Metformin Does Not Affect CancerRisk: A Cohort Study in the U.K.Clinical Practice ResearchDatalink Analyzed Like anIntention-to-Treat TrialDiabetes Care 2014;37:2522–2532 | DOI: 10.2337/dc14-0584

OBJECTIVE

Meta-analyses of epidemiologic studies have suggested that metformin may re-duce cancer incidence, but randomized controlled trials did not support thishypothesis.

RESEARCH DESIGN AND METHODS

A retrospective cohort study, Clinical Practice Research Datalink, was designed toinvestigate the association between use of metformin compared with other anti-diabetes medications and cancer risk by emulating an intention-to-treat analysisas in a trial. A total of 95,820 participants with type 2 diabetes who started takingmetformin and other oral antidiabetes medications within 12 months of theirdiagnosis (initiators) were followed up for first incident cancer diagnosis withoutregard to any subsequent changes in pharmacotherapy. Cox proportional hazardsmodels were used to estimate multivariable-adjusted hazard ratios (HR) and95% CI.

RESULTS

A total of 51,484 individuals (54%) were metformin initiators and 18,264 (19%)were sulfonylurea initiators, and 3,805 first incident cancers were diagnosedduring a median follow-up time of 5.1 years. Compared with initiators of sulfo-nylurea, initiators of metformin had a similar incidence of total cancer (HR 0.96;95% CI 0.89–1.04) and colorectal (HR 0.92; 95% CI 0.76–1.13), prostate (HR 1.02;95% CI 0.83–1.25), lung (HR 0.85; 95% CI 0.68–1.07), or postmenopausal breast (HR1.03; 95% CI 0.82–1.31) cancer or any other cancer.

CONCLUSIONS

In this large study, individuals with diabeteswho usedmetformin had a similar riskof developing cancer compared with those who used sulfonylureas.

The effects of antidiabetes medications on cancer risk have recently attractedsignificant public interest. In particular, meta-analyses of observational studieshave found that the biguanide metformin, which is an insulin sensitizer and themost commonly used first-line therapy for type 2 diabetes, may reduce cancerincidence (1,2). However, a meta-analysis of randomized controlled trials did notsupport this hypothesis (3).

1Department of Hygiene and Epidemiology, Uni-versity of Ioannina School of Medicine, Ioannina,Greece2Cancer Epidemiology Unit, University of Oxford,Oxford, U.K.3Clinical Trial Service Unit, University of Oxford,Oxford, U.K.4Lipid Disorders Clinic, Department of InternalMedicine, University Hospital of Ioannina, Ioan-nina, Greece5Division of Epidemiology, University of TexasSchool of Public Health, Houston, TX6Department of Epidemiology and Biostatistics,School of Public Health, Imperial College London,St Mary’s Campus, London, U.K.7Human Genetics Foundation (HuGeF), Turin,Italy8University College London Institute of ChildHealth, Centre for Paediatric Epidemiology andBiostatistics, London, U.K.9Imperial Clinical Trials Unit, School of PublicHealth, Imperial College London, St Mary’s Hos-pital, London, U.K.10Stanford Prevention Research Center, Depart-ment of Medicine; Department of Health Re-search and Policy, Stanford University School ofMedicine; and Department of Statistics, StanfordUniversity School of Humanities and Sciences,Stanford, CA

Corresponding author: Konstantinos K. Tsilidis,ktsilidi@cc.uoi.gr, ktsilidis@gmail.com.

Received 6 March 2014 and accepted 24 April2014.

This article contains Supplementary Data onlineat http://care.diabetesjournals.org/lookup/suppl/doi:10.2337/dc14-0584/-/DC1.

© 2014 by the American Diabetes Association.Readers may use this article as long as the workis properly cited, the use is educational and notfor profit, and the work is not altered.

Konstantinos K. Tsilidis,1,2

Despoina Capothanassi,1 Naomi E. Allen,3

Evangelos C. Rizos,4 David S. Lopez,5

Karin van Veldhoven,6–8

Carlotta Sacerdote,7 Deborah Ashby,9

Paolo Vineis,6,7 Ioanna Tzoulaki,1,6 and

John P.A. Ioannidis10

2522 Diabetes Care Volume 37, September 2014

EPIDEM

IOLO

GY/HEA

LTHSERVICES

RESEA

RCH

Many of the published pharmacoepi-demiologic studies are difficult to inter-pret, however, because they may besubject to several biases (4). The pro-gressive nature of type 2 diabetes re-quires changes in pharmacotherapyover time and makes very difficult theassessment of the independent associa-tion of a specific medication and cancerrisk. The lack of randomization makesobservational effect estimates vulnera-ble to confounding by indication due tothe different prognoses of individualsbetween treatment groups. Moreover,most randomized clinical trials are notdesigned or sufficiently powered to ex-amine cancer outcomes due to shortfollow-up periods and very few cancerevents.To overcome these limitations, we

emulated the design and analysis of atrial in a large retrospective cohort studywithin the U.K. Clinical Practice Re-search Datalink (CPRD), one of theworld’s largest electronic medical re-cord databases, to investigate asso-ciations of cancer risk among usersof metformin compared with users ofsulfonylurea or other first-line oral hy-poglycemic agents (OHAs) among indi-viduals with newly diagnosed diabetes.This “incident diabetes drug” cohort de-sign and intention-to-treat (ITT) analysiscan be regarded as the equivalent of anonrandomized “trial” that avoids manyof the biases of traditional observationalstudies (5).

RESEARCH DESIGN AND METHODS

Data SourceThis study was conducted using the U.K.CPRD, previously known as GeneralPractice Research Database (6), andwas approved by the Medicines andHealthcare Products Regulatory AgencyIndependent Scientific Advisory Com-mittee. The CPRD was established in1987 and currently encompasses morethan 5million people enrolled from over600 general practitioners nationwide.The people enrolled in the CPRD are rep-resentative of the U.K. population withregard to age, sex, and geographical dis-tribution (7). Prior studies have foundcomplete agreement between prescrip-tion information received from generalpractitioners and that recorded in thedatabase (7,8). The diagnostic codescontained in the CPRD have shown ex-cellent agreement with those from

medical records, with 95% of neoplasmsand 88% of endocrine andmetabolic dis-orders identified in the CRPD confirmedwith alternate data sources (8,9).

Study PopulationWe extracted data on all participantswith type 2 diabetes aged 35 to 90 yearswho were prescribed at least one anti-diabetes agent between 1 January 1987and 31 December 2010, provided thatthe first prescription was dated at least6 months after the CPRD registrationdate of the participant. This was to en-sure that most of the included partici-pants with diabetes would be new usersof antidiabetes drugs and to excludeprevalent users with unknown typeand duration of treatment. In addition,eligible participants had never a historyof cancer. We also excluded the initial12 months of follow-up after the firstantidiabetes prescription, because anycancer diagnoses occurring within thattime are unlikely to be attributable tothe medications. Of the initial 113,301participants with type 2 diabetes identi-fied, 95,820 were available for statisticalanalysis (Fig. 1).

Exposure AssessmentExposure to antidiabetes medicationswas assessed by the presence of antidia-betes prescription records in the CPRDusing British National Formulary codes.Treatment-naive participants with dia-betes at CPRD enrollment who firststarted using metformin (metformin ini-tiators) were compared with initiatorsof other OHAs. Exposure to certain anti-diabetes drugs was classified based onthe initial 12-month treatment period,and an individual’s treatment patternwas categorized into one of the follow-ing mutually exclusive groups:

1. monotherapy with metformin, ifthey were only exposed to metfor-min during their initial 12-monthtreatment period;

2. monotherapy with sulfonylureas,either first generation (tolbuta-mide, chlorpropamide, tolazamide,acetohexamide) or second genera-tion (gliclazide, glibenclamide, glipizide,glimepiride, gliquidone, glibornuride,glymidine sodium);

3. monotherapy with thiazolidinediones(rosiglitazone, pioglitazone);

4. monotherapy with other OHAs, in-cluding a-glucosidase inhibitors

(acarbose), meglitinide analogs(nateglinide, repaglinide), GLP-1receptor antagonists (exenatide),dipeptidyl peptidase 4 inhibitors(sitagliptin, vildagliptin);

5. monotherapy with insulin;6. combination therapies with metfor-

min if the participants were simulta-neously exposed to two or moreOHAs including metformin duringthe initial 12-month treatmentperiod;

7. combination therapies with insulin;and

8. other combination therapies.

The main analysis was based on a com-parison of those who started using met-formin compared with those who usedsulfonylureas because these two medi-cations were the most common and be-cause sulfonylureas had similar clinicalindications to metformin.

Outcome AssessmentFirst incident cancer cases were definedby the presence of National Health Ser-vice Read codes in the CPRD. The Readcode dictionary was first searched toidentify malignant neoplasms, excludingnonmelanoma skin cancers, using sev-eral sensitive and specific algorithmsand code lists. Two researchers (K.K.T.and D.C.) manually reviewed and con-firmed the codes identified by the initialsearches and excluded codes for border-line, in situ, or suspected malignancies.A similar process was followed to definetype 2 diabetes.

Covariate AssessmentInformation on sociodemographic data,lifestyle characteristics, anthropometricvariables, medical conditions, and treat-ments was extracted from the CPRD atapproximately the time of the first anti-diabetes prescription (index date). Ageand sex were recorded at the time ofthe first prescription. BMI measure-ments and smoking status (never, for-mer, current) were retrieved for the1 year before the index date. Alcoholconsumption (never, former, current)was defined from information derivedfrom up to 5 years before the indexdate to diminish the proportion ofmissing values for this variable. Treat-ment with aspirin or nonsteroidal anti-inflammatory drugs (NSAIDs), statins, andexogenous hormones in women was re-trieved from the number of prescriptions

care.diabetesjournals.org Tsilidis and Associates 2523

for the year before the index date. Partic-ipants who had amammogram, a colono-scopy, or a prostate-specific antigenmeasurement during the 10 years beforethe index datewere considered as havingever received a cancer-screening test.Duration of diabetes was calculated

from the date of diabetes diagnosis tothe index date. Diabetic complicationsor hospitalizations any time beforethe index date were also considered.Glycemic control was defined as atime-weighted average of glycosylatedhemoglobin (HbA1c) concentrations forthe year before the index date. Becausemenopausal status is not directly re-corded in the CPRD, we used the ageof 49 years, previously reported asthe median age of menopause amongBritish women (10), to definemenopausalstatus for the breast cancer analysis. A

sensitivity analysis was conducted using55 years as the cutoff value to increasethe specificity of the definition.

Statistical AnalysisWithin this retrospective cohort, cancerrisk was compared among those whofirst started using metformin and thosewho started using other OHAs, and thenparticipants were followed up withoutregard to any subsequent changes inpharmacotherapy (i.e., akin to an ITTanalysis). Cox proportional hazardsmodels were used to estimate hazardratios (HR) and 95% CI for cancer riskusing the date since start of follow-upas the underlying time scale. Follow-up time began at the end of the 12-month period after the first prescriptionand ended at the diagnosis of the firstcancer, death, loss to follow-up, or cen-soring at the end of the follow-up (31

December 2011), whichever came first.Proportionality of hazards was verifiedbased on the slope of the Schoenfeldresiduals over time. All models werestratified by age (in 5-year groups from35 to 90 years) at the index date and sex.Models were adjusted for known or sus-pected risk factors for cancer, includingsmoking status (never, former, current),BMI (,18.5, 18.5–25.0, 25.0–30.0,$30.0 kg/m2), alcohol consumption sta-tus (never, former, current), use of aspi-rin or NSAIDs (based on percentiledistribution of the number of prescrip-tions: no, sparse [1–5 prescriptions],medium [6–16 prescriptions], or fre-quent use [.16 prescriptions]), use ofstatins (no, sparse [1–6 prescriptions],medium [7–15 prescriptions], or fre-quent use [.15 prescriptions]), use ofexogenous hormones in women (no,

Figure 1—Selection criteria for the study population in the CPRD.

2524 Antidiabetes Drugs and Risk of Cancer in CPRD Diabetes Care Volume 37, September 2014

sparse [1–2 prescriptions], medium [3–7prescriptions], or frequent use [.7 pre-scriptions], for colorectal, breast, andendometrial cancer), diabetes duration(in days), and year of the first antidiabe-tes prescription. Missing values wereassigned to separate categories forsmoking status (12.5% missing data),BMI (7.5%), alcohol consumption(28.4%), and duration of diabetes(9.1%). Analyses that excluded partici-pants with missing values for any ofthese covariates and analyses that in-cluded further adjustments for familyhistory of cancer, cancer screening his-tory, diabetes complications, and HbA1cproduced very similar results and arenot presented here.Although following the design and the

statistical analysis (ITT) of a trial maycorrect for confounding by indication,it may misclassify exposure over timeif participants change medication,thereby biasing the results toward thenull. Therefore, adherence curves dur-ing follow-up were plotted to evaluatethe amount of nonadherence, andadherence-adjusted estimates werecalculated using inverse probabilityweighting. Participants were censoredat the discontinuation of their initialtreatment, and the uncensored person-years were weighted by the inverse oftheir estimated probability of remaininguncensored. The probabilities of re-maining on treatment for each partici-pant at years 1, 3, and 5 after the start offollow-up were estimated through lo-gistic regression models adjusted forall baseline covariates (outlined above)and the most recent postbaseline valuesof these covariates in a time-dependentfashion. Separate regression modelswere used for initiators of metforminand initiators of other OHAs. To improveprecision of the models, the inverseprobability weights were stabilized bymultiplying the weights by the probabil-ity of censoring given the baseline val-ues of the covariates, as describedelsewhere (5).Several other sensitivity analyses

were performed. First, to better definecontinuous antidiabetes medicationuse, participants who had a treatmentbreak of more than 90 days during theinitial qualifying 12-month exposure pe-riod were excluded. Second, the expo-sure to an antidiabetes drug class wasredefined based on the first 6 months of

each participant’s prescription recordinstead of the 12 months used in theprimary analysis. Third, we excludedparticipants with any treatment breaksof more than 90 days during the latterqualifying 6-month exposure period.Fourth, we excluded the initial 36months of follow-up after the first anti-diabetes prescription (instead of just 12months) in case cancers diagnosedwithin 36 months were still not due tothe antidiabetes treatment.

Prespecified stratified analyses wereconducted according to age at recruit-ment (,62 vs. $62 years), BMI (,30vs. $30 kg/m2 or ,25 vs. 25–30 vs.$30 kg/m2), smoking status (nevervs. ever), NSAID/aspirin use (never vs.ever), and statin use (never vs. ever),because these variables may influencecancer risk and the decision to start oradhere to a certain antidiabetes phar-macotherapy. Tests for interactionwere done by using the relevant expo-sure variables, indicator variables forthe potentially modifying factors, andproduct terms of the two variables.The P values for interaction are inter-preted in light of the 90 (15 cancers 36 modifying variables) comparisonsmade. All P values were two-sided, andall analyses were performed usingSTATA 12 software (StataCorp LP, Col-lege Station, TX).

RESULTS

Baseline Characteristics of the StudyPopulationIncluded were 95,820 participants withtype 2 diabetes and at least one antidia-betes prescription recorded in CPRD(Fig. 1). Of these, 51,484 individuals(54%) started taking metformin mono-therapy, 18,264 (19%) started sulfonyl-urea monotherapy, and a smallernumber started taking each of the othertherapies (Table 1). The study did notfurther analyze 4,124 participants (4%)who started antidiabetes therapy withinsulin because they likely had type 1diabetes or long-standing type 2 diabe-tes that required insulin treatment.

Table 1 also reports the baselinecharacteristics according to the firstantidiabetes treatment prescription.Compared with those who started usingsulfonylurea monotherapy, those whostarted using metformin monotherapywere, on average, younger (mean age:61.1 vs. 65.3 years), had a higher BMI

(mean: 32.1 vs. 27.7 kg/m2), weremore likely to be current alcoholdrinkers (57% vs. 44%), were more likelyto frequently use aspirin/NSAIDs (12%vs. 8%) or statins (15% vs. 6%), andhad a longer median duration of diabe-tes before the time of their first antidia-betes prescription (99 vs. 54 days), andtheir median year of the first prescrip-tion was more recent (2005 vs. 2000). Inearly study years, patients were morelikely to initiate oral hypoglycemic ther-apy with a sulfonylurea. More than 75%of the metformin monotherapy groupinitiated treatment after 2003. Themedian follow-up time for the metfor-min and sulfonylurea monotherapy ini-tiators was 5.1 years (interquartile range2.9–8.1 years; maximum 24 years) andwas not different in the two groups. Atotal of 3,805 first incident cancers wereidentified, of which 599 were colorectalcancers, 580 were prostate cancers, 468were lung cancers, and 460 were post-menopausal breast cancers.

ITT Estimates for the Effect ofMetformin on CancerCompared with those who started usingsulfonylurea monotherapy, those whostarted using metformin had a similarincidence of total cancer (HR 0.96; 95%CI 0.89–1.04; Fig. 2) after adjustment forsmoking status, BMI, alcohol consump-tion, use of aspirin or NSAIDs, statins,diabetes duration, and year of first anti-diabetes prescription. The HR was 1.06(95% CI 0.93–1.22) when the analysiswas restricted to the first 3 years offollow-up, 1.09 (95% CI 0.98–1.22) forthe first 5 years, and 1.02 (95% CI 0.93–1.12) for the first 8 years. Compared withusers of sulfonylurea monotherapy,users of metformin also had similar risksfor colorectal (HR 0.92; 95% CI 0.76–1.13), prostate (HR 1.02; 95% CI 0.83–1.25), lung (HR 0.85; 95% CI 0.68–1.07),and postmenopausal breast (HR 1.03;95% CI 0.82–1.31) cancers and anyother cancer types (Fig. 2). When theage of 55 years was used as the cutoffvalue to define postmenopausal status,the association for postmenopausalbreast cancer remained the same (HR1.01; 95% CI 0.79–1.30) and was alsosimilar for total breast cancer risk (HR1.04; 95% CI 0.82–1.32).

We further performed an analysis be-tween metformin and sulfonylureausers who first started taking these

care.diabetesjournals.org Tsilidis and Associates 2525

Table

1—Base

linech

aracteristicsbyantidiabetestreatm

entin

theCPRD

Metform

inmono-tx

Metform

incombo-tx

Sulfonylureamono-tx

TZDsmono-tx

Other

OHAmono-tx

Other

OHAcombo-tx

Insulin

mono-tx

Insulin

combo-tx

n=51

,484

n=20

,967

n=18

,264

n=29

2n=26

3n=42

6n=1,72

3n=2,40

1

Sex Male

28,893

(56.1)

11,437

(54.6)

10,578

(57.9)

170(58.2)

156(59.3)

246(57.7)

979(56.8)

1,33

0(55.4)

Female

22,591

(43.9)

9,53

0(45.4)

7,68

6(42.1)

122(41.8)

107(40.7)

180(42.3)

744(43.2)

1,07

1(44.6)

Age,years

35–49

9,02

4(17.5)

4,32

9(20.7)

1,90

1(10.4)

38(13.0)

45(17.1)

71(16.7)

547(31.8)

767(31.9)

50–64

22,086

(42.9)

9,12

0(43.5)

6,32

7(34.6)

120(41.1)

111(42.2)

162(38.0)

641(37.2)

941(39.2)

65–79

17,616

(34.2)

6,50

3(31.0)

7,94

6(43.5)

112(38.4)

91(34.6)

160(37.6)

469(27.2)

609(25.4)

$80

2,75

8(5.4)

1,01

5(4.8)

2,09

0(11.5)

22(7.5)

16(6.1)

33(7.7)

66(3.8)

84(3.5)

BMI,kg/m

2

,18

.546

(0.1)

36(0.2)

124(0.7)

2(0.7)

3(1.1)

3(0.7)

11(0.6)

37(1.5)

18.5–24

.94,03

7(7.8)

2,11

2(10.1)

4,43

5(24.3)

49(16.8)

42(16.0)

125(29.3)

306(17.8)

563(23.5)

25–29

.916

,321

(31.7)

6,48

5(30.9)

6,79

7(37.2)

105(36.0)

103(39.2)

142(33.3)

460(26.7)

638(26.6)

$30

29,098

(56.5)

11,024

(52.6)

3,83

3(21.0)

128(43.8)

94(35.7)

109(25.6)

468(27.2)

932(38.8)

Unknown

1,98

2(3.9)

1,31

0(6.2)

3,07

5(16.8)

8(2.7)

21(8.0)

47(11.1)

478(27.7)

231(9.6)

Smoking

Never

21,214

(41.2)

8,05

2(38.4)

6,51

4(35.7)

135(46.2)

93(35.4)

168(39.4)

547(31.8)

811(33.8)

Form

er18

,225

(35.4)

7,07

5(33.7)

4,19

4(23.0)

108(37.0)

57(21.7)

118(27.7)

379(22.0)

734(30.6)

Curren

t8,38

9(16.3)

3,54

5(16.9)

2,53

(13.8)

42(14.4)

41(15.6)

56(13.2)

283(16.4)

516(21.5)

Unknown

3,65

6(7.1)

2,29

5(11.0)

5,02

6(27.5)

7(2.4)

72(27.3)

84(19.7)

514(29.8)

340(14.1)

Alcohol

Never

8,54

7(16.6)

3,63

8(17.4)

2,86

8(15.7)

45(15.4)

45(17.1)

71(16.7)

215(12.5)

390(16.2)

Form

er1,42

6(2.8)

575(2.7)

283(1.6)

6(2.0)

4(1.5)

10(2.3)

50(2.9)

95(3.9)

Curren

t29

,427

(57.1)

10,641

(50.7)

8,05

6(44.1)

169(57.9)

129(49.1)

190(44.6)

595(34.5)

1,10

8(46.2)

Unknown

12,084

(23.5)

6,11

3(29.2)

7,05

7(38.6)

72(24.7)

85(32.3)

155(36.4)

863(50.1)

808(33.7)

Aspirin/N

SAID

use*

No

20,414

(39.7)

9,02

7(43.0)

9,33

3(51.1)

114(39.0)

136(51.7)

211(49.5)

889(51.6)

1,05

7(44.0)

Sparse

13,391

(26.0)

5,82

9(27.8)

4,35

4(23.8)

65(22.3)

69(26.2)

110(25.8)

419(24.3)

717(29.9)

Med

ium

11,382

(22.1)

4,06

4(19.4)

3,12

4(17.1)

74(25.3)

31(11.8)

75(17.6)

319(18.5)

423(17.6)

Freq

uen

t6,29

7(12.2)

2,04

7(9.8)

1,45

3(8.0)

39(13.4)

27(10.3)

30(7.1)

96(5.6)

204(8.5)

Statin

use*

No

19,352

(37.6)

9,70

9(46.3)

13,422

(73.5)

81(27.7)

178(67.7)

249(58.5)

1,12

5(65.3)

1,23

8(51.6)

Sparse

10,783

(20.9)

4,65

1(22.2)

1,75

8(9.6)

74(25.3)

28(10.6)

66(15.5)

283(16.4)

585(24.4)

Med

ium

13,54(26.3)

4,33

4(20.7)

1,99

1(10.9)

79(27.1)

27(10.3)

79(18.5)

233(13.5)

397(16.5)

Freq

uen

t7,80

9(15.2)

2,27

3(10.8)

1,09

3(6.0)

58(19.9)

30(11.4)

32(7.5)

82(4.8)

181(7.5)

Horm

oneuse*†

No

19,193

(85.0)

7,83

0(82.2)

6,84

1(89.0)

106(86.9)

91(85.0)

149(82.8)

624(83.9)

821(76.7)

Sparse

1,28

9(5.7)

633(6.6)

316(4.1)

5(4.1)

5(4.7)

10(5.6)

54(7.3)

89(8.3)

Med

ium

1,38

9(6.1)

697(7.3)

341(4.4)

6(4.9)

8(7.5)

16(8.9)

52(6.9)

115(10.7)

Freq

uen

t72

0(3.2)

370(3.9)

188(2.5)

5(4.1)

3(2.8)

5(2.7)

14(1.9)

46(4.3)

Con

tinuedon

p.25

27

2526 Antidiabetes Drugs and Risk of Cancer in CPRD Diabetes Care Volume 37, September 2014

medications between 2000 and 2005when the transition from sulfonylureato metformin became evident and indi-viduals with diabetes could equallylikely start on either drug; but again,the results remained the same and notstatistically significant for any cancer(HR 0.95 for total cancer; 95% CI 0.86–1.06). The strongest confounder in themodels for total cancer risk and for mostseparate cancer sites was the year of thefirst antidiabetes prescription, althoughinclusion of this covariate in the modelsdid not change the inference from therisk estimates (Supplementary Table 1).

There was also no difference in cancerrisk when initiators of metformin werecompared with initiators of any otherOHA monotherapy (sulfonylureas, thia-zolidinediones, other OHAs) or whencombination treatments of metforminor any other OHA were added to themonotherapy regimens (SupplementaryFigs. 1 and 2). Very few participants (785[4.3%]) used a first-generation sulfonyl-urea, and the results were identicalwhen metformin monotherapy initia-tors were compared with initiators ofsecond-generation sulfonylureas (datanot shown).

Adherence-Adjusted Effect EstimatesApproximately half of the participantsadhered to their initial therapy 3 yearsafter the start of follow-up, whereasonly 20% remained on their initial ther-apy after 5 years (Fig. 3). The inverseprobability weighted HRs for total can-cer comparing metformin with sulfonyl-urea monotherapy were 0.96 (95% CI0.89–1.03) for the first year of follow-up, 0.95 (95% CI 0.87–1.04) for the first3 years, and 0.94 (95% CI 0.85–1.04)for the first 5 years. Similar estimateswere observed for the associations bycancer site, with few exceptions(Supplementary Table 2). Use of met-formin was associated with a higherrisk of postmenopausal breast cancer(first year: HR 1.13; 95% CI 0.91–1.41;first 3 years: HR 1.33; 95% CI 1.02–1.74; first 5 years: HR 1.47; 95% CI1.07–2.02) and endometrial cancer(first year: HR 2.51; 95% CI 1.36–4.65;first 3 years: HR 2.04; 95% CI 1.01–4.11; first 5 years: HR 2.29; 95% CI0.98–5.38), and with a lower risk ofpancreatic cancer (first year: HR 0.66;95% CI 0.45–0.97; first 3 years: HR0.70; 95% CI 0.44–1.10; first 5 years:

HR 0.60; 95% CI 0.37–0.98) comparedwith sulfonylurea monotherapy after ad-justment for nonadherence, but these in-triguing findings were not statisticallysignificant for all evaluated follow-uptime points.

Sensitivity AnalysesAfter we excluded participants with anytreatment breaks of more than 90 daysduring the initial qualifying 12-monthexposure period, the results (HR 0.95;95% CI 0.88–1.03) were almost identicalto themain analysis (Fig. 4). Similarly, allother sensitivity analyses, including re-defining the exposure groups based onthe first 6 months of each participant’sprescription record or excluding the ini-tial 36 months of follow-up, led to al-most identical results for total cancer(Fig. 4) or by cancer site (SupplementaryFigs. 3 and 4).

Subgroup AnalysesWe next examined potential multiplica-tive interactions, stratifying our ITT esti-mates by age at recruitment, BMI,smoking status, and NSAID/aspirin andstatin use. Eight interactions of 90 werestatistically significant at the 0.05 level,but none of these remained significantwhen we used an adjustment for mul-tiple comparisons with a Bonferroni-corrected P value of 0.0006 (Fig. 5 andSupplementary Figs, 5–9). The strongestsuggestion of an interaction was bysmoking status: the HR of total cancercomparing metformin with sulfonylureamonotherapy was 1.15 (95% CI 0.99–1.33) in never smokers and was 0.88(95% CI 0.78–0.98; P = 0.001 for interac-tion) in ever smokers (Fig. 5).

CONCLUSIONS

Main FindingsHere, we have reanalyzed the CPRD datausing an approach that tried to emulate aclinical trial in all regards other than ran-domization. Treatment-naive participantswith diabetes who first started usingmetformin or sulfonylurea monotherapyduring their initial 12-month treatmentperiod were compared and followed upwithout regard to any subsequentchanges in their treatment. We foundthat initiators of metformin had a similarcancer incidence as the initiators of sulfo-nylurea over a maximum follow-up of24 years. Similar results were obtainedwhen metformin users were comparedwith users of other OHAs and after

Table

1—Continued

Metform

inmono-tx

Metform

incombo-tx

Sulfonylureamono-tx

TZDsmono-tx

Other

OHAmono-tx

Other

OHAcombo-tx

Insulin

mono-tx

Insulin

combo-tx

n=51

,484

n=20

,967

n=18

,264

n=29

2n=26

3n=42

6n=1,72

3n=2,40

1

Diabetes

duration,d

ays‡

Med

ian(IQR)

99(1–61

8)35

(0–36

0)54

(0–41

4)

259(5–86

8)

206(6–76

5)

30(0–34

7)

18(0–10

5)

30(0–34

7)

Missing,%

5.6

8.1

17.0

9.2

14.8

15.6

39.0

8.7

Firstyear

oftherap

yMed

ian(IQR)

2005

(200

3–20

07)

2004

(200

1–20

07)

2000

(199

6–20

02)

2005(200

4–20

06)

2000(199

8–20

03)

2002(199

9–20

04)

2001(199

7–20

04)

2004(200

1–20

07)

Figu

resaren(%

)unless

otherwisespecified

.Combo-tx,combinationtherapy;IQR,interquartilerange;m

ono-tx,monotherapy;TZDs,thiazolidined

iones.*Based

onthepercentiledistributionofthenumber

of

prescriptionsas

nouse,sparse

use

(1–10

%ofallprescriptions),m

edium

use

(10–

50%ofallprescriptions),andfreq

uen

tuse

(.50

%ofallprescriptions).Theexactdistributionalcutpointswereuse

ofaspirinor

NSA

IDs:no,sparse

[1–5prescriptions],m

edium

[6–16

prescriptions],frequen

tuse

[.16

prescriptions];use

ofstatins:no,sparse

[1–6prescriptions],med

ium

[7–15

prescriptions],frequen

tuse

[.15

prescriptions];u

seofexogenoushorm

ones

inwomen

:no,sparse

[1–2prescriptions],m

edium

[3–7prescriptions],frequen

tuse

[.7prescriptions].†Exogenoushorm

oneuse

(oralcontracep

tivesandhorm

one

replacemen

ttherap

y)in

women

.‡From

timeofdiabetes

diagnosisto

timeoffirstantidiabetes

treatm

entprescription.

care.diabetesjournals.org Tsilidis and Associates 2527

adjustment for nonadherence to theinitial treatment.

Comparison With Other StudiesMetformin has been shown to re-duce the concentrations of circulating

glucose and insulin by reducing the pro-duction of glucose by liver cells and in-creasing insulin sensitivity, respectively(11). Metformin has also direct antipro-liferative effects because it activatesthe liver enzyme AMPK, which has

been shown to inhibit the growth ofcancer cells in several in vitro and invivo studies (12,13). The upstream reg-ulator of AMPK is the protein kinaseLKB1, which is a well-recognized tumorsuppressor.

Figure 2—Multivariable-adjusted HRs and 95% CIs for the association between initiators of metformin monotherapy and initiators of sulfonylureamonotherapy and cancer risk using the ITT principle in the CPRD. NHL, non-Hodgkin lymphoma.

Figure 3—Proportion of participants who adhered to their initial antidiabetes treatment in the CPRD.

2528 Antidiabetes Drugs and Risk of Cancer in CPRD Diabetes Care Volume 37, September 2014

In accordance to the mechanistic ev-idence, results from some epidemio-logic studies conducted in the 2000ssuggested that treatment with metfor-min was associated with a reduced risk ofcancer incidence or mortality (14–21).However, these studies were generallylimited in their ability to assess an asso-ciation with specific cancer types, andmost of them compared metforminwith any other antidiabetes drug regard-less of disease severity. Thus, the inverseassociation observed for metformin andcancer risk might be at least partially dueto an increased risk of cancer in long-term patients with diabetes who needto take insulin rather than metformin

to manage their disease. In addition,some studies have shown that the reduc-tion in risk with the use of metforminstarts from the first year of follow-up,which seems biologically implausible(16,21). Several of these studies also suf-fered from immortal time bias (15,16),whereby misclassification of unexposedtime initially in a cohort as metforminexposure during follow-up may createspurious reductions in risk (4).

However, some other more recentstudies have found no association be-tween metformin use and cancer risk,which is consistent with our findings.A retrospective cohort in the Kaiser Per-manente Northern California Diabetes

Registry found no association with cancerrisk between ever users ofmetformin, de-fined as two prescriptions in a 6-monthperiod, versus never users, although themaximum duration of follow-up in thatstudy was less than 6 years (22).

The association of antidiabetes med-ication on cancer risk was also previ-ously analyzed within the CPRD dataset (23–28). Qiu et al. (23) analyzedthe data from 1995 to 2008 using theITT principle and reported that sulfo-nylurea use versus metformin mono-therapy use, defined as at least sixsequential prescriptions of each drug,was not significantly associated withthe risk of total cancer (HR 1.07; 95% CI

Figure 4—Multivariable-adjusted HRs and 95% CIs for the association between initiators of metformin monotherapy and initiators of sulfonylureamonotherapy and total cancer risk using the ITT principle in the CPRD by type of sensitivity analyses. (Sensitivity analysis 1: after excludingparticipants with any treatment breaks of more than 90 days during the initial qualifying 12-month exposure period. Sensitivity analysis 2: afterredefining the exposure to an antidiabetes drug class based on the first 6 months of each participant’s prescription record compared with the 12-month treatment period used in the primary analysis. Sensitivity analysis 3: after excluding participants with any treatment breaks of more than 90days during the latter qualifying 6-month exposure period. Sensitivity analysis 4: after excluding the initial 36 months of follow-up after the firstantidiabetes prescription).

care.diabetesjournals.org Tsilidis and Associates 2529

0.98–1.15) or colorectal, breast, or pros-tate cancer, but this analysis included ashorter follow-up time, reported sepa-rate results only for a few cancer sites,and did not adjust for nonadherence. Incontrast, a nested case-control analysisusing data from 1988 to 2009 found apositive association between ever useof metformin, defined as at least oneprescription between cohort entry andend of follow-up, versus never use andrisk of prostate cancer (HR 1.23; 95% CI0.99–1.52), but this analysis did not usethe ITT principle and had a potentiallybiased definition (never vs. ever use) ofthe exposure groups (28). Other reportsfrom the CPRD have generally observednull associations for metformin and can-cer risk but reported results for a fewcancer sites. Some of these studiesused potentially biased exposure defini-tions as ever use versus never use ofmetformin without an ITT analysis ap-proach, and other studies conducted

time-varying exposure analyses and/orincorporated dose and duration of anti-diabetes medications in their statisticalmodels but without adjusting for time-varying confounders (25,26,28).

Our findings agree with two large ran-domized trials. The ADOPT (A DiabetesOutcome Progression Trial) reported nodifference in cancer risk between partic-ipants randomized to metformin andthose randomized to glibenclamide (rel-ative risk 0.78; 95% CI 0.53–1.14) or ro-siglitazone (HR 0.92; 95% CI, 0.63–1.35),based on 155 cases during 4 years offollow-up (29). The RECORD (Rosiglita-zone Evaluated for Cardiovascular Out-comes and Regulation of Glycemia inDiabetes) also found no difference incancer risk between individuals receiv-ing metformin and sulfonylurea versusrosiglitazone and sulfonylurea (relativerisk 1.22; 95% CI 0.86–1.74), based on125 cases over a mean of 5.5 years offollow-up (30).

ITT estimates are generally attenu-ated toward the null due to changes inadherence to medication over time. Ourstudy is the first, to our knowledge, totake account of likely changes in adher-ence. Although only 20% of the metfor-min or sulfonylurea monotherapy userswere still using these regimens after5 years, the adherence-adjusted esti-mates were not appreciably different,with few exceptions. Compared withsulfonylurea users, metformin usersseem to have a higher risk for breastand endometrial cancer and a lowerrisk for pancreatic cancer, althoughthese associations were not alwaysnominally statistically significant whenwe corrected for nonadherence at sev-eral follow-up cut points, and the CIswere often very wide. CPRD did nothave consistent time-varying con-founder information, and this mighthave affected the validity of theadherence-adjusted estimates. Future

Figure 5—Multivariable-adjusted HRs and 95% CIs for the association between initiators of metformin monotherapy and initiators of sulfonylureamonotherapy and total cancer risk using the ITT principle in the CPRD by subgroups.

2530 Antidiabetes Drugs and Risk of Cancer in CPRD Diabetes Care Volume 37, September 2014

prospective studies and large consortiashould attempt to clarify this issue.

Strengths and WeaknessesOur study is one of the largest to date toinvestigate the association betweenmetformin use and risk of cancer. Thecomparison of drug initiators ensuredthat our estimates were not explainedby unknown antidiabetes treatmentand/or other variables that occurred be-fore entry to the cohort. Sulfonylureawas used as the main comparison groupand should have reduced confoundingby indication because sulfonylureashad similar indications to metforminand were prescribed as first-line antidia-betes therapy in the 1980s and most of1990s in the U.K. The follow-up spannedup to 24 years, and 50% of our studypopulation had follow-up time formore than 5 years.Our study is not without potential

limitations. Patients were not assignedto therapy in a random manner, andtherefore, potential residual confound-ing or residual confounding by indica-tion is possible. Several potentialconfounders were included in our statis-tical models, most of which did notgreatly change the effect estimates.The CPRD lacked information for someother potential confounders, such asdiet, physical activity, race/ethnicity,and reproductive factors, but these var-iables are not considered the strongestconfounders for pharmacoepidemio-logic cancer outcomes. Moreover,CPRD did not have consistent time-varying confounder information, andthis was a reason for not performing atime-varying analysis of actual prescrip-tion to complement the ITT analysis, be-cause such an analysis or an analysis thatincorporates information on durationand dose of antidiabetes treatmentswithout proper adjustment for time-varying confounders would have aggra-vated the potential for bias due toconfounding by indication. Emulatingthe design and ITT analysis of a clinicaltrial has better performance in areasmore susceptible to confounding byindication (5).

ConclusionIn summary, we found no evidence thatmetformin use was associated withcancer risk compared with sulfonylureause in a cohort of newly diagnosed di-abetic patients designed and analyzed

emulating a clinical trial. Several largetrials are currently planned or are inprogress testing metformin for the sec-ondary prevention of several cancers(e.g., NCT01864096 plans to enroll408 men with low-risk localized pros-tate cancer, and NCT01101438 enrolled3,649 women with early breast cancer)after the evidence from the mechanis-tic studies and the results from someearly epidemiologic studies suggestedan inhibitory effect of metformin oncancer risk or progression. However,several of the latter observational stud-ies were potentially biased, and morerigorous study designs and analyses areneeded in future observational studiesthat could assist in avoiding the futureinitiation of potentially unnecessarytrials.

Funding. This work was supported by theSeventh Framework Programme of the Euro-pean Union (PIEF-GA-2010-276017 to K.K.T.and J.P.A.I.).Duality of Interest. No potential conflicts ofinterest relevant to this article were reported.Author Contributions. K.K.T. formulated thehypothesis, designed the study, supervised andconducted the statistical analysis, andwrote themanuscript. D.C. carried out the data manage-ment and statistical analysis. N.E.A. participatedin formulating the hypothesis and reviewed andedited the manuscript. E.C.R. provided valuableideas for sensitivity and subgroup analysesand reviewed and edited the manuscript. D.S.L.,K.v.V., and C.S. reviewed and edited the manu-script. D.A. and P.V. contributed to the discussionof the research hypothesis and reviewed andedited the manuscript. I.T. formulated the hy-pothesis, designed the study, and reviewed andedited the manuscript. J.P.A.I. wrote the manu-script and supervised the research. All authorsapproved the final version submitted. K.K.T. is theguarantor of this work and, as such, had fullaccess to all the data in the study and takesresponsibility for the integrity of the data and theaccuracy of the data analysis.Prior Presentation. An abstract of this studywas presented as a poster at the 105th AnnualMeeting of the American Association for CancerResearch, San Diego, CA, 5–9 April 2014.

References1. Zhang P, Li H, Tan X, Chen L, Wang S. Associ-ation of metformin use with cancer incidenceand mortality: a meta-analysis. Cancer Epide-miol 2013;37:207–2182. Thakkar B, Aronis KN, Vamvini MT, Shields K,Mantzoros CS. Metformin and sulfonylureas inrelation to cancer risk in type II diabetes pa-tients: a meta-analysis using primary data ofpublished studies. Metabolism 2013;62:922–9343. Stevens RJ, Ali R, Bankhead CR, et al. Canceroutcomes and all-cause mortality in adults

allocated to metformin: systematic review andcollaborative meta-analysis of randomised clin-ical trials. Diabetologia 2012;55:2593–26034. Suissa S, Azoulay L. Metformin and the risk ofcancer: time-related biases in observationalstudies. Diabetes Care 2012;35:2665–26735. Hernan MA, Alonso A, Logan R, et al. Obser-vational studies analyzed like randomized ex-periments: an application to postmenopausalhormone therapy and coronary heart disease.Epidemiology 2008;19:766–7796. Walley T,Mantgani A. TheUKGeneral PracticeResearch Database. Lancet 1997;350:1097–10997. Garcıa Rodrıguez LA, Perez Gutthann S. Useof the UK General Practice Research Databasefor pharmacoepidemiology. Br J Clin Pharmacol1998;45:419–4258. Jick SS, Kaye JA, Vasilakis-Scaramozza C, et al.Validity of the general practice research data-base. Pharmacotherapy 2003;23:686–6899. Herrett E, Thomas SL, Schoonen WM,Smeeth L, Hall AJ. Validation and validity of di-agnoses in the General Practice Research Data-base: a systematic review. Br J Clin Pharmacol2010;69:4–1410. Cooper R, Mishra G, Clennell S, Guralnik J,Kuh D. Menopausal status and physical perfor-mance in midlife: findings from a British birthcohort study. Menopause 2008;15:1079–108511. Shaw RJ, Lamia KA, Vasquez D, et al. Thekinase LKB1 mediates glucose homeostasis inliver and therapeutic effects of metformin. Sci-ence 2005;310:1642–164612. Buzzai M, Jones RG, Amaravadi RK, et al. Sys-temic treatment with the antidiabetic drug met-formin selectively impairs p53-deficient tumorcell growth. Cancer Res 2007;67:6745–675213. Zakikhani M, Dowling R, Fantus IG,Sonenberg N, Pollak M. Metformin is an AMPkinase-dependent growth inhibitor for breastcancer cells. Cancer Res 2006;66:10269–1027314. Bodmer M, Meier C, Krahenbuhl S, Jick SS,Meier CR. Long-term metformin use is associ-ated with decreased risk of breast cancer. Di-abetes Care 2010;33:1304–130815. Bowker SL, Majumdar SR, Veugelers P,Johnson JA. Increased cancer-related mortalityfor patients with type 2 diabetes who use sulfony-lureas or insulin. Diabetes Care 2006;29:254–25816. Currie CJ, Poole CD, Gale EA. The influence ofglucose-lowering therapies on cancer risk in type2 diabetes. Diabetologia 2009;52:1766–177717. Evans JM, Donnelly LA, Emslie-Smith AM,Alessi DR, Morris AD. Metformin and reducedrisk of cancer in diabetic patients. BMJ 2005;330:1304–130518. Hall GC, Roberts CM, Boulis M, Mo J,MacRae KD. Diabetes and the risk of lung can-cer. Diabetes Care 2005;28:590–59419. Landman GW, Kleefstra N, van Hateren KJ,Groenier KH, Gans RO, Bilo HJ. Metformin associ-ated with lower cancer mortality in type 2 diabe-tes: ZODIAC-16. Diabetes Care 2010;33:322–32620. Li D, Yeung SC, Hassan MM, Konopleva M,Abbruzzese JL. Antidiabetic therapies affect riskof pancreatic cancer. Gastroenterology 2009;137:482–48821. Libby G, Donnelly LA, Donnan PT, Alessi DR,Morris AD, Evans JM. New users of metforminare at low risk of incident cancer: a cohort studyamong people with type 2 diabetes. DiabetesCare 2009;32:1620–1625

care.diabetesjournals.org Tsilidis and Associates 2531

22. Ferrara A, Lewis JD, Quesenberry CP Jr,et al. Cohort study of pioglitazone and cancerincidence in patients with diabetes. DiabetesCare 2011;34:923–92923. Qiu H, Rhoads GG, Berlin JA, Marcella SW,Demissie K. Initial metformin or sulphonylureaexposure and cancer occurrence among pa-tients with type 2 diabetes mellitus. DiabetesObes Metab 2013;15:349–35724. van Staa TP, Patel D, Gallagher AM,de Bruin ML. Glucose-lowering agents and thepatterns of risk for cancer: a study with theGeneral Practice Research Database and sec-ondary care data. Diabetologia 2012;55:654–665

25. Smiechowski B, Azoulay L, Yin H, Pollak MN,Suissa S. The use of metformin and colorectalcancer incidence in patients with type II diabe-tes mellitus. Cancer Epidemiol Biomarkers Prev2013;22:1877–188326. Smiechowski BB, Azoulay L, Yin H, PollakMN, Suissa S. The use of metformin and the in-cidence of lung cancer in patients with type 2diabetes. Diabetes Care 2013;36:124–12927. Redaniel MT, Jeffreys M, May MT, Ben-Shlomo Y, Martin RM. Associations of type 2diabetes and diabetes treatment with breastcancer risk and mortality: a population-basedcohort study among British women. CancerCauses Control 2012;23:1785–1795

28. Azoulay L, Dell’Aniello S, Gagnon B, PollakM, Suissa S. Metformin and the incidence ofprostate cancer in patients with type 2 diabetes.Cancer Epidemiol Biomarkers Prev 2011;20:337–34429. Kahn SE, Haffner SM, Heise MA, et al.;ADOPT Study Group. Glycemic durability of ro-siglitazone, metformin, or glyburide monother-apy. N Engl J Med 2006;355:2427–244330. Home PD, Pocock SJ, Beck-Nielsen H, et al.;RECORD Study Team. Rosiglitazone evaluatedfor cardiovascular outcomes in oral agent com-bination therapy for type 2 diabetes (RECORD):a multicentre, randomised, open-label trial.Lancet 2009;373:2125–2135

2532 Antidiabetes Drugs and Risk of Cancer in CPRD Diabetes Care Volume 37, September 2014