Estimation of Potential Deaths Averted From Hypothetical US Income Support Policies

Key Points Question How many deaths among working-age US adults can hypothetical income support policies, such as universal basic income, the modified LIFT Act, poverty alleviation, and negative income tax, potentially avert? Findings In this multicohort modeling study that simulated US adults age 18 to 64 years over 5 to 40 years, broad income support policies, like universal basic income, were estimated to avert the most deaths among working-age adults, although targeted approaches, like poverty alleviation, may also avert thousands of deaths among low-income populations. Results were sensitive to several inputs, primarily the income group–specific mortality rates used. Meaning The results of this study suggest that income support policies may prevent thousands of deaths among working-age US adults.


VARIABLES AND INDICES
Variable population number proportion of population in income group mortality rate mortality incident rate ratios by income group (from NLMS) scenario or policy age year all-cause mortality rate (from CDC life table) annual household income Δ change in mortality risk for each unit of income gain (from PSID)
The second study used data from the Panel Study of Income Dynamics (PSID), a longitudinal study of a representative sample of US individuals and their families that started in 1968, to characterize the relationship of total family income and mortality. 9 The authors found that individuals in 1990 (the most recent year in their analysis) who had household incomes below the 32 nd percentile ($33 080) experienced a 54.7% reduction in mortality for every $10 000 increase in household income. At household incomes above this threshold, individuals experienced a more modest reduction (5.3%) in mortality per $10 000-increase in household income, and this finding was not statistically significant.
Estimation of base mortality rates NLMS The study that used data from NLMS reported the mortality incident rate ratios by gender for each household income group (adjusted to 1990 US$). We inflated these household income groups into 2019 US$ using the Consumer Price Index (CPI). 10 In order to calculate age-, gender-, and household income group-specific mortality rates, we first calculated the proportion of people in each household income group by age and gender using data from ASEC. 7 The ASEC reports the breakdown of people in households by total household income in $5 000 increments so we assumed that people were uniformly distributed within each $5 000 household income bracket. eTables 5 and 6 report the proportion of people in each NLMS household income group by age and gender.
Using the proportion of people in each household income group and the reported mortality incident rate ratios for each household income group, we calculated the age-, gender-, and household income group-specific mortality rates. Specifically, for each age and gender , we estimated the mortality rate for each household income group, , by solving the following equality expressed in matrix form (Eq. 1): where indicates household income group, is proportion of people in each household income group, is mortality incident rate ratios by household income group (relative to the seventh household income group, which is the reference group), and is the overall probability of death based on the CDC life table. 11 eTables 7 and 8 show base estimate of age-, gender-, and household income group-specific mortality rates using the NLMS, and Figure 1 in the main manuscript shows how the calculated household income-based, all-cause mortality rates among females differ from the CDC-reported all-cause mortality rates. The mortality rates were estimated using the statistical program R Version 4.0.5 (R Foundation for Statistical Computing, Vienna, Austria).

PSID
To calculate age-, gender-, and household income group-specific mortality rates using the PSID, we first inflated costs from the PSID to 2019 US$ using the CPI 10 ; thus, the minimum unit of household income that is associated with a decline in mortality risk was set at $14 897 (2019 US$) instead of the original $10 000 (2000 US$) reported in the study. 9 We then used the equation (Eq. 2): to estimate the risk of mortality at any household income level ( ), where 0 is baseline mortality, Δ is the change in mortality risk for each unit of household income gain estimated through the PSID, and is a person's annual household income. Baseline mortality ( 0 ), which represents the level of mortality risk that people in each age group experience regardless of household income, was estimated using the following formula (Eq. 3): where is the overall probability of death at a certain age based on 2017 CDC life tables, is the proportion of the population in each household income group below the knot (eTables 5 and 6), is the median household income in each household income group, and is the proportion of the population with household incomes above the knot ($49 122 in 2019 US$), which is the household income threshold where any additional household income has no significant effect on mortality based on the PSID study. (A knot is a point that joins piecewise polynomial functions like linear splines.) Thus, we assumed individuals in the household income group that included the knot experienced mortalities equal to the sum of group mortalities above and below the knot, weighted by the proportion of that group's population above and below the knot. For our base-case analysis, Δ was set to 0.547. eTables 9 and 10 show the base estimates age-, gender-, and household income group-specific mortality rates using the PSID. All analyses were done in Microsoft Excel (Microsoft Corp, Redmond, Washington, USA).

Estimation of mortality rates under various scenarios
We estimated two additional sets of age-, gender-, and household income group-specific mortality rates using both the NLMS and PSID studies, which we refer to as the high-and low-effect scenarios. These scenarios reflect the uncertainty in the estimated relationship between household income and mortality in the NLMS and PSID studies. In the higheffect scenario, we assumed that household income has a bigger effect on mortality (i.e., larger difference in mortality between household income groups), while in the low-effect scenario we assumed that household income has a smaller effect on mortality (i.e., smaller difference in mortality between household income groups). Under the high-effect scenario, individuals with lower household income would have greater mortality than in the base case. In contrast, under the low-effect scenario, individuals with lower household income would have smaller mortality than in the base case. We used the mortality rate estimates under the high-and low-effect scenarios in sensitivity analyses to generate different estimates of the number of deaths averted from each policy we modeled.
Using the NLMS, we calculated the age-, gender-, and household income group-specific mortality rates for the high-and low-effect scenarios by applying the 95% confidence intervals reported for the mortality incident rate ratios ( ) for each household income group to Eq. 1. If household income had a greater effect on mortality (as in the high-effect scenario), we would expect the rate ratios to be further from the reference income group ($26 837-32 466 1990 US$ for males) as the absolute magnitude of the income effect would be greater, and vice versa for the low-effect scenario. For the high-effect scenario, the mortality incident rate ratios ( ) used in Eq. 1 for the lowest six household income groups were set equal to the higher incident mortality rate ratios reported in the 95% confidence intervals, and the mortality incident rate ratios ( ) for the top three household income groups were set equal to the lower incident mortality rate ratios reported in the 95% confidence intervals, signifying the greater impact household income may have on mortality. On the other hand, for the low-effect scenario, the mortality incident rate ratios ( ) used in Eq. 1 for the lowest six household income groups were set equal to the lower incident mortality rate ratios reported in the 95% confidence intervals, and the mortality incident rate ratios ( ) for the top three household income groups were set equal to the higher incident mortality rate ratios reported in the 95% confidence intervals, signifying the smaller impact household income may have on mortality. The mortality rates under high-and low-effect scenarios using the NLMS are show in eTables 11 and 12.
Using the PSID, we calculated the age-, gender-, and household income group-specific mortality rates for the high-and low-effect scenarios by applying the 95% confidence intervals reported for the mortality hazard rate ratios (Δ) in Eq. 2. For the high-effect scenario, Δ was set at 0.615; for the low-effect scenario, Δ was set at 0.468. Similar to the NLMS study, under the high-effect scenario, the resulting age-, gender-, and household income group-specific mortality rates must be magnified in lower income individuals than the base case, but the adjusted mortality rates must be reduced in higher income individuals, however this change is less apparent as it is logarithmic in the PSID study. The income gain from Policy 1 is $12 000 per year per adult. We assumed that there is at least one adult in each household that would receive UBI; as a result, household incomes increase by $12 000 per year. However, this is likely a conservative estimate since more than one qualifying adult may be living in a household, which means some household incomes may increase by at least $24 000 or more per year.
Policy 2: Modified LIFT Act In Policy 2, we model a smaller monthly transfer of $500 per adult with household incomes less than $100 000 per year; this policy is akin to the LIFT (Livable Incomes for Families Today) Act proposed by Vice President Kamala Harris when she was a senator in the 116 th Congress. 19 As a result of Policy 2, eligible adults receive $6 000 annually in income support; in contrast, Vice President Harris's LIFT Act capped the benefit to $3 000 per individual and $6 000 for married couples filing income taxes jointly.

Policy 3: Poverty alleviation
In Policy 3, we simulate a scenario where all US adults are lifted out of poverty. We used the federal poverty level (FPL) for one individual as the threshold for poverty, which in 2019 was set at $12 760 per year. 20 As a result of Policy 3, individuals with annual household incomes below the FPL across all ages and genders move to the next income group where they experience lower all-cause mortality rates. We assume that each eligible individual receives enough government transfer to move them above the FPL.
We do not specify the combination of policies that will lift adults out of poverty, which may include work access programs, minimum wage policies, and housing support. 21 To reduce poverty, experts believe that social welfare programs and programs that encourage employment will have to be implemented concurrently since employment alone will likely be insufficient to eliminate poverty, as evidenced by the 7 million wage-earning adults who live in poverty (so-called "working poor"); additionally, some adults such as the disabled and elderly may not be able to engage in full-time work. 22-24 * We conservatively assumed that 0 is the lowest possible household income in our population, as is reported in the ASEC, though in reality households may have negative incomes, such as when they go into debt or own a business with losses.
The number of subgroups was calculated using the formula (Eq. 4) where and are the upper and lower limits of each household income group (denoted by the index ), respectively, and rounding down the result to the nearest one. For example, in the second household income group ( 2 =$11 677, 2 =$19 915), the number of subgroups is 9 ((19 915-11 677)/1 000 + 1 = 9.238 ≈ 9). The number of subgroups represents the maximum number of equal increments between the lowest and highest household incomes in each band that are at least $1 000. These increments were the basis of the subgroup household incomes shown in eTable 15, which were the household incomes used to estimate the impact of the modeled policies.
The distribution of the population by household income under Policies 1-4 and the status quo is shown in the eFigure, and the specific changes in household income groups following the implementation of Policies 1-4 are detailed in eTable 18. As both the eFigure and eTable 18 illustrate, Policies 1 and 2 lead to the most shifts in household income and income groups, while Policies 3 and 4 lead to minimal changes.

eFigure. Distribution of the US population by household income under different scenarios
This figure shows the distribution of the US working-age population by annual household income under a status quo scenario (curve) and different hypothetical income support policies (bars). LIFT, Livable Incomes for Families Today; NIT, negative income tax; UBI, universal basic income.

eMethods, continued E. Analysis
Our main outcome of interest is deaths averted, which we calculated by taking the difference between the total deaths in the no-intervention scenario and the total deaths from each modeled policy ( ), as summarized in the following equation (Eq. 5):

Sensitivity analysis
Sensitivity analyses allow us to test the influence of our assumptions on the results of our analysis. In this study, we conducted a series of deterministic sensitivity analyses where we varied key assumptions and used different input parameter estimates in the model. First, we varied the distribution of the population by age under the equal and median assumptions discussed previously. Second, we varied the time lag or delay between when a policy is implemented and when individuals experience the mortality rates of their new household income group. Studies on the health effects of income support policies have documented benefits after three years [27][28][29] ; additionally, research on the relationship between income inequality and mortality suggests that the timing of the effect on mortality is not immediate and may vary by population and setting. 30 In the base-case analysis, we assumed that individuals experience the benefit associated with higher household incomes after three years, and in sensitivity analyses, we built in 5-, 10-, and 15-year lags between changes in household income and mortality benefit. With a lag, populations transition to their new household income group but experience the same risk of death as their previous or original household income group until after the lag has lapsed. Third, we used two additional estimates of mortality rates by household income, gender, and age (the high-effect and low-effect scenarios) to generate results for each policy. Fourth, we changed the time horizon to 5, 10, 30, and 40 years, which represent different time horizons that policymakers may use when evaluating income support policies. All-in-all, varying our assumptions in all possible combinations allowed us to generate 180 estimates of deaths averted for each modeled policy and understand the most influential parameters on our results.

F. Limitations
There are several limitations to our study that are worth noting and elaborating. First, we used cross-sectional estimates of the nonlinear association between income gains and mortality, and we assumed that individuals who receive additional income experience reductions in their mortality risk after a lag. While several quasi-experimental and longitudinal cohort studies have shown that higher incomes or increases in income reduce mortality in the US 8,9,27,33,34 and elsewhere 35 , there is still disagreement about the magnitude of the effect. Additionally, because the NLMS and PSID focused on adult mortality, we excluded children in our study. Future analyses should also look at the benefit of increased household income on mortality among children, especially since income-based policies like the EITC and raising the minimum wage have both been associated with reductions in infant mortality. 36,37 Second, we used non-equivalized national household income estimates from ASEC which does not provide additional information on household composition, or the proportion of household income earned from various sources such as wages or government transfers. Thus, in calculating the effect of policies on household income, we had to rely on simplifying assumptions to reasonably estimate the resulting total household income. For example, in modeling the effect of UBI which should benefit every adult aged 18 years and older, we assumed that there was only one adult in each household, leading to a $12 000 annual increase in income. In reality, there may be two or more adults in a household; additionally, household composition may be dynamic over time, which we were not able to model in this study. Similarly, in modeling NIT, we assumed that household income was equal to earned income on which government benefits would be based. Future studies can use income tax return data which provides a more detailed picture of personal and household income, as previous studies have done. 31 Third, we only modeled household income increases that result from income support policies, and we did not model the effect of potential mechanisms such as progressive taxation that may be used to fund these redistributive policies. For example, UBI may require raising taxes on high-income individuals to implement the policy, and this may shift their household income. However, the policies we modeled may also be funded through other means, such as reallocating existing government funding, reforming existing social safety net programs, or government borrowing. We also did not model these policies' effect on jobs, prices, and other economic domains, and a "general equilibrium" analysis that considers the effect of income support policies on the whole system may elucidate these issues.
Fourth, we assumed that individuals' household incomes are constant over time, and that any secular changes (i.e., changes not due to the policy) in income will be captured by the income range or bands that individuals are assigned to. This is a major limitation of the study, though it may have not have had a significant effect on our results because of (1) the wide household income bands we used which captures small changes in income over time and (2) research suggesting that intragenerational economic mobility in the US has remained stable/stagnant since the 1980s. 32 Additionally, while there is upward economic mobility, there is also downward economic mobility that may cancel out some of the benefits of the former; for example, between 1994-2004, 22.8% of people in the bottom (poorest) income quintile were able to move up to the second income quintile in 10 years, 21.5% of people who are in the second income quintile moved down to the bottom income quintile over the same time period. 32 Fifth, we did not model the effect of income inequality on mortality. There is evidence that the magnitude of income inequality-often measured using the Gini coefficient-is positively associated with mortality and other poor health outcomes 30,[38][39][40][41] , and previously published model-based studies exploited this relationship to estimate the health benefits of various income support and redistributive policies. 31,42 However, the literature on income inequality and mortality is mixed, so we opted to exclude it in this analysis. Future studies can explore the effect of both income and income inequality on population health outcomes including mortality.
Finally, we focused exclusively on household income in this study and did not include the effects of other measures of socioeconomic position, such as education, employment, or wealth, which also have documented effects on health outcomes. 42,43 Future studies should look at the independent effects of these social determinants and their intersections on mortality.

eResults eTable 19. Deaths averted from Policy 1 (Universal basic income) under various assumptions (in thousands)
Time horizon

PSID mortality rates NLMS mortality rates
Equal age assumption Median age assumption Equal age assumption Median age assumption