Correlates of testosterone change as men age

Abstract Objective: The literature on testosterone (T) in men reports diverse correlates of T, some with minimal empirical support and most with little indication of how they change with advancing age. We test eight putative correlations across age. Method: Correlations were tested on a large sample of British men. Results: Seven of eight correlations replicated. Most change across men's life courses. The diurnal cycle of T is considerably weaker among older than younger men. Single men have higher T than married men of the same age; however, this difference lessens as men get older. Elevated T among smokers is less pronounced as men age. The inverse relationship between obesity and T is sustained across the adult age range. The lessening of T with age is well established, however there is disagreement about the course of decline. We find T having a steep decline around age 30, with possibly a rebound around age 50, after which levels remain roughly constant. Correlations involving health become stronger among older men. After age 30 or 40, the inverse relationships between T and HbA1c, diabetes, and metabolic syndrome all become increasingly significant, though not necessarily strong in magnitude. Conclusion: Most putative correlates of T are replicated. There is a basis here for the generalization that among older men, those healthy have higher T than those who are not, but not a lot higher.


Introduction
Testosterone (T) plays a dynamic role in men's lives [1]. Empirical correlations that have been reported include diurnal decline in T from morning to afternoon, decline in T with obesity, declining T with advancing age, and associations with morbidity and mortality. These (and other) findings are sometimes based on modest samples, especially for the oldest men, leaving in doubt their validity across ages, e.g. [2]. Possibly the dynamics of T lessen in older men. "Endocrine function undergoes major changes during aging … Specifically, alterations in hormonal networks and concomitant hormonal deficits/excess, augmented by poor sensitivity of tissues to their action, take place" [3].
Here, we present eight putative correlates of T, derived from various studies, some well-grounded, others less so. The literature is sometimes sparse and not entirely consistent. Discrepant reports may invalidate generalizations, or they may simply be due to inadequate sampling, whether serum or saliva was assayed, and the method of assay. Especially suspect is enzyme immunoassay of T in saliva, compared to the "gold standard," liquid chromatography-tandem mass spectrometry (LCTMS) [4,5]. Thus, the robustness of our generalizations is variable. Most have not adequately been tested, and their application to elderly men is especially dubious. Here we test proposed correlations on the large United Kingdom Household Longitudinal Study (UKHLS), a general-purpose longitudinal dataset of the UK population that collects a range of socioeconomic and health information, little exploited for this purpose [6]. Between 2010 and 2012, a one-off nurse visit involved the collection of additional health and biomarker information that we use for this analysis (see) [7]. We give special attention to the effects of aging.
Relatively low female T makes measurement difficult so we exclude females from the analysis. Also of note, only a small portion of total T is biologically active as free T, a quantity not included in the dataset, so we do not address free T. All results are for total serum T. The time of nurse's visit approximates the time of blood drawing.

Putative correlates of male T
Our enumeration of propositions is pragmatic, focusing on those explicit in the literature and testable with the large dataset at hand. Potential associations of T with male sexuality or prostate cancer, for example, are well known but cannot be tested for lack of relevant data in this dataset. We address these eight: � 1. Male T has a diurnal rhythm, high in the morning, lower as the day goes on (e.g.) [8,9]. However Bremner et al. [10] reported a loss of circadian rhythm with aging. � 2. Low T is associated with obesity as measured by body mass index (BMI) [11][12][13][14]. � 3. In industrialized societies, men past the age of thirty on average show declining T [15][16][17][18][19][20][21][22]. It is uncertain whether this decline is gradual and continuous into old age or occurs mainly in middle age with little further reduction [18,23,24]. Of related interest are studies of subsistence societies, suggesting that male T is normally lower than in industrial societies and has little or none of the age-related decline found in industrial societies [25][26][27]. � 4. Married men have lower levels of T than single men [28,29]. In one longitudinal study with four waves over 10 years, T declined when a marriage occurred but rose in the years around a divorce [4]. � 5. Cigarette smoking raises T [30]. � 6. T has been associated with diverse risk-taking behaviors [31][32][33]. Some affirmative studies depended on the ratio of the second-to-fourth finger as a proxy for T in utero, however, that association is questionable [34][35][36], measuring salivary T using LCTMS, found no consistent relationship between T and economic risk-taking, and a review of earlier work raised doubts about the association [37].
Low T is associated with morbidity and mortality [38][39][40][41]. The popularity of testosterone replacement therapy (TRT) in the twenty-first century has pushed this concern to the fore, with reports of both beneficial and adverse effects on cardiovascular disease (CVD) and type 2 diabetes [42][43][44][45][46] . Therefore, we address two health-related correlations:  [46,[51][52][53]. However, Bhasin et al. [54] found no independent relationship, and Blaya et al. [55] assert that central obesity is the primary factor of metabolic syndrome that is related to T.

Methods
Our analysis is restricted to men aged 16-90 years old from the UKHLS [56]. These are publicly available data from the UK Data Service and require no ethics review for secondary analysis. The age range is collapsed into 23 age groups, their maximum sample sizes shown in Table 1.
In the years 2010-2012 (Waves 2 and 3), T was assayed in blood drawn from adults during home visits by a nurse [57]. T was measured once per subject and not measured in later waves. Nurse visits occurred throughout the day, providing an opportunity to replicate the diurnal effect on T. The mean time of day when nurses visited men over sixty years of age was about two hours earlier than the mean time of visits to younger men, so raw T values were adjusted as if all blood had been drawn at 10 am. Hereafter, "T" refers to T adjusted to 10 am.
Blood drawn by the nurse was measured for total T by electrochemiluminescent immunoassay on the Roche Modular E170 analyzer, with the lowest detectable T > 1 nmol/L. Few men had undetectable T, but the majority of women did. An undetectable T value was scored as a missing value. Intra-and inter-assay coefficients of variation were less than 4% [7]. We follow the UKHLS biomarker user guide (see) [7], which states the normal range of T in men is 9-25 nmol/L.
Metabolic syndrome (MS) is a measure of risk for CVD, typically diagnosed by the presence of three or more medical conditions: abdominal obesity, high blood pressure (multiple readings), high blood sugar, high triglycerides, and low high-density lipids. Details of measurement vary among medical organizations (see Wikipedia's entry for MS), and we use a modified definition here, taking HbA1c to measure high blood sugar because fasting glucose test results are not available. We exclude high blood pressure because our dataset did not include multiple readings, replacing this with another risk factor for CVD, C-reactive protein, discussed as a supplemental indicator of MS (e.g.) [58]. Typically, abdominal obesity is measured by waist size, but the International Diabetes Federation takes BMI � 30 kg/m 2 as implying central obesity. In the UKHLS data, high BMI and high waist size are redundant measures, 96% consistent across men. Analytic results are nearly identical no matter which one is used. Our dummy for MS is scored 1 if at least three of the following risk factors are present, 0 otherwise. Triglycerides � 1.7 mmol/L High density lipids < 1.04 nmol/L BMI � 30 C-reactive protein > 3 mg/L HbA1c � 48 mmol/mol (� 6.5%).
For each putative correlation, we display a plot (or plots) of mean adjusted T against age. For tests of the early propositions, we instead plot standardized regression coefficient (beta) against age.
Each correlation is further tested with regression models, usually for a restricted age range, using a suite of control variables that may explain away the putative correlation. Standard control variables are BMI, ethnicity, marital status, employment status, education level, socioeconomic class of job (separate category for retired), and whether or not the respondent is a smoker. These data are measured at the nurse visit or the most recent main-stage survey wave prior to the nurse visit.

Results
The UKHLS sample is broadly representative of the UK adult population in 2009-2010. Mean T for 5940 men ¼ 15.7 nmol/L (sd ¼ 5.8) with few missing values. Assays were conducted for 7333 women, producing usable values (i.e. > 1 nmol/L) for only 1787 women, whose mean T ¼ 1.89 nmol/L (sd ¼ 2.1). The ratio of male to female T is 8.3, however, if values had been available for the majority of women with undetectable T (i.e. < 1 nmol/L), that ratio would have been correspondingly higher.

Diurnal variation in T decreases with age
Nurse visits were scattered throughout the day, allowing a test of diurnal change by regressing (unadjusted) T against the time-of-day blood was drawn. A negative beta, or standardized regression coefficient, indicates that T is lower the later in the day that blood was drawn. Separate regressions were run for each age group. Figure 1 plots beta across age groups. (Vertical lines show 95% confidence intervals for each beta.) We see that beta is negative at each age; however, this effect is the strongest in young men, greatly diminishing in middle age and the elderly.
As a simple indication of the diurnal effect, consider that if men aged 22-24 years had blood drawn at 4 pm, the assay would show their mean T to be 17.83 nmol/L, but if the same men had blood drawn at 10 am, their mean T would have been 23.16 nmol/L, a difference of 5.33 nmol/L. If men aged between 58 and 60 had blood drawn at 4 pm, the assay would show their mean T to be 14.94 nmol/L, but if the same men had blood drawn at 10 am, their mean T would have been 17.52 nmol/L, a difference of only 2.57 nmol/L.
To evaluate the effect of possibly confounding variables, we estimate two regression models. The first only considers the time of nurse visit (approximating time of blood draw) on unadjusted T. In this case, for a 1-minute increase in the time of the nurse visit (past 10 am), unadjusted T declines by À 0.015 ��� nmol/L for men aged between 22 and 24. In the second regression, including the suite of covariates (listed under Methods), for a one-minute increase in the nurse visit, unadjusted T declines to À 0.013 ��� nmol/ L. Thus, the inclusion of controls barely changes the association of T with the time of the nurse's visit. 1 The remainder of the analysis uses "adjusted T," statistically adjusting the time of nurse visit as if it were 10 am.

Obesity is associated with lower T
Adjusted T was regressed on BMI for each age group (after dropping extreme values of BMI < 15 and BMI > 35, less than 1% of the sample). Figure 2 plots beta across age groups, showing the inverse relationship between T and BMI at each age. This inverse correlation becomes more negative after age 30, then remaining relatively constant across as men age.
To evaluate the effect of control variables on the association of BMI with adjusted T, as before, we estimate two regression analyses. The first considers the effect of BMI alone on adjusted T, showing if BMI increases by one unit (on the standard scale), average adjusted T declines by À 0.17 nmol/L for men aged between 22 and 24. In the second regression, including our suite of control variables, the effect of BMI on T becomes increasingly negative, to À 0.38 � nmol/L. Thus, for young men, the inclusion  of control variables makes the inverse association of BMI with T stronger. For older men, the inverse correlation (beta ca. À 0.5) is sizable regardless of controls.

T decreases with age, especially for males in their thirties
In Figure 3, we plot mean adjusted T across age groups, separately for men with normal BMI (18.5-<25) and for men who are obese (BMI ¼ 30-40). T declines steeply among men during their 30 s, with possibly a "bounce back" for men in their forties, then staying more or less level, independently of BMI. BMI does not explain this age trend.
To evaluate the influence of control variables on the relationship of age to T, we again estimate two regressions. The first, including only the effect of age on T, shows that if age increases by five years, from 50 to 55, average T declines by 1.50 �� nmol/L. Adding our suite of control variables shows that if age increases from 50 to 55, average T decreases by 2.65 ��� nmol/L. Thus, the inclusion of control variables strengthens the inverse relationship between age and T.

Married men have lower T, especially among young men
In Figure 4, age groups have been collapsed to ensure adequate samples of single men, the smallest group now being men over age 60 (N ¼ 110). For individuals in their twenties (N ¼ 342), the mean T of single men is far higher than that of married men. This difference remains significant in later ages but diminishes in magnitude, perhaps reflecting the independent decline in T with age.
We again estimate two regressions to evaluate the effect of marital status on T. The first, with only marital status as independent variable, shows that among men aged between 30 and 39, those married have average T that is 1.96 ��� nmol/L lower than single men. The second, including our suite of control variables, again for men aged 30-39, shows those married have average T that is 1.34 �� nmol/L lower than single men. Thus, the qualitative nature of the correlation remains significant, but its magnitude declines by 30% in the presence of controls. Figure 5 replicates the previously reported finding that smokers have higher T than non-smokers. We see the difference persisting into old age, even as T generally decreases. As before, we evaluate the effect of control variables by comparing two regressions, the first including only smoking as an independent variable, showing smokers with average T to be 2.50 ��� nmol/L higher than non-smokers of the same age. If we include our set of covariates, the excess T Figure 4. Testosterone of single men is higher than that of married men at all ages. Data were taken from Understanding Society: Waves 2 and 3 Nurse Health Assessment, 2010-2012. The blue (red) dots refer to the age-group-specific adjusted T level for single (married) men, and the vertical lines show the 95% confidence intervals. � , �� , and ��� signify statistical significance at the 10, 5, and 1% levels, respectively. Figure 5. Testosterone of smokers is higher than that of non-smokers at all ages. Data were taken from Understanding Society: Waves 2 and 3 Nurse Health Assessment, 2010-2012. The blue (red) dots refer to the age-group-specific adjusted T level for smoking (not smoking) men, and the vertical lines show the 95% confidence intervals. � , �� , and ��� signify statistical significance at the 10, 5, and 1% levels, respectively. among smokers is a bit less, on average T ¼ 2.33 ��� nmol/L over non-smokers. Thus, the qualitative nature of the correlation remains unchanged, though the magnitude of the coefficient decreases by around 7%.

T is not significantly related to risk-taking
To test the association of T with risk-taking behavior, we use a particular question in the UKHLS asked at wave 1: "Are you generally a person who is fully prepared to take risks or do you try to avoid taking risks?" Answers are scaled from 0 to 10, where 0 ¼ avoid taking risks and 10 ¼ fully prepared to take risks. We plotted mean T by age group, comparing men who are risk tolerant with men who are risk averse. At no age is T significantly higher for the risk tolerant than for the risk averse. Furthermore, there is no consistent or significant pattern across all ages that show T relatively higher among risk tolerant men.
The effect of covariates is rather small in the case of risk assessment. Without control variables, for men aged 60-69, increasing the 0-10 risk scale by one unit, average adjusted T increases by 0.20. Including all covariates the effect attenuates slightly to 0.17.

Low T is related to high HbA1c and type 2 diabetes
High levels of HbA1c (>42 mmol/mole) are diagnostic of prediabetes. Men over age 50 with high HbA1c have significantly lower T than their age mates with normal HbA1c (Figure 6).
However, even positing that raising T would lower HbA1c, the estimated differences are not especially large. For example, among men aged 50-59, if we statistically raise the T level of those diabetic and prediabetic from their mean level of 15.1-17.6 mmol/mol, the average T for non-diabetic men of that age, the diabetic men's Hba1c would be reduced from 39.2 to 38.6.
We again estimate two regression models to test the influence of control variables on the relationship. In the first regression, with HbA1c the sole independent variable, diabetic and prediabetic men of ages 50-59 have an average T that is 2.40 ��� nmol/L lower than the T of normal men of the same age. The suite of covariates is added in the second regression, reducing the deficit in average T to 1.51 ��� nmol/L, still highly significant.
Type 2 diabetes, infrequent among young men, is usually diagnosed after age 40. Figure 7 compares T (by age) of 357 men who self-reported being diabetic, Figure 6. Over the age of 50, testosterone is significantly lower for men with high HbA1c. Data were taken from Understanding Society: Waves 2 and 3 Nurse Health Assessment, 2010-2012. The blue (red) dots refer to the age-group-specific adjusted T level for men with an HbA1c level > 42 nmol/mol (� 42 nmol/mol), and the vertical lines show the 95% confidence intervals. � , �� , and ��� signify statistical significance at the 10, 5, and 1% levels, respectively. compared to T of 3201 men who did not report the disease.

Metabolic syndrome is associated with low T
MS (here modified, as explained under Methods), is an aggregate risk factor for heart disease and stroke. Men with and without metabolic syndrome are shown separately in Figure 8, where mean T is plotted across age groups. After age 50, T levels are significantly and sizably lower for men with metabolic syndrome than for those without that risk. 2 As before, we estimate two regression models. The first considers only the effect of metabolic syndrome on T, showing that for men of ages 58-60, those with the syndrome have an average T that is 3.95 ��� nmol/ L lower than those without metabolic syndrome. Including our suite of covariates (except BMI) increases this deficit in T to 3.44 ��� nmol/L.

Discussion
Using the UKHLS, a representative large-scale survey of UK households, we replicated for men all putative correlates of T, except its putative relationship to risktaking (see summary Table 2). These correlations are robust after controlling for pertinent covariates. Furthermore, these successful replications effectively validate electrochemiluminescent immunoassay as a trustworthy method for measuring T in blood. A question remaining is whether these correlates apply to women with their much lower levels of T.
Most of the examined correlations change across men's life course, sometimes but not always attenuating. One novel finding is the attenuation of the diurnal cycle of male T, which is considerably weaker among older than younger men (Figure 1). An implication is that the medical practice of morning blood draws to assess T is not crucial for older patients. Another instance of "aging attenuation" is the lessening difference in T between married and single men, a direct result of single men's T falling more quickly with age ( Figure 4). Also, elevated T among smokers is less so as men age ( Figure 5). These examples of diminishing effect may follow from the general diminution of T with age. On the other hand, the inverse relationship between BMI and T is sustained across the adult age range (Figure 2).
Of particular interest is the lessening of T with age. Though well established in the literature, at least for industrial societies, there is disagreement about the course of the decline. A common view is that after the prime adult years, there is a continual decline with age. Empirically, we find evidence to the contrary, that after a steep decline around age 30, and a possible "bounce back" for men in their forties, T remains roughly level during the remaining years ( Figure 3). . Among older men, T is lower among those reporting they have diabetes. Data were taken from Understanding Society: Waves 2 and 3 Nurse Health Assessment, 2010-2012. The blue (red) dots refer to the age-group-specific adjusted T level for men self-reporting being diabetic (not being diabetic), and the vertical lines show the 95% confidence intervals. � , �� , and ��� signify statistical significance at the 10, 5, and 1% levels, respectively. Furthermore, we see the same shape of decline in comparing plots for men of normal weight and those obese. Obesity is clearly associated with lower T, but it does not explain the general decline with age.
Apparently, T of young men is remarkably high in industrial society, far higher than in subsistence societies. Perhaps the steep post-age 30 decline is simply a "correction" from abnormal heights during men's twenties.
Unlike the correlations just noted, those involving health become stronger among older men. After age 30 or 40, the inverse relationships between T and HbA1c (Figure 7), and diabetes (Figure 8), and metabolic syndrome become increasingly significant, though not necessarily strong in magnitude. There is some basis here for the generalization that among older men, healthy men have higher T than those who are not, however not a lot higher.
Correlations become truly interesting when they display causal links, an issue beyond the scope of this article. We have begun to see randomized, doubleblind, placebo-controlled, preregistered studies with sizable samples, aimed at therapeutic applications of T (e.g.) [59]. This is an active area of inquiry that requires robust correlations as the first step that must be taken. Notes 1. We check whether the diurnal change in T is sensitive to functional form and alternative specifications that use dummies to identify time of interview. All variations are consistent with those presented here. 2. Corona et al. [52] report that T drops with the number of MS components, which we confirm. Results available on request.
a preliminary version of the article, as well as anonymous reviewers for later comments.

Disclosure statement
Authors report no conflicting interests.

Funding
This research is unfunded.