The non-linear risk of mortality by income level in a healthy population: US National Health and Nutrition Examination Survey mortality follow-up cohort, 1988–2001

Despite longstanding knowledge of an inverse association between income and mortality in the United States [1, 2] and calls to reduce socioeconomic gradients in health [3], few studies have examined whether the higher risk of mortality at lower incomes is uniform across the income distribution or whether threshold points exist in this association. Several studies from the UK and Scandinavia suggest that a threshold point does not exist in the association between social class and mortality [4‐6], and this has been used to inform policy for reducing health disparities worldwide [7]. A linear risk of mortality across the income distribution implies that there are etiologic pathways from income to mortality for all individuals, regardless of income level. Conversely, studies from the US and New Zealand have found primarily non-linear associations between income and mortality [8‐10]. If the association is non-linear and the income-mortality gradient exists mainly at lower income levels, research to understand income differences in health should be among this population. In addition to guiding etiologic research of income-mortality associations, the shape of association has implications for efforts to reduce income-based disparities: either providing evidence in support of targeting the full population (if a linear association) or evidence for support of policies focused on the population where such associations are found to exist.

Within the US, only two studies have explicitly examined the shape of the relationship between income and mortality [9, 10] and both found suggestive evidence of non-linear associations, with a higher level of income more strongly associated with increased mortality at the lower end of the income distribution. These studies were limited, however, by: (1) a lack of access to finely-measured income data (thus unable to determine if threshold points exist in the risk of mortality); (2) an inability to exclude persons whose declining health status may have reduced their income level [11], and (3) methods that did not allow for the testing of statistically significant non-linear associations and threshold points in change in mortality risk.

The aim of this paper is to describe the shape of the income and all-cause and cause-specific mortality associations among US adults age 18 to 64 at baseline (who were age 25–77 by the end of follow-up). We examined the association of income and mortality, restricting our analysis only to those individuals who were free from health conditions that caused them to change jobs, change to part time work, or retire early due to health reasons. By using data with a large number of income categories and by modelling the association without using a pre-specified functional form or pre-specified inflection points we are able to more accurately estimate the shape of the income and cause-specific mortality associations. We also compare the fit of models with baseline covariates and either a linear income term, a log-income term, or a smoothed spline income term in order to determine which income-mortality model provides the best fit to the data.

The US Third National Health and Nutrition Examination Survey (NHANES III), 1988–1994, was designed to be representative of the non-institutionalized population of the U.S. when analyzed using weights to account for over-sampling and non-response [12]. Our analysis used data from the National Death Index (NDI), for deaths through December 31, 2001 linked to NHANES III data, for a length of follow up of up to 13 years. The NDI is a well documented and validated method of matching deaths in the United States to population datasets [13, 14]. The definition of cause of death has been validated to have a discrepancy rate of approximately 5% [15]. Only NHANES III participants at least 18 years of age at the time of interview were eligible for follow-up, and we limited our analyses to individuals that were under the age of 65 at baseline (n = 14,798) (individuals missing at the time of follow up (0.3%) were excluded). We excluded the following individuals to minimize causation from health to income: individuals who self-reported that they "changed permanently to an easier job," "changed temporarily to an easier job," "cut down to part-time work only," "have to stop working for a few months," or "retire because of a disability," for the reason of "a disability or health problems" (10%) [16]. We also excluded individuals who died after one year or less of follow up (0.3%), as illness in the year prior to death may have reduced family income. Individuals who were missing income (7%), education (0.05%) or occupation (1%) were also excluded. The sum of these exclusions resulted in a final analytic sample of 11,733 individuals (106 037 person-years) with 349 deaths (table 1).

We examined all-cause mortality and three cause-specific categories of adult mortality as defined by the following ICD-10 classifications: 1) heart disease (I00-I09, I11, I13, I20-I51), 2) cancer (C00-C97) and 3) injury (both intentional and unintentional) (V01-Y34, Y85-Y86, Y87.0, Y87.1, Y87.2, Y89.9).

Total combined pre-tax family income for the 12 months prior to the survey included wages, salaries, income from self-employment, veteran's benefits, interest dividends, rental income and public assistance. Family income data were available in 28 income categories (none, less than $1000, $1000 to $20,000 in $1000 dollar increments, $20,000 to $50,000 in $5000 dollar increments, and greater than $50,000). Income from each half of the survey (1988–1991, 1991–1994) was adjusted to 1991 dollars using the Urban Consumer Price Index. For all analyses we used the midpoint of each income category and calculated the mid-point of the upper category of income ($66,800) by assuming a Pareto distribution of family income per standard methodology [17]. Income (as well as reference points of median income and poverty line) is made equivalent for economies of scale by dividing family income by the square root of family size. In this study we conceptualize income as a measure of the level of earnings of a household at a particular point in time, a value that has been empirically shown to have a strong association with lifetime earnings[18]

Additional covariates included: (a) education (0–17 or more years), (b) race/ethnicity (using the available categories of white non-Hispanic, Hispanic, black, other), (c) age (in years), and (d) occupation, referring to the longest held occupation, divided into 5 categories: (1): white collar and professional (e.g. executive, management, professional, supervisors); (2): white collar, semi-routine (e.g. technicians, sales workers, secretaries); (3): blue collar, high skill (e.g. mechanics, construction trades, military); (4): blue collar, semi-routine (e.g. personal services, waiters and waitresses, food preparation, farm workers, motor vehicle operators); and (5): never worked. Detailed NHANES III occupational categories were used to create this variable [see Additional file 1]. Labor force participation was defined as working at a job or business or having a job or business within the last two weeks, not including work around the house.

In order to be consistent with prior work on the shape of the association of income and mortality [6, 8, 9] we modelled mortality for women and men separately. The three models that we present include the following covariates: model 1) income, age, race/ethnicity, marital status, model 2) income, age, race/ethnicity, marital status, education, occupation, model 3) income, age, race/ethnicity, marital status, education, occupation, labor force participation. The estimated income coefficient from model 1 is not conditional on other measures of socioeconomic position, and thus may not adequately control for confounding of the income-mortality association. The estimated income coefficient from model 2 is conditional on other measures of socioeconomic position (i.e. education and occupation), and is the primary model for inference and for which associations are shown in the figures. Model 3 additionally controls for current labor force participation, as this may confound the association of income and mortality, but may also be on the causal pathway between income and mortality and thus should not be included as a control variable [8].

The income and cause-specific mortality associations were modelled with penalized splines (with a cubic basis) in proportional hazard survival models in order to allow for possible non-linear dependence of mortality hazard on income (as well as for education when included as a covariate)[19, 20]. We use age as the time-scale of the baseline hazard rather than time-on-study as this allows for less biased model estimates with respect to age, and allows for non-linear associations between age and mortality [21, 22]. Penalized splines model the association by fitting a large number of regression splines joined at evenly spaced knots, with a penalty on coefficients of adjacent knots [23]. The starting point for the model fitting procedure is to fit a saturated model of a large number of splines, and then to penalize the fit based on a smoothing parameter (theta) to adjust the penalty term to avoid over fitting the data. Prior work has shown that by using this method smoothed estimates of survival are insensitive to the number and location of knots [23]. Theta can vary between 0 and 1, where 0 is close to an exact fit to the data, and 1 is a straight line. The initial values of theta were chosen by generalized cross validation – an automated model fitting procedure where the optimal value of theta is determined by repeatedly fitting the model on random subsets of the data and then testing how well the model predicts the data that was not used to create the model [20]. Based on subject matter knowledge, we then increased the thetas to be slightly closer to 1 (a more conservative model) to avoid models that overfit the data, although fit was very similar in both sets of models (model fit from the theta chosen by generalized cross validation results available from the authors). All models we present are thus more conservative models for identifying non-linearity then those selected by generalized cross-validation.

We first present the unadjusted incidence rates of all-cause mortality by gender and income (table 2). We then present the Chi-square statistics and associated two-sided p-values from adjusted spline models in order to allow for assessment of the degree of difference in spline slopes for income parameters, values that may be interpreted as tests of non-linearity (table 3) [20]. We also present likelihood ratio tests comparing fitted models with a spline term to: 1) models controlling for baseline covariates, 2) models controlling for baseline covariates and a linear income term, and 3) models controlling for baseline covariates and log transformation of income (table 4). A likelihood ratio test comparing each fitted model to a null model is used to determine overall model fit. The null hypothesis for this comparison is no difference in model fit, therefore p-values of less than 0.05 reject the null hypothesis and indicate that the fitted model is a better description of the data. Due to the different degrees of freedom in each of the models, it is most appropriate to compare p-values across models to evaluate relative model fit. All analyses were conducted using R version 2.5.1[24] The proportional hazard model was implemented using the function coxph(), and the penalized spline fit was implemented using the pspline() function[25] Plotted standard errors were calculated by a grouped jackknife estimate implemented using the cluster() function, and all models employed weighting to account for the complex sampling design. Survey adjusted population frequencies and histograms were calculated using the survey package [26, 27]. This research was approved by the Harvard School of Public Health's Committee on Human Subjects.

Among US women and men age 18 to 64 at baseline, with follow-up of up to 13 years, we found evidence of a generally stronger associations of income with all-cause mortality at the lower end of the income distribution, i.e., under median income. Similar patterns occurred for deaths due to cancer and injury; by contrast, a more linear association across the full income range was evident for death due to heart disease. These results are unlikely to be substantially driven by contemporaneous effects of illness on income because of the restrictions of our sample to individuals with more than one year of follow up who had not ever changed jobs, changed to part-time work, or retired early due to health reasons. In fact, our results are likely a conservative estimate of the association due to the potential effects of income on illness, given that we restricted our analysis to a healthy sample that has not left the labor force due to health reasons.

Our results for all-cause mortality are generally consistent with the previously reported logarithmic functional form of association with income [9, 10]. However, our results expand on this observation in three important ways. First, we find that the shape of association varies by type of mortality. Second, based on the overall fit of models with either a log function of income or a non-linear function of income, the appropriateness of a log function for modelling income and mortality holds most closely for only cancer mortality among men and injury mortality among women. Third, based on visual inspection of the plotted income-mortality hazard ratios we identify median income level as a critical point of inflection for outcomes where substantial non-linear associations were found.

These results should be considered within the context of several study limitations. First, the data we used lack specific income categories above $50,000 a year, which limits our understanding of the impacts of income for the 8% of families that had the highest equivalized income. Our estimates at the upper end of the income distribution are less precise, as indicated by the widening confidence intervals in the plots, and the point estimates in these regions should be interpreted cautiously. A second limitation of this analysis is that income is measured at only one point in time, thus not capturing household income dynamics that influence health outcomes [29‐31]. This also results in a potential mismatch between income and social class related exposures and the relevant time periods for disease etiology. However, as income level of households are correlated across time, a recent measure of income does capture an important aspect of households' socioeconomic position [30], and one that is strongly correlated with permanent income. A third limitation is that as with any non-experimental study the associations observed may be due to residual confounding rather than a causal relationship between income and mortality. For example, we lack information on wealth and debt, additional markers of socioeconomic position that have been demonstrated to affect health [28, 32]. This may result in some residual confounding by social class that may contribute to bias in our estimates of the associations between income level and mortality risk [33].

Based on a qualitative inspection of the smoothed plots of the income-mortality association, and the overall tests of model fit, we have shown that a non-linear association with a stronger association below median income is the most prevalent pattern of association. There were however variations in this association by cause and by gender. For heart disease, in particular among men, there appears to be less of a threshold at median income. While this may be due to mortality risk for heart disease more evenly distributed across the income distribution, an alternative explanation is that we have limited power to detect the shape of the association due to a relatively small number of heart disease events above median income (as confidence intervals indicate). Supporting this later speculation is prior work examining a two slope model of income and cardiovascular mortality that has shown there is a stronger association at lower income levels [10].

While significant associations were observed for both men and women, and the shape of the relation was similar for all-cause, heart disease and injury mortality, there were different associations observed among women and men for cancer above median income – no association for men, and a slightly increasing association among women. Tests of overall model fit showed that while a linear or logarithmic model of income was an equally good fit to the data among men, a non-linear model was a better fit among women. This difference may be due to the positive association between socioeconomic position and rates of breast cancer mortality in women that does not exist as strongly for any site of cancer in men [34, 35]. These differing results by cause and gender are consistent with the context-, time-, and cause-contingent nature of social gradients; empirical findings indicate that income-mortality associations are not fixed, but rather vary across cultures and centuries [36].

Our results are generally consistent with the two other US studies that examined the shape of the income-mortality association [9, 10], despite the fact that these studies were not able to eliminate individuals who had health problems that may have effected their income. While our results not adjusting for labor force partipation are similar to those observed by Blakely et al. in New Zealand, this study finds that after adjusting for labor force participation the association of mortality with income is markedly attenuated, in contrast with our findings [8]. This may be due to our restricted healthy sample, our non-linear method of analysis, or due to stronger effects of labor force participation on mortality in New Zealand as compared to the United States. Our finding of a non-linear shape of association contrast with findings from Finland, where Martikainen et al report a linear shape of association between income and mortality [6]. While this may be due to differing access to benefits and social services in the U.S. and Finland, Blakely et al. note that if an absolute level of income was used to assess the shape of the association, the results of Martikainen et al are consistent with a flattening of risk at upper income levels[8]

The results presented have implications for understanding the etiological links between income and mortality. Based on the observed associations, income disparities in mortality chiefly among lower income populations (below the median income) appear to be driving the commonly reported socioeconomic gradients in all-cause, cancer among men and injury mortality. These findings also underscore why efforts to address income disparities in mortality cannot be restricted simply to persons below the US poverty line but instead should include persons with income at least up to the median level. The difference in size of these two populations is large: in 1991, the mid-point of the income data collection, 13% of families in the U.S. were below the poverty line (equal to $10,860 for a three person family) as compared to 50% of families below the median family income (equal to $35,940), an absolute difference of 37% of US families, and similar to what we have in our study population (Table 1). While anti-poverty policies and programs that include individuals up to either 125%, 150% or 185% of poverty level are common (185% of the poverty level is equivalent to $20,090 for a family of three), the additional population between the 185% poverty level and median income, equivalent to approximately 22% of US families, constitute a large segment of the population that potentially is not benefiting from efforts to reduce health and economic disparities. Suggesting this could be done, benefits from the earned income credit, which specifically address income disparities, reach beyond 185% of the poverty level [37].

In the US context, in adults aged 18–64 at baseline, the non-linear risk of mortality with income arises from the stronger relationship between income level and mortality among lower compared to higher income populations. This evidence is supportive of the hypothesis that policies to improve the health of individuals among the lower half of the income distribution will have the most impact on reducing US income-based disparities in mortality, although from our data we cannot establish why this association exists. Second, if the associations presented are not due to residual confounding, measurement error, or unaccounted for reverse causation, they may also have implications for the importance of income in contributing to premature mortality. Future studies determining the pathways connecting income and mortality will benefit from a consideration of where in the income distribution the burden of disparity exists, and that this association varies by cause.