PSC. Research Report. Out of the Wallet and into the Purse: Modeling Family Expenditures to Test Income Pooling P OPULATION STUDIES CENTER

Transcription

1 Jennfer Ward-Batts Out of the Wallet and nto the Purse: Modelng Famly Expendtures to Test Income Poolng Report No Research Report PSC P OPULATION STUDIES CENTER AT THE INSTITUTE FOR SOCIAL RESEARCH U NIVERSITY OF MICHIGAN

2 The Populaton Studes Center at the Unversty of Mchgan s one of the oldest populaton centers n the Unted States. Establshed n 1961 wth a grant from the Ford Foundaton, the Center has a rch hstory as the man workplace for an nterdscplnary communty of scholars n the feld of populaton studes. Today the Center s supported by a Populaton Research Center Core Grant from the Natonal Insttute of Chld Health and Human Development (NICHD) as well as by the Unversty of Mchgan, the Natonal Insttute on Agng, the Hewlett Foundaton, and the Mellon Foundaton. PSC Research Reports are prepublcaton workng papers that report on current demographc research conducted by PSC assocates and afflates. The papers are wrtten by the researcher(s) for tmely dssemnaton of ther fndngs and are often later submtted for publcaton n scholarly journals. The PSC Research Report Seres was begun n 1981 and s organzed chronologcally. Copyrghts are held by the authors. Readers may freely quote from, copy, and dstrbute ths work as long as the copyrght holder and PSC are properly acknowledged and the orgnal work s not altered. PSC Publcatons Populaton Studes Center, Unversty of Mchgan PO Box 1248, Ann Arbor, MI USA

3 Out of the Wallet and nto the Purse: Modelng Famly Expendtures to Test Income Poolng Jennfer Ward-Batts * Unversty of Mchgan November 2000 Abstract: Untary models, assumng a sngle objectve functon and unfed budget constrant, are tradtonally used to model household behavor. Most emprcal tests of untary models rely on endogenous regressors. Ths paper uses an exogenous change n the ntrahousehold dstrbuton of ncome, provded by a change n U.K. Famly Allowance polcy. Two approaches address problems nherent n makng nference about consumpton from expendture data. Expendture shares and levels for a range of goods are estmated. Shfts n assgnable goods suggest that chldren benefted at the expense of men when ths polcy change shfted ncome wthn households from men to women. JEL codes: D12: Household Behavor and Famly Economcs Consumer Economcs: Emprcal Analyss D19: Household Behavor and Famly Economcs Other D79: Analyss of Collectve Decson-Makng Other I38: Welfare and Poverty Provson and Effects of Welfare Programs J18: Demographc Economcs Publc Polcy Keywords: ncome poolng, ntrahousehold resource allocaton, chld beneft, collectve model, untary model, famly polcy, household demand *Populaton Studes Center, PO Box 1248, Unversty of Mchgan, Ann Arbor, MI , emal: jwbatts@umch.edu. I am grateful to Shelly Lundberg, Rchard Startz, Robert Pollak, Terry Wales, Yoram Barzel, Elana Rose, Jutta Joesch, Duncan Thomas, and partcpants of the Labor/Development Brown Bag Seres at the Unversty of Washngton for useful dscussons and comments. Any remanng errors are my own. I am grateful for support from an Alfed P. Sloan Doctoral Dssertaton Fellowshp whle work for an earler draft was beng done. Materal from the Famly Expendture Survey s Crown Copyrght, has been made avalable by the Central Statstcal Offce through the ESRC Data Archve, and has been used by permsson. Nether the CSO nor the ESRC Data Archve bear any responsblty for the analyss or nterpretaton of the data heren.

4 In the past decade and a half, there has been much dscusson of, and many attempts to test, the poolng hypothess mpled by untary models of household decson-makng aganst the set of alternatve collectve models. 1 Untary models assert that a household behaves lke an ndvdual, maxmzng a sngle objectve functon subject to a unfed budget constrant. One mplcaton of these models s that household members pool ther ncome, and that who controls what proporton of that ncome does not affect household demands. Emprcal tests of ncome poolng are typcally plagued by potental bases due to the endogenety of measures of control over ncome n the household. In cross secton data, the magntude and dstrbuton of ncome n the household may be correlated wth the preferences of and prces faced by famly members. Clearly what s needed for a test of poolng s an experment generatng varaton n ntra-household ncome control. In the late 1970s, the U.K. changed the form of ts unversal chld beneft scheme, essentally shftng recept of transfer ncome from fathers to mothers n two-parent famles. Ths natural experment provdes an exogenous source of varaton n the control of resources wthn the famly. Usng aggregated data, grouped by household composton and ncome, Lundberg, Pollak, and Wales (1997) fnd that ratos of chldren s to men s and of women s to men s clothng expendtures n households wth chldren n the U.K ncreased after ths polcy change took effect. In ths paper, I use mcro data to perform a more extensve set of tests for changes n expendture patterns for a broader range of goods. I use prce ndces for each of the three narrowly defned clothng categores, whch s preferred to the aggregate clothng prce ndex used by Lundberg, Pollak, and Wales, whch cannot control for changes n relatve prces of the three goods. 1 Lundberg and Pollak (1996) provde a comprehensve dscusson of both the theoretcal and emprcal lterature. 2

5 Household-level data provde better controls for tme trends, seasonal effects, and household composton than do the aggregated data. However, the hgh frequency of zeroes n household-level expendture data presents a modelng challenge. Zero expendtures may be observed n survey data for several reasons. One possblty s a corner soluton n whch a household may choose not to consume a partcular good gven ther preferences due to realzed values of prces and ncome. The tobt model s the approprately specfed model to deal wth ths possblty. A second possble explanaton for zeroes n the data s nfrequent purchase of some goods whch are durable or non-pershable to some extent. Such goods may be consumed durng a survey perod, but not purchased durng that perod. Budget surveys often collect nformaton about whether households possess certan durable goods n order to deal wth ths problem. However, current expendture s the only nformaton collected for many goods whch are purchased more frequently than, say, refrgerators, but less frequently than every two weeks (the typcal survey perod). If we are assured of some consumpton durng the survey perod, as s the case wth clothng, then the tobt model s not approprate snce t mples exstence of non-consumers. By contrast, purchase-nfrequency models treat zero expendtures as resultng from the durable or non-pershable nature of the good as opposed to non-consumpton due to prce and ncome consderatons. However, as I wll show, the purchase-nfrequency model wll not be adequate n ths case, and an nstrumental-varables approach wll be used to correct for what can be thought of as measurement error. I estmate budget shares for a number of goods categores. I fnd evdence that sgnfcant changes n household expendtures were assocated wth ths polcy change. In examnng a comprehensve set of eleven broad goods categores, the hypothess that expendture shares were unaffected by ths shft s strongly rejected. Whle any sgnfcant change n expendture patterns 3

6 attrbutable to ths polcy change warrants a rejecton of poolng, testng the poolng hypothess usng goods whch are assgnable to partcular household members s comparatvely more appealng. We cannot assgn consumpton of most goods to any one person, and the possblty of nterdependent preferences makes t even more problematc to assgn utlty from consumpton. However, complete assgnablty s not requred. Examnng goods whch are of greater nterest to some household members than to others, such as the three clothng categores examned by Lundberg, Pollak and Wales, provdes an ntutve means of testng the poolng hypothess, as well as questonng whether any shfts n expendture patterns are consstent wth what we would expect n the context of collectve models. I fnd shfts n expendtures on some assgnable goods whch are generally n the drecton we would expect gven the drecton of change n control over household ncome. For nstance, men's clothng and a men's tobacco category - consstng of cgars, ppe tobacco, and snuff products - declne n the budget, whle chldren's clothng, toys, and restaurant and take-away meals ncrease as a share of total expendture. Ths paper s organzed as follows: Secton one provdes some theoretcal and emprcal background motvatng ths research, ncludng a dscusson of the change n the Famly Allowance polcy. Secton two dscusses data; secton three ntroduces the models; secton four presents results and ther mplcatons; and secton fve concludes. I. Background Tradtonal economc theory avods the dffcult ssue of how decsons are made n households by assumng each household maxmzes a sngle objectve functon subject to a unfed budget constrant. Samuelson (1956) suggests that ths objectve functon mght be arrved at through consensus among 4

7 famly members. Becker (1981) proposes t may represent the preferences of an altrust (dctator) n the household. The restrctons mposed by the untary models of Samuelson and Becker are strong ones. The household objectve functon s assumed to have the propertes of a utlty functon, and to be nvarant to who controls resources n the household. One restrcton mpled by such models s that household members pool ther resources. Ths pooled ncome s then used to maxmze the household objectve functon. Ths mples that only total household ncome, not ts dstrbuton, affects household demands. Ths has mportant practcal mplcatons for polces targeted at mprovng the welfare of partcular household members, such as women or chldren. Income poolng results n transfers to targeted household members beng neutralzed by the household allocaton mechansm. The welfare of the targeted member may be mproved, but no more or less than f the transfer had been gven to another household member. Cash transfers, as well as transfers n knd, may be subject to ths reallocaton. The concept of ndvdual ratonal choce, whch les at the heart of mcroeconomc theory, compels us to consder the possblty that each person may have dstnct preferences, and that there may therefore be competng nterests of ndvduals wthn a household. Socal choce theory has demonstrated that aggregatng preferences s no smple matter, and that the aggregate functon arrved at may be senstve to the mechansm of aggregaton. In lght of these crtcsms, f untary models make poor predctons, then alternatve models must be used to descrbe household behavor. 5

8 Recent theoretcal work allows for the possblty that each adult n the household has dstnct preferences, and analyzes how ther competng nterests are reconcled. 2 Models whch do not mpose poolng have been ntroduced, ncludng general collectve models (Chappor 1988, 1992), and both cooperatve and non-cooperatve barganng models due to Manser and Brown (1980), McElroy and Horney (1981), Bergstrom (1996), and Lundberg and Pollak (1993, 1994), among others. These models allow ncome accrung to dfferent famly members to affect household demands dfferently. For nstance, n Nash barganed solutons, each person's (potental) ncome affects hs or her reservaton level of utlty, or threat pont, and therefore hs or her equlbrum level of utlty n the household. Thus, these models allow for more effectve targeted transfer polces. Endogenety of the dstrbuton of ncome wthn the household creates obstacles to testng the poolng hypothess. An obvous example s earned ncome. Its dstrbuton and magntude depend on hours worked by each spouse, whch are determned jontly wth household expendtures. Unearned ncome s also lkely to be endogenous wth respect to household behavor. For nstance, ncome from assets may be correlated wth past labor supply, and thus wth current labor supply as well. Transfers, both publc and prvate, are typcally condtoned on some set of crtera over whch household members exercse some control. Sources of unexpected unearned ncome whch are not subject to these concerns tend to be sporadc and nsgnfcant. Prevous attempts to test the poolng hypothess, ncludng Thomas (1990), Schultz (1990), Phpps and Burton (1998), and Bourgunon, Brownng, Chappor, and Lechene (1993) have used some of these problematc sources of ncome. These tests have generally rejected poolng, but possble endogenety bases call those results nto queston. 2 Carolyn Moehlng (1997) extends ths to nclude chldren as household decson-makers. 6

9 In the Unted Kngdom n the late 1970s, the chld beneft scheme was altered n such a way as to shft ncome from fathers to mothers n two-parent households. Pror to 1977, the unversal chld beneft scheme n the U.K. conssted of a small taxable Famly Allowance payment to the mother, and a more sgnfcant Chld Tax Allowance, whch reduced the amount of taxes wthheld from earned ncome. The latter generally would have ncreased the father s take home pay. Ths two-part program was phased out over the perod Aprl 1977 to Aprl 1979 and replaced by the Chld Beneft, a non-taxable cash payment to the mother. Whle the average amount of the total beneft stayed roughly constant, ths polcy change shfted apparent control of a porton of famly ncome from fathers to mothers. The amount of ncome nvolved s a sgnfcant fracton of the average famly's budget. For example, for a famly wth two chldren n Aprl 1980, the Chld Beneft was approxmately 445 Brtsh pounds per year 3 (approxmately 180 pounds n 1974 currency), or about 8% of male manual earnngs n the UK. 4 The real value of that beneft s 7.5% of average total expendture among famles wth one to three chldren n the data used for ths study. In Aprl 1974, the Famly Allowance pad to the mother n a two-chld famly was 47 pounds. Snce Famly Allowance was taxable and Chld Beneft was not, the dfference of 134 pounds can be vewed as an upper lmt of the amount of ncome shfted to the wfe. Based on ths polcy shft, 5 whch s clearly exogenous wth respect to household expendture decsons, Lundberg, Pollak and Wales (1997) present emprcal evdence aganst the poolng hypothess usng grouped data on household expendture n the U.K. Usng expendture data from before and after the polcy change, they fnd strong evdence that a shft towards relatvely greater expendtures on women s goods and chldren s goods concded wth ths ncome redstrbuton. (p.1) 3 Socal Securty Statstcs (1991), Table G1.01, p U.K. House of Commons Hansard, 14 January 1980, pp See Lundberg, Pollak, and Wales (1995) for a more detaled dscusson of the polcy change. 7

10 Lmtatons of the aggregated data dctated some shortcomngs, such as a concentraton on clothng wthout consderaton of other goods, and lmted demographc controls. Each cell n ths data conssts of the mean of expendtures on a category of goods for all famles n the sample whch fall nto a partcular ncome famly-sze group. Because of the small number of cells n the data, t s dffcult to dstngush effects of the polcy change from tme trends. Another potental confounder n ths study s the 1979 change n the value-added tax (VAT) rate. Ths rate rose from 8% to 15%, and at the same tme, clothng and footwear for young chldren ganed zero-rated status. Ths change, occurrng around the tme of the change n the chld beneft, lkely affected the relatve prce of adult and chldren s clothng. Lundberg, Pollak, and Wales use a prce ndex for the aggregate category, clothng and footwear. Thus, ther results on the rato of chldren s clothng expendtures to men s clothng expendtures may be attrbutable n part to changes n relatve prces for whch they are unable to control. I use separate prce ndces for the three clothng categores n estmates presented here. Snce a change n the relatve prce of chldren's clothng s of prmary concern n connecton wth the VAT change, ths addresses whether mothers appear to allocate more resources to chldren than do fathers, as has been suggested by a number of other studes examnng ntra-household allocaton ssues. Usng mcro data from the Famly Expendture Survey, I estmate budget shares for some narrow goods categores whch may be to some extent assgnable to certan members of the household, ncludng men s, women s, and chldren s clothng and footwear. Snce the polcy change gves wves control over a greater proporton of household ncome, we would expect to see a shft n consumpton toward goods whch are of relatvely greater nterest to wves. Such shfts would warrant a rejecton of poolng, and thus provde evdence aganst the untary model. The nature of these shfts also have 8

11 mportant mplcatons for the effectveness of polces ntended to mprove welfare for targeted household members. II. Data and Sample Constructon The Famly Expendture Survey s conducted annually n the U.K. Ten thousand households are randomly selected each year, approxmately seventy percent of whch complete the survey. One or more face to face ntervews are conducted wth one or more members of each household. In addton to these ntervews, each spender 6 n the household completes a personal spendng dary for a perod of two weeks. From the ntervews and the dares, detaled nformaton on household expendtures, ncome, and household demographcs s compled. Some expendture tems, such as food and clothng, are covered only n the dares. For other tems, such as fuel and housng expendtures, detaled ntervew questons supplement the nformaton from the dares. Expendtures are reported as weekly values. The survey s spread over the entre the year. Ths has two advantages. One s that those collectng and processng the data do so full-tme, year-round. Thus they are generally experenced at collectng data, and are famlar wth ths partcular survey. Ths provdes a hgher level of consstency n the data than that n many other data sets. The second advantage s that seasonal effects can be controlled for. The data also nclude the regon n whch the household s located. There are a total of twelve regon categores: Northern Ireland, Scotland, Wales, and nne regons n England. 6 A spender s defned as anyone at least age 15 pror to 1973, or at least age 16 begnnng n

12 The data are collected by the Offce of Populaton Censuses and Surveys Socal Survey Dvson for the Department of Employment. The data archve s mantaned by the Economc and Socal Research Councl Data Archve at the Unversty of Essex. 7 The prmary purpose of the survey s to provde the weghts for constructon of the Retal Prce Index. However, the data are ntended to also be useful for research purposes. They nclude detaled ncome nformaton by source and by household member. The age and sex of each household member are reported, as well as the number of persons by sex n each of several age categores. I use these categores for chldren (aged under 2 years, aged 2 but less than 5, aged 5 but less than 16, and aged 16 but less than 18) as a general gude for constructng dummy varables to control for household composton. I create ten categores defned by the number of chldren n the household and ther ages. Approxmately 7,000 households complete the Famly Expendture Survey n any gven year. For ths analyss, the sample s lmted to households wth one man, one woman, and one to three chldren. Ths results n a total sample sze of 16,131 households for 8 years of data. I use data from 1973 to 1983 but drop the ntermedate years, , durng whch the chld beneft polcy change was phased n. I create three bnary varables to control for the post-polcy-change perod, one for each famly sze. Ths wll allow for the polcy to affect famles of dfferent szes n a general way. Prce ndces from the Abstract of Natonal Statstcs are used. These ndces are avalable quarterly for 11 broadly defned goods categores, as well as for all goods combned (the Retal Prce Index). In estmates of narrowly-defned clothng expendtures, I also use detaled prce ndces obtaned from the Offce for Natonal Statstcs n the U.K. These ndces are avalable monthly for a 7 Materal from the Famly Expendture Survey s Crown Copyrght, has been made avalable by the Central Statstcal Offce through the ESRC Data Archve, and has been used by permsson. Nether the CSO nor the ESRC Data Archve bear any responsblty for the analyss or nterpretaton of the data reported here. 10

13 number of narrowly defned goods, ncludng men's clothng, women's clothng, and chldren's clothng. The two sets of prce ndces come from the same source, but the latter s aggregated to a lesser degree. Unfortunately, nether set s avalable by regon. III. Models The basc emprcal model s an adaptaton of the Almost Ideal Demand System (AIDS) of Deaton and Muellbauer (1980a,b). The budget share for each good can be estmated equaton by equaton usng: C w = α + γ log p + β log + f( D, H, Q, t ) + u I g g g g (1) where w g s the budget share for good category g for household ; p g s the relatve prce for good category g, computed by dvdng the natonal prce ndex for that category by the Retal Prce Index. 8 C s total household consumpton or expendture. I s the household-specfc prce ndex for all goods, approxmated by log I = w log p. 9 D s a set of dummy varables representng the polcy change, g g g and H s a vector of household demographc varables. Q s the quarter of the year n whch the household was surveyed, t measures a tme trend, and u g s an error term wth standard propertes. Ths model can be estmated equaton by equaton usng ordnary least squares regresson. However, OLS s generally not consstent, for reasons I wll presently explan. Thus, we must explore alternatve estmaton methods. 8 Whle ncluson of all prces s preferred, multcolnearty results f all prces are ncluded due to the short tme seres and absence of cross-sectonal prce varaton. 9 Accordng to Deaton and Muellbauer (1980b) ths approxmaton was used by J. R. N. Stone. No partcular work s cted. I use broad category expendtures and correspondng prce ndces to calculate ths ndex. 11

14 I begn wth clothng as an example. Estmatng clothng consumpton presents a modelng challenge. Whle consumpton of clothng servces s smooth over tme, clothng s purchased nfrequently due to ts durable nature. Ideally we would observe consumpton, but t s not typcally drectly measured, and s arguably not measurable. So we must fnd a way to make nference about consumpton from expendture data. A number of approaches are avalable n the lterature. Purchase-nfrequency models allow factors other than prce and ncome (or total expendture) to generate zeroes n the data through nfrequent purchase of goods whch are non-pershable or durable. More precsely, they allow the probablty of purchase of a good durng the survey perod to be generated by a separate vector of factors from that whch determnes the expendture on the good, or budget share, gven that a purchase occurs (Cragg 1971). The purchase-nfrequency model allows us to make nference about consumpton, whch occurs more-or-less contnuously, n cases where purchases are lumped and therefore nfrequent. Such a model s developed by Blundell and Meghr (1987): y represents observed expendture, whle y * s actual consumpton of a partcular good by household. Let D represent the latent decson to purchase, so that D > 0 f and only f y > 0. P s the probablty of purchase. Thus we have E ( y ) = E( y D > 0) P + E( y D 0)( 1 P ) Assumng E(y ) = E(y * ), ths reduces to * ( y D > 0) P E( y ) E = The relatonshp of y to y * can then be wrtten. (2). (3) P y = y + v = xβ + e + v = x β + u. (4) 12

15 Assumng both error terms, e and v, have zero mean and constant varance, we can wrte the censorng rule for y as follows: y = ( y + v ) P > 0 = 0 otherwse f D (5) Note the dfference between ths censorng rule and that for the tobt model: y = y = 0 f y > 0 otherwse (6) Makng some dstrbutonal assumptons, ths purchase nfrequency model can be estmated by maxmum lkelhood. The model does not allow for corner solutons. Rather, t assumes that the good s consumed n the survey perod, mplyng that * y s strctly postve for all. It s possble to nest ths model wth the tobt model, allowng for corner solutons and n prncple allowng for a test of nonconsumpton. Whle ths may be a desrable strategy for some goods, we can safely assume clothng consumpton to be strctly postve among all households n ths sample durng the survey perod for each of the three clothng categores I consder here. It appears the purchase-nfrequency model would be deal for applcaton to ths problem. However, purchases made on credt pose a serous problem. The total purchase amount s not recorded n the FES data for purchases made on credt. Only cash outlays, such as down-payments and subsequent payments, are observed. Lkely due to the rse n the avalablty of revolvng credt accounts, such as Vsa and store credt card accounts, there s a rsng ncdence of use of credt for clothng purchases n the data over the perod studed here. Furthermore, ths trend s coupled wth a shft n the type of household usng credt for clothng purchase. The mean real total expendture of households who buy clothng on credt and smlar pay-over-tme mechansms n the post-polcy-change 13

16 perod s substantally hgher than those who do so before. It may be that lower ncome households are usng somethng lke layaway plans n the earler perod (pay over tme and consume later), whle more affluent households are choosng to buy under more contemporary credt arrangements (consume now and pay later) n the later perod. Because only cash outlay and not total purchase s reported for credt expendtures, the purchase-nfrequency model s msspecfed f we use all of the clothng expendture data avalable. However, lmtng the sample to non-credt purchases, or alternatvely to households who do not record any credt clothng expendture, results n constructon of an nvald artfcal panel dataset for the purpose of testng for an effect of the polcy change on consumpton. Due to the dfferental trends n use of credt across ncome groups, we cannot be assured that a non-credt sample would consst of smlar households before versus after the polcy change. As above, let us assume that expendture and consumpton are equal on average for a suffcently long tme perod, n essence that E(y g ) = E(y * g ) for household, good g. If we observe expendture by a household for a random two-week perod, we can wrte the relatonshp of observed expendture to unobserved consumpton as y * g=y g +e g, where e g can be thought of as a mean-zero measurement error. Ths mples that total expendture wll be hgher n perods when clothng s purchased than n perods when t s not. Normally, measurement error n the dependent varable s not a concern, but e g wll also enter n total consumpton, whch s a regressor. For smplcty, suppose only consumpton of good 1 s measured wth error, so that total consumpton equals total expendture plus e 1 : Total Consumpton C y = y + e where Total Expendture E y * g g 1 g g g g 14

17 We can substtute nto equaton (1) and rearrange terms to obtan: y E + e E + e = α + γ log p + β log + f( D, H, Q, t ) + ν E E P where v e = u ( u 1) 1 1 E Ths gves us the measured budget share on the left-hand sde. Note that the regressor, total consumpton, s correlated wth the error term through the presence of e 1 n both terms. We can nstrument for total expendture where t s used as a regressor ntended to measure total consumpton, arguably gettng closer to a true measure of total consumpton and dong away wth some of the measurement error. However, some correlaton between the regressor and error term wll reman because total expendture appears n the denomnator of the latter. Further, estmated coeffcents on regressors wll be scaled by total consumpton over total expendture. However, as long as these are equal on average, ths rato s unty. Because the specfcaton of the AIDS model calls for the consumpton share and not the level of expendture to be estmated, elmnaton of measurement error bas n estmates s not assured by nstrumentng for total expendture. Note that f we estmate consumpton levels nstead, our equaton would be the followng: E + e y = α + γ logp + β log + f( D, H, Q, t ) + ε I where ε 1 = u 1 e 1 In ths case, nstrumentng for total expendture s suffcent to elmnate the bas, gven that nstruments used are approprate. Keen (1986) uses normal ncome (based on usual rather than current earnngs) to nstrument for total expendture n estmatng Engel curves usng a sngle year of FES data, and shows that ths produces consstent estmates. If macroeconomc condtons are constant, then consumpton should be 15

18 approxmately proportonal to normal ncome. However, changes n macroeconomc condtons may affect savng rates and therefore the proporton of normal ncome consumed n a gven perod. I use normal ncome and regonal quarterly unemployment rates as nstruments for total expendture. Because normal ncome s zero n some cases, a log specfcaton, to match that for total expendture, s not feasble. I fnd that a ffth-order polynomal n normal ncome and a thrd-order polynomal n unemployment rate provdes the best ft. These varables are used as nstruments n twostage least-squares estmates of both the budget share and the level of expendture for each of the three clothng categores If clothng s separable from other goods, and f clothng s the only good for whch consumpton and expendture durng the survey perod may dffer, then standard OLS and Tobt models wll suffce for other goods. These are strong assumptons. Clearly, other goods also suffer from purchase nfrequency. These may nclude durables, harcuts, toys, and books to name a few. Therefore nstrumental varables methods wll also be used n estmatng expendture shares and levels for eleven broadly defned goods categores and nne addtonal narrowly defned goods.. In cases where a tobt model should be used, I estmate total expendture usng the nstruments dscussed above as regressors n an OLS model and enter the estmated total expendture n the Tobt model. I ll refer to ths as a twostage tobt model. 10 OLS estmates of total expendture usng the nstruments as regressors s shown n Appendx A. All regressors are sgnfcant at the one percent level or better, wth the excepton of the quadratc normal ncome term. The adjusted R squared for the model s Recall that n order for the measurement error to be mean zero, we requre a suffcently long 10 Standard errors have not been corrected n two-stage tobt estmates. Ths s unlkely to affect nference for the polcy change varables whch are the focus of ths paper. 16

19 tme perod so that average consumpton equals average expendture. The length of ths tme perod may dffer for dfferent goods. In partcular, I wll assume that three months s not a suffcently long perod for clothng, and so wll not control for quarter of the year n clothng estmates. 11 Though I do control for quarter n estmates of other goods, I check whether effects of the polcy change are affected by ncluson versus excluson of these varables for goods for whch a quarter s arguably too short to assume expendture equals consumpton, such as durables, books, and toys. An addtonal method wll also be used whch allows estmaton of the budget share whle mnmzng concerns about appearance of total expendture n the denomnator of the error term. I subtract clothng and durables expendtures from total expendture and use the result to proxy for total expendture, both on the rght-hand sde and n the share calculaton. Standard OLS and Tobt models are then used to estmate budget shares. These results are compared to IV results as a check of robustness. IV. Results I estmate budget shares and expendture levels for eleven broadly defned goods categores and twelve narrowly defned goods, ncludng three assgnable clothng categores. Table 1 presents summary statstcs, ncludng means of total expendture and ncome measures, and expendture shares and levels for the goods of nterest. The number of households reportng postve expendtures n each good category s also gven. The dstrbuton of famles across the three famly-sze categores s shown. 11 Clothng consumpton need not be greater n colder months. We may wear more clothes at once n colder months, but may use as many clothng servces n warm months due to summer actvtes and more frequent washng, especally n the case of chldren s clothes. 17

20 Almost half of famles n the sample have two chldren. Thus, the omtted household composton category n estmates s the most common chld-age category among two-chld famles. Table 2 presents results for nstrumental varable estmates of the eleven broad goods categores. Panels (a) and (c) show budget share estmates whle panels (b) and (d) show expendture level estmates. Because more than fve percent of households report zero expendtures on alcohol, clothng, food out and tobacco, both two-stage least sqares and two-stage tobt estmates are reported for these two categores (panels c and d). TSLS estmates are reported for the remanng seven categores (panels a and b): durables, food for home preparaton (food n), fuel and power (not ncludng fuel for automobles), housng, servces, transportaton (ncludng petrol), and mscellaneous goods. The three polcy shft varables (for one-, two-, and three-chld famles) had a postve and sgnfcant effect on expendture shares and levels for clothng, food out, and mscellaneous goods. It had negatve and sgnfcant effects on expendture shares and levels for housng. There s also some evdence of declne n alcohol expendtures, partcularly n three-chld famles, although ths result s not sgnfcant n tobt models, whch are theoretcally more appealng gven that there may be nonconsumng households. Estmates for broadly-defned goods are done equaton by equaton, and no constrant on addng up of shares s mposed. 12 F-tests and lkelhood rato tests were used for each of the broad goods to test the jont hypothess that the three polcy shft dummes had no effect. These tests ndcate that the polcy shft dd sgnfcantly affect budget shares for several of the categores, ncreasng clothng, food out, and mscellaneous goods expendtures and decreasng expendtures on housng and food n. The durables estmates were also done excludng controls for quarter of the year and the results on 18

21 polcy varables, whch are not statstcally dfferent from zero, were vrtually unchanged. Table 2 also reports a ch-sqare statstc based on a seemngly unrelated regressons system of equatons (estmated wth nstruments usng three-stage least quares) testng the jont hypothess that effects of the three polcy varables n all eleven equatons are zero. Ths hypothess s strongly rejected, ndcatng that the polcy shft dd sgnfcantly change overall expendture patterns. Phpps and Burton (1998) and Hoddnott and Haddad (1991) found negatve effects of the wfe s relatve ncome on tobacco and alcohol. I fnd no such effects for tobacco, but fnd margnal negatve effects for alcohol. Phpps and Burton also found postve effects of the wfe's ncome on restaurant meals. Ths could be nterpreted as a prce effect snce the study uses earned ncome as the key explanatory varable. The exogenety of the polcy change n the present analyss rules out such an nterpretaton here of the strong postve effect on food out, whch ncludes restaurant meals and takeaway food. An alternate method n whch total expendture less clothng and durables expendtures s used as a proxy measure of total consumpton n estmatng budget shares yelds somewhat smlar results (shown n Table B1 of Appendx B) to those on broadly-defned goods usng IV methods. In general, sgns of coeffcents on the polcy varables are the same, but statstcal sgnfcance n most cases s stronger. Some dfferences from the IV models should be hghlghted: Here, the negatve coeffcents for tobacco n one- and two-chld famles s sgnfcant n the OLS model, but not the tobt model. There are clearly many households whch are non-consumers, so ths result s not convncng. Alcohol does declne sgnfcantly here for one- and three-chld famles n the tobt model, and the negatve coeffcents n the food n equaton are sgnfcant for all famly szes. Two stark dfferences emerge 12 The sum of TSLS polcy change coeffcents for budget shares for a gven sze famly s approxmately zero 19

22 effects on durables here are postve and sgnfcant and effects on fuel and power negatve and sgnfcant. Ths dffers markedly from the IV models, and nvolves swtchng sgns of coeffcents, although they were far from statstcally dfferent from zero n the earler models. Agan, a seemngly unrelated regressons system of equatons s used to test the jont hypothess that polcy varables n all equatons have zero effect, and agan the hypothess s strongly rejected (p<.001). Table 3 reports results from expendture share equatons on nne narrowly defned goods. All of these goods had more than fve percent zero expendtures reported, so both OLS and tobt results are presented. For goods whch are arguably consumed by most or all households n the sample but whch are purchased nfrequently, least squares estmates are theoretcally more appealng, as dscussed earler. However, t may be that, for some of these goods, there s both non-consumpton and nfrequent purchase (e.g., pets expendtures). A model whch dstngushes these emprcally cannot be convncngly dentfed wth ths data. I therefore present both two-stage least quares and two-stage tobt estmates so that the reader can judge dfferences and smlartes n results. Cgarettes expendtures do not appear to have been affected by the polcy change, but other tobacco expendtures dd declne, sgnfcantly for two-chld famles, whch comprse almost half of the sample. The jont test of all three polcy varables havng no effect on other tobacco s rejected n the levels estmates. Ths category ncludes cgars and ppe tobacco, almost exclusvely consumed by men, whereas cgarettes are a farly gender-neutral good. Domestc servces, whch mght be vewed as a substtute for the wfe s home producton, appear to have ncreased for one-chld famles per the tobt results, but declned for two-chld famles per the least squares model. There are most certanly non-consumers, but purchase frequency may also play a role. The sgnfcant jont test statstcs are puzzlng and not very nformatve gven that coeffcent 20

23 sgns dffer for dfferent famly szes. Cosmetcs expendtures declne sgnfcantly. Whle these goods are drectly consumed mostly by women, one could make arguments for ether a declne or ncrease n consumpton n the context of collectve models. Hardressng would nclude trps to the barber or stylst for all famly members, and does not appear to be affected by the polcy change. In panels (c) and (d), we see a postve effect on books, but ths s only sgnfcant n tobt models. Chldren s books cannot be dstngushed from adult books n the data. The books category also ncludes some other readng materals, such as perodcals. Whle t s unlkely there are many households n ths sample who don t consume n ths category, consumpton may be va borrowed lbrary materals, so that expendture and consumpton do not relate to one another n the way I have specfed above. There s also most probably an mportant frequency of purchase element here. Pets expendtures appear unaffected, whle toys are ncreasng for one- and two-chld famles. There are lkely few non-consumers of toys n ths sample, but purchase frequency s lkely to be mportant. Chldren s pocket money s ncreasng, sgnfcantly for one- and (n one model) two-chld famles. Results from proxed OLS models for these nne narrowly-defned goods are qute smlar to those n nstrumented estmates (see Table B2 of Appendx B). Table 4 reports two-stage least squares shares and levels estmates, and proxed OLS estmates for the three narrow clothng categores. All three models show strong and statstcally sgnfcant postve effects of the polcy change on chldren's clothng. Polcy change coeffcents are consstently negatve for men's clothng n all three models, and the effect s statstcally sgnfcant for three-chld famles n both nstrumented models. The coeffcents are consstently postve for women s clothng across all three models, and are statstcally sgnfcant n two- and three-chld famles n the proxed 21

24 OLS model. A test of the jont hypothess for all three polcy varables for women s clothng shows a statstcally sgnfcant effect n both the nstrumented share estmates and the proxed share estmates. Results found here on men s, women's, and chldren s clothng are consstent wth those found by Lundberg, Pollak, and Wales (1997). They found that the rato of chldren's clothng expendtures to men's clothng expendtures ncreased when the chld beneft polcy changed. They also found an ncrease n the rato of expendtures on women's clothng to that on men's clothng. As dscussed earler, ther results may be attrbutable n part to the 1979 change n VAT rates on chldren s relatve to adult clothng. 13 Ther declne n men s clothng relatve to women s clothng s unlkely to have been drven by the change n tax rates, snce the tax rate change on both goods was the same, but the tax nclusve prce of chldren's clothng relatve to adult clothng may have declned sgnfcantly. Thus, ther result on chldren's relatve to men's clothng expendtures s of partcular concern. The results I have presented here are compellng gven that narrow prce ndces wth VAT ncluded were used to control for prces of the three separate clothng categores. These results demonstrate that the ncrease n the relatve share of chldren's goods they fnd was not solely due to tax rate changes, but s attrbutable n part to the chld beneft polcy change, and to the apparent tendency of mothers to allocate more of the famly budget to chldren's goods. The magntudes of changes n budget shares these results suggest are certanly plausble gven the amount of ncome beng shfted from men to women. For nstance, for the average two-chld famly, the change n expendtures on the broad category "food out" attrbutable to the polcy change s 26 per year. As a result of the polcy change, the same famly decreased ts annual housng expendture by VAT rates were 10% n 1973, changng to 8% n 1974, and to 15% n Young chldren's clothng and footwear became zero-rated n (Prest, 1980) 22

25 and men s tobacco expendture by 3, and ncreased chldren s clothng expendtures by 16 and toys expendtures by 5. These are reasonable magntudes gven that approxmately 134 per year was shfted from husband to wfe. 14 V. Concluson I use mcro data to examne how budget shares changed n two-parent famles wth chldren n the U.K. when a change n Famly Allowance polcy essentally shfted transfer ncome from fathers to mothers. Shfts n expendture patterns due to ths change are nconsstent wth the untary model of household decsons. I fnd sgnfcant changes n expendtures on broadly defned goods, as well as on some narrowly defned assgnable and quas-assgnable goods. Although the broad goods categores do not represent assgnable goods, a dscrete change n shares n the before versus after perod whch s not attrbutable to other factors ndcates that there has been a shft of power over decson makng n the household, and that a systematc dfference n preferences over allocaton of household ncome exsts between husbands and wves. Closer examnaton of expendtures on narrow goods wthn these broad categores produces further nsghts. These shfts ndcate that women and chldren benefted at the expense of men when ths new polcy took effect. Results on narrow goods generally pont to an ncrease n the budget share allocated to some women s and chldren s goods, along wth a declne n that for some men s goods. 15 Such changes n consumpton patterns resultng from a shft n the control of ncome n the household adds evdence to 14 The quoted magntudes of change are n 1974 currency, and are based on TSLS budget share estmates. 15 One noteable excepton s the negatve effect on cosmetcs expendtures. 23

26 the growng case aganst the untary model evdence whch s partcularly convncng gven the exogenety of ths polcy shft. Rejecton of the untary model has mportant mplcatons for the effectveness of polces amed at mprovng welfare of targeted members of households. The untary model mples neutralzaton of targeted transfers. Some alternatve collectve models allow for more effectve transfer polcy. As Alderman, et. al. put t, t may be "tme to shft the burden of proof." Analyzng the household allocaton process wthn the framework of collectve models wll allow the development and applcaton of more effectve polces n areas of health, educaton, and welfare. References Alderman, Harold, Perre-Andre Chappor, Lawrence Haddad, John Hoddnott, and Rav Kanbur. "Untary versus Collectve Models of the Household: Is It Tme to Shft the Burden of Proof?" World Bank Research Observer 10:1, February 1995, Becker, Gary (1981), A Treatse on the Famly. Cambrdge, Mass.: Harvard Unversty Press; enlarged edton Bergstrom, Theodore (1996), "A Survey of Theores of the Famly," n Handbook of Populaton and Famly Economcs, ed. Mark R. Rosenzweg and Oded Stark, Amsterdam: North-Holland. Blundell, Rchard and Costas Meghr (1987), "Bvarate Alternatves to the Tobt Model," Journal of Econometrcs 34: Bourgugnon, Francos, Martn Brownng, Perre-Andre Chappor, and Valere Lechene (1993), " Intra Household Allocaton of Consumpton: A Model and Some Evdence from French Data," Annales d'econome et de Statstque 29: Chappor, Perre-Andre (1988), Ratonal Household Labor Supply, Econometrca, 56, Chappor, Perre-Andre (1992), Collectve Labor Supply and Welfare, Journal of Poltcal Economy, 100,

27 Cragg, John C. (1971), Some Statstcal Models for Lmted Dependent Varables wth Applcaton to the Demand for Durable Goods, Econometrca 39(5): Deaton, Angus and John Muellbauer (1980a), An Almost Ideal Demand System, Amercan Economc Revew, 70, Deaton, Angus and John Muellbauer (1980b), Economcs and Consumer Behavor. New York, N. Y.: Cambrdge Unversty Press. Deaton, Angus and Margaret Irsh (1982), "Statstcal Models for Zero Expendtures n Household Budgets," Journal of Publc Economcs 23: Hoddnott, John and Lawrence Haddad (1995), Does Female Income Share Influence Household Expendtures: Evdence from The Cote d Ivore, Oxford Bulletn of Economcs and Statstcs, 57:1 (February), Kay, J.A., M.J. Keen, and C.N. Morrs (1984), "Estmatng Consumpton from Expendture Data," Journal of Publc Economcs 23: Keen, Mchael (1986), Zero Expendtures and the Estmaton of Engel Curves, Journal of Appled Econometrcs 1, Kemsley, W.F.F., Famly Expendture Survey - Handbook on the Sample, Feldwork, and Codng Procedures. (1969, and revsed edton 1980), London: HMSO. Lundberg, Shelly J., and Robert A. Pollak, (1993), Separate Spheres Barganng and the Marrage Market, Journal of Poltcal Economy, 101:6 (December), Lundberg, Shelly J., and Robert A. Pollak, (1994), Noncooperatve Barganng Models of Marrage, Amercan Economc Revew Papers and Proceedng 84:2, Lundberg, Shelly J., and Robert A. Pollak (1996), Barganng and Dstrbuton n Marrage, Journal of Economc Perspectves 10:4, Lundberg, Shelly J., Robert A. Pollak, and Terry J. Wales (1997), Do Husbands and Wves Pool Ther Resources? Evdence from the U.K. Chld Beneft, Journal of Human Resources 32 (Summer). Manser, Marlyn and Murray Brown (1980), Marrage and Household Decson makng: A Barganng Analyss, Internatonal Economc Revew, 21:1 (February), McElroy, Marjore B. and Mary Jean Horney (1981), Nash Barganed Household Decsons, Internatonal Economc Revew, 22:2 (June),

28 Meghr, Costas and Jean-Marc Robn (1992) "Frequency of Purchases and the Estmaton of Demand Systems," Journal of Econometrcs 53: Moehlng, Carolyn (1997) "The Impact of Chldren's Income on Intrahousehold Resource Allocatons," prepared for the NBER/DAE 1997 Summer Insttute. Phpps, Shelley and Peter Burton (1998), What s Mne s Yours? The Influence of Male and Female Incomes on Patterns of Household Expendture, Economca 65(260): Pollak, Robert (1994), Takng Power Serously, mmeo, Unversty of Washngton. Prest, A.R. (1980) Value Added Taxaton: The Experence of the Unted Kngdom, Washngton, D.C.: Amercan Enterprse Insttute for Publc Polcy Research. Robn, Jean-Marc (1993) "Economc Analyss of the Short-Run Fluctuatons of Households' Purchases," Journal of Economc Studes 60: Samuelson, Paul A. (1956), Socal Indfference Curves, Quarterly Journal of Economcs 70 (February): Schultz, T. Paul (1990), Testng the Neoclasscal Model of Famly Labor Supply and Fertlty, Journal of Human Resources, 25(4) Thomas, Duncan (1990), Intra-Household Resource Allocaton: An Inferental Approach, Journal of Human Resources, 25(4) U.K. Central Statstcal Offce, Annual Abstract of Statstcs, , HMSO, London. U.K. Department of Employment and Productvty, Famly Expendture Survey, , , HMSO, London. U.K. Department of Health and Socal Securty, Socal Securty Statstcs, 1991, HMSO, London. U.K. House of Commons Hansard, January Wales, T.J. and A.D. Woodland (1983) "Estmaton of Consumer Demand Systems wth Bndng Non- Negatvty Constrants," Journal of Econometrcs 21:

29 Table 1: Summary Statstcs Share of Budget Real Level # postve Varable Mean Std Dev Mean Std Dev observatons Broadly-defned goods Alcohol (0.047) (2.745) Clothng (0.076) (5.205) Durables (0.091) (9.932) Food n (0.087) (3.759) Food out (0.032) (2.035) Fuel, power, & lght (0.041) (1.991) Housng (0.079) (7.173) Msc (0.056) (4.458) Sevces (0.082) (9.520) Tobacco (0.043) (1.675) Transport (0.111) (11.530) Narrowly-defned goods Women s clothng (0.042) (2.740) Men s clothng (0.041) (2.695) 5906 Kd s clothng (0.039) (2.176) Cgarettes (0.043) (1.659) 9429 Other tobacco (0.010) (0.434) 2457 Domestc servces (0.015) (0.899) 2362 Cosmetcs (0.009) (0.509) 9816 Hardressng (0.010) (0.519) 5805 Books (0.010) (0.694) 5676 Toys (0.022) (1.239) 5954 Pets expendture (0.018) (0.978) 8167 Chldren s pocket money (0.009) (0.483) 5987 Real total expendture (28.086) Log real total exp (0.436) Predcted log real expendture (0.262) Real total expendture less clothng & durables (18.205) Real normal ncome (from earnngs) (37.864) N = # chldren n famly Frequency Percent Means and standard devatons are for full sample. Expendture levels are per week n 1974 currency. Mean shares are computed usng total expendture. Means of shares computed usng expendture excludng clothng and durables are proportonally larger, and are not constraned to be less than or equal to unty. 27