From 1996 to 1999, DPP enrolled 3234 adults at high risk of developing diabetes as described elsewhere¹¹. Key inclusion criteria were adults age 25+, impaired glucose regulation (fasting plasma glucose 100–125 mg/dl and 2-hour 75g OGTT glucose 140–199 mg/dl) and the presence of overweight or obesity (BMI ≥25 kg/m²). Participants were randomly assigned to one of three interventions: intensive lifestyle (ILS) aimed to achieve and maintain a 7% weight loss and 150 min or more per week of moderate-intensity physical activity, or double-blind metformin 850 mg twice per day, or matching placebo. The lifestyle intervention has been previously described in detail¹⁴ and is available online through the Centers for Disease Control. The DPP study was ended in 2001 one year ahead of schedule when it was established that ILS reduced the incidence of diabetes by 58% and metformin reduced the incidence by 31% compared to placebo¹⁵. Study interventions were safe, as previously reported^15,16. During a one-year bridge period, participants were informed of the main study results, those assigned to metformin or placebo were unblinded, and all participants were offered a group-administered version of the lifestyle curriculum prior to initiation of DPPOS in 2002¹⁷. DPPOS enrolled 2779 (88%) of the original surviving participants, including 767 participants who had been diagnosed with diabetes. Group lifestyle sessions were offered to all participants every 3 months until 2014, and booster lifestyle sessions were offered to participants in the original ILS group twice annually through 2014. Those previously assigned to the metformin group continued to receive metformin 850 mg twice daily open label, unless the drug was discontinued for safety reasons or participants had developed diabetes with a fasting glucose of 140 mg/dl or more in DPP or glycated hemoglobin (HbA1c) of 7% or more in DPPOS, at which time, according to the protocol, medication management deferred to their own primary care provider.¹⁸. In 2012, incident cancer was added as a study outcome. Written informed consent was obtained from all participants, consistent with the Declaration of Helsinki and the study amendments were approved by each clincal center’s responsible institutional review board. Participants provided written informed consent after discussion with study staff. Physical activity was recorded for all participants using the MAQ to estimate the leisure physical activity as met hours per week¹⁹. Trial registration: ClinicalTrials.gov Identifier: NCT00004992, NCT00038727.

Following randomization in DPP, incident cancer cases were ascertained at each annual follow-up visit and/or were reported as a serious adverse event if there was hospitalization or death. In 2012, retrospective collection of medical records associated with the reported events was undertaken. Medical records were collected prospectively from 2012 onward. Detailed clinical records, including pathology reports, were sought from treating physicians or healthcare institutions. A National Cancer Institute (NCI) staff oncologist, blinded to participant treatment group and metformin use, conducted adjudication according to the Surveillance, Epidemiology, and End Results (SEER) Program guidelines 2012²⁰. In addition, cancer cases reported through death certificate were included. Reports of cancer were categorized as not cancer/not reportable, primary incident case, primary non-incident case, recurrence of a non-incident case, or recurrence of an incident case. Primary incident cases (primary events that occurred after randomization) were included in this analysis. If a participant experienced cancer at additional anatomical sites, only their first occurrence contributed to the composite cancer analyses, while their first occurrence of each specific anatomical-type contributed to the individual anatomical-type analysis. A questionnaire about cancer risk factors and cancer screening habits was conducted in 2012 and screening history was subsequently collected annually. Obesity-related cancer was defined as meningioma, multiple myeloma, or cancer of the esophagus, kidneys, uterus, ovaries, thyroid, breast (postmenopausal), liver, gallbladder, upper stomach, pancreas, colon, and rectum. The remainder of cancer types were considered as nonobesity-related². Only cancer sites where the total number of cases across the groups was greater than 20 were analyzed individually.

Primary analyses used the “intention-to-treat” paradigm including all participants and data as of the data lock on February 23, 2020. Cumulative incidence rates accounted for competing risk due to non-cancer deaths using Fine-Gray’s estimates²¹ in Cox proportional hazards models to assess treatment and covariate effects that may mediate or moderate treatment effects. Censoring date for the analysis was the date of last participant contact or the date of death, whichever was later. Follow-up after the last visit was included only in the event of participant death, which provides a better estimate of cancer incidence among participants who are deceased but overestimates incidence overall if mortality is associated with cancer²². The primary analysis used the first incident cancer event based on adjudication. It was determined a priori that this analysis would be performed when 170 events had occurred in the placebo group, such that the primary comparison of metformin vs. placebo had 85% power for a 2-sided significance level of 0.05, to detect a 28% risk reduction in total cancer based on prior observational data using a log-rank test³. ILS compared to placebo was a secondary outcome, as were treatment comparisons of site-specific cancer incidence such that individuals could have contributed to multiple cancer sites.

Interaction of each subgroup x treatment assessed the heterogeneity in treatment effects in prespecified baseline subgroups (sex, age, race/ethnicity, and baseline BMI) as well as post-randomization defined time-dependent subgroup for diabetes status without adjustment for multiplicity.

Sensitivity analyses were conducted to address factors that may have affected treatment response: 1) exclusion of participants with a personal history of cancer prior to randomization, 2) exclusion of cancers diagnosed within a year of randomization and 3) exclusion of cases adjudicated based solely on a death certificate. Extended follow-up increased the likelihood that time-varying confounders due to intercurrent event of diabetes may have attenuated the hypothesized effect of randomized metformin compared to placebo. Additional Cox models assessed the effect of lifestyle and metformin compared to placebo after adjustment for cancer risk factors and out-of-study metformin use to explore these potential sources of confounding²³. The etiologic effect of metformin was explored using marginal structural models (MSM) with truncated inverse probability censoring weighting (IPCW) and instrumental variable (IV) approach to account for out-of-study metformin and discontinuation of randomized metformin^24–26. Each method makes different assumptions about the observed data and their relationships and may reach different conclusions about the metformin effects to evaluate the complex and heterogeneous metformin effects.

Notation. We start by defining notation.

The IPCW method along with MSM attempts to correct the imbalance between actual metformin takers versus non-takers in 2 steps²⁴. The first step of the IPCW method censors participants when they switch treatments, i.e., participants who were randomized to the metformin group who stopped taking study metformin or placebo participants who started taking out-of-study metformin. This treatment censoring may be informative (i.e., not occurring at random) due to factors that are also associated with the outcomes. The second step of the IPCW method is to correct this informative censoring using weighted analyses. The weights are calculated as the inverse probability of having remained on the randomized assignment given a set of specified factors which are assumed to be related to both the probability of switching treatment and the outcome. The weights create a pseudo- or counterfactual population as if the participants had complied with their treatment assignments. The last step is to use a Cox MSM to estimate the treatment effect in the weighted population where participants who are censored due to switching off their assigned treatment are represented by other uncensored participants who are similar in terms of the specified confounders. Assuming all measured confounders are specified and there are no unmeasured confounders, and correct model specifications, the obtained estimator represents the counterfactual effect of the treatment on the outcome in a hypothetical setting where all participants remained on their randomized treatment. This method, however, does not account for informative loss to follow-up, i.e., not switching treatments but withdrawing or becoming lost. The MSM approach used the Stage 1 model to derive the weights using switch defined as either start of metformin in the Placebo group or stopping metformin in the metformin group with adjustment for age, race/ethnicity, sex, LDL meds, lipid, smoking, fasting glucose, 2 hour glucose, Hba1c, fasting insulin, alcohol use, family history of diabetes, hypertension, BMI, urine albumin-creatinine ratio and eGFR by ckd-epi. Stage 2 MSM included time dependent HbA1c and The IV method uses two-stage-residual-inclusion to derive the counterfactual effects of taking metformin, considering the presence of unmeasured confounding caused by treatment crossover²⁵. The first stage is a generalized linear model using metformin randomization, I, as the instrumental variable to predict the observed metformin usage X(t), adjusting for characteristics Z(t). Residuals in such models are independent from the metformin randomization and can be viewed as a summary of the confounders leading to extra or less metformin taking, C(t). In the second stage, the Cox models for the time to incident cancer, outcome Y(t), are fitted adjusting for both X(t) and residuals from the first stage. By adjusting for the residuals, we adjust for unmeasured confounders leading to individual-level less or extra metformin taking compared to their original assignments. Therefore, the effects of the observed metformin indicator in the second stage can be interpreted as the causal effects of metformin taking, i.e., the difference in the outcomes between participants taking or not taking metformin when potential confounders are adjusted through the residuals. Models in Stages 1 and 2 of the IV approach adjusted for age, race/ethnicity, sex, family history of diabetes, smoking, alcohol use, and BMI. All p-values are two-sided, with nominal statistical significance level of 0.05. All statistical analyses were performed with SAS 9.4 and R 3.5.2.

The CONSORT flow of the randomized cohort is shown in Figure 1. Baseline characteristics of the participants are presented in Table 1. Treatment groups were well balanced in terms of established cancer risk factors including smoking, alcohol use, BMI, and family history of cancer. At baseline, 141 (4.4%) participants reported a diagnosis of cancer prior to randomization. During DPP, adherence to metformin, defined as taking at least 80% of study prescribed pills, was about 70% and fell to ~50% during DPPOS¹⁸ (Figure 2). Metformin exposure across all study arms increased over time as some participants developed diabetes and were treated by their primary care provider outside of the study protocol. The mean metformin exposure time in the metformin arm after 21 years of median follow-up was 11.3 years (9.9 years of which were study metformin) compared to 2.7 years in the placebo group and 2.1 years in the ILS group (Figure 2). Participants in the ILS group achieved the weight loss and physical activity goals during the first study year, although there was subsequent weight re-gain and treatment group differences were attenuated over time (Figure 2). Of note, weight declined in all treatment groups during the later years of follow-up (~10–20), consistent with an aging cohort. Cancer screening in this cohort was very high when first collected in 2012, with over 96% of participants reporting screening for cancer at least once, mammogram and pap smears reported in 97% and 86% of women respectively, PSA tests reported by nearly three-quarters of men, and colorectal screening reported by 85% of the cohort (Table 1). There was no significant difference in cancer screening across randomized treatment groups.

During the retrospective collection of medical records, more than 98% of cases were adjudicated and approximately 96% of adjudicated events were classified as cancer. During the prospective collection of medical records, approximately 93% of cases were adjudicated and 96% of adjudicated events were classified as cancer. A death certificate was available as the only source of information in 7% of adjudications. During a median follow-up of 21 years, 173 participants in the metformin group had a first incident cancer compared to 191 participants in the placebo group and 182 participants in the ILS group (Table 2). The cumulative incidence (Figure 3) was 20.3% in the metformin group compared with 22.3% in the placebo group and 21.6% in the ILS group. Of all cancer events, approximately one half were categorized as obesity-related.

Randomization to metformin was not associated with incidence of total cancer risk (HR=0.90, 95% CI=0.73 to 1.10, p=0.31), obesity-related cancer risk (HR= 0.89, 95% CI=0.66 to 1.18, p=0.41), or non-obesity-related cancer risk (HR=0.93, 95% CI 0.69 to 1.24, p=0.61) compared to placebo (Table 2). Supplemental Table S1 examines the effect of randomization to metformin and ILS compared to placebo on the incidence of individual anatomical cancer sites. Randomization to metformin compared to placebo was not associated with a reduced risk in any of the individual cancer sites except kidney (HR=0.27, 95% CI=0.09 to 0.80, p=0.02) (Supplemental Table S1).

Randomization to ILS was not associated with a decrease in total cancer risk (HR=0.96, 95% CI=0.79 to 1.18, p=0.72), obesity-related cancer risk (HR= 0.93, 95% CI=0.70 to 1.23, p=0.61), or non-obesity-related cancer risk (HR=1.0, 95% CI=0.75 to 1.34, p=0.97) compared to placebo. ILS was not associated with a reduced risk of cancer in any of the individual cancer sites analyzed compared to placebo.

The effect of metformin and ILS stratified by age group, sex, race, BMI group, and time dependent diabetes status are shown in Figure 4. The effect is similar across the prespecified subgroups with no interaction between treatment effect and subgroup (Supplemental Table S2). The time varying effects of the interventions were also examined through a test of proportional hazards assumption of the ILS or metformin effects and through an interaction of treatment by the time periods defined by DPP (3.2 average years) or post-DPP. In both sets of analyses, the effect of ILS or metformin did change over time.

To further evaluate potential confounding due to demographics and cancer risk factors, adjusted models (Supplemental Table S3) showed that the risk of cancer in the metformin or ILS groups compared to placebo remained unchanged (metformin vs. placebo HR=0.86, 95% CI=0.69 to 1.06, p=0.16; ILS vs placebo HR=0.93, 95% CI=0.76 to 1.14, p=0.47). In sensitivity analyses, adjustments for personal cancer history, cases occurring within a year of randomization, and method of adjudication did not appreciably change the results. Due to the increasing out-of-study metformin use in all three randomized groups over time, models were constructed to explore the potential mediating effects of diabetes status and duration, and cumulative glycemic and metformin exposure (Supplemental Table S4). These models also did not change the overall result of the study. The lowest estimated hazard ratio based on a marginal structural model accounting for time dependent HbA1c for the effect of metformin vs. placebo was HR=0.81, 95% CI=0.61 to 1.09, p=0.16). The instrumental variable approach resulted in HR=0.83, 95% CI=0.63 to 1.08) for total cancer.

A significant limitation to this intention to treat analysis is the increasing use of metformin in the ILS and placebo groups over time. However, adjusted analyses and marginal structural models including this “cross-over treatment” did not appreciably change our results. Cross-over has the potential to impact any trial but a study where the prevention intervention under investigation is also the first line intervention for the disease being prevented faces increasing risk to the randomized analyses over time. In addition, adherence to study metformin declined over time, although there remained a wide difference in total metformin exposure between the groups. The intensity of the ILS intervention was not sustained after the initial 3.2 years and limited lifestyle intervention was offered to all treatment groups during DPPOS, leading to reduced differences in weight between the groups. Finally, the study was powered to detect a 28% reduction in cancer cases (metformin vs. placebo) and was underpowered to detected smaller differences as well as underpowered to examine individual anatomical sites or subgroups.

DPPOS represent the longest metformin treated cohort of individuals at risk for diabetes and a cohort in which there was significant and modestly sustained weight loss with the ILS intervention. However, in these participants, assignment to metformin or ILS did not reduce the risk of total cancer or obesity-related cancers. Metformin use in both the placebo group and the ILS group increased over time due to the development of diabetes and the intensity of the ILS intervention was not sustained. These features may have diluted the effect of the interventions.