How do policy outcomes – and federal spending – vary across the states?

Research on state-to-state differences in public policy outcomes takes three broad forms: work that emphasises features of states, work that emphasises the actions of national elected official and bureaucrats and work that reveals how state actors can influence the flow of resources from the federal government. The state-centred work analyses differences related to state-level public policy choices directly under the control of state legislatures and governors. State-level variation has been linked to diverging paths of historical development, diverse political cultures, the presence of different racial and ethnic majorities, the ideologies of political leaders and citizens, different types of dominant economics interests and elites, and the economic health of the state (Elazar Reference Engelhardt and Gruber1966; Jacoby and Schneider Reference Jacoby and Schneider2001; Amberg Reference Amberg2008; Lieske Reference Lieske2010). The variety of policies that this research encompasses is quite broad: from income support programmes and Medicaid to child support enforcement actions (for the latter, see Keiser and Soss Reference Keiser and Soss1998). States also react in important ways to the policy choices of neighbouring states (Peterson and Rom Reference Peterson and Rom1990). The DC-centred work analyses how federal choices – in the White House, Congress or federal administrative agencies – may concentrate spending in particular states. Berry et al. (Reference Berry, Burden and Howell2010) examine the geographic distribution of federal spending (direct expenditures and grants) focussing on how spending varies across congressional districts and counties. They find that districts with a more senior representative from the president’s party tend to see higher federal expenditures of this form. In work designed to determine how much federal spending in various forms translates into economic activity, Nakamura and Steinsson (Reference Nakamura and Steinsson2014) demonstrate that federal spending – specifically, military procurement – can vary and have quite different impacts across states and regions. A third set of work investigates how state-level actions impact the level of federally administered programme spending (Rich Reference Rich1989; Nugent Reference Nugent2009). States can, for example, invest resources to assist eligible residents in applying for a variety of benefits that are fully or partially funded by the federal government – supplemental nutrition assistance, unemployment insurance, Medicaid and even veteran’s disability benefits (Miller and Kaiser Reference Miller and Kaiser2012). States can also vary in the way administrative determinations make it easier or harder to qualify for federal assistance: the SSA has identified substantial state-to-state variation in the rejection of initial disability claims (Strand Reference Strand2002) and the Department of Agriculture has documented state variation in Temporary Assistance for Needy Families payments – which impacts eligibility for, and amount of, supplemental nutrition assistance (Hanson and Andrews Reference Hanson and Andrews2009).

The aggregate effect of these federal and state choices generates quite remarkable differences in the “return” on tax dollars across states. The ratio of fiscal year 2013 federal spending to gross revenue collections for the entire United States was about $1.10 (Internal Revenue Service 2014; Pew Reference Pew2014). The expenditure total excludes interest payments on federal debt, classified expenditures and international transactions, but remains above $1.00 as some of the spending is financed by borrowing. For some states, this ratio is over $3.00 (Mississippi, New Mexico and West Virginia), whereas other states are large net contributors to the fiscal union, with ratios below $0.75 (Minnesota, Delaware, New Jersey and Nebraska). This disparity – and the fact that Red States are more often net beneficiaries – has attracted attention from journalists and other observers (e.g. see Gilson Reference Gilson2012; Tierney Reference Tierney2014).

The federal OA programme may be unique in that none of the literature on state policy variation applies. Eligibility and benefit levels are strictly a function of federal rules – there is no state administrative discretion to alter benefit levels. There is also no federal agency discretion that could result in higher benefits in particular states. State or federal administrative actors, members of Congress, state governors or legislators cannot influence the average benefit for workers only in particular states. Early DC decisions can, and did, impact the scope of coverage and eligibility in ways that disadvantaged some states, particularly Southern states, but today nearly 95% of US workers are covered by the OA programme. The OA programme is interesting precisely as it operates outside the influence of state-level actors but accounts for a remarkably large proportion of the overall flow of federal expenditures to the states. It may be possible for states to attempt to lure seniors – and their Social Security checks – by offering low or no income taxes on benefits, but there are few other steps that state leaders could take to influence the level of benefits that flow to their state. The one exception is whether or not participants in the states’ public sector retirement plan are also covered by Social Security – in seven states all public sector workers participate in Social Security, and in 33 states no public sector workers participate. Federal legislative choices are also unlikely to result in cross-state differences. Members of Congress can change the way that the benefit levels are determined or adjusted upwards each year, but those choices apply to all recipients in the same way, regardless of place of residence.

Despite the absence of direct opportunities to influence the flow of resources to particular states, the state-to-state variation in OA programme outcomes remains important for two reasons. First, the sheer size of the programme suggests that even small state-to-state variation may be important. The combined benefit expenditures of the Old-Age and Survivors Insurance components of Social Security approached $700 billion in fiscal year 2014, nearly 20% of the entire federal budget. With the exception of Medicaid, the OA programme is much larger than the sum of all the various grant and transfer payments addressed in the literature on state-to-state variation in federal spending. Second, the literature on state variation highlights differences in the economy of the states – states have different dominant industries, levels of income and unemployment levels. These differently situated states could experience widely divergent policy outcomes even under a standardised approach to the calculation of benefits and eligibility. The link between particular state features and OA expenditure across states is poorly understood, at best.

What is the OA programme and how does it impact individual states?

In 2014, nearly 42 million Americans received Social Security OA benefits on a monthly basis – 39 million retirees, two million spouses and almost one million children (SSA 2015). As perhaps the most effective anti-poverty programme in the United States, the OA component lifted more than 14 million seniors out of poverty in 2011 (Van de Water and Sherman Reference Van de Water and Sherman2012). In fact, OA payments are largest source of retirement income for persons aged 65 and older [Government Accountability Office (GAO) 2012].

The OA programme is financed through employee and employer payroll taxes and provides cash retirement benefits to workers upon retirement at a designated age. Workers are eligible for OA benefits if they have a work history that includes Federal Insurance Contributions Act (FICA) withholdings over a period of roughly 10 years. OA benefits are then determined by a formula that includes 35 years of earnings history. The SSA computes an average monthly indexed wage – indexed to account for inflation – and calculates benefit based on that wage. Both benefits and FICA withholdings are capped. This combination of uniform eligibility requirements and formula-based benefits means that similarly situated workers will receive identical benefits independent of where they live. There is no adjustment of OA benefits based on place of residence. In addition, there is no link between OA benefits and receipt of other benefits or eligibility for other services. OA benefits are the same if a recipient lives at home or is a resident of a skilled care nursing facility. In addition, OA benefit levels do not vary depending on eligibility for, or participation in, Medicare or Medicaid. The most important exceptions to this uniform treatment of earnings are the Windfall Elimination Provision and the Government Pension Offset: OA benefits levels (for worker or survivor) are reduced if an individual worked for an employer who did not withhold Social Security taxes. Social Security benefits may also be reduced if individuals file for benefits before full retirement age and may be further reduced if benefits are claimed early and the recipients continue to earn income from work.

If individual benefits are strictly a function of earnings and there are no discretionary choices about eligibility or benefit levels, then the primary factor that should explain variation in OA expenditures is the proportion of the state population aged 65 or over. Figure 1 summarises the link between senior population and the level of benefits flowing to each of the 50 states – the relationship is obviously positive but also very noisy.

Figure 1

Total real Old-Age (OA) benefits per capita and state senior population, 1999–2013.

Different metrics tell a similar story. One convenient way to describe the impact of OA expenditures in individual states is to look at the ratio of OA expenditures to state gross domestic product (GDP). This metric is summarised, nationally and for Michigan and Louisiana, in Figure 2.

Figure 2

Old-Age (OA) programme expenditures as a percentage of state gross domestic product (GDP), 1999–2013.

As suggested in the introduction, the OA programme has a much larger impact in Michigan than in the Louisiana – and this large difference is quite persistent over time. One obvious candidate reason for this difference is the size of the senior population in each state and, of course, this matters. Similar to Figure 1, Figure 3 summarises OA expenditures as a percentage of state GDP and the proportion of the state population over age 65 for each state from 1999 to 2013.

Figure 3

Old-Age (OA) expenditures as percentage of state gross domestic product (GDP) and state senior population, 1999–2013.

The figure demonstrates the link between the senior population and the economic importance of OA expenditures, but also highlights the tremendous variation in the impact of OA expenditures on the state economy in states with very similar senior populations – states with 12–14% of the population aged over 65 are observed with impacts from 2 to 4% of state GDP. Why is this important? As part of efforts to understand the impact of federal spending on the economy, a study by Romer and Romer (Reference Romer and Romer2014) is one of a handful of attempts to assess the macroeconomic impact of Social Security benefits. Their focus is on the aggregate impact of legislated changes in benefit levels (such as the Tax Reform Act of 1969) on consumption. They find a large effect of benefit increases – a one-to-one increase in consumer spending (a permanent 1% increase in benefits translates into an immediate 1% increase in consumption in the same month, an effect that persists for at least five months). The implication is clear: Social Security transfer payments translate into direct and immediate economic benefits. Consistent with this implication, state-level work on the impact of Social Security on poverty among the elderly indicates that the programme reduces the elderly poverty rate from 48 to 7% in Michigan, but from 50 to 15% in Louisiana. Despite the uniform features of the programme, the elderly poverty rate remains twice as high in Louisiana as in Michigan (Van de Water and Sherman Reference Van de Water and Sherman2012).

Why do outcomes vary across states?

Across states that have similar over-65 populations, there are two factors specific to the OA programme that drive expenditures and account for most of the differences across states: variation in coverage (the proportion of the over-65 population receiving OA benefits) and variation in the average benefits (total OA payments divided by the number receiving benefits). Average benefits and coverage are summarised for the sample period in Figures 4 and 5.

Figure 4

Percentage of population aged 65 and over receiving Old-Age (OA) benefits.

Figure 5

Average annual Old-Age (OA) benefit.

Michigan has a high and rapidly growing average benefit and a consistently high coverage relative to Louisiana – those persistent differences are the source of the discrepancy identified in the introduction.

But what explains these differences and why do they persist? We suggest two types of answers: state senior populations are not homogeneous and state economic performance varies.

A statistical model of state-level OA expenditures

To determine the impact of state demographic and economic features on OA coverage and benefits, we pool data over a 14-year period – 2000–2013 – from the 50 states. We combine census data on state demographics (total population, poverty levels,Footnote ¹ education levels, minority population and senior population) with data from the SSA on expenditures and beneficiaries in each state, as well as data on state public pension plans. We also merge data on per capita income from the Bureau Economic Analysis and data on unemployment rates from the Bureau of Labor Statistics.Footnote ² The data sources and descriptive statistics for each variable are provided in Appendix 1. The technical objective of the model is to understand how the coverage of the OA programme and the average benefit in each state are associated with, or affected by, economic and demographic characteristics of the states. Although we assume that the states share a common adjustment process – the parameters of the model are common across states – the fact that we measure each variable annually at the state level means that the modelling approach accommodates the fact that unemployment and income levels may rise or fall at different times in different states.

We use a panel data estimator to determine the link between OA expenditures and coverage and state demographic and economic characteristics. This modelling strategy requires a number of methodological choices. With one exception, we assume that all effects are contemporaneous. Changes in unemployment or poverty this year, for instance, will impact average benefits this year as new filers enter the pipeline. As the ultimate focus of the article is the role of unemployment, we chose to keep the model as straightforward as possible and to introduce only one lagged independent variable: real per capita income. As OA benefits are calculated based on a 35-year earnings history, states that have had high-wage jobs in the past are likely to have retiring workers who have high OA benefits. We introduce a 15-year lag of real per capita income to capture earnings experience near the midpoint of a working career for a worker nearing full retirement age. (The results are the same – almost identical –if we use any lag longer than 12 years and we evaluated models with lags up to 25 years.)

We transform the data in two ways before estimating the coefficients. Any data that are in dollars (expenditures, income) are adjusted to real 2013 dollars using the Consumer Price Index. Second, with the exception of coverage data on state and local government workers, all of the variables introduced in the model are the difference between the current year value and the prior year value (the first difference). Using the variables without taking this step is problematic as these series are not stationary – the mean of the series changes over time. A fairly broad literature reveals that using variables that are not stationary can – or will – generate spurious results (see e.g. Phillips Reference Phillips1986). We test each variable to ensure that the first difference is stationary using conventional tests implemented in most statistical software (see Im et al. Reference Im, Pesaran and Shin2003). Using the first difference also mitigates problems of collinearity. Unemployment rates, poverty rates and real income are closely correlated at the state level, but the first differences in these series are not highly correlated. Canonical tests for collinearity (the variance inflation factor) revealed no problems.

To estimate the models of OA coverage and benefit levels, we use a generalised least squares (GLS) panel estimator with an autoregressive order 1 heteroskedastic error structure (see Wooldridge Reference Wooldridge2002, for an explication of this First Difference approach). The coefficients describe how benefit levels or coverage change as demographic and economic characteristics change in a particular state. Therefore, as unemployment rises or falls within a state, what happens to coverage and average benefits? The key results were quite robust to the choice of estimator: whether we used the first difference estimator with simple ordinary least squares, fixed effects or the GLS estimates reported in the tables.Footnote ³ The estimates and standard errors for models of both programme features – coverage and average benefits – are summarised in Table 1.

Table 1

Old-Age (OA) programme coverage and expenditures by state

Note: n=650 (50 states×13 years). Standard errors in parentheses.

Feasible generalised least squares panel estimator with heteroskedastic panels and a common AR(1) error term. Estimates from STATA 13.0.

**p<0.05, *p<0.10.

Average benefit level

Which states have high benefit levels? The estimates suggest that five of the variables included in the model are important, but the impact of unemployment is particularly large. States with historically high growth in real per capita income are associated with higher average benefit levels, as expected. States with rising unemployment levels experience substantial growth in average benefits. The effect of unemployment swamps the other factors in the model: a 1 SD change in the first difference of unemployment (~1%) is associated with a more than $100 increase in annual average benefit levels (about a 0.75% increase in the average benefit). This finding reveals that the earnings history of new filers supports a higher benefit than the earnings history of earlier cohorts, even with any penalty related to claiming before full retirement age. The data are simply unambiguous about this result. Figure 6 plots the average difference in average benefits each year and the average difference in unemployment rates each year. The average benefit level surges upwards in periods of high unemployment and moves up at a slower rate when unemployment is low.

Figure 6

Unemployment and average real old-age benefit levels, first difference, 2000–2013

Using data for individual states can reveal how this process unfolds in different ways in different states. Figures 7 and 8 summarise the path of first difference in unemployment level and average benefits in Louisiana and Michigan. Note the different paths of each indicator in 2006. In Louisiana, unemployment falls and average real benefit falls by $18 (more accurately, changes in nominal benefits fail to keep pace with inflation). In Michigan, unemployment increases by a small amount and the average real benefit increases by over $200.

Figure 7

Unemployment and average real old-age benefit levels, first difference, 2000–2013. Note: State of Louisiana.

Figure 8

Unemployment and average real old-age benefit levels, first difference, 2000–2013. Note: State of Michigan.

To test the robustness of this result outside of the sample period, we collected data on average national unemployment and average national benefit increases from 1983 to 2014. The observed relationship between the first difference in real benefits levels and the first difference in unemployment was very close to the results in Table 1: an increase of just over $100 in average annual benefits for each 1% increase in unemployment. Finally, we also tested to see whether it was simply the huge spike in unemployment in 2009 that accounted for the impact of unemployment. We found that the impact of unemployment is positive and statistically significant if the sample is restricted to the period ending in 2007.

Other variables in the model also matter, but the impact is much lower. States with growing minority populations have benefits that grow faster than other states – a 1 SD increase in the first difference of the state proportion minority implies a $30 increase in average annual benefits. States with an increase in college completion will also see lower average benefits, inconsistent with our expectations. This is perhaps the most puzzling result we uncover – in places and at times that overall educational attainment falls, OA average benefits rise. Part of this is due to one year – 2011 – where Social Security average benefits decline in nearly all states, whereas the college completion measure increases by a large amount in almost all states. Overall, states with poorly educated and minority residents experiencing growing levels of unemployment will see average benefits climb; the new and large cohort of workers who elect to begin receiving Social Security OA payments qualifies for sometimes much higher benefits than current recipients. The growth rate of the senior population has no link to the level of average benefits.

Overall, the combination of insights from the two models reveals what explains state variation. In most cases, if a variable is positively associated with average benefits, it is negatively associated with coverage and vice versa. This is the case with real per capita income, minority population, poverty, senior population and college completion. When a high proportion of the state is covered, the average benefit is lower. When a smaller proportion of the state is covered, the average benefit is higher. The exceptions to this are unemployment and public sector worker participation. States where public sector employees are covered have both broader coverage and higher average benefits. Unemployment is positively associated with both coverage and benefit levels: states with rising unemployment have broader coverage and higher benefit levels and the effect of unemployment in both models is large.