Testosterone and type 2 diabetes in men

Abstract

Objective: To assess from observational data if low testosterone in men is an independent risk factor for high fasting glucose (FG) and for a diagnosis of type 2 diabetes (T2D).

Methods: Multivariate analysis of data from 991 male US Air Force veterans who completed six medical examinations over 20 years.

Results: Low testosterone was moderately related to high FG, independent of age and obesity. Low testosterone is a very weak predictor of a diagnosis of T2D.

Conclusions: In men, low testosterone is an independent risk factor for high FG, comparable to aging and obesity. Low testosterone is a weak predictor of a diagnosis of T2D.

• Aging
• fasting glucose
• obesity
• testosterone
• Type 2 diabetes

Introduction

As men age and gain weight, their risk of higher fasting blood glucose and type 2 diabetes mellitus (T2D) increases. Several studies of hypogonadism in men with T2D have inferred that low testosterone (T) is an independent risk factor for T2D, suggesting that T supplementation therapy may be of benefit in some diabetic males. Most studies that have been done to date examining T2D in hypogonadal men have been in cross-sectional samples, often small in size, and the few longitudinal studies that have been done were of short duration. Small clinical trials of T replacement therapy for hypogonadal men with T2D have reported some success in reducing fasting glucose (FG) [1–13].

Since obesity, particular central adiposity, and aging are associated with low T in men [14–17], it is difficult to determine the independent causal effects of these factors on T2D or on its principle clinical manifestation, high blood glucose. T supplementation in an uncontrolled observational cohort of hypogonadal men was accompanied by the loss of body weight and waist circumference [18]. Free T in obese men after bariatric surgery was shown to increase in parallel with insulin sensitivity, suggesting a causal link between T and obesity [19]. Androgen deprivation therapy for prostate cancer increases obesity, decreases insulin sensitivity and may be associated with a greater incidence of diabetes [20]. Rapid lowering of circulating T, either by ending T replacement therapy in hypogonadal men or by experimentally blocking the action of endogenous T in normal men, is quickly followed by elevations in FG [21]. These glucose responses occur too quickly to be explained simply by a T-induced buildup of muscle mass and/or reduction of central adiposity. Kelly and Jones [4] suggested a number of cellular mechanisms that are beneficially modulated by T; however, these processes are not well understood, and T replacement therapy remains an issue of debate. For hypogonadal men with T2D, there are some data suggesting that low T predicts increased general mortality, although follow-up periods have not exceeded six years [22–24]. Also, it is difficult to determine the effect of T therapy on individual risk factors for T2D since T affects multiple factors.

The Air Force Health Study (AFHS) was initiated to evaluate health effects of exposure to Agent Orange during the Vietnam War. It compared air and ground crewmen involved in wartime herbicide spraying (in Operation Ranch Hand) with matched Air Force veterans involved in other transport aircraft missions in Southeast Asia [25]. This analysis examined the relationship between T, FG, and the occurrence of T2D among 991 US Air Force veterans who participated in six examination cycles over 20 years of the AFHS.

Methods

AFHS “Ranch Hands” and designated comparison subjects were invited in 1982 for a baseline personal interview, physical examination, and psychological testing [25]. At that time, the men ranged in age from 32 to 68 years. They were invited again in 1985, 1987, 1992, 1997, and 2002 (some examinations were conducted slightly before or after these years). There was little difference between men who participated in the physical examination and those who refused in terms of reported health status, medication use, and days lost from work. There were few discernible differences in the health of Ranch Hands and comparison subjects [25] (see Michalek and Pavuk [26] for possible exceptions).

Comparison subjects lost from the panel were replaced, and additional Ranch Hands were located and added, so that over 4000 men participated in at least one cycle, with about 2000 men participating in any single cycle. This analysis included 991 men who completed all six cycles and gave permission for their data to be used or who had since died, in which case permission was not required for inclusion. The US Institute of Medicine now holds these data and provided limited access for this research.

Compared to the 1881 men who completed all of the first four cycles (reported by Mazur and Michalek [27]), the 991 veterans included in this analysis were about one year younger, slightly more educated (45% versus 42% college graduates), slightly more likely to be married at each of the first four cycles (ca. 87% versus ca. 84%), and less likely to be black (4% versus 6%). Mean T levels (in ng/dl) for the samples of 991 and 1881 men differed by at most 10 ng/dl in each of the first four cycles. By these measures, this sample of 991 men was similar to the 1881 men who completed the first four cycles. The sample of 991 was biased at least in terms of being sufficiently healthy to still be alive for the sixth cycle of the study. By 1992, 27% of the 991 men in the sample were obese (BMI ≥ 30 kg/m²), and by 2002 35% were obese. These percentages of obesity were similar to those reported by the US Centers for Disease Control [28] for the age-adjusted adult population (23% for 1988–1994 and 35% for 2005–2008).

The AFHS does not describe the criteria by which men were diagnosed with T2D. Whether or not a man was designated as diabetic in the AFHS may not be fully reliable because men whose records were marked “diabetic” in one cycle often were not so marked in all subsequent cycles, which is inconsistent with the presumption that T2D is a chronic condition. Men diagnosed with T2D during the study are coded 1, others coded 0, forming a binary dependent variable in logistic regressions run for each cycle. Year of first diagnosis ranged from 1958 to 2003. For convenience, these were grouped by the cycle year closest to the date of diagnosis.

FG was measured once in the morning, before breakfast, for all men in all cycles. Hemoglobin A1c (HbA1c), indicating adequacy of blood glucose control over the past 6 to 12 weeks, was measured for all men in cycles 4 and 5. (In cycle 6, HbA1c was measured only for those diagnosed or suspected to be diabetic, which was not useful for this analysis)

T was assayed in duplicate from morning blood samples after an overnight fast. Quality control procedures required that when the coefficient of variation (CV) between duplicates was greater than 8%, assays were retaken. Mean CV between duplicates is about 5%. About 20 of nearly 6000 T values were either missing or were extraordinarily high (over 1600 ng/dl) or low (under 100 ng/dl) for the sample as a whole and compared to the other five T values recorded for the subject. Since these were most likely errors of measurement or recording, they were replaced by the mean of the subject’s two T values in adjacent cycles, or in the case of anomalously low numbers, by the otherwise lowest T value for that subject. Means and standard deviations for T levels, none of them unusual, are shown by year of cycle in Table 1 with other descriptive statistics.

Table 1. Mean (and standard deviation) or percentage of relevant variables by year of cycle (n = 991).

	1982	1985	1987	1992	1997	2002
Age (yr)	43 (6.9)	46 (6.9)	48 (6.9)	53 (6.9)	58 (6.9)	63 (6.9)
Age range (yr)	31–63	34–66	36–68	41–73	46–78	51–83
>High school	54%	Same	Same	Same	Same	Same
Black	4%	Same	Same	Same	Same	Same
Ranch Hands	46%	Same	Same	Same	Same	Same
Married	88%	88%	83%	87%	85%	86%
BMI (kg/m^2)	26.6 (3.5)	27.3 (3.8)	27.6 (3.9)	28.2 (4.1)	28.5 (4.5)	28.9 (4.3)
Obese (BMI ≥ 30)	13%	19%	22%	27%	30%	35%
Weight (kg)	84.4 (12.2)	85.5 (12.6)	86.7 (13.0)	88.8 (13.9)	90.0 (15.1)	91.0 (14.8)
Cumulative number (and percent) diagnosed T2D	37 (3.7%)	46 (4.6%)	54 (5.5%)	87 (8.8%)	141 (14.2%)	183 (18.5%)
Fasting glucose (mg/dl)	97.5 (15.8)	98.5 (17.5)	100.2 (18.9)	104.9 (26.4)	102.9 (29.1)	102.9 (26.1)
HA1c (%)	–	–	–	7.1 (1.3)	6.5 (1.3)	–
Testosterone (ng/dl)	638 (176)	611 (203)	535 (160)	520 (189)	440 (179)	431 (200)

FG distributions had considerable right-hand skew (=6.0), so regression analysis was performed on log-transformed values as well as raw values. T distributions had slight right-hand skew (about 0.6) and were not transformed.

Results

Table 1 shows descriptive statistics for relevant variables across the six cycles. As expected, there were gradual increases in obesity, FG, and T2D, and a decline in mean T, over the 20 years of study. Pearson correlations between BMI in one cycle and BMI in any other cycle ranged from r = 0.74 to 0.90 (median r = 0.84, p < 0.0001). For T, correlations among cycles ranged from r = 0.45 to 0.72 (median r = 0.57, p < 0.0001).

Table 2 shows a correlation matrix of FG in all six cycles, HbA1c in cycles 4 and 5, and a diagnosis of T2D at any time during the study. There was a significant (all p < 0.001) and substantial clustering of these three variables across the 20-year study period.

Table 2. Pearson correlation matrix of indicators of type 2 diabetes.

	FG1	FG2	FG3	FG4	FG5	FG6	HbA1c4	HbA1c5
FG2	0.61
FG3	0.60	0.70
FG4	0.52	0.72	0.83
FG5	0.44	0.52	0.59	0.64
FG6	0.33	0.32	0.40	0.36	0.52
HbA1c4	0.49	0.69	0.71	0.81	0.59	0.34
HbA1c5	0.42	0.48	0.51	0.57	0.78	0.45	0.60
T2D diagnosis	0.38	0.42	0.44	0.50	0.60	0.55	0.48	0.54

FG1 = fasting glucose in cycle 1, etc. All p < 0.0001.

Time course of type 2 diabetes

Figure 1 shows mean FG (with standard errors) across cycles for non-diabetic men (n = 808), for men who were diagnosed with diabetes early in the study (i.e. by cycle 4, n = 87)), and for those diagnosed with diabetes later in the study (after cycle 4, n = 96). Among non-diabetic men, mean FG remained stable across cycles with little variation. The mean FG of early diagnosed diabetics was high throughout the study, although decreasing in cycle 6, probably due to medication. It is noteworthy that men who would not be diagnosed with T2D until cycles 5 or 6 already showed by cycle 1 that their mean FG was significantly (if not greatly) higher than normal.

Figure 1. Mean fasting glucose for early and late-diagnosed diabetics and for normal men, by cycle.

Early diagnosed diabetics were slightly, but not significantly, older than late-diagnosed diabetics (45.1 years versus 44.4 years) and were not significantly different in BMI or T, so they were henceforth combined (n = 183). The mean age of these diabetics at cycle 1 was about three years older than non-diabetic men (44.7 years versus 42.2 years, p < 0.0001). The distributions of FG values for all diabetic men compared with non-diabetic men are shown in Figure 2. FG in non-diabetic men were tightly clustered around 100 mg/dl, while FG values in diabetic men were highly skewed and dispersed, with a few values falling below 100 mg/dl while others exceeding 300 mg/dl.

Figure 2. Cycle 5: Scattergram of fasting glucose for normal and diabetic men.

Mean BMI (with standard errors) across cycles for all diabetics (i.e. all recognized as T2D during the study) and non-diabetic men are shown in Figure 3. Throughout the study, the mean BMI of diabetic men exceeded that of non-diabetic men by 2–3 kg/m². BMI trends for early and late-diagnosed diabetics were not significantly different.

Figure 3. Mean BMI for diabetic and normal men, by cycle.

BMI distributions for diabetic and non-diabetic men in cycle 5 are shown in Figure 4. The BMI distributions in this cycle were similar except that BMI levels in diabetic men were generally higher than in non-diabetic men, although there was much overlap. Of 201 men with BMI < 25 kg/m², only 9% had T2D, whereas among 288 men with BMI > 30 kg/m², 31% had T2D.

Figure 4. Cycle 5: Scattergram of BMI for normal and diabetic men.

Figure 5, comparing mean T (with standard errors) across cycles for diabetics and non-diabetic men, shows non-diabetic men having consistently higher T, a difference approaching 100 ng/dl (There were no significant differences in T trends for early and late-diagnosed diabetics).

Figure 5. Mean testosterone for diabetic and normal men, by cycle.

T distributions for diabetic and non-diabetic men are similarly skewed and overlapped considerably, but the proportion of men with normal T levels was generally higher in the normal group compared with the diabetic group (Figure 6). Of 151 men with cycle 5 T > 600 ng/dl, only 10% had T2D, whereas among 140 men with cycle 5 T < 300 ng/dl, 35% had T2M.

Figure 6. Cycle 5: Scattergram of testosterone for normal and diabetic men.

High BMI was confounded with low T (in this sample, Pearson correlations between BMI in any cycle and T in any cycle range from r = −0.21 to −0.41, with median r = −0.33, p < 0.0001). Pearson correlations of age with BMI in any cycle ranged from r = −0.10 to 0.08 (median r = 0.03, ns); correlations of age with T in any cycle ranged from r = −0.05 to −0.27 (median r = −0.13, p < 0.001).

Low T as an independent risk factor for T2D

The percentage of diabetics for each combination of trichotomized T and trichotomized BMI (<25, 25–30, >30) were examined. In cycle 4, among men with high T and low BMI, only 8% were diabetic, whereas among men with low T and high BMI, 42% were diabetic. The marginal percentages showed that both factors were associated independently with diabetes. This pattern was consistent and highly significant across all cycles (p < 0.0001). There was no consistent indication of a BMI-by-T interaction effect on diabetes (Table 3).

Table 3. Cycle 4: Percent of men with type 2 diabetes as a joint function of (trichotomized) testosterone and BMI.

	Testosterone
	Low	Medium	High	All
BMI
Low	16% (n = 75)	4% (n = 98)	7% (n = 161)	8% (n = 334)
Medium	18% (n = 94)	15% (n = 117)	13% (n = 112)	15% (n = 323)
High	42% (n = 162)	25% (n = 115)	19% (n = 57)	32% (n = 334)
All	29% (n = 331)	15% (n = 330)	11% (n = 330)	18% (n = 991)

A parallel analysis of mean FG in cycle 4 is shown in Table 4. Decreasing T and increasing BMI were both associated with higher FG, each independent of the other. Men with high T and low BMI had a mean FG level of 100 mg/dl, while those with low T and high BMI had a mean FG level of 114 mg/dl. Although this difference was small, the pattern was consistent and significant across cycles (p < 0.0001). There was no consistent indication of a BMI-by-T interaction effect on FG.

Table 4. Cycle 4: Mean fasting glucose (mg/dl) as a joint function of (trichotomized) testosterone and BMI.

	Testosterone
	Low	Medium	High	All
BMI
Low	106 (n = 75)	98 (n = 98)	100 (n = 161)	101 (n = 334)
Medium	111 (n = 94)	102 (n = 117)	101 (n = 112)	104 (n = 323)
High	114 (n = 162)	105 (n = 115)	104 (n = 57)	109 (n = 334)
All	111 (n = 331)	102 (n = 330)	101 (n = 330)	105 (n = 991)

Logistic regressions were run for each cycle, with independent variable T and BMI at each cycle, and birth year. (In preliminary models, whether a man was a Ranch Hand versus a control was statistically insignificant and was omitted here) Table 5 shows odds ratios (with significance level and 95% confidence interval) for all independent variables, by cycle, and a pseudo R² for the full logistic regression. (Note: An association of low T or advanced age (i.e. lower birth year) with T2D is indicated by an odds ratio <1.0; an association of high BMI with T2D is indicated by an odds ratio >1.0)

Table 5. Logistic regressions predicting type 2 diabetes, by cycle.

	Independent variables
	Testosterone	BMI	Birth year	Pseudo R^2
Cycle 1	0.998** (CI = 0.997–0.999)	1.21** (CI = 1.15–1.28)	0.956** (CI = 0.933–0.980)	0.12
Cycle 2	0.998** (CI = 0.997–0.999)	1.22** (CI = 1.16–1.28)	0.962* (CI = 0.938–0.987)	0.13
Cycle 3	0.998** (CI = 0.996–0.999)	1.22** (CI = 1.16–1.28)	0.960* (CI = 0.936–0.986)	0.14
Cycle 4	0.999* (CI = 0.998–0.999)	1.20** (CI = 1.14–1.25)	0.948** (CI = 0.925–0.972)	0.12
Cycle 5	0.998** (CI = 0.996–0.999)	0.112** (CI = 1.08–1.16)	0.947** (CI = 0.924–0.971)	0.10
Cycle 6	0.998 (CI = 0.998–1.00)	1.13** (CI = 1.08–1.18)	0.939** (CI = 0.916–0.962)	0.08

Odds ratios are for independent variables testosterone, body mass index, and birth year. CI = 95% confidence interval.

*p ≤ 0.05.

**p ≤ 0.001.

Comparing results from cycle to cycle, the pseudo R² decreased over time, indicating that the model as a whole lost predictive power toward the end of the study. The predictive power of BMI and birth year (age), well-known risk factors for T2D, also weakened in the later cycles. With regard to the impact of T on T2D, the coefficients were statistically significant in all but the last cycle (because of the large n), but the odds ratios were not sufficiently different from 1.0 to be important.

Taking the analysis further, FG was regressed on T, BMI, and birth year (i.e. age) for each cycle. (Ranch Hand versus control was not significant in any model and was excluded here) To reduce the right-hand skew of glucose distributions, regressions were run after log transformation; however, the resulting coefficients are not as easy to interpret as coefficients obtained for raw glucose values. Therefore, results shown in Table 6 are for untransformed glucose. Raw and log glucose models give essentially the same results except that using log glucose explained more variance (R²) and gave more significant p values for coefficients.

Table 6. Linear regression coefficients (B), standardized coefficients (β), and R² for the dependent variable fasting glucose, by cycle.

	Independent variables
	Testosterone	BMI	Birth year	R^2
Cycle 1	B = −0.01**, β = −0.14	B = 0.30*, β = 0.07	B = −0.32**, β = −0.14	0.06
Cycle 2	B = −0.01*, β = −0.11	B = 0.76**, β = 0.17	B = −0.29**, β = −0.08	0.06
Cycle 3	B = −0.01*, β = −0.08	B = 0.71**, β = 0.15	B = −0.21*, β = −0.10	0.05
Cycle 4	B = −0.02**, β = −0.11	B = 0.58*, β = 0.09	B = −0.40**, β = −0.11	0.04
Cycle 5	B = −0.02**, β = −0.09	B = 0.58, β = 0.09	B = −0.37*, β = −0.14	0.05
Cycle 6	B = −0.01, β = −0.06	B = 1.39**, β = 0.23	B = −0.23*, β = −0.06	0.07

*p ≤ 0.05.

**p ≤ 0.001.

Results in Table 6 were fairly consistent across cycles, without consistent weakening in predictive power for individual variables or for the model as a whole. All three independent variables were significant predictors of FG, each at least partly independent of the others. Comparing the values of β across cycles, no one risk factor consistently dominated the others. However, even taken together, they did not explain much variance in FG, with R² ranging from 0.04 to 0.07. (In the models with log glucose, R² ranged from 0.05 to 0.09) An increase in FG of 10 mg/dl (roughly the marginal percentage changes seen in Table 4) was equivalent to about an 800 ng/dl decrease in T, or about 15 kg/m² increase in BMI, or about 30 years of aging.

Correlations among T and FG were examined across all cycles for consistent leading or lagging correlations that might indicate the time sequence of causal change, for example, decreases in T followed by increases in FG. No obvious pattern was identified. Substituting free T (only in cycles 4, 5, and 6) for total T did not affect any of the models.

Discussion

In this study of 991 US Air Force veterans followed through six examinations over a 20-year period, modest support was found for the hypothesis that declining T is a risk factor (in men) for high levels of FG. Comparing 20-year profiles of non-diabetic men with those of men diagnosed with T2D, non-diabetic men have lower BMI and higher T from the outset (though with considerable overlap), and BMI generally increases while T generally decreases with advancing age. These results were consistent with those of previous reports.

Since T varies inversely with BMI (and age), it is difficult to determine the independent association of these factors with FG. Our multivariate analyses showed that T was inversely related to glucose, independently of BMI and age, but its effect on glucose was small. In cycle 4, for example, among men of approximately similar BMI, those in the upper third of T levels had mean FG of about 10 mg/dl less than men in the lower third of T levels. The difference was greater than the difference in FG between men in the upper and lower thirds of BMI (holding T approximately constant).

A comparison of standardized coefficients in multiple regression models showed that the independent association of T to FG was roughly similar to the independent associations of BMI and age with FG (Table 6). Overall, the tabular and regression results suggest that low T may be regarded as a risk factor for high glucose on a par with BMI and aging.

We used similar multivariate analyses to test for a unique association of T with a diagnosis of T2D. Our trichotomized T and BMI analysis showed that T was modestly predictive of T2D. In cycle 4, for example, among men of approximately similar BMI, those in the upper third of T levels were 18% less likely to have T2D than those in the lower third of T levels. A greater difference (24%) in the proportion of men who had T2D was observed between men in the upper and lower thirds of BMI (holding T approximately constant) for the same cycle.

T’s independent predictive power for T2D did not hold up well in multivariate logistic models, barely altering the odds ratio for being diabetic versus non-diabetic. T was a poorer predictor of T2D than BMI or age (birth year). T did reach statistical significance in nearly every cycle, but we discount this because of our large sample size, which inflated the importance of a small effect.

T therapy may improve multiple risk factors for T2D. Besides raising T, T supplementation improves BMI and may lower serum glucose. The full effect of these multiple benefits may reduce the risk of T2D or, among diabetic men, may work in concert to lower serum glucose. We compared men in the AFHS who were in the highest third of the sample for BMI as well as in the lowest third of the sample for T (i.e. the worst combination of BMI and T), with men who were in the lowest third for BMI as well as in the highest third for T (i.e. the best combination). While 42% of men in the worst combination were diagnosed with T2D, only 7% of those with the best combination of BMI and T were diagnosed with T2D. Men with the worst combination had a mean level of FG that was about 14 mg/dl higher than the mean FG of men with the best combination of BMI and T.

We emphasize the most important limitations of the present study. As noted above, we are wary of the coding of a diagnosis of T2D, because it is often inexplicably variable from cycle to cycle, which should not happen for a chronic disease, so our results may understate the strength of the relationship between low T and T2D. FG among diabetic varies from day to day, and since it was assessed on only one morning per examination, we assume that there was a fair amount of error variation in the measurement. Probably, the effects reported here understate the association of glucose with T (as well as with BMI). Possibly our sample of Air Force veterans is atypical of non-veterans in some way that invalidates the external validity of the study, although we are not aware of any factor that is relevantly different. Observational studies do not test causality. Whether T therapy (when endogenous T is low) will reduce the risk of developing diabetes, or whether it will help glucose control in diabetic men, can be definitively answered only with large, prospective, well-controlled clinical trials.

Some small or nonrandomized clinical trials have already suggested that T therapy improves glycemic control [29,30]. We conclude that our long-duration (20 years) observational follow-up data are in agreement with the hypothesis that T supplementation may improve T2D risk in hypogonadal men.

Declaration of interest

A.M. and R.W. report no conflicts of interest. A.W. and U.M. recently completed a type 2 diabetes managed-care study, funded by Intercomponent Ware Inc. and Roche Diagnostics Ltd Germany, a subsidiary of Hoffmann La Roche Ltd. Switzerland, ClinicalTrials.gov Identifier NCT01556529. The authors alone are responsible for the content and writing of this article.