BMI and other antidiabetic drugs as time-dependent confounders affected by previous treatment (Box 1); (v) adjustment for baseline (study entry) confounders (smoking, diabetes severity, age, gender); (vi) immortal time (cohort studies only); (vii) missing data; and (viii) censoring methods (cohort studies only).
Low risk meant that there was small possibility of bias but the potential magnitude of the bias was unlikely to materially affect the overall study conclusions. However, the timing and accuracy of the confounder were still considered as sources of bias, since these aspects could have resulted in its incorrect omission.
Medium and high risk of bias meant that there was potential for some or substantial effect of bias on the study estimate, respectively. Bias from outcome and exposure definition encompassed both misclassification bias, biases induced by timing of measurement, and whether the definitions may introduce selection bias.