Is It Better to Be a Boy? A Disaggregated Outlay Equivalent Analysis of Gender Bias in Papua New Guinea

Transcription

1 Department of Agrcultural and Resource Economcs Unversty of Calforna Davs Is It Better to Be a Boy? A Dsaggregated Outlay Equvalent Analyss of Gender Bas n Papua New Gunea by John Gbson and Scott Rozelle Workng Paper No December, John Gbson and Scott Rozelle All Rghts Reserved. Readers May Make Verbatm Copes Of Ths Document For Non-Commercal Purposes By Any Means, Provded That Ths Copyrght Notce Appears On All Such Copes. Calforna Agrcultural Experment Staton Gannn Foundaton for Agrcultural Economcs

2 Is It Better to Be a Boy? A Dsaggregated Outlay Equvalent Analyss of Gender Bas n Papua New Gunea John Gbson 1 Unversty of Wakato Scott Rozelle Unversty of Calforna, Davs Abstract Dscrmnaton n the allocaton of goods between boys and grls wthn households n Papua New Gunea s examned usng Deaton s (1989) outlay-equvalent rato method. Addng a boy to the household reduces expendture on adult goods by as much as would a nne-tenths reducton n total outlay per member, but grls have no effect on adult goods expendture. The hypothess of Haddad and Reardon (1993) that gender bas s nversely related to the mportance of female labour n agrcultural producton s not supported. There s no evdence of bas aganst grls n the urban sector. JEL: D12, J16 Keywords: Boy-grl dscrmnaton, Gender bas, Outlay-equvalent analyss 1 Address for correspondence: Department of Economcs, Unversty of Wakato, Prvate Bag 3105, Hamlton, New Zealand. Fax: (64-7) E-mal: jkgbson@wakato.ac.nz.

3 IS IT BETTER TO BE A BOY? A DISAGGREGATED OUTLAY EQUIVALENT ANALYSIS OF GENDER BIAS IN PAPUA NEW GUINEA I. Introducton The queston of unequal dvsons between men and women and boys and grls nsde the household s the topc of much recent research. Unequal allocatons cause poverty and nequalty to be understated when we use measures that assume that every household member s treated evenly (Haddad and Kanbur, 1990). The demographc evdence of mssng women n Bangladesh, Chna, Inda, and Pakstan provdes further mpetus for ths lne of research (Dreze and Sen, 1989). However, the dffculty of observng the nner workngs of households restrcts many studes to externally observable outcomes lke health and nutrton (Sen, 1984). Studes of ndvdual consumpton wthn the household, such as Ptt, Rosenzweg and Hassan (1990) on food n Bangladesh, are less common because they need data that are dffcult and expensve to obtan n a relable manner. Because of the lack of accurate ntra-household data, nferental methods of detectng gender bas usng just standard household expendture are attractve. Deaton (1989) ntroduced a method for usng household expendture data to nfer dscrmnaton n the allocaton of goods between boys and grls. The method starts, somewhat paradoxcally, wth expendtures on goods that are not consumed by chldren--for example beer, gamblng, and tobacco products. The addton of a chld to the famly s treated as a pure (negatve) ncome effect on the demand for these goods, snce the chld does not contrbute to famly ncome and also s unlkely to partcpate n the consumpton of these goods. Evdence for gender bas can then be found by (statstcally) answerng the queston: Is the reducton n spendng on these adult goods larger when the addtonal chld s a boy rather than a grl? 1

4 Despte the cleverness of Deaton s nferental method, the results have been dsappontng. Curously, n ts few applcatons, the method sometmes fnds bas aganst grls n locatons where t s not expected but no bas n places where other evdence strongly suggests that males are favoured. For example, Ahmad and Morduch (1993) fnd no evdence of gender bas n Bangladesh household expendture data, whereas the mbalance n sex ratos suggests consderable dscrmnaton aganst grls. In Indan states where excess mortalty and low lteracy pont to bas aganst grls, Subramanan (1994) fnds no bas n household expendtures aganst grls; however, n Maharashtra, a state n whch socal ndcators suggest the status of grls s relatvely better, the expendture method ndcates there s a bas. Deaton (1997) lsts other examples from Pakstan and Tawan where the expendture method fnds no bas despte other studes ndcatng a preference for boys n these countres. Several explanatons have been advanced for the falure of the adult goods method to detect bas. One econometrc reason may be sample truncaton n cases where grls have been so dscrmnated aganst that they have ded (Udry, 1997). Alternatvely, grls may have hgher needs than boys so that equal allocatons stll adversely affect grls (Ahmad and Morduch, 1993). Crtcal nterventons mght not be made for grls when they are made for boys. A further reason for the falure of the method s that t has been used n some places that are not lkely canddates for bas, such as Côte d Ivore (Deaton, 1989) and Burkna Faso (Haddad and Reardon, 1993). In these countres women are economcally productve so grls are not seen as a burden on ther parents. Haddad and Reardon advance ths argument further by searchng for dfferences n the degree of gender bas across agroecologcal zones. Ther hypothess, derved from Boserup (1970) and Rosenzweg and Schultz (1982), s that dscrmnaton aganst grls wll be less as the economc 2

5 opportuntes for women ncrease (such as n urban areas or n agrcultural settngs wth hgh ncome potental), suggestng that the greatest dscrmnaton wll be n rural areas where the potental ncome from agrculture s low. Accordng to Haddad and Reardon (1993), dscrmnaton n households s expected to declne as the study area moves to hgher potental rural areas; declnng even further when the study examnes urban areas. In ths paper we provde an example n whch the adult goods method detects gender bas n a place, Papua New Gunea (PNG), where, ex ante, we expect there to be bas. As wll be argued below, there are several reasons for our strong results. Frst, a number of economc ndcators and the weght of other socal scence work suggest that males are favoured n PNG. Second, adult goods consumpton, for a varety of reasons, s mportant n ths settng (more than 12 percent of the budget), and ts hgh level may make the results more easly detectable. Thrd, the enormous heterogenety n the sample data may facltate dentfcaton of these effects; we have a natonally representatve sample n a country that s almost certanly home to the most dverse set of ethnc and socal groups of humans n the world. Fourth, unlke prevous studes (to the best of our knowledge), we desgned and managed the collecton of the data ourselves, plannng from the start to undertake adult goods analyss. As such, the questonnare contans a set of well-defned categores for adult goods, and enumerators were taught to exercse extra care when collectng these data. Fnally, unlke some prevous analyses, we use econometrc technques and statstcal tests that control for the survey desgn effects n our data. To examne dscrmnaton n PNG, the rest of the paper s organzed as follows. Secton II brefly examnes gender bas n PNG. The two next sectons revew the methodology and descrbe the data and estmaton approach. Secton V contans the results. We look frst at gender 3

6 dscrmnaton, n general, and then further examne ths queston by searchng for dfferences n the degree of gender bas across agroecologcal zones and between rural and urban areas. II. Gender Bas n PNG A preponderance of evdence has been amassed by socal scentsts that boys are hghly favoured n PNG. The male share of the populaton s 51.6 percent, a demographc mbalance comparable to Inda and Pakstan (UN, 1996). The lteracy rate for adult females, at only 44 percent, s 18 percentage ponts lower than for males (Gbson, 2001) and 49 percent of women have never attended school (nearly twce the rate for men World Bank, 1999). Women and grls n PNG also appear to be dsadvantaged n health and nutrton (Groos and Garner, 1988). Ths bas s nterestng, and as such perhaps deep-seated, because t exsts n a country where women have tradtonally been thought to be economcally productve female labour s often consdered to be a scarce factor n agrcultural producton n relatve land-rch PNG. Postve brde-prces and the prevalence of polygamy are symptoms of ths scarcty. Thus, PNG appears to be a counterexample to the vew advanced by Haddad and Reardon that bas aganst grls may be ratonal n the sense that t can be explaned by factor endowments and producton relatons n agrculture. The mportance of adult goods n PNG household budgets and the concentraton of cash earnngs n the hands of adult males may mean that the adult goods method gves an unusually clear vew of any tensons over ntra-household consumpton allocatons. Adult goods potentally account for 12 to 13 percent of the household budget n PNG, a fgure that s robust across agro-ecologcal zones (Table 1, row 1). In comparson, n one country where adult goods analyss has prevously been used -- Burkna Faso -- the budget share of canddate adult goods s 4

7 reported as less than 10 percent but lkely s even lower snce many of these goods do not meet the statstcal crtera for true adult goods (Haddad and Reardon, 1993). Budget shares for a smlar array of adult goods n poor areas of other countres, such as Chna, are also sgnfcantly lower (below 8 percent--zhang, 1999). Whle true n many countres, the cash for purchasng both adult and other goods s unequally concentrated n the hands of male adults n PNG, even though women do play an mportant role n many economc actvtes. For example, women make up 87 percent of the partcpants n sweet potato producton, the domnant staple crop, and 83 percent for other food crops (World Bank, 1999). However, men more frequently perform cash earnng jobs (comprsng over 80 percent of the partcpants), and undertake more rewardng cash croppng actvtes (comprsng well over half of the partcpants). Even wthn cash croppng actvtes there appears to be dscrmnaton aganst women; the returns to men s own-account coffee producton are up to twce as hgh as the returns to women (Overfeld, 1998). Costs assocated wth rasng chldren n PNG, especally those who are school-aged (typcally from age 7 onwards), may exacerbate tensons over the budgetary needs of chldren. Although school and health fees at one tme were hghly subsdsed by the government, by the tme of the survey, 1995, most of these had been elmnated through fscal cuts, structural adjustment shfts, and bureaucratc neglect (Jarrett and Anderson, 1989). For example, the average outlay of a poor famly for a sngle secondary school student s equvalent to 36 percent of the household s non-food expendture (World Bank, 1999). Wth a devalung currency, chldren s clothng and shoes, most of whch are mported, have become more expensve. In short, chldren, as every parent knows, can be expensve to rase, and almost certanly mean 5

8 sacrfce n other areas. Hence, gven these hgher costs, n a socety wth a revealed preference towards boys (as shown by other ndcators), wth a hgh propensty to consume adult goods, and wth a large part of the cash n the hands of males, f the adult goods analyss can be statstcally mplemented, we should expect to fnd t revealng gender bas n PNG. However, n areas n whch an nvestment n a grl can rase her future ncome earnng potental, the extent of the dscrmnaton aganst grls may perhaps be attenuated. III. Methods The frst step s to dentfy the set of adult goods. Followng Deaton, Ruz-Castllo and Thomas (1989) one can use the lnear model p q = b 0 +b 1 x G + c j n + d? z + v j (1) where pq s expendture on canddate adult good, xg s total expendture on adult goods, nj s the number of people n each of eght age and gender classes, z s a vector of control varables, and v s a random error. Chldren affect only the total expendture allocated to adult goods, not the allocaton wthn the adult goods group once total group expendture s gven. Therefore, f good s a genune adult good, the age and gender of chldren should play no part n equaton (1), whch can be tested wth an F-test on the relevant estmated cj coeffcents (.e., those coeffcents on the varables assocated wth the age categores for chldren). An alternatve test of whether a lst of canddate goods contans only genune adult goods uses the concept of "outlay-equvalent ratos." For any normal commodty and demographc category r, the outlay-equvalent rato r s: 6

9 ƒ( p q )/ ƒ n = ƒ( p q )/ ƒx r r? n x (2) where r s an expresson that measures the effect of an addtonal person of type r on the demand for good, measured as the percentage change n outlay (expendture) per person that would have been necessary to produce the same effect on demand. The effect of an addtonal chld should be that adult goods expendture falls, gven the extra needs that the chld brngs. Because ths effect s exactly lke a reducton n ncome, the reduced expendture on each ndvdual adult good ought to be n proporton to the margnal propenstes to spend on each good (Deaton, 1989). Hence, the π-ratos for any partcular type of chld should be the same across all of the adult goods, so ths mplcaton provdes another test for whether the canddate goods are ndeed adult goods. The outlay equvalent ratos can be calculated from the coeffcents produced for any estmated Engel curve. Followng Deaton (1989), the Engel curve used here s specfed as w = p x q = + x ln + n lnn+ J -1 j=1 j n j + n? z +u (3) where w s the budget share for the th adult good, x s the value of total household consumpton, n s household sze, nj s the number of people n the jth demographc group, and z s a vector of control varables. The estmated parameters of equaton (3) can then be used to calculate r ( = - )+ r - + J j=1 w j ( n j / n) (4) 7

10 for r = 1,.., J, where γ J s defned to be zero. Sample means can be used for the w and the nj /n ratos. The calculated π-ratos can then be used to test the null hypothess of even treatment for boys and grls of a certan age class, H 0 : j = k (5) for all adult goods, where j refers to boys and k refers to grls. In addton to ths test across demographc categores, the mplcaton of equalty of π-ratos across adult goods and wthn (chld) demographc categores can also be tested: H 0 : r = jr (6) for all and j that refer to adult goods and for all r demographc categores. General procedures for dervng the test statstcs and the standard errors of the π-ratos are descrbed by Deaton, Ruz- Castllo, and Thomas (1989). IV. Data and Estmaton Issues Data used n ths paper come from the Papua New Gunea Household Survey (PNGHS), the frst naton-wde household consumpton survey ever conducted n PNG. The survey desgn and enumeraton, whch were carred out by the authors n 1995 and 1996, covered a random sample of 1200 households, resdng n 120 rural and urban communtes ( clusters ), who were ntervewed between January and December The survey team selected clusters from the enumeraton areas of the 1990 Census, stratfyng the sample by sector (urban and rural), by envronmental condtons (elevaton and ranfall), and by the level of agrcultural development. 2 2 Ths was establshed from an agrcultural mappng project (Allen, Bourke and Hde, 1995). 8

11 The stratfcaton allows us to create sub-samples that correspond to the agro-ecologcal zones used by Haddad and Reardon (1993). The dvson of the sample nto urban, and rural zones wth hgh and low ncome-earnng potental n agrculture facltates the study s ablty to answer hypotheses regardng ncome effects. We also generate a set of household weghts generated on the bass on varaton between the 1990 Census estmates of the sze of each cluster and the actual sze found n 1996, and on the devaton of the actual number of households surveyed n each cluster from the target number. All results presented below take account of the clustered, weghted and stratfed nature of the sample. Enumerators ntervewed each household twce, wth the start of the two-week consumpton recall perod sgnalled by the frst ntervew. Expendture data were collected on all food (36 categores) and other frequent expenses (20 categores) durng the recall perod. The expendture estmates nclude the mputed value of own-producton, 3 net gfts receved, and stock changes, so they should be a good measure of consumpton durng the recall perod. An annual recall covered 31 categores of nfrequent expenses. An nventory of durable assets was used to estmate the value of the flow of servces from these assets, ncludng rental servces from owneroccuped dwellngs. We also broke out sx categores desgned especally to capture adult goods. Enumerators were traned to be sure to explan to respondents the scope of the goods ncluded n each commodty category, especally those for adult goods. The data requred for our analyss, the estmaton of equatons (1) and (3), are avalable for 1144 households. The data set ncludes sx commodtes that may plausbly be canddates for adult 3 The monetary values for self-produced foods were the values used by respondents. Estmates of average expendture are unchanged f these respondent-reported unt values are replaced by ether cluster medans of the unt values or cluster averages of market prces (Gbson and Rozelle, 1998). 9

12 goods: adult clothng, alcohol, betelnut (a mld narcotc smlar to pan), gamblng and lotteres, meals eaten away from home, and tobacco and cgarettes. All except for betelnut and gamblng have been ncluded as adult goods by Deaton (1989). Eght demographc groups were created: the number of males and females n each of four age groups, 0-6 years, 7-14 years, years, and over 50 years. To control for other household trats that could affect household expendtures, our specfcaton ncludes a dummy varable for whether the household head was a prmary school graduate (sx years of schoolng), three dummes for the man ncome source of the head (formal busness or wages, tree crop agrculture, food crops and lvestock), and four regonal dummes. Table 1 contans budget shares for the canddate adult goods along wth means for the other Engel curve regressors, all dsaggregated by agroecologcal zone. As dscussed above, the total budget share of the canddate adult goods s rather constant across the zones, at approxmately 12 percent. The largest shares for the canddate goods are betelnut and tobacco products, whle alcohol shows the greatest varaton n budget shares across the zones. Comparng the urban and two rural zones demonstrates a pattern of per capta expendtures, educaton of the household head, and nvolvement n the formal sector that are consstent wth those expected wth changng ncome levels. Consstent wth the aggregate evdence of females havng a smaller share of the PNG populaton (UN, 1996), the household data show that at each age group, females have a smaller demographc share than males. The gap between male and female demographc shares s especally apparent for chldren of age 7-14 years n all regons, and n the rural zone wth low ncome potental. The uneven sex ratos are least apparent n the urban sector, and n the youngest age group (0-6 years). 10

13 One possble econometrc complcaton for the estmaton of equaton (1) arses when there are only a few adult goods; a regresson for one adult good on the total would be rather lke regressng somethng on tself (Deaton, 1989). Observatons wth a large unexplaned resdual expendture on a partcular adult good could have a large total adult goods expendture, wth bas arsng from the correlaton between the equaton error and the explanatory varable. To guard aganst ths, we could adopt an nstrumental varables approach, usng total expendture, x, as an nstrument for xg. Of course, f the bas s not statstcally mportant, there would be a loss of effcency when usng nstrumental varables, affectng the test of whether good s a genune adult good. To resolve the trade-off between potental bas and loss of effcency, we use ether OLS or IV to estmate equaton (1), dependng on the outcome of a Durbn-Wu-Hausman test. There s also a complcaton n testng the hypothess that the π-ratos for a partcular type of chld are equal across adult goods (equaton 6) when the data come from a clustered survey, as they do here. Because the π-ratos are non-lnear transformatons of the OLS parameter vector b, Deaton, Ruz-Castllo and Thomas (1989) use the delta method to calculate the varances both wthn and across equatons, usng: 1 { V( ˆ r )} = J r ( X X) Jr j j where J r s the 1 k Jacoban matrx of the transformaton from the b s nto the scalar r, X s the matrx of explanatory varables whch are common n each adult goods equaton, and j s the resdual covarance between the th and jth equatons, estmated from: ˆ j = 1 ( n k) e e j where e s the vector of resduals from the th equaton and ( n k) s the degrees of freedom n each 11

14 regresson. The problem wth clustered surveys s that the observatons are not sampled ndependently, so the covarance matrx of the cross-equaton parameters, E[ ( b b )( bˆ j b j )] not smply the ( X X) 1 ˆ s j that s used n the procedure outlned by Deaton, Ruz-Castllo and Thomas (1989). To address ths feature of our data, we took several steps to modfy the exstng testng procedures. Frst, we note that the robust (or Huber/Whte/Sandwch ) estmator of the varance s approprate for sngle equaton estmaton wth clustered survey data (StataCorp, 1999). Hence, a statstcally approprate covarance matrx for testng the equalty of the π-ratos across adult goods can be obtaned by stackng the equatons for each adult good and usng a fully nteracted dummy varable model (that s, a dummy varable specfed for each type of adult good s nteracted wth all of the varables n the stacked model). 4 To the extent that ths modfcaton of the testng procedure allows us to control for the survey desgn effects, and n partcular the reduced precson that clustered samples brng, we wll be less lkely to reject null hypotheses (both wthn and across equatons) than wll prevous studes that have gnored survey desgn effects. In other words, the falure of prevous outlay-equvalent studes to detect gender bas becomes more surprsng because most studes appear to have gnored survey desgn effects, makng them more lkely to reject the null of no gender bas. 4 We are grateful to Roberto Guterrez of Stata Corporaton for suggestng ths procedure. Because repeatng the same observatons nduces a correlaton n the dsturbances of the stacked model, the clusters also have to be redefned as the nteracton of the orgnal clusters wth the dummy varable for each replcaton of the dataset. 12

15 V. Results Table 2 contans the results of testng the canddate adult goods, usng the lnear expendture model descrbed n equaton (1). The nstrumental varables estmator was used for all of the expendture tems except for betelnut and gamblng, where ths choce reflected the outcome of the Durbn-Wu-Hausman tests. The hypothess that the age and gender of chldren plays no part n explanng the allocaton of expendtures wthn the adult goods group s accepted for all canddate goods. Ths result s consstent wth evdence from other countres, although gamblng has typcally not been tested as a potental adult good whle betelnut s a specalsed tem for PNG (although the related pan has been used as an adult good n South Asa). The Engel curves for these sx adult goods and for the aggregate adult goods budget share explan between 3.3 percent (tobacco products) and 13 percent (meals out of the home) of the varaton n budget shares. The Engel curve for the aggregate adult goods group explans 7.8 percent of varaton n budget shares. In all cases, F-tests ndcate that the overall sgnfcance of the varables n the regresson, even when adoptng the conservatve approach of usng degrees of freedom based on the number of clusters rather than the number of households. 5 To save space, these Engel curves are not presented because nterest s not so much n the ndvdual coeffcents but n the combnatons of the coeffcents gven by equaton (4). Table 3 contans the estmated outlay-equvalent ratos for each of the sx adult goods and for the aggregate good formed from the sum of expendtures on the sx ndvdual goods. To nterpret these ratos, note that the π-rato of for the effect of young boys on adult clothng means that 5 Specfcally, we used an adjusted Wald (W) test for zero slopes: ( d k + 1 kd) W ~ F( k, d k + 1), where d s the number of clusters mnus the number of strata (105), and k s the number of slope varables (StataCorp, 1999). 13

16 the addton of a boy of age 0-6 years to the household has the same effect on adult clothng expendture as would a 49 percent reducton n total outlay per household member. The frst four rows of Table 3 gve the π-ratos for chldren. These should be negatve, whch they are n 23 of 28 cases. The only anomales are for the effect of young boys on gamblng expendtures, the effect of young grls on meals consumed out of the home, and the effect of older grls on the consumpton of adult clothng, betelnut, and tobacco products. In all fve cases of postve π-ratos for chldren the pont estmates are surrounded by wde standard errors and the π-ratos are not statstcally dfferent from zero. Although these anomales occur wth a smlar low frequency to that n the orgnal studes by Deaton (1989), they rase the queston of whether all of the tems ncluded n Table 3 are genune adult goods. 6 The results of the alternatve test (that s, the test that the π-ratos are equal across adult goods categores wthn age and gender groups or, along the rows n Table 3) are reported n the top panel of Table 4 (rows 1 to 4). Whle the null hypothess s never rejected at conventonal levels, lendng support to ths choce of goods, there s a potentally worryng aspect of the test results. 7 For 7 to 14 year-old boys, the test s on the borderlne of rejecton, wth p= On closer nspecton of our results, however, there appears to be a way to allay ths concern. The good that appears to contrbute most to the nequalty n π-ratos for 7 to 14 year old chldren 6 Three of 32 π-ratos were postve n the results for Côte d Ivore and 8 of 48 n the results for Thaland, whch ncluded some tems such as men s and boy s clothng whch were known to nclude chld goods. 7 Actually, there may be a second, although n our opnon, more mnor concern. In contrast to the results n Deaton (1989), the equalty of the π-ratos across goods for each of the adult demographc groups s also not rejected. Whle ths equalty s less nformatve for adults than t s for chldren, the falure to reject may ndcate low test power, although once agan, f the survey desgn effects are gnored the equalty across the goods s rejected at the 0.02 level, for at least two of the adult groups. 8 If the survey desgn effects are gnored, as they seem to have been n some prevous studes, the p-value for 7-14 year old grls would also fall nto the rejecton zone, at p=0.02 (from 0.12 wth survey desgn effects ncluded). 14

17 s betelnut. It s possble that some older chldren n ths group actually do consume some betelnut. In PNG, there s no legal age lmt on ts consumpton, and n some communtes there are no real norms aganst ts use by chldren. In contrast, even n PNG, there are mnmum age requrements and communty norms n most areas aganst alcohol and tobacco use by chldren. The results n the bottom panel of Table 4 (rows 5 to 8) show that the equalty of the π-ratos s more dffcult to reject once betelnut s removed from the lst of adult goods. Hence, to ensure the robustness of our results, we present the tests for boy-grl dscrmnaton both wth and wthout betelnut ncluded n the lst of adult goods. If boys are favoured over grls, the π-rato for a gven age category should be a bgger negatve number for boys than for grls. It s apparent by lookng down the columns of Table 3 (comparng rows 2 and 4) that for the older age group (7 to 14 year olds) ths pattern holds for all sx ndvdual adult goods, as well as for the aggregate group. The pont estmate for the aggregate adult goods group suggests that the addton of an older boy to the household reduces adult goods consumpton by as much as would a 90 percent reducton n total outlay per member. In contrast, the addton of a smlarly aged grl to the household s equvalent to only a two percent reducton n per capta outlay. The dfference n these π-ratos for the aggregate good s statstcally sgnfcant (p<0.01), although dfferences for the ndvdual adult goods are less precsely measured (row 18, columns 1 to 6). 9 If betelnut s excluded from the lst of adult goods, the same concluson of bas n favour of boys s reached: the π-rato for the aggregate adult goods group s 1.11 for 7 to 14 year-old boys 9 The mportance of controllng for survey desgn effects can be seen by notng that f the test of equal π-ratos for the effect of older boys and older grls on betelnut (alcohol) consumpton s carred out wth the clusterng and stratfcaton gnored, the p-values fall from 0.08 to 0.02 (0.30 to 0.13). 15

18 (standard error of 0.21) and 0.28 for smlarly aged grls (standard error of 0.27) and the hypothess that these two are the same s stll rejected, at the p<0.02 level. In contrast to older chldren, for the 0 to 6 year-olds, only one ndvdual adult good (meals out of the home) generates the pattern of a more negatve π-rato for boys than for grls. Moreover, nspecton of the standard errors shows that the dfferences n π-ratos for boys and grls n the 0 to 6 years age group are all statstcally nsgnfcant. Hence, f there s any gender bas amongst the young chldren, t does not appear to be detectable usng the adult good method. The π-ratos n Table 3 also suggest that dfferences n the effects that adult men and women have on adult goods consumpton are usually statstcally nsgnfcant. 10 Even f they were not, conclusons about gender bas for adults cannot be made on the bass of these results because they may just reflect gender dfferences n preferences (ths explanaton can be ruled out for chldren, who should exert only ncome effects on adult goods demand). Dsaggregatng by agroecologcal zone Our data also demonstrate the same pattern of gender dscrmnaton n rural zones of PNG when usng the outlay equvalent ratos for the adult goods aggregate, although a dfferent pattern s found for urban households (Table 5, columns 1 to 3). The results suggest that parents n the rural sector favour boys over grls n the 7 to 14 year old age category n the hgh-ncome potental areas (p<0.00), and also although less precsely measured n the low-ncome potental areas (p<0.06). In contrast, we can not reject the hypothess that the nfant grl and nfant boy π-ratos are the same 10 The p-value for the dfference between π-ratos for prme age males and females n the aggregate adult goods group drops from 0.28 to 0.06 f betelnut s excluded, probably because adult women are equal partcpants n consumpton of betelnut n contrast to alcohol and tobacco whch are predomnantly consumed by men. 16

19 n ether of the rural zones (row 9). In urban areas, however, there s no sgn of gender dscrmnaton when examnng ether nfants or adolescents (column 1, rows 9 and 10). Ths pattern of results across dfferent areas appears to provde only partal support to the hypothess of Haddad and Reardon (1993) that dscrmnaton aganst grls wll be less detectable as the economc opportuntes for women ncrease. Whle the lack of apparent bas n urban areas may be consstent wth greater economc opportuntes n urban areas, t s puzzlng that gender bas does not appear to vary across the two rural zones. Most observers would expect woman s labour n agrcultural producton to be more valuable n the rural zone wth hgh agrcultural ncome potental and therefore the costs to households of dscrmnatng aganst grls to be much greater. Moreover, the hypothess that the π-ratos are equal across the urban-rural and agroecologcal zones s rejected (at the p=0.03 level) only for boys (Table 5, column 4). Further tests show that ths sgnfcant result s drven by the dfference between the urban sector and each of the two rural zones; the comparson between the hgh- and low-ncome rural zones brngs an nsgnfcant result (p=0.24). Thus, whle parents n both hgh- and low-ncome rural zones of PNG sgnfcantly reduce ther consumpton of adult goods when they have an addtonal male adolescent, urban parents do not seem to make smlar adjustments n ther consumpton allocatons. Because bas n favour of boys seems to be restrcted to rural areas, whch are presumably more tradton-bound, t mght be construed from these fndngs that n PNG gender bas may arse more from cultural rather than strctly economc factors, although further examnaton of ths hypothess s needed. 17

20 VI. Conclusons In our paper we have tested for bas n the allocaton of goods between boys and grls wthn households n Papua New Gunea usng Deaton s (1989) outlay-equvalent rato method. Addng a 7-14 year old boy to the household reduces expendture on adult goods by as much as would a nne-tenths reducton n total outlay per member, but smlarly aged grls have no effect on adult goods expendture. The hypothess of Haddad and Reardon (1993) that gender bas s nversely related to the mportance of female labour n agrcultural producton s not supported. There s no evdence of bas aganst grls n the urban sector, nor of bas aganst younger grls n the rural sector. Perhaps the most surprsng aspect of our research s that t worked so well. In the past, adult goods-based approaches have not yelded results consstent wth fndngs of other socal scence research and other economc studes based on alternatve approaches. Consequently, there has been some queston about whether expendture-based methods can be used as a relable tool for nvestgatng the nature and extent of gender dscrmnaton (Deaton, 1997). Although our results cannot explan why the adult goods method has faled to work n other countres, the successful applcaton n ths settng at least provdes some grounds for belevng that the method can be a useful tool. Perhaps the combnaton of factors n PNG that make bas greater and easer to pck up, or the hgher qualty of our data, whch were collected wth ths purpose n mnd, may have helped to provde greater success for the method on ths occason. If our results have dentfed the exstence of a deep-seated cultural-socal bas aganst grls n Papua New Gunea, educatonal and other development polces counterng gender dscrmnaton could lead to both gans to socal welfare and more rapd, equtable growth. For 18

21 example, nterventons through PNG s wdespread (though flounderng) health system mght target young grls because expendture on females may have a greater effect on welfare than f t were spent on boys. Moreover, the ntra-household bas revealed here may mean that the ncdence of poverty s understated because average household data wll mss the fact that females n some household have access to suffcently few resources such that ther ndvdual consumpton level falls below the poverty lne. However, further work s needed n PNG to explan why bas n favour of boys s restrcted to the rural areas and s also restrcted to chldren n the older age groups. One plausble hypothess, whch we have not tested, s that the needs of younger chldren -- prmarly food and care are met usng resources controlled manly by women. In contrast, the needs of older chldren partcularly school fees and clothng requre cash, whch s controlled manly by adult males, who may be the ones who favour boys over grls. 19

22 REFERENCES Ahmad, A., and Mordoch, J. (1993) Identfyng sex bas n the allocaton of household resources: evdence from lnked household surveys from Bangladesh, mmeo, Harvard Unversty. Allen, B., Bourke, M., and Hde, R. (1995) The sustanablty of Papua New Gunea agrcultural systems: the conceptual background Global Envronmental Change 5(4): Boserup, E. (1970) Womens Role n Economc Development, London: Earthscan Press. Deaton, Angus (1997), The Analyss of Household Surveys: A Mcroeconometrc Approach to Development Polcy Johns Hopkns, Baltmore. Deaton, A. (1989) Lookng for boy-grl dscrmnaton n household expendture data World Bank Economc Revew, 3(1): Deaton, A., and Muellbauer, J.(1986) On measurng chld costs: wth applcatons to poor countres Journal of Poltcal Economy, 94(4) Deaton, A., Ruz-Castllo, J., and Thomas, D. (1989) The nfluence of household composton on household expendture patterns: theory and Spansh evdence Journal of Poltcal Economy, 97(1):

23 Gbson, J. (2001) Lteracy and ntrahousehold externaltes. World Development 29(1): Gbson, John, and Rozelle, Scott (1998), Results of the Household Survey Component of the 1996 Poverty Assessment for Papua New Gunea mmeo Populaton and Human Resources Dvson, The World Bank. Groos, A. D., and Garner, P. A. (1988) Nutrton, health and educaton of women n Papua New Gunea Papua New Gunea Medcal Journal, 31(1): Haddad, L., and Kanbur, R. (1990) How serous s the neglect of ntra-household nequalty? Economc Journal, 100(Sept): Haddad, L., and Reardon, T. (1993) Gender bas n the allocaton of resource wthn households n Burkna Faso: a dsaggregated outlay equvalent analyss Journal of Development Studes, 29(2):, Jarrett, F. G., and K. Anderson (1989) Growth, Structural Change, and Economc Polcy n Papua New Gunea: Implcatons for Agrculture Canberra: Natonal Centre for Development Studes, Australan Natonal Unversty. Overfeld, D. (1998) An nvestgaton of the household economy: coffee producton and gender 21

24 relatons n Papua New Gunea Journal of Development Studes, 34(1): Rosenzweg, M. R., and T. P. Schultz (1982) Market Opportuntes, Genetc Endowments, and Intrafamly Resource Dstrbuton: Chld Survval n Rural Inda, Amercan Economc Revew 72(4): Sen, A. K. (1984) Famly and food: sex bas n poverty, n Resources, Values, and Development (Ed.) A. K. Sen, Harvard Unversty Press, Cambrdge, Mass, pp StataCorp (1999), Stata Statstcal Software: Release 6.0, College Staton: Stata Corporaton. Subramanan, S. (1994), Gender dscrmnaton n ntra-household allocaton n Inda, Department of Economcs, Cornell Unversty. Udry, C. (1997) Recent advances n emprcal mcroeconomc research n poor countres: an annotated bblography. Journal of Economc Educaton 28(1): Unted Natons (1996) Sex and Age Dstrbuton of the World Populaton: The 1996 Revson Populaton Dvson, Department of Economc and Socal Affars of the Unted Natons Secretarat. 22

25 World Bank. (1999) Papua New Gunea: Poverty and Access to Publc Servces Washngton, DC: World Bank. Zhang, L. (1999) Poverty and nutrton n rural Chna Center for Chnese Agrcultural Polcy, Chnese Academy of Agrcultural Scences, Bejng. 23

26 Varable Table 1: Descrpton of the Data (Means) Papua New Gunea Urban Zone Rural Hgh Zone Rural Low Zone total canddate adult goods budget share adult clothng budget share alcohol budget share betelnut budget share gamblng and lotteres budget share meals eaten out of home budget share tobacco and cgarettes budget share log per capta total expendture log of household sze share of males, 0 6 years, n household share of females, 0 6 years, n household share of males, 7-14 years, n household share of females, 7-14 years, n household share of males, years, n household share of females, years, n household share of males, >50 years, n household share of females, >50 years, n household =head completed prmary school, 0=other =man ncome of head from formal sector =man ncome of head from tree crops =man ncome of head from food crops Number of Observatons (percentage of natonal household numbers) (100) (13) (64) (23) Note: Means are calculated usng household samplng weghts. 24

27 Adult good Table 2: Identfyng Adult Goods by Testng for Excluson of Chld Demographcs Hausman Test Approprate Estmator Wald Test for Excludng Chld Demographcs p-value Adult Clothng 2.78 IV Alcohol 3.29 IV Betelnut 1.72 OLS Gamblng 1.79 OLS Meals out 3.56 IV Tobacco 2.58 IV Note: Test statstcs and p-values are corrected for clusterng, samplng weghts and stratfcaton. The Wald test s dstrbuted as F (k, d-k+1). where d s the number of clusters mnus the number of strata (105), and k s the number of restrctons (4). The Hausman test s dstrbuted t d. 25

28 Table 3: Outlay Equvalent Ratos for Adult Goods Gender and age Adult clothng Alcohol Betelnut Gamblng Meals out Tobacco All adult goods π-ratos Chldren Male Male Female Female Adults Male Male > Female Female > Chldren Standard errors Male Male Female Female Adults Male Male > Female Female > p-values for test of equal π-ratos for males and females Infants Adolescents Prme Adults Old Adults Note: Standard errors and p-values are corrected for clusterng, samplng weghts and stratfcaton. 26

29 Table 4: Adjusted Wald Tests for Equalty of π-ratos across Adult Goods Chldren Adults Gender and age Test p-value Test p-value Males Males Males Males > Females Females Females Females > Excludng Betelnut Males Males Males Males > Females Females Females Females > Note: The test statstcs are adjusted Wald (W) tests, ( d k + 1) W kd, where d s the number of clusters mnus the number of strata n the stacked model and k s the number of restrctons beng tested. The tests are approxmately F dstrbuted wth k and d k + 1 degrees of freedom under the null hypothess that the π-ratos are the same across all canddate adult goods (wth betelnut ncluded F(5, 630) and wth betelnut excluded F(4, 525)). 27

30 Table 5: Outlay Equvalent Ratos for the Aggregate Adult Goods Group by Agroecologcal Zone Gender and age Urban Zone Hgh Income Potental Rural Zones Low Income Potental Test for equalty across zones (p-value) Chldren π-ratos Male (0.52) Male (0.03) Female (0.81) Female (0.23) Adults Male (0.01) Male > (0.43) Female (0.28) Female > (0.76) p-values for test of equal π-ratos for males and females Infants Adolescents Prme Adults Old Adults Note: The test for equalty of the π-ratos across zones s an adjusted Wald test, dstrbuted as F (2,105) under the null hypothess. The p-values for the hypothess tests are corrected for sample desgn effects. 28