NHANES is a cross-sectional, complex, multistage probability sampling survey conducted by the Centers for Disease Control and Prevention’s (CDC) National Center for Health Statistics (NCHS) designed to assess the health and nutritional status of non-institutionalised civilians living in the US.²⁸ During 1999–2002, 3569 women aged 20–49 were eligible for the survey, 2935 (82%) were interviewed and 2771 (78%) participated in the examination component.²⁹ The NCHS Ethics Review Board at the CDC approved the NHANES data collection and no specific additional review was required for this analysis, which used data from public-use files.

Persons aged 20 and over at the time of the interview were asked to provide whole blood for DNA analysis. At the Division of Health Statistics Laboratory, CDC, DNA was extracted from specimens and stored at −80°C; purified DNA samples were then coded and shipped to an outside laboratory (Dr. Elizabeth Blackburn at the University of California, San Francisco) for analysis as part of a surplus specimen project.¹⁶ The leucocyte telomere/standard (T/S) ratio was measured for each sample three times (in duplicate) using the quantitative polymerase chain reaction (PCR) method, resulting in six measurements which were used to calculate the mean and standard deviation of the T/S ratio for each participant.¹⁵ The T/S ratio (also referred to as ‘relative telomere length’) is directly proportional to mean TL and will be referred to as ‘telomere length’ throughout this manuscript for ease of understanding.³⁰ The mean T/S ratio can be converted to number of base pairs using the following formula: 3274 + 2413 * (T/S ratio).¹⁶

In addition to whole blood collection, the NHANES examination component consisted of physical measurements and serum and urine collection. Laboratory methods for biospecimen analysis have previously been described.^31,32 Approximately half of the participants were instructed to fast overnight before their examination appointment in order to ascertain fasting blood glucose and lipid levels.

Allostatic load has been operationalised in many different ways. Following previously published algorithms that have been applied to NHANES 1999–2002 data among reproductive-aged adults, we operationalised allostatic load scores in six different ways to examine whether associations with TL might differ according to the method of defining allostatic load. These six different scoring methods used either quartile-based^{22,24,26,27,33,34} or clinical-based^25,35 cutpoints to categorise individuals as low risk (score of 0) or high risk (score of 1) for each of the component biomarkers included its score. These component scores were then summed across the included biomarkers to generate a total allostatic load score. All six methods of defining allostatic load scores were therefore count-based, though each was constructed using a different set of biomarkers and/or cut-points (see Table S2 for biomarkers and cut-points used in each allostatic load score). The total number of biomarkers measured across all of the previously published scoring methods was 14, though each score summed over a subset of 9 or 10 biomarkers.

High-risk quartile cut-points were determined from the weighted distribution of each of the following biomarkers in our analytic sample: C-reactive protein (CRP), mg/dL; serum albumin, g/dL; body mass index (BMI), kg/m²; glycohaemoglobin (HbA1c), %; systolic blood pressure, mm Hg; diastolic blood pressure, mmHg; high-density lipoprotein (HDL) cholesterol, mg/dL; total cholesterol, mg/dL; triglyceride, mg/dL; homocysteine, μmol/L; pulse, beats/min; serum creatinine, mg/dL; urine creatinine, mg/dL and creatinine clearance, mL/min (estimated using the Cockcroft-Gault equation).³⁶ For the following biomarkers, high-risk groups were also determined based on clinically significant or empirically defined cut-points:^25,35 CRP, serum albumin, BMI, HbA1c, systolic blood pressure, diastolic blood pressure, HDL, total cholesterol, and pulse.

Participants reporting medication use for diabetes, hypertension, or high cholesterol, were assigned to the high-risk group for HbA1c; systolic and diastolic blood pressure; and total cholesterol, respectively. Ten-point allostatic load scores were calculated by summing membership in high-risk groups with each biomarker equally weighted (possible range 0–10); prior to summation, 9-point scores were rescaled to 10-point scores for comparison purposes.^25,26,35 Allostatic load scores were only calculated for women with non-missing values for all component biomarkers for each score.

Participant characteristics examined included age, Hispanic origin and race, marital status, smoking history, educational attainment, and household income as a percentage of poverty level. Pregnancy status was ascertained by combining information from the interview with results from a spot urine pregnancy test.

We limited our analysis to women 20–44 years old who provided DNA specimens for TL measurement and who completed the examination component of the survey. Pregnant women were excluded from our analysis because many allostatic load biomarkers are affected by pregnancy.²⁴ For the comparisons of TL and allostatic load component biomarkers, women missing individual biomarkers were excluded from that biomarker’s analysis but retained in other bio-marker analyses if information was available. All analyses accounted for the multistage, complex sampling design and used either the mobile examination centre weights or the fasting morning subsample weights (for analyses concerning triglycerides).³⁷ No adjustment was made for fasting, as recent evidence suggests fasting time shows little association with total cholesterol and HDL levels.³⁸

Telomere length was log-transformed. We used unadjusted linear regression to calculate mean TL by characteristics of study participants and for component biomarkers (low-risk quartile, 25th to 75th percentile, high-risk quartile). We performed significance testing of the difference in mean TL between the highest and lowest risk quartiles for each component bio-marker. In addition, differences in mean TL were estimated per 1 or 10 unit increase (depending on the range of biomarker values) in each allostatic load bio-marker using linear regression. Similarly, we estimated the difference in TL per 1 unit increase in allostatic load score. All regression models were further adjusted for age, which is positively associated with higher allostatic load scores and shorter TL, by including continuous age (in years) as a covariate. Linear regression coefficients were exponentiated to calculate the percent change in TL on the original scale for ease of interpretation. All P-values for general linear F tests were determined using the Satterth-waite adjusted F-test.³⁹

To assess differences in the relationship between TL and allostatic load biomarkers and scores by race/ethnicity, interaction terms for race/ethnicity (Mexican American, non-Hispanic white, non-Hispanic black) and allostatic load were added to regression models adjusted for age and race/ethnicity. We used the general linear Satterthwaite adjusted F-test to determine the significance of adding the interaction term to the model. To evaluate the effect of assigning high-risk group status based on medication use, we reran the models of TL and allostatic load scores after excluding women who had been assigned to high-risk groups based solely on their medication use. We also used multiple imputation to assign values to individual missing biomarkers so that allostatic load scores could be constructed for all women in our analysis; imputations used chained equations and predictive mean matching with demographics and non-missing biomarkers as predictor variables. Results using allostatic load scores based on these imputations were compared with the main results in a sensitivity analysis.

All analyses were conducted with SAS 9.3 (SAS Institute, Cary, NC, USA) and SAS-callable SUDAAN 11.0 (RTI International, Research Triangle Park, NC, USA).

There were 2386 women between the ages of 20 and 44 years at the time of the interview who participated in NHANES 1999–2002 and took part in the examination component. Of those, 1954 (82%) had leucocyte TL measured, 451 of whom were excluded from our analysis (444 were pregnant and 7 were 45 years old at time of the exam), leaving 1503 in our analytical sample. Among those eligible for TL measurement, non-Hispanic white women were more likely than non-Hispanic black women to provide specimens (Table S3). No significant differences in provision of specimens for TL measurement were observed by age, marital status, smoking, education, and poverty level.

Geometric mean TL (reported as T/S ratio), was 1.12 (95% confidence interval (CI): 1.08, 1.16), which was equivalent to 6045 base pairs (95% CI 5951, 6139). After adjustment for age, mean TL varied by race/ethnicity and marital status (Table 1). Non-Hispanic black women had longer telomeres compared with non-Hispanic white and Mexican American women. Never married women had longer telomeres compared with married, living with partner, or no longer married (separated, divorced or widowed) women.

There were 37, 72, and 13 participants (not mutually exclusive) reporting medication use for diabetes, hypertension, or high cholesterol, respectively, who were subsequently assigned to the high-risk group for the corresponding biomarkers. In most of these instances (75/122), the women were already categorised in the high-risk group for the respective biomarker.

Mean TL was shorter in the high-risk quartile compared with the low-risk quartile for CRP, BMI, and diastolic blood pressure (Table 2). After adjustment for age, mean TL was shorter in the high-risk quartiles for CRP and HDL. For the remaining biomarkers, there was no significant difference in TL between the high- and low-risk quartiles.

In models with component biomarkers as continuous linear variables, increases in BMI, HbA1c, systolic blood pressure, diastolic blood pressure, and total cholesterol were associated with shorter TL and an increase in HDL cholesterol was associated with longer TL (Table 3). However, all associations were null after adjustment for age except for HDL cholesterol, which showed a 1.2% longer TL (β^ (95% CI: 0.004, 0.020)) per 10 mg/dL increase and creatinine clearance, which gained significance showing a 0.4% shorter telomere length (β^ (95% CI: −0.008, 0.000)) per 10 mL/min increase. Interactions with race/ethnicity were significant for two biomarkers (Figure 1): HbA1c, which showed a negative relationship with TL for non-Hispanic whites compared with a flat slope for Mexican Americans and non-Hispanic blacks; and triglycerides, which showed a negative relationship with TL for non-Hispanic blacks compared with nearly flat slopes for non-Hispanic whites and Mexican Americans. For the remaining 12 biomarkers, differences in the relationship between TL and biomarker level were not observed among race/ethnicity groups.

Clinical cut-points were generally more stringent than quartile-based cut-points, which resulted in lower average allostatic load scores for clinical-based scoring methods (scoring method 2: mean = 1.34 (95% CI:1.25, 1.43); scoring method 4: 1.74 (95% CI: 1.64, 1.85)) compared with quartile-based scores (scoring method 1: 2.52 (95% CI: 2.34, 2.70); scoring method 3: 2.55 (95% CI: 2.41, 2.68); scoring method 5: 2.54 (95% CI:2.42, 2.65); scoring method 6: 2.31 (95% CI: 2.20, 2.42)). The standard error was largest for scoring method 1, in part because this allostatic load score included triglycerides which were only available from the fasting morning subsample and reduced the number of observations by approximately half. The allostatic load scores based on clinical cut-points shared the same nine biomarkers (but not the same cut-points), which differed from the quartile-based cut-point scores by not including measures of creatinine, triglycerides, or homocysteine (Table S2). Mean allostatic load scores were higher for non-Hispanic black compared with white women for all scoring methods (range of difference: 0.58–0.89); and Mexican Americans had a lower mean allostatic load score compared to non-Hispanic white women for scoring method 5 (difference = −0.40 (95% CI: −0.63, −0.16)); no other differences in allostatic load score by race/ethnicity were observed.

With the exception of scoring method 1, higher allostatic load scores were significantly associated with shorter TL in the unadjusted analysis (Figure S1). After adjustment for age, only allostatic load scores based on clinical-based cut-points remained associated with TL (−1.3% for scoring method 2 and −1.2%, for scoring method 4) (Figure 2). Differences by race/ethnicity for the relationships between TL and allostatic load score were not significant (all interaction term P-values >0.26).

After the exclusion of women with at least one high-risk grouping determined solely by medication use, the associations between TL and allostatic load became stronger for all allostatic load scores (all β^ were farther from zero). This exclusion resulted in three of the four quartile-based allostatic load scores (scoring methods 3, 5, 6) becoming significantly associated with shorter TL after adjustment for age. The estimated relationships between TL and allostatic load scores using imputed biomarkers were 0.1–0.2 percentage points stronger and estimated with greater precision compared to the main analysis (Figure S2).