FIGHTING LONE MOTHERS POVERTY THROUGH IN-WORK BENEFITS METHODOLOGICAL ISSUES AND POLICY SUGGESTIONS. Chiara Daniela Pronzato

Transcription

1 FIGHTING LONE MOTHERS POVERTY THROUGH IN-WORK BENEFITS METHODOLOGICAL ISSUES AND POLICY SUGGESTIONS Chiara Daiela Prozato

2 Fightig Loe Mothers Poverty through I-Work Beefits Methodological Issues ad Policy Suggestios Chiara Daiela Prozato Jue 2013 ABSTRACT Loe mothers are overrepreseted amog poor people i may Europea coutries, with detrimetal cosequeces for themselves ad their childre. Also i Norway, which is kow as a coutry of ecoomic ad welfare success, loe mothers were at least three times more likely to be poor tha married mothers with childre. I 1998, a welfare reform icreased the amout of beefits ad itroduced workig requiremets. Usig a quasi-experimetal model, Mogstad ad Prozato (2012) fid a positive effect of the reform o loe mothers labour supply ad a small reductio i poverty. Is the best result that policy makers could obtai i terms of poverty reductio give the amout of public resources ivested? To aswer this questio, I estimate a discrete choice model of earigs ad welfare participatio decisios, ad use the behavioural estimates to derive the policy parameters which would have miimized poverty amog loe mothers. I order to provide policy suggestios based o robust results, the discrete choice model is validated by comparig its predictios with the estimated effects of the reform obtaied with a quasi-experimetal desig (Mogstad ad Prozato, 2012). JEL classificatio: I38, J22, C25 Keywords: loe mothers, i-work beefits, poverty, discrete choice models, compariso of methods Affiliatios ad cotact: Uiversita degli Studi di Torio, Collegio Carlo Alberto, Statistics Norway, Dodea, ISER, ad IZA. Postal address: Lugodora Siea 100, Torio, Italy. Tel chiaradaiela.prozato@uito.it Ackowledgemets: The Norwegia Research Coucil has provided fiacial support for this project. I am grateful to Rolf Aaberge, Richard Bludell, Ugo Colombio, Joh Ermisch, Fracesco Figari, Marco Fracescoi, Marco Guerzoi, Aa Lo Prete, Mage Mogstad, Adreas Peichl, Steve Pudey, Trie Vattø, Katharia Wrohlich for their commets as well as participats to semiars at ISER, DIW Berli, Dodea, the departmet of Ecoomics of the Uiversity of Turi, the departmet of Ecoomics of the Uiversity of the Italia Switzerlad, ad participats to the XXIII ESPE coferece i Seville, ad the III IMA coferece i Stockholm. Ay error should be attributed to the author. 1

3 1 Itroductio Loe mothers are overrepreseted amog poor people i may Europea coutries, with detrimetal cosequeces for themselves ad their childre. Also i Norway, which is kow as a coutry of ecoomic ad welfare success, loe mothers were at least three times more likely to be poor tha married mothers with childre i the same age-rage. To this aim, i 1998, a reform of loe paretal welfare was udertake. The mai chages ivolved the most geerous beefits, the so-called trasitioal beefit. The maximum amout of the beefits was icreased, workig requiremets were itroduced, ad ew time limits were imposed. Both loe mothers ad loe fathers were eligible for the beefit, but the policy discussio cocerig the trasitioal beefit ad how to reform it to avoid work disicetives was carried out primarily thikig about loe mothers work ad poverty rates. The reasos are twofold. First of all, as much as 9 loe parets i 10 were wome at the time of the reform. But more importatly, the huma capital levels ad socio-ecoomic status of loe fathers have bee show to differ substatially from those of loe mothers, presumably due to a strog selectio process for loe fathers to actually get daily custody of their childre (Kjeldstad ad Røse, 2004). Mogstad ad Prozato (2012), usig a quasi-experimetal model, fid a positive effect of the reform o loe mothers labour supply ad a small reductio i poverty. Is the best result that policy makers could obtai i terms of poverty reductio give the amout of public resources ivested? Usig the reform as a istrumet, i a quasi-experimetal settig, we caot aswer this questio. We ca uderstad whether loe mothers behaviour is iflueced by public policies, without strog assumptios ad referrig oly ituitively to the ecoomic theory. However, we caot distiguish the effects of the differet parts of the reform, caot uderstad the mechaisms, ad caot predict what kid of policy would have made loe mothers better off. To kow what policy would make loe mothers less likely to be poor, we eed a more structural approach: thikig what matters for loe mothers decisios (icome, hours of work, age of the childre, ), estimate what weight they give to icome,.. whe takig decisios, ad use the estimates behavioural parameters to predict ew policy scearios. Nevertheless the advatages of a structural model i providig policy suggestios, structural models are based o relatively stroger assumptios compared to quasi-experimetal methods. How ca be sure that the ecoomic model I costruct ad estimate ca reproduce how wome take their decisios i a realistic way? What I do i this paper is to simulate the chages brought by the 1998 reform i Norway usig the behavioural estimates ad the to compare its predictios with the effects of the reform 2

4 estimated with a quasi-experimetal desig whose assumptios are cosidered less strog. Oce validated the discrete choice model, I ca use it for policy suggestios. The compariso betwee quasi-experimetal methods ad structural models for policy evaluatio seems to be a area of research ivestigated oly by a few papers, but ecessary to give credibility to both the approaches, ad to recocile them. This is the appeal to youg ecoomists made by Keae (2006, 2010), durig his keyote lecture at the Duke Coferece o Structural Models i Labor, Agig ad Health (2005), titled Structural versus Atheoretic Approaches to Ecoometrics. He uderlies the ecessity of cosiderig descriptive statistics, reduced ad structural forms as well as experimetal methods as complemetary approaches to the study of the effects of policy chages. He ecourages researchers to perform validatio exercises to test the extet to which structural models give reasoable predictios of the reality. The adjective reasoable may be still judged i a subjective way, but via multiple validatio exercises cosesus may be reached. Recetly, as a part of the Mirrlees Review, Bludell (2012) has uderlied the importace of differet empirical strategies to evaluate the effects of earigs taxatio o labour market decisios i order to desig better tax policy reforms. Examples are offered by Todd ad Wolpi (2006), Bludell (2006), Brewer et al. (2006), Keae ad Wolpi (2007) ad, more recetly, by Beral a Keae (2010), Hase ad Liu (2011), Geye et al. (2012), Thorese ad Vattø (2012). Todd ad Wolpi (2006) use data from a radomized social experimet i Mexico to study ad validate a dyamic behavioural model of paretal decisios about fertility ad child schoolig. The PROGRESA is a radomized social experimet implemeted by the Mexica govermet, i which aroud 500 rural villages were radomly assiged to participate or ot i the program which provided paymets to parets who regularly sed their childre to school. They estimate the behavioural model without usig observatios from the treated villages ad predict the potetial fertility ad child schoolig of families i utreated villages. The impact of the program predicted by the behavioural model tracks the experimetal results. Keae ad Wolpi (2007) adopt aother approach to validate a behavioural model. They costruct ad estimate a dyamic structural model of female behaviour, i which work, welfare participatio, marriage ad fertility decisios are joitly cosidered. I order to validate the model, they use a holdout sample, a sample which differs from the sample used i the estimatio ad whose policy regime is well outside the support of the data. They use data from some US states to estimate the model, ad from others to predict ad validate the model. Beral a Keae (2010) evaluate the effects of materal work ad childcare use o cogitive child developmet usig a sample of sigle mothers i the Natioal Logitudial Survey of Youth. I order to take ito accout the selectio process i work ad childcare use, they develop a model of 3

5 mothers employmet ad childcare decisios. To idetify the model, they use exogeous variatios i welfare rules ad local demad coditios across States ad over time. They also employ the same istrumetal variables for a straight liear IV regressio. The estimated effects o childre s cogitive developmet are very close whe comparig the IV strategy ad the structural model. While the above studies costruct ad estimate structural models which are dyamic, there is a umber of empirical works which validate static structural models with quasi-experimetal evidece ad which maily look at labour market chages due to a chage i the welfare. Brewer, Duca, Shephard ad Suarez (2006) estimate a static structural model of labour supply ad programme participatio usig data from before ad after the itroductio of the Workig Families Tax Credit i the UK. They simulate the effect of the reform, takig ito accout all related chages i beefits ad taxes, ad compare the results with the oes obtaied from other ex-ate (Bludell et al., 2000a, 2000b) ad ex-post evaluatios (Bludell et al., 2005; Fracescoi ad Va der Klaauw, 2004, 2007; Leigh, 2005; Gregg ad Harkess, 2009). Bludell (2006) focuses o the effects of the Eared Icome Tax Credit policies o loe mothers workig decisios, by validatig a structural model of labour supply with a differece-idifferece evaluatio strategy, ad the fids the optimal policy, defied by a certai social welfare fuctio. Other recet papers compare results from quasi-experimetal methods ad structural models exploitig the itroductio of a certai reform: Geye, Ha, ad Wrohlich (2012) estimate the itroductio of a paretal leave reform i Germay by comparig workig behaviour of mothers of childre bor just before or after the reform ad compare results with the oes obtaied by a structural model of retur to work; Thorese ad Vattø (2012) evaluates a tax reform by comparig a before-after chage i labour supply with the effect predicted by a structural model of labour supply; Hase ad Liu (2011) compares the effect of a icrease of the geerosity of welfare beefits for youg people i Quebec predicted by their structural model of labour supply ad welfare participatio with the oe estimated with a discotiuity regressio model by Le Mieux ad Milliga (2008). Usig a similar approach i a comparable policy cotext as i Bludell (2006), i this paper, I validate a discrete choice model of work ad welfare participatio decisios with the results provided by a quasiexperimetal desig, usig the same data ad the same outcome variable for the two aalyses. Oce validated the discrete choice model, the behavioural parameters are the used to fid the optimal policy, defied as the policy which provides the lowest level of poverty. The paper is orgaized as follows. The 1998 reform is described i Sectio 2. The two evaluatio strategies (the quasi-experimetal model ad the discrete choice model) ad the assumptios they rely o are explaied i Sectio3, while the data are preseted i Sectio 4. After estimatig the behavioural 4

6 parameters, the reform is simulated, ad its predictios compared with the estimated effects obtaied with the quasi-experimetal method: despite the discrete choice model makig stroger assumptios, the predictios from the discrete choice model are close to the quasi-experimetal oes (Sectio 5). New policy scearios are the simulated i Sectio 6 while coclusios follow i Sectio 7. 2 The 1998 Reform of the Trasitioal Beefit I this sectio I describe the trasitioal beefit ad how has bee chaged with the 1998 reform. The reform is the directly evaluated through the quasi-experimetal desig while simulated through the discrete choice model. The trasitioal beefit used to be the most geerous beefit targeted exclusively at loe parets, maily take-up by loe mothers. Loe mothers with at least oe child youger tha 10 years old used to receive up to 700 per moth. The receipt of the beefit was idepedet of their labour market decisios but 40% of their mothly earigs exceedig 200 used to be withdraw from the maximum amout. A reform of the trasitioal beefit was udertake o the 1st of Jauary First, work requiremets were imposed: loe mothers, i order to be eligible for the beefit, were supposed to work at least part-time, to actively seek work, or to be i traiig. However, the workig requiremets were oly itroduced for loe mothers with the yougest child older tha 3 years old. Secod, the timig of the beefit was chaged: loe mothers could receive the beefit util the yougest child was 8 years old (istead of 10) 1 ad for a period up to 3 years (while it was potetially 10 years before the reform). Fially, the maximum beefit amout was icreased from aroud 700 to 800 per moth. 3 Evaluatig the Effect of the Reform I this sectio, I describe the two methods used to evaluate the effect of the 1998 reform o loe mothers workig decisios. The first strategy to evaluate the 1998 reform is a quasi-experimetal desig (a triple-differece model), where the workig behaviour is compared before ad after the 1998, 1 Before the refom a loe mother could receive the beefit util Jue of the 10 th birthday of the yougest child; after the reform, util the yougest does ot become 9 years old. The mai differece is that the ith year of life of the child, after the reform, does ot give the right to the beefit aymore. 5

7 for a group of loe mothers (treatmet group) ad a group of married mothers (cotrol group). The secod is a discrete choice model where wome are assumed to take their decisios about work ad welfare participatio uder a certai budget costrai. After estimatig the model, the chages brought by the 1998 reform (ew workig requiremets, ew age limits, icreased amout of the beefit) are icluded i the model ad the effect of the 1998 reform is simulated by usig the estimated behavioural parameters. 3.1 A Quasi-Experimetal Evaluatio Desig I this sub-sectio, I itroduce a evaluatio desig which exploits the availability of data o workig decisios of mothers, before ad after the time of the reform. However, I caot use the classical differece-i-differece model, sice I caot observe exactly the same wome both before ad after the reform. The reform, i fact, is characterized by a log phase-i period: all loe mothers who were already i welfare at the time of the reform were allowed to receive the beefit accordig to the prereform rules for aother 3 years. Therefore, loe mothers who may be observed both before ad after the reform whose observatios could be used for a classical differece-i-differece model - have basically o icetive to chage their behaviour. Therefore, I evaluate the effect of the reform by comparig the effect of becomig a loe mother i the pre-reform ad post-reform period (Mogstad ad Prozato, 2012): ζ [E(Y [E(Y v 1 v 1 Y Y v 1 v 1 S v S v 0,R 1,R v v 1) E(Y 1) E(Y v 1 v 1 Y Y v 1 v 1 S v S v 0,R 1,R v v 0)] 0)] (1) where - Y v are the aual earigs of the woma i the year v (hours of work are ot available i the Register data, see Sectio 4) - S v is equal to 1 if the woma gets separated ad becomes a loe mother i year v, 0 if remais married - R v is equal to 1 if v 1998, 0 otherwise. The ituitio behid is that a married mother who gets separated after the reform faces differet icetives to chage her labour supply: before the reform, she kows that 40% of her earigs are take 6

8 from the maximum amout but ay chage i her labour supply would ot modify her right to the beefit; after the reform, she kows that by decreasig her hours of work to less tha part-time, she would lose the right to the beefit. The assumptio of the model is that, i absece of the reform, married mothers who become loe mothers after the reform would behave the same as married mothers who became loe mothers before the reform. By cosiderig oly the flow of ew loe mothers, I ca overcome the phase-i problem 2. Not oly, ew cohorts of loe mothers - observed whe have just separated - should be of primary iterest for policy-makers, more tha a represetative sample of loe mothers at a certai poit i time, which would over represet loe mothers who have bee loe for log time A Discrete Choice Model of Work ad Welfare Participatio Decisios 4 The Norwegia register data, I use for the estimatio of the model (described i Sectio 4), provide accurate iformatio o icomes ad demographic characteristics but ot hours of work. The model outlied below takes this feature ito accout, allowig time of work to be measured with a error. A loe mother, labelled, is assumed to maximize a utility fuctio U ( x, t, w) (2) uder the budget costrait where x T l, t, y ) (3) ( - x - l is the et household icome, is the gross mothly labour icome of the loe mother i a full time job, - t is the umber of equivalet full time moths of work i oe caledar year, - w is a welfare participatio idicator, 2 Robustess checks are carried out ad show i Mogstad ad Prozato (2008, 2012) i order to make sure that compositioal chages i the sample of mothers, before ad after the reform, do ot bias the estimates. 3 See Mogstad ad Prozato (2012) for what cocers the impact of the refom o log-lastig loe mothers. 4 This paragraph follows Trai s book o Discrete Choice Methods ad Simulatios, chapter 2 (2003). Other papers used as refereces to write the model are Mc Fadde (1974), Moffitt (1983), MaCurdy et al. (1990), Ilmakuas ad Pudey (1990), Va Soest (1995), Hoyes (1996), Aaberge et al. (1999), Creedy ad Kalb (2005), Creedy et al. (2006), Keae (2011), Bargai et al. (2012). 7

9 - y is exogeous household gross icome, - T(.) is the tax-beefit fuctio which trasforms gross icome ito et icome. The loe mother faces a set of J discrete alteratives, defied by the combiatio of earigs ad welfare participatio decisios. She kows how much utility she would get from each alterative j ad chooses the alterative which provides the largest oe. We ca decompose the utility fuctio ito two parts: the determiistic part ad the stochastic part U j J (4), V where V captures the portio of utility which derives from observable characteristics, while ε the portio from uobservable oes. The determiistic part of the utility V may be see as a fuctio which relates the observable characteristics to the loe mothers utility V V z, s ) j J (5), ( where z are the observed attributes of the alteratives as faced by the loe mother, ad s the observed socio-demographic characteristics of the loe mother. I specify the determiistic part of the utility to be liear i parameters with a costat V q' k j J (6) j where q is a vector of variables that relate to alterative j as faced by the loe mother, θ are the coefficiets of these variables, ad k j is a costat that is specific to alterative j. The costat k j captures the average effect o utility of all factors ot icluded i the model. The vector z icludes the et icome available to the loe mother at alterative j ad its square, the time of work required by alterative j ad its square, a welfare participatio idicator, ad their iteractios. The sociodemographic variables s caot eter the model directly, sice they do ot vary across alteratives. They are iteracted with et icome, time of work ad the welfare idicator to allow utility from icome ad disutility from time of work ad welfare participatio to be differet for wome with differet levels of educatio, age, atioality, umbers ad ages of childre: 8

10 V k x j 1 ( x x s )' ( t 2 2 t 3 t s )' ( w 2 4 s w 5 )' x 6 t j J t 7 w x 8 w (7) where x is her et household icome, t her time of work, w her welfare participatio idicator i each alterative j, ad s are her demographic characteristics. Time of work t is ot observed i the register data. I derive the expected time of work t, expressed i equivalet full time moths of work i a year, as the ratio betwee each woma s aual earigs i the register data ( l t ) ad the predicted mothly earigs from survey data ( l ) i a full time job of a woma with same huma capital characteristics: l t (8). l t The relatioship betwee true time of work t ad expected time of work t is give by t l t t (9), l where is egatively correlated with the uobservable characteristics which make a woma ear more. If a woma ears more tha what, o average, a woma with the same observable characteristics does, it meas that she eeds to work less time tha what I predict as expected time of work. coects true ad expected time of work ad is assumed to be ormally distributed. Therefore (7) becomes V k 2 ~ ~ 2 1x 2x 3t 4t 5w ( x s )' ( t s )' ~ ( w s )' j ~ x 6 t j J ~ t 7 w x 8 w (10) where, for example, 9

11 ~ (11). 6 6 ~ The model I estimate allows disutility from time 3 to be differet for wome with differet uobservable characteristics: V x x 8 1 w 2 ~ 2 2x ( 3 ) t 4t 5w k ( x s )' ( t s )' ~ ( w j ~ x s 6 )' t ~ t 7 w j J (12). coicides with oly if there is o differece i tastes due to uobservables amog wome 5. However, I do ot eed to idetify because the mai aim is to take ito accout that time of work is measured with a error. 4 The Data The data used for the empirical aalysis are from the register data of the Norwegia populatio i the period , which cotais household ad demographic iformatio, ad is merged with detailed icome data from the Tax Assessmet Files through uique idividual idetifiers. The icome data are collected from tax records ad other admiistrative registers rather tha iterviews ad self-assessmet methods. The coverage ad reliability of Norwegia register data are cosidered to be exceptioal, as is documeted by the fact that the quality of such atioal data sets received the highest ratig i a data quality survey i the Luxembourg Icome Study database (Atkiso et al., 1995). The populatio of study comprises married, cohabitig ad loe mothers who i each year were at least 18 years old ad ot more tha 55, with the yougest child betwee 4 ad 9 years old. From ow o, for simplicity, I cosider married ad cohabitig mothers together, referrig geerally to all of them as 5 The model is estimated usig the software Stata (commad: mixlogit). 4, 6, 7, should be also allowed to vary amog wome but, i practice, the model does ot coverge whe allowig uobservable heterogeeity i may parameters. For more details, see a loger versio of the model i Prozato (2012). 10 ~ ~ ~ ~

12 married 6. Self-employed wome, studets, as well as wome receivig permaet disability beefits, are excluded from the aalysis. The sample for the triple-differece model is composed of 1,121,898 wome: becomig loe mothers (treatmet group) ad stayig married mothers (cotrol group), before ad after the reform. The sample for the discrete choice model is composed of the sub-sample of loe mothers observed before the reform (7,921), represetig what oe would use i a typical ex-ate evaluatio. The effect of the 1998 reform of the trasitioal beefit is evaluated o aual gross earigs rather tha o hours of work. The reaso for focusig o earigs to evaluate the effects of the reform o labour market participatio is that I do ot have data o workig hours. To limit measuremet error whe usig aual gross earigs, I use the cosumer price idex to make icomes from differet periods comparable; throughout this paper the referece year is 1998, ad 1 is set equal to NOK 8.4. Details o the sample selectio are described i the ext sectios. 4.1 Data for the Estimatio of the Quasi-Experimetal Evaluatio Desig The way the iformatio is registered is very importat for uderstadig the evaluatio desig: we kow the marital status of the woma o Jauary 1 st of each year while gross earigs are measured aually, from Jauary 1 st to December 31 st of each year. The sample is the selected as follows. The treatmet group after the reform is composed of the cohort of mothers who are married o Jauary 1 st 1997 (with the yougest child 2-7 years old), who are still married o Jauary 1 st 1998 (with the yougest child 3-8 years old), who are loe o Jauary 1 st 1999 (with the yougest 4-9 years old), therefore gettig separated ay day betwee Jauary 2 d ad December 31 st By comparig their earigs betwee 1999 ad 1997, we estimate a chage i earigs which may be due to may factors, amog which the fact of becomig a loe mother, the work-icetives provided by the welfare system, ad other time varyig factors 7. The compariso with married mothers o Jauary 1 st 1997 (with the yougest child 2-7 years old), who are still married o Jauary 1 st 1998 (with the yougest child 3-8 years old), who stay married o Jauary 1 st 1999 (with the yougest 4-9 years old) helps to clea the total chage i earigs by the other time-varyig factors (maily, the child who is growig ad the ecoomic tred) ad to idetify the causal effect of becomig a loe mother o earigs after the 6 Register data do ot allow for idetificatio of cohabitat couples directly, but Galloway ad Aaberge (2007) costructed a household type variable, derived from a large variety of iformatio which ca help to idetify cohabitats idirectly. 7 For this cohort of mothers, we do ot use iformatio about their earigs i 1998, sice we caot kow for what part of the year the mother was married ad for what part of the year the mother was loe. Moreover, this temporal lag allows mothers to have time to adjust their behaviour. 11

13 reform. By doig the same exercise before the reform (married mothers i 1995, i 1996, who may or may ot be loe i 1997), I ca idetify the causal effect of becomig a loe mother o earigs before the reform. The differece betwee the effect of becomig a loe mother before ad after the reform idetifies the casual effect of the reform. I the aalysis, I exploit all cohorts available i the data 8, ad iclude local uemploymet rate ad year time dummies as furthers cotrol of the ecoomic cycle. 4.2 Data for the Estimatio of the Discrete Choice Model From the sample described i the previous sessio, I select ew loe mothers before the reform (i the years 1995, 1996, 1997). I assume loe mothers face at most 8 alterative choices, give by the joit decisio of how much to work (4 alteratives) ad whether or ot participatig i the welfare (2 alteratives). As explaied i Sectio 3, expected time of work is obtaied comparig aual earigs observed i the register data with potetial mothly full time earigs from survey data. I order to costruct potetial earigs, I use the Norwegia part of the Europea Uio Survey of Icome ad Livig Coditios for the year 2004, I select wome i the same age-rage (18-55), ad I estimate a Heckma regressio. The depedet variable is hourly gross labour icome. I the outcome equatio I iclude two dummy variables for educatio (secodary ad tertiary educatio), a variable for potetial workig experiece (age - years of schoolig - 7), its square, ad a part time dummy. 9 I the selectio equatio, I also cosider the presece of depedet childre, other household icome, whether beig i a couple, ad livig i a city. Results are reported i Table A1. I order to make survey earigs comparable to earigs i the register data, predicted hourly earigs are multiplied by typical hours of work i a full time job (38) ad umber of weeks i a moth, ad adjusted i order to take ito accout omial ad real growth. 10 The 4 work alteratives are defied i the followig way: 8 Mothers married i 1993/1994/1995/1997/1998/1999, who are still married i 1994/1995/1996/1998/1999/2000, who may or may ot be loe mothers i 1995/1996/1997/1999/2000/2001. The first three cohorts are ot iflueced by the reform, while the last three are iflueced by the reform. Time dummy variables are icluded i the regressio model. 9 I iclude a part time dummy to test whether the wage rate ca be cosidered costat over time of work. Part time wage rate is ot sigificatly differet from full time wage rate, as show i Table A1. 10 Prices are deflated to Real growth is take ito accout lookig at the variatio, year by year, of the basic amout (grubeløp), which is the official referece amout used for the up-ratig of beefits ad pesios. 12

14 1) First work alterative (which will be called o work ): ratio betwee aual observed earigs ad expected mothly earigs i a full time job smaller tha 3, which correspods to less tha 9.5 hours a week (o average, i the data, 1 hour ad half per week). 2) Secod work alterative (which will be called short part time ): ratio betwee observed aual earigs ad expected mothly earigs i a full time job larger or equal to 3 ad smaller tha 6, which correspods to hours a week (o average, i the data, 13 hours per week). 3) Third work alterative (which will be called part time ): ratio betwee observed aual earigs ad expected mothly earigs i a full time job larger or equal to 6 ad smaller tha 9, which correspods to hours a week (o average, i the data, 22 hours per week). 4) Fourth work alterative (which will be called full time ): ratio betwee observed aual earigs ad expected mothly earigs i a full time job larger or equal to 9, which correspods to more tha 28.5 hours a week (o average, i the data, 33 hours per week). I the observed choice, the three objects of the utility fuctio are defied as follows: (i) the observed welfare participatio decisio, (ii) the et icome which derives from observed earigs through the tax-beefit fuctio (2) ad (iii) the expected umber of moths of work, obtaied dividig the observed aual earig by potetial mothly earigs i a full time job. For the other 7 alteratives I costruct couterfactuals. Suppose her observed earigs are 17,500 ad she participates i the welfare (see example i Table 1). Give her huma capital characteristics, she is supposed for example to ear 2,500 per moth i a full time job. I classify her as workig part time (17,500 / 2,500 = 7 equivalet full time moths; 22 hours per week). I costruct three other earig alteratives: o work, workig short part time, workig full time (see Table 1, first five colums). The umber of moths chose for each utake work alterative (o work / short part time / full time) is draw from the distributio of moths of people choosig that alterative (o work / short part time / full time) (Aaberge et al., 2009). Predicted earigs are the imputed. I the example, Table 1, the draw umbers of moths are 0, 4 ad 12, ad earigs are, respectively, 0, 10,000, ad 30,000. For each earig alterative, she ca decide whether to participate i the welfare. The trasitioal beefit is calculated as follows. The maximum aual amout of the beefit is aroud 8,000 per year. From this maximum amout, 40% of earigs exceedig 2,500 are subtracted. I Table 1, 6 th colum, we ca see the correspodig amouts. For this woma, the 7 th alterative is dropped, sice the related full time earigs are too large to be still eligible for the beefit. 13

15 I the simulate the childcare beefit, aother beefit which depeds o labour supply, give as a reimbursemet for extra-costs for childcare, occurred whe the mother works. All other remaiig beefits are oly available i the data as a total amout. However, oe of them depeds o her workig decisios. Fially, I simulate taxes, ad obtai the total et icome she ca have i differet work/welfare alteratives (8 th colum, Table 1). Poverty is defied by a dichotomous variable takig the value of 1 if the loe mother s household has aual equivalet disposable icome below 60 percet of the media aual equivalet disposable icome i the overall populatio, ad 0 otherwise. The 9 th colum (Table 2) idicates i which alteratives the household is cosidered poor: i the example, the household would be poor if the mother decided ot to work, or worked short part time ad did ot take-up the beefit. The variable, i the 10 th colum (Table 1), idicates the decisio observed. Descriptive statistics of the sample of loe mothers before the time of the reform are show i Table 2. This also represets the sample of referece for the simulatio of the optimal policy (which miimizes poverty) show i Sectio 6. 5 Comparig the Estimated Effects of the Reform 5.1 Quasi-Experimetal Estimated Effects ad Compariso with the Simulated Effects of the Discrete Choice Model Results of the triple-differece model are show i Table 3. I geeral, I fid a positive effect of the reform: loe mothers icrease their earigs of 384 per year. Whe distiguishig wome by level of educatio, results appear positive ad sigificat oly for low ad medium high educated wome, osigificat for highly educated wome. We observe that the itroductio of the reform has reduced the distace betwee loe ad married mothers: before the reform becomig a loe mother implies, o average, a decrease of 702 i earigs; after the reform, it implies a decrease of oly 320 ( ). Results are robust to differet specificatios ad to the iclusio of more cotrol variables Estimated Behavioural Parameters from the Discrete Choice Model I estimate the effects of icome, time of work, welfare participatio ad their iteractios with other socio-demographic variables, o the probability of choosig oe of the alteratives, usig a mixed logit specificatio with the coefficiet of time of work treated as radom coefficiet, assumed to be ormally 11 See Mogstad ad Prozato (2008, 2012). 14

16 distributed, as outlied i Sectio 3. Results are reported i Table 4. The model is estimated without ay restrictio imposed o the utility fuctio. To check that the utility fuctio respects the cocavity ad mootoicity properties, I check the derivatives with respect to the utility argumets. The first derivative with respect to icome is positive for the whole sample as well as the first derivative with respect to time of work is egative for the whole sample. Secod derivatives are i the expected directio, as show i Table 4. Utility is decreasig i welfare participatio for 96% of the sample. The stadard deviatio of the radom coefficiet is sigificatly differet from zero, revealig a importat role of uobserved heterogeeity ad/or measuremet error. The iteractio betwee icome ad time of work is positive, which may be explaied by the presece of better positios i the labour market that, eve if imply loger hours of work, icrease woma s utility. The iteractio betwee welfare participatio ad icome is positive: sice a large part of loe mothers icome comes from other beefits i Norway, the positive iteractio could reveal that the cost of participatig is lower for wome who also participate i other welfare programs. The iteractio betwee welfare ad time of work is also positive: wome who work more may be more iformed because they are more likely to talk with other people at the place of work or they may suffer less from welfare stigma because they feel they do ot completely deped o welfare 12. Results cocerig umber ad age of childre are i the expected directio: o the oe had, havig more ad youger childre icreases the cost of workig; o the other had, it icreases utility from icome. Immigrat wome have more disutility from time of work. This fidig could result from the fact that, give their level of educatio, they are i poorly paid jobs. Youger wome have less disutility from participatig i the welfare while the cost of the welfare is ot liear by years of educatio. Wome with secodary schoolig have less disutility from participatig i the welfare tha higher ad lower educated wome. This may capture differet aspects of welfare participatio: o the oe had, if iformatio is eeded the better educated wome may be more prompt to apply for the beefit; o the other had, better educated wome may suffer more to be depedet from welfare. 5.3 Simulatig the Effect of the Reform I order to simulate the reform with the discrete choice model, I eed to parameterize the trasitioal beefit accordig to the ew rules. There are three importat chages. First, workig requiremets are 12 Geerally speakig, i the Norwegia cotext, where applicatios ca be doe o lie ad trasfers ca be received i the bak accout without frieds ad family ecessarily kowig, we could expect the role of the welfare stigma to be relatively less importat. 15

17 imposed. Secod, the age limit for eligibility o the yougest child is lowered, ad time limits o welfare participatio are itroduced. Ad third, i-work beefit levels are raised. The icrease of the maximum amout (from aroud 8,300 to 9,500 per year) ot oly makes the trasitioal beefit more geerous but also makes wome more likely to be eligible: before the reform, oly wome earig less tha 1,900 per moth ca receive the beefit while, after the reform, wome earig util 2,200 per moth are also eligible. This results i a larger umber of alteratives i the choice set for those wome ow eligible to receive the beefit. Accordig to the chage of the age limit, wome with the yougest child aged 9 years old are ot allowed to receive the trasitioal beefit aymore. The reform requires loe mothers to be i traiig, to work at least part time, or to seek work. The law does ot give details about the traiig ad seekig work activities. I do ot have ay iformatio o what these activities cosist of, whether it was difficult to have a traiig period, what wome were asked i case they were seekig work. Ad I do ot kow whether these activities were easily approved by the public admiistratio, ad for how log they were compatible with beig eligible for the beefit. Moreover, register data do ot have iformatio o traiig ad periods of seekig work, so that I caot kow who was takig this decisios, eve after the reform. What I ca do is to assume that loe mothers receivig the maximum amout of the trasitioal beefit after the reform are egaged i oe of the two activities. I fact, the maximum amout is give oly to wome earig less 200 per moth ad it is reasoable to assume there is o part time job i Norway paid less tha 200 a moth. The percetage of o-workig wome o welfare after the reform (supposed to be i traiig or seekig-work activities) is 7.0% while the percetage of o-workig wome o welfare is 18.5% before the reform. While before the reform, the possibility of receivig the trasitioal beefit ad ot workig was a woma s decisio, after the reform it is the result of the woma s decisio ad the ew costraits imposed by the law. We may expect wome with lower level of educatio to be more likely to be observed i traiig or seekig work activities after the reform, but this does ot seem to be the case whe lookig at post-reform data. The percetage of o-workig wome receivig the beefit before the reform was 26.4% amog low educated wome, 14.0% amog medium educated wome, 11.0% amog highly educated wome. The percetage of o-workig wome receivig the beefit after the reform (assumed to be i traiig or seekig work activities) is 8.7% amog low educated wome, 6.3% amog medium educated wome, 7.6% amog high educated wome. I order to reproduce what I observe i the data, I drop radomly the alteratives of o-workig ad participatig i the welfare for a umber of wome so that the percetage of wome takig this decisio - with the ew rules - is 7.0%. The simulated effects of the reform from the structural model are 16

18 calculated as weighted earigs, give by the sum of earigs i each alterative times the probability of choosig that alterative. Results are show i Table 5. The simulated effect of the whole reform is positive: o average, loe mothers icrease their earigs of per year. The effect appears heterogeeous for wome with differet level of educatio: positive ad stroger for low ad medium educated wome, while egative ad smaller for high educated wome 14. We ca separate the effects of the itroductio of the workig requiremets ad ew age limits from the effect of more geerous beefits. The bottom part of Table 5 summarizes the results. The itroductio of workig requiremets has icreased wome s earigs, as expected. The effect is larger for low ad medium educated wome tha for highly educated wome. Also the ew age limit has a positive but small effect o work decisios. Makig the beefit more geerous has the expected egative effect o aual earigs. The effect is relatively large also for highly educated wome. I fact, for highly educated wome, the icrease i the maximum amout has made them eligible i more work alteratives. Results from differet specificatios of the model are icluded i the Appedix (Table A2). The compariso betwee tables 3 ad 5 represets a first cotributio of the paper. Predictios from the discrete choice model, usually cosidered ecoometrically more fragile, are cofirmed by the tripledifferece model. Results of the two methods are, i fact, positive sigificat ad poit estimates are rather close. Not oly, the discrete choice model predicts well also by level of educatio: the positive effect is larger for low-medium educated wome, while slightly egative for highly educated wome. 6 New Policy Scearios The compariso betwee the results of the discrete choice model ad the triple-differece model makes me cofidet i usig the behavioural estimates to fid which policy chages to the trasitioal beefit would have miimized poverty amog loe mothers. Before the reform, the percetage of poor loe mothers is 11.8, as show i the 1 st colum of Table 6. At the bottom of the Table, the parameters of the reform are reported. 13 I order to calculate the cofidece itervals aroud the predictios from the structural model, I employ the bootstrap method: I draw 100 ew samples from the origial oe, each of them cotaiig the origial umber of observatios (N = 7,921), where each observatio may be repeated more tha oce (with replacemet); I re-estimate the model usig each of the 100 ew samples; I parameterize the reform ad get 100 sets of predictios. From these predictios I calculate meas, stadard errors, ad cofidece itervals. 14 Imposig the proportio of wome i traiig or seekig work activities to 7.0% implies, accordig to the discrete choice model estimates, to grat the possibility of receivig the beefit without workig to 40 wome out of

19 The aim is to fid the policy parameters which miimize poverty 15. I look at two situatios: whe the workig requiremets are those implemeted at the time of the reform: i order to be eligible for the beefit, wome are required to be i part-time work, to seek work, to be o traiig (case 1); whe the workig requiremets are itroduced without the possibility of traiig or seekig work (case 2). These two scearios should be iterestig bechmarks for policy makers. The first sceario represets the case i which policy makers wat to itroduce workig requiremets but allow wome to ivest time i traiig ad seekig work. The secod sceario ca be see as a log term realizatio of the reform: after the first period spet i traiig or seekig work, wome have to work to be still eligible for the beefit. I this last sceario I also allow time of work to vary i order to choose the workig requiremet which miimizes poverty. I the first sceario, istead, workig requiremets are reproduced as observed at the time of the reform. Aother differece betwee the two scearios is the amout of resources ivolved: stricter workig requiremets (case 2) imply a lower public expediture, which derives from givig less geerous beefits (due to the withdrawal rate) ad from cacellig welfare for wome who evetually decide ot to work. I the 2 d ad 3 rd colums of Table 6, I report the simulated effects of the actual reform o poverty i the two policy scearios. Poverty decreases to 8.6% (case 1) ad to 9.4 (case 2) while the average cost per woma is, respectively, 3,163 ad 1,920. I order to fid the optimal policies, uder reveue eutrality, I vary the maximum amout of the beefit, the withdrawal rate, the disregarded amout, the age limit ad, oly for case 2, the workig requiremets. I order to fid the parameters of the reform I proceed with a two-step maximizatio procedure 16. The results are show i Table 6. If we cosider the case where workig requiremets are 15 As covetioally doe, a household is cosidered poor whe the equivalet household icome is below 60% of the media equivalet household icome i the geeral populatio. 16 I the first step, I wide the iterval aroud each parameter i tur to try all possible combiatios of the parameters, util I caot fid ay additioal combiatio that gives a lower level of poverty. Whe I arrive to this stage, the policy parameters itervals are: maximum amout: 6,672 13,344 (case 1), 3,336 10,008 (case 2), withdrawal rate: 0 64 % (case 1), % (case 2), disregarded amout: 1,005 5,026 (case 1), 0 8,042 (case 2); age limit: 7 10 (case 1, case 2); workig requiremets: 0 8 equivalet full time moths of work (case 2). I the secod step, withi the above itervals for each parameter I try all the possible combiatios cosiderig small variatio i the parameters each time, i order to fid the optimal solutios which miimize poverty. 18

20 implemeted as i 1998 with the possibility of traiig ad seekig work (case 1), with a average expediture of 3,163 per loe mother, we observe a further decrease i poverty to 7.2%. Comparig the parameters of the beefit betwee the actual reform ad the optimal policy, we see that the reductio i poverty is a cosequece of the reductio of the withdrawal rate ad of the disregarded amout, while the maximum amout ad age limit are the oes observed, respectively, i the pre ad post reform period. I the secod sceario (case 2), we also observe a declie i poverty which is ow equal to 9.0%. Also i this case, the withdrawal rate is the parameter which is more distat from the observed oe. The maximum amout is still aroud 8,000 per year while, the disregarded amout has icreased. I this sceario, I allowed the required time of work to vary. However, the optimal oe is cofirmed to be 6 moths a year (part time). Also the age limit is cofirmed to be 9 years old. Could they have reformed the trasitioal beefit i a more efficiet way ivestig the same amout of public resources? Yes, the paper shows that this would have bee possible, ot by icreasig the geerosity of the beefit, but through a reductio of the withdrawal rate so that decisio to work could have bee more attractive, leadig to higher icome, ad lower poverty. 7 Coclusios I this paper, I compare the effect of the 1998 Norwegia welfare reform o loe mothers earigs estimated usig a triple-differece model ad a discrete choice model of earigs ad welfare participatio decisios. The reform icreases the maximum amout of the trasitioal beefit, itroduces ew workig requiremets ad chages time limits i order to be eligible for it. A first cotributio of the paper is to compare two differet ways of doig policy evaluatio. From both the evaluatio methods, we observe a positive effect o loe mothers earigs, drive by behavioural resposes of lower ad medium educated wome. The two strategies help the uderstadig of the policy impact i a complemetary way: while the focus of the triple-differece model is to measure what really happeed, the challege of the discrete choice model is to predict what potetially ca happe. Both aspects are importat from a policy poit of view. The fact that predictios provided by the discrete choice model track the results of the triple-differece aalysis gives credibility to both the approaches. From a policy poit of view, the availability of structural models gives policy makers the opportuity to pla how to use ratioally the resources at disposal to pursue their social objects. The 19

21 mai cotributio of the paper is to suggest what would work better for fightig loe mothers (ad their childre) poverty. I the studied case, we observe that uder reveue eutrality loe mothers poverty could be more efficietly reduced by lowerig the withdrawal rate. This would give wome the icetive to work ad ear more, ad to reach a level of icome beyod the poverty-lie threshold. 20

22 Refereces Aaberge R., Colombio U., ad Strom S. (1999), Labour Supply i Italy: a Empirical Aalysis of Joit Household Decisios, with Taxes ad Quatity Costraits, Joural of Applied Ecoometrics, 14: Aaberge R., Colombio U., ad Weemo T. (2009), Evaluatig Alterative Represetatios of the Choice Sets i Models of Labour Supply, Joural of Ecoomic Surveys, 23(3): Atkiso A., Raiwater L., ad Smeedig T. (1995): Icome Distributio i OECD Coutries, Paris, OECD. Bargai O., Orsii K., ad Peichl A. (2012), Comparig Labor Supply Elasticities i Europe ad the US: New Results, IZA Discussio Paper Beral R. ad Keae M. P. (2010), Quasi-structural estimatio of a model of childcare choices ad child cogitive ability productio, Joural of Ecoometrics, 156: Bludell R. (2006), Eared Icome Tax Credit Policies: Impact ad Optimality. The Adam Smith Lecture 2005, Labour Ecoomics, 13: Bludell R. (2012), Tax Policy Reform: The Role of Empirical Evidece, Joural of the Europea Ecoomic Associatio, 10(1): Bludell R., Brewer M., ad Shephard A. (2005), Evaluatig the Labour Market Impact of Workig Families Tax Credit usig Differece-i-differeces, Ope Access from Uiversity College Lodo, Bludell R., Duca A., McCrae J., ad Meghir C. (2000a), Evaluatig I-Work Beefit Reform: the Workig Families Tax Credit i the UK, JCPR Workig Paper o 160. Bludell R., Duca A., McCrae J., ad Meghir C. (2000b), The Labour Market Impact of the Workig Families Tax Credit, Fiscal Studies, 21: Brewer M., Duca A., Shephard A., ad Suarez M. J. (2006), Did Workig Families Tax Credit Work? The Impact of I-work Support o Labour Supply i Great Britai, Labour Ecoomics, 2006: Creedy J. ad Kalb G., (2005), Discrete Hours Labour Supply Modellig: Specificatio, Estimatio ad Simulatio, Joural of Ecoomic Surveys, 19: Creedy J., Kalb G., ad Scutella R. (2006), Icome Distributio i Discrete Hours Behavioural Microsimulatio Models: A Illustratio, Joural of Ecoomic Iequality, 4: Fracescoi M. ad Va der Klaauw W. (2004), The Cosequeces of I-work Beefit Reform i Britai: New Evidece from Pael Data, ISER Workig Paper o 13. Fracescoi M. ad Va der Klaauw W. (2007), The Socioecoomic Cosequeces of I-work Beefit Reform for British Loe Mothers, Joural of Huma Resources, 42(1): Geyer J., Ha P. ad Wrohlich K. (2012), Labor supply of mothers with youg childre: Validatig a structural model usig a atural experimetal, mimeo. Gregg P. ad Harkess S. (2009), Welfare Reform ad Loe Parets i the UK, Ecoomic Joural, 119(535): F38-F65. Hase J. ad Liu X. (2011), Estimatig Labor Supply Resposes ad Welfare Participatio: Usig a Natural Experimet to Validate a Structural Labor Supply Model, IZA Discussio Paper Hoyes H. V. (1996), Welfare Trasfers i Two-Paret Families: Labor Supply ad Welfare Participatio Uder AFDC-UP, Ecoometrica, 64: Ilmakuas S. ad Pudey S. (1990), A Model of Female Labour Supply i the Presece of Hours Restrictios, Joural of Public Ecoomics, 41: Keae M. P. (2006), Structural vs. Atheoretic Approaches to Ecoometrics, Keyote Address at the Duke Coferece o Structural Models i Labor, Agig ad Health 2005, mimeo. Keae M. P. (2010), Structural vs. Atheoretic Approaches to Ecoometrics, Joural of Ecoometrics, 156(1): Keae M. P. (2011), Labour Supply ad Taxes: A Survey, Joural of Ecoomic Literature, 49(4):