Study population
The present study is a cross-sectional analysis conducted with baseline data from the participants of the Women's Health Study (WHS). The WHS is a nation-wide randomized-controlled trial of the efficacy of aspirin and vitamin E in the prevention of cardiovascular disease and cancer among women [17, 18]. Methods for participant recruitment have been described previously [19]. The baseline WHS cohort consists of 39,876 healthy middle-aged and elderly women aged 38 and older who were without cardiovascular disease or cancer at study entry between 1993 and 1996. At study entry, all participants completed questionnaires to survey their baseline demographic, clinical, and lifestyle/behavioral characteristics. Baseline blood samples were obtained from 28,296 participants to quantify levels of traditional and novel cardiovascular risk factors [18, 20]. The Partners Instituional Review Board reviewed and approved this study.
Analytic sample
Our initial sample consisted of all 28,296 partcipants with a baseline blood sample. Data were missing in 12.8% of the study cohort (n = 3,632). The largest source of missing data was item non-response due to missing survey data on the personal annual household income of WHS participants (n = 1,504 accounting for 5.3% of the data). Missing data on income were correlated with participant's age, cholesterol levels, body mass index and hsCRP, where those missing income data tended to be older, have lower total cholesterol levels, lower body mass index and lower hsCRP. Thus we imputed missing data for income in multivarible models, using procedures for multiple imputation [21] in SAS® version 9.2 (SAS Institute, Cary, NC) with the PROC MI and PROC MIANALYZE procedures.
Non-imputed samples (n = 24,664) excluding the 3,632 participants with missing income data were used in descriptive analyses. The final analytic cohort used in multivarible analysis consisted of 26,029 participants with complete data after imputing missing data on income.
Measures
Outcome measures: biomarkers of cardiovascular inflammation
The study outcome measures were biomarkers of inflammation as quantified by blood plasma levels of (1) hsCRP (mg/L), (2) sICAM-1 (ng/ml) and (3) fibrinogen (mg/dL), each assessed separately. Assays used to quantify biomarker levels have been described [20].
Key predictor variables: state-level macro socioeconomic conditions
The primary predictor variables were measures of state-level socioeconomic conditions: (1) state-level wealth and prosperity, (2) state-level labor productivity, (3) state-level poverty, (4) state-level income inequality and (5) state-level average annual economic growth. State-level wealth and prosperity were assessed with two separate measures (a) 1990 state-level real per-capita gross domestic product (GDP) from the Bureau of Economic Analysis, and (b) the 1990 US Census state-level median household income. GDP at the state-level in 1990 was calculated by the Bureau of Economic Analysis using the Office of Management and Budget Standard Industrial Classification (SIC) [22]. GDP data were computed from all industry activity within that year, accounting for inflation ("real GDP"), and scaled for state population ("per capita GDP") to allow comparison across states of different sizes [22, 23]. Median household income was taken from the 1990 US Decennial Census, and represents money income received in the 1989 calendar year from related and non-related household members aged 15 years and over. The US Census estimated median income at the state-level as calcuated from wages and salary income, self-employment income, interest income, dividends, rental and royalty income, and money income from social security, public assistance and welfare income [24].
State-level labor productivity describes the value of workers' output - i.e., what workers do - as a contribution to the wealth of the economy, [25, 26] in contrast to wealth gained through capital income [25, 27]. State per employee earnings has been suggested as a measure of labor productivity as it captures both the value of goods and services produced by workers, as well as the resources that accrue back to the employed population through wages and salaries [25, 28, 29]. Per employee earnings were calculated by the Morrison Institute for Public Policy using Bureau of Economic Analysis data on earnings from wages and salaries, proprietors' income, employer contributions to employee pensions and insurance payments, as distributed across the employed population of the region. We obtained state-level data on per employee earnings from the PEW Center on the States for this analysis [25, 26, 29].
Poverty at the state-level was obtained from the 1990 US Decennial Census long form survey, measured as the percentage of the total state-population with an annual household income under the 1989 federal poverty threshold, accounting for household size and age of the householder. State-level income inequality was measured as the Gini index of inequality. State-level Gini coefficients were obtained from the 1990 Decennial Census based on 1989 household income data from the Census long form survey [30].
The average annual growth in state-level real per capita GDP was used as the measure of economic growth at the state-level. Average annual growth statistics were calculated by the Bureau of Economic Analysis across the decade for which inflammatory biomarker data were collected in the study, between 1990 and 1996.
Individual-level covariates: cardiovascular risk factors and personal annual household income
Individual-level covariates thought to correlate with inflammation included in the analysis were: age, race/ethnicity (non-Hispanic White versus non-White race/ethnicity), body mass index (normal weight versus overweight and obese body mass index), low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), systolic blood pressure category, the presence of diabetes (defined by participant self-report and measured blood hemoglobin A1C equal or greater than 6.5%), frequency of exercise (recreational physical activity performed rarely or never, less than once per week, 1 to 3 times per week, four or more times per week), average daily caloric intake, smoking status (never smoked, prior smoker, current smoker), and the personal annual household income of the participant. Assays used to quantify the blood-derived measures were certified by the National Heart, Lung, and Blood Institute/Centers for Disease Control and Prevention Lipid Standardization Program [20].
Statistical analysis
Descriptive data
We present descriptive means, medians and percentages of demographic, behavioral and clinical characteriscs of WHS participants by state-level median household income, as well as ranges of all the state-level socioeconomic measures. We used Spearman rank correlation coefficients calculated in SAS® to assess correlations among state-level socioeconomic measures.
Multilevel associations with personal income and state-level socioeconomic conditions
We hypothesized that biomarkers levels would vary depending on both state-level socioeconomic conditions as well as the participant's personal household income [16]. Thus, we present figures describing the median values of biomarkers of inflammation within quartiles of state-level socioeconomic conditions, and across categories of personal household income. To test for a multilevel effect of state-level socioeconomic conditions on inflammatory biomarker levels in excess of personal household income, we used the PROC MIXED procedure in SAS® to estimate associations between biomarkers of inflammation and state-level measures, adjusted for individual-level personal income and covariates. Standardized beta coefficient fixed effect estimates with 95% confidence intervals are reported from multilevel models. Due to the known skew toward lower values, hsCRP was log-transformed prior to statistical analysis.
Adjustment for covariates and handing of missing data in multivariable models
In multivariable models, we used a propensity score predicting body mass index (normal weight versus overweight and obese status) to account for the causal relationships between adiposity and several metabolic and behavioral variables (exercise, caloric intake, HDL-C, LDL-C, diabetes, systolic blood pressure) as they relate to inflammation [31]. We report multivariable analyses with imputed data for income. Sensitivity analyses in multivariable multilevel models showed that effect estimates were not substantively different in models with and without missing data on income.