Policy Research Working Paper 9244 Cross-Region Transfers in a Monetary Union Evidence from the US and Some Implications Steven Pennings Development Economics Development Research Group May 2020 Policy Research Working Paper 9244 Abstract US federal transfers to individuals are large, countercyclical, multiplier of around 1.5 for permanent transfers and 1/3 vary geographically, and are often credited for helping stabi- for temporary transfers. Results are broadly consistent with lize regional economies. This paper estimates the short-run an open-economy New Keynesian model. At business-cycle effects of these transfers using plausibly exogenous regional frequencies, cross-region transfer multipliers are not large, variation in temporary stimulus packages and earlier per- suggesting only modest gains in regional stabilization from manent Social Security increases. States that received larger US federal automatic stabilizers. transfers tended to grow faster contemporaneously, with a This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The author may be contacted at spennings@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team CROSS-REGION TRANSFERS IN A MONETARY UNION: EVIDENCE FROM THE US AND SOME IMPLICATIONS STEVEN PENNINGS* Date : 11 May 2020 (Updated December 2020). JEL: E62, F45, F41; Keywords: Fiscal Multiplier, Fiscal Transfer, Monetary Union. *Development Research Group, The World Bank. Email: spennings@worldbank.org or steven.pennings@nyu.edu; Address: World Bank, 1818 H St NW, Washington DC 20433. The views expressed here are the author's and do not necessarily reect those of the World Bank, its executive directors, or the countries they represent. For helpful comments I thank several anonymous referees, Emi Nakamura, Virgiliu Midrigan, Mark Gertler, Bill Easterly, Aart Kraay, Emmanuel Farhi, Glenn Follette, Hyun Oh, Pierre Bachas, Jenny Guardado, Arthur Mendes, Roberto Fattal, Thuy Lan Nguyen, Antonn Park and seminar participants at NYU, the World Bank, Federal Reserve Board, the NY Fed, the Bank of England, Midwest Macro, UNSW, Monash, ANU, Georgetown U, CIDE, JHU-SAIS, CESifo Venice Summer Institute, and the ECB. An online appendix is available at: https://drive.google.com/le/d/1-Dldgp2KY5IyxsxGxQXZHwFx5ddmZXl6/view. 1. Introduction US federal expenditure increasingly involves making transfers to individuals rather than purchasing (or producing) output. In 2017, US federal transfers to individuals were around $1.4$2.1 trillion (depending on the denition used), which is larger than government consumption. Transfers to individuals were the cornerstone of countercyclical scal stimulus in 2001 and 2008 and ac- counted for the majority of the increase in expenditure around the 20072009 nancial crisis in the US and other countries (Oh and Reis 2012). More re- cently targeted transfers to households are part of the economic response to COVID-19 in more than two dozen countries, with the US federal government spending $250 billion on one-o payments (IMF 2020). In developing coun- tries, transfers are increasingly popular antipoverty programs, with around 400 million people receiving regular cash payments (Filmer et al. 2019). Federal transfer policies naturally redistribute income across regions, with benets often tilted to areas with low incomes or to regions receiving negative shocks. Automatic stabilizers mean individuals in regions entering recessions receive more federal benets and pay lower federal taxes. A number of papers have found that these (and other) net federal transfers are about $0.20$0.40 of every dollar fall in regional income in the US, which are thought to help stabilize regional economies (Feyrer and Sacerdote (2013), Sala-i-Martin and Sachs (1991), and Bayoumi and Masson (1995)). As such, The Economist writes that America's scal union is so good at absorbing [regional] shocks 1 that it is often cited as a model for the more accident-prone euro zone. Despite their importance, little is known about the eect of transfers to individuals on short-run regional economic growth, which I call the size of the cross-region transfer multiplier. My main contribution is to ll this gap using several natural experiments in which permanent Social Security benet 1See The Economist For richer, for poorer (28 November 2015). This perception has been around for more than a quarter-century: Sala-i-Martin and Sachs (1991, p. 20) write that Some economists...argue that [the] regional insurance scheme provided by the federal government is one of the key reasons why the system of xed exchange rates within the United States has survived without major problems. Other benets of scal unions in addition to automatic stabilizers, such as centralized deposit insurance and discretionary stimulus, are beyond the scope of this paper. See Appendix 2 for background on the size and countercyclicality of transfers. 1 increases and temporary stimulus payments allocate transfers across US states in a way that is plausibly unrelated to regional business cycles. My ndings are related to literatures on marginal propensity to consume (MPC) transfers (e.g. Parker et al. (2013); Johnson, Parker, and Souleles (2006) (henceforth JPS); Romer and Romer (2014); Hausman (2016)) and the cross-sectional purchase multiplier (e.g. Nakamura and Steinsson (2014) (henceforth NS); Clemens and Miron (2012); Chodorow-Reich et al. (2012)). But the cross-region transfer multiplier is conceptually very dierent from both the MPC and cross-sectional purchase multiplier in New Keynesian (NK) and Neoclassical (NC) models. 2 Showing this analytically is my second contribution. The cross-region transfer multiplier diers from the MPC due to local gen- eral equilibrium (GE) eects. Local GE eects amplify the MPC in NK mod- els with sticky prices/wages if incomes are mostly spent on locally produced goods, but they can dampen the MPC in more open economies. In NC models, cross-region transfer multipliers are always negative due to wealth eects on labor supply in GE. The cross-region transfer multiplier diers from the cross-region purchase multiplier due to a smaller boost to local demand from transfersdepending on the MPC and opennessand from stronger wealth eects. These mean the cross-region transfer multiplier is smaller than the purchase multiplier in both NK and NC models and is more variable in sign and absolute size. The permanent income hypothesis suggests transfer multipliers much more sensitive to the persistence of scal shocks, which I explore empirically by studying both temporary and permanent transfer policies. The wide variety of theoretically possible cross-region transfer multiplierslarge, small, and negativeunderscores the need for empirical evidence to discipline the choice of models and parameters. Providing this guidance is my third contribution. There are two key challenges to identifying the cross-region transfer mul- tiplier empirically. The rst challenge is reverse causality, where changes in state-level growth can induce changes in countercyclical transfer policies to 2Most of the recent theoretical literature has focused on determinants of the MPC (Kaplan and Violante 2014) or the (large) dierences between cross-region and aggregate purchase multipliers (Farhi and Werning (2016) and NS). Farhi and Werning (2017) focus on optimal transfers rather than the size of multipliers. 2 residents of the aected states. To address this problem, my identication strategy combines a transfer policy change at the aggregate level that af- fects all statesand so is unlikely to be driven by developments in a par- ticular statewith a predetermined and slow-moving cross-state allocation of the transfer based on regulations that determine eligibility and their relative importance in dierent states. This is a similar idea to a Bartik (1991) instru- ment, though is constructed separately for each aggregate policy change. For permanent transfers, the aggregate policy change is a series of ad hoc increases in the monthly stipend of Social Security (old age pension) recipients over 19521974. An increase in the monthly Social Security stipend naturally leads to a larger increase in transfers to states like Florida, relative to Alaska (which is less popular with retirees). Romer and Romer (2016) provide narra- tive evidence that this sample of Social Security increases was legislated mostly to compensate for past ination and never out of any desire to stimulate the economy. As such, these Social Security increases are likely even exogenous at the aggregate level, though I only require the weaker claim that they are exogenous in the cross-section. For temporary transfers, the aggregate policy changes are one-o stimulus payments in 2001 and 2008 (a check in the mail) of $300$600. While the relatively xed dollar value of payments means that these transfers are me- chanically more important in poorer states, other eligibility criteriasuch as having a tax liability in 2001reduce the progressivity of the transfer and mean that the cross-state allocation varies across policies. It is dicult to argue that these policies were enacted to help specic states in particularly deep recessions given that the benets were widely spread across states and eligibility rules were a simple function of prior-year individual taxable income. The second identication challenge is omitted variable bias, where other variables might aect both transfers and regional economic growth. I address this by controlling for state xed eects (FEs) to remove state-level trends and time FEs to remove all aggregate variation (e.g., the US business cycle, monetary policy, or international shocks), leaving only potential confounding factors that vary both across states and over time. The remaining omitted variable bias is reduced by (i) a research design that involves many policy 3 changes (reducing the chance of a coincidental correlation) and (ii) a battery of robustness tests, such as other time-and-state-varying controls, dropping states and quarters one at a time, and a number of placebo tests. I focus on the contemporaneous eect (impact multiplier) of cross-state transfers on quarterly state GDP, or non-transfer labor income (W × L), which is available over a longer sample. But these impact multipliers also prove to be similar to cumulative multipliers over longer horizons. My key empirical nding is that states receiving larger transfers tended to have faster short-run growth in non-transfer labor income or GDP, with a much larger cross-region transfer multiplier for permanent than for temporary 3 transfers. Specically, I nd that a state that received an extra $1 tempo- rary transfer experienced an increase in labor income or GDP in the quarter by around $1/3 ($0.2$0.9, depending on the specication). For permanent trans- fers, I nd that states that received an extra $1 in Social Security payments increased their labor income by around $1.5 contemporaneously ($0.9$1.9, depending on the specication). Romer and Romer (2016) nd that almost all of the permanent Social Security increases were consumed and JPS and Parker et al. (2013) argue that a modest fraction of the temporary transfers were spent contemporaneously on non-durablesboth of which are broadly consistent with my results (though those papers do not investigate GE eects across regions or look at eects on GDP or non-transfer income). These empirical ndings have implications for the ability of the US federal scal system to smooth shocks to regional economic growth through auto- matic stabilizers, as that ability depends, in part, on the size of the cross- region transfer multiplier. Regional business cycles are temporary, and so the relevant multipliers are closer to those on temporary transfers and hence are modest in size. Combined with the countercyclicality of federal net transfers to the residents of US states estimated in the literature, back-of-the-envelope 3Note that other dierences across policies, beyond persistence, might be relevant. Most important, the one-o transfers were in the 2000s when regional economies were more open than when Social Security payments were increased in the 1950s1970s, which would reduce multipliers in a NK model. Households in the 1950s1970s likely also had lower access to credit, but this is less relevant for permanent transfers that are spent even by Ricardian households. The payment size, age of beneciaries, and behavioral framing also dier across policies. 4 calculations suggest that the US federal system would only smooth about 10% of regional asymmetric shocks to output, which is perhaps less than in the popular perception. My empirical estimates are also related to the size of the aggregate closed- economy transfer multiplier, albeit indirectly as an identied moment that guides the choice of theoretical models (Nakamura and Steinsson 2018). Ag- gregate scal multipliers are sensitive to monetary policy and tax responses and so are very dierent, in general, to cross-sectional multipliers estimates here, which dierence out these factors. My cross-sectional multiplier esti- mates are broadly consistent with a standard open-economy NK model that features home bias in consumption and a share of hand-to-mouth households, and they are less consistent with a canonical NC model. That NK model produces aggregate purchase multipliers similar to those in NS: large (≥1) if monetary policy is accommodating and small (<1) otherwise. Closed-economy transfer multipliers at business-cycle frequencies in this model are similar to purchase multipliers if transfers are targeted at hand-to-mouth households but only a third of the size if transfers are untargeted. The rest of this paper is organized as follows. Section 2 outlines the empirical methodology, including describing the natural experiments, data construction, identication issues and specication. Section 3 provides the main empirical results, robustness tests and some extensions. Section 4 provides analytical expressions for cross-region transfer multipliers in NK and NC models, and compares them to the MPC and cross-region purchase multipliers. Section 5 compares theoretical and empirical multiplier estimates quantitatively. Section 6 presents policy implications on the ability of the US federal scal system to smooth regional shocks and the size of the aggregate closed economy transfer multiplier. Section 7 concludes. 2. Empirical Methodology 2.1. Transfer Policies and Variable Construction. In this paper, I study the eect of the 19521974 permanent Social Security benet increases and 2001 and 2008 one-o stimulus payments on state-level income growth. This 5 subsection describes these policies and how the transfer variables are con- structed. 2.1.1. Social Security Increases (19521974). Before 1975, Social Security pay- ments (largely the old-age pension) were not indexed to ination and were increased by Congress on an ad hoc basis (Wilcox 1989). These were mostly permanent increases in transfersi.e., a higher monthly stipend received by the elderly and their dependentsand so are more likely to be spent by Ricar- dian consumers (by the permanent income hypothesis). As transfer increases varied in size and timing, they would not be subsumed into seasonal factors. Wilcox (1989) and Romer and Romer (2016) study these transfers at the ag- gregate level and nd that increases in permanent Social Security payments signicantly increase consumption. This is consistent with my nding of large income multipliers in the cross-section, though their results do not imply mine and vice versa. From 1975 onward, Social Security payments were indexed to the CPI and adjusted annually, making them much more predictable, and so 4 they are excluded from my analysis. 2.1.2. One-o Transfer Stimulus Payments (2001 and 2008). I consider two temporary stimulus payments, the Economic Growth and Tax Relief Reconcil- iation Act (EGTRRA) stimulus payments in 2001 and the Economic Stimulus Act payments in 2008. My main results pool across these two temporary payments, though the payments are analyzed separately in Section 3.5. In 2001Q3, the Bush Administration transferred $38 billion to households in a one-o stimulus payment as part of the broader Economic Growth and Tax Relief Reconciliation Act (EGTRRA). Individuals paying net taxes mostly re- ceived $300 per capita, though unlike the 2008 transfers, there was no payment for those with no tax liability and no phase out for those on high incomes. Ex- ploiting the randomization of payment dates, JPS nd that 20%40% of the payment was spent on non-durables in the months that it was received, with a higher MPC for the poor and credit constrained, and no response for durables. 4Romer and Romer (2016) and Wilcox (1989) nd smaller or insignicant, respectively, responses of consumption from 1975 onward. An earlier version of this paper used Wilcox's (1989) shorter sample of pre-1975 Social Security increases, which produced broadly similar results. Romer and Romer's sample only covers benet increases for existing Social Security recipients and so excludes expansions in eligibility and other rule changes. 6 In 2008Q2Q3, around $95 billion was transferred to households as one-time payments as part of the Bush Administration's Economic Stimulus Act. The vast majority (85%) of the payments were made during 2008Q2 (Parker et al. 2013). These economic stimulus payments (ESPs) had two main components: (i) $300 per-capita payments made to those paying no net taxes but with at least $3,000 in eligible annual income (around $30 billion, which I call the low-income rebate component, which was refundable) and (ii) $600 per-capita payments made to those paying net taxes with a phase-out for those earning over $75,000 (around $65 billion, which I call the middle-income tax refund component). Most of the eects of the 2008 ESP turn out to be driven by the low-income component. Parker et al. (2013) exploit randomization in the timing of the ESPs and nd that about 12%30% of the payments were spent on non-durable consumption in the months they were received (50%90% including durables). 2.1.3. Construction of Exogenous Transfer Variables. The changes in exoge- nous transfers at the state level used in regressions (∆tri,t ) are only pub- licly available for the 2008 low-income rebate component (see BEA 2009), but otherwise have to be calculated based on the size of the aggregate trans- fer change ∆trU S,t (across the whole US) multiplied by a state-specic share 5 statesharei,t−j : (1) ∆tri,t = ∆trU S,t × statesharei,t−4 where t denotes quarters, i denotes states, and ∆ is the rst dierence operator. On the surface, this appears similar to a Bartik (1991) instrument, but diers because the aggregate component ∆trU S,t is based on individual policy changes 5The cross-state allocation of the low-income rebate component from BEA (2009) is pri- marily based on the geographic distribution of recipients of refundable Earned Income Tax Credits (EITCs). According to the BEA (2009), $28 billion of the low-income component was paid in 2008Q2, with only $1.35 billion paid in 2008Q3 (the annualized gures in BEA (2009) divided by 4). Combined with Parker et al.'s (2013) quarterly prole, this suggests a middle-income payout of around $50 billion in 2008Q2 and $14 billion in 2008Q3. All of the 2001 stimulus payments were made in 2001Q3. Hence ∆tri,t > 0 for 2001Q3 and 2008Q2, ∆tri,t < 0 for 2001Q4, 2008Q3Q4 (small in 2008Q4), and ∆tri,t = 0 for all other quarters. The rest of the sample helps to estimate state FEs and controls. 7 (rather than general aggregate variation), and statesharei,t−4 is based on the eligibility rules for each specic policy change (rather than average transfer shares). Combined, this means that ∆tri,t reects actual changes in transfers to dierent states very accurately  for example, in an IV specication rst- stage coecients are close to one and F-stats are above 500 rather than being an instrument correlated with transfer changes. For the 2001 transfers and the 2008 middle-income tax refund the aggregate size of the payments ∆trU S,t in Equation 1 are taken from JPS and Parker et al. (2013) respectively (subtracting the size of the low income rebate from BEA (2009) in the latter case). Each state's share (statesharei,t−4 ) is calcu- lated using IRS state-level data on individual tax returns for the previous tax year (2000 and 2007 respectively), using eligibility rules for each specic pay- ment. Eligibility for the 2001 and 2008 stimulus payments largely depended on income in 2000 and 2007, as full-year income was not known in 2001Q3 or 2008Q2 when the rst payments were made. The cross-state distribution of one-o transfers is plotted below in Figure 1. For permanent Social Security benet increases, the aggregate transfer change 6 ∆trU S,t is taken from Table 1 of Romer and Romer (2016) over 19521974. I allocate the exogenous increase in aggregate Social Security payments across states in proportion to that state's share of Social Security payments a year before, statesharei,t−4 as in Equation 1 (see Appendix 1 for details). As the policy change essentially involves scaling up existing Social Security trans- fers, Equation 1 is exact using a contemporaneous stateshare. In practice statesharei is almost unchanged year-to-year, so Equation 1 is likely to be 7 very accurate (see Appendix Figures 1.1 and 3.1B). 6As Romer and Romer's data are monthly, whereas mine are quarterly, I spread the ad- justments over two quarters if the permanent increase in payments occurred mid-quarter. For example, a permanent $1 increase on a $10 monthly payment starting on June 1 will become a 3.3% increase in Q2 and a ≈6.6% increase in Q3. 7For example, in the case of an increase in Social Security benets in the rst month of the SS quarter,∆trU S,t is the dollar value of Romer and Romer's permanent transfer change, and SS 50 SS statesharei,t−4 = tri,t −4 / i=1 trt−4 is that state's share of total Social Security transfers four quarters before. 8 There are 20 exogenous Social Security increases during 19521974, which are spread over 27 quarters. There are no increases in the remaining 65 quar- ters (∆trit = 0), but these periods are included in the sample to estimate state FEs µi and other controls. The average size of the increases is 0.2% of quarterly labor income, though they are highly heterogeneous over states and time (see descriptive statistics in Appendix Table 1). The largest permanent increases were in 1972Q4 (a 20% increase in benets) of around 1.45% of quar- terly labor income in West Virginia, Arkansas, and Florida, which have a large number of retirees as a share of the population (Appendix Figure 1.1 left side). In contrast, the increase in transfers to residents of Alaska was only 0.2% of labor income in the same quarter. 2.2. Identication Strategy. As mentioned in the introduction, there are two key challenges in identifying the eect of transfers on cross-sectional growth: reverse causality and omitted variable bias. Reverse causality is a serious problem for transfer multiplier estimates be- cause transfers are countercyclical. A simple regression of growth on transfers will pick up a combination of the transfer multiplier and the countercyclical tax-transfer system, leading to downward-biased multiplier estimates (too neg- ative), with coecients typically insignicant or negative (see Appendix Table 10). My identication strategy addresses reverse causality by studying variation in transfers across US states (as a share of income) based on an aggregate transfer change (∆trU S,t ) combined with rules determining who is eligible for the transfer (statesharei,t−4 ). For example, an increase in the monthly stipend of Social Security recipients generates a larger increase in transfers to states 8 with many retirees, such as Florida. As the transfer policy change (∆trU S,t ) is at the national level (aecting all states), it is unlikely to be implemented to address a recession or boom in a particular state. Romer and Romer (2016) 8To be clear, the identication assumption (internal validity) requires that income growth in Florida would be similar to that in other states (after controls) if Florida did not receive such a large Social Security transfer. Identication does not require that the eect of actual transfers in Florida are the same as those in other states, which is about external, rather than internal, validity. In principle, the eects of transfers could be heterogeneous across states, but identifying heterogeneity is dicult with only 50 states. 9 show that my sample of Social Security increases did not have countercycli- cal motivations. Stimulus payments were widely spread across states and had 9 simple eligibility rules, making it dicult to target them at specic states. My allocation of transfers across states (statesharei,t−4 ) cannot be aected by contemporaneous shocks because they are based on last year's eligibility char- 10 acteristics. The lagging of eligibility criteria does not prevent an accurate allocation because stimulus payments actually used last year's income for el- igibility and demographic characteristics are predetermined and slow-moving over the medium term. For example, the size distribution across states of the 1972Q4 increase in Social Security payments was almost exactly proportional to the 1970Q2 increase in Social Security payments (R-squared of 0.96) (blue circles on the left side of Appendix Figure 1.1). The secondand perhaps more seriousconcern is omitted variable bias. The rst way to address this is to include state FEs (µi ) and quarter FEs (µt ) in all specications. State FEs control for all state-level trends (like the faster growth of Sun Belt states relative to Rust Belt states), and quarter FEs remove aggregate variation (such as the US business cycle, monetary policy, aggregate expectations, or international shocks). The inclusion of state and quarter FEs means that regressions only use variation in transfers and growth both across states and over time, and so any omitted variable would also have to vary across states and over time. Including both eects in all specications drastically reduces the number of potentially confounding variables. To further reduce the risk of omitted variable bias, I use a research design that involves many transfer policy changes with dierent sizes and timings and over a long period. Through this design, any omitted variable would need to take an unlikely time-varying pattern in high- versus low-transfer states, which is a much higher hurdle for potential confounding variables than in a standard dierence-in-dierence study with a one-o permanent policy change. For example, Social Security stipends increased by 10% in June 1971 and 20% 9Voting records reveal that political support for one-o transfers was either mostly partisan (2001) or bipartisan (2008), with no evidence that legislators from slow-growing states were more likely to support the legislation (see Appendix 3.2). 10After removing trends and aggregate variation, state growth rates have little quarterly persistence, which removes the possibility of ex-ante targeting states in recession. 10 in October 1972, but not at all in 1973, and so the confounding variable would have to follow a similar pattern in high-transfer states and only boost growth in those specic quarters. For temporary transfers, the confounding variable not only has to spuriously boost growth in high-transfer states in the relevant quarter but also has to reduce growth in the following quarter when the transfer 11 is withdrawn. Moreover, the identity of high-transfer states changes over time and across transfer policies, which means that the confounding variables would also have to change. For example, the cross-state allocation of transfers in 2008 and 2001 are quite dierent, with the 2008 allocation only explaining 37% of the variation in 2001 transfers (Appendix Figure 1.1 right side). This dierence stems from the greater progressivity of the 2008 transferwhich was refund- able with a high-income phase-outrelative to the 2001 transfers, which were not refundable with no phase-out. Demographic changes over a longer period (19521972) led to a reallocation of Social Security benets across many states, such that the allocation in 1952 only explains 28% of the allocation in 1972 (Appendix Figure 1.1 left side, green triangles). Omitted variable bias is further reduced by adding other controls that vary by state and over time and by also adding a number of placebo tests in which the transfer is assumed to have occurred in a dierent period. Additional controls include state-specic sensitivities to the national business cycle or oil prices, dierential growth trends depending on industry composition, controls for state-specic quadratic trends, as well as the removal of inuential years or states. 2.3. Empirical Specication. My main specication is the following regres- sion for state i at quarter t: (2) ∆Yi,t /Yi,t−1 = β0 ∆tri,t /Yi,t−1 + δ Xit + µt + µi + eit where ∆Yi,t /Yi,t−1 is the quarterly growth rate of per-capita income (excluding the transfer), ∆tri,t /Yi,t−1 is the contemporaneous change in my constructed 11For example, in 2001 the confounding variable would have to boost growth in 2001Q3 in high-transfer states like Montana and then reduce growth in 2001Q4. 11 exogenous transfer measure as a share of income, µi and µt are state and quarter FEs, respectively and eit is the error term. Xit is vector of controls (and δ is vector of associated coecients) that varies across specications. The transfer measure is scaled by lagged income to generate a multiplier interpretation of coecients (i.e., the dollar value of extra non-transfer income produced in a state when its residents receive an extra dollar of transfers). This specication of growth rates regressed on a scaled scal shock is standard in 12 the multiplier literature. β0 in Equation 2 is known in the literature as the impact multiplier and is the main coecient of interest in this paper (I also calculate cumulative multi- pliers below, which turn out to be similar). Relative to cumulative multipliers estimated over longer horizons, impact multipliers have two advantages: rst, they require weaker identication assumptions, as there is only a short window between cause and eect that reduces the number of possible omitted variables; and second they tend to be estimated more precisely (smaller standard errors). In the spirit of Romer and Romer (2010, 2016), Equation 2 is estimated as a reduced form (by ordinary least squares (OLS)) rather than by instrumental variables (IV) (I also produce IV estimates, which turn out to be similar). The reduced form is appropriate as my exogenous variation is the variable of economic interest (the state-level transfer shock induced by an aggregate policy change), rather than being correlated with the variable of interest, as is typically the case. I use two measures of non-transfer income as left hand side variables: labor pc income per capita Yit = (W L)i,t (equivalent to the wage bill), which is available pc 13 for the whole sample, or quarterly GDP per capita Yit = GDP i,t from 2005. All variables are deated by the national quarterly Personal Consumption 12For example, a specication of GDP growth rates regressed on the change in expenditure as share of GDP (or similar) is used in Barro and Redlick (2011), Nakamura and Steinsson (2014), Miyamoto, Nguyen, and Sergeyev (2018), Kraay (2014), among others. ∆tri,t /Yi,t−1 is equivalent to the change in transfers per capita as a share of income per capita. 13Labor income data come from the Bureau of Economic Analysis (BEA) under the ocial title Earnings by place of work and mostly consist of wage and salary disbursements (70%). Labor income excludes income from transfers and is before income taxes. I focus on these dependent variables as other potential outcome variables are missing at the state level (consumption), are not available quarterly (GDP before 2005), are not seasonally adjusted, or are not available for both temporary and permanent transfer samples. 12 Expenditures (PCE) Chain-type Price Index, as there are no ocial state- 14 level price data. As state-level growth can be extremely volatile, especially for small states, I drop outliers from all specications where the annualized growth rate of per capita labor income growth or GDP is more than 20% in absolute value. Standard errors are cluster-robust (clustered at the state level), which allows for heteroskedasticity and arbitrary serial correlation. See Appendix 1 for descriptive statistics and additional information on data sources and construction. 2.3.1. Cumulative Multipliers. As an extension in Section 3.6, I estimate dy- namic cumulative multipliers several quarters after the transfer is paid. The cumulative multipliers Ch are dened as the cumulative change in an income variable relative to the cumulative spending on the transfers over a horizon of h quarters beyond the impact quarter (h +1 quarters in total). I use two specications, both of which collapse to Equation 2 for h = 0. For permanent transfers, I use a direct projections specication (Jordà 2005), which has been used by Miyamoto, Nguyen, and Sergeyev (2018) and 15 others to estimate cumulative multipliers. The estimated equation is: h h Yi,t+j − Yi,t−1 tri,t+j − tri,t−1 (3) = βh + δ Xit + µt + µi + eit j =0 Yi,t−1 j =0 Yi,t−1 h where j =0 (Yi,t+j − Yi,t−1 )/Yi,t−1 is the sum of h +1 growth rates over horizons h up to h periods (left side) and j =0 (tri,t+j − tri,t−1 )/Yi,t−1 (right side) is the sum of changes in scaled transfers over the same period. The cumulative multiplier can be estimated directly as Ch = βh . For one-o transfers, I use a distributed lag (DL) specication, which adds extra lags of the transfer variable ∆tri,t−j /Yi,t−j −1 to the right hand side of 14While the BEA does produce estimates of state-level GDP deators, these variables use national prices for dierent industries and are weighted using state-specic industry weights, which do not capture the price eects from local demand shocks. See Appendix 3.1 for robustness tests using constructed state-level ination indices. 15Kraay (2014) uses a similar specication (in changes) but takes the sum after estimation. Ramey and Zubairy (2018) apply the same specication but estimate in detrended levels. 13 16 Equation 2. A distributed lag (DL) specication is standard in the MPC lit- erature for one-o transfers, such as in JPS and Parker et al. (2013). Although the direct projections approach is more popular in the multiplier literature, regressions on simulated data reveal that it has diculty isolating the lagged eects of one-o transfers over short horizons, whereas the DL specication uncovers the correct multipliers exactly (see Appendix 3.5). Other specica- tions are reported in Appendix 3.3 and Appendix 3.4, including a specication in detrended levels, which produce broadly similar results. 3. Empirical Results 3.1. Graphical Evidence. Figure 1 provides a visual representation of the size and direction of the relationship between changes in transfers on the x- axis (permanent and temporary), and contemporaneous growth in per capita labor income and GDP (y-axis). The slope of relationship is a simple estimate of the size of the cross-region transfer multiplier. For the temporary transfers (Figure 1, Panels BD), each of the 50 dots represents the growth rate and transfer size in a dierent state in the quarter the transfer was paid. For permanent transfers (Figure 1, Panel A) there are too many Social Security increases to be plotted in this way. So instead, I group permanent transfer change residuals into 50 bins based on their size (across dierent quarters and states) and plot the average transfer size in each bin (x-axis) against the average growth rate residuals of per-capita labor income for that bin (y-axis). 17 Residuals are calculated after removing time and state FEs. . In all four subplots in Figure 1 there is a positive and signicant relationship (line slope), indicating that states that received larger transfers tended to grow faster contemporaneously. Also note that the relationship between transfers and contemporaneous growth has a steeper slope for permanent than tem- porary transfers. For permanent transfers (Panel A), the slope is about 1.9, 16Specically, β ∆tr /Y in Equation 2 is replaced by h γj ∆tri,t−j /Yi,t−j −1 , with 0 i,t i,t−1 j =0 h the cumulative multiplier being the sum of the lags for a one-o transfer Ch =γ j =0 j . 17Quarters without any Social Security increases are dropped. Figure 1, Panel A excludes bins where the normalized absolute transfer change is tiny (less than 0.015% of labor in- come). This does not aect the slope or t-statistic, but makes the relationship more visible. The graph with all the bins is shown in Appendix Figure 3.5. 14 Growth in labor income per capita, 2008Q2 (%) Quarterly Growth in Real Labor Income PC (%) A. Quantiles of Permanent Transfers and Growth B. 2008Q2 Stimulus Payments (WL) 0.8 OK 2 41 50 WV 0.4 49 LA 4045 TX 42 Line slope=1.9 PA MO KS Line slope=0.28 [t=2.02] 0 7 48 AK NE COWY [t-stat=2.8] MD UT KY AR MS 10 44 CA VA WI NM 43 0.0 8 MA OH 1 WA IL IA MI ME 11 47 HI OR DEND NC SC AL 2 NH INID 3 9 46 CT NJ VT GAFL MT TN 6 12 AZ -2 MN RI -0.4 NY SD 5 4 NV -0.8 -4 -0.4 -0.2 0.0 0.2 0.4 2 3 4 5 6 7 Permanent Social Security Transfers (% of labor income) Transfer stimulus payments, % of labor income (2008Q2) Growth in labor income per capita, 2001Q3 (%) C. 2008Q2 Stimulus Payments (and GDP) D. 2001Q3 EGTRRA Stimulus Payments Growth in state GDP per capita, 2008Q2 (%) OK NM 2 DE 4 HI WY LA WV WV MT 1 AZ SD IA ND Line slope=1.26 AK MD KS KY OKALFL LANM GA UT PA ME [t=4.03] 2 Line slope=0.9 IL MO TX TNNE RI OH SC VT AR 0 ND [t=2.53] TX NJ WI AR IN ID CO NV MI OR NH MS MA CA OR PA CA MD NC MO KY -1 MN UT NE VA AL MS NY CT CO 0 ID NC WA KS OH MT VA ILHI TN ME MA SC IN -2 MN WI GA MI CT NH FL VT NJ RI WA -2 DE IASD AZ NY NV -3 1.5 2 2.5 3 3.5 4 4.5 1.5 2 2.5 3 3.5 Transfer stimulus payments , % of state GDP (2008Q2) Transfer stimulus payments, % of labor income (2001Q3) Notes: Panel A: Each point is the mean growth rate residual or mean Social Security increase residual (share of labor income) of 50 quantiles (controlling for state and time FEs). Quantiles with tiny transfer residuals are not shown. Panels BD: each point in the scatter plot is the growth in per-capita labor income or GDP per capita (GDPPC) (y-axis) in the quarter, plotted against the size of the contemporaneous transfer (x-axis). Sources: BEA, Romer and Romer (2016). Figure 1. Scatter Plots: Transfers and Economic Growth signicant at the 1% level. This relationship is not driven by a couple of bins: an outlier-robust regression generates a similar slope and t-statistic. For temporary transfers, the line slope ranges from 0.31.2, depending on the quarter and income variable, and the scatter plots help us to understand the drivers of cross-state variation in transfers. Figure 2 Panel B plots 2008 trans- fers (share of labor income) against growth, which have a relative multiplier (line slope) of about 0.28 (t -stat = 2.0)remarkably similar to the pooled temporary transfer multiplier estimated later in regressions. The cross-state variation in transfers is striking: The 2008 ESP ranges from around 2.3% of 15 quarterly labor income in Connecticut (CT) to 6.4% of quarterly labor income in Mississippi (MS). Two factors are at play here: First, the level of per-capita labor income is much lower in MS than CT, and so xed dollar payments are mechanically more important. Second, much of the variation is driven by the low-income rebate (those paying no net taxes), which is focused toward poor states. Of the 4.1 percentage point gap in stimulus transfers in 2008Q2 be- tween MS and CT, 3.4 percentage points are due to the cross-state allocation of the low-income component. Panel C of Figure 2 shows the same stimu- lus package and quarter but for quarterly growth in GDP per capita rather than labor income, which generates a steeper slope of 0.9, but one that is less precisely estimated (t -stat = 2.5). The nal scatter plot (Panel D) displays the relationship between the 2001 EGTRRA stimulus payments (as a share of labor income, x-axis) and contemporaneous per-capita labor income growth in 2001Q3 (y-axis). Again, states that received a larger payment (as a share of labor income) tended to grow faster contemporaneously, though here the multiplier is large at 1.25 (t -stat = 4). I do not put too much weight on these scatter plots quantitatively as they have no controls. 3.2. Main Empirical Results. The main empirical results are presented in Panel A of Table 1, which pool across all transfer changes grouped by persistence and dependent variable. The rst row presents a parsimonious specication, with no controls beyond state and quarter FEs. The second row presents a benchmark specication, which includes controls for US GDP growth × state FEs, and population growth, and is used for placebo tests (Figure 2) and used to compare with the theoretical model in Section 5. The 50 US GDP growth × state FEs are to control for dierential state sensitivities to the aggregate business cycle  potentially important as the 2001 and 2008 transfers were explicitly countercyclical at the aggregate level  discussed further below. Population growth is important because sunbelt states with more retirees tend to have faster growing populations than other states, which might aect per capita income growth and also be correlated with transfer size (for example retirees claiming Social Security benets migrate to warmer states). 16 Table 1. Cross Region Transfer Impact Multiplier Estimates TABLE 1—CROSS-REGION TRANSFER IMPACT MULTIPLIER ESTIMATES Permanent Tr. Temporary Transfers (Social Security) (2001 and 2008 Stimulus Payments) Dependent Variable is quarterly growth in real per capita: Labor Income Labor Income GDP (1952-1974) (2001–2008) (2005–2008) Specification varies by row: Multiplier Obs. Multiplier Obs. Multiplier Obs. Panel A. Impact multiplier estimates for main specifications A1. Parsimonious Specification 1.47 4413 0.26 1588 0.39 742 (no additional controls) (0.49) (0.08) (0.16) A2. Benchmark Specification 1.29 4413 0.31 1588 0.41 742 (controls for population growth & (0.54) (0.09) (0.18) US GDP Growth × State FEs) Panel B. Impact multiplier estimates with additional economic controls (added one-by-one) B1. Controlling for US GDP Growth 1.14 4413 0.31 1588 0.42 742 × State fixed effects (0.53) (0.09) (0.18) B2. Controlling for Oil Prices (and lag) 1.57 4413 0.27 1588 0.42 742 × State fixed effects (0.50) (0.08) (0.15) B3. Controlling for state-specific 1.71 4413 0.25 1588 0.41 742 quadratic trends (0.45) (0.08) (0.17) B4. Controlling for state manufacturing 0.92 4413 0.25 1588 0.41 742 GDP share × quarter fixed effects (0.45) (0.09) (0.17) B5. Controlling for Social Security 1.51 4413 - - back payments (0.51) B6. Controlling for 2001/08 recession - 0.26 1588 0.39 742 depth × aggregate transfer size (0.08) (0.16) Panel C. Impact multiplier estimates with additional dynamic controls (added one-by-one) C1. Controlling for lagged log income 1.53 4413 0.26 1588 0.47 742 (0.57) (0.08) (0.16) C2. Controlling for lagged dependent 1.18 4281 0.23 1577 0.37 686 variable (0.51) (0.08) (0.15) C3. Controlling for lagged transfers 1.36 4413 0.21 1588 0.45 742 (0.48) (0.09) (0.24) Panel D. Impact multiplier estimates for other samples or IV specifications D1. Excluding influential states or 1.43 4051 0.26 1556 0.43 727 years (0.69) (0.10) (0.21) D2. Instrumenting Social Security 1.88 4413 0.20 1588 0.36 742 transfers or other current transfers (0.59) (0.07) (0.14) IV first stage F-statistics: 522 882 843 State and quarter fixed effects All Specifications All Specifications All Specifications Notes: Each cell reports the impact multiplier of a regression of growth in per capita labor income (Columns 1 and 2) or GDP (Column 3) on the scaled change in permanent Social Security benefits (Column 1) or temporary stimulus transfers (Columns 2 and 3), as in Equation 2. All specifications have state and quarter fixed effects (FEs). The main results are in Panel A: Row A1 is a parsimonious specification with no further controls, and Row A2 is a benchmark specification used in figures (except scatter plots) with controls for US GDP growth × state FEs and population growth. Panels B and C add controls one-by-one to the parsimonious specification. Row B1 adds US GDP growth × state FEs; Row B2 adds state FE × log(OilPr t), log(OilPrt-1); Row B3 adds state FEs × t, t2; Row B4 adds the manufacturing share of state GDP × time FE; Row B5 controls for ad-hoc temporary Social Security backpayments (t and t-1). Row B6 controls for the interaction between the depth of the 2001/08 recessions in each state and the mean transfer variable each quarter. Panel C Row 1 controls for the lagged log level of per capita labor income (Columns 1 and 2) or per capita GDP (Column 3). Panel C Row 2 adds a control for the lagged dependent variable. Panel C Row 3 adds a control for the lagged transfer variable. Panel D Row 1 drops influential states or years from the parsimonious specification, as identified by leave-one-out regressions in Appendix 3 and scatter plots; for permanent transfers: years 1952 and 1972; for temporary transfers: Mississippi. Panel D Row 2 reports estimates from an instrumental variables 17 (IV) specification, where all BEA social security transfers (Column 1) or “All other personal current transfer receipts” (Columns 2 and 3) are instrumented with the constructed transfer measures used in the other rows. First stage coefficients are close to one (across Columns 1–3: 0.8, 1.1 and 1.1, respectively). Coefficients on controls are not reported. Outliers |growth|>20% (annualized) dropped. Robust std errors are in parentheses (clustered by state). Table 1 column 1 reports cross-region transfer multiplier estimates for per- manent Social Security benet increases. In general, I nd that a $1 increase in permanent Social Security transfers to the residents of a state increases that state's relative labor income by around $1.5 contemporaneously, though the size of the multiplier can vary from 0.91.9 (about one standard error) across all specications in Table 1. In the parsimonious specication (row A1), the relative permanent transfer multiplier is almost exactly 1.5 (signicant at the 1% level). For the benchmark specication (row A2), the relative multiplier falls slightly to 1.3 (also signicant at 1%). This reects partially osetting eects of the two controls; if added individually, US GDP growth × state FEs tend to reduce the permanent transfer multiplier to around 1.14 (row B1, signicant at the 1% level), whereas population growth tends to increase the 18 multiplier to 1.61 (not reported, also signicant at the 1% level). Table 1 Columns 2 and 3 reports cross region transfer multiplier estimates for one-o stimulus transfers, using per capita labor income or GDP (respec- tively) as the income measure. Unlike the scatter plots in Figure 1, the regres- sions control for state and time FEs, impose that the withdrawal of transfers reverses any earlier increase in income, and for labor income, pools across the 2001 and 2008 stimulus payments (recall GDP data are only available from 2005 so excludes the 2001 transfer). In general, I nd that $1 one-o trans- fers to the residents of a state increases that state's relative labor income or GDP by around $1/3 contemporaneously, with multipliers around $0.4 for GDP (range 0.350.47), and $0.25 for labor income (range 0.200.31). In the parsimonious specication (row A1), the temporary cross-region transfer mul- tiplier is around 0.26 (signicant at the 1% level), rising slightly to 0.31 for the benchmark specication (also signicant at 1%). When the dependent variable is per-capita GDP growth, the multiplier for the parsimonious specication is larger at 0.39 (row A1), increasing marginally to 0.41 in the benchmark speci- cation (row A2). Both GDP multipliers are signicant at the 5% level because standard errors are larger for the GDP specication. 18The coecient of population growth is negative (0.37) and is statistically signicant (not reported). Population growth has little eect on temporary transfer estimates and is often insignicant. 18 3.3. Robustness Tests (Controls, Samples and Estimation Methods). To address potential omitted variable bias in the main results, I conduct a suite of robustness tests to dierent controls, dynamics, specications and estimation methods, which are presented in the remaining panels of Table 1. 3.3.1. Additional Controls. Table 1 Panel B controls for additional economic variables (one-by-one) that might be correlated with both state-level growth and transfer size. As mentioned above, the 2001 and 2008 one-o transfers were explicitly countercyclical measures. While the aggregate business cycle is subsumed into quarter FEs, dierent states might have dierent sensitivities to the national cycle (e.g. if they specialized in cyclically sensitive industries like manufac- turing). If more of these more-sensitive states happened to receive relatively smaller transfers, that could drive a positive multiplier. I address this con- cern in three ways. First, for the regression in Table 1 row B1, I include US GDP growth × state FEs (also included in the benchmark) which increases the estimated temporary transfer multipliers marginally but does not aect sig- nicance. Second, I interact time FEs with the average manufacturing share of the state's GDP, which allows manufacturing-intensive states to grow more slowly in recessions  which also has little eect (row B4). Finally, I construct a peak-to-trough measure of the depth of the 2001 and 2008 recessions in each state and interact it with the mean size of the transfer variable each quarter 19  again multipliers are unchanged (row B6). Although my sample of Social Security benet increases are not explicitly countercyclical (by Romer and Romer's narrative), I also control for industrial structure and dierential state sensitivities to the aggregate business cycle for robustness. As mentioned above, adding 50 control variables for US GDP growth × state FEs (Table 1, row B2 rst column), reduces the estimated multiplier to 1.14 (signicant at the 5% level). Controlling for state average manufacturing share of GDP × quarter xed eects (92 variables, Table 1, row B5 rst column) reduces the permanent transfers falls to 0.92 (signicant at the 19The peak (trough) is the maximum (minimum) per-capita income in the two years before (after) the NBER recession, 19992000 and 20062007 (20022-03 and 20091910). These controls turn out to be insignicant at the 5% level. 19 5% level), though I cannot reject that the coecient diers from 1.5 (as in the parsimonious specication). The low coecient can partly be explained by the saturation of controls rather than economic trends; interacting manufacturing share with annual rather than quarterly FEs yields a multiplier of 1.57 (not reported). Oil prices fell rapidly during the 2008 recession, and were also volatile in the early 1970s. Oil prices are naturally subsumed into time FEs, but growth in oil- producing states might be stimulated (hurt) by higher (lower) oil pricesand those states might have coincidentally received higher or lower transfers. To assuage this concern, I add 100 controls for state-specic sensitivities to log real oil prices (50 contemporaneous variables and 50 rst lags, which exibly controls for price levels and changes). For both permanent and temporary transfers, the estimated multiplier is very similar to that in the parsimonious specication, and signicant at the 1% level (row B2). To cover any other general time-varying state-specic covariates, I also con- trol for a quadratic state-specic polynomial trend (some terms of which are dropped due to collinearity). This yields a permanent transfer multiplier of 1.71 (signicant at the 1% level, row B3 rst column)slightly higher than that in the parsimonious specication, but has little eect on multiplier esti- 20 mates for temporary transfers. Finally, I control for an omitted variable specic to the permanent transfer sample: three one-o back payments (in 1965, 1970, and 1971) in compensation for delayed increases in Social Security benets (Romer and Romer 2016). These temporary increases coincided with the increases in permanent Social Security benets in those yearsand had similar eligibility requirements. In Table 1 row B5, I control for these temporary transfers (and their rst lag), 21 which has little eect on the permanent transfer multiplier. 20Appendix Table 3, presents several robustness tests with dierent standard errors (ho- moskedastic, robust without clustering and allowing for spatial error correlation). Perma- nent transfer multipliers are still signicant at the 5% level, and temporary transfer mul- tipliers signicant at 1% or 5%, except for homoskedastic errors in the GDP specication (but in any case, homoskedasticity is an extreme assumption). 21The temporary Social Security increases and rst lag are insignicant (coecients of 0.58 t ( = 1.33) and 0.16 ( t = 0.28), respectively). I do not put much weight on these temporary transfer estimatesrelative to those of the stimulus paymentsbecause they are imprecise: The standard errors here are around ve times larger. 20 3.3.2. Dynamic Controls. Table 1 Panel C addresses the concern that the par- simonious specication in Equation 2 might miss important dynamics. The rst row (C1) adds the lag log level of income per capita to control for the eect of unconditional convergence on growth rates. Although this control is negative and highly signicant (perhaps also reecting mean reversion), it does not much aect the size or signicance on the transfer multiplier estimates. The second row (C2) allows for persistence in state level growth rates, by con- trolling for a lagged dependent variable, which has little eect on the size or signicance of temporary transfer estimates. However, the lagged dependent variable reduces the estimated permanent transfer multiplier to 1.2 (still sig- nicant at 5%), though this is mostly due to sample selection (excluding the quarter after a removed outlier). A nal concern regarding dynamics is that the specication in Equation 2 imposes a symmetric up-down prole of the eect of temporary transfers: if a one-o transfer causes a rise in output by X% output when paid, then withdrawing those transfers a quarter later should lead to a fall in output of X %. If the restriction does not hold then impact multiplier estimates could be biased. Here I add an extra lag of transfers as an additional control in row C3, which partially relaxes this restriction. Temporary transfer multipliers for labor income are similar. The multiplier for GDP is slightly larger but is less precisely estimated, causing the p-value to fall to 6%. Permanent Social Security transfer increases are never withdrawnso are not aected by the symmetry assumptionbut the extra lag provides a robustness test against mis-specication of the timing of transfer payments across quarters. The im- pact multiplier is mostly unchanged (signicant at the 5% level) and the rst transfer lag is only signicant at the 10% level. 3.3.3. Alternative Samples. So far I have tested robustness to dierent con- trols, but it is also possible the estimated multipliers are driven by specic states or years and might not be reective of a general relationship. For per- manent transfers, no individual states are inuential when dropped one-by-one (see Appendix Figure 3.3), but the multiplier does move by more than one standard error if either 1952 or 1972 are excluded. In Table 1 row D1, I re- port estimates excluding both these inuential years, which has little eect on 21 the multiplier, as the eect of these years is mostly osetting (standard errors are slightly larger, and signicance is now at the 5% level). For temporary transfers, the scatter plots in Figure 2 suggest that Mississippi (MS) might be inuential. In the right two columns of row D1, I drop MS and show that it has little eect on estimated multipliers, which are still signicant at 5%. No other states are inuential either when dropped one-by-one (see Appendix 22 Figure 3.2). 3.3.4. Instrumental Variable Estimates. As discussed in Section 2.3, the spec- ication (Equation 2) is a reduced form as the exogenous changes in transfers are essentially the variable of interest. However, it is worth estimating an instrumental variables specication (row D2) in case the constructed transfer variables are measured with error (which could attenuate multiplier estimates), or there are other correlated transfer payments that I am not capturing (lead- ing to upward bias). For permanent transfers in the rst column, I instru- ment all Social Security transferswhich are potentially endogenoususing my sample of exogenous Social Security increases. The multiplier is around 23 1.9, signicant at the 1% level (the rst-stage F-stat is above 500). For one-o transfers in the second and third columns, I instrument changes in the potentially endogenous BEA category All other personal current trans- fer receipts (as a share of labor income or GDP) using the pooled 2008 ESP 22The second last column of Appendix Table 3 adds back in previously dropped extreme outliers, where |growth|>20% (annualized). For permanent transfers that reduces signi- cance to a p -value of 8%, and it also increases the estimated multiplier to 2. For temporary transfers, there is little eect on the labor income specication, but these extreme observa- tions make the GDP specication go to zero (and become insignicant). Dropping the two smallest states in 1980 (Alaska and Wyoming), with extremely volatile growth rates and a combined population of <1m increases the permanent transfer multiplier substantially (pos- itive and signicant at the 1% level), and results in multipliers for the GDP specication similar to those estimated above (signicant at 5%). See Appendix 3.01 for further details. 23The size of the IV coecient is unsurprising, as a $1 increase in my exogenous permanent Social Security transfers series increases all BEA Social Security payments by about 80 cents (after controlling for time and state FEs). The strong rst-stage relationship is plotted in Panel B of Appendix Figure 3.1 and provides an additional validation of the cross-state allocation of transfers and Romer and Romer's aggregate narrative. 22 transfers (as above) as instruments. Multipliers are similar to the parsimonious 24 specication and are signicant at the 5% level. 3.4. Placebo Tests. Even if I cannot control for potential confounding vari- ables, I can test for omitted variable bias by running a placebo regression using counterfactual growth rates when the confounding variable might be inuen- tial, but no transfers were actually paid. A non-zero multiplier during these periods might indicate a spurious relationship. In the rst set of placebo re- gressions, shown in Panel A of Figure 2, I test whether states receiving large one-o transfers always grew faster during recessions by regressing growth rates during the three quarters of the 19901991 recession on the 2001 and 25 2008 transfers. The estimated multipliers are always insignicant and are usually close to zero. Alternatively, the potential confounding variable might be inuential in the years around the time the transfer is actually paid. To test for this, I counter- factually move transfers backward or forward by up to six quarters (using the Benchmark with controls specication from Table 1). In addition to detect- ing spurious results, it can also pick up anticipation eects (for short leads, t < 0) or delayed eects (for short lags, t > 0). For permanent transfers, the results are shown in Figure 2, Panel B. One can see that the largest t -statistic is at t = 0 (when the actual transfers occurred), and all other leads and lags are insignicant at the 5% level (t -statistic less than the red cuto of 2), which 26 is what one would expect to see. 24This endogenous BEA category includes all transfers that are not Social Security Benets, Medicare or Medicaid payments or state unemployment compensation. The BEA classies tax rebates as transfers if they are at least partially refundable, which includes all 2008 transfers but excludes the 2001 transfers and so the latter are not included in the IV estima- tion. Multipliers are similar in the IV and reduced form specication because growth in All other personal current transfer receipts in 2008Q2 is extremely well explained by growth in ESP transfers as constructed above (see Appendix Figure 3.1, Panel B), with rst-stage coecients close to one and rst-stage F-stats over 800. This provides an external validation of my construction of the cross-state allocation of the 2008 ESPs. 25According to the NBER's denition, the 19901991 recession lasted through 1990Q3, 1990Q4, or 1991Q1, so I report placebo payouts during all of those dates. I thank an anonymous referee for suggesting this placebo test. The sample period for these regressions is 19901997 and they use the benchmark specication. 26The estimated multiplier (green line) appears to be negative in the couple of quarters in anticipation of the Social Security rate increase, positive in the few quarters following the 23 A. Placebo during 1990-91 Recession B. Permanent Transfers Placebo/Lead/Lag Actual 2001 Transfers Actual 2008 Transfers 3 7 2 6 -1.0 -0.5 0.0 0.5 1.0 1.5 1.5 1 5 Beta coeff & CI Absolute t-stat Multiplier and 95% CI 1.0 0 4 0.5 -1 3 t-stat=2 cutoff -2 0.0 2 -3 1 -1.0 -0.5 -4 0 -6 -4 -2 0 2 4 6 Lag (Quarters) 1990Q3 1990Q4 1991Q1 1990Q3 1990Q4 1991Q1 Beta Coeff. (LHS) Absolute t-stat (RHS) Placebo timing first transfer received C. Temp Transfers (WL) - Placebo/Lead/Lag D. Temp Transfers (GDP) - Placebo/Lead/Lag .5 5 1 6 .5 5 0 4 Beta coeff & CI Beta coeff & CI Absolute t-stat Absolute t-stat 0 4 -.5 3 -.5 3 t-stat=2 cutoff -1 2 t-stat=2 cutoff -1 2 -1.5 -1.5 1 1 -2 -2 0 0 -6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6 Lag (Quarters) Lag (Quarters) Beta Coeff. (LHS) Absolute t-stat (RHS) Beta Coeff. (LHS) Absolute t-stat (RHS) Notes: Panel A: multiplier and 95% condence intervals (CIs) from regression of real labor income growth during the 19901991 recession on actual state-level transfers from 2001 or 2008 using the benchmark specication. Panels BD: Regressions of growth on transfers using the benchmark specication (Permanent Transfers, Temporary Transfers for labor income and Temporary Transfers for GDPPC, respectively), but moving the growth variable backward/forward by up to 6 quarters. Source: BEA, Romer and Romer (2016), author's calculations. Figure 2. Placebo Tests For temporary transfers (Figure 2 Panels CD), the largest t -statistic is always at t = 0 (when the actual transfers were paid), and there is little increase, and sometimes signicant at the 10% level. Romer and Romer (2016) nd that Social Security increases were typically legislated roughly a quarter in advance of payment, though they nd no evidence of an increase in aggregate consumption at that time. Antici- pated demand shocks in NK models often reduce output as rms raise prices and markups in advance, a possible rationale for negative leads. However, I do not want to emphasize those results given their insignicance and general lack of robustness (not reported). 24 27 evidence of consistently positive (or negative) coecients at other times. However, there are some marginally signicant lags at t = 4 (a year before the transfer was paid) and t = 1 (the quarter after the transfer was paid) using the labor income specication and t = 1 (one quarter lead, GDP specication), and some other marginally insignicant leads and lags at the 10% level. The negative rst lead in the GDP specication seems to be due to the overlap of the placebo withdrawal of stimulus in 2008Q2 and the eect of the actual stimulus. If I run the same specication and omit 2008Q2 from the sample (when most of the transfers were actually paid) but keep 2008Q1 and the other quarters, the coecient on the lead halves, and the p -value increases to 0.4. The other marginally signicant coecients are to be expected. Under a random allocation with no true eect, the ≈40 placebo regressions in Figure 2 would produce two false positives on average when testing at the 5% level. Moreover, the chance of no false positives is also very low at around 12% (= 0.9542 under random allocation). Having now established the robustness of the main results, I now consider two extensions involving heterogeneity and dynamics. 3.5. Extension 1: Heterogeneity by Quarter and Transfer Policy. So far, I have presented regression results that restrict the temporary transfer multiplier to be identical across payment and withdrawal quarters and across policies. In this subsection I relax those assumptions, with results presented in 28 Table 2 (using the benchmark specication). I rst allow for heterogeneity by payment or withdrawal quarter, and then by transfer component, and discuss the results in the context of the large literature on the MPC of one-o transfer payments. Table 2 Column A relaxes the symmetry restriction in Equation 2, by al- lowing multipliers to vary across payment and withdrawal quarters. Despite substantial heterogeneity in estimated coecients, I fail to reject the restriction 27Both stimulus payments were legislated a quarter in advance, but there is little robust evidence of signicant negative anticipation eects. In any case, Ricardian households who respond to news also save temporary transfers. 28Results using the parsimonious specication are shown in Appendix Table 4 and are generally similar. 25 Table 2. Heterogeneity by Quarter and Transfer Policy TABLE 2—TEMPORARY TRANSFERS MULTIPLIERS -- HETEROGENEITY BY QUARTER AND TRANSFER POLICY A. Heterogeneity by quarter B. Heterogeneity by transfer component Labor Income GDP Labor Income GDP 2008Q2 Δ Transfer (paid) 0.27 0.90 2008 Δ Tax rebates 0.36 0.64 (0.14) (0.38) (low-income) (0.10) (0.25) 2008Q3 Δ Transfer 0.31 -0.19 2008 Δ Tax refunds 0.07 -0.18 (withdrawn) (0.12) (0.33) (middle-income) (0.31) (0.57) 2001Q3 Δ Transfer (paid) 0.84 2001 Δ Tax refunds 0.44 (0.34) (0.24) 2001Q4 Δ Transfer 0.03 (withdrawn) (0.41) P-value (equal coeff.) 0.41 0.10 P-value (equal coeff.) 0.59 0.24 State & quarter FEs Yes Yes State & quarter FEs Yes Yes Benchmark controls Yes Yes Benchmark controls Yes Yes Observations 1588 742 Observations 1588 742 Notes: Each column represents a regression of the growth rate of real labor income per capita or real GDP per capita on the change in normalized temporary transfers. The regressor is disaggregated by quarter of payment/withdrawal (Column A), or the transfer policy (Column B). All specifications have state and quarter fixed effects, with “benchmark” controls for population growth & US GDP Growth × State FE. Outliers >20% (annualized, absolute value) dropped. Robust std errors are in parentheses (clustered by state). that payment and withdrawal multipliers are equal (in both labor income and GDP specications; p-values around 0.4 and 0.1), which justies the pooled TABLE 2—TEMPORARY TRANSFERS MULTIPLIERS -- HETEROGENEITY BY QUARTER AND TRANSFER POLICY specication used in the main results in Table 1. For labor income, the 2008 multipliers A. areHeterogeneity around 0.3 by quarter B. Heterogeneity for both payment and withdrawal by transfer component quarters and are Labor Income GDP Labor Income GDP 29 remarkably (paid) to the similar 2008Q2 Δ Transfer pooled estimates 0.27* 0.90** in 2008 Δ Table 1. Tax rebates For trans- the 20010.64** 0.36*** (0.14) (0.38) (low-income) (0.10) (0.25) fers (labor income) and 2008 transfers (GDP), the point estimates are larger 2008Q3 Δ Transfer 0.31** -0.19 2008 Δ Tax refunds 0.07 -0.18 (withdrawn) when transfers are paid (0.12) than when (0.33) they are (middle-income) withdrawn, though (0.31)these(0.57) dier- 2001Q3 Δ Transfer (paid) 0.84** 2001 Δ Tax refunds 0.44* ences are not statistically signicant. Specically, the 2001Q3(0.24) (0.34) (labor income) 2001Q4 Δ Transfer 0.03 2008Q2 (GDP) payment-quarter multipliers are around 0.9, but multipliers (withdrawn) (0.41) onP-value (equal coeff.) withdrawal 0.41 are close to 0.10 means zero. This P-value (equal coeff.) states receiving a larger0.24 0.59 trans- State & quarter FEs Yes Yes State & quarter FEs Yes Yes fer controls to have Yes may continue Benchmark higher relative Yes Benchmarkeven income, after that controls transfer Yes Yeswas Observations 1588 742 Observations 1588 30742 withdrawn, though the evidence is not strong enough to be denitive. Notes: Each column represents a regression of the growth rate of real labor income per capita or real GDP per Column capita on the B of Table change 2 breaks in normalized down temporary pooled transfers. The temporary transfers by regressor is disaggregated into three quarter of payment/withdrawal (Column A), or the transfer policy (Column B). components described in Section 2.12008 refundable (low-income) rebates,All specifications have state and quarter fixed effects, with “benchmark” controls for population growth & US GDP Growth × State FE. Outliers >20% (annualized, absolute value) dropped. Robust std errors are in parentheses (clustered by state). 29When pooling across 2001 and 2008 payment quarters, and across 2001 and 2008 with- drawal quarters, the estimates are similar to the 2008 multipliers due to their greater leverage (not reported). 30Dierence-in-dierence studies often include a graph demonstrating that the outcome variable exhibits a similar trend in high vs low treatment states pre-treatment, with a gap in the outcome variable opening up post-treatment. Appendix Figure 3.4 shows this pattern. 26 2008 (middle-income) tax refunds, and 2001 transfersusing the same sample period and specication as the main results in Table 1 (which also imposes up- down symmetry). Despite substantial heterogeneity in estimated multipliers for dierent policies, I fail to reject the restriction that coecients are equal at the 59% (labor income) or 24% (GDP) level, justifying the pooled specication. Overall, the pooled results seem to be driven by the 2008 low-income re- bate, which is consistent with Parker et. al.'s (2013) nding that low-income households had a large and statistically signicant MPC out of the same 2008 31 transfers. With labor income as a dependent variable, the 2008 low-income rebates have a relative multiplier of 0.36 (signicant at the 1% level)similar to the pooled multiplier from Table 1. The 2001 EGTRRA stimulus payments also have a multiplier of 0.44 but have much wider standard errors (leading to signicance at the 10% level). For GDP growth as a dependent variable, coecient on the low-income rebates is even larger at around 0.64 (signicant at the 1% level), which also drives the pooled results. The 2008 middle-income tax refund seems to have little eect on growth. Specically, the coecient is close to zero for the labor income specication and negative for GDP specication, but insignicant in both cases. This might be because a larger fraction of this payment was saved, as middle-income households receiving the bulk of this transfer were not nancially constrained. Alternatively, it could be due to the transfer being spent on durable goods produced in other regions, which in turn depends on the transfer size (recall the 2008 mid-income rebate was around $600 per recipient, double the size of the other one-o transfers). Larger transfers can be spent on big ticket durable goods with less local content, and indeed Parker et al. (2013) nd a larger fraction of the 2008 rebate coecient was spent on consumption of 32 durables. 31JPS also found that the 2001 transfers were more likely to be spent by low-income house- holds, though the 2001 transfer was non-refundable and so cannot be disaggregated as easily in my sample. This could be because low-income households are more likely to be nan- cially constrained. However, Misra and Surico (2014) argue that both low- and high-income earners had the highest MPC, with the latter being the wealth hand-to-mouth of Kaplan and Violante (2014). 32Hausman (2016) also reports that a sizable fraction of the 1936 veterans' bonus was spent on durables, including cars, though that payment was much larger as a share of recipient income. Other evidence on the eect of transfer size on the MPC is mixed. Hseih (2003) 27 3.6. Extension 2: Cumulative Multipliers with Flexible Dynamics. In this section, I estimate cumulative multipliers Ch over the rst few quarters following each transfer shock, which allow for dynamics beyond the rst quar- ter (Figure 3 in black circles, with 95 percent condence interval (CI) bars). I use the benchmark specication with state and time FEs, and so C0 (h = 0) 33 represents the impact multipliers in the second row of Table 1. The speci- cations estimated are presented in Section 2.3.1. As permanent transfers are likely to have longer-lasting eects than one-o transfers, I estimate over a longer horizon (as do others in the literature). For permanent transfers, I fol- low Romer and Romer (2016) and estimate over a horizon of a year (h + 1 =4 quarters), and for one-o transfers I follow Parker et al. (2013) and estimate 34 over a horizon of half a year (h +1=2 quarters). For permanent transfers (Figure 3, Panel A), the cumulative multipliers are roughly constant at about 1.5, and are similar to impact multipliers presented in Table 1, Panel A (within the 95 percent condence interval). While the multiplier rises to C3 = 1.7 after a year, this is less than 1/2 of a standard error from 1.5. For temporary transfer shocks and labor income (Panel B(ii)), C1 = 0.26 is very similar to the impact multiplier estimate of 0.31 (and insignicantly dierent). For temporary transfers and GDP (Panel B(iii)), the point estimate C 1 = 0 .8 is larger than the impact multiplier estimate of 0.4, but they are insignicantly dierent. In part this is because C1 is imprecisely estimated: the standard errors triple in width, even leading to a loss of statistical signicance 35 (from zero). nds that households saved large Alaska permanent fund dividends (PFD), but tended to spend small tax refunds. But Kueng (2018) nds that the PFD are spent, with an MPC of 0.25, mostly by high-income households. 33For h > 0, h j =0 (Yi,t+j − Yi,t−1 )/Yi,t−1 for I drop the top and bottom 1% observations of permanent transfers using the projection method. Due to mean reversion these outliers are not well captured dropping extreme quarterly growth rates. 34Longer horizons introduce other confounding variables, like the 2009 scal stimulus. 35Of the 0.4 increase in the GDP multiplier from C to C 0 1 in Figure 3, about 2/3 is due to a larger impact multiplier in a specication that includes a lagged transfer variable. However, the larger estimates of C1 are not robust, they only occur in the benchmark specication and not the parsimonious specication (see Appendix Table 7, Panel C). Additionally, lagged transfers have little eect on the impact multiplier in the parsimonious specication (Table 1 row C3 and Appendix Table 7, Panel C). 28 In sum, my main nding in this subsection is that the impact multipliers reported in Table 1 are a good summary statistic of cumulative multipliers over subsequent quarters; points estimates are often similar and any dierences are not statistically signicant. Ramey (2019) also nds that cumulative spending multipliers do not vary greatly by horizon. A. Permanent Transfers B. Temporary Transfers 3.0 Measured multiplier type: (i) Labor income (ii) Labor income (iii) GDP 2.5 Empirical estimates (measured multiplier) New Keynesian model (measured multiplier) 2.0 Neoclassical model (measured multiplier) Cumulative Multipliers New Keynesian model (actual GDP multiplier) 1.5 Neoclassical model (actual GDP multiplier) 1.29 1.21 1.0 1.10 0.5 0.31 0.41 0.49 0.31 0.31 0.0 0.01 0.01 -0.5 -0.50 -1.0 0 1 2 3 0 1 0 1 Horizon (h, Quarters after shock) Notes: This gure compares cumulative multipliers (y-axis) estimated in the data and theoretical models over dierent horizons. Panel A reports cumulative multipliers for permanent transfers, with temporary transfers in Panel B. The black line (circles) are estimated cumulative multiplier using projection methods for permanent transfers, and a distributed lag model for one-o transfers (see Section 2.3.1). Error bars are 95% condence intervals (CI)). The rst quarter is the impact multiplier, as in Table 1 (benchmark specication). Red lines (squares) and green lines (triangles) report cumulative measured multipliers (deated using national prices) in the New Keynesian model and Neoclassical models, respectively. Measured multipliers (theory and empirical) have a dependent variable of per capita labor income in sub-panels (i) and (ii), and GDP in sub-panel (iii). Dashed lines are actual (not measured) GDP multipliers over the same horizon in the NK model (blue diamonds), and Neoclassical model (orange crosses). Figure 3. Transfer Cumulative Multipliers (Data and Models) 29 4. Understanding Cross-region Transfer Multipliers in an Analytical Model This section shows analytically the determinants of the cross-region transfer impact multiplier estimated in Section 3: when it will be large, small, or nega- tive and how it diers from the MPC and cross-region purchase multiplier. To do this, I use two simplied open-economy models that I can solve analytically: (i) the limit of a full New Keynesian (NK) model as prices/wages become rigid (an ultra Keynesian model) and (ii) a canonical Neoclassical (NC) model. The NC model and full NK model are used in Section 5 to quantitatively in- terpret the empirical estimates. So long as scal shocks are not too persistent, the analytical impact multipliers presented here are informative about those 36 presented in the empirics and in the full quantitative models. As the models are relatively standard, I provide an overview here, with details in Appendix 4. The full NK model is similar to NS's separable preferences, incomplete mar- kets NK model (Section IV, D of their paper) but with two main dierences: ω of households consume their income hand-to-mouth as they can- rst, a share not borrow/save (the remaining fraction 1 − ω are Ricardian and borrow/save through a risk-free bond), and second, wages are sticky and cannot be changed each quarter with probability θw , as in Erceg, Henderson, and Levin (2000) 37 (θp is the analogous Calvo stickiness of prices). Hand-to-mouth households are needed to match the multiplier on temporary transfers, and sticky wages 38 are needed to match the impact multiplier for permanent transfers. The 36For highly persistent shocks, some dierences arise as relative prices move. Specically, persistent shocks allow for substantial movement in relative prices in the full NK model in the long run, which result in lower multipliers than predicted by the simple NK model with rigid prices and wages. In the NC model with persistent shocks, regional prices adjust so empirical measured multipliers using national price deators are higher than from actual multipliers (using regional deators)see Section 5. 37As is common in models with hand-to-mouth households and sticky wages, (i) forward- looking nominal wage setting only takes into account the Ricardian household's labor-leisure rst order condition, and (ii) the labor supply of the two households move proportionately. This shuts down the wealth eects for hand-to-mouth households receiving transfers studied in Giambattista and Pennings (2017). 38I also abstract from capital, and so assume output is produced using only labor with constant returns (equivalent to a=1 in NS). The simple NK model is the limit of the full NK model when prices and wages become perfectly sticky (θw , θp → 1). 30 NK model is also similar to Galí and Monacelli (2005) but with sticky wages, hand-to-mouth households, and incomplete markets for Ricardian households. The NC model is the NK model with exible prices and wages (θw , θp → 0) and no hand-to-mouth households (ω = 0). As in NS, the economy is a monetary union consisting of a small home region (representing a small US state with population n) and a large foreign region with population 1 − n, each of which produces their own variety of imperfectly substitutable goods that can be consumed locally, consumed abroad or used for local government purchases. In the simple models, n → 0, so the small region becomes innitesimal. The complete model is presented in Appendix 4. While the model is relatively standard, the application to cross-region trans- fer multipliers (rather than purchase multipliers) is mostly new. The closest paper is Farhi and Werning (2016), who produce related expressions in their analysis of the nancing of the cross-region purchase multipliers, which overlap with mine when there are no hand-to-mouth households (ω = 0). 4.1. Analytical Cross-Region Transfer Impact Multipliers. Proposi- tion 1 presents cross-region transfer impact multipliers (denoted MT r ) in the simple rigid price/wage NK model and simple NC model. MT r is the increase in real output in the rst quarter in a small home region whose households receive a $1 lump-sum transfer from the federal government. The federal government here funds the $1 transfer by levying lump sum taxes on the residents of the rest of the monetary union. The transfer is untargeted, so hand-to-mouth agents receive a fraction ω of the transfer, equal to their pop- ulation share. As long as the transfer is nanced federally, further nancing details (decit nancing, distortionary taxation, etc.) do not aect M, as the 39 small home region is innitesimal in size. Here α is home bias in consump- tion, β (close to 1) is the household's quarterly discount factor, and ρ is the quarterly persistence of the transfer (which follows an AR(1) process). Proofs are available in Appendix 7. 39See Appendix 6.1 for how federally nanced cross-region multipliers compare to locally nanced multipliers. MT r can also be interpreted as the response of output in the small region relative to the rest of the monetary union (like in empirical Section 2). 31 Proposition 1. Cross-region transfer impact multipliers in the rigid-price/wage NK model and NC models are given by: α 1−β α 1−β (a) MN K Tr = ×1 × + × ω × [1 − ] 1−α 1 − βρ 1 − αω 1 − βρ Local GE (perm) M P C (perm) Local GE (temp) M P C (temp) 1 1−β (b)MN C Tr = − ×1 × + 0 1+ϕ 1 − βρ Local GE (perm) M P C (perm) Proposition 1 decomposes the cross-region transfer multipliers in each model into two terms relating to the permanent and temporary components of the transfer. The permanent component of a $1 cross-region transfer payment is given by the annuity value of the transfer $(1 − β )/(1 − βρ). The temporary component is the excess of the initial $1 payment over its annuity value: $(1 − (1 − β )/(1 − βρ)). As the transfer shock becomes perfectly persistent (ρ → 1), the permanent component equals unity and the temporary component is zero (and vice versa when ρ=0 for one-o payments). The cross-region transfer multipliers depend on the fraction of each component that is spent, as well as the local GE eects, which I discuss in turn. The permanent component is spent by all households (Ricardian and hand- to-mouth) according to the permanent income hypothesis, yielding a MPC of 1 × (1 − β )/(1 − βρ). This is the same in simple NK and NC models. The temporary component is spent only by hand-to-mouth households. These households receive a fraction ω of the temporary component of the transfer $(1 − (1 − β )/(1 − βρ)), yielding a MPC of 1 × ω [1 − (1 − β )/(1 − βρ)]. The micro literature on the MPC (e.g. JPS) focuses on the consumption response to one-o transfers (ρ = 0) and so only identies M P C (temp) ≈ ω , which is only one part of the overall transfer multiplier. In NC models, ω = 0, yielding M P C (temp) ≈ 0, and a cross-region transfer multiplier that is trivially close to zero for one-o shocks. Local GE eects amplify or dampen the MPC in the expressions in Propo- sition 1. In NK models, output is demand-determined in the short run, so local demand and output initially increase by the fraction α that is spent on locally produced goods (α is consumption home bias for an atomistic region, in the numerator in the expression for local GE eects). This increase is then 32 amplied by later-round eects. For the permanent component, an extra α 2 of local demand/income generates α local demand/income (and so forth), amplifying the initial eect by the traditional Keynesian multiplier 1/(1 − α). For the temporary component, the amplication mechanism is similar, but in each round the extra income is only spent by hand-to-mouth households, yielding a smaller increase in local demand of αω , (αω )2 , ... and hence a smaller traditional Keynesian multiplier of 1/(1 − αω ). In the NC model, local GE eects are negative. Any increase in consumption demand for local goods (from a transfer) will result in an increase in the relative price of the home good and will cause a shift in expenditure toward foreign goods until the extra demand is exhausted. Output then falls as households spend some of their extra income on leisure. With the preferences in my model, the size of wealth eects on labor supply depend on the size of the Frisch elasticity ϕ−1 . The larger the Frisch elasticity, the stronger are wealth eects and the more negative the local GE eects −1/(1 + ϕ). As such, in the NC model, cross-region transfer multipliers range from close to zero (for one-o transfers) to −1/(1 + ϕ) for permanent transfers. 4.2. Relation to Cross-Region Purchase Multipliers. Cross-region pur- chase multipliers estimated in the literature are very dierent from cross-region transfer multipliers, with purchase multipliers being larger in both the NK and NC models. In the simple NK model, federal purchases of local goods have multipliers exactly one unit greater than those of the cross-region transfer multipliers above (see Proposition 2a in Appendix 7). The reason for this is that for a purchase, the initial $1 of extra income is generated as payment for an extra unit of local output rather than as a windfall. In a very open small region (α ≈ 0), the cross-region transfer multiplier would be close to zero, but the purchase multiplier would be around one. Cross-region transfer multipliers are also much more sensitive than purchase multipliers (in proportional terms) to model parameters such as shock persistence ρ, consumption home bias α, and the hand-to-mouth household share ω . In the NC model, cross-region government purchase multipliers are always positive (see Proposition 2b in Appendix 7) because they increase demand for 33 local goods, and hence prices and wages, without any wealth eects on labor supply (I assume government purchases are not valued by households). Again, this is very dierent from the cross-region transfer multiplier, which is negative in NC models. 5. Quantitative Cross-Region Transfer Multipliers In this section, I investigate the extent to which standard NK or NC models can rationalize the multipliers estimated in the data. The full models used here are similar to the analytical models presented at the start of Section 4, but the full NK model has Calvo sticky prices/wages rather than rigid prices/wages. The population share of the home region is calibrated to represent a typical US state (n = 1/50 rather than n → 0 as in the analytical model). Other parameters are taken from NS, who present a very similar model except for parameters relating to two features not in their model: the fraction of hand- to-mouth household (ω = 1/3; based on evidence from Kaplan, Violante, and Weidner 2014) and sticky wages that adjust once a year, on average (θw = 0.75). See Appendix Table 11 for a full list of parameters. Before comparing model-based and empirical multipliers, it is important to ensure that these multipliers are calculated in the same way. As there are no ocial state-level price data, empirical estimates in Section 3 are produced 40 by deating nominal labor income or nominal GDP by national ination. I call these measured multipliers to distinguish them from the actual multi- pliers in Proposition 1 (actual multipliers are quantities, which are nominal values deated by local (state-level) producer prices). Measured and actual impact multipliers are very similar in NK models (due to sticky prices), but in NC models with exible prices and persistent shocks, demand shocks increase 40Using my own constructed proxy of quarterly state-level prices, I nd little robust evi- dence that the transfers studied above increased local ination. Regressions of my quarterly state ination proxy on transfers are always insignicant in the 2000s, and over 19521974 transfers are always negative and insignicant once I control for the state-specic eects of oil price shocks. See Appendix 3.2 for further details. NS nd that there is very little movement in the local CPI in response to local military purchases and (consequently) get similar results when they deate by either national or state annual CPIs. 34 local prices leading measured multipliers to substantially larger than actual 41 multipliers. 5.1. Comparing Measured Multipliers in the Models and in the Data. Figure 3, presented in Section 3.6, compares measured cumulative multipliers in the models and data over the rst few quarters following the transfer shock. In sum, I nd the NK model is more consistent with the empirical evidence than the NC model, which is why I use it in Section 6 for the policy impli- cations. Also note that cumulative multipliers over the rst few quarters are very similar to impact multipliers in canonical NK and NC models, suggesting that the impact multiplier is a good summary statistic of later dynamics. Figure 3, Panel A compares permanent transfer multipliers in the model and data. The empirical impact multiplier of 1.29 (black circles, same as in Table 1) are almost the same as the impact measured multiplier in the NK model of 1.21 (red squares). Despite missing the rise in point estimates in the nal quarter, cumulative NK multipliers are close to the center of the 95% CIs, even over longer horizons. In contrast, the measured cumulative multiplier in the NC model is 0.5 over all horizons, which is on the border of the 95% CI. Panels B(i) and B(ii) of Figure 3 compare temporary transfer multipliers for labor income and GDP dependent variables, respectively. For labor income, the cumulative measured multipliers in the data are close to 0.3 and are close to the multipliers in the NK model. For GDP, the multiplier of 0.3 in the NK model is a little too low, but is always towards the center of the 95% CIs. In contrast, the lack of a response to a temporary transfer shock in the NC model is outside the 95% condence interval for labor income and GDP multipliers. This is largely because the NC model lacks hand-to-mouth households, and as a result, temporary transfers are saved and have little eect on the regional economy. 41Measured GDP and labor income in the home region can be rewritten as Y ˆ ˆ ˆ meas. ≈ (Yh +Ph ) and ˆ L)meas. ≈ (w (W ˆh + P ˆh + Y ˆh ), ˆ ˆ as n ≈ 0, Yh = Lh , and there are no aggregate or foreign shocks (hats denote deviations from steady state, Pˆh is the producer price of the home good). ˆ ˆ In the NC model, Ymeas. = (W L)meas. , as w ˆh = 0. Lower values of the Armington elasticity θT generate larger movements in P ˆh in the NC model and hence generate potentially larger increases in measured labor income or GDP in response to a permanent transfer. However, the fact that I am calibrating to small regions within a monetary union, rather than large countries, suggests a higher value of θT is appropriate. 35 5.2. Cross-Region Transfer Multipliers (Quantities) in the Full Model. The dashed lines in Figure 3 report actual GDP multipliers (not measured mul- tipliers) in the NK and NC models for comparison. For temporary transfers, actual and measured multipliers are almost identical in both NK and NC mod- els (as prices and wages do not move much): about 0.3 in the NK model and zero in the NC model. For permanent transfers, actual GDP multipliers in the NK model (blue dashed lines, impact multiplier of 1.1) are marginally smaller than measured labor income multipliers as wages and prices increase slightly in response to a permanent transfer. However, in the NC model a 1% of GDP permanent transfer generates an actual multiplier of -0.5 (= −(1 + ϕ)−1 with −1 ϕ = 1), which is much lower than the measured multiplier of +0.5 because home GDP producer prices increase by 1%. This illustrates the importance of measuring empirical and theoretical multipliers in a consistent way. 6. Some Policy Implications In this section I rst discuss some implications of my empirical estimates for the ability of federal automatic stabilizers to dampen regional shocks in a monetary union, and then relate my ndings to the size of the aggregate transfer multiplier in a closed economy. The primary policy implication of my cross-region transfer estimates is that the automatic stabilizers built into the US federal scal system can only pro- 42 vide modest stabilization of regional output at business cycle frequencies. The ability of the US federal system to smooth regional shocks has become prominent of late, as Eurozone policy makers have looked across the Atlantic for an alternative scal structure following the deep recessions in a number Eurozone countries (see The Economist's quote in the introduction). The lit- erature, also cited in the introduction, suggests that when a US region enters a recession, the residents of that region receive extra federal social benets and 42Related, my modest cross-region transfer multiplier estimates (at business cycle frequen- cies) suggest that federally nanced (cross-region) purchase multipliers are, at most, only slightly larger than their locally nanced counterparts. This corroborates theoretical argu- ments in Nakamura and Steinsson (2014), Farhi and Werning (2016), and Chodorow-Reich (2019) and means that cross-region multipliers can be used as a rough lower bound for closed-economy ZLB estimates, as in Chodorow-Reich (2019). See Appendix 6.1 for further discussion on self-nanced versus cross-region multipliers. 36 pay fewer federal taxes, generating a cross-region transfer of around 30 cents 43 for every $1 fall in income (known as the normalized tax change, NTC). This is eectively a temporary cross-region transfer from the rest of the monetary union to the residents of the aected region, and so one can apply my esti- mates of a temporary cross-region transfer multiplier of around 1/3. As such, the automatic stabilizers from a federal tax-transfer system would only stabi- lize around 10% (= 0.3 × 1/3) of any fall in output in a short-lived regional recession. A more complete calculation adjusts for (i) the eect of smoothing on the size of the cross-region transfer generated and (ii) asymmetric regional recessions being more persistent than a one-o transfer, which raises the rele- vant multiplier to around 0.42 (calculated using the NK model with ρ = 0.935 in Appendix 6.2). Combined, these only increase the fraction smoothed to 11%, with simulated recessions in the NK model producing similar results (see Appendix 6.2). The fraction smoothed ranges from 6-18% with alternative 44 assumptions. My cross-region transfer multiplier estimates are also related to the size of the aggregate closed economy transfer multiplier, albeit indirectly. Aggregate and cross-region multipliers are very dierent in general, as monetary policy and tax responses are dierenced out in the cross-section but are relevant for the aggregate multiplier, and there are no demand leakages in a closed economy. Nonetheless, my empirical estimates can be used as an identied moment (Nakamura and Steinsson 2018) to distinguish between models, with the favored model being used to calculate aggregate closed-economy transfer multipliers. This is the approach taken by NS but for purchases rather than transfers. 43Auerbach and Feenberg (2000) use NBER taxsim (rather than regressions) to estimate a lower NTC for the US as a whole. For my claims, a higher NTC is a more conservative assumption. Automatic stabilizers can also change marginal tax rates, the eects of which are not considered here (see Auerbach and Feenberg 2000). 44The smoothing fraction is S = N T C × MT r /(1 + N T C × MT r ) with MT r = 0.42 and NTC = 0.3. With Auerbach and Feenberg's NTC = 0.25 (applied to adjusted gross income), the smoothing fraction falls to 10%, and with their NTC = 0.15 (applied to GDP), the smoothing fraction falls to 6%. On the other hand, with a hand-to-mouth share of 50% and home bias of 75%, the NK model produces cross-region transfer multipliers of 0.6 for temporary transfers and 1.66 for permanent transfers. At business cycle frequencies, the transfer multiplier is 0.73, which produces a smoothing fraction of 18% (with NTC = 0.3). 37 As mentioned in Section 5.1, my empirical estimates are best captured by a NK model with sticky wages and a share of hand-to-mouth households. In that NK model, the aggregate present-value transfer multiplier MT r AggP V at busi- ness cycle frequencies is simply proportional to the aggregate present-value purchase multiplier MAggP G V , scaled by the fraction of transfers targeted at AggP V V 45 hand-to-mouth households ωT such that MT r = ωT MAggP G . Consis- tent with the ndings of NS, this type of model produces aggregate purchase AggP V multipliers MG at business cycle frequencies that are small ( 1) when monetary policy leans against the wind but are large (> 1) when monetary policy is more accommodating of ination (using a constant real interest rate 46 here for simplicity). My empirical estimates of one-o cross-region transfer multipliers are consistent with a modest share of hand-to-mouth households ω , which also pins down the scaling factor for untargeted transfer multipliers (ωT = ω ). Hence my NK model produces large aggregate transfer multipli- AggP V ers (MT r > 1) at business-cycle frequencies when both monetary policy is accommodating and transfers are targeted at hand-to-mouth households (ωT ω ), modest multipliers MAggP Tr V ≈ 1/2 for untargeted transfers with AggP V constant real interest rates, and small transfer multipliers MT r < 1/2 oth- erwise (see Appendix 6.3 for details). 7. Conclusion In this paper, I investigate the size of the cross-region transfer multiplier in the US. This is novel, as the cross-region transfer multiplier is conceptually dierent from the MPC and cross-region purchase multipliers estimated in the literature, and is relevant, as transfers are countercyclical and are the largest component of federal spending. I nd that cross-region transfers signicantly 45Aggregate transfers boost demand if they are spent as well as possibly reducing the labor supply of households receiving the transfer (see Giambattista and Pennings 2017). In this paper, my simplifying assumptions for sticky wage setting abstract from the second channel, leaving only the eects on demand. 46Multipliers can be substantially larger when monetary policy is more accommodating, for example when the Zero Lower Bound binds for an extended period. Párraga Rodríguez (2018) estimates aggregate transfer multipliers using Romer and Romer's (2016) series of Social Security increases as an instrument. She nds a large transfer multiplier of around unity after a year, which might be rationalized by the high persistence of Social Security increases and a high MPC. 38 boost short-run growth in the states receiving them, with impact multipliers around 1/3 for one-o transfers as part of stimulus packages ($0.2$0.9, de- pending on the specication) and 1.5 for permanent Social Security transfers ($0.9$1.9, depending on the specication). The size of the multipliers can be roughly rationalized by a standard NK model with sticky wages and a fraction of hand-to-mouth households. My estimates of modest temporary cross-region transfer multipliers, in turn, imply that the automatic stabilizers built into the US federal tax-transfer system only have a limited ability to smooth output in regional downturns, perhaps less than in the popular perception. However, this is not to say that other federal policies, such as discretionary federal purchases in regions in recessions, or intergovernmental transfers, might not be more eective (see Chodorow-Reich 2018 for a survey). Moreover, transfers might help to smooth regional consumption, which would enhance welfare even if output smoothing were modest. References [1] Auerbach, Alan, J., and Daniel R. Feenberg. 2000. The Signicance of Federal Journal of Economic Perspectives, 14 (3): 37-56. Taxes as Automatic Stabilizers. [2] Bartik, Timothy J. 1991. Who Benets from State and Local Economic Development Policies? Kalamazoo, MI: W.E. Upjohn Institute for Employment Research. [3] Barro, Robert J., and Charles J. Redlick. 2011. Macroeconomic Eects from Government Purchases and Taxes. Quarterly Journal of Economics, 26 (1): 51102. [4] Bayoumi, Tatami, and Paul R. Masson. 1995. Fiscal ows in the United States and Canada: Lessons for monetary union in Europe. European Economic Review 39: 253274 [5] BEA 2009. Regional Quarterly Report (January 2009), Economic Stimulus Payments box https://apps.bea.gov/scb/pdf/2009/01%20January/0109_regqtrlyreport.pdf [6] Chodorow-Reich, Gabriel. 2019. "Geographic Cross-Sectional Fiscal Spending Mul- tipliers: What Have We Learned?" American Economic Journal: Economic Policy, 11 (2): 1-34. [7] Chodorow-Reich, Gabriel, Laura Feiveson, Zachary Liscow, and William Gui Woolston. 2012. "Does State Fiscal Relief during Recessions Increase Employ- ment? Evidence from the American Recovery and Reinvestment Act." American Eco- nomic Journal: Economic Policy, 4 (3): 118-45. 39 [8] Clemens, Jerey, and Stephen Miran. 2012. Fiscal Policy Multipliers on Subna- tional Government Spending. American Economic Journal: Economic Policy, 4 (2): 46-68. [9] Erceg Christopher J., Dale W. Henderson and Andrew T. Levin. 2000., Opti- mal Monetary Policy with Staggered Wage and Price Contracts. Journal of Monetary Economics 46 (2): 281313. [10] Filmer, Deon, Jed Friedman, Eeshani Kandpal, and Junko Onishi. 2018. Cash Transfers, Food Prices, and Nutrition Impacts on Nonbeneciary Children. World Bank Policy Research Working Paper 8377 [11] Farhi Emmanuel, and Iván Werning. 2016. Fiscal Multipliers: Liquidity Traps and Currency Unions, Handbook of Macroeconomics Vol 2: 24172492 [12] Farhi, Emmanuel, and Iván Werning. 2017. "Fiscal Unions." American Economic Review, 107 (12): 3788-3834 [13] Feyrer, James, and Bruce Sacerdote. 2013. How Much Would US Style Fiscal Integration Buer European Unemployment and Income Shocks? (A Comparative Empirical Analysis). American Economic Review, 103 (3): 125-28. [14] Galí, Jordi and Tommaso Monacelli. 2005. Monetary Policy and Exchange Rate Volatility in a Small Open Economy. Review of Economics and Statistics 72 (3): 707734 [15] Giambattista, Eric M. and Steven M. Pennings. 2017. When is the government transfer multiplier large? European Economic Review, 100: 525543 [16] Hsieh, Chang-Tai. 2003. Do Consumers React to Anticipated Income Changes? Evidence from the Alaska Permanent Fund. American Economic Review, 93 (1): 397- 405. [17] Hausman, Joshua K. 2016. "Fiscal Policy and Economic Recovery: The Case of the 1936 Veterans' Bonus." American Economic Review, 106 (4): 1100-1143. [18] International Monetary Fund. 2020. Fiscal Monitor (2020 April) Chapter 1: Poli- cies to Support People During the COVID-19 Pandemic. [19] Johnson, David S., Jonathan A. Parker, and Nicholas S. Souleles. 2006. "Household Expenditure and the Income Tax Rebates of 2001." American Economic Review, 96 (5): 1589-1610. [20] Jordà, Òscar. 2005. Estimation and Inference of Impulse Responses by Local Pro- jections. American Economic Review, 95 (1): 161-182. [21] Kaplan, Greg, and Giovanni L. Violante. 2014. A Model of the Consumption Response to Fiscal Stimulus Payments., Econometrica 82 (4): 11991239 [22] Kaplan, Greg, Giovanni L. Violante and Justin. Weidner. 2014. The Wealthy Hand-to-Mouth. Brookings Papers on Economic Activity, (Spring 2014): 77138 40 [23] Kraay, Aart. 2014. Government Spending Multipliers in Developing Countries: Ev- idence from Lending by Ocial Creditors. American Economic Journal: Macroeco- nomics, 6 (4): 170-208. [24] Kueng, Lorenz. 2018. Excess Sensitivity of High-Income Consumers The Quarterly Journal of Economics, 133(4): 16931751 [25] Miyamoto, Wataru, Thuy Lan Nguyen, and Dmitriy Sergeyev. 2018. Gov- ernment Spending Multipliers under the Zero Lower Bound: Evidence from Japan. American Economic Journal: Macroeconomics, 10 (3): 247-77. [26] Misra, Kanishka, and Paolo Surico. 2014. "Consumption, Income Changes, and Heterogeneity: Evidence from Two Fiscal Stimulus Programs." American Economic Journal: Macroeconomics, 6 (4): 84-106. [27] Nakamura, Emi, and Jón Steinsson. 2014. Fiscal Stimulus in a Monetary Union: Evidence from US Regions. American Economic Review, 104 (3): 753-92. [28] Nakamura, Emi, and Jón Steinsson. 2018. Identication in Macroeconomics. Journal of Economic Perspectives, 32 (3): 59-86. [29] Oh, Hyunseung and Ricardo Reis. 2012. Targeted transfers and the scal response to the great recession. Journal of Monetary Economics 59(S): S50S64. [30] Parker, Jonathan A., Nicholas S. Souleles, David S. Johnson, and Robert McClelland. 2013. Consumer Spending and the Economic Stimulus Payments of 2008. American Economic Review, 103 (6): 2530-53. [31] Párraga Rodríguez, Susana. 2018. The dynamic eects of public expenditure shocks in the United States, Journal of Macroeconomics 56: 340360 [32] Ramey, Valerie A. 2019. "Ten Years after the Financial Crisis: What Have We Learned from the Renaissance in Fiscal Research?" Journal of Economic Perspectives, 33 (2): 89-114. [33] Ramey, Valarie A. and Sarah Zubairy. 2018. Government spending multipliers in good times and in bad: Evidence from US historical data. Journal of Political Economy 126(2): 850901 [34] Romer, Christina D., and David H. Romer. 2010. "The Macroeconomic Eects of Tax Changes: Estimates Based on a New Measure of Fiscal Shocks. American Economic Review, 100 (3): 763-801. [35] Romer, Christina D., and David H. Romer. 2016. "Transfer Payments and the Macroeconomy: The Eects of Social Security Benet Increases, 1952-1991." American Economic Journal: Macroeconomics 8 (4): 142. [36] Sala-i-Martin, Xavier. and Jerey Sachs. 1991. Fiscal Federalism and Optimal Currency Areas: Evidence for Europe from the United States, NBER Working Paper No. 3855 [37] Wilcox, David. 1989. Social Security Benets, Consumption Expenditure, and the Life Cycle Hypothesis, Journal of Political Economy 97 (2): 288304. 41