《国际政治经济学》文献资料（Political Economy of Migration）Welfare and the Multifaceted Decision to Move

American Political Science Review Vol.99.No.1 February 2005 Welfare and the Multifaceted Decision to Move MICHAEL A.BAILEY Georgetown University hether poor single mothers move in response to welfare benefits has important implications for social policy in a federal system.Many scholars claim that welfare does not affect migration. These claims are not definitive,however,because the models underlying them rely on problem- atic assumptions and do not adequately control for nonwelfare determinants ofmigration.I address these shortcomings with an improved statistical model of individual-level migration.The results indicate that welfare does affect residential choice.Although the effects of welfare are much smaller than the effects of family ties,they are real and have the potential to cause nontrivial changes in welfare populations and welfare expenditures. re poor single mothers more likely to stay in 2000,519)or,more typically,that welfare-induced mi- or move to states with higher welfare benefits? gration is a"myth"(Allard and Danziger 2000;Schram This question has important implications on at and Soss 1999.83).However.we should be cautious least two levels.As a policy matter,the answer will about accepting this emerging conventional wisdom. enlighten us about the effects of welfare on society In a variety of ways,these studies fail to account for and will assist efforts to understand whether states the complexities of migration and consequently run systematically lower benefits in order to avoid becom- the risk of either obscuring the effect of welfare or, ing"welfare-magnets"(Peterson and Rom 1989;Rom, even worse,conflating the effect of welfare with the Peterson,and Scheve 1998:Volden 2002).This answer effects of other unmeasured factors.As I show be- will also help us anticipate how welfare policies may low,this is precisely what has happened,the case in evolve in other areas of the world where it is-or point being the link between welfare benefits granted is becoming-as easy to move across jurisdictional through the Aid to Families with Dependent Children boundaries as in the United States. (AFDC)program in the late 1980s and the residential As a theoretical matter,understanding the relation- choices of poor single mothers,the program's primary ship between welfare and migration can help us bet- constituency. ter understand how the increasing mobility of people, firms,and capital affects governmental capacities to provide welfare and other redistributive benefits.If THREE HOLES IN THE EXISTING generous government benefits prompt people who re- LITERATURE ceive them to flow in and the people who pay for them to flow out,the benefits will become increasingly diffi- Assessing whether welfare affects migration is no sim- cult to sustain.This is true not only for a federal system ple task,as attested by a vast,highly contested liter- such as the United States,but also for the international ature (for a review,see Brueckner 2000).To do this system in which political,social,and economic barriers convincingly,researchers must account for all the fac- to migration have fallen dramatically in recent years tors other than welfare that affect people's decisions to Finding that welfare-induced migration occurs in the move.Researchers have made progress in this respect, United States would enhance concern about govern- but problems persist.Three issues undermine the re- mental capacity for social services;finding no such be- cent wave of research that downplays or dismisses the havior,on the other hand,would make us less inclined effect of welfare on migration. to believe that migration constrains governments in First,many studies risk distorting the effect of wel- other,less likely contexts. fare by inadequately accounting for state attributes that Lately,most scholars researching the question have affect migration.These studies typically consist of sta- found very little or no support for the idea that wel- tistical analyses of either aggregated migration flows fare affects migration in the United States (Allard and or individual migration choices.They control for state- Danziger 2000;Levine and Zimmerman 1999;Schram, level influences on residential choice through variables Nitz,and Krueger 1998;Schram and Soss 1999).They measuring such attributes as state economic perfor- conclude that the evidence is"at best mildly in favor' mance and differences in state climates (e.g.,Allard of the idea that welfare affects migration (Brueckner and Danziger 2000:Frey et al.1996:Schram.Nitz.and Krueger 1998). This seemingly straightforward enterprise is actu- Michael A.Bailey is Associate Professor,ICC.Suite 681,Depart- ally remarkably difficult.Consider,for example,the ment of Government,Georgetown University,Washington,DC variables Schram,Nitz,and Krueger use to charac- 20057(baileyma@georgetown.edu). terize nonwelfare components of state attractiveness. I appreciate helpful comments from Scott Allard,Arik Levinson. Florida-the quintessential high-population growth Bob Lieber,Forrest Maltzman,Dan Nexon,Mark Rom,George state-averaged 5.5%unemployment from 1985 to Shambaugh,Joe Soss,Michele Swers,and anonymous reviewers.I also wish to thank Craig Volden for sharing data.In addition,I grate- 1990;its median income averaged $34,931 in nominal fully acknowledge the support of the Hoover Institution National terms and $29.090 in state cost-of-living adjusted terms. Fellowship Program,Stanford University. Many states that were less attractive to potential 125

American Political Science Review Vol. 99, No. 1 February 2005 Welfare and the Multifaceted Decision to Move MICHAEL A. BAILEY Georgetown University Whether poor single mothers move in response to welfare benefits has important implications for social policy in a federal system. Many scholars claim that welfare does not affect migration. These claims are not definitive, however, because the models underlying them rely on problem￾atic assumptions and do not adequately control for nonwelfare determinants of migration. I address these shortcomings with an improved statistical model of individual-level migration. The results indicate that welfare does affect residential choice. Although the effects of welfare are much smaller than the effects of family ties, they are real and have the potential to cause nontrivial changes in welfare populations and welfare expenditures. Are poor single mothers more likely to stay in or move to states with higher welfare benefits? This question has important implications on at least two levels. As a policy matter, the answer will enlighten us about the effects of welfare on society and will assist efforts to understand whether states systematically lower benefits in order to avoid becom￾ing “welfare-magnets” (Peterson and Rom 1989; Rom, Peterson, and Scheve 1998; Volden 2002). This answer will also help us anticipate how welfare policies may evolve in other areas of the world where it is—–or is becoming—–as easy to move across jurisdictional boundaries as in the United States. As a theoretical matter, understanding the relation￾ship between welfare and migration can help us bet￾ter understand how the increasing mobility of people, firms, and capital affects governmental capacities to provide welfare and other redistributive benefits. If generous government benefits prompt people who re￾ceive them to flow in and the people who pay for them to flow out, the benefits will become increasingly diffi- cult to sustain. This is true not only for a federal system such as the United States, but also for the international system in which political, social, and economic barriers to migration have fallen dramatically in recent years. Finding that welfare-induced migration occurs in the United States would enhance concern about govern￾mental capacity for social services; finding no such be￾havior, on the other hand, would make us less inclined to believe that migration constrains governments in other, less likely contexts. Lately, most scholars researching the question have found very little or no support for the idea that wel￾fare affects migration in the United States (Allard and Danziger 2000; Levine and Zimmerman 1999; Schram, Nitz, and Krueger 1998; Schram and Soss 1999). They conclude that the evidence is “at best mildly in favor” of the idea that welfare affects migration (Brueckner Michael A. Bailey is Associate Professor, ICC, Suite 681, Depart￾ment of Government, Georgetown University, Washington, DC 20057 (baileyma@georgetown.edu). I appreciate helpful comments from Scott Allard, Arik Levinson, Bob Lieber, Forrest Maltzman, Dan Nexon, Mark Rom, George Shambaugh, Joe Soss, Michele Swers, and anonymous reviewers. I also wish to thank Craig Volden for sharing data. In addition, I grate￾fully acknowledge the support of the Hoover Institution National Fellowship Program, Stanford University. 2000, 519) or, more typically, that welfare-induced mi￾gration is a “myth” (Allard and Danziger 2000; Schram and Soss 1999, 83). However, we should be cautious about accepting this emerging conventional wisdom. In a variety of ways, these studies fail to account for the complexities of migration and consequently run the risk of either obscuring the effect of welfare or, even worse, conflating the effect of welfare with the effects of other unmeasured factors. As I show be￾low, this is precisely what has happened, the case in point being the link between welfare benefits granted through the Aid to Families with Dependent Children (AFDC) program in the late 1980s and the residential choices of poor single mothers, the program’s primary constituency. THREE HOLES IN THE EXISTING LITERATURE Assessing whether welfare affects migration is no sim￾ple task, as attested by a vast, highly contested liter￾ature (for a review, see Brueckner 2000). To do this convincingly, researchers must account for all the fac￾tors other than welfare that affect people’s decisions to move. Researchers have made progress in this respect, but problems persist. Three issues undermine the re￾cent wave of research that downplays or dismisses the effect of welfare on migration. First, many studies risk distorting the effect of wel￾fare by inadequately accounting for state attributes that affect migration. These studies typically consist of sta￾tistical analyses of either aggregated migration flows or individual migration choices. They control for state￾level influences on residential choice through variables measuring such attributes as state economic perfor￾mance and differences in state climates (e.g., Allard and Danziger 2000; Frey et al. 1996; Schram, Nitz, and Krueger 1998). This seemingly straightforward enterprise is actu￾ally remarkably difficult. Consider, for example, the variables Schram, Nitz, and Krueger use to charac￾terize nonwelfare components of state attractiveness. Florida—–the quintessential high–population growth state—–averaged 5.5% unemployment from 1985 to 1990; its median income averaged $34,931 in nominal terms and $29,090 in state cost-of-living adjusted terms. Many states that were less attractive to potential 125

Welfare and the Multifaceted Decision to Move February 2005 in-migrants looked similar or better in these terms: The omission of race in many studies raises simi- Rhode Island averaged 3.9%unemployment and lar concerns.Individuals are more likely to move to $38,492 in nominal median income.South Dakota states with larger numbers of racially similar people averaged 4.4%unemployment and $30,460 in cost- (Frey et al.1996).In the data discussed below,there adjusted median income.Of course.one could add are about 60,000 poor white single mothers and about variables(e.g,“average temperature,.”“murder rate") 40,000 poor black single mothers.Of the whites,510 and all manner of nonlinearities and interactions (e.g., lived in North Dakota,South Dakota,or Vermont;two “temperature squared,”“temperature x income"). of the black single mothers lived in those states.If race- Nonetheless,one cannot help but suspect that signifi- specific attraction to states correlates with welfare (as is cant aspects of state attractiveness resist measurement likely if,for example.African Americans are relatively The danger is that studies with inadequate state-level more attracted to low-benefit southern states that have controls will conflate the effect of welfare on migration relatively large African American populations),failure with other factors.Recent demographic trends make to account for such variables may introduce yet another this a particular concern.Americans tend to move source of omitted variable bias that can distort the es- from northern("rust belt")states with relatively high timated effects of welfare on migration. welfare benefits to southern ("sun belt")states with Third,many studies aggregate away important state- relatively low welfare benefits.Failing to account for level differences.Levine and Zimmerman estimate a the complicated mixture of economic and social fac- model in which the dependent variable is whether an tors behind such moves results in analyses in which individual moved out of state.By ignoring whether the states where welfare is high are also the states the person left for a high-benefit state such as where the unmeasured attractiveness of living is low, California or a low-benefit state such as Louisiana, and vice versa.The statistical result is that unmea- this approach limits the ability of the method to as- sured disincentives to migrate to a state get lumped certain the role of welfare.Meyer (2000)estimates in with the observed (and correlated)welfare mea- a model in which migration across regions is the de- sures,leading to estimates in which the effect of wel- pendent variable,thereby treating states as identi- fare appears to be small or inconsequential,even if it cal within regions.Depending on the specification, is not. Meyer assumes that there are two or nine regions in Second,existing research fails to account adequately the entire United States,implying,for example,that for individual-level factors that influence migration. Maryland is identical to West Virginia and that New Many individuals want to move "home"to the state in Hampshire is identical to New York.This assump- which they were born because,that is,where they are tion of intraregional homogeneity creates a chronic more likely to have family and to know the neighbor- error in variables problem that likely will obscure hoods,schools,and industries.Moving home may have relationships between variables such as welfare and a particularly powerful appeal for single mothers,who migration. often depend on the housing,childcare,financial assis- Each of these problems potentially obscures or dis- tance,and psychological support of parents,siblings, torts the estimated effect of welfare on migration.Ev- and friends (Allard and Danziger 2000,358:Schram. ery recent study that dismisses welfare effects suffers Nitz,and Krueger 1998;Vartanian et al.1999).In what from more than one of these problems,meaning that follows,I refer to the attractions of home as "family the true effect of welfare is buried under multiple layers ties";some scholars refer to them as"social capital." of specification error.To get a better sense of the true The data described below bear out these expecta- relationship between welfare and migration,I develop tions.Home is not just another variable:it is a funda- an analytical approach that directly addresses each of mental influence on migration.Fully one-third of all these issues interstate moves by poor single mothers were to the individuals'birth states.For many states,the propor- tion of in-migrants who were born there is extremely A MORE COMPREHENSIVE APPROACH high:54%of poor single mothers moving into Alabama At the heart of the analysis is a random utility model of from out of state had been born in Alabama.The individual-level migration choices.The model charac- comparable numbers were 57%for Louisiana.58% terizes the utility for every individual of living in every for Mississippi,and 51%for West Virginia.(At the single state.Specifically,the utility of living in state s other extreme,only 12%of poor single mothers mov- for person i currently in state j consists of a deter- ing into Florida or Nevada were returning to their state ministic component viis and a stochastic component of birth. E订s Failure to account for the special attractiveness of birth states can lead researchers to understate or even Us=vs+es· (1) reverse the true effect of welfare on migration.The reason is that single mothers were born disproportion- I estimate the model with a conditional logit setup ately in poor,low-benefit states.If we fail to control (Greene 2000.858).In the model.each individual se- for the attraction of home states for these women,we lects the state that offers the highest utility.Assuming may mistake their fairly common moves home with a that the random shocks are independently and identi- complete disregard for the low welfare benefits in their cally distributed Extreme Value Type I random vari- states of birth. ables,the probability that person i living in state j 126

Welfare and the Multifaceted Decision to Move February 2005 in-migrants looked similar or better in these terms: Rhode Island averaged 3.9% unemployment and $38,492 in nominal median income. South Dakota averaged 4.4% unemployment and $30,460 in cost￾adjusted median income. Of course, one could add variables (e.g., “average temperature,” “murder rate”) and all manner of nonlinearities and interactions (e.g., “temperature squared,” “temperature × income”). Nonetheless, one cannot help but suspect that signifi- cant aspects of state attractiveness resist measurement. The danger is that studies with inadequate state-level controls will conflate the effect of welfare on migration with other factors. Recent demographic trends make this a particular concern. Americans tend to move from northern (“rust belt”) states with relatively high welfare benefits to southern (“sun belt”) states with relatively low welfare benefits. Failing to account for the complicated mixture of economic and social fac￾tors behind such moves results in analyses in which the states where welfare is high are also the states where the unmeasured attractiveness of living is low, and vice versa. The statistical result is that unmea￾sured disincentives to migrate to a state get lumped in with the observed (and correlated) welfare mea￾sures, leading to estimates in which the effect of wel￾fare appears to be small or inconsequential, even if it is not. Second, existing research fails to account adequately for individual-level factors that influence migration. Many individuals want to move “home” to the state in which they were born because, that is, where they are more likely to have family and to know the neighbor￾hoods, schools, and industries. Moving home may have a particularly powerful appeal for single mothers, who often depend on the housing, childcare, financial assis￾tance, and psychological support of parents, siblings, and friends (Allard and Danziger 2000, 358; Schram, Nitz, and Krueger 1998; Vartanian et al. 1999). In what follows, I refer to the attractions of home as “family ties”; some scholars refer to them as “social capital.” The data described below bear out these expecta￾tions. Home is not just another variable; it is a funda￾mental influence on migration. Fully one-third of all interstate moves by poor single mothers were to the individuals’ birth states. For many states, the propor￾tion of in-migrants who were born there is extremely high: 54% of poor single mothers moving into Alabama from out of state had been born in Alabama. The comparable numbers were 57% for Louisiana, 58% for Mississippi, and 51% for West Virginia. (At the other extreme, only 12% of poor single mothers mov￾ing into Florida or Nevada were returning to their state of birth.) Failure to account for the special attractiveness of birth states can lead researchers to understate or even reverse the true effect of welfare on migration. The reason is that single mothers were born disproportion￾ately in poor, low-benefit states. If we fail to control for the attraction of home states for these women, we may mistake their fairly common moves home with a complete disregard for the low welfare benefits in their states of birth. The omission of race in many studies raises simi￾lar concerns. Individuals are more likely to move to states with larger numbers of racially similar people (Frey et al. 1996). In the data discussed below, there are about 60,000 poor white single mothers and about 40,000 poor black single mothers. Of the whites, 510 lived in North Dakota, South Dakota, or Vermont; two of the black single mothers lived in those states. If race￾specific attraction to states correlates with welfare (as is likely if, for example, African Americans are relatively more attracted to low-benefit southern states that have relatively large African American populations), failure to account for such variables may introduce yet another source of omitted variable bias that can distort the es￾timated effects of welfare on migration. Third, many studies aggregate away important state￾level differences. Levine and Zimmerman estimate a model in which the dependent variable is whether an individual moved out of state. By ignoring whether the person left for a high-benefit state such as California or a low-benefit state such as Louisiana, this approach limits the ability of the method to as￾certain the role of welfare. Meyer (2000) estimates a model in which migration across regions is the de￾pendent variable, thereby treating states as identi￾cal within regions. Depending on the specification, Meyer assumes that there are two or nine regions in the entire United States, implying, for example, that Maryland is identical to West Virginia and that New Hampshire is identical to New York. This assump￾tion of intraregional homogeneity creates a chronic error in variables problem that likely will obscure relationships between variables such as welfare and migration. Each of these problems potentially obscures or dis￾torts the estimated effect of welfare on migration. Ev￾ery recent study that dismisses welfare effects suffers from more than one of these problems, meaning that the true effect of welfare is buried under multiple layers of specification error. To get a better sense of the true relationship between welfare and migration, I develop an analytical approach that directly addresses each of these issues. A MORE COMPREHENSIVE APPROACH At the heart of the analysis is a random utility model of individual-level migration choices. The model charac￾terizes the utility for every individual of living in every single state. Specifically, the utility of living in state s for person i currently in state j consists of a deter￾ministic component vijs and a stochastic component
ijs : Uijs = vijs +
ijs . (1) I estimate the model with a conditional logit setup (Greene 2000, 858). In the model, each individual se￾lects the state that offers the highest utility. Assuming that the random shocks are independently and identi￾cally distributed Extreme Value Type I random vari￾ables, the probability that person i living in state j 126

American Political Science Review Vol.99.No.1 chooses state s is ner that state unemployment and state climate were above).The state fixed effect will "soak up"the wel- Ps=Prob(Us>UkHk≠s), (2) fare effect and leave it statistically unidentified.I avoid =Prob(∈k-∈s<s-vHk卡S), 3) this problem by using a quasi-experimental research design,sometimes referred to as a comparison group e method (Levine and Zimmerman 1999;Meyer 2000). (4) ∑et This design requires that I include in the sample a "control group"that is not eligible for welfare but oth- where K is the total number of states to which an erwise resembles the"treatment group"of poor single individual can move.The computationally convenient mothers.General state attributes (captured by state form makes estimation conceptually straightforward fixed effects)influence individuals in the control and (even as it is practically difficult,given that a very treatment groups;welfare,however,influences only in- large number of individuals are choosing among a large dividuals in the treatment group.Given the inclusion number of discrete choices).By explicitly modeling all of the control group,the welfare variable is no longer a the state choices,I reduce the possibility that errors in constant for all individuals for any given state (that is, variables obscure the effect of welfare on migration. welfare benefits are zero for individuals in the control I control for state attributes by using state-level fixed group and the measured value for individuals in the effects,i.e.,by using state-level dummy variables to treatment group).The welfare variable now is statisti- control for all state attributes that are the same for all cally identified:it allows us to see whether differences individuals in the analysis.For example,these variables in welfare benefits explain any differences in behavior control for state unemployment and state climate be- by the treatment and control groups. cause for any given state,the values of these variables I also control for,among other factors,the gravi- will be the same when modeling the probability that tational pull of birth states and potential differences any individual will move to the state.2(A variable not in the attraction whites and African Americans have encompassed by fixed effects varies for a given state toward states.Including these variables not only serves across individuals;for example,only some people were important statistical control purposes,but also human- born in New York,meaning that when modeling the izes the analysis by moving beyond the caricature of utility of New York,the state-of-birth variable would welfare recipients as solely motivated by financial gain be one for some individuals and zero for others.)The (see the excellent discussion on this point in Schram, real advantage of fixed effects comes from their ability Nitz.and Krueger 1998). to subsume unmeasured variables and unspecified in- teractions.That is,fixed effects control for any attribute of a state-measurable or not-that affects all individ- DATA uals in the same way.Thus the fixed-effect approach Individual-level data are from the Census Bureau's controls for state-level factors at least as well as-and Public Use Microdata Series (PUMS)1990 5% usually better than-any approach relying on state- sample as accessed via Integrated PUMS (IPUMS) level covariates. (Ruggles et al.1997).This data set provides individual When using fixed effects,one must make special ef- information on age,marital status,number and ages forts to distinguish the effect of welfare on migration- of children,income,race,education,birth state,and if any exists-from the more general attractiveness of state of residence in 1985 and 1990.I work with the states measured by fixed effects.If the sample includes 1990 data for two reasons.First,most studies of welfare only poor single mothers who are all eligible for Aid migration assess AFDCin the late 1980s or early 1990s. to Families with Dependent Children("AFDC"),then I use data from that period in order to ensure that it the welfare generosity of each state will be the same is the methods-and not changes in reality-behind for all individuals in the sample (in the same man- any new results.Second,the highly variable welfare environment from 1996 to 2000 makes it hard to draw inferences about migration based on average levels of 1 The model automatically satisfies the"independence of irrelevant benefits over that time period.In contrast,AFDC was alternatives"(IIA)condition.This condition implies that the ratio of quite stable from 1985 to 1990. probabilities of choosing one option to another is the same,whether or not a third option is included in the choice set.In an appendix The welfare population consists of 110,243 single available upon request,I discuss alternative estimation strategies mothers between 25 and 53 with children between 4 and present results that indicate that the results are very similar in and 18 who had an income less than 125%of the models that do not satisfy the IIA condition. poverty level.3 Of these,8.9%moved across state 2 To see this,first suppose that the utility of a state depends only on a single variable(say"unemployment rate")and that the coefficient on this variable is negative one.For every individual,the utility of living in any given state would be negative one times the unemploymen 3 The earlier literature sometimes focuses on individuals who ac. rate for the state.A state-level fixed effect completely captures this tually receive welfare.Meyer(2000,5)details how doing so biases amount.If we add another state-level variable with a coefficient of the results in favor of the welfare migration hypothesis.For exam- two,say,the utility for all individuals of living in the state would be ple,some of the people who would not receive welfare in a low- negative one times the unemployment rate plus two times the value benefit state could move to a higher-benefit state and receive welfare of the new variable.Again,a state-level fixed effect would capture simply because eligibility is easier in the higher-benefit state.This the utility value of a state.This reasoning directly extends to any dynamic will exaggerate the flow of welfare recipients into high- number of state-level variables. benefit states and nonrecipients out of low-benefit states.This paper 127

American Political Science Review Vol. 99, No. 1 chooses state s is Pijs = Prob(Uijs > Uijk ∀ k = s), (2) = Prob(
ijk −
ijs < vijs − vijk∀ k = s), (3) = evijs
K k evijk , (4) where K is the total number of states to which an individual can move. The computationally convenient form makes estimation conceptually straightforward (even as it is practically difficult, given that a very large number of individuals are choosing among a large number of discrete choices).1 By explicitly modeling all the state choices, I reduce the possibility that errors in variables obscure the effect of welfare on migration. I control for state attributes by using state-level fixed effects, i.e., by using state-level dummy variables to control for all state attributes that are the same for all individuals in the analysis. For example, these variables control for state unemployment and state climate be￾cause for any given state, the values of these variables will be the same when modeling the probability that any individual will move to the state.2 (A variable not encompassed by fixed effects varies for a given state across individuals; for example, only some people were born in New York, meaning that when modeling the utility of New York, the state-of-birth variable would be one for some individuals and zero for others.) The real advantage of fixed effects comes from their ability to subsume unmeasured variables and unspecified in￾teractions. That is, fixed effects control for any attribute of a state—–measurable or not—–that affects all individ￾uals in the same way. Thus the fixed-effect approach controls for state-level factors at least as well as—–and usually better than—–any approach relying on state￾level covariates. When using fixed effects, one must make special ef￾forts to distinguish the effect of welfare on migration—– if any exists—–from the more general attractiveness of states measured by fixed effects. If the sample includes only poor single mothers who are all eligible for Aid to Families with Dependent Children (“AFDC”), then the welfare generosity of each state will be the same for all individuals in the sample (in the same man- 1 The model automatically satisfies the “independence of irrelevant alternatives” (IIA) condition. This condition implies that the ratio of probabilities of choosing one option to another is the same, whether or not a third option is included in the choice set. In an appendix available upon request, I discuss alternative estimation strategies and present results that indicate that the results are very similar in models that do not satisfy the IIA condition. 2 To see this, first suppose that the utility of a state depends only on a single variable (say “unemployment rate”) and that the coefficient on this variable is negative one. For every individual, the utility of living in any given state would be negative one times the unemployment rate for the state. A state-level fixed effect completely captures this amount. If we add another state-level variable with a coefficient of two, say, the utility for all individuals of living in the state would be negative one times the unemployment rate plus two times the value of the new variable. Again, a state-level fixed effect would capture the utility value of a state. This reasoning directly extends to any number of state-level variables. ner that state unemployment and state climate were above). The state fixed effect will “soak up” the wel￾fare effect and leave it statistically unidentified. I avoid this problem by using a quasi-experimental research design, sometimes referred to as a comparison group method (Levine and Zimmerman 1999; Meyer 2000). This design requires that I include in the sample a “control group” that is not eligible for welfare but oth￾erwise resembles the “treatment group” of poor single mothers. General state attributes (captured by state fixed effects) influence individuals in the control and treatment groups; welfare, however, influences only in￾dividuals in the treatment group. Given the inclusion of the control group, the welfare variable is no longer a constant for all individuals for any given state (that is, welfare benefits are zero for individuals in the control group and the measured value for individuals in the treatment group). The welfare variable now is statisti￾cally identified; it allows us to see whether differences in welfare benefits explain any differences in behavior by the treatment and control groups. I also control for, among other factors, the gravi￾tational pull of birth states and potential differences in the attraction whites and African Americans have toward states. Including these variables not only serves important statistical control purposes, but also human￾izes the analysis by moving beyond the caricature of welfare recipients as solely motivated by financial gain (see the excellent discussion on this point in Schram, Nitz, and Krueger 1998). DATA Individual-level data are from the Census Bureau’s Public Use Microdata Series (PUMS) 1990 5% sample as accessed via Integrated PUMS (IPUMS) (Ruggles et al. 1997). This data set provides individual information on age, marital status, number and ages of children, income, race, education, birth state, and state of residence in 1985 and 1990. I work with the 1990 data for two reasons. First, most studies of welfare migration assess AFDC in the late 1980s or early 1990s. I use data from that period in order to ensure that it is the methods—–and not changes in reality—–behind any new results. Second, the highly variable welfare environment from 1996 to 2000 makes it hard to draw inferences about migration based on average levels of benefits over that time period. In contrast, AFDC was quite stable from 1985 to 1990. The welfare population consists of 110,243 single mothers between 25 and 53 with children between 4 and 18 who had an income less than 125% of the poverty level.3 Of these, 8.9% moved across state 3 The earlier literature sometimes focuses on individuals who ac￾tually receive welfare. Meyer (2000, 5) details how doing so biases the results in favor of the welfare migration hypothesis. For exam￾ple, some of the people who would not receive welfare in a low￾benefit state could move to a higher-benefit state and receive welfare simply because eligibility is easier in the higher-benefit state. This dynamic will exaggerate the flow of welfare recipients into high￾benefit states and nonrecipients out of low-benefit states. This paper 127

Welfare and the Multifaceted Decision to Move February 2005 lines between 1985 and 1990.The nonwelfare control averaged across 1985-90.I adjust for cost-of-living dif- groups reasonably match the welfare population in ferences using Meyer's (2000,14)state price index all respects except for eligibility for welfare.Follow- (which focuses on variation in housing costs)and the ing Meyer(2000)and Levine and Zimmerman (1999), national consumer price index. I use three different control groups.The first con- I control for moving costs with several variables.I sists of 69,270 childless single women between 25 and measure the "fixed cost"of moving across state lines 53 years of age who had less than three times the with a dummy variable called move that takes on a poverty income and no college degree.4 The second value of one if js.I measure the "variable cost"of control group consists of 96,684 childless single males moving,which depends on the distance of the move, who had less than three times the poverty income and with a variable called distance.which is the log of dis- no college degree.The third control group consists of tance between state s and state j.Interaction terms 122.681 married women with children with household allow for the possibility that the effect of moving costs incomes greater than three times the poverty level differs between the welfare and the nonwelfare popu- and below the lesser of five times the poverty level or lations. $50,000.No group perfectly matches the welfare pop- I also control for the possibility that the welfare and ulation,but all match in some way the skill profiles and nonwelfare populations respond differently to state economic circumstances of poor single mothers.Using characteristics.For example,individuals in the welfare multiple specifications should increase confidence in population may care less about wages and unemploy- the robustness of the results. ment if they are expecting to rely on government or The focal variable is welfare benefits measured as family assistance.Therefore I include interactions of the sum of maximum AFDC benefits for a family of state-level wage and unemployment variables with an four and Food Stamps for each state.The Food Stamp individual-specific indicator variable for individuals in data are from the U.S.House Committee on Ways and the welfare population.Although the general effects of Means(various years).I restrict the welfare effect to wages and unemployment are not identified (because be zero for the control group by multiplying welfare they are soaked up by the fixed effects),I can estimate benefits times a dummy variable indicating whether the differential effect of these variables on the con- an individual is in the welfare population.This creates trol and treatment populations with these interaction within-state individual-level variation in the welfare terms. variable and allows it to be included in a model with state-level fixed effects.This is the critical variable for the welfare migration hypothesis. RESULTS State wage data are the average retail wages for food stores from the Census Bureau (2000):data from The analysis proceeds in two steps.First,I replicate this sector of the economy reflect the earnings poten- and extend the analysis by Schram,Nitz,and Krueger tial of low-skill women (Berry,Fording,and Hanson (1998)to make two points:(1)that models with no 2003).State unemployment data are from the Bureau or few nonwelfare controls show no welfare effects of Labor Statistics(2001).All state-level variables are and (2)that better accounting for state-level and nonwelfare determinants of migration produces initial evidence of a welfare effect.I then present results for follows Meyer's recommendation(8)of using an "at-risk group(sin the more flexible and powerful random utility model gle mothers or.better yet,low-educated single mothers)."He also of migration. notes that "a substantial fraction of any at-risk group may not be likely welfare recipients,and thus effects on the overall group are likely to be watered down estimates of the effects on likely partic- Revisiting Schram,Nitz,and Krueger ipants."Given the findings of this paper,it is reassuring that the welfare population is identified in a manner that biases against the welfare migration rather than in favor of it.Following the convention Schram,Nitz,and Krueger (1998)model migration of this literature,I include only individuals who started and ended patterns of poor single mothers as a function of wel- up in the continental United States;including Alaska and Hawaii fare,income,and employment differentials.(Allard produces essentially the same results.The limits on children's ages and Danziger (2000,361)provide,among other anal- limit the sample to only those women who had children during the yses,a similar analysis with no controls.)In two of entire period from 1985 to 1990;earlier versions of this paper allowed for younger children and had similar results.The poverty level varies four specifications,Schram,Nitz,and Krueger find a based on number of children in the family and other factors;the significant negative relationship between welfare ben- average poverty threshold in 1989 was S12,674 (IPUMS codebook efits and migration.This odd result suggest either that Ruggles et al.1997],225). high welfare benefits repel poor single mothers(which For all control groups I exclude individuals who have served in the seems unlikely and would constitute a major paradigm military in the last five years,as their mobility may be very different from that of civilians.Also,I exclude disabled individuals from the shift if true)or that nonwelfare factors correlated with control groups,as they may be more eligible for,or more interested welfare benefits have been omitted and are causing a in,welfare than others in the group. spurious negative relationship. This is the standard measure of welfare generosity in the literature. To investigate whether omitted variable bias is the Other aspects of welfare generosity such as eligibility standards are problem,Table 1 revisits Schram,Nitz,and Krueger's correlated,but distinct.See Bailey and Rom 2004 for further discus- sion of the multiple dimensions of welfare generosity.Estimating the model.The dependent variable is Census Bureau data model using a measure of spending per poor person-a measure that on the proportion of poor,single women with chil- taps eligibility as well-produces similar results. dren moving from one state to another between 1985 128

Welfare and the Multifaceted Decision to Move February 2005 lines between 1985 and 1990. The nonwelfare control groups reasonably match the welfare population in all respects except for eligibility for welfare. Follow￾ing Meyer (2000) and Levine and Zimmerman (1999), I use three different control groups. The first con￾sists of 69,270 childless single women between 25 and 53 years of age who had less than three times the poverty income and no college degree.4 The second control group consists of 96,684 childless single males who had less than three times the poverty income and no college degree. The third control group consists of 122,681 married women with children with household incomes greater than three times the poverty level and below the lesser of five times the poverty level or $50,000. No group perfectly matches the welfare pop￾ulation, but all match in some way the skill profiles and economic circumstances of poor single mothers. Using multiple specifications should increase confidence in the robustness of the results. The focal variable is welfare benefits measured as the sum of maximum AFDC benefits for a family of four and Food Stamps for each state.5 The Food Stamp data are from the U.S. House Committee on Ways and Means (various years). I restrict the welfare effect to be zero for the control group by multiplying welfare benefits times a dummy variable indicating whether an individual is in the welfare population. This creates within-state individual-level variation in the welfare variable and allows it to be included in a model with state-level fixed effects. This is the critical variable for the welfare migration hypothesis. State wage data are the average retail wages for food stores from the Census Bureau (2000); data from this sector of the economy reflect the earnings poten￾tial of low-skill women (Berry, Fording, and Hanson 2003). State unemployment data are from the Bureau of Labor Statistics (2001). All state-level variables are follows Meyer’s recommendation (8) of using an “at-risk group (sin￾gle mothers or, better yet, low-educated single mothers).” He also notes that “a substantial fraction of any at-risk group may not be likely welfare recipients, and thus effects on the overall group are likely to be watered down estimates of the effects on likely partic￾ipants.” Given the findings of this paper, it is reassuring that the welfare population is identified in a manner that biases against the welfare migration rather than in favor of it. Following the convention of this literature, I include only individuals who started and ended up in the continental United States; including Alaska and Hawaii produces essentially the same results. The limits on children’s ages limit the sample to only those women who had children during the entire period from 1985 to 1990; earlier versions of this paper allowed for younger children and had similar results. The poverty level varies based on number of children in the family and other factors; the average poverty threshold in 1989 was $12,674 (IPUMS codebook [Ruggles et al. 1997], 225). 4 For all control groups I exclude individuals who have served in the military in the last five years, as their mobility may be very different from that of civilians. Also, I exclude disabled individuals from the control groups, as they may be more eligible for, or more interested in, welfare than others in the group. 5 This is the standard measure of welfare generosity in the literature. Other aspects of welfare generosity such as eligibility standards are correlated, but distinct. See Bailey and Rom 2004 for further discus￾sion of the multiple dimensions of welfare generosity. Estimating the model using a measure of spending per poor person—–a measure that taps eligibility as well—–produces similar results. averaged across 1985–90. I adjust for cost-of-living dif￾ferences using Meyer’s (2000, 14) state price index (which focuses on variation in housing costs) and the national consumer price index. I control for moving costs with several variables. I measure the “fixed cost” of moving across state lines with a dummy variable called move that takes on a value of one if j = s. I measure the “variable cost” of moving, which depends on the distance of the move, with a variable called distance, which is the log of dis￾tance between state s and state j . Interaction terms allow for the possibility that the effect of moving costs differs between the welfare and the nonwelfare popu￾lations. I also control for the possibility that the welfare and nonwelfare populations respond differently to state characteristics. For example, individuals in the welfare population may care less about wages and unemploy￾ment if they are expecting to rely on government or family assistance. Therefore I include interactions of state-level wage and unemployment variables with an individual-specific indicator variable for individuals in the welfare population. Although the general effects of wages and unemployment are not identified (because they are soaked up by the fixed effects), I can estimate the differential effect of these variables on the con￾trol and treatment populations with these interaction terms. RESULTS The analysis proceeds in two steps. First, I replicate and extend the analysis by Schram, Nitz, and Krueger (1998) to make two points: (1) that models with no or few nonwelfare controls show no welfare effects and (2) that better accounting for state-level and nonwelfare determinants of migration produces initial evidence of a welfare effect. I then present results for the more flexible and powerful random utility model of migration. Revisiting Schram, Nitz, and Krueger Schram, Nitz, and Krueger (1998) model migration patterns of poor single mothers as a function of wel￾fare, income, and employment differentials. (Allard and Danziger (2000, 361) provide, among other anal￾yses, a similar analysis with no controls.) In two of four specifications, Schram, Nitz, and Krueger find a significant negative relationship between welfare ben￾efits and migration. This odd result suggest either that high welfare benefits repel poor single mothers (which seems unlikely and would constitute a major paradigm shift if true) or that nonwelfare factors correlated with welfare benefits have been omitted and are causing a spurious negative relationship. To investigate whether omitted variable bias is the problem, Table 1 revisits Schram, Nitz, and Krueger’s model. The dependent variable is Census Bureau data on the proportion of poor, single women with chil￾dren moving from one state to another between 1985 128

American Political Science Review Vol.99,No.1 TABLE 1.Determinants of Interstate Mobility Rates of Poor,Single Mothers All State Dyads Interstate Move Dyads Only 2 Welfare benefits difference 0.00004 0.0001* 0.00004 0.0001* (0.03) 2.21) (0.68) (2.86) Unemployment difference 0.00005 0.0001* 0.00005 0.0001* (0.04) (2.93) (0.97) (3.81) Income difference -0.00003 0.00014 -0.00003 0.0001 (0.03) (3.21) (0.76) (3.69) Nonwelfare migration 0.99+ 0.83* (1306.68) (57.58) Intercept 0.02* 0.0003* 0.002 0.0006* (7.88) (2.56) (22.58) (8.15) Observations 2,304 2.304 2.256 2.256 R2 0.000001 0.999 0.001 0.596 Note:Figures are OLS coefficients for a model in which the dependent variable is the proportion of poor single mothers moving from state j to state k for all continental state pairs;see text for details.t-statistics are in parentheses.*p <0.05;**p <0.01; *p<0.001. and 1990 for all state pairs.The independent variables Conditional Logit Results measure welfare,unemployment,and income differ- ences.Column 1 reports results for a sparse specifica- Tables 2 through 4 present the results for the more tion as in Schram,Nitz,and Krueger.The results echo compelling tests based on the individual-level model theirs:no variable is significant and the R2 is hardly of state choice.I estimate but do not report state fixed measurable.Column 2 reports results when I add con- effects for all specifications. trol for non-welfare state attractiveness with a vari- Table 2 indicates that welfare benefits exert a pos- able measuring the proportion of poor,non-college- itive and highly significant effect on migration.The educated single women without children who moved first specification includes only the distance,move,and from state j to state k.The same economic,social,and welfare variables.The second specification adds birth cultural attributes of states affect these women and the state variables.I proceed in this fashion in order to welfare population with one important exception:the highlight how omitting birth state effects attenuates women without children were not eligible for AFDC. the estimated effect of welfare on migration.Note that Their migration patterns therefore embody(and con- including the birth state variables causes the coeffi- trol for)the nonwelfare attractiveness of states. cient on welfare benefits to almost double.The welfare Including better controls dramatically changes the benefits variable is significant-and hardly changed- results.Most importantly for our purposes,the welfare in the third specification,which adds wage and unem- variable is now positive and significant,as predicted by ployment interactions for the welfare population.In the welfare migration hypothesis.One discordant note the last two specifications,I assess whether welfare is the extremely high R2.This occurs because the both benefit levels interact with birth state and distance the dependent variable and the nonwelfare migration The results indicate that both interactions matter,but variable are close to one for the 48 own-state pairs and that neither substantially changes the results.Col- close to zero for the 2,256 other pairs.Therefore the umn 4 reports the results when birth state and benefits next two columns look only at interstate move dyads variables are allowed to interact.The coefficient on by excluding own-state pairs.Again,the sparsely spec- welfare is higher than in the other specifications for ified model performs abysmally and the model with non-birth states,while the coefficient on welfare for improved controls performs much better.Here again, birth states(the sum of the main effect and the interac- welfare benefits are positively and significantly associ- tion)is essentially zero.This result implies that welfare ated with migration. and family effects are substitutes.not complements Other individual-level variables measure systematic Column 5 reports the results when distance and bene- determinants of individual-specific attraction to or re- fits interact.This tests whether the magnetic effect of pulsion from certain states.An excellent proxy for welfare diminishes across space.The results indicate family ties and social capital is the birth state of an that this is indeed the case.as the interaction is signi- individual.The variable birth state is one if person i was ficant. born in states and zero otherwise.Simply put,this vari- The other noteworthy result in Table 2 is the over- able controls for the possibility that-all else equal-a whelming statistical significance of the birth state vari- person born in Mississippi derives greater utility from ables,which consistently has a t-statistic over 90.The living in Mississippi than someone born in Vermont.I interaction with the indicator variable for poor single also interact this variable with an indicator for individ- mothers is also significant,indicating a stronger birth- uals in the welfare population in case birth state effects state attraction for poor single mothers relative to the differ for the welfare and nonwelfare populations. control group of women without children.Even taking 129

American Political Science Review Vol. 99, No. 1 TABLE 1. Determinants of Interstate Mobility Rates of Poor, Single Mothers All State Dyads Interstate Move Dyads Only 1 2 12 Welfare benefits difference 0.00004 0.0001∗ 0.00004 0.0001∗∗ (0.03) (2.21) (0.68) (2.86) Unemployment difference 0.00005 0.0001∗∗ 0.00005 0.0001∗∗∗ (0.04) (2.93) (0.97) (3.81) Income difference −0.00003 0.0001∗∗∗ −0.00003 0.0001∗∗∗ (0.03) (3.21) (0.76) (3.69) Nonwelfare migration — 0.99∗∗∗ — 0.83∗∗∗ — (1306.68) — (57.58) Intercept 0.02∗∗∗ 0.0003∗∗ 0.002∗∗∗ 0.0006∗∗∗ (7.88) (2.56) (22.58) (8.15) Observations 2,304 2,304 2,256 2,256 R2 0.000001 0.999 0.001 0.596 Note: Figures are OLS coefficients for a model in which the dependent variable is the proportion of poor single mothers moving from state j to state k for all continental state pairs; see text for details. t-statistics are in parentheses. ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001. and 1990 for all state pairs. The independent variables measure welfare, unemployment, and income differ￾ences. Column 1 reports results for a sparse specifica￾tion as in Schram, Nitz, and Krueger. The results echo theirs: no variable is significant and the R2 is hardly measurable. Column 2 reports results when I add con￾trol for non–welfare state attractiveness with a vari￾able measuring the proportion of poor, non-college￾educated single women without children who moved from state j to state k. The same economic, social, and cultural attributes of states affect these women and the welfare population with one important exception: the women without children were not eligible for AFDC. Their migration patterns therefore embody (and con￾trol for) the nonwelfare attractiveness of states. Including better controls dramatically changes the results. Most importantly for our purposes, the welfare variable is now positive and significant, as predicted by the welfare migration hypothesis. One discordant note is the extremely high R2. This occurs because the both the dependent variable and the nonwelfare migration variable are close to one for the 48 own-state pairs and close to zero for the 2,256 other pairs. Therefore the next two columns look only at interstate move dyads by excluding own-state pairs. Again, the sparsely spec￾ified model performs abysmally and the model with improved controls performs much better. Here again, welfare benefits are positively and significantly associ￾ated with migration. Other individual-level variables measure systematic determinants of individual-specific attraction to or re￾pulsion from certain states. An excellent proxy for family ties and social capital is the birth state of an individual. The variable birth state is one if person i was born in state s and zero otherwise. Simply put, this vari￾able controls for the possibility that—–all else equal—–a person born in Mississippi derives greater utility from living in Mississippi than someone born in Vermont. I also interact this variable with an indicator for individ￾uals in the welfare population in case birth state effects differ for the welfare and nonwelfare populations. Conditional Logit Results Tables 2 through 4 present the results for the more compelling tests based on the individual-level model of state choice. I estimate but do not report state fixed effects for all specifications. Table 2 indicates that welfare benefits exert a pos￾itive and highly significant effect on migration. The first specification includes only the distance, move, and welfare variables. The second specification adds birth state variables. I proceed in this fashion in order to highlight how omitting birth state effects attenuates the estimated effect of welfare on migration. Note that including the birth state variables causes the coeffi- cient on welfare benefits to almost double. The welfare benefits variable is significant—–and hardly changed—– in the third specification, which adds wage and unem￾ployment interactions for the welfare population. In the last two specifications, I assess whether welfare benefit levels interact with birth state and distance. The results indicate that both interactions matter, but that neither substantially changes the results. Col￾umn 4 reports the results when birth state and benefits variables are allowed to interact. The coefficient on welfare is higher than in the other specifications for non-birth states, while the coefficient on welfare for birth states (the sum of the main effect and the interac￾tion) is essentially zero. This result implies that welfare and family effects are substitutes, not complements. Column 5 reports the results when distance and bene- fits interact. This tests whether the magnetic effect of welfare diminishes across space. The results indicate that this is indeed the case, as the interaction is signi- ficant. The other noteworthy result in Table 2 is the over￾whelming statistical significance of the birth state vari￾ables, which consistently has a t-statistic over 90. The interaction with the indicator variable for poor single mothers is also significant, indicating a stronger birth￾state attraction for poor single mothers relative to the control group of women without children. Even taking 129