Men Sustain Higher Dysregulation Levels Than Women Without Becoming Frail

Abstract

The aging process differs in important ways between the sexes, with women living longer but at higher risk for frailty (the male–female health-survival paradox). The underlying biological mechanisms remain poorly understood, but may relate to sex differences in physiological dysregulation patterns. Here, using biomarkers from two longitudinal cohort studies (InCHIANTI and BLSA) and one cross-sectional survey (NHANES), we assess sex differences in trajectories of dysregulation globally and for five physiological systems: oxygen transport, electrolytes, hematopoiesis, lipids, and liver/kidney function. We found higher dysregulation levels in men, both globally and in the oxygen transport and hematopoietic systems (p < .001 for all), though differences for other systems were mixed (electrolytes) or absent (lipids and liver/kidney). There was no clear evidence for sex differences in rates of change in dysregulation with age. Although risk of frailty and mortality increase with dysregulation, there was no evidence for differences in these effects between sexes. These findings imply that the greater susceptibility of women to frailty is not simply due to a tolerance for higher dysregulation; rather, it may actually be men that have a greater tolerance for dysregulation, creating a male–female dysregulation-frailty paradox. However, the precise physiological mechanisms underlying the sex differences appear to be diffuse and hard to pin down.

Worldwide, women outlive men (1). While it may possible to understand sex differences in life expectancy by decomposing mortality into causes (eg, heart disease kills more men and strikes earlier than some other major causes) (2), women show lower mortality rates for 12 of 15 leading causes of death in the United States, with a higher mortality rate only for Alzheimer’s disease (3). Health behaviors may be part of the answer (eg, women’s propensity to choose healthier food) (4), but are unlikely to fully explain the difference. Despite lower mortality rates, women are more susceptible to becoming frail with aging and become frail younger (5), but are better able to survive in a frail state for a long time (6–8). These findings have been corroborated across diverse populations from all around the world, including Chinese, Mexicans, Europeans, North Americans, and Australians (7). Even when taking frailty criteria one at a time, women show higher incidence than men for all but one (9).

This leads to the characterization of male aging as more “catastrophic” than female aging, in the sense that men tend to experience more rapid declines than women. This male–female health-survival paradox (MFHSP) has been broadly replicated (7), even at age 85 and over (10). Moreover, the MFHSP was found in a Mormon population (11), a population in which women have higher fertility rates (and thus a higher mortality risk) and men have fewer unhealthy behaviors, supporting the existence of underlying biological mechanisms rather than solely lifestyle or behavioral causes. Yet, this mechanistic understanding is still lacking (12).

More generally, age-related changes in biomarkers, as well as their associations with mortality risk and frailty, may be sex-specific (13–16), and a number of studies have noted sex differences in age-related changes of biomarkers (17,18), certain mechanisms (oxidative stress (19), telomeres (20), etc.), as well as in immune response (reviewed in ref. 21). For instance, while inflammatory markers are better predictors of frailty incidence for women (15), adiponectin (a marker of energy metabolism) is associated with the number of frailty criteria in men but not women (16). Women also tend to have higher levels of inflammatory markers, but a slower increase with age, compared to men (22). Metabolomic data also show clear evidence for sex differences in aging (23). However, a study on sex differences in blood biomarkers from Russia (a country where the MFHSP is pronounced) showed mixed results, specifically with regard to lipids (24). Other studies aimed to dissect the MFHSP through different measures of physiological state. Using three summary indices of dysregulation, including allostatic load, Yang and Kozloski demonstrated that women sustain higher dysregulation levels than men (22). In contrast, results from the epigenetic clock revealed higher aging rates in men compared to women in three tissues (blood, brain, and saliva) and multiple racial groups, including Tsimane Amerindians and rainforest hunter-gatherers (25). These varied results are hard to interpret without a framework to understand how individual mechanisms relate to aging more generally.

At a mechanistic level, there appears to be clear evidence for a role of sex hormones in determining sex differences in aging, beyond the tautological sense implied by the developmental role of these hormones. Women who undergo early menopause have more “male-like” health and mortality profiles during aging (26,27), and testosterone levels were associated with all-cause mortality risk in older men, whereas estradiol levels were not (28). Nonetheless, attributing the difference to sex hormones seems broadly inadequate absent an understanding of what the hormones might change downstream, such as oxidative stress or immune function. While some evidence suggests that oxidative stress might differ between males and females (12), patterns of these differences do not always agree with the hypothesis (12), and oxidative stress is no longer considered a sufficient mechanistic explanation for aging (29). The process of immunosenescence does differ markedly between men and women (21). Both estrogens and androgens modulate the immune system, and these hormones also change with age and menopause. The numerous male–female differences in immune function during aging could explain the different mortality and frailty patterns, but not all the differences favor male mortality, and sex hormones have many effects beyond the immune system. Much more work would be needed to fully support or refute this hypothesis.

We have recently developed an approach to measure physiological dysregulation during aging by integrating the signal of many clinical biomarkers using Mahalanobis distance (D_M) (30), which quantifies how aberrant an individual’s biomarker profile is relative to the norm (31). D_M increases with age within individuals, predicts numerous health outcomes after controlling for age, including both frailty and mortality (17,32), and can be measured in specific physiological systems (33). Together, these results imply that dysregulation of specific systems is an important part of the aging process, with or without additional upstream causes. Such dysregulation might underlie men’s accelerated mortality relative to women, or women’s increased risk of frailty relative to men, though presumably not both at the same time. A previous study integrating this dysregulation measure into stochastic process models showed that women deviate farther from this physiological norm with age, but the mortality risk associated with dysregulation is stronger in men and increases more quickly with age (34).

Understanding how dysregulation of different systems differs between men and women and relates to frailty and mortality risk could help clarify the physiological mechanisms underlying the MFHSP. Here, we used D_M to assess how dysregulation differs between sexes during aging by comparing dysregulation levels and trajectories with age in older men and women. Specifically, we wanted to test two plausible hypotheses for the link between dysregulation and the MFHSP. First, if dysregulation underlies frailty, women should have higher dysregulation levels and/or rates, globally and across systems. This would agree with the existing hypothesis that women are more robust, that is, able to survive levels of dysregulation that would kill a man, though in a weakened state. Alternatively, perhaps men and women differ in which systems become dysregulated. Under this scenario, perhaps the systems that tend to dysregulate faster in men generally lead to mortality, whereas the systems that dysregulate faster in women lead to frailty. We thus measured D_M globally and in five physiological systems to be able to test these hypotheses.

Method

Data Sets

We used data from two study populations and one cross-sectional survey (see Supplementary Method for details): Invecchiare in Chianti (InCHIANTI) (35), the Baltimore Longitudinal Study of Aging (BLSA) (36), and the National Health and Nutrition Examination Survey (NHANES) (37). All aspects of research were approved by the ethics committees at the institutions responsible for data collection, or by the National Institute of Environmental Health Services Internal Review Board for BLSA, and this secondary analysis was approved by the ethics committee (Comité d’éthique de la recherche en santé chez l’humain) at the Centre de recherche clinique du CHUS, project # 14–059. Participants signed informed consent for data collection and analysis. Although the data used in these analyses cannot be freely shared due to confidentiality constraints related to human medical data, they are all available to researchers submitting an appropriate research proposal: InCHIANTI at http://www.inchiantistudy.net/obtain_data.html and BLSA at http://www.blsa.nih.gov/researchers. For all data sets, we used only individuals aged 55 years or older. This was to avoid problems with different age distributions across data sets (few young in InCHIANTI and BLSA) and because D_M increases nonlinearly with age; starting at age 55 allowed us to use linear models as a reasonable approximation, facilitating interpretation.

Biomarker Selection

Biomarkers were chosen based on availability in sufficient sample size across the three studies, though previous studies have shown limited sensitivity of D_M to precise choice of biomarkers (17,34). We used a set of 30 biomarkers, divided into five physiological systems as validated in a previous publication (33) (see Supplementary Table 1 and Supplementary Figure 1): (a) oxygen transport, which includes ferritin, hemoglobin, hematocrit, iron, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, red blood cell count, and red cell distribution width; (b) hematopoiesis, which includes basophils (%), lymphocytes (%), monocytes (%), leukocytes (%), and white blood cell count; (c) electrolytes, which includes calcium, chloride, potassium, and sodium; (d) lipids, which includes triglycerides, high-density lipoprotein, low-density lipoprotein, cholesterol, and cholesterol to high-density lipoprotein ratio; and (e) liver and kidney function (“liver/kidney”), which includes albumin, alkaline phosphatase, alanine transaminase, aspartate transaminase, γ-glutamyl transpeptidase, lactate dehydrogenase, total protein, and uric acid.

Dysregulation Scores

Age at each visit was calculated in years (decimal format to the nearest day, except for NHANES for which only age in years was available). Dysregulation scores were calculated, both globally with all 30 biomarkers and by physiological system, as described elsewhere (33). Briefly, D_M applied to biomarkers is a measure of how aberrant an individual’s profile is relative to everyone else in the population, and greater distance should thus measure greater dysregulation. Our previous work indicates that a younger and healthier reference population to calculate the centroid generally provides a better signal of dysregulation (17). Hence, to construct our reference population, we used an equal number (n = 421) of our youngest subjects (under 70 years old) from each data set. The reference population serves as the centroid in D_M calculation and to standardize the biomarkers (see below); subjects included in the reference population are not excluded from subsequent analyses. D_M was calculated with all data sets combined to ensure a homogeneous scale for comparisons across data sets. The logarithm of this dysregulation score, divided by the standard deviation, was used as the key variable in the analyses. Before D_M calculation, all biomarkers were log- or square-root-transformed as necessary to approach normality (using the same transformation for all data sets), and were centered at the mean of the reference group and divided by the standard deviation of the reference group.

Data Analysis

Overview

Because some of our analyses are statistically technical, we provide a conceptual overview here. We wanted to model how D_M changes with age, and to do this, we used regression models with D_M as a linear function of age. Actual increase appears to be exponential (32), but nonlinear models are hard to fit for technical reasons, and also harder to interpret, so we chose to use the linear model, which is a reasonable approximation for the older cohort used here. Within our longitudinal data sets, we were able to do this using multilevel models that permit us to estimate slopes of change for individuals over time, and thus to estimate an average individual slope that represents rate of increase in dysregulation with age. The key parameters we sought to estimate were (a) the impact of sex on baseline levels (intercept) of D_M (Table 1) and (b) the impact of sex on the rate of change (slope) of D_M (Table 2). The former is estimated via the sex term in the model, and the latter via the sex-age interaction term. Interaction terms permit for a formal test of the difference between men and women, which is not possible with stratified analyses. We nonetheless present stratified analyses in figures as a visual indication of the magnitude of the differences. We replicated this approach in different data sets and a merged data set, as well as in different age subsets. We then used regression models to assess whether sex influenced the effect of D_M on either mortality or frailty, again using interaction terms. Many of the technical details below are to harmonize analyses across data sets with different structures.

Table 1.

Effect of Being Male on Dysregulation Level

	Allc		Young (55–74)		Old (75+)
	βb	p	β	p	β	p
Global
Alla	0.172	<.0001	0.185	<.0001	0.159	.0005
InCHIANTI	0.243	<.0001	0.161	.01	0.269	<.0001
BLSA	0.083	.14	0.102	.13	0.032	.73
NHANES	0.182	<.0001	0.230	<.0001	0.092	.30
Oxygen transport
All	0.175	<.0001	0.211	<.0001	0.137	.0006
InCHIANTI	0.335	<.0001	0.301	<.0001	0.337	<.0001
BLSA	−0.081	.16	−0.055	.44	−0.106	.25
NHANES	0.190	<.0001	0.245	<.0001	0.074	.22
Electrolytes
All	−0.018	.25	−0.009	.62	−0.032	.25
InCHIANTI	0.050	.23	−0.044	.46	0.119	.02
BLSA	0.043	.34	0.012	.82	0.104	.17
NHANES	−0.039	.03	−0.008	.72	−0.108	.002
Hematopoiesis
All	0.220	<.0001	0.206	<.0001	0.237	<.0001
InCHIANTI	0.029	.51	0.017	.78	0.038	.51
BLSA	0.283	<.0001	0.241	.0008	0.356	.0001
NHANES	0.235	<.0001	0.223	<.0001	0.265	<.0001
Lipids
All	0.000	.99	0.001	.99	0.010	.77
InCHIANTI	0.039	.45	0.015	.83	0.080	.20
BLSA	0.040	.44	−0.017	.80	0.089	.28
NHANES	−0.035	.20	−0.002	.96	−0.112	.02
Liver/kidney
All	0.012	.47	0.024	.24	−0.008	.75
InCHIANTI	−0.002	.96	−0.059	.34	0.031	.60
BLSA	0.064	.14	0.085	.11	0.027	.71
NHANES	0.014	.48	0.024	.30	−0.008	.80

Note: ^aRegressions were performed using all individuals (n = 13,909), or stratifying by data set: InCHIANTI (n = 1,147), BLSA (n = 1,024), and NHANES (n = 11,738). Positive β coefficients indicate higher dysregulation in males in the population indicated by the column (age subset) and row (data set). ^bThe table shows regression coefficient estimates (β) adjusted for age, with the corresponding p value. Positive coefficients indicate higher dysregulation in men. ^cAnalyses were repeated with different age strata: with all individuals (n ranging from 993 to 13,830 depending on data sets included and physiological systems analyzed), with younger individuals (between 55 and 74 years old; n ranging from 686 to 9,518), with older individuals (aged 75 years and over; n ranging from 401 to 4,823). Values with p < .05 are shown in bold.

Table 2.

Effect of Being Male on Dysregulation Rate

	Allc		Young (55–74)		Old (75+)
	βb	p	β	p	β	p
Global
Alla	0.002	.55	−0.002	.75	−0.003	.72
InCHIANTI	0.004	.48	−0.008	.46	0.000	.97
BLSA	−0.002	.72	−0.008	.47	0.014	.31
Oxygen transport
All	0.002	.57	0.003	.70	−0.009	.30
InCHIANTI	−0.001	.86	−0.011	.30	−0.001	.91
BLSA	−0.003	.60	−0.005	.63	−0.004	.75
Electrolytes
All	0.002	.52	−0.010	.15	−0.004	.62
InCHIANTI	0.007	.19	−0.006	.63	−0.002	.85
BLSA	0.000	.92	−0.010	.25	−0.008	.55
Hematopoiesis
All	0.008	.04	0.010	.18	0.015	.07
InCHIANTI	0.006	.23	0.018	.11	0.011	.30
BLSA	0.008	.15	0.010	.37	0.006	.66
Lipids
All	0.006	.09	−0.003	.67	0.008	.31
InCHIANTI	−0.002	.78	−0.009	.44	−0.007	.52
BLSA	0.009	.08	−0.006	.55	0.036	.007
Liver/kidney
All	0.001	.85	0.001	.87	−0.001	.92
InCHIANTI	0.008	.12	0.009	.42	0.002	.85
BLSA	−0.001	.89	−0.002	.86	0.014	.26

Note: ^aRegression were performed using all available subjects (n = 2,171), or stratifying by data set: InCHIANTI (n = 1,147) and BLSA (n = 1,024). Positive β coefficients indicate higher dysregulation slopes (faster increases) in males in the population indicated by the column (age subset) and row (data set). ^bThe table shows regression coefficient estimates (β) of an interaction term between age and sex, with the corresponding p value. Positive coefficients indicate faster dysregulation in men. ^cAnalyses were repeated with different age strata: with all individuals (n ranging from 993 to 2,169 depending on data sets included and physiological systems analyzed), with younger individuals (between 55 and 75 years old; n ranging from 686 to 1,412), with older individuals (aged 75 and over; n ranging from 401 to 1,269). Values with p < .05 are shown in bold.

Details

To assess the effect of sex on D_M level, we included sex as a covariate either in linear mixed-effects models of D_M changes with age using the lmer function (lme4 package) with individual as a random effect (longitudinal data sets), or in linear models (lm function) for cross-sectional data. Analyses were performed with all data sets included in the same model (All) or stratifying by data set (Table 1), and were repeated using age subsets: a younger subset (“young”; individuals aged 55–74 years), or an older subset (“old”; individuals aged 75 years and over). Results obtained with a model controlling for age with a cubic spline (bs function from the fda package) were nearly identical (Supplementary Table 2). To measure the effect of sex on yearly changes in D_M with age (individual slope), we included an interaction term of sex with age in linear mixed-effects models (lmer function), with individual as a random effect (Table 2). Again, analyses were performed by merging or stratifying data sets (in this case longitudinal data sets; InCHIANTI and BLSA), and repeated for different age subsets as described above. Note that we present coefficients for effect of sex on D_M (Table 1) based on models without sex-age interaction terms to aid with interpretability of coefficients. Figure 1 shows unadjusted mean D_M levels calculated separately for each sex. When longitudinal D_M measures were available, we present the mean of all observations per subject. Figure 2 shows D_M yearly change with age calculated separately for each sex, using the lmer function. All analyses were performed in R v3.2.2. All code is available upon request.

Figure 1.

Unadjusted D_M levels by sex, globally and by physiological system. Mean D_M levels were calculated using all individuals (black; n = 13,909), or stratifying by data set: InCHIANTI (blue; n = 1,147), BLSA (red; n = 1,024), and NHANES (green; n = 11,738). For longitudinal cohorts, we used the mean of all available D_M. Analyses were repeated using all individuals (n ranging from 993 to 13,830 depending on data sets included and physiological systems analyzed), or stratifying by age at first visit: “young”, individuals between 55 and 74 years old (n ranging from 686 to 9,518), and “old”, individuals aged 75 and over (n ranging from 401 to 4,823). Variance explained by age, sex, and data set were calculated with an ANOVA.

Figure 2.

D _M rates by sex, globally and by physiological system. D_M longitudinal changes (slopes) were calculated using all available subjects (black, n = 2,171), or stratifying by data set: InCHIANTI (blue; n = 1,147) and BLSA (red; n = 1,024). Analyses were repeated using all individuals (n ranging from 993 to 2,169 depending on data sets included and physiological systems analyzed), or stratifying by age at first visit: “young”, individuals between 55 and 74 years old (n ranging from 686 to 1,412), and “old”, individuals aged 75 and over (n ranging from 401 to 1,269).

To evaluate the effect of sex on the association between D_M and mortality, we performed Cox proportional hazards regression models with the coxph function (survival package). We used D_M measures and age at first visit (with complete biomarker data) with age at last contact or death, using a counting process formulation with and without an interaction term between D_M and sex. Analyses were performed both by merging and stratifying data sets (InCHIANTI and BLSA; mortality outcome was not available for NHANES).

To evaluate the effect of sex on the association between D_M and frailty, we used a Poisson regression model (glm function) with and without an interaction term between D_M and sex, controlling for age by cubic spline (bs function) with four knots. We did not have frailty measures for BLSA and NHANES, and only at first visit for InCHIANTI. We used the number of frailty criteria (between 0 and 5), according to Fried’s phenotype definition (5), which are: (a) unintentional weight loss, (b) fatigue, (c) reduced grip strength, (d) reduced physical activity, and (e) low gait speed. Fulfilling three or more of these criteria indicates clinical frailty. We used the number of frailty criteria (0–5) rather than frailty state (Y/N) in order to maximize statistical power. Numerous previous analyses by our group have shown that results are qualitatively similar between dichotomized and count operationalizations of frailty in this context, with the latter simply providing more statistical power (17,32), despite the potential conceptual difference from the original phenotypic definition of frailty (5).

Results

Effect of Sex on Dysregulation Levels

InCHIANTI included 1,147 individuals of which 642 (56.0%) were women, and men were aged 72.6 ± 8.0 in average at baseline (55.0–94.3 years old), compared to 73.9 ± 8.6 for women (55.0–98.4 years old). BLSA included 1,024 individuals of which 496 were women (48.4%), and men were aged 71.3 ± 9.5 in average at baseline (55.0–96.6 years old), compared to 68.0 ± 10.0 for women (55.0–99.3 years old). NHANES included 11,738 individuals of which 5,921 were women (50.4%), and men were aged 69.1 ± 8.6 in average (55–85 years old), compared to 69.2 ± 8.8 for women (55–85 years old). As expected, D_M levels increase with age, as can be seen by higher means in old than young (Figure 1) and by positive slopes with age (Figure 2). Most but not all slopes were significantly positive across physiological systems, sexes, and data sets, though the lipid system does not appear to have an overall increase with age, and may even decline with age in some analyses (Figure 2).

Table 1 presents regression coefficients of sex predicting D_M, adjusting for age. To better visualize differences in D_M levels between sexes, we also present mean D_M levels by sex, unadjusted for age (Figure 1). Men had higher D_M levels, globally (β = 0.172, p < .001) and in two physiological systems, namely the oxygen transport (β = 0.175, p < .001) and hematopoiesis (β = 0.220, p < .001) systems. In each case, results could be replicated across age strata and in two of the three data sets. Total variance explained in our models was not high, but variance explained by sex was of similar magnitude to that explained by age and data set (Figure 1).

Effect of Sex on Dysregulation Rate

However, the rate of dysregulation did not show such clear associations with sex (Table 2 and Figure 2). Global dysregulation, as well as dysregulation in the oxygen transport, electrolytes, and liver/kidney systems, revealed no effect of sex for any analysis performed, whereas the other systems showed mixed results. The lipid system revealed only one significant association and had differing signs of effect between data sets, suggesting that the significant result may be an artifact of multiple testing. However, the hematopoietic system shows some trend for higher dysregulation rates in men compared to women, with similar coefficients in the two data sets (Table 2 and Figure 2), although only the model including all individuals reached statistical significance (Table 2). There is thus a possibility that men dysregulate faster than women in the hematopoietic system, but no evidence for differences in other systems. The small effect sizes (Table 2) and our ability to detect differences between data sets but not sexes (Figure 2) suggests a true absence of major sex differences in dysregulation rates, rather than insufficient power to detect important effects.

Effect of Sex on D_M Association With Mortality and Frailty

Table 3 shows mortality hazard ratios (HRs) of D_M and its interaction with sex (D_M × sex). We chose to present only results obtained by merging data sets because we lacked statistical power in the BLSA data set due to the low number of mortality events (n = 96). However, results were always replicated when using only individuals from InCHIANTI (data not shown). As expected, males had higher mortality risk, globally and in all physiological systems analyzed (Table 3). D_M was associated with higher mortality risk globally (HR = 1.13 per unit log-D_M, p < .001), as well as in three systems (oxygen transport, lipids, and liver/kidney; Table 3). However, this association appears to be unmodified by sex, with only one of six analyses showing a marginally significant interaction, a lower susceptibility to D_M in men (liver/kidney; HR = 0.83, p = .04).

Table 3.

Effect of D_M and Its Interaction With Sex on Mortality and Frailty Criteria Number (men compared to women)

Mortality	D M		Sex		D M × sex
	HRa	p	HR	p	HR	p
Global	1.13	.0006	1.75	<.0001	0.95	.46
Oxygen transport	1.15	.0009	1.77	<.0001	1.07	.44
Electrolytes	0.99	.75	1.81	<.0001	1.08	.44
Hematopoiesis	0.99	.81	1.81	<.0001	0.90	.32
Lipids	1.12	.01	1.79	<.0001	0.94	.44
Liver/kidney	1.15	.0003	1.77	<.0001	0.83	.04
Frailty	βb	p	β	p	β	p
Global	0.155	<.0001	−0.187	.01	0.010	.85
Oxygen transport	0.124	.0001	−0.183	.01	0.092	.15
Electrolytes	0.082	.02	−0.155	.03	0.032	.65
Hematopoiesis	0.047	.24	−0.144	.04	0.083	.30
Lipids	0.087	.005	−0.153	.03	0.009	.89
Liver/kidney	0.170	<.0001	−0.165	.02	−0.041	.47

Note: ^aThe table shows proportional hazard ratios (HR) of D_M, sex (being male compared to female), and the interaction term between D_M and sex (D_M × sex) obtained from a counting process Cox regression model, with their respective p value (n ranging from 1,841 to 1,872 depending on physiological system analyzed). For ease of interpretation, we present HR associated with D_M obtained from a model excluding the interaction term, but results were similar in the model including it. ^bThe table shows regression coefficient estimates (β) of D_M, sex (being male compared to female), and of an interaction term between D_M and sex (D_M × sex), and their respective p value, obtained through Poisson regression models predicting the number of frailty criteria, controlling for age. Regressions were performed using the InCHIANTI data set, for which frailty criteria number was only available at baseline (n ranging from 976 to 984 depending on physiological systems analyzed). For ease of interpretation, we present β associated with D_M and sex alone obtained from a model excluding the interaction term. Values with p < .05 are shown in bold.

Again, as expected, males had fewer frailty criteria after control for DM globally or by system (Table 3). D_M was strongly associated with number of frailty criteria globally (β = 0.150, p < .001), for the liver/kidney system (β = 0.199, p < .001), and to a lesser extent in two other systems (oxygen transport and lipids, see Table 3), though we note that we only had access to frailty data for InCHIANTI. As for mortality, sex did not appear to modify the associations between D_M and frailty, with none of the interaction terms significant. However, the large effect sizes for the interaction terms in the oxygen transport and hematopoiesis systems suggest that we may have lacked power to detect these effects. Five of six interaction coefficients were positive, leaving open the possibility that dysregulation increases risk of frailty in men more than women.

In additional mediation analyses, we used InCHIANTI to ask whether the effect of D_M on mortality was attenuated by including frailty in the model (Supplementary Table 3). Among 12 analyses (6 systems, 2 sexes), 7 show significant effects of D_M on mortality, and 6 of these show significant effects of D_M on frailty. All six of these show diminished hazard ratios for D_M on mortality once frailty is included in the model, though the magnitude of the decline varies from 0.03 (eg, HR = 1.13 → HR = 1.10) to 0.15 (eg, HR = 1.35 → HR = 1.20). This suggests that some but not all of the impact of dysregulation on mortality is attributable to its relationship with frailty, and this was equally true in both sexes.

Discussion

Our results show that physiological dysregulation levels are higher in men than women, both globally and in two of five physiological systems examined: oxygen transport and hematopoiesis. The other three systems showed no sex differences. Sex appears to be about as important as age and data set as a determinant of dysregulation levels in these systems. While replication of this result was not perfect across all data sets and age subsets, effects were strong, highly significant, and consistent in sign across systems. We were not able to detect a similar effect for dysregulation rates (ie, no faster dysregulation increases with age in men), though results were ambiguous for the hematopoiesis system.

These results appear to refute both of our a priori hypotheses. Clearly, women are not more dysregulated than men, either in levels or rates, refuting the hypothesis that women become frail rather than die because they are better able to tolerate high levels of dysregulation than men, and that this dysregulation leads directly to frailty. In other words, our findings suggest that while dysregulation is associated with frailty, it is not in itself a simple underlying physiological reflection of frailty state.

Our second hypothesis was that some systems dysregulate faster in men while others dysregulate faster in women, and that which systems dysregulate in which determines the MFHSP. While we cannot exclude the possibility that this hypothesis might have been supported had we measured many more systems, it does not seem to agree with our finding that when dysregulation is higher in one sex than the other, it is higher (or maybe faster) in men. A clear majority of our nonsignificant findings also show a trend toward more dysregulation in men, which is in line with results from the epigenetic clock (25).

This second hypothesis is also not supported by our finding that while dysregulation of most systems increases the risk of both mortality and frailty, as previously shown (33), there was no clear evidence that these effects were larger in one sex than the other. This agrees with a previous study by Fried et al. (38), which found that sex did not modify the effects of other risk factors on mortality. However, it disagrees with a previous study by Arbeev et al. (34), which found that a global measure of dysregulation based on a limited set of biomarkers has greater effects on mortality in men than women, and that this difference accelerates with age. Women also had faster dysregulation and a greater ability than men to withstand dysregulation in that study. Potential reasons for the different results of Arbeev et al. include (a) differences between the Framingham cohort and our cohorts or (b) a very different set of biomarkers in their dysregulation measure (blood pressure, heart rate, glucose, BMI, hematocrit, and total cholesterol). Arbeev et al.’s larger sample size and long follow-up permitted sophisticated statistical models that we could not apply here. However, we do not feel that lack of power in our study is sufficient to explain the difference, as we were able to detect large effects of D_M and sex on mortality and frailty (Table 3), but no modification of D_M effect by sex; additionally, the effects we detected went in the opposite direction of those in Arbeev et al. Taking together our negative result and the discrepancies across studies, patterns in global dysregulation levels do not appear to confirm either of our hypotheses concerning MFHSP.

Rather, our findings support a surprising alternative hypothesis: that MFHSP is inverted at a physiological level, with men being more robust to higher physiological dysregulation levels than women, and women robust to higher clinical frailty than men. In other words, in addition to the MFHSP, there may also be a male–female dysregulation-frailty paradox, where men become dysregulated without becoming frail (Figure 3). This agrees with our findings and those of others that men have higher dysregulation levels, women are more frail, and men are more subject to mortality. Possibly this higher dysregulation in men leads to their earlier mortality (in agreement with Arbeev et al. (34)), but without as large an accompanying change in quality of life or health status. We attempted to test this alternative hypothesis post-hoc by assessing a D_M × sex interaction term for prediction of physical functioning measures in InCHIANTI, but results were ambiguous (Supplementary Table 4). Interestingly, mediation analyses showed that the effect of D_M on mortality is partially attributable to its effect on frailty, equally in both men and women.

Figure 3.

The male–female dysregulation-frailty paradox. Men have higher dysregulation levels than women, but despite this women are more frail. This parallels the male–female health-survival paradox, where women are more frail but despite this men are more susceptible to mortality.

In related work, Mitnitski et al. (39) found that health deficit counts, as measured by the FI, vary greatly between sexes, while Kulminski et al. (40) found no differences between sexes using an index of cumulative deficits. However, both studies revealed accelerated deficit accumulation in healthier individuals, suggesting some kind of threshold in health deficit accumulation beyond which death occurs, and men and women may have different thresholds. Hubbard (41) uses the safety factor concept borrowed from engineering to demonstrate that women may be intrinsically conceived to tolerate higher stress levels than men.

Additional insight might be found by considering which systems are more dysregulated in men. We found no sex differences in dysregulation of electrolytes, lipids, or liver/kidney function, but we did find greater male dysregulation in oxygen transport (largely composed of markers of red blood cell function from a standard blood panel) and hematopoiesis (composed of white blood cell counts). Several hypotheses might explain links between these systems and sex differences in aging patterns: (a) Dysregulation of these two systems might lead directly to mortality, whereas dysregulation of other systems might be compatible with living in a frail state. However, we previously found that these two systems were not markedly different from others in predicting mortality and frailty (33). (b) Dysregulation of these two systems might have disproportionate effects on certain causes of death (eg, cardiovascular) that might lead to sudden death, and less effect on other, less fatal pathologies. However, this is not supported by our previous finding that these two systems had some of the weakest predictive power for cardiovascular disease incidence (33). (c) The systems measured here are but a small fraction of actual systems and were based on convenience sampling of biomarkers. There is a strong possibility that the key sex differences underlying mortality–frailty patterns are found in systems we did not measure. An inverse morbidity paradox depending upon health dimensions assessed has been proposed (42), and averaging all dimensions might drown out any sex-specific difference (40).

Our finding of greater dysregulation in the hematopoietic system provides some support to the hypothesis that immune differences underlie the male–female health-survival paradox (21). However, this hypothesis would not explain the difference we found in oxygen transport dysregulation, and we suggest that immune differences are likely part of the explanation but not sufficient in themselves. Generally, our findings show physiological differences in aging in men and women, but do not support clear links between these physiological differences and the MFHSP. This is likely partially due to a lack of power for these analyses. Figure 2 shows that male–female differences in slopes are small compared to the confidence intervals of the sex-specific estimates, and neither our methods nor our power are ideal to detect very subtle interaction effects. For example, the fact that D_M increases are not linear makes the use of slope estimates problematic (32), though the alternatives (higher-order regression or splines) would have been underpowered and hard to interpret. Nonetheless, based on our findings, we predict more broadly that the physiological mechanisms underlying the MFHSP will be highly diffuse and hard to pin down. This would in fact agree with the evolutionary hypothesis that different selective pressures on males and females in polygynous or sexually dimorphic species create sex-specific physiologies that optimize sex-specific demographies (43). In fact, our finding that dysregulation levels but not rates differ between the sexes parallels the finding of Tidière et al. (44) that onset but not rate of senescence differs in relation to polygyny across mammal species. Presumably, there is some canalization of the physiological mechanisms underlying these demographic patterns, but there is no reason to think the effect should be constrained to one or a few physiological systems, and no reason to expect the mechanisms to be identical across species.

Supplementary Material

Supplementary data is available at The Journals of Gerontology, Series A: Biological Sciences and Medical Sciences online.

Funding

This work was supported by catalyst grant on biological differences between the sexes from the Canadian Institutes of Health Research (CIHR) (grant number SVB-145585). AAC is also supported by a CIHR New Investigator Salary Award and is a member of the Fonds de recherche du Québec – Santé funded Centre de recherche du CHUS and Centre de recherche sur le vieillissement.

Conflict of Interest

None reported.