Weight

The included trial did not report on weight.

HbA1c

The included trial reported final HbA1c measurements of 7.8% (SD 0.4) and 7.9% (SD 0.5) in the M+S group compared with the metformin monotherapy, respectively. They did not report the number of participants included in the analysis.

All‐cause mortality

Both included trials reported data on all‐cause mortality (low‐certainty evidence because of serious imprecision; Analysis 1.1). One trial reported 6 deaths out of 307 participants (2%) and 1 death out of 101 participants (1%) in the M+S combination group compared with the metformin plus placebo group, respectively (Ahrén 2014). However, the number of deaths was unclear due to varied reporting. We contacted the trial authors for clarification but did not receive a reply. To be sure to account for all deaths, we extracted data from the publication that reported the highest number. One trial reported data after an 18‐month, open‐label extension: there were no deaths in 242 participants and no deaths in 121 participants in the M+S combination group compared with the metformin plus placebo group, respectively (Nauck 2013).

Health‐related quality of life

None of the included trials reported on this outcome.

Serious adverse effects

Both included trials reported that a total of 84 participants experienced a serious adverse event; in the M+S group 60/549 (10.9%) participants had a serious adverse event compared with 24/222 (10.8%) participants in the metformin plus placebo group (RR 0.97, 95% CI 0.59 to 1.61; P = 0.91; 2 trials; 771 participants; very low‐certainty evidence because of attrition bias and serious imprecision; Analysis 1.2). One trial reported data after an 18‐month, open‐label extension phase (Nauck 2013).

TSA showed that 2.45% of the diversity‐adjusted required information size to detect or reject a 10% relative risk reduction (RRR) had been accrued. Diversity was 21%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Cardiovascular mortality

Both included trials reported data on cardiovascular mortality (low‐certainty evidence because of serious imprecision; Analysis 1.3). One trial reported 1 cardiovascular death in 307 (0.3%) participants in the M+S group compared with 1 cardiovascular death in 101 (1%) participants in the metformin plus placebo group (Ahrén 2014). One trial reported that no participants died in either intervention group after an 18‐month, open‐label extension phase (Nauck 2013).

Non‐fatal myocardial infarction

Both included trials reported that a total of three participants experienced a non‐fatal myocardial infarction; in the M+S group 2/549 (0.4%) participants had a myocardial infarction compared with 1/222 (0.5%) participants in the metformin plus placebo group (RR 0.63, 95% CI 0.08 to 5.10; P = 0.67; 2 trials; 771 participants; very low‐certainty evidence because of attrition bias and serious imprecision; Analysis 1.4). One trial reported data after an 18‐month, open‐label extension phase (Nauck 2013).

TSA showed that 0.13% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was zero. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Heart failure

Both included trials reported data on heart failure (Analysis 1.5). One trial reported one heart failure in 307 participants (0.3%) and no heart failure in 101 participants in the M+S group compared with the metformin plus placebo group, respectively (Ahrén 2014). One trial reported that no participants experienced a heart failure in either intervention group after an 18‐month, open‐label extension phase (Nauck 2013).

Non‐fatal stroke, amputation of lower extremity, blindness or severe vision loss

None of the included trials reported on this outcome.

End‐stage renal disease

One trial reported that no participants had end‐stage renal disease in either intervention group after an 18‐month, open‐label extension phase (very low‐certainty evidence because of attrition bias and serious imprecision; Nauck 2013).

Non‐serious adverse events

Both included trials reported that a total of 505 participants experienced a non‐serious adverse event: in the M+S group 386/549 (70.3%) participants had a non‐serious adverse event compared with 119/222 (53.6%) participants in the metformin plus placebo group (random RR 1.25, 95% CI 0.96 to 1.64; P = 0.10; fixed RR 1.24, 95% CI 1.10 to 1.41; P < 0.001; 2 trials; 771 participants; Analysis 1.6). None of the trials provided a detailed definition of the outcome. One trial reported data after an 18‐month, open‐label extension phase (Nauck 2013).

TSA showed that 5.8% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 78%. The TSA‐adjusted 95% CI was 0.47 to 3.32.

Hypoglycaemia

Both included trials reported that a total of 181 participants experienced a mild or moderate hypoglycaemic episode; in the M+S group there were 160/549 (29.1%) participants with hypoglycaemic episodes compared with 21/222 (9.5%) participants in the metformin plus placebo group (random RR 3.93, 95% CI 0.71 to 21.88; P = 0.12; fixed RR 2.87, 95% CI 1.90 to 4.34; P < 0.001; 2 trials; 771 participants; Analysis 1.7 in favour of metformin plus placebo). Ahrén 2014 defined mild or moderate hypoglycaemia by blood glucose levels 3.9 mmol/L or lower and Nauck 2013 used blood glucose values 3.1 mmol/L or lower. One trial reported data after an 18‐month, open‐label extension phase (Nauck 2013).

TSA showed that 0.18% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 93%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Both included trials reported data on serious hypoglycaemia (Analysis 1.8). One trial reported one participant with serious hypoglycaemia out of 307 participants (0.3%) and no participants with serious hypoglycaemia out of 101 participants in the M+S group compared with the metformin plus placebo group, respectively (Ahrén 2014). One trial reported that no participants had serious hypoglycaemia in either intervention group after an 18‐month, open‐label extension phase (Nauck 2013).

Socioeconomic effects

None of the included trials reported on this outcome.

Weight

Both included trials reported weight change in favour of metformin plus placebo (mean difference (MD) 3.4 kg, 95% CI 1.4 to 5.4; P = 0.001; 2 trials; 476 participants; Analysis 1.9). Nauck 2013 reported data after an 18‐month, open‐label extension phase.

HbA1c

Both included trials reported change in HbA1c in favour of M+S (random MD −0.5%, 95% CI −1.1 to 0.1; P = 0.13; fixed MD −0.6%, 95% CI −0.8 to −0.4; P < 0.001; 2 trials; 472 participants; Analysis 1.10). One trial reported data after an 18‐month, open‐label extension phase (Nauck 2013).

All‐cause mortality

Three trials reported that a total of 22 participants died: in the M+S group 11/1057 participants (1.0%) died compared with 11/1537 participants (0.7%) in the metformin plus GLP‐1 analogue group (RR 1.15, 95% CI 0.49 to 2.67; P = 0.75; 3 trials; 2594 participants; low‐certainty evidence; Analysis 2.1). Calculation of the 95% prediction interval did not provide a meaningful estimate. One trial reported data after an 18‐month, open‐label extension phase (Nauck 2013).

Derosa 2011a reported that no participants died but did not provide the number of participants included in the analysis.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to trials with low risk of selection bias did not change the direction of the effect estimate (RR 0.93, 95% CI 0.30 to 2.93; P = 0.91; 2 trials; 1985 participants). We could not perform sensitivity analysis excluding large trials because none of the included trials for this comparison had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company since all trials received funding from a pharmaceutical company.

TSA showed that 0.61% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was zero. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Health‐related quality of life

We did not identify trials with data on health‐related quality of life for this comparison. One trial reported on 'impact of weight on quality of life' (IWQOL; Nauck 2013). They reported the result separately for each intervention group.

Serious adverse effects

Three trials reported that a total of 322 participants experienced a serious adverse event: in the M+S group 128/1057 participants (12.1%) had a serious adverse event compared with 194/1537 participants (12.6%) in the metformin plus GLP‐1 analogue group (RR 0.90, 95% CI 0.73 to 1.11; P = 0.32; 3 trials; 2594 participants; very low‐certainty evidence; Analysis 2.2). The 95% prediction interval ranged between 0.23 and 3.51. One of the trials reported data after an 18‐month, open‐label extension (Nauck 2013). Derosa 2011a reported that no participants experienced serious adverse events but did not provide the number of participants included in the analysis.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis according to trials with low risk of selection bias and duration of intervention (excluding trials with duration of intervention longer than 104 weeks) contained the same two trials (Gallwitz 2012a; Nauck 2013). The effect estimate did not change substantially (RR 0.94, 95% CI 0.73 to 1.20; P = 0.60; 2 trials; 1985 participants). We could not perform sensitivity analysis excluding large trials because no trials had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 12.4% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was zero. The TSA‐adjusted 95% CI was 0.38 to 2.15.

Cardiovascular mortality

Three trials reported on cardiovascular disease (Ahrén 2014; Derosa 2011a; Nauck 2013). Derosa 2011a reported that no participants died due to cardiovascular disease but did not provide the number of participants included in the analysis. Nauck 2013 reported that no participants died due to cardiovascular disease in any of the intervention groups referring to the 18‐month, open‐label extension phase.

Ahrén 2014 reported that in the M+S group 1/307 (0.3%) participants died of cardiovascular disease compared with 1/302 (0.3%) participants in the metformin plus GLP‐1 analogue group (Analysis 2.3; low‐certainty evidence).

Non‐fatal myocardial infarction

Two trials reported that a total of eight participants experienced a non‐fatal myocardial infarction: in the M+S group 2/549 (0.4%) participants had a non‐fatal myocardial infarction compared with 6/1026 (0.6%) participants in the metformin plus GLP‐1 analogue group (RR 0.57, 95% CI 0.12 to 2.82; P = 0.49; 2 trials; 1575 participants; very low‐certainty evidence; Analysis 2.4).

Nauck 2013 reported data after an 18‐month, open‐label extension phase. Derosa 2011a reported that no participants experienced a non‐fatal myocardial infarction but did not provide the number of participants included in the analysis.

We could not perform sensitivity analyses due to lack of data.

TSA showed that 0.4% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was zero. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Heart failure

Four trials reported that a total of five participants developed heart failure: in the M+S group 1/1057 (0.1%) participants developed heart failure compared with 4/1537 (0.3%) participants in the metformin plus GLP‐1 analogue group (RR 0.54, 95% CI 0.10 to 2.77; P = 0.46; 3 trials; 2594 participants; Analysis 2.5). Calculation of the 95% prediction interval did not provide a meaningful estimate.

Nauck 2013 reported data after an 18‐month, open‐label extension phase. Derosa 2011a reported that no participants developed heart failure but did not provide the number of participants included in the analysis.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to trials with low risk of selection bias and duration of intervention (excluding trials with duration of intervention more than 104 weeks) contained the same two trials (Gallwitz 2012a; Nauck 2013). The effect estimate did not change substantially (RR 0.58, 95% CI 0.06 to 5.54; P = 0.63; 2 trials; 1985 participants). We could not perform sensitivity analysis excluding large trials because no trials had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 0.26% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was zero. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Non‐fatal stroke

Derosa 2011a reported that no participants experienced a stroke during the intervention period but did not provide the number of participants included in the analysis.

Amputation of lower extremity

Derosa 2011a reported that no participants had amputation of lower extremity during the intervention period but did not provide the number of participants included in the analysis.

Blindness or severe vision loss

Derosa 2011a reported that no participants had blindness or severe vision loss during the intervention period but did not provide the number of participants included in the analysis.

End‐stage renal disease

One trial provided data on end‐stage renal disease, however only after the 18‐month, open‐label extension phase (Nauck 2013). In the M+S group no participants out of 242 participants had end‐stage renal disease compared with 1 participant out of 724 (0.1%) participants in the metformin plus GLP‐1 analogue group (Analysis 2.6).

Derosa 2011a reported that no participants had end‐stage renal disease during the intervention period but did not provide the number of participants included in the analysis.

Non‐serious adverse events

Three trials reported that a total of 1621 participants experienced a non‐serious adverse event: in the M+S group 634/1057 (60.0%) participants had a non‐serious adverse event compared with 987/1537 (64.2%) participants in the metformin plus GLP‐1 analogue group (Analysis 2.7). Due to substantial heterogeneity, aggregating the trials in a meta‐analysis was not appropriate. With regard to the M+S groups, the trial with most of the non‐serious adverse events had two placebo interventions (GLP‐1 analogue as subcutaneous injection and placebo DPP‐4 inhibitor as tablet) in addition to M+S (Ahrén 2014).

One trial was open‐label and did not apply placebo (Gallwitz 2012a), and one trial (Nauck 2013), reported data after an 18‐month, open‐label extension. With regard to the metformin plus GLP‐1 analogue treatment arms, each trial used a different GLP‐1 analogue: one administered albiglutide (Ahrén 2014), and one administered exenatide (Gallwitz 2012a), or liraglutide (Nauck 2013).

Derosa 2011a reported that no participants experienced non‐serious adverse events but did not provide the number of participants included in the analysis.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analyses restricted to trials with low risk of selection bias and duration of intervention (excluding trials with duration of intervention longer than 104 weeks) contained the same two trials (Gallwitz 2012a; Nauck 2013). The effect estimate was in favour of M+S combination therapy (RR 0.84, 95% CI 0.76 to 0.92; P = 0.0001; 2 trials; 1985 participants). We could not perform sensitivity analysis excluding large trials because no trials had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

Hypoglycaemia

Three trials reported that a total of 569 participants experienced a mild or moderate hypoglycaemic episode: in the M+S group 400/1057 (37.8%) participants had a hypoglycaemic episode compared with 169/1537 (11.0%) participants in the metformin plus GLP‐1 analogue group (RR 3.24, 95% CI 2.05 to 5.13; P < 0.001; 3 trials; 2594 participants; Analysis 2.8; in favour of metformin plus GLP‐1 analogue). Calculation of the 95% prediction interval did not provide a meaningful estimate. The trial with mostly mild or moderate hypoglycaemic events defined hypoglycaemia as any sign or symptom of hypoglycaemia or blood glucose 3.9 mmol/L or less (Gallwitz 2012a), whereas the other trials defined hypoglycaemia as symptoms of hypoglycaemia plus blood glucose values 3.9 mmol/L or less, or blood glucose values 3.9 mmol/L or less without symptoms (Ahrén 2014), or symptoms of hypoglycaemia or blood glucose values 3.1 mmol/L or less (Nauck 2013). Nauck 2013 reported data after an 18‐month, open‐label extension phase.

Derosa 2011a reported two events of mild or moderate hypoglycaemia in the glimepiride group and one in the exenatide group but did not provide the number of participants included in the analysis.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analyses restricted to trials with low risk of selection bias and duration of intervention (excluding trials with duration of intervention longer than 104 weeks) contained the same two trials (Gallwitz 2012a; Nauck 2013). We could not perform sensitivity analysis excluding large trials because no trials had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

Only three participants reported serious hypoglycaemia: in the M+S group 1/1057 (0.1%) participants had a serious hypoglycaemic event compared with 2/1537 (0.1%) participants in the metformin plus GLP‐1 analogue group (RR 1.00, 95% CI 0.16 to 6.30; P = 1.00; 3 trials; 2594 participants; Analysis 2.9).

Nauck 2013 reported data after an 18‐month, open‐label extension phase. Derosa 2011a reported that no event of serious hypoglycaemia was observed but did not provide the number of participants included in the analysis.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analyses restricted to trials with low risk of selection bias and duration of intervention (excluding trials with duration of intervention longer than 104 weeks) contained the same two trials (Gallwitz 2012a; Nauck 2013). The effect estimate did not change substantially (RR 0.58, 95% CI 0.06 to 5.54; P = 0.63; 2 trials; 1985 participants). We could not perform sensitivity analysis excluding large trials because no trial had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA for severe hypoglycaemia showed that 0.09% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was zero. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Socioeconomic effects

We did not identify trials with data on socioeconomic effects for this comparison.

Weight

Five trials reported weight change in favour of metformin plus GLP‐1 analogue (MD 5.5 kg, 95% CI 3.6 to 7.5; P < 0.001; 5 trials; 1777 participants; Analysis 2.10). Calculation of the 95% prediction interval did not provide a meaningful estimate. Heterogeneity in the findings could have been caused by the various durations of follow‐up, ranging from 52 weeks to 156 weeks, and various doses of glimepiride ranging from 1 mg/day to 6 mg/day. Hollander 2017 did not report the number of participants included in the analysis.

Nauck 2013 reported data after an 18‐month, open‐label extension phase.

HbA1c

Five trials reported change in HbA1c (MD 0.01%, 95% CI ‐0.2 to 0.2; P = 0.91; 5 trials; 2346 participants; Analysis 2.11). The 95% prediction interval ranged between 0% and 1.0%.

Nauck 2013 reported data after an 18‐month, open‐label extension phase. Trials with 52 weeks of treatment reported at greater decrease in Hba1c with sulphonylurea treatment, whereas trials with 104 weeks of treatment or more reported a greater decrease in HbA1c with GLP‐1 analogue treatment.

TSA showed that 0.61% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was 91%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

All‐cause mortality

Nine trials reported that a total of 59 participants died: in the M+S group 33 participants died out of 5387 (0.6%) participants, compared with 26 deaths in 6307 (0.4%) participants in the metformin plus DPP‐4 inhibitor group (RR 1.32, 95% CI 0.76 to 2.28; P = 0.32; 9 trials; 11,694 participants; low‐certainty evidence; Analysis 3.1). The 95% prediction interval ranged between 0.68 and 2.55.

A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.77).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to trials with low risk of selection bias did not substantially change the effect estimate (RR 1.29, 95% CI 0.41 to 4.10; P = 0.66; 5 trials; 4363 participants). Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 1.52, 95% CI 0.81 to 2.88; P = 0.19; 8 trials; 8595 participants). We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials had received funding from a pharmaceutical company.

TSA showed that 2.4% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was zero. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Health‐related quality of life

We did not identify trials reporting data on health‐related quality of life for this comparison.

Serious adverse effects

Nine trials reported that a total of 1514 participants experienced a serious adverse event: in the M+S group 735/5387 (13.6%) participants had a serious adverse event compared with 779/6307 (12.4%) participants in the metformin plus DPP‐4 inhibitor group (RR 1.07, 95% CI 0.97 to 1.18; P = 0.17; 9 trials; 11,694 participants; very low‐certainty evidence; Analysis 3.2). The 95% prediction interval ranged between 0.95 and 1.20.

A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.82).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to only trials with low risk of selection bias did not substantially change the effect estimate (RR 1.09, 95% CI 0.93 to 1.28; P = 0.27; 5 trials; 4363 participants). Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 1.06, 95% CI 0.94 to 1.20; P = 0.33; 8 trials; 8595 participants). We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 61% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was zero. The TSA‐adjusted 95% CI was 0.93 to 1.23.

Cardiovascular mortality

Six trials reported that a total of 20 participants died due to cardiovascular disease: in the M+S group 11 participants died out of 2989 (0.4%) participants compared with 9 out of 3885 (0.2%) participants in the metformin plus DPP‐4 inhibitor group (RR 1.54, 95% CI 0.63 to 3.79; P = 0.34; 6 trials; 6874 participants; low‐certainty evidence; Analysis 3.3). The 95% prediction interval ranged between 0.43 and 5.52. Two trials reported a composite outcome of cardiovascular and cerebrovascular events (Filozof 2010: M+S 12/493 (2.4%) participants, metformin plus DPP‐4 inhibitor 7/510 (1.4%) participants; Matthews 2010: M+S 60/1546 (3.9%) participants, metformin plus DPP‐4 inhibitor 59/1553 (3.8%) participants).

We could not perform a test for subgroup differences according to duration of follow‐up due to lack of data.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to only trials with low risk of selection bias did not substantially change the effect estimate (RR 1.72, 95% CI 0.44 to 6.69; P = 0.43; 4 trials; 3645 participants).

We could not perform sensitivity analysis excluding large trials because no trials had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 0.92% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was zero. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Non‐fatal myocardial infarction

Six trials reported that a total of 28 participants experienced a non‐fatal myocardial infarction: in the M+S group 15/2989 (0.5%) participants had a myocardial infarction compared with 13/3885 (0.3%) participants in the metformin plus DPP‐4 inhibitor group (RR 1.45, 95% CI 0.69 to 3.07; P = 0.33; 6 trials; 6874 participants; very low‐certainty evidence; Analysis 3.4). The 95% prediction interval ranged between 0.50 and 4.20.

We could not perform a test for subgroup differences according to duration of follow‐up due to lack of data.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to trials with low risk of selection bias did not substantially change the effect estimate (RR 1.44, 95% CI 0.55 to 3.77; P = 0.45; 4 trials; 3645 participants). We could not perform sensitivity analysis excluding large trials because no trials had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 0.95% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was zero. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Heart failure

Eight trials reported that a total of 29 participants developed heart failure: in the M+S group 15/4894 (0.3%) participants developed heart failure compared with 14/5797 (0.2%) participants in the metformin plus DPP‐4 inhibitor group (RR 1.05, 95% CI 0.47 to 2.34; P = 0.90; 8 trials; 10,691 participants; Analysis 3.5). The 95% prediction interval ranged between 0.39 and 2.86.

A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.08).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to only trials with low risk of selection bias did not substantially change the direction of the effect estimate (RR 1.21, 95% CI 0.38 to 3.78; P = 0.75; 5 trials; 4363 participants). Sensitivity analysis excluding large trials did not substantially change the direction of the effect estimate (RR 1.07, 95% CI 0.42 to 2.69; P = 0.89; 7 trials; 7592 participants). We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

Non‐fatal stroke

Four trials reported that a total of 22 participants experienced a non‐fatal stroke: in the M+S group 14/2098 (0.7%) participants had a non‐fatal stroke compared with 8/2995 (0.3%) participants in the metformin plus DPP‐4 inhibitor group (RR 2.21, 95% CI 0.74 to 6.58; P = 0.15; 4 trials; 5093 participants; very low‐certainty evidence; Analysis 3.6). The 95% prediction interval ranged between 0.12 and 40.89.

We could not perform subgroup analysis due to lack of data.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis restricted to trials with low risk of selection bias due to lack of data. We could not perform sensitivity analysis excluding long trials because all trials had a follow‐up of 104 weeks or less. We could not perform sensitivity analysis excluding large trials because no trials had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 0.76% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was 23%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Amputation of lower extremity

Dei Cas 2017 did not observe amputation of the lower extremity in either intervention group (64 participants; very low‐certainty evidence).

Blindness or severe vision loss

Dei Cas 2017 did not observe blindness or severe vision loss in either intervention group (64 participants; very low‐certainty evidence).

End‐stage renal disease

Dei Cas 2017 did not observe end‐stage renal disease in either intervention group (64 participants; very low‐certainty evidence).

Non‐serious adverse events

Seven trials reported that a total of 4751 participants experienced a non‐serious adverse event: in the M+S group 2156/3348 (64.4%) participants had a non‐serious adverse event compared with 2595/4244 (61.1%) participants in the metformin plus DPP‐4 inhibitor group (RR 1.18, 95% CI 1.03 to 1.35; P = 0.02; 7 trials; 7592 participants; Analysis 3.7 in favour of metformin plus DPP‐4 inhibitor). The 95% prediction interval ranged between 0.73 and 1.91. There was substantial heterogeneity, which could be caused by various definitions of the outcome (most trials did not adequately specify this outcome measure).

A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.11).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to trials with low risk of selection bias did not substantially change the effect estimate (RR 1.27, 95% CI 1.06 to 1.52; P = 0.01; 5 trials; 4363 participants). We could not perform sensitivity analysis excluding large trials because no trials had more than 1000 participants randomised to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 21.5% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 94%. The TSA‐adjusted 95% CI was 0.88 to 1.58.

Hypoglycaemia

Seven trials reported that a total of 1544 participants experienced a mild or moderate hypoglycaemic event: in the M+S group 1359/4535 (30.0%) participants had a mild or moderate hypoglycaemic episode compared with 185/5438 (3.4%) participants in the metformin plus DPP‐4 inhibitor group (RR 7.42, 95% CI 4.77 to 11.53; P < 0.001; 7 trials; 9973 participants; Analysis 3.8 in favour of metformin plus DPP‐4 inhibitor). The 95% prediction interval ranged between 1.86 and 29.66. There was substantial heterogeneity, probably caused by various definitions of the outcome. All trials that stated a definition of outcome defined mild/moderate hypoglycaemia in different ways (Ahrén 2014; Del Prato 2014; Göke 2013; Matthews 2010).

The test for subgroup differences analysing trials according to duration of intervention showed a statistically significant difference between subgroups (P = 0.04; Analysis 3.8). However, CIs overlap indicating that in fact there was no true interaction.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to trials with low risk of selection bias did not substantially change the effect estimate (RR 6.71, 95% CI 3.94 to 11.44; P < 0.001; 4 trials; 3645 participants). Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 6.00, 95% CI 4.33 to 8.32; P < 0.001; 6 trials; 6874 participants). We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 1.4% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was 88%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Eight trials reported that a total of 55 participants experienced a serious hypoglycaemic event: in the M+S group 51/4894 (1.0%) participants had a serious hypoglycaemic episode compared with 4/5797 (0.1%) participants in the metformin plus DPP‐4 inhibitor group (RR 8.04, 95% CI 3.31 to 19.53; P < 0.01; 8 trials; 10,691 participants; Analysis 3.9 in favour of metformin plus DPP‐4 inhibitor). The 95% prediction interval ranged between 2.66 and 24.35.

A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.78).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis restricted to trials with low risk of selection bias did not substantially change the effect estimate (RR 8.62, 95% CI 2.81 to 26.43; P = 0.0002; 5 trials; 4363 participants). Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 6.93, 95% CI 2.72 to 17.65; P < 0.001; 7 trials; 7592 participants). We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 0.62% of the diversity‐adjusted information size had been accrued to detect or reject a 10% RRR. Diversity was 0%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Socioeconomic effects

Two trials performed economic analyses of trial data (Del Prato 2014; Göke 2013).

One trial presented an economic analysis using the IMS Core Diabetes Model (Del Prato 2014). Treatment with alogliptin 12.5 mg or 25 mg compared with sulphonylurea was associated with an incremental cost‐effectiveness ratio of GBP 10,959 or GBP 7217 per quality adjusted life year, respectively.

One trial presented an economic analysis using the Cardiff Stochastic Simulation Cost‐utility Model (Göke 2013). The overall mean cost per quality adjusted life year gained with saxagliptin plus metformin compared with M+S was GBP 7888.

Weight

Nine trials reported weight change. Metformin plus DPP‐4 inhibitor compared with M+S combination therapy resulted in a weight loss of 2.2 kg (95% CI 1.7 to 2.6; P < 0.001; 9 trials; 10,228 participants; Analysis 3.10). The 95% prediction interval ranged between 0.8 kg and 3.5 kg.

HbA1c

Nine trials reported HbA1c (random MD −0.1%, 95% CI −0.1 to 0.03; P = 0.25; 9 trials; 9320 participants; fixed MD −0.1%, 95% CI −0.1 to −0.04; P < 0.001; Analysis 3.11 in favour of M+S). The 95% prediction interval ranged between −0.3% and 0.2%.

All‐cause mortality

In the M+S group no participants died out of 375 participants, compared with 2 deaths in 375 (0.5%) participants in the metformin plus long‐acting DPP‐4 inhibitor group (RR 0.20, 95% CI 0.01 to 4.15; P = 0.30; very low‐certainty evidence because of indirectness and very serious imprecision; Analysis 4.1).

Health‐related quality of life

Handelsman 2017 did not report on this outcome.

Serious adverse effects

In the M+S group 18/375 (4.8%) participants experienced a serious adverse event compared with 24/375 (6.4%) participants in the metformin plus long‐acting DPP‐4 inhibitor group (RR 0.75, 95% CI 0.41 to 1.36; P = 0.34; very low‐certainty evidence because of very serious imprecision, Analysis 4.2).

Cardiovascular mortality

In the M+S group no participants out of 375 died compared with 1 death in 375 (0.3%) participants in the metformin plus long‐acting DPP‐4 inhibitor (RR 0.33, 95% CI 0.01 to 8.16; P = 0.50; very low‐certainty evidence because of indirectness and very serious imprecision; Analysis 4.3).

Non‐fatal myocardial infarction

In the M+S group 1/375 (0.3%) participants experienced a non‐fatal myocardial infarction compared with 0/375 participants in the metformin plus long‐acting DPP‐4 inhibitor (RR 3.00, 95% CI 0.12 to 73.41; P = 0.50; very low‐certainty evidence because of indirectness and very serious imprecision, Analysis 4.4).

Heart failure, non‐fatal stroke, amputation of lower extremity, blindness or severe vision loss, end‐stage renal disease

Handelsman 2017 did not report on these outcomes.

Non‐serious adverse events

In the M+S group 125/375 (33.3%) participants experienced a non‐serious adverse event compared with 43/375 (11.5%) participants in the metformin plus long‐acting DPP‐4 inhibitor (RR 2.91, 95% CI 2.12 to 3.99; P < 0.001; Analysis 4.5), favouring metformin plus long‐acting DPP‐4 inhibitor.

Hypoglycaemia

In the M+S group 110/375 (29.3%) participants experienced a mild or moderate hypoglycaemic episode compared with 21/375 (5.6%) participants in the metformin plus long‐acting DPP‐4 inhibitor (RR 5.24, 95% CI 3.36 to 8.17; P < 0.001; Analysis 4.6 in favour of metformin plus long‐acting DPP‐4 inhibitor).

In the M+S group 6/375 (1.6%) participants experienced a serious hypoglycaemic episode compared with 1/375 (0.3%) participant in the metformin plus long‐acting DPP‐4 inhibitor (RR 6.00, 95% CI 0.73 to 49.60; P = 0.10; Analysis 4.7).

Socioeconomic effects

Handelsman 2017 did not report on this outcome.

Weight

Change from baseline in body weight increased by 1.5 kg (SD 4.0) in 375 participants in the M+S group compared with a weight reduction of 0.4 kg in 375 participants in the metformin plus long‐acting DPP‐4 inhibitor group (MD 1.9 kg, 95% CI 1.3 to 2.5; P < 0.001; Analysis 4.8).

HbA1c

In the M+S group there was a MD in HbA1c change of −0.5% in 375 participants compared with −0.3% in 375 participants in the metformin plus long‐acting DPP‐4 inhibitor group (MD of −0.2%, 95% CI −0.3 to −0.1; P = 0.006; Analysis 4.9).

All‐cause mortality

Six trials reported that a total of 237 participants died: in the M+S group 123/3300 (3.7%) participants died compared with 114/3354 (3.4%) participants in the metformin plus thiazolidinedione group (RR 1.09, 95% CI 0.85 to 1.40; P = 0.51; 6 trials; 6654 participants; low‐certainty evidence; Analysis 5.1). The 95% prediction interval ranged between 0.75 and 1.55. One of the trials provided data from a time‐to‐event analysis on all‐cause mortality: the HR for metformin plus pioglitazone versus M+S was 1.10 (95% CI 0.75 to 1.61, P = 0.63; Vaccaro 2017).

In one trial, comparing M+S with metformin plus pioglitazone, the investigators reported that no participants died but did not provide the number of participants included in the analysis (Derosa 2011b).

A test for subgroup differences comparing rosiglitazone with pioglitazone did not indicate interaction (P = 0.84; Analysis 5.1). A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.91; Analysis 6.1).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis according to trials with low risk of selection bias did not substantially change the effect estimate (RR 1.08, 95% CI 0.84 to 1.38; P = 0.57; 5 trials; 6570 participants). Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 2.08, 95% CI 0.49 to 8.80; P = 0.32; 4 trials; 1404 participants). We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company due to lack of data.

TSA showed that 7.7% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. The TSA‐adjusted 95% CI was 0.39 to 3.01.

Health‐related quality of life

None of the included trials reported on this outcome.

Serious adverse effects

Six trials reported that a total of 1337 participants experienced a serious adverse event: in the M+S group 666/3300 (20.2%) participants had a serious adverse event compared with 671/3354 (20.0%) participants in the metformin plus thiazolidinedione group (RR 1.01, 95% CI 0.93 to 1.11; P = 0.80; 6 trials; 6654 participants; very low‐certainty evidence; Analysis 5.2). The 95% prediction interval ranged between 0.88 and 1.16.

Derosa 2011b compared M+S with metformin plus pioglitazone and reported that no participants experienced a serious adverse event but did not provide the number of participants included in the analysis.

A test for subgroup differences comparing rosiglitazone with pioglitazone did not indicate interaction (P = 0.84; Analysis 5.2). A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.28; Analysis 6.2).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis according to trials with low risk of selection bias did not substantially change the effect estimate (RR 1.01, 95% CI 0.92 to 1.10; P = 0.90; 5 trials; 6570 participants). Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 1.13, 95% CI 0.69 to 1.86; P = 0.63; 4 trials; 1404 participants). We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company due to lack of data.

TSA showed that 10.6% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 81%. The TSA‐adjusted 95% CI was 0.41 to 2.22.

Cardiovascular mortality

Four trials reported that a total of 78 participants died due to cardiovascular disease: in the M+S group 37/2946 (1.3%) participants died compared with 41/2994 (1.4%) participants in the metformin plus thiazolidinedione group (RR 0.78, 95% CI 0.36 to 1.67; P = 0.52; 4 trials; 5940 participants; low‐certainty evidence; Analysis 5.3). The 95% prediction interval ranged between 0.07 and 8.92. Vaccaro 2017 provided data from a time‐to‐event analysis on cardiovascular mortality: metformin plus pioglitazone versus M+S had a HR of 2.24 (95% CI 0.69 to 7.28, P = 0.18). Derosa 2011b, compared M+S with metformin plus pioglitazone and reported that no participants died due to cardiovascular disease but did not provide the number of participants included in the analysis.

A test for subgroup differences comparing rosiglitazone with pioglitazone did not indicate interaction (P = 0.40; Analysis 5.3). A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.36; Analysis 6.3).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias since all trials were evaluated as low risk of selection bias. We could not perform sensitivity analyses excluding large trials and trials funded by a pharmaceutical company due to lack of data.

TSA showed that 1.3% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 66%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Non‐fatal myocardial infarction

Three trials reported that a total of 46 participants experienced a non‐fatal myocardial infarction: in the M+S group 25/1841 (1.4%) participants had a non‐fatal myocardial infarction compared with 21/1877 (1.1%) participants in the metformin plus thiazolidinedione group (RR 1.21, 95% CI 0.68 to 2.14; P = 0.51; 3 trials; 3718 participants; very low‐certainty evidence; Analysis 5.4). The 95% prediction interval ranged between 0.03 and 48.76.

Vaccaro 2017 provided data from a time‐to‐event analysis on non‐fatal myocardial infarction: M+S versus metformin plus pioglitazone had a HR of 0.87 (95% CI 0.48 to 1.55; P = 0.63).

Derosa 2011b compared M+S with metformin plus pioglitazone and reported that no participants experienced a non‐fatal myocardial infarction but did not provide the number of participants included in the analysis.

A test for subgroup differences comparing rosiglitazone with pioglitazone did not indicate interaction (P = 0.58; Analysis 5.4). A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.58; Analysis 6.4).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. Sensitivity analyses excluding large trials and trials funded by a pharmaceutical company could not be performed due to lack of data.

TSA showed that 1.72% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Heart failure

Five trials reported that a total of 83 participants developed heart failure: in the M+S group 33/3259 (1.0%) participants developed heart failure compared with 50/3311 (1.5%) participants in the metformin plus thiazolidinedione group (RR 0.67, 95% CI 0.43 to 1.04; P = 0.08; 5 trials; 6570 participants; Analysis 5.5). The 95% prediction interval ranged between 0.33 and 1.37.

Vaccaro 2017 provided data from a time‐to‐event analysis on heart failure: the HR comparing participants in the sulphonylureas group with participants in the pioglitazone group was 1.57 (95% CI 0.76 to 3.24; P = 0.22).

Derosa 2011b compared M+S with metformin plus pioglitazone and reported that no participants developed heart failure but did not report the number of participants included in the analysis.

A test for subgroup differences comparing rosiglitazone with pioglitazone did not indicate interaction (P = 0.65; Analysis 5.5). A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.79; Analysis 6.5).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 0.51, 95% CI 0.12 to 2.07; P = 0.35; 3 trials; 1320 participants). We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company due to lack of data.

TSA showed that 3.3% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Non‐fatal stroke

Two trials compared M+S with metformin plus thiazolidinedione: in the M+S group 20/1540 (1.3%) participants had a non‐fatal stroke compared with 16/1583 (1%) participants in the metformin plus thiazolidinedione group (RR 1.29, 95% CI 0.67 to 2.47; P = 0.45; 2 trials; 3123 participants; very low‐certainty evidence; Analysis 5.6). Vaccaro 2017 administered pioglitazone combined with metformin and also provided data from a time‐to‐event analysis on non‐fatal stroke: M+S compared with metformin plus pioglitazone showed a HR of 0.79 (95% CI 0.41 to 1.53; P = 0.49; 3028 participants). Only Vaccaro 2017, in his trial of long duration, observed non‐fatal strokes (Analysis 6.6).

Derosa 2011b administered pioglitazone combined with metformin and reported that no participants experienced a non‐fatal stroke but did not provide the number of participants included in the analysis.

Amputation of lower extremity

Two trials with 3123 participants provided data on amputation of lower extremity (very low‐certainty evidence; Analysis 5.7). Neither of the trials reported any events.

Blindness or severe vision loss

One trial with 95 participants provided data on blindness or severe vision loss (very low‐certainty evidence; Analysis 5.8). They did not report any events.

End‐stage renal disease

One trial with 95 participants provided data on end‐stage renal disease (very low‐certainty evidence; Analysis 5.9). They did not report any events.

Non‐serious adverse events

Five trials reported that a total of 3072 participants experienced a non‐serious adverse event: in the M+S group 1510/2987 (50.6%) participants had a non‐serious adverse event compared with 1562/3037 (51.4%) participants in the metformin plus thiazolidinedione group (RR 0.94, 95% CI 0.44 to 2.01; P = 0.87; 5 trials; 6024 participants; Analysis 5.10). The 95% prediction interval ranged between 0.05 and 16.49.

Derosa 2011b compared M+S with metformin plus pioglitazone and reported that two and three participants experienced non‐serious adverse events, respectively. However, they did not provide the number of participants included in the analysis.

A test for subgroup differences comparing rosiglitazone with pioglitazone did not indicate interaction (P = 0.36; Analysis 5.10). A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.94; Analysis 6.7).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis according to trials with low risk of selection bias did not substantially change the effect estimate (RR 0.92, 95% CI 0.36 to 2.35; P = 0.87; 4 trials; 5940 participants). Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 0.97, 95% CI 0.76 to 1.25; P = 0.83; 3 trials; 774 participants). Sensitivity analysis excluding trials funded by a pharmaceutical company did not substantially change the effect estimate (RR 0.93, 95% CI 0.82 to 1.05; P = 0.24; 2 trials; 3123 participants).

TSA showed that the cumulative z‐curve crossed the futility boundaries suggesting that a 10% or greater RRR could be rejected at this point. Diversity was 0%. The TSA‐adjusted 95% CI was 0.74 to 1.19.

Hypoglycaemia

Five trials reported that a total of 926 participants had a mild or moderate hypoglycaemic episode: in the M+S group 721/2999 (24.0%) participants had a mild or moderate hypoglycaemic episode compared with 205/3060 (6.7%) participants in the metformin plus thiazolidinedione group (RR 3.63, 95% CI 2.98 to 4.44; P < 0.001; 5 trials; 6059 participants; Analysis 5.11 in favour of metformin plus thiazolidinedione). The 95% prediction interval ranged between 2.30 and 5.73. Derosa 2011b compared M+S with metformin plus pioglitazone and reported two and one events of mild/moderate hypoglycaemia, respectively. However, they did not provide the number of participants included in the analysis.

A test for subgroup differences comparing rosiglitazone with pioglitazone did not indicate interaction (P = 0.62; Analysis 5.11). A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.48; Analysis 6.8).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. Sensitivity analysis according to trials with low risk of selection bias did not substantially change the effect estimate (RR 3.63, 95% CI 2.92 to 4.52; P < 0.00001; 4 trials; 5975 participants). Sensitivity analysis excluding large trials did not substantially change the effect estimate (RR 5.99, 95% CI 2.43 to 14.76; P = 0.0001; 3 trials; 809 participants). Sensitivity analysis excluding trials funded by a pharmaceutical company did not substantially change the effect estimate (RR 3.37, 95% CI 2.84 to 3.99; P < 0.00001; 2 trials; 3123 participants).

TSA showed that 8.0% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 45%. The TSA‐adjusted 95% CI was 1.63 to 8.09.

Five trials reported that 36 participants experienced serious hypoglycaemia: in the M+S group 30/3259 (0.9%) participants had a serious hypoglycaemic episode compared with 6/3311 (0.2%) participants in the metformin plus thiazolidinedione group (random RR 3.98, 95% CI 0.34 to 46.01; P = 0.27; fixed RR 4.77, 95% CI 2.05 to 11.09; P < 0.001; 5 trials; 6570 participants; Analysis 5.12; in favour of metformin plus thiazolidinedione). However, there was substantial heterogeneity, probably caused by various definitions of serious hypoglycaemia. Only one trial provided a clear description of serious hypoglycaemia (Vaccaro 2017).

Derosa 2011b compared M+S with metformin plus pioglitazone and reported no serious hypoglycaemic events. However, they did not provide the number of participants included in the analysis.

A test for subgroup differences comparing rosiglitazone with pioglitazone showed a statistically significant difference between subgroups (P = 0.009; Analysis 5.12). However, CIs overlap slightly indicating that in fact there was no true interaction. A test for subgroup differences according to duration of follow‐up did not indicate interaction (P = 0.85; Analysis 6.9).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. We could not perform sensitivity analysis of large trials and trials funded by a pharmaceutical company due to lack of data.

TSA showed that 0.08% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 79%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Socioeconomic effects

None of the included trials reported on this outcome.

Weight

Seven trials reported weight change (random MD −0.6 kg, 95% CI −2.8 to 1.6; P = 0.62; fixed MD −2.0 kg, 95% CI −2.4 to −1.6; P < 0.001; 7 trials; 6877 participants; Analysis 5.13; in favour of M+S). The 95% prediction interval ranged between −8.3 kg and 7.2 kg.There was substantial heterogeneity probably caused by two of the rosiglitazone trials (Derosa 2005; Home 2009), and by duration of follow‐up (ranging from one year to 5.5 years), various sulphonylureas (glimepiride, glibenclamide, gliclazide) and various doses of rosiglitazone ranging from 4 mg/day to 8 mg/day.

HbA1c

Ten trials reported change in HbA1c (random MD 0.2%, 95% CI 0.04 to 0.3; P = 0.01; fixed MD 0.2%, 95% CI 0.1 to 0.2; P < 0.001; 10 trials; 7020 participants; Analysis 5.14; in favour of metformin plus thiazolidinedione). The 95% prediction interval ranged between −0.2% and 0.5%. There was substantial heterogeneity probably caused by duration of follow‐up (ranging from 1 year to 5.5 years), various sulphonylureas (glimepiride, glibenclamide, gliclazide) and various doses of sulphonylureas and thiazolidinediones.

Derosa 2009a reported HbA1c final values of 7.8% (SD 0.4) in the M+S group and 7.2% (SD 0.3) in the metformin plus pioglitazone group. However, they did not provide the number of participants included in the analysis.

All‐cause mortality

Three trials with 874 participants reported data on all‐cause mortality and one person died in each intervention group (low‐certainty evidence; Analysis 7.1).

Health‐related quality of life

None of the included trials reported on this outcome.

Serious adverse effects

Three trials reported that a total of 61 participants experienced a serious adverse event: in the M+S group 34/424 (8%) participants had a serious adverse event compared with 27/450 (6%) participants in the metformin plus thiazolidinedione group (RR 1.68, 95% CI 0.54 to 5.21; P = 0.37; 3 trials; 874 participants; low‐certainty evidence; Analysis 7.2).

In one trial serious adverse events were reported for up to six months and six to 12 months separately. For up to six months 5/126 (4.0%) participants in the M+S group had a serious adverse event compared with 7/130 (5.4%) participants in the metformin plus glinide group. For six to 12 months, 7/101 (6.9%) participants and 2/112 (1.8%) participants experienced a serious adverse event in the M+S group compared with the metformin plus glinide group, respectively (Ristic 2007).

Cardiovascular mortality

Two trials with 446 participants provided data on cardiovascular mortality (Derosa 2009b; Ristic 2007). No cardiovascular death was observed in either trial (low‐certainty evidence; Analysis 7.3).

Non‐fatal myocardial infarction

Two trials provided data on non‐fatal myocardial infarction in 446 participants. In total two non‐fatal myocardial infarctions were reported in 2/215 (0.9%) participants in the M+S group compared with 0/231 participants in the metformin plus thiazolidinedione group (2 trials; 446 participants; low‐certainty evidence). Derosa 2009b stated that no participants experienced a non‐fatal myocardial infarction. One trial reported non‐fatal myocardial infarction for up to six months and six to 12 months separately (Ristic 2007). For up to six months no event occurred in either intervention group. For six to 12 months, 2/101 (2%) participants had non‐fatal myocardial infarction in the M+S group compared to 0/112 participant in the metformin plus glinide group (RR 5.54, 95% CI 0.27 to 114.02; P = 0.27).

Heart failure

Two trials provided data on heart failure. Derosa 2009b stated that no heart failure occurred. One trial reported data on heart failure for up to six months and six to 12 months separately. For up to six months no event occurred in either intervention group. For six to 12 months, 0/101 participants and 1/112 (0.9%) participants developed heart failure in the M+S group compared to the metformin plus glinide group, respectively (Ristic 2007).

Non‐fatal stroke

Derosa 2009b stated that no non‐fatal stroke occurred (233 participants; very low‐certainty evidence).

Amputation of lower extremity

Derosa 2009b stated that no amputation of lower extremity occurred (233 participants; low‐certainty evidence).

Blindness or severe vision loss

Derosa 2009b stated that no blindness or severe vision loss occurred (233 participants; low‐certainty evidence).

End‐stage renal disease

Derosa 2009b stated that no end‐stage renal disease occurred (233 participants; low‐certainty evidence).

Non‐serious adverse events

Derosa 2009b stated that no non‐serious adverse events occurred.

Hypoglycaemia

Two trials provided data on mild or moderate hypoglycaemia. Gerich 2005 reported data from a subgroup analysis of participants 65 years and older: in the M+S group 7/40 (17.5%) participants had mild or moderate hypoglycaemia compared with 1/35 (2.9%) participants in the metformin plus glinide group (Analysis 7.4). Derosa 2009b reported two mild or moderate hypoglycaemic episodes in the M+S group compared with three events in the metformin plus glinide group.

Two trials provided data on serious hypoglycaemia. Gerich 2005 reported that 2/209 (1%) participants in the M+S group compared with 0/219 participants in the metformin plus glinide group experienced a serious hypoglycaemic episode (Analysis 7.5). Derosa 2009b stated that no serious hypoglycaemia occurred.

Socioeconomic effects

Neither of the included trials reported on this outcome.

Weight

Two trials reported weight change (MD 1.1 kg, 95% CI −0.1 to 2.3; P = 0.06; 2 trials; 619 participants; Analysis 7.6).

HbA1c

Three trials reported change in HbA1c (random MD 0.2%, 95% CI ‐0.6 to 1.0; P = 0.69; fixed MD 0.4%, 95% CI 0.3 to 0.5; P < 0.00001; 3 trials; 852 participants; Analysis 7.7; in favour of metformin plus glinide). Calculation of the 95% prediction interval did not provide a meaningful estimate. There was substantial heterogeneity probably caused by various durations of follow‐up (ranging from 52 weeks to 104 weeks), various sulphonylureas (glibenclamide and gliclazide) and various doses of nateglinide (ranging from 180 mg/day to 540 mg/day).

All‐cause mortality

Four trials reported that a total of 32 participants died: in the M+S group 13/2107 (0.6%) participants died compared with 19/3027 (0.6%) participants in the metformin plus SGLT‐2 inhibitor group (RR 0.96, 95% CI 0.44 to 2.09; P = 0.91; 4 trials; 5134 participants; very low‐certainty evidence; Analysis 8.1). The 95% prediction interval ranged between 0.17 and 5.30.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. Sensitivity analysis excluding long trials did not substantially change the effect estimate (RR 0.72, 95% CI 0.30 to 1.76; P = 0.47; 3 trials; 4320 participants). One trial reported data for 104 weeks (Ridderstråle 2014). We could not perform sensitivity analysis excluding large trials because all trials randomised fewer than 1000 participants to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 1.04% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Health‐related quality of life

None of the included trials reported on this outcome.

Serious adverse effects

Four trials reported that a total of 690 participants experienced a serious adverse event: in the M+S group 315/2107 (15.5%) participants had a serious adverse event compared with 375/3027 (12.4%) participants in the metformin plus SGLT‐2 inhibitor group (RR 1.02, 95% CI 0.76 to 1.37; P = 0.90; 4 trials; 5134 participants; very low‐certainty evidence; Analysis 8.2). The 95% prediction interval ranged between 0.30 and 3.51.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. Sensitivity analysis excluding long trials did not substantially change the effect estimate (RR 0.93, 95% CI 0.35 to 2.49; P = 0.88; 2 trials; 2775 participants). We could not perform sensitivity analysis excluding large trials because all trials randomised fewer than 1000 participants to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 12.4% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 78%. The TSA‐adjusted 95% CI was 0.31 to 3.36.

Cardiovascular mortality

Three trials reported that a total of 10 participants died due to cardiovascular disease: in the M+S group 4/1327 (0.3%) participants died compared with 6/2262 (0.3%) participants in the metformin plus SGLT‐2 inhibitor group (RR 1.22, 95% CI 0.33 to 4.41; P = 0.77; 3 trials; 3589 participants; very low‐certainty evidence; Analysis 8.3). Calculation of the 95% prediction interval did not provide a meaningful estimate.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. Sensitivity analysis excluding long trials did not substantially change the effect estimate (RR 1.02, 95% CI 0.25 to 4.18; P = 0.98; 2 trials; 2775 participants). We could not perform sensitivity analysis excluding large trials because all trials randomised fewer than 1000 participants to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 0.36% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Non‐fatal myocardial infarction

Two trials reported that a total of 15 participants experienced a non‐fatal myocardial infarction: in the M+S group 7/890 (0.8%) participants had a non‐fatal myocardial infarction compared with 8/1374 (0.6%) participants in the metformin plus SGLT‐2 inhibitor group (RR 1.43, 95% CI 0.49 to 4.18; P = 0.52; 2 trials; 2264 participants; very low‐certainty evidence; Analysis 8.4).

We could not perform sensitivity analyses due to lack of data.

TSA showed that 0.46% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Heart failure

Three trials reported that a total of four participants developed heart failure: in the M+S group 4/1670 (0.2%) participants developed heart failure compared with 0/2139 participant in the metformin plus SGLT‐2 inhibitor group (Peto OR 9.21, 95% CI 1.26 to 67.24; P = 0.03; 3 trials; 3809 participants; Analysis 8.5).

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. Sensitivity analysis excluding long trials did not substantially change the effect estimate but widened the CI (Peto OR 11.84, 95% CI 0.68 to 205.34; P = 0.09; 2 trials; 2264 participants). We could not perform sensitivity analysis excluding large trials because all trials randomised fewer than 1000 participants to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

None of the participants in the comparator group (metformin plus SGLT‐2 inhibitor) developed heart failure. TSA could therefore not be performed as it was not possible to calculate an event rate in the comparator group. However, applying 0.5 events in each of the trials in the comparator group showed that 10.9% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. The TSA‐adjusted 95% CI did not provide a meaningful estimate.

Non‐fatal stroke

Two trials reported that a total of 10 participants experienced a non‐fatal stroke: in the M+S group 3/919 (0.3%) participants had a non‐fatal stroke compared with 7/1856 (0.4%) participants in the metformin plus SGLT‐2 inhibitor group (RR 0.87, 95% CI 0.22 to 3.34; P = 0.83; 2 trials; 2775 participants; very low‐certainty evidence; Analysis 8.6).

We could not perform sensitivity analyses due to lack of data.

TSA showed that 0.37% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Amputation of lower extremity

Hollander 2017 reported one amputation of lower extremity in 437 (0.2%) participants in the M+S group compared with one amputation in 888 (0.1%) participants in the metformin plus SGLT‐2 inhibitor group (very low‐certainty evidence, Analysis 8.7).

Blindness or severe vision loss

None of the included trials reported on this outcome.

End‐stage renal disease

None of the included trials reported on this outcome.

Non‐serious adverse events

Three trials reported that a total of 2284 participants had non‐serious adverse events: in the M+S group 1139/1670 (68.2%) participants experienced a non‐serious adverse event compared with 1145/2139 (53.5%) participants in the metformin plus SGLT‐2 inhibitor group (RR 1.27, 95% CI 1.01 to 1.59; P = 0.04; 3 trials; 3809 participants; Analysis 8.8; in favour of metformin plus SGLT‐2 inhibitor). The 95% prediction interval ranged between 0.07 and 23.77. None of the trials provided a detailed definition of this outcome measure.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. Sensitivity analysis excluding long trials did not substantially change the direction of the effect estimate (RR 1.38, 95% CI 1.06 to 1.80; P = 0.02; 2 trials; 2264 participants). We included one trial with a duration of intervention longer than 104 weeks of treatment in the analysis because they reported data for non‐serious adverse events after 52 weeks of treatment (Del Prato 2015). We could not perform sensitivity analysis excluding large trials because all trials randomised fewer than 1000 participants to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 6.2% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 96%. The TSA‐adjusted 95% CI was 0.53 to 3.01.

Hypoglycaemia

Three trials reported that a total of 603 participants experienced mild or moderate hypoglycaemia: in the M+S group 514/1670 (30.8%) participants had a mild or moderate hypoglycaemic episode compared with 89/1639 (5.4%) participants in the metformin plus SGLT‐2 inhibitor group (RR 5.60, 95% CI 2.38 to 13.14; P < 0.001; 3 trials; 3309 participants; Analysis 8.9; in favour of metformin plus SGLT‐2 inhibitor). Calculation of the 95% prediction interval did not provide a meaningful estimate.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. We could not perform sensitivity analysis excluding long trials due to lack of data. We could not perform sensitivity analysis excluding large trials because all trials randomised fewer than 1000 participants to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 0.44% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 93%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Four trials reported that a total of 38 participants had serious hypoglycaemia: in the M+S group 30/2107 (1.4%) participants had a serious hypoglycaemic episode compared with 8/3027 (0.3%) participants in the metformin plus SGLT‐2 inhibitor group (RR 6.16, 95% CI 2.92 to 12.97; P < 0.001; 4 trials; 5134 participants; Analysis 8.10; in favour of metformin plus SGLT‐2 inhibitor). The 95% prediction interval ranged between 1.20 and 31.58. For one of the trials included in the meta‐analysis, the number of serious hypoglycaemic events was unclear due to varied reporting (Leiter 2015). We contacted the trial authors for clarification but did not receive a reply. To be sure to have included all serious hypoglycaemic events, we extracted the highest number of serious hypoglycaemic events reported.

Because all trials were published in English, we could not perform sensitivity analyses according to publication status or language of publication. We could not perform sensitivity analysis according to trials with low risk of selection bias because all trials were evaluated as low risk of selection bias. Sensitivity analysis excluding long trials did not substantially change the effect estimate (RR 6.39, 95% CI 2.89 to 14.12; P < 0.001; 2 trials; 2775 participants). We could not perform sensitivity analysis excluding large trials because all trials randomised fewer than 1000 participants to each intervention group. We could not perform sensitivity analysis excluding trials funded by a pharmaceutical company because all trials received funding from a pharmaceutical company.

TSA showed that 0.56% of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued. Diversity was 0%. As only a minor fraction of the diversity‐adjusted required information size to detect or reject a 10% RRR had been accrued, we could not calculate the TSA‐adjusted 95% CI.

Socioeconomic effects

One trial performed three economic analyses of trial data using the Cardiff Diabetes Model (Del Prato 2015). One economic analysis aimed to assess the cost‐effectiveness of SGLT‐2 inhibitor compared with sulphonylurea when added to metformin for treatment of people in the UK with diabetes mellitus inadequately controlled on metformin alone. There was a mean incremental benefit of 0.47 quality‐adjusted life years (95% CI 0.42 to 0.67), calculated for dapagliflozin plus metformin, the incremental cost‐effectiveness ratio point estimate was GBP 2671 per quality‐adjusted life year.

Another economic analysis aimed to assess the cost‐effectiveness of SGLT‐2 inhibitor compared with sulphonylurea when added to metformin for treatment of Nordic people with diabetes mellitus inadequately controlled on metformin alone. The mean lifetime gain in quality‐adjusted life years for metformin plus dapagliflozin compared to M+S was 0.25 in Denmark, 0.27 in Finland, 0.24 in Norway and 0.28 in Sweden. The cost per quality‐adjusted life year gained was EUR 7944 in Denmark, EUR 5424 in Finland, EUR 4769 in Norway and EUR 6093 in Sweden.

The third economic analysis aimed to assess the cost‐effectiveness of SGLT‐2 inhibitor compared with sulphonylurea when added to metformin for treatment of Spanish people with diabetes mellitus inadequately controlled on metformin alone. Dapagliflozin was a cost‐effective option compared with sulphonylureas and resulted in a cost per quality‐adjusted life year gained of EUR 3560.

Weight

Three trials reported weight change (MD 4.4 kg, 95% CI 4.1 to 4.8; P < 0.001; 3 trials; 3294 participants; Analysis 8.11; in favour of metformin plus SGLT‐2 inhibitor). The 95% prediction interval ranged between 2.7 kg and 7.3 kg.

HbA1c

Four trials reported HbA1c (random MD 0.1%, 95% CI −0.1 to 0.2; P = 0.26; fixed MD 0.1%, 95% CI 0.02 to 0.1; P = 0.005; 4 trials; 4182 participants; Analysis 8.12; in favour of metformin plus SGLT‐2 inhibitor). There was substantial heterogeneity, probably caused by various durations of intervention. For one trial (Hollander 2017), we extracted data after 52 weeks of treatment showing benefit of sulphonylurea treatment, whereas for the remaining trials we extracted data after 104 weeks or more showing benefit of SGLT‐2 inhibitor treatment. Furthermore, heterogeneity could have been caused by various sulphonylureas (glipizide and glimepiride) administered in various doses (1 mg/day to 8 mg/day) and various SGLT‐2 inhibitors (dapagliflozin, ertugliflozin, canagliflozin, empagliflozin).