B R E A D Working Paper

Transcription

1 Do entry costs provde an emprcal bass for poverty traps? Evdence from Mexcan mcroenterprses Davd McKenze and Chrstopher Woodruff BREAD Workng Paper No. 020 February 2003 Copyrght 2003 Davd McKenze and Chrstopher Woodruff B R E A D Workng Paper Bureau for Research n Economc Analyss of Development

2 Do entry costs provde an emprcal bass for poverty traps? Evdence from Mexcan mcroenterprses Davd McKenze and Chrstopher Woodruff BREAD Workng Paper No. 020 February 2003 JEL Codes: O12, J23, J62, D31 Keywords: non-convextes, poverty-trap, start-up costs, mcroenterprses ABSTRACT Recent theoretcal lterature n development economcs has shown that non-convex producton technologes can result n low-growth poverty traps. Ths paper uses detaled mcroenterprse surveys n Mexco to examne the emprcal evdence for these non-convextes at low levels of captal stock. Whle theory emphaszes non-dvsble start-up costs whch exceed the wealth of many potental entrepreneurs, ths paper shows start-up costs to be very low n some ndustres. Sem-parametrc methods are then used to flexbly estmate returns to captal n mcroenterprses. Much hgher returns are found at low levels of captal stock than at hgher levels, and ths remans true after controllng for frm characterstcs and measures of entrepreneural ablty. Overall, lttle evdence of producton non-convextes s found at low levels of captal. The absence of nonconvextes s a sgnfcant fndng because t suggests that access to startup captal does not determne the ultmate sze of the enterprse. Davd McKenze Department of Economcs Stanford Unversty mcken@stanford.edu Chrstopher Woodruff Graduate School of Internatonal Relatons and Pacfc Studes UCSD

3 Do Entry Costs Provde an Emprcal Bass for Poverty Traps? Evdence from Mexcan Mcroenterprses # Davd J. McKenze * and Chrstopher Woodruff ** * Department of Economcs, Stanford Unversty ** Graduate School of Internatonal Relatons and Pacfc Studes, UCSD Abstract Recent theoretcal lterature n development economcs has shown that non-convex producton technologes can result n low-growth poverty traps. Ths paper uses detaled mcroenterprse surveys n Mexco to examne the emprcal evdence for these non-convextes at low levels of captal stock. Whle theory emphaszes non-dvsble start-up costs whch exceed the wealth of many potental entrepreneurs, ths paper shows start-up costs to be very low n some ndustres. Sem-parametrc methods are then used to flexbly estmate returns to captal n mcroenterprses. Much hgher returns are found at low levels of captal stock than at hgher levels, and ths remans true after controllng for frm characterstcs and measures of entrepreneural ablty. Overall, lttle evdence of producton non-convextes s found at low levels of captal. The absence of non-convextes s a sgnfcant fndng because t suggests that access to startup captal does not determne the ultmate sze of the enterprse. JEL Classfcatons: O12, J23, J62, D31. Keywords: non-convextes, poverty trap, start-up costs, mcroenterprses. # We thank partcpants at semnars at the Unversty of Southern Calforna and Stanford Unversty for ther helpful comments and suggestons and Bll Maloney for help usng the ENAMIN data

4 1. Introducton There s a growng body of theoretcal work concerned wth the nteracton between captal market mperfectons and non-convex producton technologes. 1 Banerjee and Newman (1993), for example, develop a dynamc model wth rsky nvestment projects requrng an ndvsble start up cost whch exceeds the wealth level of at least some of the agents n the economy. Low-wealth entrepreneurs are unable to fnance proftable ventures. Entrepreneural actvty nvolvng a mnmum nvestment level s also a feature of the models of Aghon and Bolton (1997) and Lloyd-Ells and Berhardt (2000). In combnaton wth poorly functonng credt markets, the nonconvextes have several mplcatons. Frst, there s too lttle entry of new frms. Second, producton s neffcent because some of the most talented entrepreneurs cannot obtan the amount of captal necessary to overcome mnmum scale nvestments. Thrd, the ncentve for the poor to save s reduced because nonconvextes result n low returns to low levels of nvested captal, and poorly functonng captal markets result n low returns to fnancal savngs. The models, then, have mportant mplcatons for the path of development. They mply that an economy may land n a low-growth poverty trap from whch those wth lttle captal to nvest can nether bootstrap ther way up by renvestng profts over tme nor borrow ther way above the mnmum scale. Relaxng ether of the assumptons generally elmnates the poverty trap outcome. 2 In partcular, f producton technologes are convex, then ntal condtons wll not affect the long run path of the economy. 1 Banerjee (2001) provdes an excellent recent survey. 2 Mookherjee and Ray (2001) show that poverty traps can arse even wth convex producton technologes. They develop a model wth moral hazard n whch prncpals are matched wth agents each perod. When prncpals have all of the barganng power n the relatonshp, they must provde agents wthout any wealth rents n order to nduce effort. An agent s wealth acts as collateral for promsed effort levels, removng the need for rents. Snce wealth accumulaton by agents s effectvely taxed n ths manner, return to wealth accumulaton s low

5 In ths paper, we examne the emprcal relevance of the assumptons that mnmum start-up costs are hgh relatve to wealth and that returns to captal are low at low nvestment levels. That s, do producton non-convextes prevent entry to selfemployment or result n low returns from enterng wth small amounts of captal? Usng data from an extremely detaled survey of mcroenterprses n Mexco, we frst show that the medan nvestment levels of new frms n some sectors are very low, less than US $100 n constructon and personal servces, an amount less than half of monthly earnngs of even a low wage worker. Next, we show that the margnal return to captal nvested n these sectors s qute hgh for low levels of nvested captal. We have a large sample of mcroenterprses and so are able to use sem-parametrc estmaton methods whch allow the returns to vary flexbly over the range of captal stock. We fnd that returns to captal are above 15 percent a month for nvestment levels below $US 200, substantal for nvestment levels below $US 1000, and more moderate for nvestment levels above $1000. Returns are always postve, and whle our pont estmates ndcate ncreasng returns between $1000 and $2000 of captal, the estmates are not precse enough to statstcally reject that returns are non-ncreasng. Our estmates of returns to nvested captal at captal stock levels n excess of $1000 are n lne wth nterest rates offered by mcrofnance organzatons and other lenders. The much hgher returns experenced by low captal frms are more puzzlng. We consder several possble explanatons for the very hgh rates of return. Frst, we consder the possblty our estmates may be based upward due to a lack of controls for dfferences n entrepreneural ablty. The mcroenterprse data we use are cross sectonal, and hence we are not able to control for ablty wth ndvdual fxed effects. Instead, we use the panel labor market survey from whch our mcroenterprse survey s drawn to obtan three measures of entrepreneural ablty, whch are dscussed n more detal later n the paper. Although there s a postve correlaton between measures of the entrepreneur s ablty and frm earnngs, ncluson of the ablty controls does not result n - 3 -

6 a marked change to our estmated of return to captal. Moreover, we fnd some confrmaton of the hgh rates of return n the fact that ncome growth n the quarters followng the ENAMIN sample perod s hgher among those wth low captal stocks than among those wth hgher captal stocks. Next, we explore the possblty that captal constrants, learnng, or dfferences n rsk, may explan the hgh returns to captal. Entrepreneurs recevng the hghest returns are found to be those wth lowest wealth and for whom captal constrants may apply. We also fnd that the varaton n earnngs ncreases and the return to captal decreases wth the level of nvested captal stock, suggestng that the hgh returns do not reflect compensaton for ncreased rsk. The modest medan nvestment levels and the hgh returns to small captal nvestments suggest that non-convex producton technologes play no role n preventng entry nto the mcroenterprse sector. The absence of non-convextes s a sgnfcant fndng because t suggests that access to startup captal does not determne the ultmate sze of the enterprse. Gven the pattern of returns to captal we fnd, entrepreneurs wthout startup captal can pull themselves up by ther bootstraps by renvestng ther profts. We note two caveats to these rather optmstc fndngs. Frst, the data suggest that non-convextes may be mportant n certan sectors of actvty, especally transportaton and professonal servces. These are the sectors n whch returns to captal appear to be hghest at hgher levels of nvested captal. Thus, the bootstrappng strategy may have lmts f there are costs of swtchng sectors of actvty. Second, as we noted above, there s some suggeston n the data of non-convextes n the $1000-$2000 range. Ths suggests the possblty of a hurdle whch must be overcome along the frm s growth path. Levy (1993) suggests the possblty of a threshold effect, whereby fscal and bureaucratc costs are only faced by frms above a mnmum sze. Banerjee (2001) notes that non-convex costs may also arse from hrng labor, due to montorng costs. Ths - 4 -

7 paper s unable to rule out the possblty that such non-convextes prevent small busnesses from becomng larger, suggestng that future emprcal work on poverty traps should examne larger frms. In comparson to the theorsts level of nterest n nonconvextes, there has been relatvely lttle emprcal work examnng the returns to nvestment among mcroenterprses. Ths s due at least n part to a lack of data. Surveys whch nclude data on nvestments and profts of mcroenterprses are stll relatvely rare. Hll and Kalrjan (1993), Bgsten et al. (2000), and Hernández, Pagán and Paxton (2002) are the closest studes we are aware of to our own. Hll and Kalrjan (1993) estmate returns to captal n the Indonesan garment ndustry, usng a sample of 2250 frms taken from that country s Census of Small Industry. Bgsten et al (2000) use data from surveys of manufacturers n 5 Afrcan countres. Hernández, Pagán and Paxton (2002) use the same data we do to estmate the techncal effcency of the frms. All of these studes fnd hgh rates of return to nvested captal. These studes do not speak drectly to the queston of nonconvextes, however, because each specfes a translog producton functon. Gven the range n sze of frms ncluded n ther studes, the translog specfcaton smoothes the data too much, potentally maskng mportant varaton n margnal returns to captal across levels of captal nvestment. A second emprcal lterature examnes how the occupatonal choce of workers s affected by access to nvestment captal. Evans and Jovanovch (1989), and Blanchflower and Oswald (1998) fnd an assocaton between entry nto self employment and an ndvdual s assets usng data from the US and Great Brtan, respectvely. Paulson and Townsend (2000) fnd that havng more household assets makes entry nto self employment more lkely n Thaland. These studes may be nterpreted as provdng ndrect evdence of the presence of a mnmum scale of producton. However, these studes do not necessarly mply non-convex producton as there may be credt-related or - 5 -

8 rsk averson explanatons for the relatonshp found. Moreover, these studes do not provde estmates of the margnal returns to captal at low levels of nvestment. The remander of ths paper s structured as follows. Secton 2 examnes the mportance of non-convextes for entry to self-employment. Secton 3 dscusses semparametrc estmaton of margnal returns to captal and provdes results at the aggregate and ndustry level. In Secton 4 controls are made for ablty bas and evdence s provded on ncome growth for dfferent captal stock ranges. Several explanatons for the hgh returns to captal at low captal stocks are offered n Secton 5, whle Secton 6 concludes. Appendx A provdes more detal on constructon of the data. 2. How Important are Startup Costs? We use data from Mexco s Natonal Survey of Mcro Enterprses (ENAMIN), gathered n 1992 (1 st quarter), 1994 (1 st quarter), 1996 (1 st quarter), and 1998 (4 th quarter). The survey samples frms wth 5 or fewer workers, except n manufacturng, where the upper lmt s 15. The survey s representatve of all urban areas wth populatons of 100,000 or more (INEGI, 2000). The ENAMIN sample s drawn from Mexco s quarterly urban labor survey (the ENEU), whch s tself a household-based survey. About 10,000 frms were surveyed n each of the 4 years. We focus n ths paper on frms owned by males, whch represent about two-thrds of the frms n the sample, and whose owners are between 18 and 65 years of age and who work full tme (35 or more hours per week). Whle there are mnor dfferences n the survey nstruments, the data appear to be comparable across the 4 surveys. We translate all peso value responses to 1998 $US, usng the Mexcan GDP deflator and the peso/dollar exchange rate for the 4 th quarter of The data are descrbed n more detal n Appendx A. 3 The data are frst translated to 1998 pesos usng the GDP deflator and then translated to dollars so that the volatlty of the peso/dollar exchange rate does not affect the data

9 There s lttle formal fnancal ntermedaton among the ENAMIN frms. Less than 1% report recevng start-up fnancng from a bankng nsttuton and less than 6% report ever havng had a bank loan. A pror, then, there s reason to expect that captal constrants may be bndng. Table 1 examnes how mportant start-up costs are n generatng non-convextes for mcroenterprses n Mexco. We take frms whch have been n exstence for two years or less, and n Panel A nspect the dstrbuton of ther captal stock by ndustry. Large dfferences are seen across ndustres. In constructon and personal servces, 24% and 47% of frms report that no ntal fundng was needed to begn the busness, wth the lower quartle of frms n these ndustres havng total captal stock less than $US 47. In contrast, manufacturng, trade, and transportaton requre much bgger ntal outlays. Less than seven percent of frms n these ndustres report that no ntal fundng was needed, whle the 25th percentle of captal stock s $US 281 for manufacturng, $US 329 for trade, and $US 2070 for transportaton. To llustrate the mportance of these start-up costs relatve to ncome levels, we calculate for each ndustry the number of months of earnngs the 10th and 25th captal stock percentles represent. The last four columns of Panel A report ths n terms of both the tenth percentle (low wage) and the medan of the earnngs dstrbuton for wage workers n the fourth quarter of In personal servces and constructon, the 25th percentle of ndustry captal amounts to only 0.1 and 0.4 months of low wage earnngs, compared to 2.6 months for manufacturng and 19.3 months for transportaton. In other words, a low wage worker savng 10% of hs ncome would requre only one to four months to acqure suffcent captal to enter the constructon and personal servces sector at the 25th percentle of captal stock level, 9 months to enter repar servces, 31 months to enter trade, 50 months to enter professonal servces, and 193 months to enter the - 7 -

10 transportaton ndustry 4. Under a 10% savngs rate, even a medan wage worker would requre 41 months of savng to enter the transportaton ndustry at only the 10th percentle of ndustry captal. In addton to the costs of physcal captal, start-up costs may nclude labor and montorng costs for hred workers (Banerjee and Newman, 1993), search and permt costs for leasng premses, contractng costs for acqurng nventores 5, and the bureaucratc costs of regsterng and dealng wth government and other agences (Levy, 1993). The dollar amounts of these types of start-up costs can be dffcult to measure, so n Panel B of Table 1 we smply report the percentage of frms n each ndustry who do not face these costs. Agan constructon and personal servces seem to be the lowest startup cost ndustres, wth the majorty of frms n these sectors not subject to any of these addtonal costs. Transportaton, professonal servces, and trade are all more subject to bureaucratc requrements, and n addton may requre acqurng premses and nventores pror to start-up. From Table 1 we conclude that fxed costs are unlkely to be an mportant source of non-convexty n terms of provdng barrers to entry nto self-employment. However, access to startup captal may affect the partcular ndustry a gven entrepreneur chooses to enter. Entry nto the constructon, personal servces and repar servces ndustres requres very lttle physcal captal, nvolves low bureaucratc obstacles, and does not requre hrng pad workers or obtanng formal busness premses. 3. Returns to captal Havng entered self-employment, entrepreneurs may be confronted wth nonconvextes n producton. Banerjee (2001, p.34) remarks that non-convextes n 4 We gnore nterest on savngs n constructng these estmates, whch may be reasonable gven Mexco s underdeveloped fnancal system. 5 The drect costs of nventores and leasng premses are ncluded n the physcal captal amounts n Panel A

11 producton are certanly very plausble, snce most machnes have a mnmum effcent scale, and rental markets are at best an mperfect substtute. The presence of these nonconvextes can lead to poverty traps n theoretcal models. Furthermore, f returns to captal are low at low levels of nvestment, then the poor get low returns, reducng ther ncentve to save and thereby renforcng the poverty trap. We begn by examnng the returns to captal n the cross-secton of entrepreneurs. It s mportant to allow the returns to captal to vary accordng to the level of captal employed, and so earnngs of the self-employed are modeled semparametrcally. A ffth-order polynomal n captal s also estmated for comparson purposes. Let π denote the monthly earnngs of self-employed worker, gven by the response to the survey queston: How much dd you make n profts, after subtractng expendtures. Let K be the non-rented captal stock employed by the frm, measured as the replacement cost of real estate, machnery, tools, vehcles, and other equpment owned by the enterprse, plus the value of any nventores and work n progress. 6 Fnally, let Z be a vector of other varables whch potentally affect earnngs, such as the age and educaton of the entrepreneur. We motvate the varables ncluded n Z below. The selfemployment earnngs functon s then modeled by a partal lnear model: ( K ) + Z β ε π = h ' + (1) Sem-parametrc estmaton of equaton (1) s carred out n two steps. Frst, the parameter vector β s estmated usng the hgher-order dfferencng method of Yatchew (1997), wth a dfferencng order of fve. The data s sorted n ascendng order of the captal stock K. Let d 0, d 1,,d 5 denote the optmal dfferencng weghts gven n Yatchew (1998, p. 697). Then β s obtaned by least squares estmaton of the equaton: 6 We exclude equpment and real estate whch s rented rather than owned, snce a sgnfcant porton of the return to ths captal s the rental prce whch s not ncluded n the frm s profts. Inventores average 8 percent of the value of captal stock for frms wth captal stock below $200 and 10 percent of captal for captal stock between $1000 and $5000. Most busness captal s owned; rented captal averages 4.3 percent of total captal for frms wth below $200 of captal stock and 14.7 percent for frms wth $

12 5 j= 0 5 d jπ j d j Z = j β + η j= 0 (2) Ths ffth-order dfferencng acheves 91 percent effcency relatve to the asymptotc effcency bound and does not requre nonparametrc estmaton to estmate β. Let βˆ be the estmate of β obtaned from (2). Then the second step n estmaton s to use nonparametrc methods to estmate the functon h(.) and ts dervatve n the equaton: ( K ) v π Z ' ˆ β = h + (3) Estmaton of equaton (3) s carred out usng the local lnear regresson of Fan (1992). In addton to certan optmalty propertes, ths estmator s partcularly desrable for our purposes as t provdes for easy estmaton of the dervatve of the functon h(k) along wth the functon tself, provdng the margnal returns to captal. We use crossvaldaton to select the optmal bandwdth and the Epanechnkov kernel followng the recommendatons of Fan and Gjbels (1996). Some further dscusson of the varables ncluded n the vector Z s warranted. The cross secton of enterprses n our sample s the outcome of ndvduals decsons to enter nto and reman self employed. The alternatve to self employment s work n the wage sector. Hence, we expect the level of profts to be related to wages n the labor market. We measure the alternatve opportunty wth the average wage of full tme workers n the ndustry and state n whch the ndvdual works. (The ndustres are the nne shown n Table 1.) Proft rates wll also be affected by the ablty of the entrepreneur. Our prmary specfcaton measures the ndvdual s productve ablty wth years of schoolng and age. We nclude a varable ndcatng that the owner s marred as labor economsts have found that marred workers generally earn hgher wages, after controllng for other characterstcs. Gnther and Zavodny (2001) revew the varous explanatons and, usng data on shotgun weddngs, fnd that most of the male marrage premum represents a

13 drect productvty effect of marrage rather than a selecton effect. We also nclude the number of hours the owner normally works n the enterprse. About 19% of the enterprses employ unpad famly members. Snce these workers are unpad, we expect the profts from ther efforts to accrue to the owner. We nclude the number of hours normally worked n the enterprse by unpad workers. Returns to manageral ablty are measured by ncludng the number of hours worked by pad employees. Fnally, frms are more lkely to reman n busness f they are more proftable. Therefore, we nclude a measure of the age of the frm and ts square. We also nclude ndcators of the year n whch the survey was undertaken (wth 1998 as the base year) and the ndustry n whch the frm operates (wth constructon as the base ndustry). Table 2 reports the estmates of the controls n equaton (2) usng data from all ndustres and years combned. 7 Returns are estmated separately for low captal ($0- $1000 of captal) and hgher captal frms ($500-$5000). We have about 7000 observatons n the low-captal sample and 5000 observatons n the hgh-captal sample. 8 Parametrc estmaton of equaton (1) usng a ffth-order polynomal n captal n place of h(k) s also provded. 9 Results for two specfcatons are provded. The frst (Model ) specfcaton ncludes the measures of the owner s age, educaton and martal status, the age of the frm, and the number of hours worked by pad and unpad employees. The second (Model ) adds year and ndustry effects to Model. The labor hours of the owner has at most a weakly sgnfcant postve mpact on earnngs, whle the owner s age only has a sgnfcant effect on earnngs for low captal frms. Earnngs ncrease wth frm age up untl the frm s years n busness, after whch earnngs start to fall. Ths may reflect ether learnng by young frms, or selecton as less successful frms have hgher 7 When we estmate returns separately by year we fnd smlar returns each year, and so combne years n order to ncrease the sample sze avalable. 8 Above $5000, the densty of observatons falls, makng relable estmates more dffcult to obtan. 9 For the parametrc estmaton, we add roughly the next 10% of observatons beyond the captal stock cutoff of nterest n order to mtgate boundary effects. Results from parametrc estmaton can be senstve to the nterval over whch the global polynomal n captal s estmated

14 ext rates. The educaton of the owner, whch partally reflects ablty, has a strong postve and nonlnear effect on earnngs. College educaton results n substantally hgher earnngs than any other educaton level. Marred owners earn more. Both unpad and pad labor ncreases earnngs, though pad labor has a much larger mpact. Each (weekly) hour of pad labor ncreases the entrepreneur s earnngs by about $US1 a month. The prevalng wage n the labor market has a hghly sgnfcant postve effect as expected. The ndustry and year effects are sgnfcant and so we work wth Model for the remander of the paper. 10 Fgure 1 dsplays the returns to captal found from estmatng equaton (1) wthout any controls and returns estmated from the models n Table 2. Addng controls reduces the estmated returns, wth Models and provdng smlar results. Returns to captal are found to be extremely hgh at low levels of captal. Measured characterstcs of the owner, workers, and frm account for only a porton of ths hgh return. We fnd returns of above 20 percent per month for captal stock below $200 even after controllng for these measured characterstcs. Returns to captal fall to around 5 percent per month for captal stock of $ Returns to captal are ncreasng over captal stock ranges $ and $ , provdng some evdence of non-convextes n these captal ranges. However, returns reman postve over the $ range of captal, and the hgh returns at low levels of captal do not suggest a pattern whch would be expected to produce a poverty trap. The lower half of Fgure 1 provdes returns to captal when the captal stock s n the $500-$5000 range. Returns are decreasng over the captal stock range $ , but then ncrease to around 5 percent for captal stock of $2000. Some evdence of nonconvex technology s agan seen n returns ncreasng over the ranges $ and 10 The absolute magntudes of these effects are however much lower when the state-ndustry wage varable s ncluded n the regresson, suggestng that ths varable captures a large porton of the year and ndustry effects

15 $ The parametrc returns show the same general shape as the sem-parametrc returns after $1500, although wth more varablty and more dependence on the range of estmaton. Returns reman non-negatve, suggestng that non-convextes are not strong enough to cause a range of negatve returns to captal. The results thus suggest that ncremental changes n captal stock wll generally provde postve returns. The local lnear returns n Fgure 1 show some regons of ncreasng returns. To examne whether these non-convextes are sgnfcant, we carry out 100 bootstrap replcatons of the sem-parametrc estmaton of model, and plot the estmated 95% pontwse confdence ntervals n the top half of Fgure 2. Although the bandwdth used for smoothng s chosen by cross-valdaton to be the optmum, we see qute wde confdence ntervals around the estmated returns, and one can not reject that returns are non-ncreasng over the entre captal stock range up to $5000. In the lower half of Fgure 2 we trade off some bas for lower varance, oversmoothng the returns. Over the low captal regon we fnd the non-convexty smoothed out, and narrow bands around the hgh returns at captal levels below $200, wth returns averagng 4.5% n a confdence nterval of 2-7% for captal stock $ For the returns wth hgher captal, there s stll some slght suggeston of ncreasng returns between $1000 and $2000 after more smoothng, but one can stll not reject non-ncreasng returns over ths regon. We therefore conclude that there s lttle support for non-convextes n the returns to captal below $1000, and only weak evdence for non-convextes n returns to captal above $1000. One possble explanaton for the pattern of returns found n Fgure 2 s that returns dffer across ndustres and that combnng dsparate ndustres mght produce msleadng estmates of returns to captal. We nvestgate ths by reestmatng model n Table 2 separately for sx dfferent ndustres. 11 Local lnear returns were estmated 11 The sx ndustres chosen were those wth most observatons avalable for estmaton

16 between the 5 th and 80 th percentles of ndustry captal stock, and the parametrc returns between the 5 th and 85 th percentles of captal stock. The dollar value of the upper cutoff thus dffers by ndustry, rangng from $673 for constructon to $14,641 for transportaton. The number of observatons avalable for estmaton ranged from 2448 for trade to only 691 for professonal servces. Fgure 3 plots the estmated returns by ndustry. In four of the sx ndustres constructon, repar servces, trade and manufacturng--the pattern of returns s very smlar to that found n the entre sample. Returns to margnal nvestments at low levels of captal exceed 20% per month. The returns declne precptously, fallng below 10% per month at nvestment levels of $ In trade, we fnd a range of ncreasng returns, as was found when ndustres were combned. Returns n the more hghly captal ntensve ndustres--transportaton and professonal servces show a markedly dfferent pattern, remanng always n the 2-3% range. Ths mples a much lower rate of return n these ndustres among frms nvestng small amounts, and comparable or hgher rates of returns among frms nvestng larger amounts. The fact that we observe very few frms n constructon and repar servces wth hgh levels of captal stock would seem to suggest that returns to captal are very low at hgh captal levels n these ndustres, however ths same lack of observatons prevents us from estmatng returns at hgh captal levels. It may be necessary for entrepreneurs to swtch to hgh captal ndustres such as transportaton and professonal servces n order to earn returns on larger amounts of captal stock. Overall, the results from ndustry-level estmaton do not suggest that the returns patterns found n Fgure 2 arse from aggregaton across ndustres, and, n partcular, hgh returns at low levels of captal stock are found n several ndustres. 12 In constructon, the 80 th percentle of captal stock s only $390. The parametrc and local lnear returns dffer most for ths ndustry, but both fnd monthly returns above 10 percent for captal stock below $300. The fact that we do not observe many constructon frms wth captal stock above $500 prevents us from estmatng the returns to captal at hgher levels for ths ndustry

17 Of course, a non-convexty could occur precsely at the pont of zero captal as well, and thus not be reflected n the margnal returns to captal. That s, earnngs n self employment could be lower than wage earnngs n wage work, even though the margnal return to captal n self employment s hgh. To check for ths possblty, we calculate a predcted wage for each entrepreneur n the sample, gven hs age, educaton, and sector of employment. We fnd no evdence of a self employment earnng penalty. Even entrepreneurs wth very low levels of nvested captal (less than $50) earn more on average than wage earners wth smlar measured characterstcs. 4. Returns and Entrepreneural Ablty The returns estmated above are based on dfferences n captal nvestment across enterprses. In ths secton, we explore the nature of the possble bas, and descrbe the manner n whch we address the ssue. There are two possble sources of bas. The frst comes from the possblty that entrepreneurs operatng n more proftable markets wll nvest more n ther enterprses. The measured ncrease n profts assocated wth an ncrease n captal stock, then, reflects both the true margnal productvty of captal and the correlaton between captal stock and the proftablty of the market. The dffculty n measurng the proftablty of the nvestment opportuntes of a gven frm s a central concern n the lterature on captal constrants and nvestment (see Fazzar, Hubbard and Petersen, 1988). The credt constrant lterature has generally used data from frms whch are orders of magntude larger than the ones n our sample. Large frms mght mantan compettve advantages for long perods of tme, whle our frms are unlkely to do so. Gven the small sze of the frms n our sample, and the absence of barrers to entry, we expect the long run proftablty of frms n our sample to converge to some common level. The opportunty to make proftable nvestments wll be affected by short term shocks whch mght vary across the locatons and ndustres n whch the frms operate. We have controlled for

18 ndustry/locaton specfc shocks to some extent wth the ncluson of the average wage n the state and ndustry n whch the frm operates. 13 The sgnfcance of ths varable ndcates that shocks to wages are reflected n proft rates. A second source of bas, whch may be of greater concern n our data, comes from the possble correlaton between the nvestment level and the unmeasured ablty of the entrepreneur. 14 In the presence of captal constrants, ths bas may produce ether an overestmate or an underestmate of the returns to captal. We start wth the assumpton that hgher ablty entrepreneurs earn hgher profts. If, as s commonly assumed, ablty and physcal captal complement one another, then the margnal returns to captal wll be ncreasng n entrepreneural ablty. Formally, ths s represented by addng ablty to the proft equaton as follows: (1a) π = θ h( K ) + Z ' β + ε where θ measures the entrepreneural ablty of the th ndvdual. We make two further assumptons here. The frst s that the return to captal s decreasng n the level of captal nvestment, as we have found above and as s generally assumed n the entrepreneurshp lterature. The second s that all ndvduals face the same market rate of nterest. Under these assumptons, the optmal captal stock wll be ncreasng n ablty when entrepreneurs equate the cost of captal wth ts margnal product. Now assume all entrepreneurs n our sample have the optmal captal stock gven ther ablty. Then n the absence of measures of ablty, our estmates of returns to captal wll be based upward, as the ncreased proft resultng from an ncrease n captal wll reflect partly the margnal proftablty of captal and partly the concomtant ncrease n entrepreneural ablty. 13 We do not have panel data on the frm sde, preventng the methods of Levnsohn and Petrn (2000) from beng used to measure productvty. Moreover, the sze of our frms and the lack of nventores or electrcty expendture for many of them would prevent straght-forward applcaton of Levnsohn and Petrn even wth a panel. 14 Among the large frms examned n the credt constrants lterature just referenced, the market for manageral talent ensures some degree of homogenety n the ablty of the frms top management. Essentally all of the small frms n our sample, however, are owner-managed. Hence, unmeasured dfferences n the ablty of entrepreneurs are lkely to be more prevalent n our sample

19 Note that f ablty ncreases profts wthout ncreasng margnal returns, then the proper specfcaton of Equaton 1 s: (1b) π = h( K ) + Z ' β + θ + ε In that case, the optmal captal stock wll be uncorrelated wth ablty, and hence not a source of bas n the returns. If the entrepreneurs n the sample are constraned to operate below optmal captal stocks, however, then the ablty bas may run n the opposte drecton. The potental for negatve bas comes from the fact that at very low levels of wealth (and hence, captal avalable for nvestment), entrepreneurs of very hgh ablty are more lkely to fnd entry nto self employment proftable than are entrepreneurs of more modest ablty. As captal constrants are progressvely relaxed, entrepreneurs of more modest ablty wll fnd entry nto self employment proftable. In the presence of severe captal constrants, then, there may be a negatve correlaton between ablty and the level of nvested captal. 15 Evans and Jovanovc (1989) fnd a negatve correlaton between wealth and ablty n a sample of self employed workers n the US. We address these concerns frst by ncludng n the regressons varables whch are lkely to be correlated wth ablty. Followng the lead of Lam and Schoen (1993), we use the educaton levels of relatves by marrage as an ndcaton of the unmeasured ablty of workers. Lam and Schoen motvate the connecton between the unmeasured ablty of a worker and the educaton of hs n-laws wth a marrage matchng model, n whch hgher ablty ndvduals attract more desrable mates. (Educaton s a desrable trat for a mate to have n ther model). Usng Brazlan data, they show that the educaton level of a worker s spouse and parents-n-law are all assocated wth hgher earnngs. For 15 The model of Lloyd-Ells and Bernhardt (2000) provdes an alternatve ratonale for the negatve correlaton between captal stock and ablty. In ther model, talent lowers the fxed cost of startng a busness, so that hgher ablty entrepreneurs requre lower nvestments to start an enterprse

20 the subsample of entrepreneurs wth a spouse present (about two-thrds of the orgnal sample), we are able to nclude controls for the level of educaton of the spouse. Secondly, the ENAMIN survey asks the entrepreneur what hs man motve was for startng hs busness. Indvduals wth a famly tradton n a busness may have learned busness sklls from ther famly. Indvduals who enter busness to acheve hgher ncomes than n the wage sector may also have hgher entrepreneural ablty. In contrast, ndvduals who begn a busness because they can not fnd wage work or have been lad off from ther jobs may have lower ablty. Dummy varables for the dfferent motves for startng an enterprse therefore provde a very crude measure of ablty, but do have the advantage of beng avalable for the whole sample. Fnally, for recent entrants to self-employment we are able to provde a more drect ndcaton of the entrepreneur s unmeasured ablty. As dscussed n Appendx A, the sample to whch the ENAMIN s admnstered s drawn from Mexco s quarterly labor survey (the ENEU). Households reman n the labor survey for a total of 5 quarters. By matchng the ENAMIN sample to the ENEU, we are able to obtan nformaton about the employment and earnngs of the ENAMIN owners n the quarters leadng up to the ENAMIN survey and n the quarters followng the survey. For example, about one-ffth of the ndvduals surveyed n the ENAMIN conducted n the 4 th quarter of 1998 were frst surveyed n the ENEU n the 4 th quarter of Some of these ndvduals were wage workers n the earler quarter, shftng nto self employment durng For these ndvduals, we take the dfference between actual wages earned n the labor market and the wage predcted by a wage regresson as an ndcaton of the ndvdual s productve ablty. Ths s calculated by frst estmatng a wage regresson for all of the full tme male wage workers n the ENEU (the overwhelmng majorty of whch are not ncluded n the ENAMIN sample). The wage regresson s of the form: W = ' X γ + η (5)

21 where X ncludes the average wage n the ndustry and state n whch the ndvdual works, the ndvdual s educaton and ts square, estmated work experence (age mnus educaton mnus sx) and ts square, and the hours worked and ts square. We use the coeffcents from ths regresson to obtan a predcted wage for each of the ndvduals n the ENAMIN sample who were wage workers durng at least one of the quarters pror to the ENAMIN survey. Due to the hgh rates of entry nto self employment n Mexco (see Maloney, 1999), we have wage market data for about 2000 male owners who are ncluded n the ENAMIN surveys. The wage regresson controls for some factors affectng the productvty of workers (e.g., years of schoolng and work experence). As wth the proft functon, however, a more complete wage determnaton equaton s gven by: W = ' X γ + ω + η (6) where ω measures the ablty of the ndvdual whch s not assocated wth learnng n school or on the job. In the absence of explct measures of ω, ablty s embedded n the wage resdual η. 16 We thus take η as correlated wth labor market ablty. If we assume that labor market ablty and entrepreneural ablty are postvely correlated, then η wll also be correlated wth θ. Clearly, η s an mperfect measure of θ, but we expect the two ablty levels to be strongly correlated. 17 Table 3 reports the results of addng the educaton of the entrepreneur s spouse, the entrepreneur s reason for begnnng the busness, and ther own wage premum (for recent entrants), to the returns regressons from Table 2. Model A shows that 16 If ω s correlated wth elements of X, ts excluson wll also result n a bas n the coeffcents of those varables. 17 The decson to enter self employment should be based on ablty n self employment relatve to ablty n wage work. The wage resdual nstead provdes a measure of absolute entrepreneural ablty. Thnk of ablty (θ) as beng made up of three parts: θ = θ g + θ w + θ e, where θ g s a vector of characterstcs measurng general ablty, useful n both wage work and self employment (e.g., nnate ntellgence), θ w s a vector of abltes useful only n wage employment (e.g., the ablty to follow drectons), and θ e a vector of abltes useful only n self employment (e.g. market savvy). Then the decson to enter self employment wll depend on how θ w compares to θ e, but earnngs from self employment wll depend on θ g and θ e, and the wage resdual wll depend on θ g and θ w. As long as θ g s large relatve to the other two components, the wage resdual wll provde a measure of ablty as an entrepreneur as well

22 entrepreneurs wth more hghly educated spouses do have hgher earnngs. Ths s partcularly true f the spouse has college educaton. To the extent that ths reflects entrepreneural ablty, we do fnd that hgher ablty leads to greater earnngs. However, ncludng spousal ablty lowers medan returns for those wth captal stocks n the $0-200 only very modestly, from 16.3 percent to 16.0 percent. Estmated returns over other captal stock ranges are smlarly mpacted. Model B shows that among recent entrants, each dollar of wage premum they receved as a wage worker translates nto 13 cents of hgher earnngs as an entrepreneur. When we nclude both spouse s educaton and the average wage premum (Model C), we fnd only the latter to be sgnfcant whle returns are stll hgh at low captal stock levels. The reason for begnnng a busness also has an effect on earnngs. Model D shows that entrepreneurs who entered the busness to ncrease ncome over wage work or because of famly tradton 18 obtan hgher earnngs as entrepreneurs. As wth model A, the ncluson of these addtonal controls for ablty has very modest mpacts on the estmated returns to captal. Thus, Table 3 suggests that whle our proxes for ablty do affect earnngs, they do not have large effects on returns to captal. These results provde evdence that the hgh return to captal dentfed n the regressons s not smply a conflaton of nvestment level and entrepreneural ablty. The regressons n Table 3 mplement equaton 1b. They allow for ablty to affect the proft level, but not the margnal proftablty of captal. The prmary reason for ths s practcal. Even the relatvely large sample of frms we have s not enough to enable us to fully estmate a sem-parametrc verson of equaton 1a. We can, however, at a more aggregate level, allow for ablty to mpact the rates of return to captal by estmatng the sem-parametrc returns on subsamples of hgh ablty and low ablty entrepreneurs. We frst examne the subsamples of those owners whose spouses have less than 6 years of 18 A t-test could not reject that these two reasons had equal effects on earnngs

23 schoolng and those whose spouses have more than 6 years of schoolng. 19 Among frms wth less than $200 n captal, the returns are slghtly hgher for owners whose spouses have less than 6 years of schoolng (20.5%) than among those wth more educated spouses (18.3%). At slghtly hgher levels of nvested captal ($200-$500), returns are hgher for owners whose spouses have more educaton (14% vs. 9.6%). A smlar pattern holds when we dvde the sample by the owner s own educaton. The medan return to captal n the very low nvestment range (captal < $200) s 20.3% among owners wth lower educaton (less than 6 years), somewhat hgher than the 15.8% return among owners wth more educaton (9 or more years). Returns n the $200-$500 range are hgher among the owners wth more educaton (8.9% vs. 3.2%). The hgh ablty entrepreneurs n the lowest captal range are those most lkely to be captal constraned, and hence subject to negatve bas. In spte of ths, we fnd returns among ths group of entrepreneurs whch exceed 15%. In sum, the returns to captal among the enterprses n our sample do not ndcate the presence of nonconvextes over any broad range of captal nvestment levels. After addng controls for entrepreneural ablty, we fnd that returns to captal are hghest-- around 15% per month-- for nvestments of less than $200. Returns fall to between 7% and 10% for nvestments of $200-$500, and to around 5% for nvestments of $500- $ Evdence on Renvestment Rates: If the rates of return to captal are as hgh as those estmated above, then we should expect entrepreneurs to ncrease nvestment levels as captal becomes avalable. Snce the ENAMIN survey gathers captal nvestment data only once for each enterprse, 19 The sample s lmted to enterprses operatng wth less than $500 n captal, because that s where we fnd the hghest returns and because that s where we the data are dense enough to produce relable estmates

24 we are unable to observe renvestment rates drectly. We do have panel data on ncome and so nstead present ndrect evdence of rapd growth of low-captal frms by calculatng the rate of ncome growth n the quarters followng the ENAMIN survey. Panel data on ncome for ndvdual entrepreneurs s avalable from the quarterly labor survey (ENEU) for as many as 4 quarters followng the ENAMIN survey. About 80% of the ENAMIN sample remans n the ENEU survey for at least one quarter after the ENAMIN. Income data s avalable for these ndvduals for the ENAMIN quarter and the followng quarter. Income data s avalable for each of the 2 quarters followng the ENAMIN for 60% of the sample, for 3 quarters for 40% of the sample and for 4 quarters for 20% of the sample. These data allow us to trace ndvduals, thus controllng for ndvdual abltes whch mght be mperfectly measured n the regressons. The estmated returns to captal are hghest among frms wth nvested captal of less than $200 and lower for those wth more nvested captal. To the extent that hgher rates of return elct more rapd nvestment of captal, we should observe that the ncomes of owners of enterprses wth low levels of nvestment should grow more rapdly than the ncomes of owners of frms wth hgher levels of captal nvestment. Table 4 summarzes the growth rate of ncome of those n the ENAMIN survey, grouped by level of captal nvestment, for those ndvduals who remaned n self employment. 20 Note that snce wage rates declned rapdly n Mexco between 1994 and 1997, ncomes of those n the ENAMIN survey shrnk slghtly n aggregate n the quarters followng the survey. On average, ncomes dd not change among the 4978 owners for whom we have data one quarter out. The next columns of the frst row shows the growth n average ncome 2 quarters out (-1.2%), three quarters out (0.6%) and four quarters out (-3.9%). The average of these values, weghted by the number of 20 The ncome data s very nosy, so the top and bottom 1% of ncomes are dscarded for the ENAMIN quarter and each of the 4 quarters followng before the average ncome levels are calculated. The results are not changed much by ths truncaton

25 observatons, s 0.6%. Owners wth less than $200 nvested fared better. They saw ncomes grow very slghtly by a weghted average of 0.1%. Incomes of those wth between $200 and $1000 nvested shrank slghtly, fallng 0.3%, whle ncomes of those wth more than $1000 nvested fell by 1.1%. Thus, the ncomes of ndvduals wth less than $200 nvested grew 1.2% faster than the ncomes of those wth more than $1000 nvested n the quarters after the ENAMIN. The dfferences are even more stark f we lmt the sample to younger frms. Gven the opportunty to renvest over tme, captal constrants are lkely to bnd more tghtly on younger frms, who are therefore lkely to renvest more actvely. Among frms less than 5 years old (about half of the sample), we fnd ncome growth over the four quarters averagng 1.6% among those wth nvestments of less than $200, and a fall n ncome of 3.3% and 3.2% among those wth nvestments of $200-$1000 and $1000-$5000, respectvely. The results therefore provde evdence consstent wth the hgher returns calculated for enterprses wth very low levels of nvested captal. 5. Explanng the hgh rates of return Our best estmate s that return to captal s n the range of 15% per month for nvestment levels below $200, fallng to 3-5% for nvestment above $500. How do these returns compare wth the cost of captal n Mexco? Snce only 0.3% of the frms n our sample obtaned start-up fnancng from the formal fnancal sector and only 0.8% asked for and receved ther last loan from the formal fnancal sector, the relevant rates are those assocated wth nformal sources of fnance. Mansell Carstens (1995) descrbes a wde varety of less formal sources of fnance n Mexco, ncludng loans from famly and frends, mcrofnance organzatons, pawnshops and professonal moneylenders. Compartamos, the largest mcrofnance company n Mexco, provdes small loans to the

26 self-employed poor at nterest rates of 3% per month. 21 In 2000 ther average loan was $206 whle n 2001 the range of frst loans was $ Other mcrofnance organzatons offer loans at rates of 3-7% per month, but access to these loans, whle ncreasng, has been farly lmted. 23 The natonal pawnshop chan Naconal Monte de Pedad gves loans worth one-ffth to one-thrd of an tem s value at monthly nterest rates whch have ranged from 5.5% n 1994, 6% n 1995, 3.75% n 1997, to 3% n Mansell Carstens (1995) reports smlar rates at the other legal pawnshop, Montepío Luz Savñón, and that the majorty of loans n 1991 at these pawnshops were less than $170. The lmted number of branches and servce hours, coupled wth the need for assets to pawn, lmts ths opton for many poor self-employed. Informal pawnshops and professonal moneylenders (agotstas) offer rates from 10-15% a month up to 15% a week and even 20% per day 25 to those unable to obtan fnance from other sources. The 3-5% monthly returns to captal we fnd for margnal nvestments above $500 accord well wth the rates offered at the legal pawnshops and at mcrofnance organzatons. The sgnfcantly hgher returns we fnd for nvestments below $200 are more surprsng. What explans these returns? Gven the nature of our data, we are only able to speculate about the answer to ths queston. In ths secton, we explore three possble explanatons for the hgh rates of return. The frst s that the hgh returns represent constrants on captal. The second s that they are the result of a process by whch entrepreneurs learn both about ther own abltes and the proftablty of 21 See Mcro-loans help entrepreneurs, Busness, Pg. 1, The Houston Chroncle, August 12, Source: Accon Internatonal 2001 Annual Report, avalable at 23 Source: Conversaton wth Marc de Sousa-Shelds, Enterprsng Solutons Global Consultng, 29 July Sources: Naconal Monte de Pedad, consoldado como nsttucon de asstenca; cobra tasa de 3%, El Naconal, March 13, 1998; Bank of the poor, The Boston Globe, June 15, 1995; Frentes Poltcos, El Excelsor, January 7, 1997; Mexco s poor make do wthout natonal pawnshop, The Ottawa Ctzen, January 15, Sources: Mansell Carstens (1995, pp ); It only takes a small loan to fght Mexco s poverty, The Houston Chroncle, October 13,