Food Expenditures away from Home by Elderly Households

Transcription

1 Food Expendtures away from Home by Elderly Households Steven T. Yen Panagots Kasterds John B. Rley Department of Agrcultural and Resource Economcs The Unversty of Tennessee Knoxvlle, TN Selected Paper prepared for presentaton at the Agrcultural & Appled Economcs Assocaton s 2012 AAEA Annual Meetng, Seattle, Washngton, August 12-14, Copyrght 2012 by Steven T. Yen, Panagots Kasterds, and John B. Rley. 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.

2 Abstract Ths study nvestgates the dfferentated effects of economc and soco-demographc varables on food away from home (FAFH) expendtures by type of faclty among elderly households n the Unted States. Usng data from the Consumer Expendture Surveys, the systems of expendtures on full-servce, fast food, and other restaurants are estmated wth a multvarate sample selecton estmator whch also accommodates heteroscedastcty n the error dstrbuton. Statstcal sgnfcance of error correlatons among equatons justfes estmaton of the sample selecton systems. Income, employment statuses, race, educaton, geographc regon, and household composton are mportant determnants of FAFH expendtures. Income contrbutes to full-servce and fast-food expendtures by the elderly mplyng that the future of FAFH ndustry s ted to macroeconomc condtons. Better educaton s assocated wth greater probabltes and larger levels of expendtures at all facltes. Effects of the Supplemental Nutrton Assstance Program (SNAP) are found to be strong and negatve, nvaldatng polcy concerns for the general populaton that partcpaton n the program mght enhance consumpton of less healthy FAFH. Keywords Censorng equvalence scale elderly food away from home sample selecton system JEL D12, Q13, C31

3 1 Introducton Amercans spend nearly half of ther food dollars on food away from home (FAFH). Total spendng on FAFH by all famles and ndvduals was $433.5 bllon n 2010 and consttuted 41.3 percent of total food expendture, up from 32.0 percent n 1980 (USDA-ERS 2011). A varety of foodservce frms, ncludng full-servce restaurants, fast-food restaurants, drnkng places, and retal stores, compete n the food ndustry. Full-servce restaurants and fast-food companes, wth total sales of $219.5 bllon and $209.5 bllon, captured 72.2% of the FAFH market n 2010 (USDA-ERS 2011). By 2009, there were over 221,513 full-servce establshments and 268,459 fast-food establshments n the Unted States (U.S.) (U.S. Census Bureau, No date). Owng to rapd ncome growth, urbanzaton, and globalzaton, FAFH consumpton have also ncreased n many other countres such as Span (Mutlu and Graca 2006; Angulo et al. 2007), Greece (Mhalopoulos and Demousss 2001), Chna (2006; Ba et al. 2010), Malaysa (Tan, 2010), Tawan (Chang and Yen 2010), and many other Asan countres (Pngal 2007). The lterature has dentfed a lst of economc and soco-demographc factors that contrbute to FAFH consumpton ncludng ncome, household sze and structure, regons, employment, age, educaton, race, and ethncty (McCracken and Brandt 1987; Yen 1993; Byrne et al. 1996; Jensen and Yen 1996). Food assstance programs such as the Supplemental Nutrton Assstance Program (SNAP), have also had an mpact on FAFH (Lu et al. 2012). Durng the Great Recesson, encompassng fallng ncomes, mountng unemployment, hgh food prces, and hgh partcpaton rates n the federal food assstance programs, Amercans ate out less due to tghtened budget. Several of the demographc factors trggerng FAFH have also changed remarkably n the 2000s (Cherln 2010). For nstance, ncreasng proportons of sngle parents, Hspanc and Asan mmgrants, and the elderly n the

4 2 populaton marked sgnfcant changes n famly and ethnc composton. Contnuaton of rapd demographc changes may result n a new upward trend n FAFH consumpton n the next few years. Ths possblty escalates the concerns about the publc health mplcatons of dnng out. There s a large body of emprcal lterature pontng out that FAFH s less healthy than food at home (Mancno et al. 2009), and a number of polces have been mplemented n order to rase the publc awareness of the benefts of a healthy det. Therefore, there exsts a great nterest n consumers expendtures on FAFH among polcy makers, the academa, and food marketers. Type of faclty has been the subject of nvestgaton n consumer food purchases. Chung and Myers (1999) fnd store type more mportant n drvng food prce dspartes than the geographcal locaton of a household; such food prce dspartes can n turn affect consumer food purchases. On FAFH expendtures, prevous studes have addressed dfferences between fullservce and fast-food restaurants and found that the effects of economc and soco-demographc factors on FAFH generally dffer by type of faclty (McCracken and Brandt 1987; Byrne et al. 1996; Stewart and Yen 2004; Bnkley 2006; Lu et al. 2012). Much of the emprcal lterature on FAFH s dated except Lu et al. (2012). Further, ths lterature addresses the general populaton n a country. No study, to our knowledge, has nvestgated FAFH expendtures among the elderly. The elderly populaton may well have a dstnct expendture pattern on FAFH due to dfferences n lfestyle as well as economc condtons such as greater fnancal stablty relatve to younger households. Ths study flls ths emprcal vod by focusng on FAFH by type of faclty and for the elderly populaton n the U.S. To that end, FAFH expendtures for the younger populaton wll also be nvestgated for comparson. Further, as n other mcrodata, the data used n ths study contan censorng (zero observatons) n the expendture varables. Statstcal procedures not accountng censorng produce nconsstent

5 3 emprcal estmates. To address ths data feature, the sample selecton system, an extenson of the bvarate sample selecton model (Heckman 1979) advanced by Yen (2005) and used n prevous studes of FAFH (Stewart and Yen 2004; Lu et al. 2012), s further extended to accommodate heteroscedastcty n the error terms. Heteroscedastcty, f not accounted for, can cause statstcal nconsstency n emprcal models wth lmted dependent varables. Theoretcal framework We motvate our emprcal specfcaton by the dscrete random utlty theory. A household maxmzes the random utlty functon subject to a full-ncome constrant: max{ U( Dq, c, ; s) pq c w( T ) m}, (1) qc,, where q [ q1,..., qn ] s vector of quanttes wth postve prces p [ p1,..., pn], c s a composte commodty for other goods wth prce normalzed at unty, s s a vector of demographc characterstcs, m s non-wage earnng; and for workng members such as household head and spouse, l s a vector of lesure actvtes wth prces (wage rates) w, T s a vector of tme endowments. D dag( d1,..., d n ) s a dagonal matrx wth each bnary ndcator d ndcatng a potental consumer of q. Assume the utlty functon U( Dq, c, ; s) s strctly quas-concave and ncreasng wth respect to c, l, and postve elements of Dq. Then, solvng (1) yelds the notonal demand q * for FAFH a vector of optmal quanttes demanded wthout non-negatvty constrants, as functon of prces, wage rates, and non-wage earnng (Becker 1965). Ths constraned utlty maxmum framework motvates two alternatve specfcatons for the demand functons. Frst, assume all ndvduals are potental consumers of q n whch case d 1 for all and censorng of each q corresponds to a corner soluton governed by a Tobt mechansm.

6 4 Second, when an ndvdual can otherwse be a potental non-consumer, ether d 1 and utlty maxmum occurs n the nteror of the choce set (vz., q 0 ), or d 0 and q 0 snce prce p 0 by assumpton. In ths latter case, censorng n each q s governed by a sample selecton mechansm. Express the notonal demands as a system of equatons for latent expendtures ( * y ) y x v, 1,..., n, (2) * where x s the vector of explanatory varables, are parameter vectors, and v are random dsturbances whch reflect the unobservable (dscussed below). Data, samples and varables Data are drawn from the 2008, 2009, and 2010 Consumer Expendture Surveys (U.S. BLS 2009, 2010, 2011). Each survey provdes consecutve two-week nformaton on FAFH expendtures that are categorzed by type of faclty, and on economc and soco-demographc characterstcs of the households. The fnal sample conssts of 20,523 observatons. We focus on FAFH expendture by elderly households. For comparson, expendtures by younger households are also estmated. The sample s segmented nto an elderly sample wth household heads age 55 (n = 7,860 households) and a younger sample (n = 12,663). To accommodate economy of scale n consumpton and the role of household composton, expendtures are expressed per equvalence scale accordng to the U.S. Natonal Research Councl defnton (Ctro and Mchael 1995). By that defnton, household equvalence scale s calculated as (number of adults number of chldren) The resultng equvalence scale s used to deflate both FAFH expendtures and pre-tax non-wage earnng (henceforth, ncome). Sample statstcs of the expendtures are presented n the appendx (Table A1). The elderly spend more ($28.88) at full-servce restaurants than younger households ($27.54), whereas the

7 5 younger households spend more at fast-food restaurants ($25.50) than the elderly households ($16.08) per equvalence scale per two weeks. The elderly spend nearly twce as much at fullservce restaurants as at fast-food restaurants, whereas the younger households spend about as much at the two types of facltes. Expendtures at other facltes represent a very small proporton of the FAFH budget $1.50 among the elderly and $4.42 among the younger households. In terms of percentages, 54.61% of the elderly households consume at full-servce restaurants, 63.22% at fast food restaurant, and 19.73% at other facltes durng the two-week perod; the percentages are hgher among the younger households: at 57.04%, 76.93%, and 38.51%, respectvely. Drawng on the random utlty theory above, non-wage earnng (ncome) s used as the explanatory varable. Income s expected to ncrease FAFH consumpton for both household groups. Demographc varables nclude age, employment statuses of household heads and spouses (f present), household composton, and dummy varables ndcatng geographcal regons, seasons, home ownershp, SNAP partcpaton status, race, and educaton (Table A1). Many state agences of SNAP provde nutrton educaton to assst recpents n makng healthy food and actve lfestyles choces. SNAP partcpaton may thus decrease consumpton of FAFH as the partcpants opt for healther dets. However, the program benefts frees up resources for FAFH as well as other goods. The net effect of SNAP on FAFH s thus unclear. We nclude a dummy varable to ndcate SNAP partcpaton status. We draw on prevous studes n the selecton of addtonal explanatory varables (McCracken and Brandt 1987; Soberon-Ferrer and Dards 1991; Yen 1993; Jensen and Yen 1996; Stewart and Yen 2004; Lu et al. 2012). Household composton varables are among the most commonly used n analyss of consumer demand, FAFH, and tme-savng goods and servces.

8 6 Homeowners may consume more food away from home because of greater fnancal stablty on the one hand. On the other hand, they may have more or less cash flow dependng on mortgage payments (versus rent) whch can affect FAFH expendture. The effects of home ownershp on FAFH expendtures are therefore unclear. Tastes and eatng habts may dffer by race. Because food preferences and other unobserved characterstcs may dffer across geographc regons and seasons, dummy varables are also ncluded to account for these dfferences. Due to the absence of prces n the sngle cross secton used, these regonal and seasonal dummy varables can accommodate regonal and seasonal prce varatons as well, ameloratng bases due to omsson of prces. Educaton, gender, and age are expected to nfluence FAFH expendtures to dfferent extents by faclty. Because work competes wth home producton (meal-preparaton) for the tme endowment, and because the trade-off between tme use and det qualty can play an mportant role n food consumpton decsons (Becker 1965), households wth workng members are hypotheszed to consume FAFH more often. Dummy varables ndcatng full-tme and part-tme employment by the household head and spouse are ncluded n the selecton equatons as part of the parameter dentfcaton strategy (dscussed below). 1 Fnally, dummy varables are ncluded to ndcate the years 2009 and Defntons and descrptve statstcs of all explanatory varables are provded n the appendx for both the elderly and younger households (Table A1). There are notable dfferences n the sample statstcs between the two types of households. For nstance, the 1 The theory of condtonal demand suggests use of work hours, not wage rates, as the explanatory varable (Shaw and Feather 1999). Because our elderly sample contans a large proporton of non-workng household heads (42.8%) and spouses (75.9%), dummy varables are used nstead of work hours.

9 7 elderly households regster an average ncome as hgh as $17,540 per equvalent scale per year, compared to only $4,030 for the younger households. Household age compostons are also very dfferent: number of members age < 18 s only 0.10 among the elderly households, compared to 0.94 for the younger households. Econometrc procedure Censorng n our expendture varables need to be addressed to obtan consstent parameter estmates. Earler studes of FAFH employed models such as the Tobt model (McCracken and Brandt 1987), sngle-hurdle model (Yen 1993), and double-hurdle model (Jensen and Yen 1996; Mutlu and Graca 2006) to accommodate the censored data. The sample selecton system estmator has also been used n Stewart and Yen (2004) and, more recently, Lu et al. (2012). Except Jensen and Yen (1996), these prevous studes have not addressed heteroscedastcty n the error terms, whch can cause statstcal nconsstency n the emprcal estmates. We extend the sample selecton system procedure of Stewart and Yen (2004) and Lu et al. (2012), suggested n Yen (2005), by accommodatng heteroscedastcty of the error terms. Consder a three-good system where outcome n each FAFH expendture ( y ) s governed by a bnary sample-selecton rule log y=x +v f z +u> 0 y= 0 f z +u 0, (3) where = 1,2,3 for full servce, fast food, and other facltes, respectvely, z and x are vectors of explanatory varables, and are conformable parameter vectors, and u and v are random errors. Assume the concatenated error vector [ u,v ] [ u1, u2, u3, v1, v2, v3] s dstrbuted as sxdmensonal normal wth zero means, standard devatons [1,1,1, 1, 2, 3], and correlaton matrx

10 8 R = [ j ] such that correlaton between the error terms n the correspondng selecton and level equatons, ( u, v ), s + 3,. The sample selecton system (3) s more flexble than the Tobt system n that the bnary and level outcomes are governed by separate stochastc processes. Comparson of the sample selecton system and Tobt system s addressed n the emprcal secton below. Each dependent varable y n (3) s transformed by natural logarthm. Such transformaton s common n estmaton of endogenous selecton and swtchng regresson model and amelorates potental nonnormalty and heteroscedastcty of the error term (Yen 2005; Yen and Rosńsk 2008). Importantly, heteroscedastcty n the error terms can cause nconsstency n parameter estmates n lmted dependent varable models. To accommodate heteroscedastcty n the error terms, each of the standard devatons s parameterzed as a functon of explanatory varables h wth parameter vector : = exp( h ), = 1,2,3. (4) The sample lkelhood functon s dentcal to that presented n Yen (2005), except the addtonal parameterzaton n (4). Denote the k-dmensonal probablty densty functon (pdf) as k, cumulatve dstrbuton functon (cdf) as k, and defne standard normal varates t = (log y - x ) / for = 1,2,3. Then, for an all-postve regme, the lkelhood contrbuton s 3 = z z 2 -z 3 L = ( y ) ( t, t, t ; R ) ò ò ò h( u, u, u v, v, v ) du du du, (5) where R 3 s correlaton matrx among error terms ( v1, v2, v 3) whch s the 3 3 lower-rght submatrx of R, and 3 1 ( 1y - = ) s the Jacoban of the transformatons from [ v, v, v ] to [ t, t, t ]. In (5), the condtonal densty hu ( 1, u2, u3 v1, v2, v 3) s trvarate normal (Kotz et al. 2000, pp. 111

11 9 112) whch can be ntegrated by usual means (Yen 2005). Consder next a partally censored regme, n whch y 1 s censored at zero and ( y2, y 3) are postve. The lkelhood contrbuton s 3 = z z z 3 L= ( y ) ( t, t ; R ) ò ò ò h ( u, u, u v, v ) du du du, (6) where R 2 s correlaton matrx between error terms ( v2, v 3) whch s the 2 2 lower-rght submatrx of R. Fnally, for a sample regme wth ( y1, y 2) and y 3 > 0, the lkelhood contrbuton s z 1 - z z L = y ( t ) ò ò ò h( u, u, u v ) du du du. (7) The ntegrals n (6) and (7) can be smplfed followng a procedure smlar to that n (5) because the condtonal denstes hu ( 1, u2, u3 v2, v 3) and hu ( 1, u2, u3 v 3) are also trvarate normal. The lkelhood contrbutons for other partally censored regmes are covered n (6) and (7) by permutng the error terms. It s clear from (5), (6), and (7) that estmaton of the sample selecton system requres evaluatons of trvarate normal cumulatve dstrbuton functons (cdf s) for all sample observatons. These probablty ntegrals are evaluated wth Gaussan quadratures. The parameters to estmate nclude those from the selecton equatons ( ), level equatons ( ), heteroscedastcty specfcaton ( ) for = 1,2,3, and error correlatons ( j ) for > j. The sample selecton system nests two restrcted specfcatons: () an ndependent model whch corresponds to parametrc restrctons ρ j = 0 for all j, vz., wth all error correlatons equal to zeros; () a parwse selecton system whch corresponds to ρ j = 0 for all j except ρ 41 0, ρ 52 0, and ρ These restrcted models can be estmated by mposng the above parametrc restrctons. In addton, the ndependent model for each of the expendtures n the

12 10 system, can be estmated by the probt model based on the bnary (0/1) outcome related to y usng the whole sample, and OLS for each log( y ) usng the truncated sample (condtonal on y > 0 ). The parwse selecton system conssts of three bvarate sample selecton models (Heckman 1979) for the three expendtures whch can be estmated as such separately. Tests of the sample selecton system aganst the two nested models can be done wth the Wald, Lagrange multpler (LM), or lkelhood-rato (LR) test (Engle 1984). Margnal effects of probabltes, condtonal levels, and uncondtonal levels are calculated to explore the effects of explanatory varables. Specfcally, for each good, the probablty, condtonal mean and uncondtonal mean of y are (Yen and Rosńsk 2008, p. 5) Pr( y 0) ( z ), (8) 1 E y y x z z (9) 2 ( 0) exp( / 2) 1( 3, ) / 1( ), E y x z (10) 2 ( ) exp( / 2) 1( 3, ). Dfferentatng (dfferencng) (8), (9), and (10) wth respect to varables x, z, and h gves the margnal effects of contnuous (dscrete) explanatory varables, whch are extensons of those n Yen and Rosńsk (2008, p. 5) due to the heteroscedastcty specfcaton (4); analytc dervatves are avalable upon request. We evaluate all margnal/dscrete effects for all sample observatons and average them over the sample. Standard errors of these margnal effects are calculated by a mathematcal approxmaton procedure known as the delta method (Spanos 1999, p. 493). Specfcaton tests and estmaton results Choce of varables (h) n the heteroscedastcty specfcaton (4) s carred out emprcally. Income s found to perform better, n terms of statstcal sgnfcance, than the other varables and s retaned. Next, we test for equalty of parameters between the elderly and younger households.

13 11 Usng a lkelhood rato (LR) test, smlar to Chow test n lnear regresson models, the hypothess of equal parameters s rejected (LR = , df = 165, p-value < ), whch justfes segmentaton of the sample and our focus on the elderly households. The parameter estmates and, more mportantly margnal effects (reported below) are found to dffer notably between the elderly and younger households; these dfferences would have been masked by the use of a pooled sample and further justfy separate analyss for the two groups of households. We then carry out a seres of specfcaton tests, focusng on the elderly households. Frst, heteroscedastcty of error terms n the level equatons s tested. Results of t-tests ndcate statstcal sgnfcance of ncome n the heteroscedastcty equatons for full-servce and other facltes, both at the 5% level of sgnfcance. Further, Wald test result suggests jont sgnfcance of ncome n the heteroscedastcty equatons (Wald = 9.17, df = 3, p-value = 0.027), supportng the heteroscedastcty specfcaton. The heteroscedastc sample selecton system and s then compared to the heteroscedastc trvarate Tobt system, based on log-lkelhood contrbutons at the maxmum-lkelhood estmates for both models. Result of Vuong s (1989) test, wth standard normal statstc z = 37.60, suggests that the sample selecton system performs better than the Tobt alternatve n fttng the data. The sample selecton system s then tested aganst the nested ndependent and parwse selecton (Heckman 1979) systems dscussed above. Of the 15 error correlaton coeffcents, 14 are sgnfcant at the 1% level of sgnfcance and one at the 5% level by ndvdual t-tests (Table A2). As expected, we reject both the parwse selecton system (Wald = , LR = , LM = , df = 12) and ndependent system (Wald = , LR = , LM = , df = 15) wth p-values < In sum, results suggest separate analyss wth the elderly sample, the sample selecton system s preferred to the Tobt alternatve, and the heteroscedastc specfcaton s justfed. Maxmum-lkelhood estmates for the elderly sample are

14 12 presented n the appendx (Table A2), and estmates for younger households are avalable upon request. Margnal effects Margnal effects of varables on the probablty, condtonal level, and uncondtonal level for the elderly households are presented n Table 2. These effects dffer, notably for many varables, between the elderly households and younger households. For nstance, whereas ncome contrbutes to full-servce and fast-food expendtures by both groups, the elderly are more responsve to ncome changes than ther younger counterparts, n terms of probablty and levels of expendtures at full-servce restaurants. A dfferent pattern s found for fast foods, wth a greater effect on probablty but a smaller effects on condtonal level for the elderly households than the younger households. Income affects expendture at other facltes among the elderly households but does not the younger households. The effects of employment statuses (full-tme and part-tme) are all postve for both household groups, although the effects are generally larger for the younger households than the elderly households. Dfferences are also found n many other varables. Margnal effects for younger households are avalable upon request. Focusng on the elderly households (Table 2), the margnal effects for most varables dffer substantally, n sgns and magntudes, across facltes. Income has postve effects on expendtures at full-servce and fast-food restaurants but negatve effects on expendtures at other facltes. The magntudes dffer among facltes. Overall, the margnal effects of ncome on FAFH are relatvely small, wth a $10,000 ncrease n ncome per equvalence scale ncreasng expendture at full-servce (fast-food) restaurants by less than $3 ($1) per equvalence scale for two weeks. These postve but small effects of ncome are n agreement wth fndngs from prevous studes for the general populaton n the U.S. (e.g., McCracken and Brandt 1987; Jensen and Yen

15 ; Lu et al. 2012). Larger effects of ncome are reported for Chna (Ba et al. 2010) and Span (Angulo et al. 2007), also on the general populaton. Income has negatve but barely notceable effects on the levels of expendtures at other facltes. The roles of household composton are mxed. Numbers of adults age and age > 64 have postve and sgnfcant effects on the probabltes of consumng at all three types of facltes but, condtonal on consumpton, the levels of expendtures are sgnfcant, negatve and large. An addtonal member age (age > 64) decreases expendture at full-servce restaurants by $7.02 ($4.75) per equvalent scale per two weeks. These conflctng roles of varables on probabltes and levels of expendtures can be masked by the use of the more restrctve Tobt system and hghlght an mportant advantage of the sample selecton parameterzaton. Not surprsngly, number of younger household members (age < 18) has sgnfcant and postve effects on probablty (3.72%), condtonal level ($0.97), and uncondtonal level ($0.49) of expendture at other facltes. Age has negatve and sgnfcant effects on probabltes and levels of expendtures at both full-servce and fast-food restaurants, whle ts negatve effect on other facltes s seen only through probablty. The negatve effect of age on probablty of fast food s partcularly notable, at 6.05%. Thus, expendtures at all types of facltes decrease wth age, n terms of probabltes and, n the case of full-servce and fast-food, by way of expendture levels as well. Despte the small proportons of workng household heads (57.3%) and spouses (24.1%), employment statuses have by far the most notable effects on FAFH expendtures. Supportng our hypothess on the role of work n FAFH, full-tme and part-tme employment by both household head and the spouse have postve, sgnfcant, and relatvely large mpacts on FAFH. Full-tme work by the household head (spouse), for nstance, ncreases the probablty of consumng at full-

16 14 servce restaurants by 9.12% (9.72%), condtonal level by $9.08 ($9.88), and uncondtonal level by $9.95 ($10.97) per equvalence scale per two weeks. The effects of part-tme work by both household head and spouse are also postve and very notable n magntudes. Postve effects of employment on FAFH are found n prevous studes, despte the dfferent methodology and varables are used. Stewart and Yen (2004), also usng the sample selecton system approach, fnd postve mpact of work hour by household managers on the probabltes of consumng FAFH. Yen (1993), based on much earler Consumer Expendture Survey data, fnd that wfe s work hour has a postve effect on lunch and dnner away from home. More recently, Lu et al. (2012) fnd employment has a postve effect on fast food and full-servce expendtures for husband-wfe households wth and wthout chldren. Postve effects of employment statuses have also been reported for other countres, such as Chna (Ba et al. 2010), Span (Mutlu and Graca 2006), and Tawan (Chang and Yen 2010). Race plays a role n FAFH expendtures. Compared to Blacks, Whte households are 14.26% more lkely to consume at full-servce restaurants and spend $18.77 ($15.91) more per equvalent scale per two weeks, condtonal (uncondtonal) on consumpton. Households of other races are 9.12% more lkely to consume and also consume $37.89 ($29.49) more per equvalent scale per two weeks at full-servce restaurants, condtonal (uncondtonal) on consumpton; these households also consume more at fast-food and other facltes. Educaton has postve effects on FAFH. Compared to households headed by an ndvdual wth hgh school educaton, households less than hgh school educaton are less lkely to consume and also consume less at full-servce restaurants, whle households wth college and post-graduate educaton are more lkely to consume and also consume more at full-servce restaurants. Households wth post-graduate educaton are 10.11% more lkely to consume at full-servce

17 15 restaurants and spend $17.55 ($16.21) more per equvalence scale per two weeks than households wth only hgh-school educaton. College and post-graduate educaton also ncrease expendtures at fast-food and at other facltes. Unlke studes for the general populaton n whch the roles of SNAP are less clear (Lu et al. 2012), we fnd consstent effects of SNAP for the elderly. Partcpaton n SNAP decreases the probabltes (except fast food), condtonal levels, and uncondtonal levels of expendtures at all three facltes. The effects on full servce are partcularly notable, wth SNAP partcpaton decreasng the probablty of consumng at full-servce restaurants by 5.83%, condtonal level by $19.33, and uncondtonal level by $13.37 per equvalence scale per two weeks. Fast food s a substtute for SNAP-partcpatng elderly n terms of probablty (postve), although ts effects on condtonal level ($10.48) and uncondtonal level ($7.43) are also both negatve. Regonal dfferences are also present. Compared to households n the West, households from the Northeast are 2.51% more lkely to consume at other facltes; they also spend $2.55 ($1.82) more on fast food, condtonal (uncondtonal) on consumpton. Households n the Mdwest spend less at full-servce restaurants; they are also more lkely to spend at other facltes. Fnally, seasonal effects are scant and barely notceable, suggestng more consumpton of fast food durng summer months and more at other facltes durng the fall. Concludng remarks Much of the emprcal lterature addresses FAFH by the general populaton n the U.S. and other countres. In ths study, we nvestgate FAFH by elderly households n the U.S., focusng on expendtures by type of facltes. We fnd dfferentated effects of ncome, educaton, employment statuses, and other soco-demographc characterstcs on FAFH expendtures across facltes by the elderly households. These effects dffer from those among younger households. The elderly

18 16 households do have dfferent consumng patterns. Although not drectly comparable due to the use of dfferent econometrc approaches, measures, and varables, our results generally echo fndngs for other countres that economc and soco-demographc characterstcs play key roles n FAFH expendtures. The fndngs of ths study can nform polcy delberatons by federal and state governments. Our estmates dentfy segments of the elderly populaton that should be targeted for nutrton educaton. The educated spend more on FAFH than the less educated, as do the employed than the unemployed/retred. Therefore, nutrton educators could focus ther efforts on employed elderly wth busy workng schedules. Such efforts nclude warnngs about the relatvely hgher levels of sodum, cholesterol, and saturated fats n FAFH meals, recommendatons about healthy FAFH choces such as fruts, vegetables, mlk, and ols, and educatonal messages about moderatng consumpton of fats, added sugars, and alcohol. Our estmate of the effect of SNAP partcpaton on FAFH s also nformatve for polcymakers whose goal s more nutrtous and adequate food. Our fndng suggests that the concern about SNAP promotng consumpton of less healthy FAFH s groundless, for the elderly populaton, snce we fnd the opposte. The results of the study can also assst marketng strateges by foodservce frms. The strong mpact of household composton changes on FAFH spendng provdes valuable nformaton for the foodservce ndustry. For nstance, we fnd household composton varables generally have negatve effects on FAFH. Thus, promotonal campagns offerng quantty dscounts can be an effectve tool for restaurants to attract large famles. Senor ctzen dscounts could also be used as age reduces the probablty and levels of fast food. Such dscounts are currently avalable n many restaurants n the U.S. Some conclusons about future trends of FAFH consumpton can also be drawn. Income

19 17 contrbutes to expendtures at both full-servce and fast-food facltes. Ths suggests the restaurant ndustry wll suffer n the case of a recesson but wll prosper when the economy recovers.

20 18 References Angulo, A. M., Gl, J. M., & Mur, J. (2007). Spansh demand for food away from home: analyss of panel data. Journal of Agrcultural Economcs, 58(2), Ba, J., Wahl, T.I., Lohmar, B.T., & Huang, J. (2010). Food away from home n Bejng: effects of wealth, tme and free meals. Chna Economc Revew, 21(3), Becker, G. S. (1965). A theory of the allocaton of tme. Economc Journal, 75: Bnkley, J. K. (2006). The Effect of demographc, economc, and nutrton factors on the frequency of food away from home. The Journal of Consumer Affars, 40(2), Byrne, P.J., Capps, O., Jr., & Saha, A. (1996). Analyss of food-away-from-home expendture patterns for U.S. households, Amercan Journal of Agrcultural Economcs, 78(3), Chang, H., & Yen, S. T. (2010). Off-farm employment and food expendtures at home and away from home. European Revew of Agrcultural Economcs, 37(4), Cherln, J. A. (2010). Demographc trends n the Unted States: a revew of research n the 2000s. Journal of Marrage and Famly, 72(3), Chung, C., & Myers, S. L. (1999). Do the poor pay more for food? An analyss of grocery store avalablty and food prce dspartes. Journal of Consumer Affars, 33(2), Ctro, C. F., & Mchael, R. T. (1995). Measurng poverty: a new approach. Washngton, DC: Natonal Academy Press. Engle, R. F. (1984). Wald, lkelhood rato, and Lagrange multpler tests n econometrcs. In Z. Grlches & M. D. Intrlgator (Eds.), Handbook of Econometrcs, Vol. 2 (pp ). Amsterdam: Elsever. Heckman, J. J. (1979). Sample selecton bas as a specfcaton error. Econometrca, 47(1),

21 19 Jensen, H. H., & Yen, S. T. (1996). Food expendtures away from home by type of meal. Canadan Journal of Agrcultural Economcs, 44(1), Kotz, S., Balakrshnan, N., & Johnson, N. L. (2000). Contnuous multvarate dstrbutons, vol. 1: models and applcatons, 2nd edn. New York: John Wley & Sons. Lu, M., Kasterds, P., & Yen, S. T. (2012). Who are consumng food away from home and where? results from the Consumer Expendture Surveys. European Revew of Agrcultural Economcs 39, do: /erae/jbs012> Mancno, L., Todd, J., & Ln, B. (2009). Separatng what we eat from where: measurng the effect of food away from home on det qualty. Food Polcy, 34(6), McCracken, V. A., & Brandt, J. A. (1987). Household consumpton of food-away-from-home: total expendture and by type of food faclty. Amercan Journal of Agrcultural Economcs, 69(2), Mnhalopoulos, V. G., & Demousss, M. P. (2001). Greek household consumpton of food away from home: a mcroeconometrc approach. European Revew of Agrcultural Economcs, 28(4), Mutlu, S., & Graca, A. (2006). Spansh food expendture away from home (FAFH): by type of meal. Appled Economcs, 38(9), Pngal, P. (2007). Westernzaton of Asan dets and the transformaton of food systems: mplcatons for research and polcy. Food Polcy, 32(3), Shaw, W. D., & Feather, P. (1999). Possbltes for ncludng the opportunty cost of tme n recreaton demand systems. Land Economcs, 75(4), Soberon-Ferrer, H., & Dards, R. (1991). Determnants of household expendtures for servces. Journal of Consumer Research, 17(4),

22 20 Spanos, A. (1999). Probablty theory and statstcal nference: econometrc modelng wth observatonal data. Cambrdge: Cambrdge Unversty Press. Stewart, H., & Yen, S. T. (2004). Changng household characterstcs and the away-from-home food market: a censored equaton system approach. Food Polcy, 29(6), Tan, A. K. G. (2010). Demand for food-away-from-home n Malaysa: a sample selecton analyss by ethncty and gender. Journal of Foodservce Busness Research, 13(3), U.S. Bureau of Labor Statstcs (U.S. BLS) (2009, 2010, 2011). Consumer Expendture Survey, 2008, 2009, US Department of Labor. U.S. Census Bureau. (No date) and 2009 County Busness Patterns (NAICS). Accessed 3 June USDA-ERS (USDA-ERS) (2011) Brefng Room: Food CPI and Expendtures: Food Expendture Tables. US Department of Agrculture, Economc Research Servce. Avalable from Data/Expendtures_tables/. (Accessed 1 January 2012). Vuong, Q. H. (1989). Lkelhood rato tests for model selecton and non-nested hypotheses. Econometrca, 57(2), Yen, S. T. (1993). Workng wves and food away from home: the Box-Cox double hurdle model. Amercan Journal of Agrcultural Economcs, 75(4), Yen, S. T. (2005). A multvarate sample-selecton model: Estmaton cgarette and alcohol demands wth zero observatons. Amercan Journal of Agrcultural Economcs, 87(2), Yen, S. T., & Rosńsk, J. (2008). On the margnal effects of varables n the log-transformed sample selecton models. Economcs Letters, 100(1), 4 8.

23 21 Table 1 Sample statstcs of FAFH expendtures per equvalent scale by faclty Expendture % Full Sample Consumng Sample ($ / 2 weeks) Consumng Mean SD Mean SD Elderly households (n = 7,860) Full-servce restaurants Fast-food restaurants Other facltes Younger (n = 12,663) Full-servce restaurants Fast-food restaurants Other facltes

24 22 Table 2 Margnal effects of explanatory varables on probabltes, condtonal levels, and uncondtonal levels among elderly households Full-servce Restaurants Fast-food Restaurants Other Facltes Varable Probablty Level (C) Level (U) Probablty Level (C) Level (U) Probablty Level (C) Level (U) Contnuous explanatory varables Income / *** 2.722*** 2.960*** 1.564*** 0.553** 0.740*** *** 0.081** (0.263) (0.354) (0.262) (0.265) (0.223) (0.165) (0.231) (0.138) (0.033) Members < *** 3.968*** 3.331** *** 0.968** 0.485*** (1.293) (2.080) (1.327) (1.393) (0.864) (0.698) (0.969) (0.448) (0.127) Members *** 7.026*** 2.559*** 4.676*** 1.039* *** 1.278*** (0.808) (1.164) (0.814) (0.816) (0.557) (0.442) (0.656) (0.342) (0.085) Members > *** 4.745*** *** 2.522*** * 2.023*** 0.292** (1.029) (1.471) (1.012) (1.015) (0.731) (0.551) (0.894) (0.527) (0.126) Age / *** *** 6.051*** 2.496*** 3.100*** 2.096*** (0.821) (1.220) (0.829) (0.790) (0.599) (0.442) (0.763) (0.453) (0.110) Bnary explanatory varables Urban *** 6.617*** ** 1.869* (2.160) (2.551) (1.769) (2.129) (1.434) (1.059) (1.850) (0.837) (0.229) Homeowner 8.471*** 4.812** 6.621*** 5.058*** ** (1.440) (2.015) (1.262) (1.413) (1.033) (0.741) (1.231) (0.791) (0.193) SNAP *** *** *** *** 4.721*** *** 0.647*** (2.739) (2.747) (1.501) (2.590) (1.406) (1.025) (2.324) (0.954) (0.226) Whte *** *** *** ** 0.445** (1.842) (2.078) (1.248) (1.756) (1.210) (0.886) (1.496) (0.800) (0.192) Other race 9.117*** *** *** *** 5.039* 5.214* 6.572*** 2.082*** (3.048) (8.107) (5.835) (3.051) (2.592) (2.032) (2.839) (2.501) (0.753) < Hgh school 8.736*** *** 3.941** (1.716) (2.506) (1.530) (1.634) (1.231) (0.878) (1.421) (0.909) (0.217) College 5.472*** 9.325*** 8.089*** ** 1.846*** 2.248** * (1.289) (1.964) (1.355) (1.275) (0.925) (0.699) (1.095) (0.593) (0.153) Grad school *** *** *** * 2.068** 4.342*** ** (1.776) (3.156) (2.351) (1.813) (1.336) (1.037) (1.615) (0.879) (0.259) Northeast ** 1.817* 2.505* (1.648) (2.463) (1.697) (1.661) (1.241) (0.947) (1.483) (0.827) (0.224) Mdwest *** 5.624*** *** ** (1.619) (2.140) (1.459) (1.620) (1.136) (0.854) (1.442) (0.795) (0.217) South ** (1.512) (2.184) (1.478) (1.510) (1.066) (0.813) (1.301) (0.690) (0.181) Sprng * (1.496) (2.241) (1.527) (1.449) (1.058) (0.805) (1.260) (0.721) (0.187) Summer ** 1.775** (1.513) (2.146) (1.479) (1.469) (1.121) (0.839) (1.283) (0.704) (0.180) Fall ** * (1.493) (2.144) (1.461) (1.475) (1.051) (0.799) (1.292) (0.682) (0.185) Year (1.299) (1.889) (1.296) (1.275) (0.902) (0.677) (1.047) (0.589) (0.146) Year ** 2.379* *** 1.565** (1.303) (1.860) (1.261) (1.291) (0.931) (0.692) (1.052) (0.648) (0.155) Work FT (H) 9.122*** 9.079*** 9.947*** *** 4.188*** 5.555*** 9.810*** 1.547*** 1.097*** (1.215) (1.288) (1.400) (1.234) (0.531) (0.673) (1.282) (0.523) (0.176) Work PT (H) 3.609*** 3.568*** 3.884** 6.176*** 2.393*** 3.156*** 6.170*** 0.985** 0.716*** (1.367) (1.383) (1.519) (1.379) (0.566) (0.742) (1.413) (0.389) (0.189) Work FT (S) 9.716*** 9.878*** *** 5.428*** 2.089*** 2.757*** 5.326*** 0.850** 0.607*** (1.429) (1.577) (1.771) (1.527) (0.615) (0.812) (1.330) (0.341) (0.167) Work PT (S) 6.804*** 6.903*** 7.630*** 6.125*** 2.391*** 3.166*** 4.337** 0.694* 0.499** (1.906) (2.041) (2.304) (1.926) (0.794) (1.058) (1.833) (0.368) (0.228) Asymptotc Standard errors n parentheses. Astersks ndcate levels of sgnfcance: *** = 1%, ** = 5%, * = 10%.

25 Appendx Table A1 Defntons and sample statstcs of explanatory varables Varable Defnton Elderly Younger Contnuous explanatory varables Income Pre-tax non-wage ncome per equvalent scale n past months n $1,000 (21.47) (11.01) Members < 18 Number of chldren age < (0.44) (1.19) Members Number of adults age (1.05) (0.83) Members > 64 Number of adults age > (0.78) (0.19) Age Age of household head (9.53) (9.82) Bnary explanatory varables (yes = 1, no = 0) Urban Resded n urban area Homeowner Owned a home SNAP Any member receved food stamps last year Whte Race s whte Other race Race s of other race Black Race s black (reference) < Hgh school Has less than hgh school Hgh school Hgh school graduate (reference) College Has a bachelor s or some college Graduate Has a graduate degree Northeast Resded n the Northeast Mdwest Resded n the Mdwest South Resded n the South West Resded n the West (reference) Sprng Survey occurred durng sprng Summer Survey occurred durng summer Fall Survey occurred durng fall Wnter Survey occurred durng wnter (reference) Year 2008 Data came from year 2008 (reference) Year 2009 Data came from year Year 2010 Data came from year Work FT (H) Household head worked full-tme Work PT (H) Household head worked part-tme Not workng (H) Household head not workng (reference) Work FT (S) Spouse present spouse worked full-tme Work PT (S) Spouse present spouse worked part-tme Not workng (S) Spouse absent or not workng (reference) Sample sze 7,860 12,663 Standard devatons n parentheses. Household head s defned as the husband for a marred household, and as the reference person for a sngle-person household. All households do not have a spouse present and, therefore, spouse workng statuses reflect ther nteractons wth a spouse present dummy ndcator. 23

26 24 Table A2 ML estmaton of sample selecton system wth heteroscedastc errors: FAFH expendtures by elderly households Selecton Equatons Level Equatons Varable Full-serv. Fast-food Other Full-serv. Fast-food Other Constant 0.396** 0.844*** 0.860*** 4.069*** 3.388*** 2.249*** (0.183) (0.183) (0.226) (0.229) (0.203) (0.500) Income / *** 0.044*** *** (0.007) (0.008) (0.009) (0.008) (0.008) (0.016) Members < ** 0.142*** 0.104** (0.035) (0.039) (0.037) (0.050) (0.037) (0.065) Members *** 0.132*** 0.105*** 0.204*** 0.110*** 0.234*** (0.022) (0.023) (0.025) (0.027) (0.024) (0.052) Members > *** 0.153*** 0.066* 0.212*** 0.180*** 0.312*** (0.028) (0.029) (0.034) (0.035) (0.032) (0.072) Age / *** 0.170*** 0.080*** (0.023) (0.023) (0.029) (0.029) (0.027) (0.069) Urban *** 0.114* (0.059) (0.060) (0.068) (0.073) (0.069) (0.131) Homeowner 0.228*** 0.140*** ** 0.240** (0.039) (0.039) (0.048) (0.052) (0.045) (0.099) SNAP 0.369*** ** 0.284*** 0.444** (0.075) (0.072) (0.092) (0.110) (0.081) (0.201) Whte 0.386*** ** (0.051) (0.049) (0.059) (0.068) (0.055) (0.130) Other race 0.255*** * 0.401*** 0.211** 0.545*** (0.088) (0.087) (0.096) (0.114) (0.090) (0.196) < Hgh school 0.235*** 0.109** * 0.101* (0.046) (0.045) (0.056) (0.063) (0.053) (0.121) College 0.149*** ** 0.079* (0.035) (0.036) (0.041) (0.044) (0.040) (0.084) Grad school 0.279*** *** 0.120** (0.050) (0.052) (0.056) (0.060) (0.054) (0.114) Northeast * * (0.045) (0.047) (0.054) (0.056) (0.051) (0.113) Mdwest *** 0.118** (0.044) (0.045) (0.052) (0.055) (0.050) (0.109) South ** (0.041) (0.043) (0.049) (0.052) (0.046) (0.100) Sprng (0.041) (0.041) (0.047) (0.052) (0.045) (0.096) Summer

27 25 (0.041) (0.042) (0.049) (0.051) (0.047) (0.098) Fall ** (0.041) (0.042) (0.047) (0.051) (0.045) (0.095) Year (0.036) (0.036) (0.041) (0.044) (0.039) (0.081) Year * 0.128*** *** (0.036) (0.036) (0.042) (0.045) (0.040) (0.085) Work FT (H) 0.250*** 0.308*** 0.350*** (0.034) (0.036) (0.043) Work PT (H) 0.099*** 0.177*** 0.222*** (0.038) (0.041) (0.048) Work FT (W) 0.267*** 0.155*** 0.193*** (0.040) (0.044) (0.046) Work PT (W) 0.188*** 0.177*** 0.157** (0.053) (0.057) (0.063) Het. specfcaton Constant 0.276*** 0.181*** 0.361*** (0.019) (0.019) (0.069) Income / ** ** (0.004) (0.005) (0.011) Error correlatons (ρ j ) Fast-food (selecton) 0.529*** (0.015) Other (selecton) 0.389*** 0.490*** (0.020) (0.021) Full-serv. (level) 0.806*** 0.245*** 0.201*** (0.016) (0.019) (0.019) Fast-food (level) 0.107*** 0.699*** 0.082*** 0.195*** (0.020) (0.025) (0.021) (0.016) Other (level) 0.121** 0.227*** 0.510*** 0.181*** 0.128*** (0.054) (0.066) (0.131) (0.030) (0.028) Log lkelhood Astersks ndcate level of sgnfcance: *** = 1%, ** = 5%, * = 10%.