The Diabetes Prevention Program (DPP) is the landmark head-to-head: in people with prediabetes, it randomized an intensive lifestyle program, metformin, and placebo against each other and followed incident type-2 diabetes. It is a clean instance of the telos’s drug-as-the-realistic- alternative comparison — a behaviour change and a drug tested on the same patient-important outcome, in the same trial. The headline: an intensive lifestyle program out-prevented the drug.
What DPP tested [@knowler2002]
A 3-arm RCT, «3234 nondiabetic persons with elevated fasting and post-load plasma glucose concentrations» (prediabetes — impaired fasting glucose + impaired glucose tolerance), randomized to:
- placebo;
- metformin «850 mg twice daily»;
- an intensive lifestyle-modification program with the goals of «at least a 7 percent weight loss and at least 150 minutes of physical activity per week» (a 16-lesson curriculum, one-to-one case managers, low-calorie low-fat diet).
Mean age 51, mean BMI 34.0, 68% women, 45% minority; average follow-up 2.8 years; primary outcome diabetes incidence (ADA criteria). A fourth arm (troglitazone) was discontinued in 1998 for liver toxicity and is not reported. The blinded phase was stopped ~1 year early on the data monitoring board’s advice once efficacy was demonstrated.
The head-to-head result — lifestyle beat the drug
Incidence: 11.0 / 7.8 / 4.8 cases per 100 person-years (placebo / metformin / lifestyle).
«The lifestyle intervention reduced the incidence by 58 percent (95 percent confidence interval, 48 to 66 percent) and metformin by 31 percent (95 percent confidence interval, 17 to 43 percent), as compared with placebo; the lifestyle intervention was significantly more effective than metformin.» (Knowler, 2002)
Lifestyle vs metformin directly: 39% lower incidence (95% CI 24 to 51). The comparison as a table (the value here is the joined comparison, not two separate arms):
| Placebo | Metformin | Lifestyle | |
|---|---|---|---|
| Incidence (cases/100 person-yr) | 11.0 | 7.8 | 4.8 |
| Relative reduction vs placebo | — | 31% (17-43) | 58% (48-66) |
| Cumulative incidence at 3 yr | 28.9% | 21.7% | 14.4% |
| Weight change (kg) | -0.1 | -2.1 | -5.6 |
| NNT (3 yr, to prevent one case) | — | 13.9 (8.7-33.9) | 6.9 (5.4-9.5) |
«To prevent one case of diabetes during a period of three years, 6.9 persons would have to participate in the lifestyle-intervention program, and 13.9 would have to receive metformin.» (Knowler, 2002)
So on the absolute frame (Baseline Risk and the Relative-Absolute Split), lifestyle prevents one case per ~7 people treated 3 years vs one per ~14 for metformin — roughly twice the absolute yield. The COI direction is worth naming under symmetric standards: the trial was funded in part by Bristol-Myers Squibb — which the paper notes «sells metformin in the United States» — and Parke-Davis, and one author held BMS stock, yet the finding runs against the drug the funder sells (the study metformin and placebo were supplied by Lipha Pharmaceuticals). So the conflict does not undercut the headline. (inferred from Knowler, 2002)
The subgroup split — lifestyle works broadly, metformin works selectively
This is the decision-relevant subgroup structure, and the two arms behave differently:
-
Lifestyle was effective across essentially every stratum. «The lifestyle intervention was highly effective in all subgroups.» It did not differ significantly by sex or race/ethnic group and was «at least as effective in older participants as it was in younger participants» (numerically largest in the oldest: 48% / 59% / 71% at ages 25-44 / 45-59 / >=60). Its one significant modifier: it was stronger at lower baseline post-load glucose — «Its effect was significantly greater among persons with lower base-line glucose concentrations two hours after a glucose load than among those with higher base-line glucose values.» (Knowler, 2002)
-
Metformin’s effect was concentrated in the more obese and more hyperglycemic — genuine effect modification (significant heterogeneity), near-null at the lean / lower-fasting-glucose end:
«The effect of metformin was less with a lower body-mass index or a lower fasting glucose concentration than with higher values for those variables. Neither interaction was explained by the other variable or by age.» (Knowler, 2002)
Metformin vs placebo, by stratum Reduction (95% CI) BMI 22 to <30 3% (-36 to 30) — near-null BMI 30 to <35 16% (-19 to 41) BMI >=35 53% (36 to 65) Fasting glucose 95-109 mg/dl 15% (-12 to 36) — near-null Fasting glucose 110-125 mg/dl 48% (33 to 60) -
Consequently the lifestyle advantage over metformin was largest exactly where metformin was weakest — «The advantage of the lifestyle intervention over metformin was greater in older persons and those with a lower body-mass index than in younger persons and those with a higher body-mass index.» (Knowler, 2002)
The decision-change: metformin is a stratum-specific drug for prevention (worth most in the younger, more obese, more hyperglycemic; near-useless in the lean/near-normal-fasting), while lifestyle is a broad-spectrum lever that works across strata. The route-(b) machinery this instantiates is worked on Baseline Risk and the Relative-Absolute Split (one trial showing an arm with modified relative effect beside an arm with a roughly constant one). (inferred from Knowler, 2002)
Outcome scope — diabetes incidence is shown; CV/mortality is NOT this trial
DPP’s endpoint is diabetes incidence — a disease-onset outcome, better than a pure surrogate (it is a diagnosis, not a lab marker), but not a hard cardiovascular or mortality endpoint, and DPP is neither powered nor long enough for those. «Hospitalization and mortality rates were unrelated to treatment.» The paper is explicit that the downstream question is unanswered here and belongs to later follow-up:
«These questions should be addressed by continued follow-up of the study participants and by analysis of the main secondary outcomes — reductions in risk factors for cardiovascular disease, in the proportion of participants with atherosclerosis, and in the proportion with cardiovascular disease, which is the leading cause of death among patients with type 2 diabetes.» (Knowler, 2002)
So the honest claim from DPP alone was: preventing/delaying the diagnosis of diabetes is demonstrated; that this translates into fewer heart attacks, less nephropathy, or longer life is a separate claim not established by DPP — it depends on «whether the maintenance of these lower levels improved the long-term outcome». That downstream question is now answered by the DPP Outcomes Study (DPPOS) — see below. Diabetes-incidence prevention sits one rung below a hard outcome on the Surrogate Outcomes ladder, and DPPOS shows why the page was right not to let the incidence win stand in for a mortality win.
DPPOS — the hard-outcome answer: diabetes prevention did NOT reduce CV events over 21 years
DPPOS followed the same three arms a median 21 years for adjudicated cardiovascular events. The result is null on the hard endpoint: «Neither metformin nor lifestyle intervention reduced the primary outcome: metformin versus placebo hazard ratio 1.03 (95% CI, 0.78–1.37; P = 0.81) and lifestyle versus placebo hazard ratio 1.14 (95% CI, 0.87–1.50; P = 0.34).» The point estimates trend the wrong way for lifestyle, and «No effect of either intervention was seen on the extended cardiovascular outcome».
Effect estimates (primary MACE = nonfatal MI, nonfatal stroke, or CV death; 310 first events among 3234; studied range = a middle-aged, ethnically diverse prediabetes cohort, 21-yr median follow-up):
| Arm vs placebo | HR (95% CI) | P | Event rate /1000py (plac/arm) |
|---|---|---|---|
| Metformin — MACE | 1.03 (0.78–1.37) | 0.81 | 5.28 / 5.51 |
| Lifestyle — MACE | 1.14 (0.87–1.50) | 0.34 | 5.28 / 6.10 |
| Metformin — extended CV | 1.00 (0.80–1.25) | 0.99 | 8.73 / 8.86 |
| Lifestyle — extended CV | 1.12 (0.90–1.39) | 0.29 | 8.73 / 9.93 |
Prespecified subgroups «showed no significant heterogeneity by age, sex, race/ ethnicity, or diabetes development for either metformin or lifestyle» — so this is not a route-(b) story on the CV outcome (contrast the metformin route-(b) modification on the diabetes-incidence outcome above — the effect modification was outcome-specific, present for incidence, absent for events). The discussion flags one borderline exploratory signal — a sex difference for lifestyle (P=0.053, «potentially harmful in women yet somewhat protective in men») — which the authors themselves treat as possibly «the result of a play of chance», not a finding.
This is a bounded null, not a demonstration of no effect. The paper attributes the dilution explicitly: «Provision of group lifestyle intervention to all, extensive out-of-study use of statin and antihypertensive agents, and reduction in the use of study metformin together with out-of-study metformin use over time may have diluted the effects of the interventions.» Study metformin adherence fell «from 77% at the end of DPP to 41%»; statins reached 56–62% and antihypertensives 68–74%; «This was therefore a relatively low-risk cohort from the standpoint of the prevention of CVD.»
The baseline-risk reading (route-(a), no subgroup claim needed): a longer 30-year Da Qing follow-up did show a lifestyle MACE benefit, «hazard ratio 0.74 [95% CI, 0.59–0.92]», but «the Da Qing cohort was a higher-risk population» (more smokers, more hypertension, worse hyperglycemia, higher event rate). Absolute CV benefit scales with baseline risk, so a low-risk cohort has little absolute risk to remove and even a real relative effect is hard to detect — the Baseline Risk and the Relative-Absolute Split point. (inferred from Goldberg et al., 2022)
Decision-change: the diabetes-incidence win is real but does not license claiming a hard-CV or mortality benefit from diabetes prevention in a low-risk prediabetes stratum over two decades. Where CV risk is the concern in such a stratum, the demonstrated levers are the direct ones (statin / BP control — heavily used here); a CV payoff from diabetes prevention itself, if any, would need a higher-risk or longer-latency stratum. The Surrogate Outcomes page carries this as a worked surrogate-disconnect.
Independent replication — Finnish DPS and Da Qing (type-E robustness)
The lifestyle-prevention effect is not one trial: three randomized trials on three continents, in
three separate IGT/prediabetes populations, each found a statistically significant relative reduction in
incident diabetes for a multi-component lifestyle program versus usual-care advice. This is the value the
DPP-family ingest banks — the head-to-head above rests on a single US trial, but the lifestyle leg of
it is independently corroborated [E-independent].
The matched-parameter comparison (the same quantity? discipline, required before any cross-trial magnitude claim):
| Parameter | US DPP (Knowler 2002) | Finnish DPS (Tuomilehto 2001) | Da Qing (Pan 1997) | Same quantity? |
|---|---|---|---|---|
| Stratum | IFG+IGT, mean BMI 34, age 51, US | IGT, BMI>=25 (mean 31), age 40-65, Finland | IGT, mean BMI 25.8, age 45, China | IGT — yes; adiposity differs (34 / 31 / 26) |
| Lifestyle target | 7% wt loss, 150 min/wk activity | 5% wt loss, fat <30%E, fiber, 30 min/day | diet and/or exercise (BMI-scaled) | multi-component lifestyle — comparable class |
| Comparator | placebo | general advice | general info | usual-care/advice — yes |
| Follow-up | mean 2.8 yr | mean 3.2 yr | 6 yr | NO — horizon differs |
| Metric | Cox RRR on incident diabetes, ITT | Cox RRR (HR 0.4) on incident diabetes, ITT | Cox RRR adj. baseline BMI+FPG, ITT | Cox-derived ITT RRR — yes |
| RRR, full lifestyle vs control | 58% (48-66) | 58% (HR 0.4; CI 0.3-0.7) | 42% combined (diet 31 / exercise 46) | comparable RRR; Da Qing combined = closest match |
The three headline RRRs are the same quantity — Cox-model, intention-to-treat relative risk reductions on incident diabetes against a usual-care comparator — so they are legitimately comparable. What is not matched is the follow-up horizon (2.8 / 3.2 / 6 yr) and the baseline adiposity, and those differences explain why Da Qing’s point estimate is lower without contradicting the others: over 6 years more of the control group eventually progresses, compressing the relative reduction. The defensible claim is therefore three independent trials, concordant direction, RRR 31-58%, NOT all three showed 58%.
- Finnish DPS (Tuomilehto et al., 2001): 522 overweight middle-aged Finns with IGT; «the cumulative incidence of diabetes was 58 percent lower in the intervention group than in the control group (hazard ratio, 0.4; 95 percent confidence inter- val, 0.3 to 0.7; P<0.001)» — cumulative incidence 11% (6-15) intervention vs 23% (17-29) control at four years. Its absolute frame: «22 sub- jects with impaired glucose tolerance must be treated in this way for one year — or 5 subjects for five years — to prevent one case of diabetes.» (NNT 5 over 5 yr — a different horizon from DPP’s NNT 6.9 over 3 yr, so the two NNTs are not directly comparable; each is read against its own follow-up.)
- Da Qing (Pan et al., 1997): 577 Chinese with IGT, randomized by clinic to diet / exercise / diet+exercise / control; «In a proportional hazards analysis adjusted for differences in baseline BMI and fasting glucose, the diet, exercise, and diet-plus-exercise interventions were associated with 31% (P < 0.03), 46% (P < 0.0005), and 42% (P < 0.005) reductions in risk of developing diabetes, respectively.» Absolute frame: «The cumulative incidence of diabetes at 6 years was 67.7% (95% CI, 59.8-75.2) in the control group compared with 43.8% … in the diet group, 41.1% … in the exercise group, and 46.0% … in the diet-plus-exercise group» — an absolute cumulative measure, not the same quantity as the Cox RRRs and not to be conflated with them.
Two decision-relevant refinements Da Qing adds that the US and Finnish trials could not (type-F):
- Combining diet and exercise was NOT additive. «the efficacy of diet was similar to that of exercise, and there was no additional benefit of combining the interventions.» Da Qing is the only one of the three to randomize the components separately, and the combined arm (42%) did not beat exercise alone (46%). The decision-change: for a lever chosen on effect size, a single well-adhered component may capture most of the available prevention — stacking diet and structured exercise buys little additional incidence reduction in this stratum, though it may aid weight-loss maintenance (see mediation below).
- The relative effect held in lean and overweight alike. «The relative decrease in rate of development of diabetes in the active treatment groups was similar when subjects were stratified as lean or overweight (BMI < or >25 kg/m2).» So the lifestyle effect is not confined to the obese — it is a broad-spectrum lever across the adiposity range (consistent with DPP’s «highly effective in all subgroups»). Da Qing’s mean BMI 25.8 also extends the transportability of the effect below DPP’s mean BMI 34.
Independence classification — a genuine [E-independent] convergence, with one honest lineage caveat.
The three are separate randomized experiments in separate populations, designed in different eras (Da Qing
1986, Finnish 1993, US DPP mid-1990s), separately funded, on separate primary data — the substantive
replication independence (the effect is not one dataset re-analyzed) holds strongly. Author lists do not
overlap: no author is shared across the three, and neither Da Qing (the earliest, citing neither) nor
Finnish (which cites Da Qing only as corroborating prior evidence, not as a data source) re-analyzes
another’s cohort. The caveat, named under symmetric standards: an NIDDK-Phoenix institutional
thread does connect two of them — Da Qing co-authors Bennett and Howard were at NIH/NIDDK, the same
diabetes-epidemiology program Knowler (US DPP) belongs to, and Knowler is acknowledged «for his important
contribution to the planning of our study» in the Finnish DPS. So the three teams are not hermetically
sealed; the cleanest fully-independent pair is Finnish DPS vs Da Qing (no shared authors, different
continents, neither citing the other as antecedent). The convergence is real and raises confidence on the
lifestyle-prevention claim from single-landmark to replicated; the lineage thread is why it is marked
with the caveat rather than as three sealed-off routes.
(Pan et al., 1997; inferred from Tuomilehto et al., 2001)
Why lifestyle beat metformin — mechanism, and what DPP could NOT attribute
Metformin and lifestyle reduced fasting glucose similarly, but lifestyle did more on post-load glucose and HbA1c — consistent with metformin’s narrow mechanism: «metformin suppresses endogenous glucose production, the main determinant of fasting plasma glucose concentrations.» (Knowler, 2002) The lifestyle arm acts on the broader defect (insulin resistance via weight and activity), so it normalizes the post-load excursion the drug leaves largely untouched.
The weight-loss mediation is intuitive but NOT tested here. The lifestyle arm lost ~5.6 kg and the temptation is to read the 58% as a weight-loss dose-response — but the trial explicitly disclaims this:
«The study, however, was not designed to test the relative contributions of dietary changes, increased physical activity, and weight loss to the reduction in the risk of diabetes, and the effects of these components remain to be determined.» (Knowler, 2002)
So the widely-quoted each kg lost ~16% lower diabetes risk figure is from a later DPP mediation analysis — Hamman 2006 — not the primary trial. That analysis is now held (below).
The weight-loss mediation — Hamman 2006 (type-F, the DPP cohort re-analyzed, NOT a fourth trial)
Hamman re-analyzed the DPP lifestyle arm alone (1,079 ILS participants, mean BMI 33.9, 3.2-yr follow-up) to decompose which component of the intervention carried the prevention. This is the same DPP cohort, not an independent replication — it is a mediation/dose-response under the DPP effect (type-F mechanism), so it does not add a second data point to the type-E convergence above and must not be double-counted as one.
- Weight loss was the dominant mediator, with a clean per-kg dose-response. «Weight loss was the dominant predictor of reduced diabetes incidence (hazard ratio per 5-kg weight loss 0.42 [95% CI 0.35–0.51]; P <0.0001). For every kilogram of weight loss, there was a 16% reduction in risk, adjusted for changes in diet and activity.» (Hamman et al., 2006) In the full multivariate model «Weight loss was the dominant factor … (HR 0.42 per 5 kg; P < 0.0001; R2 = 8.09%)» — and its share of explained variance dwarfs the others: even after correcting the diet/activity measures for imprecision, «the proportion of explained variance for weight change remains substantially higher than for either percent fat or physical activity» (partial R2 ~8% for weight vs ~0.1-0.2% for percent-fat and activity). (Hamman et al., 2006)
- But physical activity has a weight-loss-INDEPENDENT effect — and this is where Hamman and Finnish DPS
converge (type-E on the sub-claim). Among DPP participants who failed the weight-loss goal, activity
still cut risk: «Among 495 participants not meeting the weight loss goal at year 1, those who achieved the
physical activity goal had 44% lower diabetes incidence.» (Hamman et al., 2006) The Finnish DPS found the same pattern in a different cohort: «Among the subjects in
the intervention group who did not reach the goal of losing 5 percent of their initial weight, the odds
ratio for diabetes in those who had achieved the goal with respect to exer- cise (more than four hours per
week) during the first year was 0.2 (95 percent confidence interval, 0.1 to 0.6)» — while losing >5%
weight itself carried «odds ratio … 0.3 (95 percent confidence interval, 0.1 to 0.7)».
(Tuomilehto et al., 2001) Two independent cohorts agreeing that activity
prevents diabetes even without weight loss is a genuine
[E-independent]sub-finding.
Reconciling the two “dominant lever” readings. Weight loss explains the most variance in who progresses (Hamman), yet activity works even in non-losers (Hamman + Finnish), and Da Qing found exercise alone at least as effective as diet and non-additive with it. The coherent decision-frame: weight loss is the primary route and the thing to steer toward, but physical activity is a partly-independent second channel — valuable in its own right for the person who cannot achieve the weight target, and the mechanism that sustains weight loss once achieved. This connects to the weight-loss cluster (Does Weight Loss Reduce Cardiovascular Events) as mechanism, not as evidence that weight loss per se is the sole mediator. (inferred from Hamman et al., 2006; Tuomilehto et al., 2001)
Adherence and adverse events
Adherence to the lifestyle goals decayed — the adherence-is-part-of-the-effect point in a landmark program: «Fifty percent of the participants in the lifestyle-intervention group had achieved the goal of weight loss of 7 percent or more by the end of the curriculum (at 24 weeks), and 38 percent had a weight loss of at least 7 percent at the time of the most recent visit» (the activity goal: 74% at 24 weeks, 58% at last visit). (Knowler, 2002) The 58% reduction was delivered despite this real-world decay — it is the effect of a program as actually adhered to, not of perfect compliance.
Harms were modest and differed by arm: gastrointestinal symptoms highest on metformin (77.8 vs 12.9 events/100 person-yr on lifestyle), musculoskeletal symptoms highest on lifestyle (24.1). No deaths were attributed to any intervention. (Knowler, 2002)
Decision relevance
(Hamman et al., 2006; inferred from Knowler, 2002; Pan et al., 1997; Tuomilehto et al., 2001)
- For prediabetes, an intensive lifestyle program is first-line — it roughly doubles the absolute prevention yield of metformin (NNT ~7 vs ~14 over 3 years) and works across strata. Frame it as the substitution it is: lifestyle versus the drug, on the same outcome, not as an add-on.
- Metformin is the realistic alternative where lifestyle is not delivered or not enough — and its value is stratum-dependent: substantial in the younger / more obese / more hyperglycemic, near-null in the lean with near-normal fasting glucose. Prescribing it to the latter buys little.
- The advice is outcome-scoped. It prevents/delays the diagnosis; whether that averts heart attacks or extends life is not shown by DPP and should not be promised on its strength (see Outcome scope above).
- Intensity matters. DPP’s lifestyle arm was «systematic and intensive» with individualized case management — the 58% is the effect of a resourced program, and a thinner real-world implementation is a different, likely weaker, exposure.
- Vitamin D is a distant third lever for this stratum, not a substitute. An IPD meta-analysis of 3 prediabetes-prevention RCTs (n=4190) found vitamin D cut diabetes incidence by only 15% (HR 0.85, 0.75-0.96), «a number of persons with prediabetes needed to treat of 30 (compared with 7 with intensive lifestyle modification and 14 with metformin in the Diabetes Prevention Program study)» (Pittas et al., 2023). On absolute yield it ranks well below both DPP arms — a cheap, low-harm add-on where 25(OH)D is low or BMI is not high, never a replacement for the lifestyle or metformin rock -> Vitamin and Mineral Supplements for Disease Prevention.
Limits and provenance
- The head-to-head vs metformin is a single landmark RCT (US DPP) — but the lifestyle leg is now
replicated: the Finnish DPS (58%, HR 0.4) and Da Qing (31-46% by arm) are held and woven above as a
[E-independent]three-trial convergence (see Independent replication). What remains single-trial is specifically the lifestyle-versus-metformin comparison — neither Finnish nor Da Qing had a metformin arm, so the claim that lifestyle out-prevents the drug still rests on DPP alone. - Prediabetes-selected, US, mean BMI 34 — transports to similar prediabetic populations; the placebo incidence (11.0/100 person-yr) was higher than observational data, from selecting people with three strong risk factors plus frequent testing, so the absolute numbers are for a high-risk stratum.
- Diabetes incidence, not hard outcomes (see Outcome scope) — the binding limit on what this page licenses.
- Component contributions unattributed (diet vs activity vs weight loss) by design.