Changes in Body Mass Index Prior to Baseline among Participants Who Are Ill or Who Die during the Early Years of Follow-up

The association between body mass index (weight (kg)/height (m) 2) and mortality may be confounded by preexisting illness. A method commonly used to control for this confounding is the exclusion of participants who have certain diseases at baseline and/or those who die during the early years of follow-up. The authors used data from the Atherosclerosis Risk in Communities (ARIC) Study (n = 14,088) to determine whether participants identified by these criteria had different changes in body mass index than other participants. Weight change was measured over a 3-year interval between study entry (1987–1989) and reexamination (1990–1993), and information on vital status was collected over the subsequent 5 years. Mean change in body mass index was –0.54 (95% confidence interval (CI): –0.90, –0.12) among participants who died in the first year of follow-up, –0.03 (95% CI: –0.18, 0.12) among those who died in the first 4 years of follow-up, and 0.36 (95% CI: 0.33, 0.39) among those who survived for at least 5 years. Participants who died during the first 4 years were over twice as likely as survivors to have changed from the obese category (body mass index ≥30) to the nonobese category (odds ratio = 2.14; 95% CI: 1.44, 3.17). Mean change in body mass index prior to baseline was not different among ill participants compared with those who were healthy, but the odds of converting from obese to nonobese were higher in ill participants than in healthy ones (odds ratio = 1. Epidemiologic studies of the relation between body mass index (BMI), usually defined as weight divided by the square of height, and longevity have formed the basis of national recommendations for healthy body weight. Therefore, the careful scrutiny that the methods used in these studies have received from investigators is merited. In 1987, Manson et al. (1) published a seminal paper that for the first time summarized potential biases in studies of body weight and mortality. Manson et al. concluded that such studies were subject to three major biases: " 1) failure to control for cigarette smoking, 2) inappropriate control for the biologic effects of obesity, such as hypertension and hyper-glycemia, and 3) failure to control for weight loss due to subclinical disease " (1, p. 353). This paper focuses on the third source of bias. As Andres stated, " The assumption is that undiscovered illness, destined to lead to early mortality, was present at enrollment and …

Epidemiologic studies of the relation between body mass index (BMI), usually defined as weight divided by the square of height, and longevity have formed the basis of national recommendations for healthy body weight.Therefore, the careful scrutiny that the methods used in these studies have received from investigators is merited.In 1987, Manson et al. (1) published a seminal paper that for the first time summarized potential biases in studies of body weight and mortality.Manson et al. concluded that such studies were subject to three major biases: "1) failure to control for cigarette smoking, 2) inappropriate control for the biologic effects of obesity, such as hypertension and hyperglycemia, and 3) failure to control for weight loss due to subclinical disease" (1, p. 353).This paper focuses on the third source of bias.As Andres stated, "The assumption is that undiscovered illness, destined to lead to early mortality, was present at enrollment and that it had already resulted in some weight loss" (2, p. 418).Data obtained from participants who had experienced disease-induced weight loss could attenuate the observed obesity-mortality association by increasing mortality at lower body weights.
The impact of preexisting illness in a study of the BMImortality association requires careful consideration of study design and the components of confounding.Classic studies of the BMI-mortality association use a cohort design in which one assessment of BMI is made (at baseline), participants are then followed, and the date (or absence) of death is observed.For confounding to occur, the confounding factor must be associated with both the exposure (BMI at baseline) and the outcome (death) and must not be a mediator of the impact of BMI on death.In this context, figure 1 illustrates four scenarios that are relevant to the issue of confounding in this type of study.
The first scenario is the one the investigator is striving to create.Here the weight measured at baseline represents the participant's "usual" weight.At some time after the baseline measurement, the participant develops disease, and this disease may lead to death.The investigator wishes to determine whether the development of disease and death occur more often among subjects who are obese than among those who are not obese.Note that weight loss caused by disease in this scenario has no impact.Weight is assessed only at baseline, and changes in weight after baseline are irrelevant.
In contrast, the second scenario illustrates confounding by illness that is present at baseline.In this scenario, the par- ticipant's "usual" weight would meet the criterion for obesity, but because of illness, his or her weight has dropped to a normal level.Thus, the weight measured at baseline does not capture this participant's "usual" weight.The subject is more likely to die because he or she has a disease, and the confounding triangle is complete.
In the third scenario, disease is also prevalent at baseline but does not cause a reduction in body weight.In this scenario, there is no confounding.This scenario applies to diseases that are unrelated to body weight and to conditions that may be associated with body weight but do not themselves cause weight loss.Hypertension and hypercholesterolemia are examples of the latter type of condition.
The fourth scenario illustrates another situation in which disease occurs before baseline.In this scenario, the participant has a chronic disease of early onset that has caused him or her to be thin throughout his/her entire life.The weight measured at baseline correctly assesses the participant's "usual" weight, but the participant might have been obese if he or she had not had the disease.However, this is unknown.Examples of this type of disease include sickle cell anemia, cystic fibrosis, congenital heart disease, and inflammatory diseases such as Crohn's disease and sarcoidosis.Since these diseases occur with relatively low frequency, the number of individuals in a community-based sample to whom this scenario applies is likely to be small.These scenarios are meant not to describe all possible scenarios but to illustrate issues pertinent to bias from preexisting illness and weight loss in studies of the BMI-mortality association.In practice, information on weight history is rarely available, information on existing disease at baseline is often not available, and information on as-yet-undiagnosed disease is never available.It has become customary to attempt to control for confounding due to disease-induced weight loss by excluding from analyses participants with cancer, cardiovascular disease, and/or other available indicators of ill health, and/or participants who die within the first few years follow-ing the baseline weight measurement.Studies which do not include one of these types of exclusions have been heavily criticized by some authors (3)(4)(5).
Recently, Allison et al. (6)(7)(8) challenged the validity of the exclusion of early deaths as a method of controlling for confounding due to preexisting illness.They used a metaanalysis ( 6), a mathematical demonstration (7), and a simulation (8) to provide evidence that exclusion of early deaths has a "minuscule" effect (6) and in theory could "exacerbate the confounding due to preexisting illness" (7).
One issue that, to our knowledge, has not been examined by these or other authors is the assumption that participants who die during the early years of vital status follow-up are more likely to have had recent weight loss than are other participants.In other words, preceding the baseline measurement, they are more likely to have experienced diseaseinduced weight loss that would cause an underestimation of their "usual" weight.The purpose of this study was to directly assess this phenomenon by comparing weight loss prior to baseline among subjects who died during the early years of follow-up with that among subjects who remained alive.In addition, we assessed the usefulness of information on prevalent disease to identify participants with weight loss prior to baseline.

MATERIALS AND METHODS
Data for this research were obtained from the Atherosclerosis Risk in Communities (ARIC) Study, a prospective, multicenter investigation of atherosclerosis and cardiovascular disease (9).Between 1987 and 1989, 15,792 men and women aged 45-64 years were examined in four US communities: Forsyth County, North Carolina; Jackson, Mississippi; the northwestern suburbs of Minneapolis, Minnesota; and Washington County, Maryland.Approximately 46 percent of eligible persons in Jackson and 65 percent of those eligible in the other three communities entered the cohort (10).Between 1990 and 1993, 91 percent of the original cohort was reexamined at a study clinic.The average amount of time elapsed between clinic examinations was 3 years.Between clinic visits, participants were interviewed annually by telephone.
For these analyses, participants who were not of either White or Black ethnicity were excluded (n ϭ 48).In addition, 55 Black participants in the Minneapolis and Washington County field centers were excluded because their numbers were too low to permit modeling with centerspecific indicators of ethnicity.We excluded cohort members who were not examined at the second visit (n ϭ 946) and those with missing data on pertinent variables (n ϭ 35) or with values out of the quality control range (n ϭ 18).Eight participants who had been lost to follow-up were excluded, as were 594 participants who had not accrued 5 years of vital status follow-up after the second clinic visit.The analysis sample included 14,088 participants (5,615 White women, 4,996 White men, 2,193 African-American women, and 1,284 African-American men).
Height was measured at the first examination, and weight was measured at both examinations.Height was measured to the nearest centimeter using a metal rule attached to a wall and a standard triangular headboard.Weight was measured in pounds using a beam balance with the subject standing in a scrub suit and no shoes.BMI was calculated as weight in kilograms divided by height in meters squared.BMI at the first examination was used as a covariate in models.Average yearly changes in BMI were calculated for each participant between the dates of his or her two clinic visits.We multiplied the average annual change by 3 to obtain the average change over a 3-year interval.
Cigarette smoking was measured by questionnaire, and participants were classified as never smokers, former smokers, or current smokers.Educational attainment was studied in three categories: less than high school, completion of high school, and more than high school.
Subjects were categorized as having "preexisting illness" if they showed evidence of heart disease, stroke, or cancer at or before visit 2. In addition, they were categorized as having preexisting illness if they described their health as "poor" (rather than "fair," "good," or "excellent") in comparison with the health of others during the annual followup telephone call that most closely preceded the second clinic visit.The categories of heart disease and stroke included prevalent cases noted at visit 1 and incident cases detected between visits 1 and 2 or at visit 2 using standard definitions developed by the ARIC investigators (11).These definitions used data from physician interviews, chart reviews, and diagnostic procedures, as well as information from self-reports.Information on the prevalence of cancer was collected at visit 2 using only self-reports.Participants who reported having been told by a doctor that they had cancer were asked about the location of the cancer.Cases who reported having had skin cancer were not included in our definition, since these cancers are usually benign and not associated with weight loss.Participants who did not meet any of these criteria for preexisting illness were categorized as "healthy." Vital status was obtained by annual telephone interviews through July 1996.Information on vital status obtained from these interviews has been compared with information obtained from death certificates and found to have high validity (99 percent agreement on events and 97 percent agreement on month of death).For the purposes of this study, vital status follow-up began on the day after a participant's second clinic visit.Here, the second clinic visit will be called "baseline," and weight change between the first and second clinic visits will be referred to as weight change "prior to baseline." Weight change was examined with adjustment for height using BMI.This was done for two reasons.First, it seemed logical to assume that a given weight change would have different meanings depending on the height of the individual.Second, most studies examining the effects of weight on mortality use BMI rather than weight.Mean changes in BMI were examined as well as categories of BMI change.We selected the categories to focus on changes sufficient to impact an analysis examining BMI and mortality.We arbitrarily chose a BMI gain or loss of 3 as a cutpoint to define major weight change over a 3-year period.A change in BMI of 3 would be approximately equivalent to a weight change of 21 pounds (9.5 kg) in a person 70 inches (179 cm) tall.We also created categories based on BMI changes that resulted in the participant's crossing the cutpoints of 25 and 30, because these cutpoints, endorsed by the International Task Force on Obesity (12) and the US National Institutes of Health (13), are often used to define categories of overweight and obesity, respectively.
Associations of year of death and preexisting illness with BMI change between visits 1 and 2 were examined using general linear models and multiple logistic regression with generalized logits.Covariates included age, ethnicity, gender, study center, education, and smoking.Data were analyzed using the PROC GLM and PROC CATMOD procedures in SAS/STAT 6.10 (SAS Institute, Inc., Cary, North Carolina).

RESULTS
Table 1 shows descriptive information from visit 1 by categories of BMI change between visits 1 and 2. Mean age was approximately 2 years lower among participants with a major weight gain (a BMI increase of >3) compared with those who had a major weight loss (a BMI decrease of >3).In all four ethnicity/gender groups, the majority of participants gained 0-3 units in BMI.White men were least likely to gain more than 3 BMI units (2.9 percent), while African-American women were most likely to have a BMI change in that category (8.8 percent).African-American women also had the highest prevalence of major weight loss (4.7 percent) over the 3-year follow-up period.Men were less likely than women to have a >3-unit decline in BMI.
There were 588 deaths over the 5-year follow-up period.Among subjects who were alive after 5 years of follow-up, approximately 2.8 percent had a major weight loss and 5.4 percent had a major weight gain resulting in a BMI change greater than 3. Compared with live subjects, subjects who died tended to have more often experienced moderate (0-3) or major (>3) reductions in BMI prior to baseline.The percentage of subjects who had experienced major weight gain also tended to be slightly higher among persons who died after the first year of follow-up, while the percentage of subjects who had experienced moderate weight gain tended to be smaller among those who died.
Table 1 shows the number of participants who met our definition of preexisting illness and the numbers with cardiovascular disease, cancer, and self-reported poor health.(Since some participants had more than one health condition, the total number for all conditions does not equal the total number of persons with preexisting illness.)Participants were defined as having missing data if variables used for disease classification were missing.Weight loss tended to be more common among participants with cancer than among healthy participants, but the differences were not as large as might have been expected (4.4 percent vs. 2.8 percent with major weight loss; 37.2 percent vs. 38.4percent with minor weight loss).In comparison with healthy subjects, there was a tendency for a larger percentage of participants who described themselves as being in poor health to have a major weight loss or gain.
Among persons with preexisting illness, 10 percent (n ϭ 219) died during the first 4 years of follow-up, whereas only   2 percent (n ϭ 232) of those classified as healthy died.Among participants with preexisting illness who also died during the first 4 years of follow-up, 11 percent had a major weight loss prior to baseline.Participants who met both of these criteria did not appear to be more likely to have a major weight gain than other participants (5.9 percent vs. 5.4 percent).

Mean changes in BMI prior to baseline
Figure 2 shows the adjusted mean changes in BMI prior to baseline according to vital status in each year of followup.Among subjects who were alive at the end of the 5-year follow-up period, mean BMI change over the 3 years prior to baseline was 0.36 (95 percent confidence interval: 0.33, 0.39).For these analyses, the BMI change among participants who died was compared with that of subjects who were alive at the end of the designated year.The difference in BMI change by vital status was largest in the first year of follow-up (BMI change was 0.36 for those who remained alive vs. -0.54for those who died).However, the differences remained statistically significant in the second and third years of follow-up (p < 0.05) and borderline significant in the fourth year (p ϭ 0.08).Combining deaths from the first 4 years of follow-up, mean BMI change prior to baseline was -0.03 in participants who died compared with 0.36 in those who remained alive at the end of 4 years (p < 0.001).Among participants who died during the fifth year of follow-up, BMI change prior to baseline was very similar to that of persons who lived throughout follow-up (0.36 in those who remained alive vs. 0.30 in those who died; p ϭ 0.59).
There was no difference in mean BMI change prior to baseline among subjects with preexisting illness compared with those classified as healthy (figure 3, top).After adjustment for covariates, BMI change in healthy participants averaged 0.36 while BMI change in participants with illness averaged 0.30 (p ϭ 0.21).There was a significant interaction (p < 0.05) between baseline health status and 4-year survival; this indicates that the difference in BMI change was larger among ill participants who died (mean difference -0.21) than among healthy participants who died (mean difference -0.07), in comparison with those who remained alive.The lower half of figure 3 shows mean BMI change among participants with preexisting illness who died within the first 4 years of follow-up.BMI change in this group was different (p < 0.001) from BMI change in the combined group of persons who were classified as healthy and persons who were classified as unhealthy but survived the first 4 years of follow-up.

Changes in BMI categories prior to baseline
Table 2 shows the numbers of participants who crossed the cutpoints for overweight (BMI ϭ 25.0) and obesity (BMI ϭ 30.0).Also shown are adjusted odds ratios for crossing these cutpoints.Participants who died during the first year of follow-up were 4.21 times as likely to have changed from overweight to normal weight than were those who remained alive.They were not significantly more likely to have changed from obese to nonobese, perhaps because there were only three deaths in this category.When data from persons who died during the first 4 years of follow-up were combined, the odds of converting to "not overweight" and the odds of converting to "not obese" were both doubled in comparison with persons who did not die during that period.Participants who died in the fifth year of follow-up were not significantly more likely to have crossed either cutpoint than were those who remained alive.
Participants with preexisting illness had 29 percent higher odds of having changed from overweight to normal weight than healthy participants, but they were not more likely to have changed from normal weight to overweight.Among participants who had preexisting illness and died during the first 4 years of follow-up, the odds of changing classification from overweight to normal weight were 2.96 times those of participants who did not have preexisting illness or did not die during the first 4 years.In both groups with preexisting illness, the odds of crossing the obesity cutpoint were increased in both directions (becoming obese and becoming nonobese).

DISCUSSION
In the ARIC cohort, mean weight change 3 years prior to baseline was positive among subjects who remained alive during follow-up.This tendency to gain weight is consistent with national cross-sectional (14) and longitudinal (15,16) studies that show that adults tend to gain weight throughout middle age.In this study, participants who died during the first 4 years of follow-up did not gain as much weight prior to baseline as did participants who did not die in that interval.In addition, participants who died during the first 4 years of follow-up and those who were classified as having disease at baseline were more likely to have lost weight such that they were no longer categorized as obese or overweight.Thus, they were at increased odds of falling into scenario 2 as it is described in figure 1, and results for these participants were more likely to be biased.
One important limitation of this work is our assumption that weight measured 3 years prior to baseline captured "usual" weight.In this context, "usual" weight must be considered not a single value but a trajectory with values changing with age, since it is "usual" to gain weight with age.It is likely that measurement of weight change over a shorter or longer time interval prior to baseline would give somewhat different results.Nevertheless, the 3-year interval used here seems reasonable.One important strength of this study is that BMI change was assessed using two measured weights rather than self-reports.
The design of this study dictated that participants who did not survive for at least 3 years after recruitment into the cohort be excluded from the sample.This may have contributed to the weak association between preexisting disease and mean BMI change observed here.Participants had to survive for at least 3 years after recruitment so that BMI change could be measured.Therefore, subjects with the most severe disease may have been eliminated from the study sample.In addition, disease-induced weight loss is likely to differ depending on the time of onset of the disease.This factor was ignored here, and we did not distinguish between diseases with recent onset and those with onset in the remote past.At visit 1, the mean BMI of participants who did not attend visit 2 was slightly higher than that among those who did attend visit 2 (28.5 vs. 27.6), but this does not inform us about potential differences in weight change.Our findings point to the need for more research on the associations of weight loss with type and severity of disease.
In their demonstration of the effects of early exclusion on bias, Allison et al. (8) did not consider the possibility that persons who die during the early years of follow-up might be more likely to have lost weight than those who die in later years of follow-up.A different type of mathematical demonstration could assume that the odds of misclassifying participants (for example, as normal weight instead of overweight) were substantially higher among subjects who died during the first 4 years of follow-up than among subjects who died later in follow-up, and that the odds of misclassifying subjects who died 4 years after baseline were equivalent in those who died and those who lived.It is possible that such a demonstration would find that eliminating the first 4 years of deaths reduces bias.* Weight (kg)/height (m) 2 .† Odds ratios were obtained from multiple logistic regression analyses using generalized logits and were adjusted for age, ethnicity, gender, study center, body mass index (BMI), education, and smoking.For these analyses, there were three outcome categories (stable, gain, loss).BMI change among participants who died was compared with that in participants who were alive at the end of the designated year(s).BMI change among participants with preexisting illness was compared with that in participants who were healthy.
‡ OR, odds ratio; CI, confidence interval.§ Referent.¶ Not applicable.Attenuation of the BMI-mortality association among participants with preexisting illness has been demonstrated in studies of two extremely large cohorts (17,18).Nevertheless, the elimination of early deaths from analyses of the BMImortality association has been shown to have little impact on results (6).This is not surprising, given that the majority of the observations in an analysis are not impacted by this maneuver.The number of deaths occurring in the first few years of follow-up and the amount of preexisting illness depend on several factors, including the age and health of the original cohort.In this sample of 14,088 adults aged 48-67 years at baseline, there were 77 deaths in the first year and 451 deaths over the first 4 years of follow-up.Assuming that weight 3 years prior to baseline is indicative of "usual" weight, our results indicate that 13.5 percent more subjects among those who died in year 1 (the difference between 16.9 percent and 3.4 percent (table 2)) would be misclassified as normal weight (rather than overweight) in comparison with those who remained alive at year 5.In the first 4 years of death, 5.5 percent more would be misclassified (8.9 percent minus 3.4 percent).Considering only the cutpoint for overweight and assuming that misclassification of 3.4 percent of the observations is unavoidable or expected, 77 deaths would be eliminated to avoid inclusion of 10 misclassified observations in the first year of follow-up, and 451 deaths would be eliminated to avoid inclusion of 25 misclassified observations in the first 4 years of follow-up.Considering only one cutpoint is an oversimplification of analyses of the BMI-mortality association, but the fact remains that most of the participants identified by early death are not misclassified by our definition.
Since the number of deaths is critical to the power of an analysis of the BMI-mortality association, the number of deaths lost through the use of different exclusion criteria should be emphasized.Elimination of subjects with preexisting illness, as it was defined here, would exclude 277 deaths at the 5-year follow-up point and additional deaths as follow-up continued.Exclusion of subjects who died during the first 4 years of follow-up would result in a loss of 451 deaths.This number can be reduced to 219 if criteria are set to exclude participants with preexisting illness who also die during the first 4 years.Loss of statistical power and generalizability must be weighed against concern that participants who die during the first 4 years of follow-up and participants with certain illnesses at baseline may be at increased risk of having experienced weight loss during the 3 years prior to the baseline measurement.
Further study of this issue is needed to aid investigators in setting criteria for avoiding bias due to preexisting illness.Nevertheless, this research indicates that investigators should examine their data before and after exclusion of the first 4 years of deaths to assure themselves that inclusion of those events is not biasing their results.Exclusion of more than 4 years of deaths does not appear to be useful.In addition, the large deviation in mean BMI change prior to baseline among participants who died during the first year of follow-up and the fact that over 10 percent of those deceased subjects were misclassified as normal weight rather than overweight lead us to recommend that deaths occurring in the first year of follow-up be routinely excluded from studies of the BMI-mortality association.

FIGURE 1 .
FIGURE 1. Scenarios illustrating possible associations between disease, body weight, and mortality.

FIGURE 2 .
FIGURE 2. Adjusted mean change in body mass index (BMI) in the 3 years prior to baseline among participants who died (white bars) compared with those who remained alive (shaded bars) at the end of the designated year of follow-up, given that they were alive at the beginning of the year, Atherosclerosis Risk in Communities (ARIC) Study cohort, 1987-1996.Covariates included age, ethnicity, gender, study center, body mass index, education, and smoking.

FIGURE 3 .
FIGURE 3. Adjusted mean change in body mass index (BMI) in the 3 years prior to baseline among participants who were healthy at baseline compared with those who had preexisting illness (top) and among participants who remained healthy or alive after 4 years of follow-up compared with those who had preexisting illness and died within 4 years of follow-up (bottom), Atherosclerosis Risk in Communities (ARIC) Study cohort, 1987-1993.Covariates included age, ethnicity, gender, study center, body mass index, education, and smoking.

TABLE 2 .
Percentages of participants with specified changes in body mass index* prior to baseline and adjusted † odds ratios for changes in body mass index, Atherosclerosis Risk in Communities (ARIC) Study cohort, 1987-1996 Percentage Overweight Stable at <25 or >25 Gain from <25 to ≥25 Loss from ≥25 to <25 Obesity Stable at <30 or >30 Gain from <30 to ≥30 Loss from ≥30 to <30 Odds ratio Overweight Gain from <25 to ≥25 Loss from ≥25 to <25 Obesity Gain from <30 to ≥30 Loss from ≥30 to <30 Change in body mass index Died in years 1-4 (n = 451)