This study shows that SRH changed with increasing age, but age affected the thinnest subjects more than all others. Underweight was associated with a higher mortality risk, and weight change increased this risk. Several authors argue that the health consequences of being underweight might be more severe in terms of premature mortality and quality of life, compared to being overweight [8,9,10, 16,17,18, 36,37,38]. Although being thin is an ideal, and despite the general advice for obese people to lose weight, our study indicates that being thin could in itself represent a health risk later in life. Inverse probability weighting (IPW) showed that loss to follow-up did not significantly affect this result, which supports the causal claim that being thin drives the later decline. It is interesting that normalizing the weight resulted in fewer comorbid conditions and lower mortality rates, and yet, lower SRH. SRH is strongly associated with biological factors, however, it is also known to be sensitive to factors such as coping skills, as well as mental distress, and social context . One plausible explanation could be that general health advice is focused on losing weight, while normalizing the weight for underweight persons will imply gaining weight. This study cannot answer questions regarding health beliefs, but this could be addressed using qualitative methods.
Previous studies have shown that persons with BMI of <18.5 kg/m2 and ≥30 kg/m2 report impaired quality of life and SRH [40, 41]. Our study shows that this depends on age. Having a BMI above 30 kg/m2 was never beneficial for SRH as compared to the normal BMI range, while those with BMI below 18.5 kg/m2 started out with more beneficial levels but had a far more negative SRH trajectory as age increased. For underweight subjects, we see a positive gap in SRH scores compared to the other categories, but they fell more rapidly with age. Consequently, a low BMI was slightly beneficial at a young age, but it represented an increasingly negative factor with age, passing below the reference category at age 38 and even below the trajectory of obese subjects at age 67 in the fully fitted model.
Previous studies have shown that subjects who reported being obese in young adulthood only or in both young and middle adulthood experienced mortality rates that were 40%–90% higher than those subjects who were non-obese at either time . Our study showed a similar association, but even higher rates for those who were very thin, with a 69% higher all-cause death risk (1.69, 95% CI: 1.38-2.06) as compared to only 12% for obese subjects (1.12, 95% CI: 1.02-1.23). The estimate is based on the baseline BMI at study entry. It indicates that starting as thin represented a mortality risk even though the body weight may change later in life, and although reaching a normal weight was beneficial for them.
The most underweight subjects were females among the youngest (<30 years) and oldest (>70 years) subjects. This is consistent with the difference we found between the group that started out as underweight and those who ended up as underweight. A Swedish study concluded that persons aged 75-90 who were overweight had a lower mortality risk than old persons with a BMI below 25 . This is consistent with our findings for persons aged >67. Our study further shows that underweight in young age also represented a risk, even though underweight in younger adults is more likely to be associated with other medical issues (e.g. % fat mass), while in older adults underweight BMI could be an indicator of sarcopenia , malnutrition or other clinical conditions [6, 12].
Our models examined how body mass affected two different aspects of health. Mortality concerns survival, while SRH reflects the quality of survival. Being underweight affected both mortality and SRH negatively, while gaining weight had a negative association with all-cause mortality but might be beneficial for SRH. Weight loss is known to have a potential negative effect on health [13, 16,17,18,19, 21, 22]. Hence, it is reasonable to ask if the findings of our study are more a question of becoming, rather than being, underweight. The negative trajectory is explained as a combination of within-subject effects (i.e. becoming underweight) and between-subject effects (i.e. being underweight). It implies that changing BMI category (i.e. adding weight for those who were underweight) can affect SRH. However, when examining the effect of weight change in individual subjects, we find that weight gain was associated with lower SRH levels, even for those with a very low body mass. Although we would advise overweight persons to lose weight towards the normal range, gaining weight as general advice for underweight persons does not necessarily lead to better SRH.
Mortality was not affected by weight change in the same manner. The baseline BMI predicted a 69% increased risk, and when we modeled weight change by updating BMI values, we found that the risk increased to 79%, while for those in the lower normal range, the risk increased from 14% to 19%. This is consistent with a previous study of the Tromsø study cohort that showed that weight loss and gain were associated with increased all-cause mortality for men and in the subgroup of women who reported no weight-loss attempts .
Our study suggests that there is a healthy and an unhealthy underweight group, and that thinness due to waste is a risk. It seems plausible to distinguish between those who were initially underweight and those who became underweight (e.g. due to malnutrition). There are studies of nutrition and BMI status among older persons , but there are no studies that identify subgroups at risk among younger subjects. Our models control for known physical diseases, but not for eating disorders. Further studies should try to identify subgroups at particular risk among underweight persons. Health-related behavior explained 17% of the variance in our data, while gender and age (21%), comorbidity (23%) and mental distress (28%) are therefore important factors for an understanding of the decline. Our study suggests following up particularly gender differences, smoking status, the effect of weight change and development of comorbid diseases later in life.
We find more persons who exercise among the underweight subjects. This is consistent with being thin as part of a healthy life, but at the same time we find more persons who lead active lives and more non-smokers among the obese than we find from <25 kg/m2. It is timely to ask whether this might be an indication that health information is primarily targeting overweight subjects.
The Tromsø study was designed to represent a general population. Our latent trajectory model utilized 48.3% of TS 3, 45.3% of TS 4, 84.5% of TS 5 and 74.9% of TS 6. We used inverse probability weighting (IPW) to examine how the missing data affected the main finding of this study. IPW lowered the estimate for thin subjects and raised it for overweight subjects. The interaction with age attenuated, but the overall effect of age increased. The difference was that the underweight trajectory was equal in the CC and IPW models at age 25 (3.03) but ended 0.06 (3.5%) lower at age 90 in the weighted model. We see that loss to follow-up had a greater effect on the overweight (4.7%) and obesity groups (5.4%). The increased decline in the IPW model is consistent with the assumption that subjects well enough to participate several times are slightly healthier, but even so, loss to follow-up did not affect our overall results. Table 3 shows that confidence intervals (CIs) widened in the weighted model, but not substantially. No variable that was non-significant has become significant, or vice versa (except for gender and blood pressure that had CIs very close to zero). The findings thus remained basically unchanged by IPW, although there is an indication that the effect of ageing may be stronger in the general population than suggested by the CC analysis for under- and overweight subjects.
The survival analysis utilized the entire TS4 panel. Being able to use updated values reduces the bias in spite of missing data in the updated measurements. 51.3% of the participants attended only in TS4, 17.6% participated in all panels, while 31.1% reappeared in either TS5 or TS6. Using multiple imputation showed that missing data did not affect the estimates at the decimal level shown. We therefore conclude that the selection bias is within reasonable limits for both models.
It is of interest that the prevalence of underweight declined, but Tromsø 5-6 did not include participants <35 years and TS3 did not include participants >70 years, where we found most of the underweight subjects. No conclusions can therefore be drawn on the prevalence of underweight from this study. The small sample size of underweight individuals makes it difficult to fully analyze the causal claim of interest, but we can conclude that there are associations between underweight, self-reported health trajectories and mortality rates and that this relationship varies with age.
Although measured on an ordinal scale, the underlying phenomenon of SRH is continuous, and the scales represent similar logical increments. Furthermore, the distribution of SRH, apart from being staggered, resembled the shape of a normal distribution. Hence, an OLS regression model could be used for the analysis of independent associations in the multivariable model .