The effects and safety of testosterone replacement therapy for men with hypogonadism: the TestES evidence synthesis and economic evaluation

Study design

Evidence was considered from randomised placebo-controlled clinical trials evaluating the effects of TRT in men with low testosterone. Only trials with a duration of at least 3 months for all intervention groups were considered suitable for inclusion; this was in line with the current recommendation of the UK British Society for Sexual Medicine (which recommends evaluating patients at 3, 6 and 12 months after TRT initiation and then every 12 months) and US Endocrine Society Clinical Practice Guidelines (which recommend evaluating men 3–6 months after TRT initiation and then annually thereafter). 30,31 Any relevant clinical setting (e.g. primary care, secondary care) was eligible for inclusion. Studies of a cross-over design were not considered suitable for inclusion, unless there were eligible groups with sufficient follow-up before cross-over occurred.

Target population

Adult men (aged 18 years or over with no upper age limit) presenting with a proven low level of serum testosterone. There has never been a consensus definition of low testosterone, which is reflected by the participant characteristics of trials in this field. However, all current clinical guidelines are in broad agreement that men with a serum level of TT > 12 nmol/l (350 ng/dl) are unlikely to have clinical features of low testosterone and do not generally require treatment. 30 This criterion was adopted in the present review.

Studies with eligibility criteria specifying free testosterone or bioavailable testosterone thresholds were eligible for inclusion in the review if they reported mean baseline TT of ≤ 12 nmol/l (or equivalent).

The original protocol for this review further specified the following criteria for trial eligibility, owing to the variability in testosterone assays:

• Clinical symptoms and/or signs of low testosterone (e.g. sexual dysfunction).

• The following information regarding testosterone samples must be available:

  • when samples were collected and assayed (since dates may differ)

  • details of any extraction method used prior to testosterone assay

  • details of the assay method and manufacturer

  • details of any local correction made to adjust assay measurements

  • relevant validation data for the assay (e.g. external quality assurance).

Initial screening of search results suggested that most studies did not report this information and, thus, would subsequently be excluded from the review. In consultation with the Advisory Group for this project, it was decided that such studies were relevant to address the research question of this evidence synthesis and had to be included, if they also fulfilled the remaining eligibility criteria. Therefore, those criteria were subsequently removed to widen inclusion of all relevant studies. Studies assessing any intervention(s) over and above testosterone and placebo were eligible for inclusion, but only data for the testosterone and placebo groups were used in the statistical analyses.

The following eligibility/inclusion criteria were further specified:

• Participants with hypogonadism caused by congenital disorders (e.g. Klinefelter syndrome) or acquired gonadal injury and participants with secondary hypogonadism (hypogonadotropic hypogonadism) were not deemed suitable for inclusion.

• Participants with concomitant medical conditions were considered suitable for inclusion only if their concomitant conditions were those that represent the constellation of low testosterone (i.e. obesity; type 2 diabetes; metabolic syndrome; osteopenia/osteoporosis; and/or history of fracture – for example, bone mineral density T-score < 2 and frailty).

Intervention

The intervention considered was TRT [also referred to as androgen replacement therapy (TRT)] with any testosterone formulation, dose, frequency and route of administration (e.g. intramuscular, subdermal, transdermal, oral and buccal preparations of testosterone). Studies involving previous treatment with finasteride or previously unsuccessful treatment with tadalafil were eligible for inclusion. Studies in which participants received other androgens apart from testosterone and studies that allowed concurrent treatment with other hormones or concomitant interventions alongside testosterone were not deemed suitable for inclusion.

Comparator

The eligible comparator treatment was placebo, which was required to be equivalent inactive treatment.

Outcomes

Outcomes of interest included: sexual function, physical health parameters, functional activities, psychological symptoms, CV and cerebrovascular (CBV) events, other comorbidities, prostate-related outcomes, physiological markers, QoL and mortality. The primary and secondary outcomes were identified and agreed upon by the members of the Advisory Group for this project before any statistical analysis was performed.

Aggregate data

A data extraction form was designed specifically for this assessment to collect aggregated data from all included studies, regardless of whether IPD could be obtained. Following piloting and further refinement of the form, one reviewer (MC) extracted details of study design, characteristics of studies, interventions and participants and outcome measures. A second reviewer (MA-M) cross-checked a random sample of 10% of selected studies. Extracted data were further checked for accuracy by the project statistician (JH). Any disagreements were resolved by consensus. Details of the items collected in the data extraction form are presented in Appendix 2.

Individual participant data

Anonymised data for each of the pre-specified variables were required for each randomised participant from as many identified studies as possible.

We established a collaborative group of the investigators of all identified trials. Contact information of authors of each eligible study was identified from the published report(s) and by electronic searches. All first authors of all eligible studies were initially contacted by email with a brief summary of the scope, rationale and objectives of the project, and invited to join the collaboration and share their anonymised IPD. Reminders were sent to non-responders after 1 week. Where reminders did not elicit a response, other methods of communication were attempted, including telephone calls, letter by post and attempts to contact other investigators of the respective studies. Once a memorandum of understanding was obtained, preferably electronically, investigators were asked to agree to share and transfer their trial data by signing a Data-Sharing Agreement form, which specified that only anonymised data were requested and accepted, stored securely and used exclusively for the purposes of the project. Trial data were considered unavailable in the event that no study authors had responded to multiple contact attempts or when the trial authors indicated that they had no longer access to the trial data. Trial authors were requested to provide data on demographic and baseline characteristics of participants and relevant outcomes (see ‘List of core items’ in Appendix 3 for further details).

A standard operating procedure (SOP) was developed specifically for the management of the IPD (see Appendix 4). A designated Gateway Manager (MC) was responsible for receiving the anonymised IPD from study collaborators and ensuring secure storage on a password-protected computer server area. Access to data at all stages of data process (checking and cleaning) and analysis was restricted to core members of the research team (MC, JH, MB, HR). Transfer of data to personal devices was forbidden at any time. Data were received in a variety of formats according to the security requirements of the countries from which data were received. In most cases, data transfer took place via an electronic system in which files were securely encrypted.

One-stage analyses

A one-stage meta-analysis approach involves fitting a regression model to the entire IPD data set rather than to each trial data set separately. This model accounts for the clustering of participants within trials. For the primary outcomes (all binary), a fixed-effect logistic regression model accounting for clustering and allowing a separate intercept per study was used. This is because a random-effects model – with a random intercept on study and random slope on treatment – failed to converge. Effect estimates were presented as odds ratios (ORs) and accompanying 95% CIs. For the secondary outcomes (all continuous), a random-effects linear regression model with a random intercept on study and a random slope on treatment was performed, accounting for clustering and allowing separate baseline adjustment per study as well as a separate residual variance using restricted maximum likelihood (REML). These effect estimates were presented as mean differences (MDs) and accompanying 95% CIs. The estimated between-study variance, τ², is reported to assess heterogeneity. The additional outcomes are summarised descriptively with a post hoc chi-squared analysis undertaken.

Two-stage analyses

Given that IPD were not available for all eligible studies, two-stage meta-analyses were also undertaken on all viable outcomes. Outcomes were analysed in their original trial and then combined in a meta-analysis to give an overall measure of effect. The first stage involves analysing IPD separately in each study to obtain aggregate data (i.e. the treatment effect in each study). For the primary outcomes, logistic regression models were fitted for each outcome. For secondary outcomes, linear regression models were similarly fitted as well as adjusting for baseline score. For studies with no IPD, we obtained effect estimates and standard errors according to current methodological recommendations, either directly from study publications or through communication with studies’ authors, as aggregated estimates. 34 In the second stage, the effect estimates from the IPD and aggregate studies were pooled together using a random-effects model with REML to produce effect estimates by study, IPD studies, aggregate data and overall. For models that would not converge using REML, then a random effects model using the DerSimonian and Laird method was used. Heterogeneity was assessed using the I² statistic.

Sensitivity analyses

Due to the low event rate for the primary outcome, mortality, a Mantel–Haenszel method was also performed for the two-stage analysis. 34,36 A sensitivity analysis was also performed for the primary outcome, CV and/or CBV events, by including unknown cause of death using the same analysis as described under the one-stage analysis.

For the secondary outcomes, glucose and HbA1c, a sensitivity analysis of removing participants with diabetes at baseline was performed as these participants could have been on medication which artificially lowers these outcomes.

Pre-specified subgroup analyses

Pre-specified subgroup analysis was conducted to explore possible treatment-modifying effects of diabetes, smoking status, testosterone and free testosterone levels on the primary outcome, CV and/or CBV events. A subgroup analysis for mortality proved unfeasible due to the limited number of reported events across trials. We originally planned to specify categories for testosterone and free testosterone; however, due to the sparse data available within each category, these were explored as continuous variables.

Subgroup by treatment interactions were assessed by including the within-interaction terms in the models outlined above. As for current methodological recommendations, continuous covariates were centred on the mean value within each trial and binary covariates were centred on the proportion within each trial. 37 Subgroup analyses were performed on the IPD studies.

A stricter level of statistical significance (two-sided 1% significance level) was applied given their exploratory nature and corresponding CIs were set at 99%.

Post hoc subgroup analyses

A post hoc analysis on possible modifying effects of age on CV and/or CBV events was undertaken. Post hoc analyses on the secondary outcomes were also undertaken following the same method described for the pre-specified subgroup analyses. It was decided to perform analyses on the most reported outcomes by the trials included in the IPD analysis or the outcomes with substantial heterogeneity (τ²) compared to other outcomes. The selected outcomes were AMS for QoL, IIEF-15 and IIEF-5 for sexual function, Hb for physiological markers and BDI for psychological symptoms.

We also perfomed a post hoc subgroup analysis based on IIEF-15 and its subscales assessing the effects of age, total serum testosterone and body mass index (BMI).

Threshold analysis

A regression analysis was performed to see whether there were any thresholds for IIEF-15 at follow-up for age, baseline total serum testosterone and BMI and for IIEF-15 at baseline for total serum testosterone. According to the number of categories identified, we performed either an analysis of variance (ANOVA) or t-test statistical analysis to confirm whether the categories were significant.

Individual participant data studies

Of the 17 studies that provided data for the IPD analyses, 6 were conducted in the USA,11,12,38–41 3 in the UK,42–44 1 in each of the Netherlands,45 Australia,46 Russia,47 Slovenia,48 Malaysia,49 Denmark50 and Norway,51 and the remaining study was conducted across nine countries (Argentina, Canada, Germany, Spain, Italy, South Korea, Puerto Rico, UK, USA). 52 The majority of studies were single centre,38,40,41,43–51 two studies involved 3 centres,12,39 one study involved 8 centres,42 one study 12 centres,11 and another study 98 centres. 52 Across studies, the total numbers of randomised participants ranged from 2751 to 79011 and duration of treatment from 12 weeks52 to 3 years. 38,39

Non-IPD studies

Seven of the 18 studies that did not provide data for the IPD analyses were conducted in the USA;53–59 3 studies were conducted in Italy;60–62 1 study was conducted in each of Mexico,63 Taiwan,64 UK,65 Canada66 and China;67 and 3 studies were conducted in multiple countries: Austria, Finland, Germany, Ireland, Italy, Spain, Sweden, UK;68 Belgium, France, Germany, Italy, Netherlands, Spain, Sweden, UK;69 and USA, Canada, Mexico. 70 Eight of the 18 studies were single-centre studies. 54–56,60,63,65 One study was conducted in 2 centres (40 randomised participants),64 and one study in each of 4 centres (58 randomised participants),66 36 centres (220 randomised participants),69 43 centres (406 randomised participants)59 and 63 centres (274 randomised participants). 57 The remaining five studies did not report the number of centres involved. Total numbers of randomised participants ranged from 1465 to 40659 and treatment duration from 3 months57,59,64,65 to 24 months. 67 Three studies that did not provide data for the IPD analysis were linked to pharmaceutical companies;59,68,69 two of these studies provided full disclosure of all serious adverse events (SAEs). 68,69

Individual participant data studies

Mean age of the testosterone and placebo groups was reported in 14 of the 17 studies that provided IPD11,12,38–40,42–45,47,49–52 and ranged from 51.6 years47 to 74 years12 years in the TRT group and from 52.8 years47 to 74 years12 years in the placebo group. One study reported an overall mean age of 60.2 years for all participants. 48 One study reported a median age of 62 years in both treatment groups46 and another study reported a median age of 68 years in the TRT group and 70 years in the placebo group. 41

Mean BMI was reported in 13 studies11,12,38,39,42,44,45,47–52 and ranged from 27.445 to 35.347 in the active treatment groups and from 27.345 to 34.247 in the placebo groups. Of these 13 studies, mean BMI was in the obese range (i.e. ≥ 30) in at least one arm of nine studies. 11,12,42,47–52 Median BMI was reported by two studies and was 31.5 and 33.4 in the active and placebo groups, respectively, in one study46 and 28.3 and 29.6, respectively, in the other study. 41 The remaining two studies did not report information on BMI. 40,43

Nine studies reported mean baseline testosterone in terms of nmol/l38,42–45,47–49,51 and five studies in units of ng/dl,11,12,39,40,52 which were converted to nmol/l for consistency. 71 Mean baseline testosterone in the TRT groups ranged from 6.747 to 11 nmol/l44,45 and the placebo groups ranged from 752 to 10.9 nmol/l. 44 Three studies reported median testosterone, ranging from 7.150 to 8.7 nmol/l46 in the TRT groups and from 8.546 to 9.4 nmol/l50 in the placebo groups.

Overall, the baseline characteristics of participants included in the IPD were well balanced between TRT and the placebo groups (see Table 5). The mean age was 65 years [standard deviation (SD) 11 years, 16 studies] with a mean BMI of 30 kg/m² (SD 5 kg/m², 17 studies). The majority of the participants were white (88%, six studies) and non-smokers (TRT 89%, placebo 87%, 10 studies). None of the participants had a diagnosis of prostate cancer. Further baseline characteristics can be seen in Appendix 6, Table 35.

Non-individual participant data studies

Mean age was reported in 17 of the 18 studies that did not provide IPD. 20,55–70 In the active treatment groups, mean age ranged from 47.9 years64 to 77.9 years,58 and in the placebo groups from 52.7 years70 to 54.5 years. 55 The remaining study did not provide information on the age of participants. 54

Mean BMI was reported by 13 studies. 53,55–61,63,66–68,70 In the active treatment groups, mean BMI ranged from 27.258 to 3955 and in the placebo groups from 26.658 to 39.4. 55 Of these 13 studies, mean BMI was in the obese range (i.e. ≥30) in at least one arm of eight studies. 55–57,59–61,66,70 Five studies did not report baseline BMI values. 54,62,64,65,69

Eight of the 18 studies reported mean baseline testosterone in units of nmol/l53,59,60,62,65,66,68,69 and nine studies in units of ng/dl,54–58,63,64,67,70 which were converted to nmol/l, as described above.

Mean testosterone in the active treatment groups ranged from 7.567 to 13.2 nmol/l,58 and in the placebo groups from 7.667 to 14.5 nmol/l. 58 The inclusion criteria in the study by Kenny (2010) stated that ‘men were selected for testosterone levels less than 350 ng/dl or bioavailable testosterone levels at least 1.5 standard deviations (SDs) lower than those of young adult men (95–350 ng/dl for men aged 40–49)’. 58 However, mean (SD) baseline testosterone levels reported were equivalent to 13.2 and 14.5 nmol/l, in the TRT and control groups, respectively. The reasons for this disparity are unclear and the study’s corresponding author did not respond to our clarification request. The remaining study did not explicitly report baseline testosterone status but stated that the mean across all groups at baseline was equivalent to < 11.1 nmol/l. 61

Individual participant data studies: published data and additional data from collaborators

A summary of the risk of bias assessments for the 17 trials that provided IPD, including data from publications and any further data received upon request from the collaborators, is presented in Figure 2. Risk of bias assessment of individual trials is presented in Figure 3.

FIGURE 2.

Summary of the risk of bias assessment of the 17 IPD studies (data derived from published reports and from contacting the collaborators).

FIGURE 3.

Risk of bias assessments of the 17 IPD studies (data derived from published reports and from contacting the collaborators).

For the ‘selection bias’ domain, the majority of trials were assessed as being at low risk of bias,11,12,38–40,44–48,50–52 with four trials reporting insufficient information on which to make a judgement. 41–43,49 For the ‘performance bias’ domain, two studies were rated to be at high risk of bias because the research pharmacist involved in the study was aware of the randomisation38,40 and four studies at unclear risk of bias because, despite being described as ‘double blind’, no clarification was provided on who was actually blinded. 41–43,52 The remaining 11 trials were assessed at low risk of bias for this domain. 11,12,39,44–51

The ‘detection bias’ domain was assessed at low risk of bias for most trials. 11,12,38–42,44–46,48–51 In three trials described as ‘double blind’, it was unclear whether the outcome assessor had been blinded. 42,43,52 For ‘attrition bias’, one trial was judged at high risk of bias due to the high number of dropouts. 12 Two trials did not provide sufficient information on which to make a robust judgement. 38,40 The remaining 14 trials were judged at low risk of bias for this domain. ‘Reporting bias’ was unclear for nine studies as it was not possible to check the respective protocols. 38–41,47,48,50–52 One study was assessed as high risk of reporting bias as some outcomes not specified in the research protocol were reported in the full publication (e.g. aerobic performance, liver volume) while other outcomes specified in the research protocol (i.e. tests of balance, reaction time, muscle volume, Psychological Well Being index, physical activity) were not. 12 The remaining seven studies were assessed as low risk of bias for this domain. 11,43–47,49 For the ‘other bias’ domain, two studies were judged at low risk of bias,38,40 while the remaining 15 studies were judged at high risk of bias due to financial or other connections with the pharmaceutical industry. 11,12,38,39,41–48,50–52 Overall risk of bias was judged to be unclear for 5 studies41–43,49,52 and low for 12 studies. 11,12,38–40,44–48,51

Individual participant data studies: published data only

Risk of bias of the 17 studies included in the IPD analysis was also conducted, based on data available in the relevant publication(s) only. A summary of the risk of bias assessments is presented in Figure 4. Risk of bias of individual trials is presented in Figure 5.

FIGURE 4.

Summary of the risk of bias assessment of the 17 IPD studies (data from published reports only).

FIGURE 5.

Risk of bias assessments of the 17 IPD studies (data from publications only).

In general, risk of bias assessment of published studies was not hugely dissimilar to that conducted after the IPD were obtained. In particular, further information obtained by the authors of six studies allowed us to re-classify a total of 15 domains originally assessed as ‘unclear’ risk of bias. 45–51 In all cases, the judgement was changed to low risk of bias. The most common domain for re-classification was ‘detection bias’, in which the blinding of the outcome assessor was unclear in the original publication but subsequently clarified by the respective authors. 45,48–51 Blinding of participants/personnel was confirmed by three authors,45,48,51 as well as random sequence generation and allocation concealment. 47,48,51

Non-IPD studies

A summary of the risk of bias assessments for the 18 trials that did not provide IPD is presented in Figure 6. Risk of bias of individual trials is presented in Figure 7.

FIGURE 6.

Summary of risk of bias assessment of the 18 Non-IPD studies.

FIGURE 7.

Risk of bias assessments of the 18 Non-IPD studies.

For the ‘selection bias’ domain, six studies were judged at low risk of bias for random sequence generation,53,54,56,63,68,70 with four of these six also being at low risk of bias for allocation concealment. 54,63,68,70 The remaining studies did not report sufficient information on which to base a robust judgement. 55,57–62,64–67,69

Three studies were judged at high risk of bias for ‘performance bias’: one study involved three randomised groups, of which one was open label and two were double blinded;59 one study was open label;67 and in a third study the pharmacist was not blinded. 66 Nine studies were judged to be at low risk of bias53,54,56–58,63,65,69,70 and the remaining six studies did not provide sufficient information on which to make a robust judgement. 55,60–62,64,68 Two studies were judged at high risk of bias for ‘detection bias’: one study involved three randomised groups, of which one was open label and two were double blinded59 and one study was open label. 67 Seven studies were judged to be at low risk in this domain53,54,56,58,63,69,70 and nine studies did not report sufficient information on which to base a robust judgement. 55,57,60–62,64–66,68 Eight studies were judged to be at high risk of bias for ‘attrition bias’ due to the numbers of dropouts. 53,55–58,64,68,69 Six studies were judged to be at low risk of bias54,60,63,65,66,70 and four studies did not provide sufficient information on which to base a robust judgement. 59,61,62,67 Two studies were judged to be at high risk of ‘reporting bias’: in one study, some outcomes that were assessed were not reported, for example, liver and kidney functions serum bilirubin, gamma-glutamyl transferase, serum glutamate oxaloacetate transaminase and serum glutamate pyruvate transaminase, albumin, and creatinine. Other outcomes were reported but were not specified as being assessed, for example, QUICKI index. 60 Similarly, blood pressure, urinalysis and urine flow rates were assessed but not reported in another study. 65 Reporting bias was unclear for the remaining 16 studies as it was not possible to check the respective protocols. 53–59,61–64,67–70 Three studies were judged to be at low risk of bias for ‘other bias’,60,63,67 three studies did not report sufficient information on which to make a robust judgement,61,62,65 and 12 studies were judged at high risk of bias due to financial or other connections with the pharmaceutical industry. 53–59,64,66,68–70 Overall risk of bias was judged to be high for two studies,59,67 unclear for 13 studies,53,55–58,60–62,64–66,68,69 and low for three studies. 54,63,70

In general, a comparison of the risk of bias assessments of the 17 IPD studies and the 18 Non-IPD studies showed a greater proportion of low risk of bias assessments for the IPD studies. However, the higher proportion of unclear risk of bias assessments of the Non-IPD studies makes for an imbalanced comparison, and no robust conclusions can be drawn regarding the relative risk (RR) of bias across the two classes of studies.

All-cause mortality

Overall effect

Based on the one-stage fixed-effects IPD meta-analysis, mortality from any cause (14 studies, 3158 men) was lower among participants treated with TRT [6/1621 (0.4%)] than placebo [12/1537 (0.8%)], but the 95% CI was wide (OR 0.46, 95% CI 0.17 to 1.24; p-value 0.13; Table 6). Causes of death included myocardial infarction, cancer and ruptured aortic aneurysm. Based on data we received, we could not determine the cause of eight deaths from three studies.

TABLE 6 - One-stage IPD meta-analysis for mortality from any cause

Of the 14 studies, 8 reported no deaths38,40,43,45,46,48,50,52 and 6 reported deaths. 11,12,39,44,47,49 For the remaining three studies we were unable to confirm whether any deaths occurred, therefore they were not included. 41,42,51

Note

Values are numbers (per cent) or numbers.

Outcome	Number of studies	TRT n/N (%)	Placebo n/N (%)	OR 95% CI	p-value
Mortality from any causea	14	6/1621 (0.4)	12/1537 (0.8)	0.46 (0.17 to 1.24)	0.13
Details		N = 6	N = 12
Myocardial infarction	3	2 (33.3)	2 (16.7)
Cancer	1	0 (0)	3 (25.0)
Ruptured aortic aneurysm	1	0 (0)	1 (8.3)
Constrictive pericarditis	1	1 (16.7)	0 (0)
Multiple organ failure	1	1 (16.7)	0 (0)
Venous thromboembolism	1	0 (0)	1 (8.3)
Unknown	3	2 (33.3)	5 (41.7)

For the two-stage analysis, the IPD studies and aggregate data gave a similar combined effect of OR 0.63, 95% CI 0.30 to 1.32; I² = 0%; however, the 95% CIs were wide for each of the individual trials for both the IPD and Non-IPD analyses (see Figure 10a). Details of the causes of death per treatment groups are shown in Appendix 6, Table 38. Sensitivity analysis using Mantel–Haenszel showed similar results (see Figure 8b).

FIGURE 8.

IPD meta-analysis for mortality from any cause; (a) two-stage analysis; (b) Mantel–Haenszel. MH, Mantel–Haenszel.

Cardiovascular and/or cerebrovascular events

Overall effect

In total, 13 studies with 3120 men reported whether CV and/or CBV events occurred within 12 months (see Table 7). There were 120/1601 (7.5%) participants in the TRT group and 110/1519 (7.2%) participants in the placebo group with no evidence of a difference between the two treatment groups (OR 1.07, 95% CI 0.81 to 1.42; p = 0.62). Results were not changed significantly by including follow-up time as a weight in the model (data not shown). Most of the participants had a CV rather than a CBV event, with some participants experiencing more than one event. The most frequent events were the following: arrhythmia, coronary heart disease (CHD), heart failure, myocardial infarction and valvular heart disease. The sensitivity analysis of including unknown cause of death did not change the results (OR 1.05, 95% CI 0.79 to 1.38; p = 0.740).

TABLE 7 - One-stage IPD meta-analysis for CV and/or CBV events

Of the 13 studies, 2 reported no CV and/or CBV events. 48,50 For the remaining four studies we were unable to confirm whether CV and/or CBV events occurred. 41–43,51

Some participants had more than one event.

One event was a ruptured aortic aneurysm.

Note

Values are numbers (per cent) or numbers.

Outcome	Number of studies	TRT n/N (%)	Placebo n/N (%)	OR 95% CI	p-value
Number of participants with a CV and/or CBV eventsa	13	120/1601 (7.5)	110/1519 (7.2)	1.07 (0.81 to 1.42)	0.62
Total number of a CV and/or CBV events	13	182	183
Number of participants with a CV event	11	107/120 (89.2)	105/110 (95.5)
Total number of CV eventsb	11	166	176
Details
Arrhythmia	6	52	47
Coronary heart disease	6	33	33
Heart failure	6	22	28
Myocardial infarction	7	10	16
Valvular heart disease	2	18	12
Peripheral vascular disease	4	8	14
Stable angina	5	7	7
Aortic aneurysmc	5	6	7
New angina	3	5	5
Unstable angina	3	2	4
Aortic dissection	1	2	0
Atherosclerosis	1	1	1
Cardiac arrest	2	0	2
Number of participants with a CBV event	11	15/120 (12.5)	7/110 (6.4)
Total number of CBV eventsc	11	16	7

For the two-stage analysis (see Figure 9), the effect estimates differed between the IPD studies (OR 1.03, 95% CI 0.77 to 1.38) and the aggregate data (OR 0.35, 95% CI 0.12 to 1.01). However, once combined, there was no evidence of a difference between the two treatment groups (OR 0.96, 95% CI 0.72 to 1.27). Further details are shown in Appendix 6, Table 37.

FIGURE 9.

Two-stage IPD meta-analysis for CV and/or CBV events.

Subgroup analyses

Figure 10 shows the pre-specified subgroup analyses (assessed at a stricter significance of 1%). Overall, there was no evidence of treatment–covariate interaction for any of the pre-specified subgroup analyses [diabetes (yes/no), smoking status (yes/no), testosterone (nmol/l) and free testosterone (pmol/l)] or for the post hoc analysis [age (years)].

FIGURE 10.

Subgroups for TRT vs. placebo.

Quality of life

Based on the one-stage meta-analysis, the SF-36/SF-12 norm-based scores (five studies) showed evidence of an overall difference in favour of TRT compared to placebo for three of the 10 subscales: the social functioning subscale (MD 1.74, 95% CI 0.14 to 3.34; p = 0.034), the role emotional subscale (MD 1.66, 95% CI 0.57 to 2.76; p-value 0.003), and the mental health composite score (MD 1.95, 95% CI 0.64 to 3.26; p = 0.004; see Table 8). There was little evidence of heterogeneity (τ² = 1.51, 0 and 0.20, respectively). For the remainder of the subscales, the MD favoured TRT but the 95% CIs were wide. The two-stage meta-analysis (see Appendix 6, Figure 23a–j) showed similar results but with varying heterogeneity (I² ranged from 0% to 68%).

Seven studies (938 men in total) provided IPD for AMS total score. The one-stage analysis (see Table 8) showed evidence of a difference in favour of TRT (MD −2.62, 95% CI −4.02 to −1.23; p < 0.001, τ² = 1.52). Of these seven studies, only five provided IPD for each of the AMS subscales, but all showed results in favour of TRT. For the two-stage analyses, four additional studies provided aggregate data for AMS total score but only two for AMS subscales (see Figure 11a–d). Overall, the results for the two-stage IPD analysis were consistent with the one-stage analysis although more heterogeneous. However, the aggregated data for AMS total score (and to some extent for the somatic and sexual subscale too) showed considerable heterogeneity compared to the IPD studies, with varying 95% CI widths, and with one study61 showing exaggerated benefit of TRT. Once combined in the final stage, this effect was dampened but the heterogeneity remained high.

FIGURE 11.

Two-stage meta-analysis for AMS. (a) Total score; (b) Somatic subscale; (c) Psychological subscale; (d) Sexual subscale. a, Data presented as change from baseline.

Other QoL outcomes are presented in Appendix 6, Table 40.

Sexual function

Five studies (1412 men in total) provided IPD for the assessment of sexual function using the IIEF-15. The one-stage analysis (see Table 9) for total score showed that compared to placebo, TRT improved sexual function (MD 5.52, 95% CI 3.95 to 7.10; p-value < 0.001, τ² = 1.17). For the individual IIEF-15 subscales, results were similar, showing evidence of a difference in favour of TRT. The two-stage analysis used four additional studies of aggregate data for IIEF-15 total score, erectile function, orgasmic function and sexual desire, and an additional three studies with aggregate data for intercourse satisfaction and two additional studies for overall satisfaction (see Figure 12a–f). In general, the two-stage analysis showed similar but more heterogeneous results (I² ranged from 0% to 97%). However, there were some differences between IPD and aggregate data with varying 95% CI widths for some of the scores.

FIGURE 12.

Two-stage meta-analysis for IIEF-15. (a) Total score; (b) Erectile function; (c) Orgasmic function; (d) Sexual desire; (e) Intercourse satisfaction; (f) Overall satisfaction. a, Data presented as change from baseline.

Sexual function was also reported by a further five IPD studies (442 men in total) using the IIEF-5, a shorter version of the IIEF-15. These studies showed no evidence of difference between treatment groups (see Table 9). No aggregate data provided further data and the two-stage analysis of the IPD studies showed similar but more heterogeneous results (I² = 75.6%) with wider 95% CIs (see Appendix 6, Figure 24).

Another sexual function scale, the ADAM, was reported by one IPD study (221 men in total). There was no evidence of a difference between treatment groups (MD −0.25, 95% CI −0.73 to 0.23; p = 0.308; Appendix 6, Table 41). Other sexual function outcomes reported by studies that provided IPD are presented in Appendix 6, Table 41.

Two-stage meta-analyses of the remaining outcomes are presented in Appendix 6, Figures 25–28. One study provided aggregate data for HED and Sexual Arousal, Interest and Drive Scale: for both instruments there was evidence of a difference in favour of TRT compared to placebo (MD 3.20, 95% CI 0.84 to 5.56 and MD 4.30, 95% CI 1.60 to 7.00, respectively).

Regarding the post hoc subgroup analysis, we did not find any treatment-modifying factors (see Appendix 6, Table 42). For the threshold analysis, we identified the following thresholds for age: 52, 70, 72 and 72.8 years, which were simplified into the following age categories < 52, 52–70 and > 70 (see Appendix 6, Table 43 and Appendix 6, Figure 29a). For total serum testosterone, a threshold was found at 9.8 nmol/l and 30.6 for BMI (see Appendix 6, Table 43 and Appendix 6, Figures 29 and 30).

Physiological markers

For the one-stage meta-analysis for level of testosterone (nmol/l), 16 studies (2308 men in total) provided IPD and showed evidence of higher testosterone levels in the treatment group compared to the placebo group (MD 7.24, 95% CI 5.07 to 9.41); p-value < 0.001, τ² = 17.01). Similar findings were observed for free testosterone (pmol/l) but with substantial heterogeneity (MD 186.40, 95% CI 115.91 to 256.90; p = < 0.001, τ² = 13741.90). For both cholesterol (mmol/l) and triglycerides (mmol/l), there was evidence of some difference and a degree of homogeneity (MD −0.15, 95% CI −0.20 to −0.10; p = < 0.001, τ² = 0.00 and MD −0.09, 95% CI −0.18 to −0.00; p = 0.044, τ² = 0.01, respectively). Similar results were observed for Hb and haematocrit. For HbA1c (%) and blood pressure, there was no evidence of a difference between treatment groups (see Table 10).

The results of one-stage meta-analyses for other physiological markers are shown in Appendix 6, Table 44.

The two-stage meta-analysis (see Appendix 6, Figures 31–43) showed similar results, but some of the individual studies had wide 95% CIs with varying heterogeneity. There were also some differences between IPD studies and aggregate data.

Psychological symptoms

Individual participant data for BDI was provided by three studies (246 men in total; see Appendix 6, Table 45). The one-stage analysis showed no evidence of a difference between the treatment groups (MD −1.10, 95% CI −2.49 to 0.30; p = 0.123, τ² = 0.71). Aggregate studies provided no further data. Regarding the remaining psychological outcomes, only one study provided IPD (see Appendix 6, Table 45). The two-stage meta-analyses are presented in Appendix 6, Figures 44–47.

Post hoc subgroup analyses

Figure 13 shows the post hoc subgroup analyses for the following outcomes: AMS, IIEF-15, IIEF-5, Hb (g/l) and BDI. Overall, there was no evidence of consistent treatment–covariate interaction between the effect of TRT and the considered subgroups [diabetes (yes/no), smoking status (yes/no), age (years), testosterone (nmol/l) and free testosterone (pmol/l)] for any of the assessed outcomes.

FIGURE 13.

Subgroups for TRT vs. placebo. (a) AMS; (b) the IIEF-15; (c) the IIEF-5; (d) Hb (g/l); (e) BDI.

Additional outcomes

Two IPD studies reporting diabetes complications showed no evidence of difference between the treatment groups (see Table 11). Similarly, no evidence of a difference between treatment groups was found for hypertension, venous thromboembolism or non-stroke CBV (see Table 11). Prostate cancer, which was assessed by four IPD studies and four aggregate data, also indicated no evidence of difference. For both oedema and high haematocrit reported by the IPD studies, there was evidence of raised levels in the TRT group compared to the placebo group.

Theme 1: Symptoms of low testosterone and impact on daily life

Within this concept, we considered any report of symptoms described by participants and perceived as associated with low testosterone. Several symptoms associated with low testosterone were reported across three of the included studies; these were broadly grouped into either physical or psychological domains. 81–84 For instance, the most frequently reported were physical symptoms (with five subthemes), which included a range of impacts from lack of energy, altered sleeping patterns, lack of strength, weight gain, altered sexual activity/desire and altered cognition (e.g. loss of memory). 28,81,82,84

Theme 2: Diagnosis of low testosterone

Any account of the challenges experienced by the participants with regard to the diagnostic strategies used to establish the presence of hypogonadism was recorded in this theme. Two studies reported the participants experience of getting a diagnosis of hypogonadism and the role of health professionals in this process. 27,28 It is worth noting that this information was obtained from the authors’ reporting of participants’ experience, rather than from quotes obtained directly from the participants. Szeinbach 2012 and Mascarenha 2016 reported that participants understood the importance of testosterone monitoring and stated it would be easy to obtain this information from their physicians. 27,28 Mascarenha 2016 also discussed the persistence of some participants, defined as ‘drug seekers’, to acquire and use TRT, irrespective of the advice of their physicians. 28 These ‘persistent’ patients were reported to have consulted multiple physicians until they found one ready to prescribe them TRT (regardless of the diagnosis). Mascarenha 2016 reported also that one participant who did not feel satisfied with his physician’s advice chose to increase his dose of TRT and subsequently, when he failed to perceive any immediate effects, requested switching products. 28

‘Both patients and providers participants mentioned that they know of primary care physicians or specialists who prescribe TRT without testing for low testosterone levels and based on informal discussions or e-mail communication.’

(Authors’ interpretation, Mascarenha 2016)

‘While only two participants were able to recall their testosterone levels, the other three participants understood the importance of testosterone monitoring and stated it would be easy to obtain this information from their physicians.’

(Authors’ interpretation,Szeinbach 2012)

Theme 3: Access to treatment information

In this theme, we recorded any account of participants’ experience of access to information on TRT. Some participants explained that access to TRT information could facilitate their eventual use of the therapy. Szeinbach 2012 observed that participants came to receive TRT via different routes: during a consultation [e.g. with their general practitioner (GP) regarding a related condition]; through posters at their pharmacy; through friends and co-workers; popular magazines; internet searching. 27

‘A couple [of] months ago, [I was] having some blood work done and read an article in Esquire magazine about testosterone. I asked my family doctor to have that checked.’

(No participant details; Mascarenha 2016)

Similarly, Mascarenha et al. reported that some participants felt that the marketing and advertisements ‘spoke to’ their perceived needs. 28 In particular, participants considered the information on improved sexual function and energy levels of particular interest. Mascarenha et al. maintained also that ‘While most patient participants found it easy to access information on the positive effects of TRT and how to acquire it, they seem to have little knowledge about its side effects or risks’. Some participants also conveyed the desire to receive more information from health professionals on the availability and effects of TRT. 28

Theme 4: Perceived effects of TRT

In this theme, we considered changes in symptoms that participants perceived as associated with TRT. Several perceived effects attributed to TRT were reported across four of the included studies; as per theme 1, these were broadly grouped into either physical or psychological symptom-modification. 27,81–84 We identified five interlinked subthemes, specifically; sexual desire/activity, strength/energy, emotional or affectional or well-being, cognitive function and general well-being. The most reported perceived effects were physical symptoms (three subthemes), which included a range of impacts including sexual activity outcomes, strength and energy improvement.

It is important to acknowledge that information on dosages, frequency and duration of TRT among participants was poorly reported or not reported across included studies, except for the study by Szeinbach 2012, where the TRT type, dosage and time of use were specified for each patient. For this reason, some perceptions regarding testosterone therapy and its effects might differ and might also influence the response of participants based on the dosing and duration of treatment.

Theme 5: Expectations, experience and preference of type of TRT

Three studies reported participants’ expectations, experience and preferences with regard to the type of TRT. Five subthemes were identified across the included studies, including: ease of administration, mode of administration, beliefs about effectiveness, perceived adverse effects, and costs.

Theme 6: Providers’ perceptions on TRT prescription.

The final theme relates to the perspectives of health professionals who prescribe TRT. Only one of the included studies, Mascarenha et al., conducted in Canada, assessed the perspectives of health providers (all types of clinicians, i.e. primary care physicians, nurses, pharmacists) about prescription of TRT. 28 In this study, providers expressed a high level of uncertainty and ‘diagnostic ambiguity’ with regard to the diagnosis of hypogonadism – particularly of late-onset hypogonadism (LOH) – and subsequent prescription of TRT, mainly due to the non-specificity of its symptoms. Some providers explained that there is little consensus on what constitutes low and normal levels of testosterone, described the diagnosis of asymptomatic patients as challenging, and suggested that cut-offs for normal ranges may vary between individuals.

‘Is your current testosterone too low for you? Or is it too low for what you are used to?’

(Primary care physician, Mascarenhas 2016).

It is worth noting that this study focused exclusively on the perspectives of health providers who prescribed TRT in their daily routine and did not include the influences of a non-prescription approach. In general, providers indicated that ‘clinical guidelines on the interpretations and administration of tests were perceived to be vague’ and described different preferences in terms of tests used to determine the level of testosterone (e.g. total serum testosterone levels versus bioavailable testosterone levels). Non-specialist providers (i.e. primary care clinicians) did not considered the timing of the test or repeat testing as critical for the diagnosis. Some providers were more inclined to request a total serum testosterone test because it was covered by the participant’s personal insurance plan and expressed some scepticism about the accuracy of bioavailable testosterone tests conducted in private laboratories. Some providers recognised that the number of patients who were keen to try TRT was increasing and reported also being aware of colleagues who prescribed TRT without testing the participants’ testosterone level.

While some providers expressed the desire to exclude other clinical conditions before starting to prescribe TRT, others preferred to start TRT straight away and monitor its effects before deciding whether testing for other clinical conditions was necessary.

In general, providers noted a rise in the availability of information on TRT, particularly the use of the concept ‘andropause’, exploited for marketing reasons. Some providers believed that there was not enough evidence to compare low levels of testosterone in men with the concept of ‘menopause’ in women.

Providers also described how the decision about the appropriateness of TRT was influenced by the evidence (or lack of evidence) on its safety. Some providers claimed that there were ‘myths’ about TRT safety, particularly about the potential harms or side effects of the use of TRT in patients with certain clinical conditions (i.e. prostate cancer). Others preferred a more cautious approach due to the lack of evidence on the long-term consequences of TRT.

‘I can see how someone might see the latest studies and say “my God, this is proof that [TRT] are dangerous”. Someone like me, who follows the literature closer, understands the potential risks and potential benefits.’

(Endocrinologist, no further detail provided)

Across providers, there were different perspectives on the ‘appropriateness of TRT’. While some specialists (e.g. oncologists) expressed the view that TRT had to be reserved only for ‘profoundly low’ cases, some primary care clinicians and specialists with an interest in men’s health believed that appropriateness of TRT could vary according to the symptoms and general health of patients and their testosterone serum results. Other primary care physicians and general urologists maintained that TRT was appropriate for treating any symptomatic patient with a low testosterone test result, regardless of their underlying clinical conditions.

Model structure

The schematic economic model structure is shown in Figure 17. Adult men with hypogonadism enter one of the two model strategies, SoC or TRT. These strategies have similar model structures. All individuals start at the No complications health state and when they experience CV or CBV events, move to the corresponding post-complication Markov state. The complications assessed in the model are the primary clinical outcomes of the IPD meta-analysis, which were organised into three main categories: cardiac pathology, pathology of the peripheral vascular system and pathology of the CBV system. The ‘cardiac pathology’ category included arrhythmia, CHD, heart failure, myocardial infarction, valvular heart disease, stable, new and unstable angina, aortic aneurysm and cardiac arrest; the ‘pathology of the peripheral vascular system’ category comprised peripheral vascular disease, aortic aneurysm, aortic dissection and atherosclerosis; and the ‘cerebrovascular system pathology’ category included stroke and transient ischemic attack. Mortality was considered with Death as an absorbing Markov health state. Markov cycle length was defined as 1 month.

FIGURE 17.

Simplified schematic of the economic model structure.

Population

The economic analysis was conducted for a cohort of symptomatic men with testosterone deficiency. To illustrate the age groups of clinical interest (under 60 years old, 60–75 years old and over 75 years old), three starting age categories were defined: 40-, 60- and 75-year-old people.

Time horizon and discounting

A 10-year time horizon was chosen for the economic modelling. Given the 3-year follow-up of the RCTs included in the synthesis of clinical effectiveness evidence and the discussions with the advisory group for this project, we believed that extrapolation of clinical effects from the IPD analysis beyond 10 years would be highly uncertain. Nevertheless, given potential for long-term differences in complications, survival and costs, a sensitivity analysis was conducted to assess the impact on cost-effectiveness of applying a lifetime horizon as recommended by UK economic evaluation methods guidelines. 114 A half-cycle correction was applied and future costs and QALYs were discounted at a rate of 3.5% per annum. 114

Clinical input parameters

Primary outcomes from the analysis of TestES data were used as key model parameters (see Table 20). The RR for all-cause mortality and the risk of CV or CBV complications from the one-stage meta-analysis were incorporated in the model. Log-normal distributions were constructed based on the 95% CI (RR mortality: 0.47, 95% CI 0.18 to 1.25; RR complications: 1.06, 95% CI 0.82 to 1.38).

The number of events in the TestES data set was used to proportionally allocate cost and reduced utilities in the model. The numbers of events for each type of complication were pooled across study arms (see Table 21) to avoid artificially biasing the cost and/or utilities towards one arm due to non-statistically significant differences in the proportional distribution of event types between treatment arms in the IPD.

The underlying risk of experiencing a CV or CBV event was derived from the British Heart Foundation Heart and Circulatory Disease Statistics 2020. 115 This compendium of statistics reports incidence rates per 100,000 adults per year for selected CV conditions (atrial fibrillation, heart failure, stroke, transient ischemic attack and peripheral vascular disease) by gender and age for the UK. The latest incidence rates for males available for 2017 were used and transformed into monthly probabilities (see Table 22) assuming a constant monthly rate.

Health state utilities

The TestES systematic literature review of RCTs identified the studies that collected QoL data. Special attention was given to those instruments from which direct measures of utility could be obtained such as EuroQol-5 dimensions (EQ-5D), SF-36 or SF-12. In addition, instruments used within the included studies related to sexual function, psychological function or QoL were checked against the Oxford database of mapping studies. 120 The Oxford database provides a readily available collection of all studies mapping patient-reported outcomes to the EQ-5D instrument. Full copies of the publications for the identified mapping algorithms were obtained with the aim of applying the mapping algorithms to estimate utilities for TRT and SoC groups using TestES IPD.

Appendix 7, Table 46 shows the data available from the received data sets that allowed the direct estimation of utility scores. Five study data sets provided SF-36 individual item data; however, one of these only presented baseline data (Snyder et al. 2016),11 leaving only data from four data sets for analysis. Data for baseline, 26 weeks (Basaria et al. 2015a, Emmelot-Vonk et al. 2008, Hildreth et al. 2013 and Magnussen et al. 2016),39,40,45,50 52 weeks (Hildreth et al. 2013),40 78 weeks and 156 weeks (Basaria et al. 2015a)39 were available. The 26-week data were regarded as the most robust as they were available for four studies, with the analysis conducted for N = 409 participants. The number of participants reduced substantially for later time points. Also, data on QoL at 1 year and beyond potentially reflect the reduction in QoL due to complications. The model accounts for utility reductions due to complications independently; therefore, there was a risk of double counting these utility reductions due to complications when using long-term trial data. Short form-6 dimensions (SF-6D) algorithms were used to estimate utility scores for each participant at baseline and 26 weeks. 121 A mixed-effect regression model (random effects on study and fixed effects on participants) was run to estimate a difference of 0.0036 (95% CI −0.012 to 0.019) in utility score between TRT and SoC. The regression predicted utility score was used for the SoC group (0.792) and the mean estimated difference was added to estimate the TRT score. Utility multipliers for TRT and SoC were calculated by dividing these utility scores by the population norm for the sample (i.e. 0.795 for 65- to 69-year-olds). 122 Finally, these utility multipliers were applied to the general population EQ-5D score formula proposed by Ara and Brazier (2010) to obtain the age- and male-specific utility score for the No complications health state for TRT and SoC, respectively. 123

In addition to the SF-36 data, the IPD sets received from the authors of three published studies presented data for the BDI score. 38,47,124 The BDI evaluates key symptoms of depression including mood, pessimism, sense of failure, self-dissatisfaction, guilt, punishment, self-dislike, self-accusation, suicidal ideas, crying, irritability, social withdrawal, indecisiveness, body image change, work difficulty, insomnia, fatigability, loss of appetite, weight loss, somatic preoccupation and loss of libido. 125 The self-rated scale comprises 21 items, each of which is scored individually from 0 (least level of difficulty) to 3 (most level of difficulty). Scores are directly summed, with the total score ranging from 0 to 63 (higher scores indicating greater depressive severity). Grochtdreis et al. 2016 provide prediction models to map from BDI score to EQ-5D utility scores based on a sample of 1074 consecutive patients with depressive disorders from a psychotherapeutic outpatient clinic in Germany. 126 The authors estimated five prediction models with varying independent variables: BDI index (model 1), BDI index and age (model 2a), BDI index and grouped age (model 2b), BDI index and gender (model 3) and BDI index, gender and age (model 4). The authors reported that Models 2a and 4 showed the best predictive abilities (lowest root mean square errors). Therefore, model 4, with explanatory variables for BDI score, gender and age, was used to estimate EQ-5D utility scores with the following equation:

Data assessed at 28,124 3047 and 3538 weeks received from the collaborators for the three relevant published studies were grouped and analysed as a single time point (N = 247). A mixed-effect regression model (random effects on study and fixed effects on participants) was run to estimate a difference of 0.0395 (95% CI 0.013 to 0.046) in utility score between TRT and SoC. The regression predicted utility score was used for SoC group (0.766) and the difference added to estimate the TRT score. Utility multipliers were obtained by dividing the utility scores for TRT and SoC by the EQ-5D population norm for the regression sample age. 123 These multipliers were applied to the Ara and Brazier general population formula to obtain the age-specific utility scores applied to the TRT and SoC No complications health states. 123 Table 23 shows the utility scores by starting age used in the model.

Further to the QoL data discussed above, five data sets presented data on the IIEF-15. While there is no mapping algorithm linking the IIEF-15 to the EQ-5D, the instrument has been used to develop health states that have been valued using the Time Trade-Off (TTO) method. 127 Stolk et al. (2000) conducted a cost-utility analysis of sildenafil compared with papaverine-phentolamine injections for the treatment of erectile dysfunction. The authors elicited 24 health states using two of the 15 questions from the International Index of Erectile Function (IIEF): the ability to attain an erection (question 3) and ability to maintain an erection (question 4). The use of these utility weights to value individuals’ health states according to IIEF data was discussed by the members of the advisory group for this project (4th Advisory Group Meeting, 2 December 2019). It was agreed that the quality-of-life dimensions relevant for the population of interest were broader than the two questions considered in the TTO exercise conducted by Stolk et al. 2000. Therefore, it was felt that the use of the IIEF data to estimate utility scores was less relevant that the directly measured sources (SF-6D) or the values obtained from the mapping algorithm through the effect on depressive symptoms.

Health service resource use and costs

Health state utilities and costs associated with complications

The unit cost and utilities associated with CV, CBV and the peripheral vascular system events were sourced through searches of technology appraisals, clinical guidelines and health technology assessments on the NICE and National Institute for Health and Care Research (NIHR) websites. These sources were favoured as they are based on comprehensive literature reviews related to the condition of interest, utilise large data sets of UK patients and have been used in the NHS decision-making process. Following the method used in NICE Clinical Guidelines (CG181), each complication is attributed a short- and long-term cost and utility multiplier. 135 Hence, patients accrue alternative costs and utilities in the short and long term for each condition depending on whether it can be considered an ongoing or instantaneous health event.

Model validation

Several steps were taken in order to secure the quality of the model. 150 The model structure was agreed with the members of the Advisory Groups for this project to secure the model structure face validity. White-box testing materialised throughout the whole model implementation (e.g. verification of formulae results with an external software) and black-box verification tests were conducted to assess the behaviour of the model specific input values (e.g. all utility scores equal to 1 to check whether total QALYs equal total life years). Markov traces were extracted from the modelling software and the cumulative proportion entering the complication states (i.e. Cardiac pathology, Pathology of the peripheral vascular system and the Pathology of the cerebrovascular system Markov states) were added up. These cumulative proportions at 10 years (120 cycles) were compared against the 10-year risk of experiencing a first CV event obtained from the QRisk3 CV risk calculator. 151 The model showed a cumulative proportion of 5%, 15.9% and 29.8% entered the model complication states at 10 years for the 40-, 60- and 75-year-old cohort, respectively. These figures were compared with a weighted average of the QRisk3 risk calculator score obtained for a white male with diagnosis of erectile dysfunction (73%) and a white man with diagnosis of erectile dysfunction and diabetes (27%). The model risk figures are higher than the weighted averages (i.e. 3.1%, 13.6% and 27.2% for 40-, 60- and 75-year-olds, respectively). However, the TestES model population included other comorbidities such as smoking resulting in a higher risk of CV and CBV events. The model data for the risk of CV and CBV event complications were sourced from the British Heart Foundation and were the latest data available, therefore no calibration was conducted.

Model analysis

The analysis captures the cumulative health and social care costs from the perspective of the NHS and QALYs accrued by patients under TRT and SoC strategies over a 10-year time horizon.

The model was run probabilistically to characterise the joint uncertainty in the modelled outputs (cost and QALYs) arising from the uncertainty in the input parameters. Monte Carlo simulation techniques were used to analyse the model many times (10,000 iterations), with sets of values drawn at random from the probability distributions assigned to each model parameter. To characterise the uncertainty in the mean parameter values, beta, gamma and log-normal distributions were used for probabilities, costs and relative effects, respectively. Normal distributions were used for the TRT utility difference estimated from the regression analysis of the TestES data and the utility multipliers attached to complications. Details of the probability distributions used are reported within Tables 20–26. Ten thousand iterations were deemed enough to obtain stable results. The outputs from these probabilistic analyses are presented as the probability of TRT and SoC being cost-effective at £10,000, £20,000 and £30,000 cost-effectiveness threshold values. Cost-effectiveness acceptability curves (CEACs) were also produced for selected scenarios to further illustrate the uncertainty around the model results. 152 CEACs present the probability of the compared strategies generating the greatest net monetary benefit for different cost-effectiveness thresholds (cost per QALY gained). In addition, incremental cost-effectiveness ratios (ICERs) were estimated to compare TRT against SoC. The ICER is defined as the ratio of the difference in expected costs over the difference in expected QALYs between TRT and SoC. As the model was run probabilistically, cost and QALYs were averaged across the 10,000 iterations and the ICER was calculated with the average difference in costs and QALYs between TRT and SoC.

Key assumptions

Results are reported for eight scenarios defined according to alternative assumptions around key model effectiveness parameters for TRT versus SoC: the definition of the utility difference between TRT and SoC, the RR of mortality and RR of CV, peripheral vascular and CBV complications. In these scenarios, the length of time at which the different effect estimates were applied varied: 12 months, 10 years, non-existent or patient lifetime. Results are reported for three age groups: 40-, 60- and 75-year-olds.

Details for the eight scenarios are defined below and summarised in Table 27:

• Scenario 1: the utility score difference between TRT and SoC was based on the SF-6D data derived from the analysis of TestES IPD. Moreover, the RR of mortality (RR mortality) was defined as 0.46 (95% CI 0.18 to 1.25) and the RR of experiencing a complication (RR complication) equal to 1.06 (95% CI 0.82 to 1.38). The model time horizon was defined as 10 years; however, all these effects were applied for the first 12 months only to be consistent with the clinical effectiveness primary outcomes.

• Scenario 2a: similar definitions for utility difference, RR mortality and RR complication as for scenario 1 were used. However, these effects were maintained for the whole 10-year time horizon.

• Scenario 2b: all definitions as for Scenario 2a except for the utility difference between TRT and SoC that was based on the BDI score mapped to the EQ-5D score.

• Scenario 3a: the utility score difference based on SF-6D was applied for 10 years. No difference in mortality or complication rates was considered (i.e. RR mortality equal to one and RR complication equal to one).

• Scenario 3b: all definitions as for scenario 3a except for the utility difference between TRT and SoC that was based on the BDI score mapped to the EQ-5D score.

• Scenario 4: the RR for mortality (0.46) was applied for 10 years. No utility score difference and no difference in complication rates (RR complication equal to one).

• Scenario 5: the RR for complications (1.06) was applied for 10 years; no utility score difference and no difference in mortality rates (RR mortality equal to one).

• Scenario 6a: the utility score difference based on the SF-6D and RR for complications (1.06) was applied for 1 year. No difference in mortality (i.e. RR mortality equal to one).

• Scenario 6b: all definitions as for Scenario 6a except for the utility difference between TRT and SoC that was based on the BDI score mapped to the EQ-5D score.

• Scenario 7a: the utility score difference based on SF-6D and the RR for complications (1.06) was applied for 10 years. No difference in mortality (i.e. RR mortality equal to one).

• Scenario 7b: all definitions as for Scenario 7a except for the utility difference between TRT and SoC that was based on the BDI score mapped to the EQ-5D score.

• Scenario 8: lifetime QoL and rate of complication differences were assumed. The SF-6D utility score difference was used as this was the smaller observed difference between TRT and SoC. No difference in mortality (i.e. RR mortality equal to one).

Modelling for 40-year-old men with MH

The discounted average costs for scenario 1 are £1126 and £4215 for SoC and TRT, respectively, giving an incremental cost of £3089. The average costs are very similar for scenarios 1–7 where a 10-year time horizon was assumed. The Monte Carlo simulation seed number was not fixed for each model run; therefore, the small variation in the SoC average costs for scenarios 1–7 is explained by sampling variation. Discounted average QALYs for scenario 1 are 7.414 and 7.424 for SoC and TRT, respectively, resulting in a small QALY gain of 0.010 from TRT. The ICER for scenarios 1 is well above the usual cost-effectiveness thresholds used by government decision-making bodies such as NICE in the UK (i.e. £20,000 per QALY gained). 114 The probability of TRT being cost-effective is zero at any threshold value because of the short term assumed for differences in the key parameters.

The differences in QoL, mortality and rate of complications were maintained for 10 years for scenario 2. While incremental cost and QALYs increased for this scenario, the increase in QALY is proportionally larger, reducing the ICERs for this scenario compared with scenario 1. However, the probability for TRT being cost-effective at £20,000 remains low (14%). BDI-based utilities were utilised in scenario 2b. This resulted in larger utility differences and greater QALY increments between TRT and SoC. Therefore, an ICER of £10,878 is obtained and the probability of TRT being cost-effective rises to 95% at £20,000 cost-effectiveness threshold. The QALY increment, ICER and the probability of being cost-effective are consistent for all the scenarios that assumed BDI-based utilities lasting 10 years (scenarios 2b, 3b, 7b).

The isolated impact of the difference in utility scores was considered in scenario 3. Cost and QALY increments reduced slightly compared with scenario 2. The ICER increased substantially for scenario 3a but not for scenario 3b; however, the probability of TRT being cost effective remains low for 3a and high for 3b, showing the limited impact that the differences in mortality and/or complications have for the cost-effectiveness of TRT in this age group. This is further illustrated in scenarios 4 and 5, where only mortality (scenario 4) or complication (scenario 5) differences were utilised. Both scenarios result in a zero probability of TRT being cost-effective at any threshold value.

A small number of deaths were reported within the TestES data set and the difference in mortality, upon discussion within the project team, was deemed as unreliable. Scenarios 6–8 show the impact of removing the mortality difference in the model. One- and 10-year QoL and rate of complication differences were assumed for scenarios 6 and 7, respectively. Results show an ICER below £20,000 only when the BDI-based utilities were assumed to last 10 years (scenario 7b). Finally, scenario 8 illustrates the effects of assuming a lifetime time horizon, and lifetime utility (SF-6D) and complication rate differences between TRT and SoC. Average cost and QALYs increased due to the longer time horizon considered. While the difference in cost doubled with respect to the shorter time horizon (scenario 7a), the difference in QALYs did not increase proportionally, resulting in a higher ICER (i.e. £119,436 compared with £91,580) and a probability of TRT being cost-effective of 11% at a threshold of £20,000.

Modelling for 60-year-old men with hypogonadism

Results for the 60-year-old group are reported in Table 10. Discounted average costs for this cohort are higher than those reported in Table 28 for the 40-year-old group due to the higher proportion of CV and CBV complications experienced by the older group. However, cost differences between TRT and SoC reduced slightly as the cohort ages. All-cause mortality increases for older groups, and therefore the 60-year-old cohort accrue relatively fewer QALYs. Crucially, as the underlying mortality for these age groups is higher, TRT appeared relatively more cost-effective for the scenarios that considered a reduction in mortality (scenarios 1, 2a, 2b and 4). Moreover, using BDI-based utilities (scenarios 2b, 3b, 6b and 7b) resulted in lower ICERs with respect to the corresponding scenarios using SF-6D utilities (scenarios 2a, 3a, 6a and 7a) and ICERs below the £20,000 threshold are obtained for the scenarios assuming long-term BDI-based utility differences (scenarios 2b, 3b and 7b). Furthermore, when the mortality difference was removed, the long-term BDI-based utility difference becomes crucial to cause the ICER to fall below the £20,000 threshold (scenario 7b). Finally, running the model for a lifetime time horizon did not improve the cost-effectiveness of TRT (scenario 8), as the cumulative difference in the incidence of complications starts to outweigh any ongoing gains in general health state utility associated with TRT use.

Modelling for 75-year-old men with MH

In Table 30, the results for the 75-year-old groups are presented. Discounted average costs are higher than those reported in Tables 9 and 10 for the 40- and 60-year-old groups, respectively. Again, this is due to the higher proportion of CV and CBV complications experienced by the 75-year-old group. However, while average costs are higher for SoC and TRT, the cost differences are reduced further compared with the younger cohorts. Moreover, the 75-year-old cohort accrues fewer QALYs as the all-cause mortality further increases for the older group. Notably, the ICER for TRT falls below £20,000 for scenarios that considered a reduction in mortality (scenarios 1, 2a, 2b and 4) as the underlying mortality for this age group further increases, and the absolute benefit associated with relative mortality reduction increases. As for the younger cohorts, the ICERs decreased when BDI-based utilities were defined (scenarios 2b, 3b, 6b and 7b) compared with the corresponding scenarios using SF-6D utilities (scenarios 2a, 3a, 6a and 7a) and they fell below the £20,000 threshold for those scenarios that assumed a long-term BDI-based utility difference (scenarios 2b, 3b and 7b). As with the younger cohorts, long-term BDI based utilities are vital for the ICERs to fall below the £20,000 threshold when the mortality difference is removed, (scenario 7b). Lastly, a lifetime time horizon model scenario (scenario 8) showed that TRT was not cost-effective when applying SF-6D utilities.

Cost-effectiveness acceptability curves

The CEACs for TRT for the alternative scenarios for the 40-, 60- and 75-year-old starting age cohorts are reported in Figures 18–20, respectively. The CEACs illustrate the decision uncertainty due to second-order uncertainty around the input parameter values. The curves show the probability of TRT being cost effective for a range of cost-effectiveness threshold values. The probability of TRT being cost-effective rises as the cost-effectiveness threshold increases, indicating that higher QALYs are obtained with TRT for almost all scenarios and age groups (see Figures 18–20). The exception is scenario 6, where only the difference in complications rates was assumed, which resulted in a greater number of TRT complications that were not overturned by a reduced mortality or a higher QoL.

FIGURE 18.

Cost-effectiveness acceptability curves for TRT for alternative scenarios – 40-year-old cohort.

FIGURE 19.

Cost-effectiveness acceptability curves for TRT for alternative scenarios –60-year-old cohort.

FIGURE 20.

Cost-effectiveness acceptability curves for TRT for alternative scenarios – 75-year-old cohort.

The CEACs for scenarios where BDI-based utilities were used were drawn with short-dashed lines. For the 40-year-old cohort (see Figure 18), these CEACs showed a high probability of TRT being cost-effective at a £20,000 threshold when the BDI-based utilities were applied for 10 years (scenarios 2b, 3b and 7b). The CEACs for the 60-year-old cohort group are presented in Figure 19. Similarly, those strategies that used BDI based utilities for 10 years showed a high probability of TRT being cost-effective at a £20,000 threshold. In addition, the scenarios assuming a RR for mortality of 0.46 for 10 years (2a and 4) resulted in an over 50% probability of TRT being cost-effective at a £20,000 threshold. A low probability of cost-effectiveness was observed for TRT for those scenarios where SF-6D and no difference in mortality were defined (i.e. scenarios 3a, 6a and 7a). Furthermore, the CEACs for the 75-year-old cohort (see Figure 20) showed a high probability of TRT being cost-effective for those scenarios where BDI utility and/or relative mortality differences were assumed to last 10 years (scenarios 2a, 2b, 3b, 4 and 7b). Lastly, a lifetime time horizon together with lifetime relative effect on the complication rate and a lifetime increment in SF-6D utility were defined for scenario 8. This scenario showed a very low probability for TRT being cost-effective regardless of the starting age of the population cohort (see Figures 18–20).

Development of the research question and outcome measures informed by patients’ priorities, experience and preferences

The research question was formulated by the NIHR HTA research panel, which included patient members. The TestES study team including two patient members, refined the research question and agreed upon outcome measures.

Involvement of patients in the design of this study

Patient members of the TestES study team were involved as research partners in all aspects of the study including refinement of the research question, identifying areas most in need for investigation and for providing ‘deciding votes and opinions’ where the study team was unable to reach consensus.

Were patients involved in the conduct of the study?

A panel of patients was convened to discuss findings of the evidence syntheses, and to guide the interpretation of findings by the research team. These opinions directly fed into the interpretation of all aspects of the study. Most notably, the patient panel members helped highlight a clear divergence in the experience of men with hypogonadism of non-sexual symptoms (e.g. poor cognition and low mood) supported by qualitative data, and the paucity of supporting evidence from several large RCTs. Furthermore, patients clearly report inconsistency and uncertainty about the threshold for treating hypogonadism with TRT. These opinions have directly influenced the conclusions and recommendations of this report.

Dissemination of findings to patients and public

The TestES investigators had originally planned to hold a public conference co-led by patients and clinicians to discuss the study findings. However, the changes imposed by the COVID-19 pandemic and feedback from our patient study team members has led us to develop a website providing textual and animated video-based information resources to highlight key findings from this NIHR-funded research. The study investigators have close links with UK-based professional and patient networks (e.g. Society for Endocrinology, You and Your Hormones), to ensure that the findings from TestES have impact far beyond the duration of this evidence synthesis.