Study population
Our study population consisted of the female participants of the Doetinchem Cohort Study. The Doetinchem Cohort is a population-based cohort, whose participants were randomly recruited from the Doetinchem area of the Netherlands in 1987 [22]. The objective of the Doetinchem Cohort Study is to observe the impact of lifestyle and biological factors on chronic disease occurrence and quality of life [22]. At the time of recruitment, participants were aged between 20 and 59 years. After the baseline visit (round 1), participants were invited for follow-up every 5 years. At the time of the study, rounds 1 through 5 had been completed, leading to an approximate follow-up time of 20 years.
At each visit, lifestyle, general health, and reproductive history were assessed through extensive questionnaires, and biometric and laboratory measurements were performed. In addition to the laboratory measurements that were performed directly after each consecutive blood withdrawal, aliquots with additional plasma samples of each participant were immediately stored for future use. All participants provided written informed consent and ethical approval was granted by the Medical Ethics Committee of the Netherlands Organization of Applied Scientific Research. The use of stored sample specimens was ethically approved by the Ethical Committee for Biobank Studies of the University Medical Center Utrecht.
Only female participants from the Doetinchem Cohort with at least one available stored plasma sample, regardless of their age or menopausal status, were eligible for the current study. Of the total number of 4128 participating women, 3326 had an available plasma sample for at least one of the follow-up rounds. Rounds 1–5 comprised plasma samples of 3133, 2914, 2507, 2324, and 2051 women, respectively.
AMH measurements
The plasma samples from round 1 were stored in EDTA aliquots at –30 °C. The samples derived from rounds 2–5 were stored in EDTA aliquots at –80 °C. Prior to the current study, the samples were thawed once for additional measurements and immediately refrozen. For the current study, stored plasma samples of rounds 1–5 were utilized. In March 2015, all the available samples of each participant had been retrieved from storage and were shipped on dry ice to AnshLabs (Webster, Texas, USA), where they were temporarily stored at –20 °C until the analyses were performed. AMH levels were measured with the picoAMH assay (AnshLabs), because of its low limit of detection and the small aliquot size necessary, which is crucial for cohort studies with a limited pool of biological samples. The plasma samples of each individual were measured in a single assay run, by a single laboratory operator. In total, two laboratory operators performed all measurements. At a mean level of 91.2 pg/mL, the coefficient of variation was 4.0 %. At 290.3 pg/mL, the coefficient of variation was 4.8 %. The limit of quantification was 3.0 pg/mL and the limit of detection 1.8 pg/mL. There were no indications of plate drift, with all coefficients of variation within plate columns and rows under 5 %.
Time to menopause
Age at the time of the final menstrual period (FMP) was assessed by taking into account questionnaire information of cycle regularity, number of menstrual periods in the prior 12 months, oral contraceptive (OC) use, pregnancy, reproductive surgery, and self-reported age at menopause. Due to slightly differing questionnaires throughout the follow-up rounds, the assessment of the timing of the FMP differed per round. The earliest estimation of the timing of the FMP was considered to be the most accurate, being the most proximate to the event. Time to menopause was calculated by subtracting a participant’s age at the FMP from her age at follow-up. Women who ever underwent a bilateral oophorectomy were excluded from this calculation in order to obtain the time to natural menopause at each follow-up round. Women who underwent a hysterectomy before the onset of natural menopause were considered to have an unknown age at menopause.
Missing data
For information on smoking, OC use, menstrual cycle regularity in rounds 1, 4, and 5, age at menarche, and body mass index (BMI), the percentage of missing information was below 2 %. In rounds 2 and 3, missing information for cycle status was 6.8 % and 14.6 %, due to missing information of the date of the last menstrual period. Missing information for hormone replacement therapy use increased with each round, and varied between 7.1 % and 59.8 %. Multiple imputation through predictive mean matching with 10 iterations was performed for these variables, including participant ID, age, and AMH levels solely as predictor variables, and all remaining variables both as predictors and outcomes. Multiple imputation was performed with R (http://www.R-project.org), using the ‘mice’ library (http://www.jstatsoft.org/v45/i03/).
Assessment of individual decline rate: parallel or non-parallel trajectories
To assess whether the decline rate of AMH differed for individuals, AMH trajectories in relation to age and time to menopause were fitted with a mixed model approach using the ‘lme4’ package in R. Mixed models enable the evaluation of multilevel longitudinal data, and are thus able to take into account multiple measurements over time for each participant, with varying AMH levels (i.e., random intercept) and decline rates (i.e., random slope) for each individual. As AMH had a skewed distribution, AMH levels were logarithmically transformed. Levels below the detection limit of 0.0018 ng/mL were set at this level for the purpose of logarithmic transformation. LogAMH was used as the outcome of the mixed models, with chronological age or time to menopause as the time variable and participant ID as the group indicator variable. We modeled age and time to menopause with non-linear natural splines and checked the significance of non-linearity (a P value of < 0.05 indicated significant non-linearity). Models were adjusted for current OC use and OC use 5 years prior, hormone replacement therapy use, and current smoking and smoking 5 years prior, as these determinants were associated with the longitudinal AMH levels (data not shown). To decide whether women had differing age-specific AMH levels, differing decline rates, or both, the multivariable-adjusted models with a random intercept, slope or both, were compared to models with only fixed terms for these two parameters. The Akaike Information Criterion (AIC) of the models was used for this purpose. A lower AIC by at least 2 points represented a significantly better fit of the data.
Assessment of the consistency of age-specific AMH levels: does high remain high and vice versa?
To get an indication of whether individual AMH levels that were relatively low or high based on age remained comparatively low or high as time progressed, women were divided into age-standardized AMH quartiles in round 1. The CG-LMS method, previously described in detail by Dólleman et al. [23], was used for age standardization. The AMH decline trajectory of women in these four quartile groups was then plotted against chronological age and time to menopause.
In order to measure whether the AMH levels of the individual participants remained on a single trajectory, and did not vary between the 95th and 5th percentile over time for example, the variance of AMH measurements within and between individuals was assessed for the final mixed models. By dividing the between-individual variance by the total variance (between-individual + within-individual variance), the intraclass correlation coefficient (ICC) was calculated. The ICC gives an indication of the correlation of AMH measurements on each individual’s trajectory, which is directly relative to the amount of variation between individuals. For AMH decline with age, we hypothesized that women would follow a consistent high or low trajectory, i.e., that the variance of their AMH levels around a trajectory would be low. We therefore postulated that most of the variance would arise from differences in AMH levels between individuals, in which case the ICC would approach 1. For AMH decline with time to menopause, we hypothesized that there would be little variance of AMH levels between individuals; for example, we expected that the AMH level at 10 years before menopause would be roughly similar across the whole group. In this case, the ICC would approach 0.
Participant involvement
Participants of the Doetinchem Cohort were not directly involved in the formulation of the study question or realization of the study design. As this was a population-based study design, patients were not involved.