J C’linEpidemiol Vol. 43, No. 9, pp. 949-960, Printed in Great Britain. All rights reserved

0895-4356/90$3.00+ 0.00 Copyright 0 1990Pergamon Press plc

1990

BODY MASS INDEX AND 15YEAR MORTALITY COHORT OF BLACK MEN AND WOMEN JAN WIENPAI-IL,‘~~~*DAVID R. RAGLAND*

IN A

and STEPHEN SIDNEY’

‘Division of Research, Kaiser Permanente Medical Care Program, Oakland, CA 94611 and *School of Public Health, University of California, Berkeley, CA 94720, U.S.A. (Received in revised form 4 May 1989; received for publicaiion 30 January 1990)

Abstract-The association between body mass index (BMI) and mortaliy was investigated in 2453 black male (aged 30-79 years) and 273 1 black female (aged 40-79 years) members of the Kaiser Foundation Health Plan. During a 15year follow-up 393 male and 283 female deaths were identified. Analyses were conducted separately in a lower and an upper range of BMI (as well as over the entire range), to isolate separate effects of low weight and high weight on mortality. Particular attention was also paid to potential bias from cigarette smoking and antecedent illness. Cox regression analyses showed that over the entire range of BMI the adjusted BMI-mortality association was significantly J-shaped for the men and essentially flat for the women. The inverse association between BMI and mortality in the lower range of BMI was statistically significant for the men; the adjusted relative hazard increasing from the 10th to the 50th percentile of BMI was 0.76 (95% confidence interval [CI] 0.59-0.98). The positive association between BMI and mortality in the upper range of BMI was highly statistically significant for the men; the adjusted relative hazard increasing from the 50th to the 90th percentile of BMI was 1.37 (95% CI 1.14-1.63). Whether controlled by multivariate analysis, by excluding the first 5 years of follow-up from the analyses, or by analyzing the BMI-mortality association in smoking-specific and/or illness-specific subgroups, smoking and antecedent illness did not have much impact on the BMI-mortality association, in either sex. The general observations on the BMI-mortality association are similar to findings in some white cohorts. Blacks

Body mass index

Cohort study

INTRODUCTION

Very few studies have investigated the relation of body weight to total mortality in black populations [l-4], although blacks, particularly black women, have a high prevalence of overweight and associated morbidity [5-71. Among other populations, mostly white and male, associations between body weight (and other *All correspondence should be addressed to: Jan Wienpahl, National Institute on Aging, - - Epidemiology, Demography and Biometry Program, Federal Bujlding, Room 618. 7550 Wisconsin Avenue. Bethesda. MD 20892. U.S.A. [Tel. 301-496-97951. 949

Men

Mortality

Women

measures of adiposity) and mortality have been inconsistent. Various studies have found a positive association in one or more subgroups of a study population [3,8-131, a negative association [l 1, 14-161 or no association [4, 15, 17-191. Most commonly, elevated mortality has been observed among both lighter-weight and heavier-weight subjects compared to subjects of intermediate weight [3,4,9, 11, 15, 17-311. Several methodological problems may account for the conflicting findings. Failure to consider potential confounders, particularly cigarette smoking and antecedent illness, could bias upward the mortality in the lower range

950

JAN WIENPAHLet al.

of the body weight distribution, leading to an incorrect conclusion that low weight is a risk factor for mortality as well as to an underestimate of the hazard of obesity [2, 10,32,33]. Another problem is the control for potential intermediates (e.g. hypertension, hyperglycemia, and hyperlipidemia) in a causal relationship between body weight and mortality [2,33]. By attenuating a statistical association between body weight and mortality, such control can lead to the erroneous conclusion that body weight does not importantly influence mortality. With these considerations in mind, we examined the association between body weight (as body mass index [BMI]) and mortality in a cohort of black men and women with a 15year follow-up period. Anticipating a U- or J-shaped crude association we formulated two study hypotheses: (i) the negative BMI-mortality association (if observed) in the lower range of the BMI distribution would markedly decrease or disappear when controlled for smoking and antecedent illness, and (ii) the positive BMImortality association in the upper range of the BMI distribution (if observed) would increase when controlled for smoking and illness. The end result should be a positive association of increasing mortality with increasing BMI over the entire BMI distribution. Accordingly our analytic strategy was to examine in detail the BMI-mortality association separately in a lower range and an upper range of the BMI distribution, followed by assessment of the association over the entire range of BMI. We did not control for factors potentially mediating an association between BMI and mortality. METHODS

Study population and variables

The study population comprised 2453 black male and 2731 black female members of the Kaiser Foundation Health Plan, a pre-paid health maintenance organization (HMO). They were part of a cohort whose members underwent three or more multiphasic health checkups at the Oakland or San Francisco Kaiser Permanente facilities between 1964 and 1973, and for which a database had been developed in connection with previous studies [34,35]. We used the third checkup (which took place between 1966 and 1973) to provide the baseline measurements and starting point for mortality follow-up, since the immediately previous study (which was on a similar topic) [35] had similarly used the third

checkup as baseline. The men were 30-79 years and the women were 40-79 years old. Women aged 30-39 years were excluded to minimize the potential influence of pregnancy or lactation (for which information was not available) on body weights. Weight and height measurements were taken with the subject partially clothed, pockets empty and shoes off. The BMI was expressed as weight (kg)/height (m)2. Smoking status was assessed from the health checkup self-administered questionnaire. Subjects were categorized as never-smoker, current cigarette-smoker of < 1 pack per day, current cigarette-smoker of l-2 packs/day, current cigarette-smoker of > 2 packs/day, ex-cigarette non-smoker, cigar- or pipe-only-smoker, or unknown. Antecedent illness status was assessed from questionnaire and laboratory data. Subjects were categorized as either “healthy” (no illness or condition) or “non-healthy” (one or more illnesses/conditions present). The illnesses/ conditions were (by questionnaire): history of heart attack (including angina), stroke, cancer, diabetes, colon or bowel disease, stomach or duodenal ulcer, past-year use of heart medicine or diabetes medicine, any past-year operations, past-6-month involuntary weight loss of more than 10 lb, current disabled work status, and (by laboratory finding): protein in the urine (2 1+), and electrocardiographic evidence of myocardial infarction. Level of education and alcohol use were two additional potential confounders [36-381 for which information was available from the questionnaire. Education level was categorized as high school or less, partial college, college or unknown. Alcohol use was graduate, categorized as none, ~2 drinks per day, > 2 drinks per day, or unknown. Physiologic variables such as blood pressure, serum cholesterol, and serum glucose were not included in the analyses. Since they are generally considered to be potential mediators rather than confounders of relationships between BMI and health outcomes, controlling for them is technically incorrect (unless the mechanisms by which BMI influences the outcome are a topic of interest). Interaction between BMI and blood pressure (in respect to cardiovascular outcomes in particular) is an hypothesis of potential interest [39], which was not investigated in the present study since our focus was the BMImortality association as influenced by smoking and antecedent illness. The latter are the two

BMI and Mortality in Black Men and Women

major factors proposed (e.g. [2]) to account for nonlinear associations between body weight and mortality. ascertainment Mortality classljication

and

vital

status

The cohort was followed for mortality through 1985. Subjects on the Health Plan roster of active members at the end of 1985 were considered to be alive (72% of the male cohort, 74% of the female cohort). Deaths ascertained in previous studies (through 1980) that included this cohort were accepted as confirmed deaths [34,35]. Deaths among remaining subjects were ascertained by computer linkage with the California death files for 1981-1985, using the CAMLIS matching program [40]. 393 male deaths and 283 female deaths were identified for the whole cohort. Subjects not matched in the death clearance were assumed to be alive (12% of the male cohort, 16% of the female cohort). Although some subjects might have been misclassified as alive under this assumption, they probably constituted under 4% of the cohort, since in a previous Kaiser Permanente studying utilizing California death files ascertainment was estimated to be 82-92% complete [41]. Further, a comparison of CAMLIS against the National Death Index (NDI), using a different Kaiser Permanente file, produced similar sensitivities (0.89 for CAMLIS, 0.94 for the NDI); while the predictive value in identifying a person who had died was 0.96 for CAMLIS (and 0.930.99 for three other study files investigated) in contrast to 0.59 for the ND1 [40]. The subjects assumed to be alive were compared to the active Kaiser members for distribution across BMI quintiles and for mean age, current cigarette smoking, and antecedent illness by BMI quintile. No significant differences were identified; therefore we believe that any misclassification of vital status would only have attenuated associations between BMI and mortality. Statistical analyses

Follow-up time for the analyses was counted from the date (month and year) of the baseline exam to the date of death for decedents and to the end of 1985 for subjects classified as alive. All analyses were conducted on men and women separately. Initially, age-standardized death rates were examined by quintiles of BMI. Standardization was by the direct method, with the person-year

951

distribution in lo-year age strata of the entire sex-specific study cohort as the reference. Cox proportional hazards regression models were then used to assess the association between BMI measured on a continuous scale and mortality. Following from the study hypotheses, detailed analyses were conducted separately on a lower range of BMI (quintiles l-3) and an upper range of BMI (quintiles 3-5). The analyses were done identically for each range. The first model included BMI and a term for age. The effect of adjusting for the other covariates was then assessed by the change in the regression coefficient for BMI (or, equivalently, the relative hazard) when the covariates were added to the model. In addition to including a term for antecedent illness in the model, possible bias from pre-existing morbidity was assessed by excluding the first 5 years of follow-up from analyses. Further, to isolate the possible impact of smoking and antecedent illness, analyses were also conducted separately on healthy subjects, never-smokers, and healthy never-smokers. Following the above assessment of the BMImortality association, the relation of BMI to mortality was modeled over the entire range of BMI. In these analyses a quadratic term for BMI (BMI x BMI [BMI*]) was included in the model in addition to the linear term. The BMI corresponding to the minimum relative hazard in the quadratic model was calculated from the formula: -b, /2b2, where b, is the regression coefficient for BMI and b2 is the regression coefficient for BMI* [23,42]. The Statistical Analysis System (SAS) computer program PHGLM was used to perform the regression analyses [43]. Age as well as BMI (and BMI’) were entered as continuous variables; the other covariates were coded as sets of binary indicator terms. Statistical significance was assessed from two-sided p-values. As assessed by the z-statistic calculated from residuals, the proportional hazards assumption of the Cox model was not inappropriate for BMI [43]. RESULTS

Table 1 presents the distributions of the study covariates by quintiles of BMI. Among both men and women, the percentage of current smokers decreased in a stepwise fashion from the first to the fourth quintile, with a slight increase in the fifth quintile. There was no consistent pattern across BMI quintiles with

952

JAN WIENPAHLet al.

Table 1. Distribution of covariates by quintiles of BMI, Kaiser Permanente black men aged 30-79 years and black women aged 4&79 years, 19661973 Men (n = 2453) Quintile BMI (kg/m*) Covariate

(17.1!24.0) (24.1226.0) (26.z27.6)

(276429.7) (29.7S47.7)

qu%iles

Age &), mean Smoking (%) Current cigarette* Never Other7

46.5

47.2

48.2

49.0

49.4

48.1

48.3 16.0 35.8

41.9 19.5 38.6

38.4 23.9 37.8

32.3 25.4 42.3

35.1 26.6 38.3

39.2 22.3 38.6

Antecedent illness (%) Present (non-healthy) Absent (healthy)

31.3 68.7

32.1 67.9

29.2 70.8

27.2 72.8

32.5 67.6

30.5 69.5

Education (%) d High school Partial college College graduate Unknown

53.8 22.1 16.4 7.8

53.9 23.4 15.0 7.7

56.3 22.0 13.1 8.6

57.9 21.3 12.9 8.0

56.8 21.1 11.4 10.8

55.7 22.0 13.7 8.6

Alcohol (%) None < 2 drinks/day > 2 drinks/day Unknown

21.3 53.6 15.1 10.0

22.8 52.2 13.2 11.8

23.5 54.1 13.5 9.0

23.5 52.4 12.5 11.7

25.2 49.5 15.0 10.3

23.2 52.3 13.9 10.6

Women (n = 2731) Quintile BMI (kg/m2) Covariate

(16.6!23.5) (23.Sz25.8) (2S.SaS.l)

(28.1431.4) (31.4SS9.0) quzles

Age (years), mean

50.0

50.0

51.1

52.3

51.8

51.1

Smoking (“7) Current cigarette* Never Others?

43.0 38.0 19.0

32.2 43.6 24.2

31.3 48.5 20.3

24.8 52.5 22.8

26.0 53.5 20.5

31.5 47.2 21.4

Antecedent illness (%) Present (non-healthy) Absent (healthy)

38.9 61.1

35.2 64.8

40.0 60.0

43.1 56.9

40.7 59.3

39.6 60.4

Education (%) C High school Partial college College graduate Unknown

53.9 25.1 14.6 6.4

56.0 18.3 17.6 8.1

58.3 21.4 13.0 7.3

63.3 17.1 9.9 9.7

65.9 17.2 7.9 8.8

59.5 19.8 12.6 8.1

Alcohol (%) None < 2 drinks/day > 2 drinks/day Unknown

37.1 43.1 5.1 14.6

40.7 43.8 4.2 11.4

44.1 40.2 3.1 12.6

51.4 34.3 3.3 11.0

53.1 33.3 3.9 9.7

45.3 39.0 3.9 11.9

*Includes smokers of c 1 pack/day, l-2 packs/day and >2 packs/day. tIncludes ex-smokers, cigar-pipe only smokers, and subjects of unknown smoking status.

respect to antecedent illness: among men, quintiles 1,2 and 5 had somewhat higher percentages of subjects with antecedent illness than quintiles 3 and 4, whereas among women quintiles 3, 4 and 5 had the higher prevalence of illness. BMI was inversely associated with education level in both sexes, a stronger association being noted for women than for men. BMI was positively associated with non-drinking, again particularly so among the women. Age-standardization did not substantially affect any of the distributions shown in Table 1.

Table 2 presents the age-standardized death rates by quintile of BMI for all men and women, and for the subgroup of healthy men and women who reported having never smoked. In the whole cohort of men, the highest mortality was in quintile 1, with a stepwise decrease in mortality to quintile 3, followed by higher rates in quintiles 4 and 5. The pattern was similar in the whole cohort of women except that the lowest mortality was in quintile 2. The rates for the healthy never-smokers were lower than for the study population as a whole. Unexpectedly,

953

BMI and Mortality in Black Men and Women Table 2. Deaths and age-standardized

death rates* by quintiles of BMI, Kaiser Permanente black men and women, 1966-1973 through 1985 All women (n = 2731) Quintile BMI

All men (n = 2453) Quintile BMI 1 Deaths Person-years Death rate* per 1000 person-years (Standard error)

2

3

4

5

13.9 (1.5)

10.1 (1.2)

(V)

‘,Z)

‘,E)

Healthy never-smoking men (n = 382) Quintile BMIt Deaths Death rate* per 1000 person-years (Standard error)

1

2

3

4

5

89 70 69 79 86 65 45 49 66 58 7267.1 7479.7 7415.6 7198.7 7180.4 8565.8 8689.3 8584.2 8382.5 8261.9 (!)

(Z)

(G)

(E)

(0”:;)

Healthy never-smoking women (n = 778) Quintile BMIt

7

5

8

11

8

10

12

6

15

11

12.5 (4.9)

5.3 (2.4)

5.5 (2.1)

8.3 (2.6)

(::;)

(;::)

(;‘:!)

(i:;)

(;‘:;)

(t?;)

*Age-standardized by the direct method, with the person-year distribution in 10 year age strata of the entire sex-specific cohort as the standard. TQuintiles were those of the entire (sex-specific) cohort.

however, healthy never-smokers of both sexes in BMI quintile 1 showed. elevated mortality compared to the healthy never-smoking men and women in the remaining quintiles. Regression analyses on lower-weight subjects (quin tiles l-3) Focusing on the low-to-middle weight component of the cohort, Table 3 presents Cox regression statistics for BMI in relation to mortality, for all men and all women, adjusted for specified covariates. The results of the analyses on healthy individuals, never-smokers and healthy never-smokers are also presented in Table 3. The analysis on all men showed a significant (p = 0.01) negative (inverse) age-adjusted association between BMI and mortality, consistent with the pattern observed among the lower three quintiles in Table 2. Adjusting for smoking and illness attenuated the association rather slightly (Table 3). Adding education and alcohol to the model had almost no additional effect. The relative hazard for the increase from the 10th to the 50th percentile of BMI was calculated: adjusted for the 5 covariates, this relative hazard was 0.76 (95% confidence interval (CI) 0.59-0.98) (Table 3). Excluding the first 5 years of follow-up (40 male deaths) from analysis had little impact on the association (multivariate-adjusted relative hazard = 0.72 [95% CI 0.54-0.951). This suggested that the initial finding was not due to uncontrolled bias from pre-existing morbidity. In the analyses restricted to healthy men and to never-smoker men (not mutually exclusive categories), the coefficient for BMI was not

appreciably different from that in the analysis on all men (Table 3). Finally, the coefficient for BMI for the healthy never-smokers was actually more negative than any of the previous (Table 3). Although the BMI-mortality relationship in these analyses was not statistically significant, this may have been due to the small size of the subgroups. The consistencies in direction and magnitude between the results for the small subgroups and the results for the whole group are compatible with the suggestion that neither smoking nor antecedent illness accounted for the elevated mortality at low BMI. The analysis on all women showed a nonsignificant (p > 0.05) negative age-adjusted association between BMI and mortality in ageadjusted and multivariate models (Table 3), which was not affected by excluding the first 5 years of follow-up (28 female deaths). The coefficients for BMI in never-smoker women and healthy women were little different from the coefficient for the whole group. As in the case of the men, the coefficient for healthy never-smoking women was strongly negative (also consistent with the pattern shown in Table 2). Regression analyses on higher-weight subjects (quintiles 3-5) Focusing on the middle-to-high weight component of the cohort, Table 4 presents Cox regression statistics for the BMI-mortality association parallel to those in Table 3. The analysis on all men showed a highly significant (p < 0.001) positive age-adjusted association between BMI and mortality, which was virtually unchanged with adjustment for the other covariates. The adjusted relative

126

1017 291 205

III. Never-smoker only?

IV. Healthy never-smoker7

-0.078 (0.010) -0.064 (0.039) -0.065 (0.036) -0.077 (0.085) -0.067 (0.442) -0.157 (0.216)

(PI

Coefficient

0.52

0.18-1.47

0.37-1.54

0.50-1.05

0.72 0.76

0.59-0.98

0.59-0.99

0.56-0.91

95% CI

0.76

0.77

0.72

Relative hazard*

men

431

711

1016

1640

1640

1640

No.

28

61

83

159

159

159

Deaths -0.051 (0.116) -0.033 (0.306) -0.037 (0.260) -0.050 (0.272) -0.040 (0.469) -0.154 (0.060)

Coefficient (P)

Low-to-middle-weight

0.48

0.83

0.79

0.84

0.85

0.78

Relative hazard*

women

0.22-1.03

0.59-1.38

OS-l.21

0.62-1.14

0.63-1.16

0.58-1.06

95% CI

129

1036 372 258

III. Never-smoker only?

IV. Healthy never-smokert

0.068 (0.0006) 0.069 (0.0006) 0.068 (0.0007) 0.041 (0.168) 0.078 (0.091) 0.016 (0.844)

(P)

men

0.51-2.28

0.94-2.17

1.43 1.08

0.92-1.57

1.141.63

1.15-1.64

1.14-1.63

95% CI

1.21

1.37

1.37

1.37

Relative hazard

502

843

962

1638

1638

1638

No.

32

79

76

173

173

173

Deaths

0.006 (0.810) -0.004 (0.921)

0.012 (0.496) 0.015 (0.377) 0.017 (0.334). (E)

(PI

Coefficient

Middle-to-high-weight

women

0.97

1.05

1.03

1.14

1.12

1.09

Relative hazard*

0.52-I .80

0.72-1.52

0.68-1.57

0.88-1.47

0.87-1.47

0.85-1.41

95% CI

*For the increase from the 50th percentile to the 90th percentile of BMI: 4.6 kg/m2 for men, 7.4 kg/m* for women. TAdjusted for applicable covariates: age, education (4 categories), alcohol (4 categories), antecedent illness (2 categories) and smoking (7 categories).

27

51

234

1472

234

234

deaths

age, smoking illness, education, alcohol use II. Healthy only?

1472

1472

age, smoking, illness

No.

Subjects in model

I. All (adjusted for): age

Coefficient

Middle-to-high-weight

Table 4. Cox regression coefficients, p-values (p), relative hazards and 95% CI for BMI in relation to mortality, Kaiser Permanente black men and women in the higher-weight (BMI quintiles 3-5) range of the study population, 19661973 through 1985

*For the increase from the 10th percentile to the 50th percentile of BMI: 4.2 kg/m2 for men, 4.8 kg/m’ for women. tAdjusted for applicable covariates: age, education (4 categories), alcohol (4 categories), antecedent illness (2 categories), and smoking (7 categories).

20

35

228

1471

228

228

Deaths

age, smoking, illness, education, alcohol use II. Healthy only?

1471

1471

I. All (adjusted for): age

age, smoking, illness

No.

Subjects in model

Low-to-middle-weight

Table 3. Cox regression coefficients, p-values (p), relative hazards and 95% CI for BMI in relation to mortality, Kaiser Pennanente black men and women in the lower-weight (BMI quintiles l-3) range of the study population, 19661973 through 1985

955

BMI and Mortality in Black Men and Women

hazard (for the 50th to 90th percentile of BMI) was 1.37 (95% CI 1.14-l .63). Excluding the first 5 years of follow-up (49 male deaths) from analysis resulted in a somewhat higher adjusted relative hazard of 1.45 (95% CI 1.19-l .77). The association for healthy individuals and for never-smokers was rather similar to that for all men (Table 4). However, the coefficient for healthy never-smokers was very small (consistent with Table 2). These analyses suggested at least that smoking and illness were not major negative confounders of the overweightmortality association. The analysis on all women indicated a very small positive age-adjusted association, which was not statistically significant (p = 0.50). Adjusting for the other covariates increased the association slightly but it remained nonsignificant (p = 0.33); the adjusted relative hazard was 1.14 (95% CI 0.88-1.47). Excluding the first 5 years of follow-up (42 female deaths) from analysis reduced the adjusted relative hazard to 1.00 (95% CI 0.73-1.37). The association was also negligible in healthy women, never-smoking women, and healthy never-smoking women analyzed separately (Table 4). Regression analyses on the whole cohort

The final regression models summarized the BMI-mortality association over the entire range of BMI, with a quadratic term (BMP) included to assess the nonlinear aspect of the association. Analyses with and without adjustment for the study covariates led to basically the same results; the multivariate results will be presented. Because the modeled relative hazard for BMI (i.e. BMI + BMI’) varied continuously with level of BMI, it is shown in plots (Fig. 1 [for men] and Fig. 2 [for women]). Relative hazards estimated from models excluding the first 5 years of follow-up are also shown (“5-year lag”). The reference level for BMI in the plots is the cohort median BMI (26.8 kg/m2 for men, 26.9 kg/m* for women). For men, the regression coefficients for BMI and BMI* were -0.294 and 0.005 respectively, and were both highly significant (p = 0.0004 and 0.0002). This model described a J-shaped relative hazards curve over the range of BMI (Fig. 1). Excluding the first 5 years from the analysis had little impact (Fig. 1). The BMI corresponding to the minimum relative hazard was 28.0 kg/m2 (95% CI 25.6, 30.4 kg/m*).

16 r

-All

years

---5-year

I

lag

f’ 54 E

/ Medm

05’ 15

I

I

I

I

I

I

I

20

25

30

35

40

45

50

Body mass index

J 55

( kg/m2)

Fig. 1. Relative hazard of BMI for mortality (from Cox regression model with BMI, BMI* and the 5 covariates). Kaiser Permanente black men; n = 2453 (393 deaths) for all years follow-up; n = 2377 (317 deaths) excluding first 5 years of follow-up (“5-year lag”).

For women, the regression coefficients for BMI and BMI* were small (-0.018 and 0.0004, respectively) and did not approach statistical significance (p = 0.83 and 0.78). Excluding the first 5 years of follow-up from the analysis made little difference (Fig. 2; p >0.90 for the regression coefficients for BMI and for BMI’). Analyses without the BMI* term also gave no evidence of a BMI-mortality association (for all years, the regression coefficient for BMI = 0.005, p = 0.70; excluding the first 5 years, b = - 0.009, p = 0.50).

The relationships between mortality and the other study covariates in this black cohort may be of interest. They are presented in Table 5 as relative hazards estimated from the multivariate model for the whole study cohort. 16

1 r

B.

se 0

z E g4

15

----

20

Auywn 5-yearlag

25

30

35

Body moss index

40

45

50

55

( kg/m2)

Fig. 2. Relative hazard of BMI for mortality (from Cox regression model with BMI, BMI* and the 5 covariates). Kaiser Permanente black women; n = 2731 (283 deaths) for all years follow-up; n = 2669 (221 deaths) excluding first 5 years of follow-up (“S-year lag”).

JAN WENPAHL et

956

al.

Table 5. Relative hazards and 95% CI* for age, smoking, antecedent illness, education and alcohol use in relation to mortality, Kaiser Permanente black men (2453 subjects, 393 deaths) and women (2731 subjects, 283 deaths), 1966-1973 through 1985 Men

Women

Relative hazard

95% CI

Relative hazard

95% CI

Age (per I year)

1.09

1.08-1.10

1.09

1.08-1.11

Smoking (reference: never) < 1 pack cigarettes/day l-2 packs cigarettes/day > 2 packs cigarettes/day Cigar/pipe only Ex-cigarette non-smoker Unknown

1.35 1.71 2.04 0.91 1.19 I .42

0.99-l .85

1.51

1.18-3.51 0.55-1.52 0.85-1.68 1.01-1.98

3.56 1.44 0.93

1.09-2.10 2.01-4.24 1.72-7.38 0.92-2.24 0.58-1.49

Antecedent illness (reference: absent) Present

1.48

1.21-1.83

1.58

1.25-2.00

Education (reference: 2 drinks/day Unknown

1.05 1.55 1.29

0.81-1.36 1.17-2.07 0.94-l .78

1.26 1.45 1.34

0.951.67 0.83-2.53 0.89-2.00

Variable and category

1.23-2.38 2.92

*Computed from Cox proportional hazards regression models with terms for BMI, BM12, and all five covariates. Relative hazards were computed from the regression coefficient (b) for the indicated category, with the formula: exp(b). Confidence intervals were calculated with formula: exp(b f 1.96 [SE,]).

DISCUSSION

Over the entire range of BMI, the relative hazards curve for mortality was significantly J-shaped for the men and was essentially flat for the women. Several possibilities were considered to explain the absence of the hypothesized overall positive association between BMI and mortality in either sex. A first possibility was underascertainment of deaths. We believe this potential source of bias could have had only a small, if any, impact on the findings (see “Methods”). Another possible biasing factor was limited follow-up: among the women, if overweightrelated mortality were delayed then 15 years of follow-up may not be long enough to detect an effect [2,24,25,28,31,33,44]. In fact, in the present study any association between high BMI and mortality in women appeared limited to the early years of follow-up. Another possibility is that age may modify the relation between BMI and mortality, such

that the hazards of overweight decrease and the BMI of minimal mortality increases with increasing age [l, 111. However, analyses stratified by baseline age (< 55 vs 2 55 years), as well as analyses with a cross-product term (age x BMI), showed no age effects sufficient to distort the major conclusions of the study. Similarly, the possibility was considered that education or alcohol modified the relationship between BMI and mortality. Analyses of the association by strata of education and alcohol use did not suggest that modification of the effect by these factors could account for either the J-shaped association of BMI with mortality in men or the absence of an association in women. Data were not available for several other [36-381. complicating variables potential Although knowledge of causal sequences is incomplete, to the extent that any such factors contribute to obesity and independently increase the risk of mortality, lack of control for them is unlikely to be responsible for an absence

BMI and Mortality in Black Men and Women

of an overall positive association between BMI and mortality. A potentially important obesity-related factor not addressed in the present study is body fat distribution, which may be directly related to mortality when BMI is not [20,45-471. Among Swedish men the association between BMI and mortality was U-shaped, with lowest mortality in the fourth quintile of BMI, whereas the association between the waist to hip ratio and mortality tended to be linear, with lowest mortality in the first quintile of waist to hip ratio [20]. Centralized body fat distribution has been associated with increased diabetes and cardiovascular risk in blacks as well as whites [5,48-501. Blacks, particularly black women, appear to have relatively more excess trunk or upper body fat than extremity or lower body fat; Gillum suggested that this factor may contribute to higher mortality from several obesity-related diseases in black compared to white women [5,48-49,50 p. 4261. Further consideration of smoking and antecedent illness

Smoking and antecedent illness are the two major factors proposed to account for observed associations of low body weight with mortality [e.g. 21. The present study did not find that smoking explained the elevated mortality at low BMI for either men or women. In contrast, the Kaiser Permanente study of white men and women found excess mortality to be associated with thinness, but only in current cigarettesmokers [35]. Increased mortality among lowweight nonsmokers (who were not in all cases never-smokers) is evident in several studies [14, 18, 19,23-25,30,44]. The numbers of low-weight nonsmokers are often small, with large sampling error [lo, 321, as in the present study. Nevertheless some researchers believe that smoking does not explain the elevated mortality often found in the lightest-weight subjects [14, 18, 29, 51, 521. The present study is compatible with that belief. The present study did not find that controlling for antecedent illness (by excluding early mortality or restricting the analysis to healthy subjects) eliminated the association of low BMI with mortality. Several other studies excluding early mortality (the first 2-12 years) from analysis have not eliminated an elevated relative risk associated with low body weight [ll, 12, 18,26,31,53]. In the Framingham cohort, an elevated relative risk for underweight

951

men (nonsmokers as well as smokers) was apparent after 30 years of follow-up, although minimum mortality was at lower relative weights when assessed over longer durations of follow-up [24]. Among Framingham nonsmoking men and women aged 65 and over the relative risk for low BMI was considerably diminished when the first 4 years of follow-up were excluded from analysis, supporting the interpretation that much of the relationship between thinness and mortality may be explained by illness present at baseline [30]. An elevated (although not statistically significant) risk associated with thinness did remain. Among non-smoking Finnish men a J- or U-shaped relation between BMI and mortality was (apparently) not much affected by excluding the first 2 years of follow-up from the analysis, but the survival curve suggested that “an appropriate length of follow up [approximately 6 years among those nonsmokers] is essential to ‘wash out’ baseline selection” [19]. In the Kaiser Permanente study of whites the relative risk for thinness in men was elevated only in the first five years of follow-up, suggesting that antecedent illness was a confounder even in that selected healthy group; however, among the women the relative risk for thinness was elevated in later years [35]. Thus, findings are mixed with respect to the influence of antecedent illness on the association between low BMI and mortality. It should be noted that long-term prior weight loss, not considered in the present study, may be an important factor in the elevated mortality at low weights [12,21,27,35]. However, the extent to which antecedent disease per se is responsible for low weights, loss and associated excess mortality is far from settled [29, 36, 38, 54, 551. Adiposity and mortality in blacks

Among black men and women of Muscogee County, Georgia, thinness and fatness were both associated with elevated 1Cyear all-cause mortality relative to intermediate adiposity [3]. However, the associations were small (relative risks derived from standardized mortality ratios were 1.05 and 1.13 for thin men and women, and 1.17 for fat subjects of both sexes [statistical significance was not indicated]), and fatnessmortality relationships were irregular among age-sex subgroups. One conclusion of the study was that “excessive mortality among fat persons was generally more marked for whites than for Negroes” [3, p. 5621.

958

JAN WIENPAHL. et al.

Among Evans County black men, no statistically significant association between BMI and 20-year all-cause mortality was detected; the observed relationship appeared to be a downward trend in mortality from low to high BMI. In contrast, statistically significant U-shaped associations between BMI and mortality were detected for white men of both low and high social status [4]. Among black women and high social status white women, no significant associations between BMI and 20-years cardiovascular mortality were detected, in contrast to a significant positive association for low social status white women. The investigators suggested that either chance or an unmeasured correlate of body weight may be the explanation for this unexpected result [56]. Coming from relatively poor, predominately rural populations in the southeastern U.S., blacks in the above cohorts were no doubt different in many respects from Kaiser Permanente blacks, who were mostly urban and employed, and who took at last three health checkups under a prepaid medical plan in northern California. A basic difference is that Evans County black men had considerably lower mean BMI than Kaiser Permanente black men (25.1 vs 26.9 kg/m2), and Evans County black women had considerably higher mean BMI than Kaiser Permanente black women (29.1 vs 26.8 kg/m2). Further, death rates for such major causes of death as heart disease and stroke are higher for blacks (and whites) in Georgia than in California [57]. The U- or J-shaped associated between BMI and mortality observed for the Kaiser Permanente black men generally resembled that in a number of non-black cohorts [4, 11, 12, 15, 17-311. As in several of those cohorts, the estimated BMI corresponding to minimum mortality (28 kg/m’) approximated or was greater than the average BMI of the study population, and was considerably higher than the “desirable” BMI calculated from the 1959 Metropolitan Life Desirable Weight Table (22 kg/m2, range from 19.8 to 25.7 for all heights and frames) [33]. (The quadratic regression method of estimating the minimum mortality BMI may yield results biased upwards [27], but it is unlikely that the “optimal” BMI in the present male cohort was less than 24 kg/m2, i.e. in the first quintile.) The lack of association between BMI and mortality in the Kaiser Permanente black female cohort appears compatible with an

hypothesis [58,59] that the mortality hazards of obesity are smaller for black women compared to white. In fact, however, several studies of white female cohorts have also failed to find an association [ 16, 18, 191. The sparse data available from cohort studies of blacks may be influenced by unknown selection biases and numerous unmeasured confounders [5, p. 381. Also, conclusions concerning low risks of obesity for black women contradict ecologic data from large populations representative of the U.S. population [5, p. 381. Given these considerations, the notion that the hazards of obesity are less for black women than for white women is premature.

CONCLUSION

Two major hypothesized confounders of the body weight-mortality association, i.e. cigarette smoking and antecedent illness, did not have a major impact in the present study. Therefore, the J-shaped association found for men, and the essentially flat association found for women, were apparently not caused by these two factors. The risk pattern (for men and albeit differently, for women) in the present black study population was similar to that in several non-black cohorts whereas, at the same time, different patterns have been found among the few black cohorts studied as well as among other nonblack cohorts. These observations suggest that the relation of body weight and adiposity to mortality is not necessarily importantly influenced by smoking, antecedent illness, or race. Various possible important confounders and effect modifiers have been noted [e.g. in 2,29, 51,601. Identification of these factors and their influence is an important goal for further study of body weight/adiposity and mortality. support and encouragement of S. Leonard Syme is gratefully acknowledged. The authors thank Gary Friedman, Beth Newman, Joseph Selby and Edmund Van Brunt for their helpful comments on the manuscript; Merril Jackson for technical assistance; and BNCJZ Fireman and Charles Quesenberry for statistical assistance. The work was supported by Grant T32 HL07365 from the National Heart, Lung and Blood Institute and Grant ROl AM30904 from the National Institute of Arthritis, Diabetes and Digestive and Kidney Diseases. Acknowledgements-The

REFERENCES 1. Andres R, Elahi D, Tobin JD, Muller DC, Brant L. Impact of age on weight goals. Ann Intern Med 1985; 103: 1030-1033.

BMI and Mortality in Black Men and Women 2.

3. 4.

5. 6.

Manson JE, Stampfer MJ, Hennekens CH, Willett WC. Body weight and longevity: A reassessment. JAMA 1987; 257: 353-358. Comstock GW, Kendrick MA, Livesay VT. Subcutaneous fatness and mortality. Am J Epidemlol 1966, 83: 548-563. Tyroler HA, Knowles MG, Wing SB et al. Ischemic heart disease risk factors and twenty-year mortality in middle-age Evans County black males. Am Heart J 1985; 108: 738-746. Kumanyika S. Obesity in black women. Epidemiol Rev 1987: 9: 31-50. Van itallie TB. Health implications of overweight and obesitv in the United States. AM Intern Med 1985; 103:

26.

27. 28. 29. 30.

983-9i38.

7. 8. 9. 10.

11. 12.

13.

14. 15. 16. 17.

18.

19.

20.

21. 22. 23.

24. 25.

Van Itallie TB. News about obesity [Letter]. N En@ J Med 1986; 315: 131. Build and Blood Pressure Study, 1959. Chicago: Society of Actuaries: 1959. Cole TJ, Gilson JC, Olsen HC. Bronchitis, smoking and obesity in an English and Danish town. Bull Physio-Path Resp 1974; 10: 657-679. Garrison RJ, Feinleib M, Castelli WP, McNamara PM. Cigarette smoking as a confounder of the relationship between relative weight and long-term mortality: the Framingham Heart Study. JAMA 1983; 249: 2199-2203. Jarrett RJ, Shipley MJ, Rose G. Weight and mortality in the Whitehall Studv. Br Med J 1982: 285: 535-537. Rhoads GG, Kagan A. The relation of coronary disease, stroke, and mortality to weight in youth and in middle age. Lancet 1983; i: 492-495. Cochrane AL, Moore, F, Baker IA, Haley TJL. Mortality in two random samples of women aged 55-64 followed up for 20 years. Br Med J 1980; i: 1131-1133. Garn SM, Hawthorne VM, Pilkington JJ, Pesick SD. Fatness and mortality in the West of Scotland. Am J Clin Nuti 1983; 38: 313-319. Keys A, Aravanis C, Blackburn H ef al. Seven Countries. A Multivariate Analysis of Death and Coronary Disease. Cambridge: Harvard University Press; 1980. Noppa H, Bengtsson, Wedel H, Wilhemsen L. Obesity in relation to morbidity and mortality from cardiovascular disease. Am J Epidemiol 1980; 111: 682-692. Pettitt DJ, Lisse JR, Knowler WC, Bennett PH. Mortality as a function of obesity and diabetes mellitus. Am J Epidemiol 1982; 115: 359-366. Vandenbroucke JP, Mauritz BJ, de Bruin A, Verheesen JHH, van der Heide-Wessel C, van der Heide RM. Weight, smoking, and mortality. JAMA 1984; 252: 2859-2860. Tuomilehto J, Salonen JT, Marti B et al. Body weight and risk of myocardial infarction and death in ihe adult population of eastern Finland. Br Med J 1987; 295, 623-627. Larsson B. Obesity and prospective risk for associated diseases. In: Vague J, Bjiirntorp P, Guy-Grand B et al., Eds. Metabolic Complications of Human Obesities. Amsterdam: Elsevier;.1985; 21-29. Avons P, Ducimetiere P, Rakotovao R. Weight and mortality [Letter]. Lancet 1983; i: 1104. Belloc NB. Relationship of health practices and mortalitv. Prev Med 1973: 2: 67-81. Dyer Ak, Stamler J, Beikson DM, Lindberg HA. Relationship of relative weight and body mass index to 14-year mortality in the Chicago Peoples Gas Company Study. J Chron Dis 1975; 28: 109-123. Feinleib M. Epidemiology of obesity in relation to health hazards. Ann Intern Med 1985; 103: 1019-1024. Lew EA, Garfinkel L. Variations in mortality by weight among 750,000 men and women. J Chron Dis 1979; 32: 563-576.

31.

32. 33. 34.

35. 36. 37. 38.

39. 40.

41. 42.

43.

44.

45. 46.

47.

959

Schroll M. A longitudinal epidemiological survey of relative weight at age 25, 50, and 60 in the Glostrup population of men and women born in 1914. Dan Med Bull 1981; 28: 106-116. Waaler HT. Height, weight and mortality: The Norwegian experience. Acta Med Stand 1984; 679 (Suppl.): l-56. Build Study 1979. Chicago: Society of Actuaries and Association of Life Insurance Medical Directors of America; 1980. Gordon T, Doyle JT. Weight and mortality in men: The Albandy Study. Int J Epidemiol 1988; 17: 77-81. Harris T, Cook EF, Garrison R, Higgins M, Kannel W, Goldman L. Body mass index and mortality among nonsmoking older persons. The Framingham Heart Study. JAMA 1988; 259: 1520-1524. Hoffmans MDAF, Kromhout D, Coulander CDL. The impact of body mass index of 78,612 Is-year old Dutch men on 32-year mortality from all causes. J Clin Eoidemiol 1988: 41: 749-756. Garrison RJ, Castelli WP. Weight and thirty-year mortality of men in the Framingham Study. Ann Intern Med 1985; 103: 1006-1009. Simopoulos AP, Van Itallie TB. Body weight, health, and longevity. AM ‘InternMed 1984; 100: 285-295. Friedman GD, Petitti DB, Bawol RD, Siegelaub AB. Mortality in cigarette smokers and quitters: Effect of base-line differepces. N Engl J Med 1981; 304: 1407-1410. Sidney S, Friedman GD, Siegelaub AB. Thinness and mortality. Am J Public Health 1987; 77: 317-322. Beasley JW. Body weight and longevity [Letter]. JAMA 1987; 257: 1895. Emsberger P. Body weight and longevity [Letter]. JAMA 1987; 257: 1895-1896. I Manson JE, Willett WC, Stampfer MJ, Hennekens CH. Body weight and longevity [Reply]. JAMA 1987; 257: 1896. Phillips A, Shaper AG. Relative weight and major ischaemic heart disease events in hypertensive men. Lancet 1989; 1: 1005-1008. Arellano MG, Petersen GR, Petitti DB, Smith RE. The California Automated Mortality Linkage System (CAMLIS). Am J Public HeaIth 1984; 74: 1324-1330. Friedman GD, Dales LG, Ury HK. Mortality in middle-aged smokers and non-smokers. N Engl J Med 1979; 300: 213-217. Cooner SP. Hardv. RJ., Labarthe DR et al. The relatjon between degree of blood pressure reduction and mortality among hypertensives in the Hypertension Detection and Follow-up Program. Am J Epidemiol 1988; 127: 387-403. Harrell FE Jr. The PHGLM procedure. In: Reinhardt PS, Ed. The SAS Supplemental Library User’s Guide, version 5 edn. Cary, N.C.: SAS Institute, Inc.; 1986: 437-466. Lew EA. Mortality and weight: Insured lives and the American Cancer Society Studies. Ann Intern Med 1985; 103: 1024-1029. Bjijrntorp P. Regional patterns of fat distribution. Ann Intern Med 1985; 103: 994-995. Lapidus L, Bengtsson C, Larsson B, Pennert K, Rybo E, Sjostrom L. Distribution of adipose tissue and risk of cardiovascular disease and death: a 12 year follow up of participants in the population study of women in Gothenburg, Sweden. Br Med J 1984; 289: 1257-1261. Larsson B, Svardsudd K, Welin L, Wilhelmsen L, Bjiirntorp P, Tibblin G. Abdominal adipose tissue distribution, obesity, and risk of cardiovascular disease and death: 13 year follow up of participants in the study of men born in 1913. Br Med J 1984; 288: 1401-1404.

960 48. 49.

50.

51. 52. 53.

JAN WI~NPAHLet al. Feldman R, Sender AJ, Siegelaub MS. Difference in diabetic and nondiabetic fat distribution patterns by skinfold measurements. Diabetes 1969; 18: 478-486. Blair D, Habicht J-P, Sims EAH, Sylvester D, Abraham S. Evidence for an increased risk for hypertension with centrally located body fat and the effect of race and sex on this risk. Am J Epidemiol 1984; 119: 526-540. Gillum RF. The association of body fat distribution with hypertension, hypertensive heart disease, coronary heart disease, diabetes and cardiovascular risk factors in men and women aged 18-79 years. J Chron Dis 1987; 40: 421-428. Andres R. Effect of obesity on total mortality. Int J Ohes 1980; 4: 381-386. Jarrett RI. Is there an ideal body_ weight? Br Med J 1986; 293: 493-495. Sorlie P. Gordon T. Kannel WB. Bodv build and mortality: the Framingham Study. JAMA 1980; 243: 1828-1831.

54. 55. 56.

57.

58. 59. 60.

Seltzer F. Discussion of Doctor Andres’ paper. Trams Aasoc Life Irtsur Med Dir Am 1981; 44: 187-197. Silver A. Body mass index and mortality in the elderly [Letter]. JAMA 1988; 260: 182-183. Johnson JL, Heineman EF, Heiss G, Hames CG, Tyroler HA. Cardiovascular disease risk factors and mortality among black women and white women aged 40-69 years in Evans County, Georgia. Am J Euidemlol 1986: 123: 209-220. National Center for Health Statistics. HeaItb, United States, 1984, DHHS Publ. No. (PHS) 85-1232. Public Health Service, Washington, D.C.: U.S. GPO; December 1984. Ernsberger P, Haskew P. News about obesity [Letter]. N Enal J Med 1966: 315: 130-131. Walk& ARP, Segal I. The puzzle of obesity in the African black female [Letter]. Lencet 1980; i: 263. Stallones RA. Epidemiologic studies of obesity. Ann Intern Med 1985; 103: 1003-1005.