Statistical adjustment for a measure of healthy lifestyle doesnt yield the truth about hormone therapy

IMS Collectio ns Probability and St atistics: Essays i n Honor o f David A. F reedman V ol. 2 (2008) 142–152 c  Institute of Mathematical Statistics , 2008 DOI: 10.1214/ 19394030 70000004 37 Statistic al adjustmen t for a measure of health y lifest yle do es n’t yield th e truth ab o ut h ormone therap y Diana B. P eti tti ∗ 1 and W ansu Chen ∗ 2 University of Southern California and Kaiser Permanente Southern California Abstract: The W omen’s Health Initiativ e randomized clinical trial of hor- mone therapy found no b enefit of hormones in pr ev en tiv e cardiov ascular dis - ease, a finding in striking co n trast with a large bo dy of observ ational re- searc h. Understanding whether better metho dology and/or statistical adjust- men t might ha v e prev en ted the erroneous conclusions of observ ational researc h is imp ortant. This i s a re-analysis of data from a case-con trol study examin- ing the relationship of p ostmenopausal horm one therapy and the risks of my- ocardial inf arction (MI) and ische mic s trok e in which we rep or ted no ov erall increase or decrease in the ri sk of either ev en t. V ariables measuring health be- ha vior/lifestyle tha t are not like ly to b e causally with the risks of MI and stroke (e.g., s unscreen use) were included in multiv ariate analysis along with tradi- tional confounders (age, hypertension, diabetes, smoking, b o dy mass index, ethnicit y , educat ion, prior coronary heart disease for MI and prior strok e/TIA for stroke) to determine whether adjustment for the health b eha vior/lifestyle v ari ables could repro duce or br i ng the r esul ts closer to the findings in a large and definitive r andomized clini cal trial of hormone therap y , the W omen’s Health Initiative. F or b oth MI and stroke, measures of health behavior/lifest yle were asso ci- ated with odds r atios (ORs) less than 1.0. Adjustmen t for traditional cardio- v ascular disease confounders did not alter the m agnitude of the OR s for MI or strok e. Addition of a subset of these v ariables selected usi ng step wise r egres- sion to the final MI or stroke models along with the traditional cardiov ascular disease confounders mov ed the ORs for estrogen and estrogen/progest in use closer to v alues observ ed in the W omen Health Initiative clinical trial, but did not r eliably repro duce the clinical trial results for these t w o endp oint s. 1. Bac kground The W omen’s Health Initia tive (WHI) clinica l tria l of hormone thera p y is a lar ge randomized trial whose primary aim was determining whether p os t-menopausal hormone ther apy preven ts corona ry heart dis e ase (Rossouw et al. [ 10 ] a nd Ander - son et a l. [ 1 ]). The study b egan in the early 19 90’s and published main results in 2002– 2004 . It inv olved recruitment a nd r andomization of more than 1 8,000 pos t- menopausal women to ho r mones or placeb o. The WHI found no overall effect, or ∗ Supported by gran t num b er R 01-H L-47043 from the National Heart Lung and Blo o d Institute. 1 Unive rsity of Southern California, Department of Preven tiv e Medicine, Keck Sch ool of Medicine, 299 E. Laurel Aven ue, Sierra M adre, CA 91024, USA, e-mail: dbpetitt i@verizo n.net 2 Kaiser Permanen te Southern Cali f ornia, Departmen t of Researc h and Ev aluation, 100 S. Los Robles, P asadena, CA 91101, USA, e- m ail: Wansu.Ch en@kp.or g AMS 2000 subje ct classific ations: Pri mary 92C60; secondary 62P10, 00B30. Keywor ds and phr ases: cerebrov ascular disorders, cerebral infarction, strok e, coronary , epi- demiological metho ds, estrogen, heart disease, hor m one replacement , my o cardial infarction. 142 Statistic al adjustment for a me asur e of he alth y lifesty le 143 per haps an increase, in the r isk of m y o cardial infar ction (MI) in women a ssigned to combined estrogen/ progestin (E/ P) therapy a nd no effect o f estrog en a lone (E) (Rossouw et al. [ 10 ] and Anderson et al. [ 1 ]). Prio r observ ational resea rch con- cluded that the risk o f corona r y hear t disease was r educed by half (Stampfer and Colditz [ 12 ]). Moreov er, the risk of str oke was increased for bo th E a nd E /P in WHI (Rossouw et a l. [ 10 ] and Anderson et al. [ 1 ]). Prio r observ a tional r esearch found no effect of hor mone therapy on s troke (Psa t y et al. [ 9 ]). Publications b y P rentice et al. sugg est s ta tistical approaches that would hav e ov ercome the discr epancy b etw e e n the o bserv ational resea r ch and the clinical tr ial (Prentice et al. [ 7 ] and Prentice et a l. [ 8 ]). Reviews of these a ppr oaches are mixed (Petitt i and F reedman [ 5 ], F reedman and Petitti [ 3 ], DeMets [ 2 ] a nd Greenland [ 4 ]). Understanding whether b etter metho dolog y and/o r statistical adjustment migh t hav e pre vented the erroneous conclusio ns made based on the obse rv atio nal r esearch is imp ortant. W e pr eviously published the re s ults of a case- control study in which we es ti- mated the relative r isks of m yocardia l infarction (MI) and str oke in current use r s of E a nd E/P that adjusted for tra ditional c a rdiov ascular disease co nfounders—age , ethnicit y , education as a measur e o f socio econo mic status, and factors known to increase the risk of MI or stro ke causally (smoking, diab etes, h yp ertension, b o dy mass index), using logistic regr ession (Petitti et al. [ 6 ] and Sidney et al. [ 11 ]). The study neither confirmed nor ruled o ut a lo w er risk of MI or stroke in hor mone users. Adjusted estimates of the relative ris k of MI and stroke were b oth, howev er, less than 1.0 in cur rent E and EP user s compar ed with never hor mo ne users , a finding inconsistent with WHI. This is a re-ana lysis of data fro m our case-control study . In the re-analy s is, v ari- ables that ass ess health b ehaviors/lifestyle were added to the multiv ariate analy s is along with the traditional c onfounders to determine whether adjustmen t for these v aria bles, as a ma rker fo r healthy lifest yle, could re pr o duce or bring the results closer to the findings in the W omen’s Health Initiative for stro ke and MI. Some of the health behavior/ lifes t yle v aria bles w ere chosen sp ecifically because they w ere NOT likely to b e ca usally rela ted to MI or s troke and whose relationship with thes e conditions would b e exp e cted to b e non-ca usal. 2. Metho ds 2.1. Overview Study metho ds a nd r esults examining the a sso ciation of hormone therapy with the risk o f str oke and MI after adjustment for traditiona l confounders a re desc rib ed in detail in tw o prior publications (Petitti et al. [ 6 ] and Sidney et a l. [ 11 ]). Briefly , an attempt was made to identify all fatal and non-fatal strokes and MIs in women age 45-74 y ears in 10 medical cent ers o f the Ka iser P ermanente Medical Care Pr ogram, Northern Ca lifornia r egion, during the p er io d, Nov em ber, 1 991 – Nov em ber, 19 9 4. A standa rd proto col was used to define s troke a nd sub classify it by t ype. Our re-analy sis includes only strokes classified as ischemic. F or each case, an attempt was made to interview one control, matched on year of birth a nd facility of us ual car e. Out of 550 ischemic stroke ca ses and 685 MI cases, we were able to obtain 34 9 stro ke cas e/controls s ets and 438 MI case/control sets for analysis, a fter ex clusions describ ed in the Analysis section. 144 D. B. Petitt i and W. Chen 2.2. Information Eligible c ases and controls were in terviewed in-p ers on by tra ined interview ers using a standar dized instrument. In terview questio ns were a sked rela tive to an index date, which was the date of symptom onset for cases and the same date for her matched control. If a c ase ha d died or was unable to co mm unicate verbally , an attempt was made to interview a pr oxy , but pr oxy resp onses are excluded from this analysis . Hyper tension was defined as a “yes” to a question ab out use o f medication for high blo o d pressure . Diabetes w a s defined as a “yes” to a question asking a bo ut use o f insulin o r pills for diab etes. W omen who stated that they had b een told by a physician that they had a high cholesterol level w ere classified as having hyp er- cholesterolemia (high c ho lesterol level was not used as a cov aria te in the original analysis). A study sub ject was defined as a nons moker if she answered “no” to the question, “ Hav e you ever smoked cigar ettes?” If she ans wered “yes” to this ques- tion, s he was catego r ized as a curr ent r egular smo ker on the bas is of her answer to the que s tion, “On [index date] were you still smoking r egularly?” (Regular ly means at least 5 c igarettes per week, almost ev ery week). Bo dy mass index was calculated from self-r epo rted height and weigh t. The questions asked ab out health b ehavior/lifestyle spanned a range of activities and b ehaviors that p eople be lieve may improve their health. W e soug ht esp ecially to identify questions ab out activities and b ehaviors that were NOT related causa lly to cardiov ascular disease r isk. W omen were asked, “Do you do the following to try to impro ve your health?” This was follow ed b y the trained in terviewer rea ding each behavior with a q ue r y for a “yes” or “ no” r esp onse. 2.3. Analysis W e defined cur r ent hormone use as “yes” for hysterectomized women who used E within 1 mo n th of the index date; current hormone use was defined as “yes” for non-hysterectomized women who us ed E/ P within 1 mon th of the index date; and “no” otherwise. Excluded fro m the a na lysis were pr e-menopausa l women and hysterectomized women who used E P , non- hysterectomized women who used E, and users o f pr ogestin o nly within 1 month of the index. All ex clusions were a pplied in the sa me way to cas es and controls. These exclus ions and definitions are the same as in our prio r published analyses. The o dds ratio (O R) was used to estimate the relative risk of MI a nd strok e. The multiv ariate ana lysis used conditional log istic regre ssion ana lysis. Ninety-fiv e per cent confidence interv als (CI) were calculated for OR estimates. W e fir st calculated age- a djusted ORs for MI and s troke in r elation to the tradi- tional v ascula r disease r is k fa c tors a nd in relatio n to use o f E and E/ P separ ately in h y sterectomized a nd non-hysterectomized women. W e then ex a mined ORs for stroke and MI in relation to each of the lifestyle/behavior questions adjusting first for age only and then for a ge and the traditional cardiov ascula r dis ease co nfounders, also separ ately in women with and without a hysterectomy . Last, we us ed stepwise logistic regr ession to scre en the list o f b ehavior/lifestyle v ariables a nd select only those meeting a significance lev el of 0.2 o r low er for both entering and staying in the mo dels while fo r cing all the traditional v ascular disea s e r isk factors to stay in the model. Thus, only those b ehavior/lifes tyle v ariables deemed imp orta nt sta tisti- cally to the outcomes are included (separ ate ana lyses for women with and without a hysterectomy). Statistic al adjustment for a me asur e of he alth y lifesty le 145 3. Results Although we attempted to repr o duce the s tudy s amples using the same exclusio n criteria describ ed in the tw o publica tions, we were unable to gather six stroke case/co n trol sets and four MI case/control sets . T ables 1 and 2 show the charac- teristics of no n-hysterectomized and hysterectomized ca ses and c ontrols and age- adjusted ORs and 95% C.I.s for MI and stro ke in relation to the traditional c on- founders a nd current use of E /P (for non-hysterectomized women) and E only (for hysterectomized women). The tra ditional c ardiov ascular disea se risk factors s how exp ected asso ciations with the risk of MI and stroke. T a ble s 3 and 4 show ORs a nd 95% C.I.s for str oke and MI in relation to each o f the v ariables that meas ur e health b ehavior in no n- hysterectomized and h ysterec- tomized women adjusting for age and then for a ge a nd the traditional confounders. F or almost all of the he a lth be havior questio ns , age-a djusted ORs for b o th stroke and MI in w omen who resp o nded “yes” to the question ar e low er than 1.0 in both non-hysterectomized and h ysterectomized women. Adjustment for the traditiona l cardiov ascular dis ease risk factors in addition to age do es not change the mag nitude of the OR for a ny v ariable by muc h, although the C.I.s often include 1.0 . The fully a djusted ORs for MI in no n-hysterectomized women who stated they regular ly used sun block or sunscreen (adjusted O R 0.3; 95% C.I. 0.2–0.5) o r who stated they w ere trying to cut down on alcohol use (adjusted OR 0.4 ; 95% C.I. 0.2–0.7 ) are o f particula r note. The fully adjusted OR for stroke in hysterectomized women who stated they regularly used sun blo ck o r sunscr een (OR 0.4; 95 % C.I. 0.2–0.9 ) is also noteworthy . The O Rs in women who a nswered yes to the behavior q uestions a re further from 1.0 for MI than for str oke in b oth non-h ysterectomized and h ysterectomized women (T ables 3–4). The adjustmen t for the traditional co nfounders changes the estimates less fo r MI than for str o ke in b oth non-hysterectomized and h ysterectomized women. T a ble 5 shows the res ults o f mo dels as s essing the ORs fo r stroke and MI in relation to current us e of E and E /P after adjustmen t. The ORs for str oke r e po rted in the pre vious publication were based on the sample that includes women with hysterectom y who us ed E/P and w o men without h ysterectomy who used E and therefore are not exactly the sa me as the r esults in this re-ana lysis. The ORs were estimated for current use o f E and E/P after adjusting for age only , for ag e and the tra ditional confounders, and then for a g e and the tra ditional confounders plus the b ehavior/lifestyle v a riables that were selected using the step wise r e gressio n pro cedure. Estimates for the r isk o f co ronar y heart dise a se a nd stroke fro m WHI in non-hysterectomized users of E/P and hysterecomized user s of E are s hown for co mparison. Adjustment for the b ehavior/lifestyle v ar iables in addition to the traditional confounders results in further changes in the ORs for curre nt use of E and E/P . Howev er, neither adjustment for the traditional confounder s no r adjustment for the b ehavior/lifestyle v aria bles in addition to the tra ditional confounder s relia bly repro duces the WHI results considering b oth disease endp oints and b oth E and E/P . 4. Discussion F or bo th MI and stroke, the v ariables that measured hea lthy behavior/lifes tyle were asso ciated w ith ORs less than 1 .0 even after adjustmen t for age. Even after adjustment fo r the traditional car diov as cular disea s e confounders in addition to age, the ORs for the v aria bles that mea sured healthy b e havior/ lifest yle remaine d mo stly 146 D. B. Petitt i and W. Chen T able 1 Women without hyster e ctomy: for demo gr aphic and other char act e ristics, p er c ent of ca ses and c ontr ols and age adjuste d o dds r atios for myo c ar dial infar ction and str oke for e ach char acteristi c Myocardial infarctio n Stroke Cases Controls Cases Controls Characteri stic ( N = 18 9) ( N = 199) OR ( 95% CI) ( N = 156) ( N = 153) OR (95% CI) Current E/P , % 20 . 1 29 . 7 0 . 6 (0 . 4–1 . 01) ∗ 18 24 . 2 0 . 7 (0 . 4–1 . 2) ∗ T reated for h ypertension, % 41 . 3 30 1 . 6 (1 . 1–2 . 5) 49 . 7 27 . 5 2 . 6 ( 1 . 6–4 . 2) T reated for diabetes, % 22 . 3 8 . 1 3 . 3 (1 . 8–6 . 1) 27 . 1 11 . 8 2 . 8 (1 . 5–5 . 1) Body Mass Index Quartile 1 (lo w est) 24 . 9 30 . 3 1 . 0 (ref) 23 . 8 30 . 5 1 . 0 ( ref) Quartile 2 17 . 8 28 . 8 0 . 7 (0 . 4–1 . 3) 22 . 5 24 . 5 1 . 2 ( 0 . 6–2 . 3) Quartile 3 25 . 4 17 . 2 1 . 8 (0 . 98–3 . 2) 29 . 2 23 . 2 1 . 6 ( 0 . 9–3 . 1) Quartile 4 31 . 9 23 . 7 1 . 6 (0 . 96–2 . 8) 24 . 5 21 . 8 1 . 4 ( 0 . 8–2 . 7) Cigarette smoking % Nev er 45 50 . 2 1 . 0 (ref) 35 . 3 53 . 6 1 . 0 (ref) Pa st 21 . 7 33 . 7 0 . 7 (0 . 5–1 . 2) 30 . 1 32 1 . 5 ( 0 . 9–2 . 5) Occasional/Curren t 33 . 3 16 . 1 2 . 5 (1 . 5–4 . 3) 34 . 6 14 . 4 3 . 8 ( 2 . 1–6 . 9) Race/Ethnicit y % White, non-Hispanic 76 . 8 75 . 7 1 . 0 (ref) 66 77 . 8 1 . 0 (ref) Hispanic 8 . 9 7 . 1 1 . 2 (0 . 6–2 . 6) 11 . 1 8 . 5 1 . 6 (0 . 7–3 . 4) African-Ameri can 5 . 4 7 . 6 0 . 7 (0 . 3–1 . 7) 12 . 4 4 . 6 3 . 3 (1 . 3–8 . 2) Asian 5 . 4 7 . 6 0 . 7 (0 . 3–1 . 7) 8 . 5 6 . 5 1 . 6 (0 . 7–3 . 9) Other or unkno wn 3 . 8 2 1 . 9 (0 . 5–6 . 6) 2 2 . 6 0 . 9 ( 0 . 2–4 . 0) Lev el of Education % Less than high school 17 . 9 11 . 7 1 . 0 (ref) 22 . 2 12 . 4 1 . 0 ( ref) High school graduate 35 . 9 33 0 . 7 (0 . 4–1 . 3) 33 . 3 31 . 4 0 . 6 ( 0 . 3–1 . 2) Some college or business or technical training 33 . 2 35 . 5 0 . 6 (0 . 3–1 . 1) 37 . 3 34 0 . 6 ( 0 . 3–1 . 2) College graduate 1 3 19 . 8 0 . 4 (0 . 2–0 . 9) 7 . 2 22 . 2 0 . 2 (0 . 1–0 . 4) History of CHD % 5 . 3 3 . 5 1 . 5 (0 . 5–4 . 0) – – Prior Stroke / TIA % – – 5 . 1 0 . 7 8 . 3 (1 . 03–67 . 0) * Refer ence group “Never” Statistic al adjustment for a me asur e of he alth y lifesty le 147 T able 2 Women with hyster e ct omy: for demo gr aphic and other char acteristics, p er c ent of c ases and c ontr ols and age adjuste d o dds r ati os for myo c ar dial infar ction and str oke for e ach char acteristi c Myocardial infarctio n Stroke Cases Controls Cases Controls Characteri stic ( N = 12 5) ( N = 122) OR (95% CI) ( N = 9 0) ( N = 85) OR (9 5% CI) Current E, % 68 74 . 6 0 . 7 (0 . 4–1 . 2) ∗ 68 . 9 74 . 1 0 . 7 (0 . 4–1 . 4) T reated for h ypertension, % 47 . 2 34 . 4 1 . 7 (1 . 04–2 . 9) 46 . 7 27 . 1 2 . 6 (1 . 4–5 . 0) T reated for diabetes, % 25 . 6 9 . 8 3 . 2 (1 . 5–6 . 5) 28 . 1 3 . 5 10 . 5 ( 3 . 0–36 . 3) Body Mass Index Quartile 1 (lo w est) 22 . 6 24 . 8 1 . 0 (ref) 20 . 6 28 . 3 1 . 0 ( ref) Quartile 2 25 32 . 2 0 . 9 (0 . 4–1 . 7) 30 . 7 23 . 5 1 . 8 (0 . 8–4 . 1) Quartile 3 29 23 . 2 1 . 4 (0 . 7–2 . 8) 23 . 9 18 . 8 1 . 8 (0 . 7–4 . 4) Quartile 4 23 . 4 19 . 8 1 . 3 (0 . 6–2 . 7) 25 29 . 4 1 . 2 (0 . 5–2 . 7) Cigarette smoking % Nev er 37 . 6 50 1 . 0 (ref) 41 . 1 47 1 . 0 (ref) Pa st 30 . 4 33 . 6 1 . 2 (0 . 7–2 . 2) 31 . 1 41 . 2 0 . 9 ( 0 . 4–1 . 7) Occasional/Curren t 32 16 . 4 2 . 8 (1 . 4–5 . 5) 27 . 8 11 . 8 2 . 6 (1 . 1–6 . 2) Race/Ethnicit y % White, non-Hispanic 87 . 9 82 1 . 0 ( r ef) 73 . 9 84 . 7 1 . 0 (ref) Hispanic 4 . 9 8 . 2 0 . 6 (0 . 2–1 . 6) 11 . 4 4 . 7 2 . 7 (0 . 8–9 . 1) African-Ameri can 3 . 2 4 . 1 0 . 7 (0 . 2–2 . 8) 5 . 7 3 . 5 1 . 8 (0 . 4–8 . 0) Asian 3 . 2 4 . 9 0 . 6 (0 . 2–2 . 2) 3 . 4 3 . 5 1 . 1 (0 . 2–5 . 5) Other or unkno wn 0 . 8 0 . 8 0 . 9 (0 . 1–14 . 6) 5 . 7 3 . 5 1 . 8 (0 . 4–7 . 8) Lev el of Education % Less than high school 21 14 . 8 1 . 0 (ref) 26 . 1 5 . 9 1 . 0 (ref) High school graduate 37 . 9 29 . 5 0 . 9 (0 . 4–1 . 9) 30 . 7 28 . 2 0 . 2 ( 0 . 1–0 . 7) Some college or business or technical training 29 . 8 33 . 6 0 . 6 (0 . 3–1 . 3) 31 . 8 43 . 5 0 . 2 ( 0 . 1–0 . 5) College graduate 1 1 . 3 22 . 1 0 . 4 (0 . 1–0 . 9) 11 . 4 22 . 4 0 . 1 ( 0 . 03–0 . 4) History of CHD % 6 . 4 4 . 9 1 . 3 (0 . 4–3 . 9) – – Prior Stroke / TIA % – – 1 . 1 1 . 2 1 . 1 (0 . 1 − 19 . 0) * Refer ence group “Never”. 148 D. B. Petitt i and W. Chen T able 3 Women without hyster e ctomy: p erc ent of ca ses and c ontr ols who r ep orte d the b ehavior and adjuste d o dds r atios for myo ca r dial infar ct ion and str oke in women who r ep orte d the beha vior Myocardial infarctio n Stroke Adjusted for Adjusted for traditi onal Age traditi onal Cases Controls Age a djusted confounders 1 Cases Controls adjusted confounders 2 Health behaviors ( N = 189) ( N = 199) OR ( 95% C I) OR (95% CI) ( N = 156) ( N = 1 53) (95% CI) OR ( 95% CI) Exercise 54 . 6 69 . 2 0 . 5 (0 . 3–0 . 8) 0 . 6 (0 . 4–0 . 99) 53 . 6 69 . 9 0 . 5 (0 . 3–0 . 8) 0 . 6 (0 . 3–1 . 0) T ry to eat more foo ds con taining fib er 75 85 . 4 0 . 5 (0 . 3–0 . 9) 0 . 6 (0 . 3–0 . 97) 73 . 7 86 . 3 0 . 4 (0 . 2–0 . 8) 0 . 6 (0 . 3–1 . 1) T ry to eat fo o ds lo w in f at 68 . 1 84 . 3 0 . 4 (0 . 2–0 . 7) 0 . 4 ( 0 . 2–0 . 7) 76 . 5 88 . 2 0 . 4 (0 . 2–0 . 8) 0 . 5 (0 . 3–1 . 1) Eat more oliv e oil 26 . 6 34 0 . 7 (0 . 5–1 . 1) 0 . 7 (0 . 5–1 . 2) 29 . 4 37 . 9 0 . 7 ( 0 . 4–1 . 1) 0 . 8 (0 . 5–1 . 5) Use sunblock or sunscreen 26 50 . 8 0 . 3 (0 . 2–0 . 5) 0 . 3 (0 . 2–0 . 5) 32 44 . 4 0 . 6 (0 . 4–0 . 9) 0 . 8 (0 . 4– 1 . 3) Cut down on alcohol consumption 11 . 9 25 . 3 0 . 4 (0 . 2–0 . 7) 0 . 4 ( 0 . 2–0 . 7) 20 . 9 22 . 9 0 . 9 (0 . 5–1 . 6) 0 . 7 (0 . 4–1 . 3) Cut down on caffeine 33 34 . 3 0 . 9 (0 . 6–1 . 4) 0 . 9 (0 . 6–1 . 5) 29 . 6 34 0 . 8 (0 . 5–1 . 3) 0 . 8 (0 . 4–1 . 3) T ak e vitamin supplemen ts 47 . 6 55 . 6 0 . 7 (0 . 5–1 . 1) 0 . 8 ( 0 . 5–1 . 3) 48 . 7 60 . 8 0 . 6 (0 . 4–0 . 96) 0 . 8 (0 . 4–1 . 3) Meditate or use other 22 . 2 38 . 4 0 . 5 (0 . 3–0 . 7) 0 . 5 ( 0 . 3–0 . 9) 22 . 2 35 . 3 0 . 5 (0 . 3–0 . 9) 0 . 6 (0 . 3–1 . 03) tec hnique to reduce stress Cut down on red meat 56 . 2 72 . 7 0 . 5 (0 . 3–0 . 7) 0 . 5 ( 0 . 3–0 . 8) 62 . 1 71 . 9 0 . 6 (0 . 4–1 . 03) 0 . 6 (0 . 3–0 . 97) T ak e calcium supplemen t 32 . 6 42 . 9 0 . 6 (0 . 4–0 . 98) 0 . 9 (0 . 5–1 . 4) 30 . 5 51 . 3 0 . 4 (0 . 3–0 . 7) 0 . 5 (0 . 3–0 . 9) T ak e fish oil supplement 4 . 3 8 . 6 0 . 5 (0 . 2–1 . 2) 0 . 4 (0 . 2–1 . 1) 5 . 2 6 . 5 0 . 8 (0 . 3–2 . 1) 0 . 7 (0 . 2–2 . 1) Any other thing to try to sta y health y 23 . 2 24 . 2 0 . 9 (0 . 6–1 . 5) 1 . 1 ( 0 . 7–1 . 8) 21 . 6 23 . 5 0 . 9 (0 . 5–1 . 5) 0 . 9 (0 . 5–1 . 7) 1 age, hypertension, diab etes, b o dy mass index (quartiles), smoking, race and ethnicit y , level of educat ion and history of coronary heart disease. 2 age, hypertension, diab etes, b o dy mass index (quartiles), smoking, race and ethnicit y , level of educat ion and history of stroke/TIA. Statistic al adjustment for a me asur e of he alth y lifesty le 149 T able 4 Women with hyster e ct omy: p er c ent of c ases and c ontr ols who r ep orte d the b ehavior and adjuste d o dds r atios for myo c ar dial infar ction and stro ke in women who r ep orte d the beha vior Myocardial infarctio n Stroke Adjusted for Adjusted for traditi onal Age traditi onal Cases Con trols Age a djusted confounders 1 Cases Con trols adjusted confounders 2 Health behaviors ( N = 125) ( N = 122) OR (95% CI) OR ( 95% CI) ( N = 90) ( N = 85) (95% CI) OR ( 95% C I) Exercise 54 . 5 69 . 7 0 . 5 (0 . 3–0 . 9) 0 . 6 (0 . 3– 1 . 06) 62 . 1 78 . 8 0 . 4 (0 . 2–0 . 9) 0 . 5 (0 . 2–1 . 2) T ry to eat more foo ds con taining fib er 81 . 5 85 . 3 0 . 8 ( 0 . 4–1 . 5) 0 . 8 (0 . 4–1 . 7) 80 . 7 89 . 4 0 . 5 (0 . 2–1 . 2) 0 . 5 (0 . 2–1 . 6) T ry to eat fo o ds lo w in f at 82 . 3 86 . 1 0 . 8 (0 . 4–1 . 5) 0 . 9 (0 . 4–1 . 8) 83 89 . 4 0 . 6 (0 . 2–1 . 5) 0 . 5 (0 . 2–1 . 5) Eat more oliv e oil 29 32 . 5 0 . 8 (0 . 5–1 . 5) 1 . 0 (0 . 5– 1 . 9) 30 4 3 . 5 0 . 6 (0 . 3–1 . 04) 0 . 8 (0 . 4–1 . 8) Use sunblock or sunscreen 37 . 1 48 . 4 0 . 6 (0 . 4–1 . 04) 0 . 8 (0 . 4–1 . 4) 27 . 3 57 . 7 0 . 3 (0 . 1–0 . 5) 0 . 4 (0 . 2–0 . 9) Cut down on alcohol consumption 22 . 6 30 . 3 0 . 7 ( 0 . 4–1 . 2) 0 . 6 (0 . 3–1 . 2) 18 . 2 25 . 9 0 . 6 (0 . 3–1 . 3) 0 . 7 (0 . 3–1 . 7) Cut down on caffeine 32 45 . 1 0 . 6 (0 . 3–0 . 96) 0 . 6 (0 . 3–1 . 00) 31 . 8 37 . 7 0 . 8 (0 . 4–1 . 5) 0 . 9 (0 . 4–2 . 0) T ak e vitamin supplemen ts 50 . 8 61 . 5 0 . 6 (0 . 4–1 . 08) 0 . 6 (0 . 4–1 . 08) 54 . 6 64 . 7 0 . 7 (0 . 4–1 . 2) 0 . 7 (0 . 3–1 . 5) Meditate or use other 28 . 2 41 . 8 0 . 5 (0 . 3–0 . 9) 0 . 5 (0 . 3–0 . 9) 29 . 9 47 . 1 0 . 5 (0 . 3–0 . 9) 0 . 5 (0 . 2–1 . 04) tec hnique to reduce stress Cut down on red meat 63 . 7 74 . 6 0 . 6 (0 . 3–1 . 04) 0 . 5 (0 . 3–0 . 97) 67 . 1 78 . 8 0 . 6 (0 . 3–1 . 1) 0 . 5 (0 . 2–1 . 1) T ak e calcium supplemen t 39 . 5 51 . 6 0 . 6 (0 . 4–1 . 02) 0 . 8 (0 . 5–1 . 4) 44 . 3 47 . 1 0 . 9 (0 . 5–1 . 7) 1 . 1 (0 . 5–2 . 2) T ak e fish oil supplement 4 . 8 7 . 4 0 . 6 (0 . 2–1 . 9) 0 . 9 (0 . 3–3 . 1) 6 . 8 7 . 1 1 . 0 (0 . 3–3 . 2) 0 . 9 ( 0 . 2–3 . 7) Any other thing to try to sta y health y 24 . 2 23 . 1 1 . 1 (0 . 6–1 . 9) 1 . 2 (0 . 6–2 . 4) 27 . 6 29 . 4 0 . 9 (0 . 5–1 . 8) 1 . 1 (0 . 5–2 . 6) 1 age, hypertension, diabetes, b ody mass index (quartiles), smoking, race and ethnicit y , l ev el of educat ion and hi story of coronary heart disease. 2 age, hypertension, diabetes, b ody mass index (quartiles), smoking, race and ethnicit y , l ev el of educat ion and hi story of strok e/TIA. 150 D. B. Petitt i and W. Chen T able 5 Odds r ati os for myo c ar dial infar ction and ischemic stro ke in curr ent users of estr o gen alone or estr o gen/pr o gestin after v arious adjustments and c omp arison with women ’s he alth initiative clinic al trial r esults W o men wi thout hysterectomy / W o men wi th hysterectomy / estrogen plus progestin estrogen only Myocardial infarctio n Isc hemic stroke Myocardia l infarctio n Isc hemic str oke Adjustmen t o dds ratio (95% C.I.) o dds ratio (95% C.I.) o dds ratio (95% C.I.) o dds ratio (95% C.I.) Current analysis Age only 0 . 62 (0 . 38–1 . 007) 0 . 71 (0 . 41–1 . 24) 0 . 71 (0 . 41–1 . 25) 0 . 72 (0 . 37–1 . 42) Age and traditional confounders 1 0 . 91 (0 . 52–1 . 58) 0 . 85 (0 . 43–1 . 66) 0 . 96 (0 . 50–1 . 82) 1 . 17 (0 . 48–2 . 90) Age, traditional confounders 1 + health behavior v ariables selected using step wise r egression 1 . 13 (0 . 62–2 . 06) 2 1 . 00 (0 . 50–2 . 01) 3 1 . 04 (0 . 52–2 . 00) 4 1 . 01 ( 0 . 40–2 . 54) 5 W o men’s Heal th Initiat ive clinical tria l 1 . 32 (1 . 02–1 . 72) 6 1 . 41 (1 . 07–1 . 85) 7 0 . 89 (0 . 70–1 . 12) 6 1 . 39 ( 1 . 10–1 . 77) 7 1 Age, hypertension, diab etes, b o dy mass index, smoking, race and ethnicit y , lev el of education, hi story of coronary heart disease for m y ocardial infarction and history of stroke for strok e. 2 T ry to eat more f oo ds l ow in fat; use sunblock or sunscreen; cut down on alcohol consumption; cut do wn on caffeine; meditate or use other tec hnique to reduce stress; cut do wn red meat; take fish oil supplement. 3 Exercise; cut do wn on red m eat; tak e calcium supplement . 4 Cut down on caffeine; take vitamin supplemen ts; meditate or use other technique to r educe str ess; cut down r ed meat. 5 Use s unblock or sunscreen; meditate or use other tec hnique to reduce stress. 6 Non-fatal MI; nominal confide nce int erv al. 7 F atal plus non-fatal; nominal confidence i n terv al. Statistic al adjustment for a me asur e of he alth y lifesty le 151 less than 1.0 . Of par ticular no te is the p ersistence of the ass o ciation of sunscr een use with low er risk of MI and stroke even a fter a djustmen t for confounders. It is highly unlikely tha t sunscreen use prev en ts MI or stroke. W e c hose to query w omen in the s tudy ab o ut this behavior pre c is ely b ecaus e there was no immediately plausible direct causa l pathw ay b etw een sunscreen use and these car diov ascular endp oints. W e hyp o thesized a prio ri that sunscreen us e a nd the other hea lth b ehaviors would be mark ers of a “healthy lifestyle” and that adjustment for a measure of healthy lifest yle would improv e inferences ab out the effect of hormone therapy on MI and stroke. Our findings with regar d to s unscreen use shows that a s tr ategy in which non- causal v aria bles ar e systema tically meas ured in a ppropriate mo dels might improv e inferences deriving fr om o bserv a tional res e arch. Educa tion, so cial s tatus, and in- come are examples of v aria bles tha t epidemiologists frequently consider as con- founders even though their relatio ns hip with disease is se ldom causa l. Rather, the v aria bles “capture” the caus a l asso ciatio ns of facto rs a s so ciated with the exp o sure that get “ mix e d” with the tr ue effect of the exp osure. Our analys is do es not include so me v ariables that are measur a ble (family history of heart disease , untreated h yper tension) and whose inclusio n might further change the estimates. W e did not include interaction terms in the mo dels and this might also hav e br ought the data clo ser to WHI. Wher e to stop when adjusting is genera lly left to the researcher’s judgmen t. When the “tr uth” is known, it may be p ossible to find the “truth.” The pro blems arise when the re s earcher s e e ks the truth through mo deling. In the cur rent ana ly sis, adjustment for the b ehaviors with the stronge s t asso ci- ation with the g iven v ascular endpoint mov ed the O Rs estimates for curr ent E and E/P closer to v a lues observed in the WHI. Of course, w e know that the adjustment mov ed the ORs in the “r ight” dir e ction only b ecause of the clinical trial. Even with the adjustment for healthy behavior/ lifes t yle in addition to a djustmen t for known confounders, we were una ble to repro duce completely the clinical trial results. Ac kno wle dgment s. Thanks to Steve Sidney who w as a co-inv estigator in the main study . References [1] Anderson, G., Limacher, M. , Assaf, A. et a l. (2004). Effects of con- jugated eq uine estr o gen in p ostmenopaus a l women with hysterectomy: The women’s health initiative r andomized controlled trial. JAMA 291 1 701–1 –2. [2] DeMets, D. (2005). Invited commentary on prentice, p ettinger and anderson. Biometrics 61 914– 918. MR22161 80 [3] Freedman, D. and P etitti, D. (20 05). In vited commen tary on pren tice, pettinge r and ander s on. Biometrics 61 918 – 920. MR22161 81 [4] Greenland, S. (2005). Invited co mmen tary on prentice, p ettinger and a n- derson. Biometrics 61 92 0–92 1. MR2216182 [5] Petitti, D. an d Freedman, D. (2005). Invited commen tary: How far can epidemiologists g et with statistica l adjustment? Am. J. Epidemiol. 1 62 415 – 418. [6] Petitti, D., Sidney, S., Quesenberr y, C. and Bernstein, A. (1998). Ischemic stro ke and use of estrog en a nd estrog en/prog estogen as hor mone re- placement therapy . Str oke 29 23–2 8 . 152 D. B. Petitt i and W. Chen [7] Prentice, R., Langer, R., Stef anick, M., How ard, B., Pettinger, M., Anderson, G ., Barad, D. , Curb, J., K otchen, J., Ku ller, L., L i- macher, M. and W act a wski-Wende, J. (200 6). Combined a nalysis of women’s hea lth initiative o bserv a tional and clinical tria l da ta on pos t- menopausal hor mone trea tment and cardiov ascular disease . Am. J. Epidemiol. 163 589– 599. [8] Prentice, R., Pettinger, M. and Anderson, G. (2005). Statistical issue s arising in the women’s hea lth initiative. Biometrics 61 89 9–911 . MR22161 77 [9] Psa ty, B., Heckber t, S., A tkins, D., Siscovick, D., K oepsell, T., W ahl, P., Longstreth, W. J., Weiss, N., W a gner, E. and P ren- tice, R. (1993 ). A review of the asso ciation of estro gens and prog estins with cardiov ascular dise a se in pos tmenopausal women. Ar ch. Intern. Me d. 153 1421– 1427 . [10] Rossouw, J., Anderson, G. , Prentice, R. et al . (2002 ). Risks and be n- efits of estrogen plus prog estin in healthy po stmenopausal women: Principal results from the women’s health initiative r andomized controlled trial. JA MA 288 321– 333. [11] Sidney, S., Petitti, D. and Quesenberr y, C. (19 97). Myo cardial in- farction and the use o f estrogen and es trogen-pr ogestog e n in po stmenopausal women. Ann. Intern. Me d. 127 501– 508. [12] St ampfer, M. and Colditz, G. (199 1). Estrog en replacement therapy and coronar y heart disea s e: A quantitativ e assess men t of the epidemiologic evi- dence. Pr ev. Me d. 20 47–6 3.