Biodemography of aging (I)
“The goal of this monograph is to show how questions about the connections between and among aging, health, and longevity can be addressed using the wealth of available accumulated knowledge in the field, the large volumes of genetic and non-genetic data collected in longitudinal studies, and advanced biodemographic models and analytic methods. […] This monograph visualizes aging-related changes in physiological variables and survival probabilities, describes methods, and summarizes the results of analyses of longitudinal data on aging, health, and longevity in humans performed by the group of researchers in the Biodemography of Aging Research Unit (BARU) at Duke University during the past decade. […] the focus of this monograph is studying dynamic relationships between aging, health, and longevity characteristics […] our focus on biodemography/biomedical demography meant that we needed to have an interdisciplinary and multidisciplinary biodemographic perspective spanning the fields of actuarial science, biology, economics, epidemiology, genetics, health services research, mathematics, probability, and statistics, among others.”
The quotes above are from the book‘s preface. In case this aspect was not clear from the comments above, this is the kind of book where you’ll randomly encounter sentences like these:
“The simplest model describing negative correlations between competing risks is the multivariate lognormal frailty model. We illustrate the properties of such model for the bivariate case.”
“The time-to-event sub-model specifies the latent class-specific expressions for the hazard rates conditional on the vector of biomarkers Yt and the vector of observed covariates X …”
…which means that some parts of the book are really hard to blog; it simply takes more effort to deal with this stuff here than it’s worth. As a result of this my coverage of the book will not provide a remotely ‘balanced view’ of the topics covered in it; I’ll skip a lot of the technical stuff because I don’t think it makes much sense to cover specific models and algorithms included in the book in detail here. However I should probably also emphasize while on this topic that although the book is in general not an easy read, it’s hard to read because ‘this stuff is complicated’, not because the authors are not trying. The authors in fact make it clear already in the preface that some chapters are more easy to read than are others and that some chapters are actually deliberately written as ‘guideposts and way-stations‘, as they put it, in order to make it easier for the reader to find the stuff in which he or she is most interested (“the interested reader can focus directly on the chapters/sections of greatest interest without having to read the entire volume“) – they have definitely given readability aspects some thought, and I very much like the book so far; it’s full of great stuff and it’s very well written.
I have had occasion to question a few of the observations they’ve made, for example I was a bit skeptical about a few of the conclusions they drew in chapter 6 (‘Medical Cost Trajectories and Onset of Age-Associated Diseases’), but this was related to what some would certainly consider to be minor details. In the chapter they describe a model of medical cost trajectories where the post-diagnosis follow-up period is 20 months; this is in my view much too short a follow-up period to draw conclusions about medical cost trajectories in the context of type 2 diabetes, one of the diseases included in the model, which I know because I’m intimately familiar with the literature on that topic; you need to look 7-10 years ahead to get a proper sense of how this variable develops over time – and it really is highly relevant to include those later years, because if you do not you may miss out on a large proportion of the total cost given that a substantial proportion of the total cost of diabetes relate to complications which tend to take some years to develop. If your cost analysis is based on a follow-up period as short as that of that model you may also on a related note draw faulty conclusions about which medical procedures and -subsidies are sensible/cost effective in the setting of these patients, because highly adherent patients may be significantly more expensive in a short run analysis like this one (they show up to their medical appointments and take their medications…) but much cheaper in the long run (…because they take their medications they don’t go blind or develop kidney failure). But as I say, it’s a minor point – this was one condition out of 20 included in the analysis they present, and if they’d addressed all the things that pedants like me might take issue with, the book would be twice as long and it would likely no longer be readable. Relatedly, the model they discuss in that chapter is far from unsalvageable; it’s just that one of the components of interest – ‘the difference between post- and pre-diagnosis cost levels associated with an acquired comorbidity’ – in the case of at least one disease is highly unlikely to be correct (given the authors’ interpretation of the variable), because there’s some stuff of relevance which the model does not include. I found the model quite interesting, despite the shortcomings, and the results were definitely surprising. (No, the above does not in my opinion count as an example of coverage of a ‘specific model […] in detail’. Or maybe it does, but I included no equations. On reflection I probably can’t promise much more than that, sometimes the details are interesting…)
Anyway, below I’ve added some quotes from the first few chapters of the book and a few remarks along the way.
“The genetics of aging, longevity, and mortality has become the subject of intensive analyses […]. However, most estimates of genetic effects on longevity in GWAS have not reached genome-wide statistical significance (after applying the Bonferroni correction for multiple testing) and many findings remain non-replicated. Possible reasons for slow progress in this field include the lack of a biologically-based conceptual framework that would drive development of statistical models and methods for genetic analyses of data [here I was reminded of Burnham & Anderson’s coverage, in particular their criticism of mindless ‘Let the computer find out’-strategies – the authors of that chapter seem to share their skepticism…], the presence of hidden genetic heterogeneity, the collective influence of many genetic factors (each with small effects), the effects of rare alleles, and epigenetic effects, as well as molecular biological mechanisms regulating cellular functions. […] Decades of studies of candidate genes show that they are not linked to aging-related traits in a straightforward fashion (Finch and Tanzi 1997; Martin 2007). Recent genome-wide association studies (GWAS) have supported this finding by showing that the traits in late life are likely controlled by a relatively large number of common genetic variants […]. Further, GWAS often show that the detected associations are of tiny size (Stranger et al. 2011).”
I think this ties in well with what I’ve previously read on these and related topics – see e.g. the second-last paragraph quoted in my coverage of Richard Alexander’s book, or some of the remarks included in Roberts et al. Anyway, moving on:
“It is well known from epidemiology that values of variables describing physiological states at a given age are associated with human morbidity and mortality risks. Much less well known are the facts that not only the values of these variables at a given age, but also characteristics of their dynamic behavior during the life course are also associated with health and survival outcomes. This chapter [chapter 8 in the book, US] shows that, for monotonically changing variables, the value at age 40 (intercept), the rate of change (slope), and the variability of a physiological variable, at ages 40–60, significantly influence both health-span and longevity after age 60. For non-monotonically changing variables, the age at maximum, the maximum value, the rate of decline after reaching the maximum (right slope), and the variability in the variable over the life course may influence health-span and longevity. This indicates that such characteristics can be important targets for preventive measures aiming to postpone onsets of complex diseases and increase longevity.”
The chapter from which the quotes in the next two paragraphs are taken was completely filled with data from the Framingham Heart Study, and it was hard for me to know what to include here and what to leave out – so you should probably just consider the stuff I’ve included below as samples of the sort of observations included in that part of the coverage.
“To mediate the influence of internal or external factors on lifespan, physiological variables have to show associations with risks of disease and death at different age intervals, or directly with lifespan. For many physiological variables, such associations have been established in epidemiological studies. These include body mass index (BMI), diastolic blood pressure (DBP), systolic blood pressure (SBP), pulse pressure (PP), blood glucose (BG), serum cholesterol (SCH), hematocrit (H), and ventricular rate (VR). […] the connection between BMI and mortality risk is generally J-shaped […] Although all age patterns of physiological indices are non-monotonic functions of age, blood glucose (BG) and pulse pressure (PP) can be well approximated by monotonically increasing functions for both genders. […] the average values of body mass index (BMI) increase with age (up to age 55 for males and 65 for females), and then decline for both sexes. These values do not change much between ages 50 and 70 for males and between ages 60 and 70 for females. […] Except for blood glucose, all average age trajectories of physiological indices differ between males and females. Statistical analysis confirms the significance of these differences. In particular, after age 35 the female BMI increases faster than that of males. […] [When comparing women with less than or equal to 11 years of education [‘LE’] to women with 12 or more years of education [HE]:] The average values of BG for both groups are about the same until age 45. Then the BG curve for the LE females becomes higher than that of the HE females until age 85 where the curves intersect. […] The average values of BMI in the LE group are substantially higher than those among the HE group over the entire age interval. […] The average values of BG for the HE and LE males are very similar […] However, the differences between groups are much smaller than for females.”
They also in the chapter compared individuals with short life-spans [‘SL’, died before the age of 75] and those with long life-spans [‘LL’, 100 longest-living individuals in the relevant sample] to see if the variables/trajectories looked different. They did, for example: “trajectories for the LL females are substantially different from those for the SL females in all eight indices. Specifically, the average values of BG are higher and increase faster in the SL females. The entire age trajectory of BMI for the LL females is shifted to the right […] The average values of DBP [diastolic blood pressure, US] among the SL females are higher […] A particularly notable observation is the shift of the entire age trajectory of BMI for the LL males and females to the right (towards an older age), as compared with the SL group, and achieving its maximum at a later age. Such a pattern is markedly different from that for healthy and unhealthy individuals. The latter is mostly characterized by the higher values of BMI for the unhealthy people, while it has similar ages at maximum for both the healthy and unhealthy groups. […] Physiological aging changes usually develop in the presence of other factors affecting physiological dynamics and morbidity/mortality risks. Among these other factors are year of birth, gender, education, income, occupation, smoking, and alcohol use. An important limitation of most longitudinal studies is the lack of information regarding external disturbances affecting individuals in their day-today life.”
I incidentally noted while I was reading that chapter that a relevant variable ‘lurking in the shadows’ in the context of the male and female BMI trajectories might be changing smoking habits over time; I have not looked at US data on this topic, but I do know that the smoking patterns of Danish males and females during the latter half of the last century were markedly different and changed really quite dramatically in just a few decades; a lot more males than females smoked in the 60es, whereas the proportions of male- and female smokers today are much more similar, because a lot of males have given up smoking (I refer Danish readers to this blog post which I wrote some years ago on these topics). The authors of the chapter incidentally do look a little at data on smokers and they observe that smokers’ BMI are lower than non-smokers (not surprising), and that the smokers’ BMI curve (displaying the relationship between BMI and age) grows at a slower rate than the BMI curve of non-smokers (that this was to be expected is perhaps less clear, at least to me – the authors don’t interpret these specific numbers, they just report them).
“To better address the challenge of “healthy aging” and to reduce economic burdens of aging-related diseases, key factors driving the onset and progression of diseases in older adults must be identified and evaluated. An identification of disease-specific age patterns with sufficient precision requires large databases that include various age-specific population groups. Collections of such datasets are costly and require long periods of time. That is why few studies have investigated disease-specific age patterns among older U.S. adults and there is limited knowledge of factors impacting these patterns. […] Information collected in U.S. Medicare Files of Service Use (MFSU) for the entire Medicare-eligible population of older U.S. adults can serve as an example of observational administrative data that can be used for analysis of disease-specific age patterns. […] In this chapter, we focus on a series of epidemiologic and biodemographic characteristics that can be studied using MFSU.”
“Two datasets capable of generating national level estimates for older U.S. adults are the Surveillance, Epidemiology, and End Results (SEER) Registry data linked to MFSU (SEER-M) and the National Long Term Care Survey (NLTCS), also linked to MFSU (NLTCS-M). […] The SEER-M data are the primary dataset analyzed in this chapter. The expanded SEER registry covers approximately 26 % of the U.S. population. In total, the Medicare records for 2,154,598 individuals are available in SEER-M […] For the majority of persons, we have continuous records of Medicare services use from 1991 (or from the time the person reached age 65 after 1990) to his/her death. […] The NLTCS-M data contain two of the six waves of the NLTCS: namely, the cohorts of years 1994 and 1999. […] In total, 34,077 individuals were followed-up between 1994 and 1999. These individuals were given the detailed NLTCS interview […] which has information on risk factors. More than 200 variables were selected”
In short, these data sets are very large, and contain a lot of information. Here are some results/data:
“Among studied diseases, incidence rates of Alzheimer’s disease, stroke, and heart failure increased with age, while the rates of lung and breast cancers, angina pectoris, diabetes, asthma, emphysema, arthritis, and goiter became lower at advanced ages. [..] Several types of age-patterns of disease incidence could be described. The first was a monotonic increase until age 85–95, with a subsequent slowing down, leveling off, and decline at age 100. This pattern was observed for myocardial infarction, stroke, heart failure, ulcer, and Alzheimer’s disease. The second type had an earlier-age maximum and a more symmetric shape (i.e., an inverted U-shape) which was observed for lung and colon cancers, Parkinson’s disease, and renal failure. The majority of diseases (e.g., prostate cancer, asthma, and diabetes mellitus among them) demonstrated a third shape: a monotonic decline with age or a decline after a short period of increased rates. […] The occurrence of age-patterns with a maximum and, especially, with a monotonic decline contradicts the hypothesis that the risk of geriatric diseases correlates with an accumulation of adverse health events […]. Two processes could be operative in the generation of such shapes. First, they could be attributed to the effect of selection […] when frail individuals do not survive to advanced ages. This approach is popular in cancer modeling […] The second explanation could be related to the possibility of under-diagnosis of certain chronic diseases at advanced ages (due to both less pronounced disease symptoms and infrequent doctor’s office visits); however, that possibility cannot be assessed with the available data […this is because the data sets are based on Medicare claims – US]”
“The most detailed U.S. data on cancer incidence come from the SEER Registry […] about 60 % of malignancies are diagnosed in persons aged 65+ years old […] In the U.S., the estimated percent of cancer patients alive after being diagnosed with cancer (in 2008, by current age) was 13 % for those aged 65–69, 25 % for ages 70–79, and 22 % for ages 80+ years old (compared with 40 % of those aged younger than 65 years old) […] Diabetes affects about 21 % of the U.S. population aged 65+ years old (McDonald et al. 2009). However, while more is known about the prevalence of diabetes, the incidence of this disease among older adults is less studied. […] [In multiple previous studies] the incidence rates of diabetes decreased with age for both males and females. In the present study, we find similar patterns […] The prevalence of asthma among the U.S. population aged 65+ years old in the mid-2000s was as high as 7 % […] older patients are more likely to be underdiagnosed, untreated, and hospitalized due to asthma than individuals younger than age 65 […] asthma incidence rates have been shown to decrease with age […] This trend of declining asthma incidence with age is in agreement with our results.”
“The prevalence and incidence of Alzheimer’s disease increase exponentially with age, with the most notable rise occurring through the seventh and eight decades of life (Reitz et al. 2011). […] whereas dementia incidence continues to increase beyond age 85, the rate of increase slows down [which] suggests that dementia diagnosed at advanced ages might be related not to the aging process per se, but associated with age-related risk factors […] Approximately 1–2 % of the population aged 65+ and up to 3–5 % aged 85+ years old suffer from Parkinson’s disease […] There are few studies of Parkinson’s disease incidence, especially in the oldest old, and its age patterns at advanced ages remain controversial”.
“One disadvantage of large administrative databases is that certain factors can produce systematic over/underestimation of the number of diagnosed diseases or of identification of the age at disease onset. One reason for such uncertainties is an incorrect date of disease onset. Other sources are latent disenrollment and the effects of study design. […] the date of onset of a certain chronic disease is a quantity which is not defined as precisely as mortality. This uncertainty makes difficult the construction of a unified definition of the date of onset appropriate for population studies.”
“[W]e investigated the phenomenon of multimorbidity in the U.S. elderly population by analyzing mutual dependence in disease risks, i.e., we calculated disease risks for individuals with specific pre-existing conditions […]. In total, 420 pairs of diseases were analyzed. […] For each pair, we calculated age patterns of unconditional incidence rates of the diseases, conditional rates of the second (later manifested) disease for individuals after onset of the first (earlier manifested) disease, and the hazard ratio of development of the subsequent disease in the presence (or not) of the first disease. […] three groups of interrelations were identified: (i) diseases whose risk became much higher when patients had a certain pre-existing (earlier diagnosed) disease; (ii) diseases whose risk became lower than in the general population when patients had certain pre-existing conditions […] and (iii) diseases for which “two-tail” effects were observed: i.e., when the effects are significant for both orders of disease precedence; both effects can be direct (either one of the diseases from a disease pair increases the risk of the other disease), inverse (either one of the diseases from a disease pair decreases the risk of the other disease), or controversial (one disease increases the risk of the other, but the other disease decreases the risk of the first disease from the disease pair). In general, the majority of disease pairs with increased risk of the later diagnosed disease in both orders of precedence were those in which both the pre-existing and later occurring diseases were cancers, and also when both diseases were of the same organ. […] Generally, the effect of dependence between risks of two diseases diminishes with advancing age. […] Identifying mutual relationships in age-associated disease risks is extremely important since they indicate that development of […] diseases may involve common biological mechanisms.”
“in population cohorts, trends in prevalence result from combinations of trends in incidence, population at risk, recovery, and patients’ survival rates. Trends in the rates for one disease also may depend on trends in concurrent diseases, e.g., increasing survival from CHD contributes to an increase in the cancer incidence rate if the individuals who survived were initially susceptible to both diseases.”
No comments yet.