Higher premorbid serum testosterone predicts COVID-19-related mortality risk in men

Objective Men are at greater risk from COVID-19 than women. Older, overweight men, and those with type 2 diabetes, have lower testosterone concentrations and poorer COVID-19-related outcomes. We analysed the associations of premorbid serum testosterone concentrations, not confounded by the effects of acute SARS-CoV-2 infection, with COVID-19-related mortality risk in men. Design This study is a United Kingdom Biobank prospective cohort study of community-dwelling men aged 40–69 years. Methods Serum total testosterone and sex hormone-binding globulin (SHBG) were measured at baseline (2006–2010). Free testosterone values were calculated (cFT). the incidence of SARS-CoV-2 infections and deaths related to COVID-19 were ascertained from 16 March 2020 to 31 January 2021 and modelled using time-stratified Cox regression. Results In 159 964 men, there were 5558 SARS-CoV-2 infections and 438 COVID-19 deaths. Younger age, higher BMI, non-White ethnicity, lower educational attainment, and socioeconomic deprivation were associated with incidence of SARS-CoV-2 infections but total testosterone, SHBG, and cFT were not. Adjusting for potential confounders, higher total testosterone was associated with COVID-19-related mortality risk (overall trend P = 0.008; hazard ratios (95% CIs) quintile 1, Q1 vs Q5 (reference), 0.84 (0.65–1.12) Q2:Q5, 0.82 (0.63–1.10); Q3:Q5, 0.80 (0.66–1.00); Q4:Q5, 0.82 (0.75–0.93)). Higher SHBG was also associated with COVID-19 mortality risk (P = 0.008), but cFT was not (P = 0.248). Conclusions Middle-aged to older men with the highest premorbid serum total testosterone and SHBG concentrations are at greater risk of COVID-19-related mortality. Men could be advised that having relatively high serum testosterone concentrations does not protect against future COVID-19-related mortality. Further investigation of causality and potential underlying mechanisms is warranted.

self-reported data, or self-reported use of anti-dementia medications, and HIV were initially considered for inclusion in analyses but were not used due to very low or zero numbers of events for participants with these conditions at baseline. Further information and ICD codes are provided in Supplemental Table S2. Medication usage (number of medications, lipid medication, anticonvulsant, glucocorticoid, opioid use) was identified from verbal interviews of participants at baseline on their use of prescription and over-the-counter medicines (UK Biobank Variable 20003). Number of medications was recoded into categories of 0, 1-2, 3-4, 5+ medications taken, consistent with recent NHS reporting. 8,9 Statistical analysis Overview.
There were two sets of analyses conducted: (i) an exploratory investigation of possible associations of endogenous hormone concentrations in UK Biobank men with incidence rate of SARS-CoV-2 infections; (ii) to investigate possible associations of endogenous hormone concentrations in UK Biobank men with incident risk of death from COVID-19. Endogenous hormone concentrations were measured at baseline (2006)(2007)(2008)(2009)(2010), with a focus on total testosterone, although analyses were replicated for SHBG, and calculated free testosterone (cFT), as additional exposures of interest.

Occurrence of infections across regions
Preliminary inspection of a heat map of the reported numbers of incident infections for the whole cohort (males and females combined) showed that the pattern of occurrence (infections) varied spatially during follow-up (Supplemental Figure S1). Accordingly, all analyses included an interaction term of spatial unit with time.

Associations of testosterone and SHBG with incidence of SARS-CoV-2 infection
Monthly incidence rates of SARS-CoV-2 infections (per 1000 person-months) were calculated by dividing the number of incident events in that month by the number at risk at the start of that month, then multiplying by 1000. For use in Figure 1 these were calculated using: where IR = incidence rate, h = hormone quintile, m = month, Count = number of participants with an incident infection, At Risk = number alive and still at risk at the start of that month.
For interpretability of Incidence Rate Ratios (IRRs), model predictors were modelled as categories in multivariable analyses of incidence rates, with the IR calculated as: where a = baseline age category (≤50, 51-60, >60 years), b = baseline BMI (<25, 25-<30, ≥30 kg/m 2 ), c = ethnicity (white, not white), d = educational attainment/qualifications (not completed college or university, college or university), e = Townsend Index quintile, h = baseline hormone (total testosterone, SHBG, or cFT) concentration modelled as quintile categories, r = Country (England, Scotland, Wales). Month was modelled using a natural cubic spline with a knot point set at the day preceding the introduction of the 'Rule of 6' social distancing measure (14 September 2020), the first of a series of national restrictions introduced to address the second wave of the epidemic. Zero-Inflated Poisson regressions were fitted to account for an excess of zero frequencies in the counts (response variable), with the logged number at risk modelled as an offset term. Kingdom  variables (history of angina, atrial fibrillation, cancer, cardiovascular disease, chronic   obstructive pulmonary disease, diabetes, hypertension, liver disease, renal impairment,   thyroid disease), and prevalent medication usage variables (anticonvulsants, lipid, glucocorticoids, opioids, number of medications used). Continuous predictors, including that for the exposure in each analysis (total testosterone, SHBG, cFT), were modelled using restricted cubic splines with boundary knots set at the 5 th and 95 th percentiles and inner knots at the 35 th and 65 th percentiles. 10

Vaccination availability in the United
The validity of the proportional hazards assumption was assessed using per-variable and global tests. 11 Plots of the Schoenfeld residuals, with estimated coefficients and 95% confidence intervals plotted against follow-up times, were inspected for statistically significant results. Test results were ignored when the detected temporal variation for proportional hazard estimates was shown to be negligible. 12 Violation of the proportional hazards assumption for the interaction term of UK region with time stratum was resolved by refitting these models instead with three time strata (Wave 1, 16 March 2020 -13 September 2020; Wave 2 pre-vaccinations, 14 September 2020 -7 December 2020; Wave 2 after commencement of vaccinations, 8 December 2020 -31 January 2021).
Hazard ratios (HRs) and 95% CIs were calculated from each of the fitted models, relative to the median of the fifth sample quintile (the reference value). Percentile bootstrap estimates of 95% CIs were calculated using 2,000 bootstrap iterations. HRs and 95% CIs relative to the reference value of the exposure variable were calculated and plotted against the exposure variable over a continuous range, to show non-linear effects in figures. HRs and 95% CIs associated with the change in hormone concentration from this reference value to the median of each of the other sample quintiles were shown in tables.  240-245 E00-E06 * = Additional data sources were also used for identifying prevalent conditions (e.g., from self-report medical conditions, self-report medication usage, physical examination and blood chemistry measurements). ** = In many cases only "I48" was provided but in others the full 4 digit code was provided. Characteristic § Table S4. Baseline characteristics of UK Biobank men stratified according to those who were tested for infection, or were infected with or died from COVID-19 during the follow-up period, and for the cohort as a whole**. Summary statistics presented for means and standard deviations of those continuous variables that are presented in Table 1. . ** = Summary data presented for data after excluding men who died or were lost to follow-up since their baseline visit but before 16 Mar 2020, with prior orchidectomy, taking androgens, anti-androgen, 5α-reductase, estrogen, anti-estrogen, progesterone medications, infertile men, men with pituitary disease, adrenogenital or testicular disorders, or variables with missing values. § = BMI, body mass index (kg/m 2 ); SHBG, sex hormone binding globulin; cFT, free testosterone calculated using the Vermeulen formula. § § = Incident infections identified for participants with a positive test result or who died from COVID-19 during the follow-up period from 16 Mar 2020 to 31 Jan 2021.     (  0.395 § = Hazard Ratios calculated for the medians of testosterone within each sample quintile (Q1-Q5), relative to the median for Q5. Quintile boundaries: Testosterone (nmol/L) Q1/2 9.0, Q2/3 10.8, Q3/4 12.5 and Q4/5 14.8 or (ng/dL) Q1/2 259, Q2/3 311, Q3/4 360 and Q4/5 427; SHBG (nmol/L) Q1/2 25.8, Q2/3 33.1, Q3/4 40.5 and Q4/5 50.8; cFT (pmol/L) Q1/2 171, Q2/3 201, Q3/4 230 and Q4/5 268. # = Model 1 included terms for testosterone and age and region, with time modelled as a 3-level stratification factor plus an interaction of region with time (see Methods). ## = Model 2 included Model 1 terms + ethnicity (white vs not white), living with partner, educational attainment, alcohol consumption, smoking status, diet (red meat: high vs low vs none), physical activity, BMI, waist circumference, cholesterol, time blood sample collected, blood type, Townsend Index quintile, hypertension, angina, atrial fibrillation, COPD, renal impairment, liver disease, thyroid disease, and use of lipid medications (a proxy for hyperlipidemia), glucocorticoids, opioids, and anticonvulsants, with the number of medications included as a proxy for overall comorbidity status. Continuous variables modelled using restricted cubic splines. 0.014 § = Hazard Ratios calculated for the medians of testosterone within each sample quintile (Q1-Q5), relative to the median for Q5. Quintile boundaries: (nmol/L) Q1/2 9.0, Q2/3 10.8, Q3/4 12.5 and Q4/5 14.8 or (ng/dL) Q1/2 259, Q2/3 311, Q3/4 360 and Q4/5 427. # = Presented estimates from Model 2, which included terms for testosterone and age and region, with time modelled as a stratification factor an interaction of region with time, ethnicity (white vs not white), living with partner, educational attainment, alcohol consumption, smoking status, diet (red meat: high vs low vs none), physical activity, BMI, waist circumference, cholesterol, time blood sample collected, blood type, Townsend Index quintile, diabetes, hypertension, angina, atrial fibrillation, COPD, renal impairment, liver disease, thyroid disease, and use of lipid medications (a proxy for hyperlipidemia), glucocorticoids, opioids, and anticonvulsants, with the number of medications included as a proxy for overall comorbidity status. Continuous variables modelled using restricted cubic splines (see Methods). 0.005 § = Hazard Ratios calculated for the medians of SHBG within each sample quintile (Q1-Q5), relative to the median for Q5. Quintile boundaries: (nmol/L) Q1/2 25.8, Q2/3 33.1, Q3/4 40.5 and Q4/5 50.8. # = Presented estimates from Model 2, which included terms for SHBG and age and region, with time modelled as a stratification factor an interaction of region with time, ethnicity (white vs not white), living with partner, educational attainment, alcohol consumption, smoking status, diet (red meat: high vs low vs none), physical activity, BMI, waist circumference, cholesterol, time blood sample collected, blood type, Townsend Index quintile, diabetes, hypertension, angina, atrial fibrillation, COPD, renal impairment, liver disease, thyroid disease, and use of lipid medications (a proxy for hyperlipidemia), glucocorticoids, opioids, and anticonvulsants, with the number of medications included as a proxy for overall comorbidity status. Continuous variables modelled using restricted cubic splines (see Methods).

End of follow-up: 28 Feb 2021 (including the tail of Wave 2)
Table S11. Fully adjusted hazard ratios estimating the relative risk of death from COVID-19 associated with baseline calculated free testosterone concentration (cFT; pmol/L), as estimated from analyses using alternative end of follow-up dates: 31 January 2021 (the main analysis); 8 December 2020; 28 February 2021. §# Figure S1. Distribution of incident SARS-2-CoV events for the whole cohort. Incident events are either the first positive test or a death from COVID-19 without prior positive test, and are calculated for males and females combined, excluding participants lost to follow-up before 16 March 2020 and events after 31 Jan 2021. Squares with darker shading indicate higher numbers of incident events and zeros indicate no incident events for individuals who attended that assessment centre at baseline (2006)(2007)(2008)(2009)(2010) during that month of follow-up. Assessment centres are ordered top to bottom by geographic location (north to south, west to east) for showing approximate spatiotemporal patterns of incident events in the cohort. Figure S2. Derivation of the study cohort.