Practice of Epidemiology Reproducibility of Estradiol and Testosterone Levels in Postmenopausal Women Over 5 Years: Results From the Breakthrough Generations Study

Prospective cohort studies examining sex hormones in relation to cancer risk have generally collected blood samples at 1 time point, with an assumption that hormone levels measured in these samples will be reliable markers of true levels at other times. In postmenopausal women, body fat is a major source of estradiol; therefore, changes in adiposity may affect the correlation of single measurements to more relevant long-term averages. To estimate the intraclass correlation coefficient (ICC) for estradiol and testosterone, we collected repeat blood samples from 119 postmenopausal women (average age = 59.4 (standard deviation, 4.7) years) from the United Kingdom during 2004 – 2005 and again during 2010 – 2011. The ICCs (adjusted for assay variation) were 0.73 (95% confidence interval: 0.63, 0.82) for total estradiol and 0.59 (95% confidence interval: 0.47, 0.72) for total testosterone. The ICCs were 3% – 5% larger after adjustment for change in body mass index (weight (kg)/height (m) 2 ) or leptin, which are 2 markers of change in adiposity. There was no increase in ICCs after adjustment for change in age, alcohol consumption, smoking, exercise, time between waking and blood collection, or season. The results suggest that other factors account for within-woman variation in these sex hormones.

Elevated levels of estradiol and testosterone are important risk factors for postmenopausal breast cancer (1)(2)(3)(4), but these hormones can be measured only if biological samples have already been collected (5). Prospective cohort studies or biobanks are, therefore, the only way to gather such information, usually by collecting and storing blood samples for the whole cohort at recruitment and later measuring sex-hormone levels on a nested case-control or case-cohort basis. However, for reasons of cost and practicality, large cohort studies have generally collected blood samples at only 1 point in time, with the implicit assumption that levels of sex hormones measured in these blood samples will be reliable markers of relevant exposure levels over a longer period if necessary after adjustment for predictable trends (e.g., estradiol levels in premenopausal women over the menstrual cycle (6)(7)(8)).
The degree to which a single exposure measurement may represent a woman's long-term exposure level or ranking at other times is usually measured by the intraclass correlation coefficient (ICC) (9), the ratio of the between-women variation to the total between-and within-woman variation, which may also be used to make corrections for measurement error (10). A number of studies have evaluated the ICC for estradiol (4,(11)(12)(13)(14)(15)(16) in repeat blood samples from postmenopausal women. The correlation is imperfect (estimates vary from 0.36 (11) to 0.69 (4)), and some of the within-woman variation might be attributable to changes in factors affecting estradiol production or metabolism. Adipose tissue is a major source of endogenous estradiol in postmenopausal women (17), but it is not known whether changes in adiposity can account for some or all of this within-woman variation. We therefore evaluated the ICCs for estradiol, testosterone, and sex-hormone binding globulin (SHBG) over a 5-to 6-year interval in a sample of postmenopausal women from the Breakthrough Generations Study, adjusting for changes in adiposity measured by changes in both body mass index (BMI) (weight (kg)/height (m) 2 ) and plasma leptin.

METHODS
Subjects were selected from a United Kingdom cohort study, the Breakthrough Generations Study (18), from more than 100,000 women from whom questionnaires and blood samples were obtained, along with informed consent, at recruitment during 2003-2013. Repeat questionnaires were administered 5-6 years later, and at the same time, repeat blood samples were collected from a subset of more than 8,600 women. The study was approved by the South-East Multicentre Research Ethics Committee (London, United Kingdom). Samples were mailed to our laboratories and centrifuged upon receipt, and the plasma was stored in liquid nitrogen tanks. Women were eligible for the current analysis if 1) they self-reported natural or surgical menopause occurring after age 45 years; 2) they were more than 2 years postmenopause at recruitment; 3) they were not taking hormone replacement therapy at the time of blood collection; 4) they had not reported cancer or other serious illness at recruitment or follow-up; 5) they were under the age of 80 years at follow-up; and 6) the number of days between blood collection and freezing was the same for their recruitment and follow-up samples. We selected a random sample of 122 subjects from eligible women.
Plasma estradiol was measured by radioimmunoassay after organic extraction (19). Assay methods for testosterone, SHBG, and leptin have been reported previously (20). To reduce variation introduced by differences between assay batches, we sequentially analyzed samples taken at recruitment and follow-up in the same batch for an individual women, with ordering of samples within pairs randomized and blinded to the technician. Three women were excluded after these assays because they had elevated estradiol levels inconsistent with postmenopausal status (>110 pmol/L), leaving 119 women in the study. For all but 4 women, the recruitment and follow-up samples both arrived at our laboratory on the day after blood collection, and for the remaining 4, they arrived 1 day later. Each blood sample was measured in duplicate (except for 2 testosterone and 1 leptin assay for which only a single measurement was available because of insufficient material for duplicate measurement). One leptin sample from recruitment could not be measured because of contamination. Concentration of free sex hormone was estimated from total hormone concentration and SHBG (21).
Duplicate assay measurements were transformed to logarithmic scale, and intraassay, within-woman, and betweenwomen variances and the ICC were estimated by using the MIXED procedure in SAS statistical software (22). The ICC between the recruitment and follow-up blood sample is ICC ¼ Variance ðbetween womenÞ Variance ðbetween womenÞ þ Variance ðwithin womenÞ : The coefficient of variation (CV) was estimated from the logarithmic scale variance (23) as The ICC was adjusted for change in BMI or leptin using previously derived associations (20) by first estimating the change in sex-hormone level associated with an individual woman's change in BMI or leptin and subtracting this change from her measured level. We used the same methods to adjust for changes in age, alcohol consumption, smoking, exercise, time between waking and blood collection, and season. We did not adjust for absolute level of BMI, leptin, age, alcohol consumption, smoking, exercise, time between waking and blood collection, or season at either recruitment or follow-up (i.e., for differences in these variables between women) because this would reduce only the between-woman variation and artefactually shrink the ICCs. All P values are 2-sided.

RESULTS
The first blood sample from the 119 women was collected during 2004-2005 when the women were, on average, 7.0 years postmenopause (range, 2-24 years) and, on average, aged 59.4 (standard deviation, 4.7) years. Further descriptive characteristics of the 119 women are presented in Web Table 1 (available at http://aje.oxfordjournals.org/). The average plasma leptin level was 11.5 (standard deviation, 8.3) ng/mL at recruitment; the change from recruitment to follow-up was −0.46 ng/mL (95% confidence interval (CI): −1.46, 0.54); and the ICC was 0.77 (95% CI: 0.69, 0.83) over the 5-to 6-year follow-up period. The average BMI at recruitment was 25.2 (standard deviation, 4.3); the average change in BMI was −0.01 (95% CI: −0.37, 0.36); and the ICC for BMI was 0.89 (95% CI: 0.85, 0.92). Table 1 shows the distribution of the estradiol, testosterone, SHBG, and calculated free sex-hormone concentrations at recruitment and at followup 5-6 years later. Figure 1 shows the correlation between the 2 measurements for total estradiol. Estradiol levels within women increased by 10.1% (P = 0.063), and there were smaller increases in testosterone and SHBG, but none of these reached statistical significance.
Web Table 2 shows the CVs attributable to laboratory assay measurement, within-woman variation, and betweenwomen variation. For all analytes, the assay CV was much smaller than the within-woman CV, which in turn was smaller than the between-women CV. Table 2 shows the ICCs over the 5-to 6-year period. The unadjusted ICCs were 0.73 (95% CI: 0.63, 0.82) for total estradiol, 0.59 (95% CI: 0.47, 0.72) for total testosterone, and 0.87 (95% CI: 0.83, 0.92) for SHBG. The ICCs were larger for the estimated free sex hormone by 5% for free estradiol and 13% for testosterone, and they were larger for greater duration since menopause (for estradiol, P = 0.024; for testosterone, P = 0.029). ICCs calculated using different methods and with and without adjustment for assay batch were similar and are presented in Web Table 3. For completeness, Web Table 4 presents ICCs calculated with no transformation (arithmetic scale).
After adjustment of analyte levels at the second blood collection for a woman's change in BMI or leptin between the 2 blood samples, all ICCs were larger, but only by 3%-5%. We also adjusted for changes between recruitment and follow-up in age, alcohol consumption, smoking, exercise, time between waking and blood collection, and season, but we did not see any increase in the ICCs.

DISCUSSION
We found that the ICC for total estradiol over a 5-to 6-year interval was higher than reported previously (4,(11)(12)(13)(14)(15)(16) (other studies generally also used log transformations (4, 12-16)), even though the interval between sample collections in our study was larger than in most prior studies (11-13, 15, 16), and our interval between menopause and first blood sampling was shorter (4,(11)(12)(13)(14)(15)(16). The large ICC we observed may be a consequence of some of the conditions and criteria we used, which may not have always been applied previously. We used stringent eligibility criteria to ensure that women in this study were truly postmenopausal and not perimenopausal or taking hormone replacement therapy. Although long-term storage of sex hormones at less than −70°C is recommended (5), we stored samples at temperatures lower than −170°C to reduce as far as possible any sample degradation. We used an indirect extraction-based radioimmunoassay for estradiol that has been shown to be more sensitive than direct assay measurements of postmenopausal levels (24), and which showed a high correlation (r = 0.94) when compared with gas chromatography tandem mass spectrometry (25). We carried out assays for the same woman in the same batch and allowed for between-women batch variation by adjusting for batch effects in the statistical analysis. Furthermore, we explicitly estimated laboratory assay measurement error and separated this from the within-woman and betweenwomen variations to avoid attenuation of the ICC (26). Therefore, it seems unlikely that our estimate of the ICC was biased by the particular group of women we selected, sample storage, limitations of the estradiol assay, or method of analysis.
The only study that used the same estradiol assay we used here reported a lower ICC of 0.56 (14) but stored blood samples at −20°C and reported an unexplained significant upward drift in estradiol (and testosterone) levels with storage time. Because the interval between samples ranged from 5-12 years, this drift may have affected their ICC estimation. Abbreviations: CI, confidence interval; IQR, interquartile range; SHBG, sex-hormone binding globulin. a Averaged over 2 duplicate measurements except in the case of 2 testosterone samples for which there was only enough material for a single measurement. Measured values were used even if they fell below the formal limit of detection.
b Limit of detection: 3 pmol/L (5 women had measurements falling below the limit of detection). c Estimated from total hormone concentration and SHBG. d Limit of detection: 0.14 nmol/L (no women had measurements falling below the limit of detection). e Limit of detection: 1.3 nmol/L (no women had measurements falling below the limit of detection).
Total Estradiol at Follow-up, pmol/L  Cauley et al. (11) reported particularly small ICCs of 0.45 and 0.36 over 4 weeks and 2 years, respectively, but in the assay they used, 50% of the estradiol concentrations were at or below assay sensitivity, and this would have distorted their results. Toniolo et al. (15) stored samples at −80°C for 1-2 years and reported an ICC of 0.51 but used an immunodirect assay that produced nonphysiological results in postmenopausal women (27). Subsequently, samples from their casecontrol study were reassayed (12), and the ICC from a larger series was 0.66, which is similar to the ICCs from the remaining published studies that stored samples at −70°C (16) or in liquid nitrogen-cooled tanks (4, 13) and used radioimmunoassay (13) or liquid chromatography mass spectrometry (4, 16) (range of reported ICCs, 0.65-0.69). The ICCs for total estradiol from these studies are consistent with the unadjusted ICC we found: 0.73 (95% CI: 0.63, 0.82). Our ICC for testosterone (0.59 (95% CI: 0.47, 0.72)) fell within the lower range of, but was consistent with, previously published estimates (range of reported ICCs, 0.57-0.91 (11)(12)(13)(14)(28)(29)(30)), suggesting that methodological differences may be more important for estradiol. Among postmenopausal women, change in BMI is associated with change in estradiol (20). Thus, changes in amount of body fat may affect the correlation between 2 estradiol measurements taken years apart in the same woman. We found that after adjustment for change in BMI, the within-woman variation was reduced (and, as expected, the assay and between-women variation were unchanged), and all ICCs became larger, but by a small amount. BMI has well-established limitations as a measure of adiposity (31); therefore, we also used plasma leptin levels as a more direct measure of body fat (32). After adjustment for change in leptin, the ICCs were similar to those after adjustment for change in BMI.
The lack of a substantial increase in the ICC after adjustment for change in adiposity suggests that changes in other factors may be responsible for fluctuations or long-term changes in postmenopausal estradiol and testosterone levels. We had information on age, alcohol consumption, smoking, exercise, time between waking and blood collection, and season (which may be related to sex-hormone levels (33)(34)(35)(36)(37)(38)), but after adjusting for changes in these factors, we did not see any increase in the ICCs. Other factors may be related to sex-hormone levels (e.g., consumption of grapefruit juice (39), dairy products (29), or exposure to light at night (40)), but we did not have information available to allow adjustment for any changes in these exposures. If, however, we could gain better measurement of long-term sex-hormone levels by reducing measurement error, we might expect to see stronger associations between sex-hormone levels and breast cancer risk. For example, pooled results from 9 prospective studies (3) suggest that the relative risk for breast cancer, when comparing total estradiol in the lowest quintile with that in the highest quintile, is 2.0, but using methods to correct for measurement error (10), we estimate it could be as large as 2.9 in the absence of within-woman variability.
In conclusion, over a 5-to 6-year period the ICCs for estradiol and testosterone were 0.76 and 0.62, respectively. The ICCs were not greatly affected by changes in adiposity or other measured characteristics. The remaining unexplained within-woman variability may be the result of the cumulative effect of many small influences that together appear to act randomly, or the variability may be due to trends in sexhormone levels influenced by factors of which we are currently unaware or for which the effect cannot yet be predicted with sufficient accuracy.

ACKNOWLEDGMENTS
Author affiliations: Division of Genetics and Epidemiology, The Institute of Cancer Research, London, United Abbreviations: BMI, body mass index; CI, confidence interval; ICC, intraclass correlation coefficient; SHBG, sex-hormone binding globulin. a ICC calculated from variances estimated on a logarithmic scale from the MIXED procedure in SAS, version 9.2, software (SAS Institute, Inc., Cary, North Carolina) using a first-order autoregressive covariance structure.
b Adjusted for assay batch as a fixed effect. c Includes 3 women for whom years from first blood sampling to menopause was not known precisely but was 2 or more. d Weight (kg)/height (m) 2 . e Estimated from total hormone concentration and SHBG.