WPS5706 Policy Research Working Paper 5706 When is Capital Enough to Get Female Enterprises Growing? Evidence from a Randomized Experiment in Ghana Marcel Fafchamps David McKenzie Simon Quinn Christopher Woodruff The World Bank Development Research Group Finance and Private Sector Development Team June 2011 Policy Research Working Paper 5706 Abstract Standard models of investment predict that credit- the median, suggesting that capital alone is not enough to constrained firms should grow rapidly when given grow subsistence enterprises owned by women. Second, additional capital, and that how this capital is provided for women they strongly reject equality of the cash and should not affect decisions to invest in the business in-kind grants; only in-kind grants lead to growth in or consume the capital. The authors randomly gave business profits. The results for men also suggest a lower cash and in-kind grants to male- and female-owned impact of cash, but differences between cash and in- microenterprises in urban Ghana. Their findings cast kind grants are less robust. The difference in the effects doubt on the ability of capital alone to stimulate the of cash and in-kind grants is associated more with a growth of female microenterprises. First, while the lack of self-control than with external pressure. As a average treatment effects of the in-kind grants are large result, the manner in which funding is provided affects and positive for both males and females, the gain in microenterprise growth. profits is almost zero for women with initial profits below This paper is a product of the Finance and Private Sector Development Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at dmckenzie@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team When is capital enough to get female microenterprises growing? Evidence from a randomized experiment in Ghana Marcel Fafchampsy David McKenziez Simon Quinnx Christopher Woodru¤{ June 2011 Abstract Standard models of investment predict that credit-constrained …rms should grow rapidly when given additional capital, and that how this capital is provided should not a¤ect decisions to invest in the business or consume the capital. We randomly gave cash and in-kind grants to male- and female-owned microenterprises in urban Ghana. Our …ndings cast doubt on the ability of capital alone to stimulate the growth of female microenterprises. First, while the average treatment e¤ects of the in-kind grants are large and positive for both males and females, the gain in pro…ts is almost zero for women with initial pro…ts below the median, suggesting that capital alone is not enough to grow subsistence enterprises owned by women. Second, for women we strongly reject equality of the cash and in-kind grants; only in-kind grants lead to growth in business pro…ts. The results for men also suggest a lower impact of cash, but di¤erences between cash and in-kind grants are less robust. The di¤erence in the e¤ects of cash and in-kind grants is associated more with a lack of self- control than with external pressure. As a result, the manner in which funding is provided a¤ects microenterprise growth. JEL Codes: O12, O16, C93 Keywords: Microenterprises; Ghana; Conditionality; Asset Integration We are grateful to Markus Mobius and conference and seminar participants at Auvergne, Bristol, the IPA Micro…nance conference, Madrid, Michigan, Oxford, Paris, Stern (NYU), Tilburg, Leuven, the World Bank, and Microsoft Research, Cambridge for useful comments. The authors thank Caroline Kouassiaman for outstanding work as project coordinator, and Innovations for Poverty Action for their support on the ground. Financial s support from the World Bank’ Gender Action Plan and Research Budget is gratefully acknowledged. y University of Oxford. Email: marcel.fafchamps@economics.ox.ac.uk z The World Bank, Email: dmckenzie@worldbank.org x University of Oxford. Email: simon.quinn@economics.ox.ac.uk { University of Warwick. Email: c.woodruff@warwick.ac.uk 1 1 Introduction Despite the emphasis placed by micro…nance organizations on lending to female business owners, evidence from three recent randomized controlled trials has cast doubt on the ability of capital alone to grow female-operated microenterprises (de Mel et al 2008, Banerjee et al 2010, and Karlan and Zinman 2010).1 The three experiments were all run in South and Southeast Asian countries: Sri Lanka, India and the Philippines. In Sri Lanka, the capital was provided as grants, while in India and the Philippines, capital was provided by increasing the availability of microloans. In Sri Lanka and the Philippines, the lack of returns in female-owned enterprises contrasted with evidence of positive returns in male-owned enterprises. One possible interpretation is that female-owned microenterprises in these countries are al- ready operating at their e¢ cient level of capital, which just happens to be very low for most …rms. Emran et al. (2007), for instance, argue that many of the women drawn into self-employment have low e¢ cient scale and are only self-employed because of labor market imperfections. Labor market imperfections for women may be particularly strong in societies like those in South Asia, as evidenced by low labor force participation rates among women. This raises the question of whether capital might be more successful in growing female-owned microenterprises in other areas of the world.2 In much of Africa, for example, female labor force participation rates are higher than in Asia, and women are more integral to household income generation. It is therefore possible that the scope for female …rm growth from more capital is higher in Africa. An alternative explanation is that the small scale of many female-owned …rms is in fact not e¢ cient, but instead arises from a lack of separation between household and business decision- making and from ine¢ ciencies in the way people allocate assets between them. One form of ine¢ ciency can arise from self-control problems, leading individuals to not undertake productive investments today that have large payo¤s in the future (Banerjee and Mullainathan, 2010; Du‡o et al, 2010). A second form of ine¢ ciency can come from ine¢ cient intra-household allocation of resources (Udry, 1996; Somville, 2011) or pressure to share with others in ones social network (Charlier, 1999; Platteau, 2000; di Falco and Bulte, 2009). Either of these cases can cause a 1 Bjorvatn et al. (2011) also …nd no e¤ect of capital grants on either male of female small businesses in Tanzania, but their sample is restricted to non-credit constrained entrepreneurs so it is unclear how relevant these …ndings are for microenterpreneurs in general. 2 This is related to the more general issue of external validity, a common refrain in recent debates about what o, the profession is learning from randomized experiments (e.g. Banerjee and Du‡ 2009; Ravallion, 2009; Deaton 2010). 2 lack of asset integration, so that it is not only how much capital, but the form that this capital comes in, which determines the extent to which it helps grow the business. This paper uses a randomized experiment in Ghana to test between competing models of microenterprise investment and growth and thereby better understand the conditions under which capital alone is enough for enterprise growth. The intervention itself is closely modeled on the design used by de Mel et al. (2008, 2009a) in Sri Lanka. A sample of both female and male microenterprise owners who had no paid employees at the time of the baseline survey were randomly allocated into treatment and control groups. The treatment group received grants of 150 Ghanaian cedis (approximately $120 at the time of the baseline). As in Sri Lanka, half the grants were provided in cash and half in kind. A key di¤erence is that the Ghanaian sample contains more than twice as many …rms as in the Sri Lankan study, providing more power to distinguish the e¤ects of providing capital in di¤erent forms. The experiment con…rms some of the …ndings from the Asian experiments, but adds con- siderable nuance to our understanding of the role of access to additional capital in determining the growth of female-owned enterprises. A one-time in-kind grant of 150 cedis is estimated to increase monthly pro…ts by 37-39 cedis for both males and females, a large average return on this grant. However, among females, in-kind grants only lead to pro…t increases for the top 40 percent of businesses in terms of initial size. Women running smaller subsistence businesses, i.e., those earning $1 per day on average, saw no gains from access to additional capital. In contrast the returns for males occur for both smaller and larger …rms.3 As in Sri Lanka, capital alone does not appear to be enough to grow subsistence businesses run by women. We …nd that cash grants of the same size had a signi…cantly smaller e¤ect, increasing pro…ts by only 10-14 cedis on average. When the sample is split between women and men, we …nd that the di¤erent in treatment e¤ect between in-kind and cash grants is signi…cant among women but not among men. In some speci…cations, but not all, men show signi…cant increases in pro…ts following the cash grants. This result is not signi…cant when we condition on our baseline data, however. We …nd that the cash grants tend to be spent on household consumption or transferred 3 The high marginal returns to the capital shocks for males are consistent with non-experimental work in Ghana which has found evidence of high returns to capital for male-owned informal enterprises. Bigsten et al. (2000) …nd much higher returns to physical capital than human capital in African small and medium scale manufacturing …rms, Udry and Anagol (2006) …nd returns to be at least 60 percent per year among purchasers of used auto parts in Accra, and Schündeln (2006) …nds strong evidence of …nancing constraints among small Ghanaian …rms using a structural modeling approach. 3 out of the house, especially when given to women whose businesses were small to begin with. The experiment allows us to test between three competing models of microenterprise invest- ment and growth. The results are not consistent with either a standard Ramsey model or with a variation of this model that incorporates time-inconsistent preferences. To explain a large di¤erence in outcomes between cash and in-kind grants, we need a model with a lack of asset integration where the form in which capital comes a¤ects the extent to which it is invested in the business. We examine two possible causes of the di¤erence between in-kind and cash grants: self-control issues caused by time-inconsistent preferences, high discount rates, or lack of ability to save; and external pressure from others to share additional resources. We …nd that the e¤ect of the cash treatment is signi…cantly more positive for individuals with the most self-control, whereas there is no evidence of treatment heterogeneity with respect to external pressure. This is consistent with the recent evidence in Spears (2009) who suggests that present-bias is a key constraint on microentrepreneurs expanding their businesses. But it contradicts results from Anderson and Baland (2002) for Kenya and Somville (2011) for Benin, whose evidence suggests that women seek to save outside the household in order to avoid contributing to household ex- penses, and …ndings by Brune et al. (2011) in Malawi, who argue that the reason for the success of a saving commitment product is the desire to escape external pressure. The remainder of the paper is structured as follows. In Section 2 we present the conceptual framework and testing strategy. Section 3 describes the experimental design and characteristics of our sample. Section 4 gives the basic experimental results, and explores heterogeneity by gender, treatment type, and randomization strata. Section 5 then asks what happens to the cash grants and what distinguishes the pro…table from less pro…table female businesses. Section 6 examines why the cash and in-kind treatments di¤er, and Section 7 concludes. 2 Conceptual framework and testing strategy In this section we present a conceptual framework that guides our empirical analysis. The model is organized around two key assumptions which characterize the population of microenterprises we study: entrepreneurs cannot borrow and have to self …nance (only 10 percent of our sample has ever had a formal loan); and they have di¤erent abilities and ability is a complement to capital. We …rst present a model of capital accumulation without time inconsistency and derive testable predictions regarding the e¤ect of a capital grant. We then introduce time consistency 4 and self-control issues and discuss how experimental predictions are a¤ected. Next we discuss asset integration and intra-household issues. The testing strategy is presented at the end. 2.1 The Ramsey model Consider an entrepreneur facing a standard accumulation problem of the form: 1 X t max u(ct ) subject to ct >0;kt 0;wt 0 t=0 ct = (kt ; ) + rwt (kt+1 kt ) (wt+1 wt ) (1) where k is capital invested in a business with total return to capital (k; ), variable is individ- ual speci…c talent, is the discount factor, and w is a …nancial asset with return r.4 We assume @ =@k 0 (positive or zero returns to capital) but @ 2 =@k 2 < 0 (decreasing returns to scale). Decreasing returns to scale may be due to the presence of …xed factors, such as entrepreneur time and family labor. We also assume that @ 2 =@k@ > 0: more talented entrepreneurs have higher marginal returns to capital.5 There are two possible treatments: a cash transfer Mt and an in-kind transfer Et at an arbitrary time t. Both can be turned into more capital k but it takes time to liquidate grant Et that comes in the form of equipment or inventories. In contrast, Mt is liquid and perfectly fungible with k or w or c. We derive model predictions about @k=@M and @k=@E. We …rst note that, by asset arbitrage, wt = 0 if 0 (k; ) > r. In this case, the …rst order t conditions are as follows: t 0 ut = t 0 t (1 + t) = t¬1 where 0 denotes the marginal return to capital and is the Lagrange multiplier associated t with the constraint. From the above we get a standard Euler equation of the form: 0 1 u0 t¬1 1+ t (k; )= u0t 4 Variable (k; ) measures value added, that is, return to capital and family labor net of intermediate input costs and other recurrent costs. Given the nature of the studied …rms, this corresponds to an accounting notion of pro…t, but not to an economic notion of pro…t/return to capital since we have not imputed the cost of the s entrepreneur’ labor. 5 It is conceivable that a minimum level of capital is needed to initiate a business. Since all households in our sample by construction have a business, we ignore this here. 5 If we ignore savings wt , there exists a steady state level of capital k such that pro…t and consumption c are constant and: 0 (k ; ) = 1¬ where . The proof follows from the fact that, without savings wt , the above is a standard Ramsey model. Given that @ 2 =@k 2 < 0 it follows that dk =d > 0 –more patient entrepreneurs have larger k . If r > , the entrepreneur stops investing in the …rm once the marginal return to capital falls below r, and invests in w instead. The optimal …rm size is then given by: 0 (k ; ) = r with k < k . Given our assumption that, @ 2 =@k@ > 0 comparative statics imply that both dk =d > 0 and dk =d > 0 –more talented entrepreneurs have larger steady state capital and …rm size. Only patient agents — that is, those with < r — ever hold non-zero savings, wt > 0. If kt < minfk ; k g, the cash and in-kind treatments are predicted to increase capital and pro…ts by the same amount.6 Their long term e¤ect is to shorten the time necessary to reach the steady state …rm size. In contrast, when a entrepreneur has reached k or k , the e¤ect of the two treatments is di¤erent. If k = k , a cash transfer has no e¤ect on capital and @kt+s =@Mt = 0 for any s 0; it raises consumption c and savings w instead. In this case we should observe no cash treatment e¤ect on pro…ts t+s (k; ): the cash treatment Mt should not be invested in …rms that have already reached their optimal size; it should be saved instead. If the in-kind treatment Et cannot be liquidated immediately, however, we expect a temporary positive e¤ect on pro…t: (k + E; ) > (k; ) since, by assumption, @ =@k 0. But this e¤ect should be short-lived: the …rm should return to its steady state capital level as soon as E can be divested. If k = k with > r, then instead of saving in asset w in order to smooth consumption of the capital grant, it is optimal for the entrepreneur to use a temporary investment in the …rm as bu¤er to smooth consumption. In this case, Mt and Et have a similar short-run e¤ect on capital and pro…ts. In all cases the model predicts that the cash and in-kind treatments will result in higher consumption. In the steady state case with > r, the household is impatient and the treatment will be consumed rapidly before consumption returns to its steady state level. In the case where 6 In the interest of space, we do not discuss the case where kt + M > minfk ; k g > kt . This case is e¤ectively a weighted average of the two cases we describe. 6 r > , there will be more smoothing, that is, part of the treatment will be saved and consumed later. In the case where kt is below its steady state, we expect an increase in consumption out of higher pro…ts. 2.2 Time-inconsistent preferences We now introduce quasi-hyperbolic preferences as in Laibson (1997). At time t the household sets kt so as to solve: 1 X s max u(ct ) + u(cs ) subject to (1) (2) fcs ;ws ;ks g s=t+1 where < 1. But once at time t + 1, the household sets kt+1 according to: 1 X s max u(ct+1 ) + u(cs ) subject to (1): (3) fcs+1 ;ws+1 ;ks+1 g s=t+2 This means that at time t + 1 the household wants to revisit decisions taken at time t and set paths for fct+1 ; ct+2 ; :::wt+1 ; wt+2 ; :::; kt+1 ; kt+2 ; :::g that di¤er from those set in period t. In Appendix 1 we show that the entrepreneur stops investing after reaching a steady state level of capital k s (for a sophisticate) or k m (for a myopic decision maker) which are, in general, smaller than k . Model predictions regarding the e¤ect of a capital grant are similar to the Ramsey model. If the …rm has already reached its steady state k s or k m , the cash transfer M will be rapidly consumed while the in-kind grant E will be divested as quickly as is feasible. If kt < k s or k m , then the additional cash M or inventories E will remain in the business and increase future pro…ts. To summarize, the standard and time inconsistent models both predict that the long-term e¤ect of the cash and in-kind transfers on capital and pro…t are nil for …rms that have already reached their steady state capital level. The short-term e¤ect of the cash transfer on capital and pro…t is also nil. For the in-kind treatment there is a short-term increase in capital and pro…t until the household is able to divest, which is expected to happen as soon as is feasible. In contrast, for …rms that are below steady state, both cash and in-kind transfers are predicted to be entirely invested and the e¤ect of the grant is to reduce the time taken to reach the optimal …rm size. 2.3 Asset non-integration and family pressure In the experiment most in-kind grants are used to purchase inventories and raw materials rather than machinery (and …rm owners could chose which of these it was). It should therefore be 7 relatively easy to de-capitalize these grants (by selling and not replacing stock) and take them out of the business. This di¤ers from the conditionality on school attendance or vaccination in traditional conditional cash transfers, which are not reversible. We should therefore think of the di¤erence between cash and in-kind treatments as earmarking the grant for a speci…c purpose and reducing its liquidity. Given that inventory turnover is quite rapid in the kind of enterprises covered by this study – e.g., one week – the reduction in liquidity is minimal. Based on the models discussed so far, we should therefore expect little di¤erence between cash and in-kind treatments. Until now we have assumed that people make decisions regarding asset accumulation in an integrated manner, i.e., that consumption c, pro…ts , capital k and savings w are regarded as fungible. Yet experimental evidence suggests that asset integration often fails. For instance, it is common for experimental participants to exhibit considerable risk aversion even though the stakes are very small stakes relative to their wealth (Harrison, Lau and Rutstrom 2007; Andersen, Harrison, Lau and Rutstrom 2008). Similarly, Camerer, Babcock, Loewenstein and Thaler (1997) …nd that cab drivers make labor supply decisions based on single-day earnings. In other words, they fail to integrate earnings over a longer time period of a week or a month when making enterprise decisions. Self-control issues arising from dynamic inconsistencies in preferences are one reason people may not undertake productive activities today that have large payo¤s in the future. For example, o Du‡ et al. (2010) …nd farmers in Kenya fail to undertake pro…table investments in fertilizer due to present-bias, but that o¤ering small time-limited discounts can induce them to do so. Banerjee and Mullainathan (2010) show that these time-inconsistency issues can be particularly important for the poor. Given this, in-kind grants may then act as a “nudge”to get …rm owners to invest in their business, and once the money is in inventories and equipment, limited illiquidity may help the …rm owner avoid impulse purchases. Asset integration may also fail for reasons external to the individual, such as disagreement over the allocation of resources between household or family members. If intra-household bar- gaining is e¢ cient, asset integration should hold. But if binding commitment is not possible, for instance because of lack of trust, intra-household allocation of resources can be ine¢ cient. Udry (1996), for instance, shows that organic fertilizer is not allocated e¢ ciently between male and female …elds in Burkina Faso. Anderson and Baland (2002) similarly show that women in ur- ban Kenya join rotating savings and credit associations (ROSCAs) to shelter money away from 8 their spouse. A similar result is reported by Somville (2011) for Benin. de Mel et al. (2009a) suggest women may ine¢ ciently over-invest in less liquid forms of business assets in order to resist spousal pressure. Pressure to redistribute resources can also be exerted from outside the household. Platteau (2000) introduces the idea of sharing norms to economics from anthropology. He notes that in many developing countries, especially in sub-Saharan Africa, individuals often live in large households and have strong links to extended family and kinship networks. Social sharing norms can make it hard for individuals to save and invest, as they are forced to share additional resources with others. These sharing norms can vary according to the source of income and how o it is stored. For example, Du‡ and Udry (2004) …nd evidence that the proceeds of di¤erent crops are used for di¤erent purposes in Côte d’Ivoire, and note that income from some crops is expected to be shared within the household and income from others is not. Charlier (1999), Ivoire, notes that as a result of sharing norms, individuals may develop based on work in Côte d’ an illiquidity preference in order to be able to resist social claims without appearing sel…sh. Suggestive evidence supporting this view comes from di Falco and Bulte (2009), who show in South Africa that households with more kinship links spend less of their income on liquid and sharable assets, and from Baland, Guirkinger and Mali (2007), who …nd individuals in Cameroon taking loans even though they have high savings balances, which their interviews reveal to be a way of resisting demands for …nancial assistance by others. Jakiela and Ozier (2011) …nd in a lab experiment in rural Kenya that women invest less when the income they earn is observable to relatives, even when this reduces their expected total earnings. However, Grimm et al. (2010) o¤er a more mixed picture, …nding in seven West-African countries that local social networks within the city actually have a positive association with business performance, whereas there is a negative association between business performance and a smaller distance to the village of origin. , In our context the existence of a “social solidarity tax” either from other household members or from extended family members, may lead to less of the cash grant being invested in the business than is the case with the in-kind grant. This could arise either due to the di¤erence in liquidity (it takes some time to decapitalize inventories and raw materials and this time is su¢ cient to resist pressure for on-the-spot transfers) or to the di¤erence in form and function (there could be an expectation to share cash coming into the household, but not to share the value of additional materials going into the business). 9 To capture these ideas, let us rewrite the law of motion of entrepreneurial capital as: kt+1 = kt + (kt ; ) ht (4) where ht ct + wt+1 (1 + r)wt represents what is taken out of the enterprise either to be consumed or to be invested in other assets. In the Ramsey and time inconsistent models – hereafter RTI models –the optimal choices of consumption ct and savings wt+1 depend on total cash-in-hand kt + (kt ; ) + (1 + r)wt . Unless kt is illiquid, increasing kt or t has the same e¤ect on cash-in-hand and thus on ht , kt+1 and t+1 . In the more general case, ht = h( t ; kt ) and asset integration requires that h( t ; kt ) = h( t + kt ). If households regard kt and t as not fungible, they are imperfect substitutes in h(kt ; t) and h( t ; kt ) 6= h( t + kt ). This simple observation forms the basis for our testing strategy. As discussed earlier, asset integration may fail because assets kt are less susceptible to internal pressure than pro…ts t. Turning working capital into inventories or equipment may serve as self-commitment device against the temptation of impulse purchases. This is akin to consumers putting money on a low-yield savings account that is less conveniently accessible. In the same vein, Fafchamps, Udry, and Czukas (1998) present evidence suggesting that, in times of duress, farming households prefer to reduce consumption than sell animals because the latter would translate into lower income in the future. The other possibility is that pressure from household members and other relatives works as @h @h a tax on the business with @ > @k 0. Money tied up in inventories or equipment is less liquid and thus partly insulated from external pressure. If successful, this tactic would yield a @h @h marginal tax rate on cash ‡ow @ that is higher than the marginal tax on capital @k . If asset non-integration signals an e¤ort to escape taxation of this kind, it is more likely to be observed among enterprises operated by more subordinate household members, such as married women. When asset integration fails, cash and in-kind treatments can have systematically di¤erent s e¤ects: the in-kind transfer may be treated as adding to the …rm’ capital, while the cash s treatment is regarded as part of the …rm’ cash ‡ow, or as never having entered the …rm in the @h @h …rst place. To illustrate, consider the simple case where @k = 0 but @ > 0. A steady state …rm 10 size k v is de…ned as a capital stock that satis…es:7 (k v ; ) = h( (k v ; )): To …x ideas, suppose that ht = a t + b with 0 < a < 1. The law of motion of capital becomes: kt+1 = kt + (1 a) (kt ; ) b (5) which resembles a Solow model with a negative drift term b. Provided that the marginal return to capital is high enough at low values of k,8 di¤erence equation (5) has two equilibria: a high, v stable equilibrium khigh similar to the steady state of a Solow model; and an low, unstable v v v equilibrium klow below which the …rm closes down. For k such that klow < kt < khigh , the …rm v is growing. For k < klow , the …rm is unstable and eventually disappears – and is thus unlikely to be part of our sample. We now introduce cash and in-kind grants. Equation (5) is rewritten: kt+1 = kt + Et + (1 a)( (kt ; ) + Mt ) b v v which implies that for initial values of k such that klow < kt < khigh , the in-kind treatment Et has a one-for-one e¤ect on capital stock kt+1 but the cash treatment only has a 1 a e¤ect on kt+1 : d kt d kt =1>1 a= dEt dMt where the notation kt denotes kt+1 kt . In other words the cash treatment is predicted to have a lower e¤ect on future capital –and hence pro…ts –than the in-kind treatment as long as @h a > 0, that is, as long as @ > 0. v Turning to long-term predictions, if the …rm was below its equilibrium size khigh , the in-kind v treatment speeds up convergence to the steady state khigh . Future additional pro…ts generated by kt + Et are subject to taxation and raise future consumption. If the …rm was at –or above v –equilibrium size khigh , then decreasing returns in capital imply (kt ; ) ht < 0 and the …rm should slowly decapitalize the in-kind treatment Et . In the special case where h( ) = b and v initial capital kt < klow but kt + Et + (1 a) (kt ; ) b > k v , the in-kind treatment pushes 7 For some functions h( ), the steady state is not stable. For instance, if h( ) = b with b a positive constant, the steady state kv is given by (kv ; ) = b, but the …rm eventually closes down if kt < kv while it expands forever for kt > kv . In contrast, if h( ) = , the law of motion of capital becomes kt+1 = kt and any capital level k is an equilibrium. 8 A standard Inada condition. 11 the …rm above the minimal threshold size and ensures its long term survival.9 In the special case where h( ) = , the in-kind treatment pushes the …rm to a new equilibrium level of capital kt + Et : future pro…ts increase but there is no further addition or subtraction to capital after t + 1. The above example can be generalized to allow ht to depend on both t and kt . For instance, let ht = a t+ kt + b with and 0 < < 1. The no-closure stable steady state k w is the (highest) value of k that solves:10 (1 a) (k w ; ) b = kw : It follows that equilibrium …rm size is a decreasing function of both a and . The in-kind treatment has a 1 e¤ect on kt+1 while the cash treatment has a 1 a, also less-than-one- for-one, e¤ect on kt+1 . Asset integration requires that a = . If investing in inventories and equipment is successful as protecting the capital of the enterprise, we should observe a > . This forms the basis of our testing strategy. 2.4 Testing strategy We estimate models of the form: i;t+s = 1 Mit + 2 Eit + ui;t+s (6) ki;t+s = 1 Mit + 2 Eit + vi;t+s (7) where t is the time of treatment, i;t+s is the pro…t of entrepreneur i at time t+s after treatment, s s ki;t+s is the capital stock, and ui;t+s and vi;t+s are error terms. Coe¢ cients ’ and ’ are the average e¤ects of each of the two treatments on capital stock and pro…ts, respectively, across the population of …rms in our sample. The RTI models predict 1 = 2 > 0 and 1 = 2 > 0 if the …rm was below its steady state at the time of the treatment. They also predict 1 = 1 = 0 if the …rm had already reached its equilibrium size at time t such that kt = k ; k m , or k s . Because the in-kind treatment is not immediately fungible, these models also predict 2 > 0 and 2 > 0 for a small time from 9 In this case, the treatment eliminates a poverty trap. 10 If b = 0, this is a Solow model in disguise. If we set (k; ) = k e , the steady state is the usual: 1 1 a 1 kw = e where 1 a is the savings rate and plays the role of depreciation. 12 treatment s, but eventually 2 = 2 = 0 for s large enough, as k returns to its steady state from above. A similar result obtains if kt = k and …rm capital is used as bu¤er to smooth consumption. In contrast, the model without asset integration makes predictions that do not in general v depend on whether the …rm is above or below steady state khigh . Predictions however depends on the form taken by the external pressure function h(:). If ht is a constant lumpsum b with a= = 0, then both treatments Mt and Et increase capital one for one, that is, 1 = 2 = 1 at time s = t + 1, that is, one period after treatment. If, in contrast, ht = t and Mt is regarded s as part of the …rm’ cash ‡ow t but Et is not, then 1 = 0 and 2 = 1 at all s t + 1. For the intermediate case where ht = a t + b with a < 1, the model predicts that Et has a one-for-one e¤ect on capital stock kt+1 , that is, that 2 = 1 but Mt only has a 1 a e¤ect on kt+1 , i.e., 0 < 1 =1 a < 1. The larger a is, the closer 1 is to 0. Finally, when ht = a t + kt + b, that is, when external pressure also puts a tax on capital, then 0 < 2 =1 < 1 while we still have 0 < 1 =1 a < 1. Asset integration requires that 1 = 2 and hence that 1 = 2. We have discussed two main reasons why household asset integration may fail: internal pressure driven by self-commitment problems; and external pressure from household and family members. If external pressure comes primarily from husbands, unmarried women should show a lower a and and thus a stronger response to treatment. If pressure comes from children or the extended family, a stronger response to treatment will be observed for entrepreneurs without children or with a smaller extended family. To implement this idea, let a = a0 + a1 z and = 0 + 1z with z a vector of proxies for di¤erent kinds of external pressure. We estimate a model of the form: i;t+s = (1 a)Mit + (1 )Eit + vi;t+s = (1 a0 )Mit a1 Mit zi + (1 0 )Eit 1 Eit zi + vi;t+s (8) If a speci…c element of zi is associated with a higher implicit tax rate on cash ‡ow, then the coe¢ cient of Mit zi should be negative and signi…cant. Similarly, if it is associated with higher taxation of capital, the coe¢ cient of Eit zi should be negative and signi…cant. We test internal pressure using a similar approach. In this case, pressure comes from the non-business minded self, that is, the self susceptible to immediate grati…cation. In this context, keeping excess liquidity in less fungible inventories and equipment can be seen as a way to insulate working capital from temptation. If this strategy is successful, we should observe that 13 1 < a1 for individuals with more self-commitment problems proxied by zi . 3 The Experiment 3.1 The Sample We purposively chose urban Ghana as the setting for this study. The choice of Ghana was motivated by the desire to provide evidence in an African context, in a country known for a history of involvement of women in business which provides a setting that is conducive to female business success. Women in Ghana have similar labor force participation rates to men, and are more likely to be self-employed. Evidence of this is seen in data from the 2000 Ghanaian Census: the labor-force participation rates for 15-60 year olds are 69.6 percent for females and 73.9 percent for males, and in urban areas 45 percent of females are non-agricultural own-account workers, compared to 33 percent of males. This contrasts sharply with Sri Lanka, the setting for the experiment in de Mel et al. (2009a), where only 7.8 percent of prime age females are self-employed, compared to 29.7 percent of prime age males. Within Ghana we chose Accra, the capital and largest city, and the nearby industrial city of Tema. A sample of microenterprises (which we term the Ghana Microenterprise Survey) was then constructed as follows. First, enumeration areas (EAs) were selected with probability proportional to the number of households in these EAs according to the 2000 census. We randomly selected 70 EAs in Accra and 30 in Tema. Then, to reduce the costs of listing, we subdivided EAs into equal areas, such that each area would contain approximately 70 to 80 households. This typically required dividing an EA into half or thirds. One of these areas was then randomly selected from each EA. Enumerators went door to door in this area to carry out a screening survey of each household. Households were screened to identify those with an individual age 20 to 55 who was self-employed and working 30 or more hours per week in a business with no paid employees and no motorized vehicle. These criteria were used to select full-time microenterprise owners who were not so large that the grants in our experiment would have little e¤ect. The gender and business sector of all individuals passing this screen were then recorded. This resulted in screening 7,567 households to identify 3,907 individuals who passed the screen. Only 19.4 percent of these individuals were male, con…rming the predominance of women among small enterprise owners in urban Ghana. Based on the gender mix of self-employed in these 14 industries in the 2000 Census, we classi…ed business sectors into male-dominated industries, identi…ed as construction, repair services, manufacturing, and shoe making and repair; female- dominated industries, identi…ed as hair and beauty care, and food and restaurant sales; and mixed industries, identi…ed as trade and retail, and sewing and tailoring. These industries cover the vast majority of the industries in which the self-employed work in Ghana. The 4.6 percent of those screened who worked in other industries such as communication services, pharmacy, photography, …shing, and agriculture were not included in the sample. Our aim was then to arrive at a sample of roughly 900 baseline …rms strati…ed by gender and sector. In order to minimize the spillovers from the treatments to be carried out, we limited the sample from each EA to no more than 5 males in male-dominated and 5 males in mixed industries, and no more than 3 females in female-dominated and 3 females in mixed industries. We also ensured that only one individual was chosen from any given household. This resulted in an initial sample of 907 …rms, consisting of 538 females and 369 males. A baseline survey of these …rms was conducted in October and November 2008. The …rm owners were asked for details of both their …rm and their household. A second pre-treatment survey of these …rms was conducted in February 2009. The purpose of a second pre-treatment round was to eliminate …rms most likely to attrit. In particular, 55 of the initial 907 …rms could not be found on at least three attempts, 15 …rm owners refused this second round, 24 …rm owners were no longer operating a business, and 20 …rms that did not provide details on their …rm pro…ts, expenses and sales were eliminated. This left a …nal sample for the experiment of 793 …rms, comprising 479 females (248 in female-dominated industries and 231 in mixed industries) and 314 males (146 in male-dominated industries and 168 in mixed industries). 3.2 Experimental design The design of the experiment closely followed that used in Sri Lanka by de Mel et al. (2008, 2009a). Firms which completed the …rst two survey rounds were randomly allocated into three groups: a control group of 396 …rms, a treatment group of 198 …rms which would receive 150 Ghanaian cedis (approximately US$120 at the time of the baseline) in cash which they could use for any purpose, and a treatment group of 198 …rms which would receive 150 cedis in equipment, materials, or inventories for their business. In the case of the in-kind treatment, the equipment or materials were selected by the …rm owner and purchased directly by our research assistants 15 with the owner. The majority of this was in the form of inventories to sell (e.g. beauty care products, electronic goods, alcohol, food) and raw materials (e.g. wood, sandpaper, cloth, oil and other cooking ingredients, shampoos and supplies for beauty salon use). Only 24 percent of those receiving the in-kind treatment elected to buy physical equipment, with the most common equipment purchased being sewing and knitting machines by tailors, hair dryers by owners of beauty salons, and drills and other carpentry equipment by …rms in woodwork. Males were more likely to get some equipment with this treatment than females (33 percent versus 19 percent). With the cash treatments, …rm owners were noti…ed by phone, or in-person, and then received the cash through money transfer at a local bank or in-person. We also randomly selected when …rms would receive their grant, staggering the timing of the grants, so that 198 …rms were assigned to receive the grants after the second round, a further 181 …rms assigned to receive the grants after the third round, and 18 …rms were assigned to receive the grants after the fourth round. This staggering was done both for the purpose of managing the logistics of making these grants, and to provide incentives for …rms to remain in the study for multiple rounds since they were told more grants would be given out after rounds 3 and 4. These grants were framed to …rms as prizes to thank …rms for participating in the survey. Participants in the survey were told that we were undertaking a study of small …rms in Ghana, and that some of the …rms would be randomly chosen to receive prizes as a token of our appreciation for their participation in the survey. Firms which were selected in either treatment group were not told they had been selected for a prize until the time their prize was being given out. Randomization was done via computer after the second round of data was collected. Firms were …rst strati…ed into 16 strata on the basis of gender and sector (males in male dominated industries, males in mixed industries, females in female-dominated industries, and females in mixed industries); baseline capital stock (above or below the raw baseline median of 181 cedis in . capital stock); and on a binary variable called “high capture” In the second survey round, …rm owners were asked on a 5 point Likert scale (ranging from 1 = strongly disagree to 5 = strongly agree) to assess how strongly they agreed or disagreed with the statements “Whenever I have , money on hand, my spouse or other family members always end up requesting some of it” and “People who do well in their business here are likely to receive additional requests from family . and friends for money to help out with some expense or another” We summed the responses 16 to these two questions, and classi…ed as “high capture” …rm owners with scores of the median of 8 or above –that is if on average they agree with both statements. Then within each strata, we ranked …rms according to January 2009 reported pro…ts (col- lected in the second round survey), and formed matched quadruplets of …rms. We used wave 2 rather than baseline pro…ts for the match since 9 percent of the …rms did not report round 1 pro…ts. Within the quadruplet one …rm was then randomly chosen to receive the cash treatment, one to receive the in-kind treatment, and two to be control …rms. We then randomly selected which quadruplets would receive their treatments after each round. In the end this resulted in the 793 …rms being matched into 195 groups, of which 4 groups ranged in size from 5 to 8 …rms and the remainder were quadruplets. This randomization design was based on the analysis in Bruhn and McKenzie (2009) who showed the potential for signi…cant increases in power and baseline balance from matched pairs (with a single treatment group) and strati…cation compared to simple randomization. The variables used for strati…cation were motivated by the results in de Mel et al. (2009a). In particular, we strati…ed by gender and industry since the ex post heterogeneity analysis in that paper found strong di¤erences by gender, and some suggestion of di¤erences according to whether women were working in female-dominated versus mixed industries. The choice of “high capture” as a stratifying variable is motivated by the literature referenced earlier that has suggested that many individuals who succeed in raising their incomes face large demands to share it from others. In addition, there was some evidence in Sri Lanka that a reason for the low returns to women is pressure from household members to redistribute resources. Strati…cation on baseline capital stock was done both because this was believed to be a variable which would be correlated with future pro…ts, and to allow for testing potential heterogeneity in treatment e¤ects for smaller and larger microenterprises. Matching of quadruplets on pro…ts was done to achieve greater balance on the pre-treatment value of the main outcome of interest as well as to investigate treatment heterogeneity in this dimension. It also enables us to eliminate quadruplets with outlier values of pre-treatment pro…ts and still be assured of balance and random allocation to treatments and control among the remaining sample. 3.3 Data collection and description of …rms The two pre-treatment survey rounds were followed up by four additional survey waves in May 2009, August 2009, November 2009, and February 2010. Of the 793 …rms which completed the 17 …rst two rounds, 730 answered the …nal wave survey. Appendix 1 details wave by wave attrition rates and shows the robustness of our main treatment e¤ects to corrections for attrition. Each follow-up round collected data on changes over the quarter in …xed capital from pur- chases, sales or repair; the current value of inventories and raw materials, and the value of the s last month’ expenses, sales, and pro…ts. The most important …rm outcome variable measured is …rm pro…ts. Pro…ts were elicited via a direct question, following the recommendations of de Mel et al. (2009b). Firm owners were asked: “After paying all expenses, what was the income of the business (the pro…ts) during January 2009? (Consider all expenses, including wages of . employees but not including any income you paid yourself as an expense)” This de…nition s of pro…t thus includes the return to the entrepreneur’ labor and managerial talent. Nominal pro…ts were converted to October 2008 real pro…ts using the Greater Accra region Consumer Price Index collected by the Ghana Statistical Service. An innovation in this experiment was the use of computerized cross-sectional and panel consistency checks. Data was collected using PDAs, and a consistency check was triggered whenever reported pro…ts exceeded reported sales in the cross-section, whenever a …rm reported sales but not pro…ts, and whenever the change in pro…ts from one quarter to the next was less than -33.3 percent or greater than +50 percent (provided the absolute change in pro…ts was at least 20 cedis). We discuss these consistency checks in more detail in Fafchamps et al. (2010), where we show that they lead to some improvements in data quality. We therefore use the pro…ts which incorporate the consistency checks in this paper. Nonetheless, our results are similar when we use the raw pro…t data. Table 1 summarizes the basic characteristics of …rms and their owners in our experimental sample, and compares the pre-treatment characteristics of …rms in the control group to those assigned to either treatment group. The top of the table shows balance for the characteristics used for strati…cation or matching, while the remaining rows compare the characteristics of other variables of interest. Mean (median) monthly pro…ts in January 2009 were 130 (68) cedis, and mean (median) capital stock at the same point in time was 452 (172) cedis. The grants of 150 cedis were therefore approximately equivalent to two months’ pro…ts and almost equal to the size of existing capital stock for the median …rm. However, since we did not explicitly cap pro…ts or capital stock when selecting …rms into the experimental sample, there are a small number of …rms with much higher levels –the maximum pro…t reported in our pre-treatment waves is over 5000 cedis per month. The inclusion of these few larger …rms does not have much e¤ect our 18 basic results, but has a larger e¤ect on our analysis of treatment heterogeneity. As discussed below, we therefore focus most of our analysis involving heterogeneity of treatment response on the …rms in quadruplets which have baseline pro…ts of 1500 cedis per month or less. Since randomization occurred within quadruplets, balance on baseline characteristics is achieved for this subsample also. Table 1 shows that overall the two treatment groups look similar to the control group in terms of pre-treatment characteristics. The exceptions are October/November 2008 pro…ts and January 2009 sales, which show signi…cant di¤erences across treatment groups in the trimmed sample, and di¤erences in magnitude, if not statistical signi…cance, in the full sample. Recall the matched randomization used the wave 2 pro…ts. However, the correlation between wave 1 and wave 2 pro…ts is only 0.19, compared to a correlation of 0.58 between wave 2 and wave 3 pro…ts, and of 0.72 for the control group between waves 5 and 6 (which is the same seasonality as between waves 1 and 2). This di¤erence in baseline pro…ts is due to pure chance, and is in a variable which the data suggests involves considerable noise and perhaps learning on the part of survey respondents as well. Imbalance on this baseline pro…t measure is thus unlikely to imply imbalance on follow-up pro…ts, particularly given the pre-treatment balance on wave 2 pro…ts (Bruhn and McKenzie, 2009). Nevertheless, we will show our results are largely robust to the use of …rm …xed e¤ects which account for any baseline imbalances, the main di¤erence being in the cash-treatment for males, which we discuss in detail later. As seen in Table 1, the mean owner in our sample is 36 years old, has almost 9 years of schooling, and has been running the …rm for 7 years. The mean number of digits recalled in a Digitspan recall test is 5.1, which is almost one digit lower than the 5.9 average among Sri Lankan microenterprise owners (de Mel et al, 2008). The majority of …rms are run out of the home, with 83 percent of women and 69 percent of men operating a business from their dwelling. Most …rms are informal, with only 14 percent registered for taxes, and only 10 percent have ever had a loan from a bank or micro…nance institution. Half of the …rm owners use a susu collector, with this more common among women (58 percent) than men (34 percent). A susu collector is an informal mobile banker, who typically collects a savings deposit daily from individuals and s returns them at the end of the month after subtracting one day’ deposit as a fee. That is, saving is at negative interest rates in exchange for safekeeping. Besley (1995, p. 2150) states that “a frequently heard rationale for the existence of this institution is that there are di¢ culties for those who have a stock of liquid assets in resisting the claims of their friends and relatives 19 (or even spouses)”. 4 Testing the predictive power of the models Before investigating the e¤ect of the randomized capital grants on sample …rms, we brie‡y examine the baseline data for evidence consistent with predictions made by the models presented in Section 2. The evidence we o¤er is non-experimental, but since we are not trying to draw causal inference, observational data remains useful. The Ramsey model predicts, other things being equal, that more patient entrepreneurs have a higher steady state capital stock. More talented entrepreneurs should also have larger …rms and thus larger capital stock. The time inconsistency model further predicts that entrepreneurs that are more hyperbolic in their discounting should also have a lower steady state capital stock. We investigate these predictions using the baseline capital stock of all respondents. The results are shown on Table 2. Consistent with the predictions of the model, we …nd some evidence that more able entrepreneurs have larger capital stocks. The regression in column 1 shows that years of schooling, results from a Raven test of non-verbal reasoning, and results from a forward digitspan test all have positive measured e¤ects, though only the digitspan score is statistically signi…cant. s The column 1 regression also includes measures of the owner’ short-term discount rate and an indication of hyperbolic attitudes. Hyperbolicity is proxied by answers to questions asked in the baseline survey involving discount rates over the two di¤erent time horizons, and patience is measured by whether a respondent has a discount rate above or below the sample median. Firm owners were asked hypothetical questions to elicit the amounts that would leave them indi¤erent between an amount today, and 100 cedis in one month; and between an amount in …ve months and 100 cedis in six months. From this we construct two proxies: an indicator variable of whether the respondent has a discount rate above the median for the one month versus today comparison (the median person would take 90 today instead of 100 in one month); and an indicator of being a hyperbolic discounter based on comparison of discount rates over the two horizons (28 percent of the sample are classi…ed as hyperbolic). We …nd that hyperbolic discounters have signi…cantly smaller businesses, consistent with the theory. Those with below median one month discount rates have insigni…cantly larger businesses. Our surveys include two additional measures which each measures of self-control and the 20 ability to save cash, both of which were measured pre-treatment. These refer to actual saving behavior as reported by respondents before treatment. The …rst is whether the respondent used a susu collector; the second, whether the respondent agrees or strongly agrees with the statement “I save regularly” (which two-thirds of respondents do). We use the principal component of these and the two discount rate measures to create an index measuring a lack of self-control. The regression in column 2 uses the index of self control in place of the discount rate measures. Those with less self control have lower baseline capital stocks. The models also predict that individuals who are more patient and less hyperbolic should save more and grow their business faster as long as they are they below their optimal size. Conditional on the log of baseline capital, we do …nd that the growth in capital stock between wave 1 and 2 (i.e., before treatment) is faster for respondents who are more patient and are not hyperbolic. All these results tally with the predictions of the RTI models. Finally, in column 3 we include a dummy for high pressure to share within the household. This variable plays in principle no role in the RTI models, but we …nd that this is also associated with signi…cantly smaller capital stock at baseline. This is consistent with the idea to pressure to share a¤ects …rm size, and suggests that factors other than those in the basic RTI framework may have important e¤ects on …rm size. All three regressions include controls for the age of the owner, and the education level of s the owner’ mother and father. The models also predict that, other things being equal, older entrepreneurs are likely to have larger …rms because they have had more time to accumulate and learn, and thus reach their long-term steady state. The data support this. Parental education is added as a control variable for initial household wealth and socioeconomic status. The regressions also include an indicator that the owners is female. We …nd that female-owned businesses are signi…cantly smaller in terms of baseline capital stock. In sum, the regressions on Table 2 suggest that baseline capital stocks are consistent with several of the predictions made by the RTI model. The signi…cance of pressure to share with others in the household suggests that factors outside the RTI framework may also be important. We now turn to the experimental data to investigate whether exogenous shocks to household capital are invested in ways which are consistent with model predictions. 21 5 Estimation of Experimental Treatment E¤ects Only 9 …rm owners assigned to receive a grant (2% of those assigned to treatment) did not receive one. One of these …rm owners had died, three women refused the grant saying their husbands would not let them accept it, and the other …ve …rms had attrited from the survey and could not be located to give them the grant. Given this, we focus on intent-to-treat e¤ects, which show the impact of being randomly assigned to receive the grant –in practice there is little di¤erence between the intent-to-treat e¤ect and the treatment on the treated e¤ect of actually receiving the grant given that compliance is almost 100%. 5.1 Impact on Pro…ts by Grant Type and Gender Figures 1 and 2 graphically show the main results of the experiment by displaying the empirical cumulative distribution functions (CDFs) of real pro…ts by gender and treatment group for the …nal two rounds of the survey. For males, Figure 1 shows that both the in-kind and cash treatments have distributions to the right of the control distribution, with separation over most of the range of pro…ts. The in-kind and cash treatments have similar distributions up to about the 80th percentile, and then separate with the distribution of pro…ts for the in-kind treatment lying to the right of the cash treatment pro…ts distribution. In contrast, the distribution of real pro…ts by treatment group for females shows two noticeable di¤erences from that of males. First, the distribution of the cash treatment group lies right on top of that of the control group, suggesting no impact of the cash treatment on pro…ts. Second, while the in-kind distribution lies to the right of the other two groups, this separation only occurs at about the 50th or 60th percentile. That is, for women, the in-kind treatment seems to have had an e¤ect only for the top half of the distribution. We then estimate the average impact of the cash and in-kind grants on …rm pro…ts. We begin by pooling together male and female business owners, and running an OLS regression of the form: X G X i;t+s = 1 Mit + 2 Eit + t Dit + g Sig + "i;t+s (9) t g=1 where Mit and Eit are dummy variables indicating whether …rm i has been assigned to receive either the cash or in-kind treatment by time t. The error term uit has been decomposed into wave …xed e¤ects Dit , quadruplet …xed e¤ects Sig , and a residual "it . The G quadruplets are the strata used in the randomization of the two treatments across entrepreneurs (see Section 22 3.2) and are included following the recommendation of Bruhn and McKenzie (2009). We test whether either treatment is signi…cantly di¤erent from zero. We also test the equality of e¤ects of the two treatments 1 = 2. We estimate equation (9) for the full sample, and then for the sub-sample which trims out matched quadruplets which have a …rm with pre-treatment pro…ts above 1500 cedis.11 In addition to OLS estimation conditional on group dummies, we also estimate equation (1) via individual …xed e¤ects. The inclusion of …xed e¤ects controls for any time invariant small-sample di¤erences between treatment groups, such as the di¤erence in baseline pro…ts seen in Table 1. We cluster errors at the …rm level in all speci…cations. The …rst four columns of Table 3 show the treatment e¤ects for the pooled sample. All four speci…cations show a large positive impact of the in-kind treatment on …rm pro…ts. Monthly …rm pro…ts are estimated to be 31-43 cedis higher as a result of the 150 cedis in-kind treatment. The cash treatment is signi…cant at the 10 percent level in the untrimmed OLS speci…cation, but becomes insigni…cant when trimming or using …xed e¤ects. The coe¢ cients are always much smaller than for the in-kind treatment, and we can reject equality of cash and in-kind grants at the 5 percent signi…cant level for three out of four speci…cations and at the 10 percent level for the other. That is, cash grants have less impact on business pro…ts than in-kind grants. These initial results pool together all waves of the survey, thereby giving the average impact of the treatments over the observed time period and improving power (McKenzie, 2011). We observe …rms at quarterly intervals, up to 12 months after treatment. Appendix 2 tests robust- ness to allowing the impact of the grants to vary with the time since treatment, and tests for equality of treatment e¤ects. There is some suggestion that the impact of the in-kind treatments are greater 9-12 months after treatment than immediately afterwards, but we reject equality of treatment e¤ects over time at the 10% level only for the in-kind treatment for females, and then only with a …xed e¤ects speci…cation. Given the sample sizes we have and lack of strong evidence to reject pooling, we therefore continue to pool all waves for the remainder of the paper. In the remainder of Table 3 we allow the impact of the grants to vary by gender. Recall the randomization was strati…ed by gender. We modify equation (9) to allow both the treatment 11 Only 7 …rms have pre-treatment pro…ts above this level, but this trimming involves dropping 28 …rms (1% of the sample) since we need to drop other …rms in the matched quadruplet. Doing this ensures that balanced randomization occurred within the trimmed sample, and prevents a few …rms with scale well above the rest of the sample exerting undue in‡uence on the results. 23 and wave e¤ects to vary by gender: i;t+s = 1 Fi Mit + 2 Fi Eit + 3 (1 Fi )Mit + 4 (1 Fi )Eit X X G X F + t Dit + t Fi Dit + g Sig + "i;t+s (10) t t g=1 where Fi = 1 if entrepreneur i is female, and 0 otherwise. Recall the randomization was strati…ed by gender. Columns 5 and 6 estimate equation (10) by OLS with quadruplet dummies, and columns 7 and 8 with individual …xed e¤ects. Finally, columns (9) and (10) restrict the OLS estimation to the last two waves of data. This corresponds to the data in Figures 1 and 2.12 For women, the estimated treatment e¤ect of the cash grant is always small (5 cedis or less) and statistically insigni…cant, whereas the treatment e¤ect of the in-kind grant is large (35-50 cedis) and statistically signi…cant. In all speci…cations we can reject equality of the cash and in-kind treatment e¤ects. This con…rms what is seen visually in Figure 2, that only the in-kind grants have a signi…cant e¤ect for women. For males, the in-kind treatment e¤ect is also large, although more sensitive to speci…cation, ranging in size from 28 to 60 cedis, and statistically signi…cant in all but one speci…cation. After trimming, the magnitude of the in-kind treatment e¤ect for males is very similar to that for females, and we cannot reject equality of in-kind treatment e¤ects by gender in any speci…cation. In contrast to females, we can never reject equality of cash and in-kind treatment e¤ects for males. This is consistent with a failure of asset integration for females but not for males. The cash treatment e¤ect for males is statistically signi…cant and large when we restrict analysis to waves 5 and 6, which is consistent with the e¤ects seen in Figure 1. However, using all waves of the data, the estimated impact varies between 5 and 29 cedis depending on speci…cation, with large standard errors. The impact of cash is larger using OLS than …xed e¤ects because of an imbalance in baseline pro…ts for males. The group assigned to the cash grant has higher wave 1 pro…ts (despite the same wave 2 pre-treatment pro…ts) than either the control group or the group assigned to the in-kind treatment. Because we balanced on wave 2 pro…ts in the randomization, the imbalance is due to chance. It is therefore not clear whether or 12 Readers may be concerned that pro…ts are arti…cially high in the quarter immediately after the equipment treatment if …rms receiving inventories to sell count this as pure pro…t. But Appendix 2 shows that, if anything, the treatment e¤ect is rising with time since treatment. Furthermore, the treatment e¤ects are still present when focusing on these …nal rounds which are six months or more removed from almost all the treatments. 24 not one should control for this pre-treatment di¤erence. If we are prepared to treat this chance imbalance as noise and not condition on it, then there is some evidence for a signi…cant cash e¤ect, at least in the last two rounds. But the con…dence interval for the male cash treatment e¤ect when we do control for it with …xed e¤ects is (-26.5, +36.7), indicating that the data really have no information about the cash treatment e¤ect for males when we condition on this di¤erence. In contrast, the in-kind treatment e¤ect for males and the cash and in-kind treatment e¤ects for females are much more robust to the choice of speci…cation, giving us more con…dence in the results for these groups. These …ndings contrast with earlier experiments which failed to reject the equality of cash and in-kind treatment e¤ects in grants to microenterprises in Sri Lanka and Mexico (de Mel et al, 2008; McKenzie and Woodru¤, 2008). Why? We …rst note that the sample sizes used in the experiments in Sri Lanka and Mexico ulti- mately resulted in low power to distinguish between cash and in-kind grants. In Sri Lanka the sample size for analysis in de Mel et al. (2008) is 385 …rms. They …nd an ITT return of 4.17 percent for the in-kind treatment, and of 6.70 percent for the cash treatment. However, the standard errors on these estimates are 2.6-2.8 percent, and one cannot reject at the 10 percent level that the in-kind treatment has twice the e¤ect of the cash treatment (p=0.102). The sam- ple size in Mexico for the analysis by McKenzie and Woodru¤ (2008) is even smaller, with 113 …rms after trimming. The ITT returns are 34.4 percent for cash and 16.1 percent for in-kind, but with standard errors of 23-24 percent. Thus it may be that Ghana is not di¤erent, but rather that the prior studies did not have su¢ cient sample to detect di¤erences between cash and in-kind grants. We return to this issue below. 5.2 Treatment Heterogeneity by Randomization Strata Next we examine treatment e¤ect heterogeneity according to the other variables used for strat- i…cation and matching. We do this separately by gender, given the di¤erences observed above. Let A and B denote the two categories of a binary variable used for strati…cation (e.g. Ai = 1 if i works in a single-sex dominated industry, and Bi = 1 if i works in a mixed-gender industry). 25 Then we estimate separately for each gender: i;t+s = 1 Ai Mit + 2 Ai Eit + 3 Bi Mit + 4 Bi Eit 6 X 6 X G X + t Dit + t Bi Dit + g Sig + "i;t+s : t=2 t=2 g=1 The results are shown in Table 4. The top two rows of the table show the categories A and B which de…ne strata. Columns (1) and (2) show the OLS and …xed e¤ects estimates of treatment heterogeneity by the gender mix of the industry …rms work in. De Mel et al. (2009a) found some evidence in Sri Lanka that the impact of grants was less for women in female-dominated industries than those in mixed industries. In Ghana, panel A of column (2) shows that with …xed e¤ects, the cash treatment has a -6.9 cedis e¤ect in female-dominated industries versus a 1.8 cedis e¤ect in mixed industries, and the in-kind treatment has a 25.4 cedis e¤ect in female-dominated industries compared to a 39.8 cedis e¤ect in mixed industries. The point estimates are therefore consistent with the idea that the grants may have more e¤ect on the businesses of women who operate in mixed industries. However, the di¤erences in treatment e¤ects by industry category are not statistically signi…cant. Likewise panel B shows no signi…cant heterogeneity by industry category for men. Columns (3) and (4) examine heterogeneity according to the baseline measure of capture. Recall that individuals in the “high capture” category state that whenever they have money on hand their family members are likely to request some of it, and that people who do well in business get requests from others for help. We do not obtain signi…cant heterogeneity according to this variable for either men or women, with large standard errors and the point estimates varying quite a lot between the OLS and …xed e¤ects speci…cations. Later in the paper we examine alternative measures of capture to see whether this lack of signi…cance is due to the particular choice of measure being used. Finally we look at heterogeneity according to the initial size of the …rm. Columns (5) and (6) consider this in terms of the initial capital stock of the …rm, as …rms were strati…ed as being above or below median baseline capital stock, while columns (7) and (8) de…ne initial size in terms of initial pro…ts. Since wave 2 pro…ts was matched to form quadruplets, we …rst calculate the maximum wave 2 pro…t within a quadruplet or group, and then de…ne …rms as being in a low pro…ts group if the maximum wave 2 pro…ts for the group is less than 138 cedis (the median of pro…ts over the whole sample). This classi…es 62 percent of females and 45 percent of males as being in the low pro…ts group. The results con…rm the visual impression in Figures 1 and 2. 26 In particular, we see that the cash grants have no signi…cant impact for any size female …rm, while the in-kind grants only have an impact for the 40 percent or so of …rms with higher initial pro…ts or higher initial capital stock. The impact of the in-kind grants is extremely large for these female …rms – monthly pro…ts increase by 77 to 96 cedis per month for the female …rms in high initial pro…ts quadruplets, compared to an insigni…cant 2 to 5 cedis per month for the low pro…ts female …rms. This di¤erence is statistically signi…cant. In contrast, there is no such pattern for male-owned …rms –the point estimates for the lower pro…ts …rms are typically just as large as those for the higher pro…ts …rms, and the di¤erence is not statistically signi…cant. Taking these results together, it appears that cash grants are not increasing pro…ts for female-owned …rms, and the in-kind grants only increase pro…ts for female-owned …rms which were larger in size to begin with. The in-kind treatments also increased pro…ts for male-owned …rms, and the e¤ect of the cash grants is inconclusive for males. There does not appear to be the same heterogeneity by initial …rm size in terms of male responsiveness to the grants. This suggests that female-owned …rms are better able to resist internal or external pressure when the treatment comes in the form of inventories or equipment, while the di¤erence in treatment is less relevant for male-owned businesses. We now scrutinize this interpretation more in detail. 6 Interpretation of the results 6.1 Where do the grants go? Table 5 examines the extent to which the grants are being used to increase the capital stock of the …rm, to make transfers to non-household members, and for household spending. In panel A we show the results of estimating equation (2) with di¤erent outcomes, while in panel B we show the results of estimating equation (3) for the female sample and the categorization of low and high initial pro…ts groups, since this is where we found large di¤erences in treatment e¤ects. For reasons of space we report the …xed e¤ects estimates only (with the exception of transfers out which were not measured pre-treatment), but note when the OLS results show large di¤erences. We begin by looking at the impact of the grants on the capital stock of the …rms. Column (1) shows this for total capital stock. In order to reduce the in‡uence of large outliers, column (2) truncates capital stock at the 99.5th percentile, which is 6130 cedis. Both speci…cations suggest that capital stock is increasing by more for the in-kind treatments than for the cash treatments, both for men and women. However, the capital stock data is noisy and the standard errors 27 are large, meaning we cannot reject equality of the e¤ect of cash and in-kind grants on capital stock. Panel B shows stark di¤erences between the women whose pro…ts were initially low and those who had higher initial pro…ts –there are large increases in capital stock for the high initial pro…ts group, and no increase in capital stock for the low initial pro…ts group that received the cash treatment. After truncating outliers, we can reject equality in treatment e¤ects for the low and high initial pro…ts groups for both cash and in-kind grants. Figures 3 and 4 show the empirical CDFs of the post-treatment capital stock distribution by treatment group and gender for the …nal two waves of the survey. For males, Figure 3 shows a similar pattern to that of pro…ts – namely that the distribution of the in-kind treatment group is shifted to the right compared to that of the control group across the distribution. The cash distribution is in between, although right at the top of the distribution crosses the control distribution curve several times, which explains the sensitivity of the cash treatment e¤ect to where we truncate the data. For females, Figure 4 shows that both treatment groups overlap with the control group for the bottom 60 percent of the distribution, a pattern similar to that seen for pro…ts. The in-kind grant distribution then separates from the control above this, with women in the in-kind treatment group having higher 70, 80, and 90th percentiles of their capital stock distributions than the control group. The cash treatment group lies in between, and, unlike in the case of pro…ts, does separate somewhat from the control group at the top of the distribution, suggesting some increases in capital for some …rms as a result of the cash treatment. Next we examine where the grants are going if not into the business. Beginning in wave 4, …rm owners were asked “During the past three months, did you make any payments in cash or goods to people living outside your household?” and if so, the value of such transfers. Columns 3 and 4 show that women who received the cash grant were more likely to have made such a transfer, and to have given more. On average they are estimated to have given 8 cedis more a quarter over the last 3 quarters of the survey. This does not account for any transfers out made in the …rst quarter after treatment by …rms treated after wave 2, since the wave 3 survey did not collect transfers data. However, restricting the analysis to the control group and …rms treated after wave 3 only marginally increases the coe¢ cient on the cash treatment, raising it to 8.9 cedis. The remaining columns report the estimated impacts on household expenditure, which was 28 collected each wave.13 Point estimates suggest higher positive impacts on expenditure for those receiving the cash treatments than those getting the in-kind treatment or the control group, especially for women with low initial pro…ts. We see a large and highly signi…cant e¤ect of the cash treatment on total quarterly spending for women as a whole, and for the subgroup of women with low initial pro…ts. The coe¢ cients are huge: women who were given a 150 cedis cash grant are estimated to be spending 120 cedis more a quarter after the grant. The magnitude of this coe¢ cient appears to be driven by a few …rm owners reporting very large spending levels –truncating at the 99th percentile of total expenditure lowers this coe¢ cient to 95, and at the 95th percentile lowers it to 76 cedis (which is still signi…cant at the 5% level). For males receiving the cash treatment, the point estimates also suggest large increases in total quarterly spending (with a coe¢ cient of 50 to 73 cedis depending on the level of truncation), but the standard error is so large that we can never reject equality with zero. These results therefore o¤er an explanation at a basic level for the pro…ts results. More of the in-kind grants ended up in the business than the cash grants. Women, especially those with lower initial pro…ts, appear to have spent most, if not all, of the grants on household expenditure and transfers to non-household members. As a result, we see more impact of in-kind grants than cash grants on business pro…ts. Note however that for women with high initial pro…t who received the cash treatment, Table 5 shows an increase in capital stock and no signi…cant change in spending, despite the lack of a signi…cant increase in pro…ts for this group. There are only 44 women who are in the high initial pro…ts group who received the cash treatment, so small sample noise is a potential explanation. Indeed, the 95% con…dence interval for the e¤ect on pro…ts of the cash treatment for the initial high pro…t group is (-26, +40). So despite the small and insigni…cant 13 s s Households were asked to recall the last week’ expenditure on food, last month’ expenditure on 9 categories (housing, fuel and light, non-durable households goods, communication, recreation, transport, household services, s personal care services, and contributions to associations) and last three month’ expenditure on a further 9 categories (clothing, footwear, ceremonies such as funerals and weddings, electronic goods, household furnishings, household appliances, vehicles, health expenses, and education expenses). We aggregate several categories to report estimates of impact on several categories of interest, as well as impacts on total quarterly spending (which s adds 13 times weekly food, and 3 times the last month’ expenses to the three month expenses). Unlike pro…ts, panel consistency checks were not programmed for these expenditure items, and the data are quite noisy. In order to ensure extreme outliers are not driving the reported results, we report results using expenditures truncated at the 99.5th percentile. Results using the untruncated expenditures are qualitatively similar with larger standard errors, and slightly larger point estimates. The impacts on speci…c household expenditure categories are not well-identi…ed due to this noise. 29 point estimate, we cannot completely rule out the possibility that the high initial pro…t women receiving the cash treatment are also investing it in the business and bene…t through higher pro…ts. 6.2 How do the low and high initial pro…t women di¤er? We have seen that the impact of the grants di¤ers greatly between women with low initial pro…ts and women with high initial pro…ts. It is therefore worth examining in more detail the t composition of these two subsamples. The …rst point to note is that these groups don’ di¤er greatly in the industry or type of business, just in the scale. The low initial pro…t group is made up of 31 percent food sales, 18 percent beauty and hair, 9 percent sewing, and 42 percent trade, compared to 37 percent food sales, 9 percent beauty and hair, 6 percent sewing and 47 percent trade for the high initial pro…t group. Even when we look more …nely within these broad sectors, we see a similar broad range of types of …rms in both subgroups: kenkey and banku (both traditional prepared foods) sellers, dressmakers, beauty salons, used clothes sellers, and retail trade. In contrast, the scale of the …rms di¤ers substantially. Table 6 compares the pre-treatment characteristics of these two subgroups of female …rms to each other and to the male-owned …rms. The …nal column also o¤ers a comparison to the sample of female microenterprises from Sri Lanka used in de Mel et al. (2009a). We see that mean and median monthly pro…ts for the low initial pro…ts female subsample is 37-38 cedis, approximately US$1 per day, while mean and median pro…ts are 4 to 6 times this level in the high pro…t group. Similarly, mean and median sales di¤ers by a factor of 5 to 6 between the low and high initial pro…t groups. Mean capital stock for the low initial pro…ts group is 251, versus 456 for the high pro…ts group. Comparing to the other two groups, we see that the high initial pro…t females have larger pro…ts than the average male-owned …rms in the sample, while the low initial pro…ts group are similar in size to the female-owned …rms in the Sri Lankan study. Table 6 also shows that women in the high initial pro…ts group are more educated, have richer households (which may be a consequence of the higher pro…ts rather than a cause), are more likely to keep accounts and to have had a formal loan, and have been in business slightly longer than the low initial pro…ts …rms. When it comes to the reasons for choosing a particular sector, women in the high pro…ts group are more likely to say they chose their sector for earnings potential and less likely to say they chose it because it had a low capital requirement. 30 Overall this paints a picture of the low pro…ts group as much smaller in size, with subsistence level income. For this group we see no impact of the grants on business pro…ts. This is consistent with the …nding in Sri Lanka, where the grants had no impact on female-owned businesses. The Sri Lankan businesses are similar in scale to the low initial pro…ts female …rms in Ghana –the 95th percentile of pro…ts is only 70 per month in the Sri Lankan sample, which is the 10th percentile of pro…ts for the high initial pro…t group in Ghana. So for the types of female-owned businesses in Ghana that are similar in scale to those in Sri Lanka, we obtain similar results. Such businesses …t the hypothesis of Emran et al. (2007) that many of the women drawn into subsistence self-employment have very low e¢ cient scale and are there only because of labor market imperfections. As a result, according to our models, neither cash nor in-kind grants should have any long-term impacts on business pro…tability for these types of low productivity …rms. The di¤erence is that the Ghanaian sample also includes a group of more successful female-owned businesses with larger scale, who do show increased pro…t growth from at least the in-kind treatment. 7 Why does the impact of cash and in-kind treatments vary? We now examine more in detail the possible reasons for the di¤erence between the impact of cash and in-kind grants, particularly for women. We …rst discuss whether the RTI models can account for this di¤erence. The only way that the RTI models we presented in the conceptual section can account for a di¤erence between the two treatments is if the in-kind treatment Et cannot be liquidated immediately and the …rm has already reached its steady-state. If this is the case, pro…ts and capital are predicted to rise above their steady state after the in-kind treatment, but only until illiquid capital can be divested. This e¤ect is unlikely to be prominent in our study. There is a relatively long lag between treatment and ex post surveys. The time lag between receiving a grant and the immediately subsequent survey is nearly 3 months. For those who received the grant after round 2, there is an interval of nearly 12 months between receiving the grant and the last survey round. For those receiving the grant after round 3, the time interval is 9 months. Given that recipients of in-kind grants invest in raw materials, inventories, or simple equipment, it would be easy to decapitalize in-kind grants between treatment and the next survey round. Furthermore, as we 31 will see below, there is no evidence of divestment of the in-kind grants. Another possibility is to modify the RTI setup to allow for learning. To see this formally, remember that in the RTI models the steady state capital stock is set by equating the marginal return to capital 0 (k ; ) to the relevant discount rate – either or r or , depending on the e version of the model. If the entrepreneur believes to be lower than its true value, say, < , then the …rm may stop growing before reaching its true steady state. Having been forced to expand thanks to an in-kind grant, the entrepreneur may realize that he or she can successfully e e handle a larger …rm, i.e., that > . If there are enough entrepreneurs with < , this could explain our results. Is there evidence in support of this hypothesis in the data? Prior to treatment, respondents were asked whether they expected sales to increase if they invested more capital. Entrepreneurs e for whom < should respond negatively to this question. If learning explains our results, the e¤ect of the in-kind treatment should be largest for entrepreneurs who expect the lowest increase in sales from increasing capital. We test this hypothesis and …nd instead that the in- kind treatment has larger – not smaller – e¤ects for …rm owners who expected to see a larger increase in sales from investing more capital. There is therefore little evidence for the learning hypothesis in the data. We thus have to turn to potential explanations that focus on asset non-integration, such as internal and external pressure. The surveys contain variables that proxy for the degree of self- control and the degree of external pressure to share facing business owners. With the exception of the “high capture”dummy which we examined in Table 4, treatment randomization was not strati…ed on any of these variables. Our analysis should therefore be considered exploratory in nature. We have four measures which each measure an aspect of self-control and the ability to save cash, all of which were measured pre-treatment. The …rst two refer to actual saving behavior as reported by respondents before treatment: whether the respondent used a susu collector and whether the respondent agrees or strongly agrees with the statement “I save regularly”(which two-thirds of respondents do). These …rst two variables have the advantage that they capture actual, not hypothetical, behavior. But they could potentially be a¤ected by external pressure, for instance if someone is unable to save because of a pressure to redistribute, or enters in a contract with a susu collector to keep money away from a prying husband, as in Anderson and Baland (2002). The other two variables we use are the discount rate and measure of hyperbolicity, 32 which are less subject to this possible confounding e¤ect, and are thus arguably more reliable. We also extract the common signal in these four variables by forming a principal component, which we call “lack of self-control” 14 High scores indicate a lower likelihood of using a susu and . of saving regularly, and a higher likelihood of being hyperbolic and of having a high discount rate. The lack of self-control index is thus the common component of using a susu and saving regularly that is correlated with hyperbolic time preferences. The …rst …ve columns of Table 7 estimate model (8) to examine the heterogeneity in the treatment e¤ect with respect to these self-control variables for the pooled sample in panel A, and for females only in panel B. Point estimates show a larger impact of the cash treatment when the …rm owner uses a susu, saves regularly, is patient, and is not hyperbolic. The signs of the coe¢ cients on the interaction with cash treatment for each of the four variables are consistent with the idea that lack of self-control is associated with a smaller increase in business pro…ts from the cash treatment. Column 5 shows that when we combine all these variables into an index of self-control using principal components, the interaction between cash treatment and lack of self-control is signi…cant at the 5 percent level. This implies that the e¤ect of the cash treatment for someone with the most self-control is 29.2, which is similar to the average impact of the in-kind treatment of 30.9 in column 4 of Table 3. The female only sample in panel B shows a similar pattern in terms of signs of the coe¢ cients, but interaction terms are not signi…cant, possibly because the sample is smaller. We also investigate whether cash and in-kind grants have a di¤erentiated e¤ect on capital retention depending on entrepreneur self-control. Figure 5 presents the evolution of capital stock after treatment, contrasting between cash and in-kind grants for entrepreneurs with and without self-control. The Figure shows that in-kind grants are retained in the …rm and lead to further accumulation of capital whether the entrepreneur lacks self-control or not –although the e¤ect on capital accumulation is strongest for those with low self-control. In contrast, cash grants have, if anything a negative e¤ect on capital accumulation among low self-control entrepreneurs while the e¤ect is positive and growing for those with high self-control. This con…rms that the reason why in-kind grants have a strong positive e¤ect on pro…t is related to self-control issues and their e¤ect on the accumulation of capital. If we further interact the self-control index with high baseline pro…ts, we …nd that self-control a¤ects the impact of treatment only for female entrepreneurs with high baseline pro…ts. For those with low initial pro…ts, treatment 14 The principal component has mean zero, and ranges from -1.67 to 2.26. 33 has no e¤ect irrespective of self-control. If we are willing to regard baseline pro…ts as proxy for ability , this suggests that self-control is an impediment to growth only for more able female entrepreneurs, who have more potential for growth. Next we examine whether the pattern of heterogeneous treatment e¤ects in the sample is consistent with external pressure. We …rst investigate capture by spouses. Unmarried individuals are not, by de…nition, subject to capture by a spouse. Consequently, if capture by a spouse is the main reason why cash treatments do not raise pro…t, this should not a¤ect unmarried entrepreneurs –especially women. To test this hypothesis, we interact treatment with a married dummy and expect the cash treatment e¤ect now to be positive and the interaction term to be negative, especially for female entrepreneurs. The percentage of unmarried individuals in our sample is 34% for females and 35% for males. Results, presented in columns (1) and (2) of Table 8, show that the cash treatment remains non-signi…cant for female entrepreneurs. None of the interaction terms is statistically signi…cant but point estimates for females are positive for cash and negative for in-kind, which is the opposite of what the external pressure hypothesis predicts. Married status is also not (negatively) correlated with baseline capital stock, as should be case if marriage leads to more capture. Next we examine the determinants of bargaining power between married entrepreneurs. The hypothesis is that less of the cash and in-kind treatment is taxed away from the business when a married entrepreneur has more bargaining power with his or her spouse. We expect the e¤ect of bargaining to be stronger for female entrepreneurs, given their subordinate status in traditional society. The …rst proxy for bargaining power is the share of assets brought to marriage by the owner. This information is available for 397 married respondents. As shown in columns (3) and (4) of Table 8, point estimates are insigni…cant and do not go in the expected direction: we would expect those who bring a higher share of assets into the marriage to have more bargaining power, yet the interaction of this with the cash treatment is negative. We also test (not shown) whether the share of assets brought to marriage is correlated with baseline capital stock, as should be the case if entrepreneurs with more bargaining power are less subject to spouse taxation and can accumulate more over time. We …nd no evidence of this. We investigate other possible determinants of intra-household bargaining power, such as di¤erences between spouses in age and education level. We expect older, more educated spouses to have more bargaining power. The results, shown, in columns (5) to (8) of Table 8, are once 34 again not signi…cant and the point estimates of the interaction terms are small. In columns (9) and (10) we use more subjective data: namely whether the respondent agrees that they can spend their income without consulting their spouse (72 percent of women and 71 percent of men say they can). By construction, this information is only available for married respondents. We again …nd no statistically signi…cant interaction e¤ect with the cash treatment and point estimates small for women. The last proxy for bargaining power that we investigate is the share of individual income in total household income or expenditure. Much of the literature on intra-household bargaining has used relative incomes as proxy for bargaining power, either because having more income gives more agency, or because earning an income outside the home results from having more bargaining power. We construct our proxy using income and consumption information at baseline. We expect spouses with a larger income share to retain more of the treatment and thus to have a larger treatment e¤ect.15 Results, presented in columns (11) to (14) of Table 8, show no statistically signi…cant interaction. We next turn to external pressure more broadly de…ned, i.e., including individuals other than the spouse. The …rst proxy we consider is the absolute size of the household. Presumably, the more members a household has, the more people there are who can put pressure on the respon- dent to share cash and in-kind grants. Pressure need not be explicit, though. If respondents are altruistic towards other members of their household, they may feel the need to share the grants with them instead of investing. In this case the needs of others are internalized and no explicit pressure need be applied. The next proxy is whether the …rm owner says they feel a lot or some pressure to share extra business income with other household members rather than invest in the business. 23 percent of females and 22 percent of males claim to feel this pressure. This information was only collected in the last survey round, after treatments, so results should be considered as suggestive only. There is, however, no di¤erence in means on this variable between treated and untreated samples. As a proxy for pressure from outside the household, we include the number of siblings in the Accra/Tema area. Presumably, the more siblings a person has, the more relatives there 15 Share of household income or consumption is not an ideal proxy for our purpose since it also re‡ects entrepre- neurial ability, and more able entrepreneurs may bene…t more from treatment. For a given ability level, a higher share of household income also implies that other household members are not high earners and hence that the entrepreneur is more capital constrained. This too would predict a higher treatment e¤ect. Given that we …nd no signi…cant e¤ect, this issue is moot. 35 are who could put pressure on the respondent to share the grants – or the more relative there are towards whom the respondent may have altruistic feelings. Results are presented in Table 9. When we look at the full sample we do not see any signi…cant interactions between the cash treatment e¤ect and proxies for external pressure. The size of the interaction terms on pressure to share is too small to generate large positive impacts of the cash treatment. When we consider the females-only sample in panel B, we get a signi…cant but positive interactions on household size and on the number of siblings in the area –i.e., the opposite sign of what we would expect from external pressure. This is nevertheless consistent with Grimm et al. (2010) who …nd a positive impact of nearby networks on …rm growth. These …ndings are also in line with the lack of sizeable or signi…cant interaction with low capture seen in Table 4 for the low capture/high capture variable we strati…ed on. Taken together, the evidence from Tables 7, 8 and 9 appears more consistent with self-control rather than external pressure being the cause of the lack of an e¤ect of the cash treatment. One might however worry that people who have trouble saving or have time-inconsistent preferences and high discount rates are also those who more external pressure. To examine this possibility, we test in Table 10 whether proxies for external pressure are signi…cantly associated with the lack of self-control index. In addition to the variables used to proxy external pressure in Table 7, we also consider: whether they think inventories and t equipment held in the business are a good way of saving money so others don’ take it (55 percent of women and 72 percent of men say yes); whether the respondent had been compelled to give money to their spouse in the three months prior to the baseline (10 percent of married women and 15 percent of married men said yes); and whether their spouse is supportive of them running a business (84 percent say yes for both men and women). We cannot reject the null hypothesis that none of the proxies for external pressure are associated with the self-control index. While measuring external pressure is di¢ cult, if self- control was merely proxying for external pressure, we would expect some relationship, yet there is none. We therefore conclude that there is more evidence to support the hypothesis that the di¤erence between cash and in-kind treatments is driven by self-control issues than by external pressure. 36 8 Conclusions We …nd that the e¤ect of in-kind and cash grants is signi…cantly di¤erent, a …nding that is di¢ cult to reconcile with models of accumulation that take either a standard Ramsey form, or allow for pure time inconsistence. These results suggest a lack of asset integration, as if entrepreneurs fail to take consumption and investment decisions jointly. The di¤erence between in-kind and cash grants is suggestive either that inventories and equipment serve as a self- commitment device against impulse purchases –or that entrepreneurs evade a social solidarity ow tax, by household members and relatives, on the cash ‡ of the …rm but not its equipment and inventories. We …nd that cash and in-kind grants have signi…cantly di¤erent e¤ects for female entre- preneurs in Ghana. In-kind grants lead to large increases in business pro…ts, but only for t female-owned …rms which were initially more pro…table – subsistence …rms don’ grow when given more capital. In-kind grants also lead to large increases in business pro…ts for men, while the e¤ect of cash grants is less robust – we …nd large positive and signi…cant e¤ects when we t don’ condition on baseline pro…ts, but smaller and insigni…cant e¤ects when we do. The di¤er- ence between cash and in-kind treatments is strongest among successful female entrepreneurs, that is, those with high pre-treatment pro…ts. We seek to identify the reason for the di¤erence in treatment between cash and in-kind grants, i.e., whether it originates in self-control di¢ culties or in pressure from household and family members. The fact that the di¤erence in treatment e¤ects is only statistically signi…cant among female entrepreneurs is suggestive of external pressure: given the social context, women entrepreneurs are expected to be subjected to pressure from husband and children. However, we fail to …nd con…rmatory evidence for this hypothesis when interact treatment with proxies for external pressure. In contrast, we …nd that individuals with more self-control di¢ culties respond better to treatment in terms of pro…ts. There is no evidence that female entrepreneurs lack self-control more than men. Ghana o¤ers a setting where women are the majority of small business owners, and in this setting we …nd the top 40 percent of women in terms of pro…tability look similar or more pro…table than the average male …rm. Such a large group of relatively high achieving women is not present in the Sri Lankan sample of de Mel et al. (2009a), and indeed the remaining group of subsistence-level Ghanaian female business owners have similar negligible business impacts from the grants as the group of women in the Sri Lankan experiment. 37 The results o¤er partially good news for advocates of directing micro…nance at women. We do …nd in Ghana a relatively large group of women whose pro…ts increase a lot when given in-kind transfers. Microcredit has been argued as allowing individuals to overcome present-bias by providing self-discipline and encouragement through regular payments and group meetings (Bauer et al, 2010). If this is true, the e¤ectiveness of micro loans in improving business outcomes is likely to resemble the e¤ect of in-kind grants in our experiment. However, our …ndings suggest this e¤ect to be more powerful for women who are already earning more to begin with, suggesting possible limits on the ability of capital alone to generate business growth among poor subsistence- level female enterprises. Moreover, as in prior work in Sri Lanka and Mexico, the results show that the average male-owned microenterprise gains a lot from being granted additional access to capital. This suggests that micro…nance programs that focus primarily on women may be ignoring a large group of enterprises with a need for more capital. Finally, our results suggest that loans and grants intended to help female enterprises grow would work better if disbursed in kind, not in cash. Banks already seem to know this, given that bank loans are nearly always in kind: when borrowing to purchase a house, car, or machine, the bank does not hand over the money to the borrower but rather pays the supplier/seller directly.16 If their objective is to foster enterprise development, and not just saving, micro- …nance organizations may want to adopt similar practices. References Andersen, Stephen, Glenn W. Harrison, Morten I. Lau, and E. Elisabeth Rutstrom (2008) "Risk Aversion in Game Shows", Research in Experimental Economics, 12: 361-406 Anderson, Siwan, and Jean-Marie Baland (2002). "The Economics of Roscas and Intra- household Resource Allocation", Quarterly Journal of Economics, 117(3): 963-95, August , Besley, Timothy (1995) “Savings, credit and insurance” pp. 2123-2207 in Jere Behrman and T.N. Srinivasan (eds.) Handbook of Development Economics Volume III, Elsevier: Amsterdam. Baland, Jean-Marie, Catherine Guirkinger and Charlotte Mali (2007) “Pretending to be , Poor: borrowing to escape forced solidarity in credit cooperatives in Cameroon” Mimeo. Uni- versity of Namur. o Banerjee, Abhijit and Esther Du‡ (2009) “The Experimental Approach to Development 16 This also applies to letter-of-credit-based lending for imports. This even applies, albeit partially, to lines of credit given that the bank observes the identity of the payee – and will object/block payment if an enterprise’s overdraft facility is used to pay for consumption items, such as a new car. 38 , Economics” Annual Review of Economics 1:151-78. o, Banerjee, Abhijit, Esther Du‡ Rachel Glennester and Cynthia Kinnan (2010) “The Miracle , of Micro…nance? Evidence from a Randomized Evaluation” BREAD Working Paper no. 278. Banerjee, Abhijit and Sendhil Mullanaithan (2010) “The Shape of Temptation: Implications , for the Economic Lives of the Poor” MIT Working Paper no 10-9. Bauer, Michal, Julie Chytilova and Jonathan Morduch (2010) “Behavioral Foundations of , Microcredit: Experimental and Survey Evidence from Rural India” IZA Working Paper no. 4901. Bigsten, Arne, Paul Collier, Stefan Dercon, Marcel Fafchamps, Bernard Gauthier, Jan- Willem Gunning, Anders Isaksson, Abena Oduro, Remco Oostendorp, Cathy Patillo, Mans Soderbom, Francis Teal, Albert Zeufack, and Simon Appleton (2000) "Rates of Return on Phys- s ical and Human Capital in Africa’ Manufacturing Sector", Economic Development and Cultural Change, 48(4): 801-27. Bjorvatn, Kjetil, Lars Ivar Oppedal Berge, and Bertil Tungodden (2011) "Human and …nan- cial capital for microenterprise development: Evidence from a …eld and lab experiment", NHH and CMI (mimeograph) Bruhn, Miriam and David McKenzie (2009) “In Pursuit of Balance: Randomization in Prac- , tice in Development Field Experiments” American Economic Journal: Applied Economics 1(4): 200-32. Brune, Lasse, Xavier Gine, Jessica Goldberg, and Dean Yang (2011). "Commitments to Save: A Field Experiment in Rural Malawi", University of Michigan, May (mimeograph). Camerer, Colin, Linda Babcock, George Loewenstein, and Richard Thaler (1997). "Labor Supply of New York City Cabdrivers: One Day at a Time", Quarterly Journal of Economics, 112: 407-41. , Charlier, Florence (1999) “Saving or Sharing: The African Households’Dilemma” Stanford Department of Economics Ph.D. Dissertation. Deaton, Angus (2010) “Instruments, Randomization, and Learning about Development”, Journal of Economic Literature, 48(2): 424–55 De Mel, Suresh, David McKenzie, and Christopher Woodru¤ (2008). “Returns to Capital: Results from a Randomized Experiment.” Quarterly Journal of Economics 123(4): 1329–72. De Mel, Suresh, David McKenzie, and Christopher Woodru¤ (2009a) “Are Women more Credit Constrained? Experimental Evidence on Gender and Microenterprise Returns.” Ameri- 39 can Economic Journal: Applied Economics 1(3): 1-32. De Mel, Suresh, David McKenzie and Christopher Woodru¤ (2009b) “Measuring microen- , terprise pro…ts: Must we ask how the sausage is made?” Journal of Development Economics 88(1): 19-31 Di Falco, Salvatore and Erwin Bulte (2009) “The Dark Side of Social Capital: Kinship, , Consumption, and Investment” Mimeo. University of Kent. o, Du‡ Esther, Michael Kremer and Jonathan Robinson (2010) “Nudging Farmers to Use , Fertilizer: Theory and Experimental Evidence from Kenya” American Economic Review forth- coming. o, Du‡ Esther and Christopher Udry (2004) “Intrahousehold Resource Allocation in Côte d’Ivoire: Social norms, separate accounts, and consumption choices” , NBER Working Paper no. 10498. Emran, M. Shahe, AKM Mahbub Morshed and Joseph Stiglitz (2007) "Micro…nance and Missing Markets", Mimeo. George Washington University. Fafchamps, Marcel, Christopher Udry, and Katherine Czukas (1998). "Drought and Saving in West Africa: Are Livestock a Bu¤er Stock?", Journal of Development Economics, 55(2): 273-305, April Fafchamps, Marcel, David McKenzie, Simon Quinn and Christopher Woodru¤ (2010) “Using PDA consistency checks to increase the precision of pro…ts and sales measurement in panels”, Journal of Development Economics, forthcoming. Grimm, Michael, Flore Gubert, Ousman Koriko, Jann Lay and Christophe Nordman (2010) “Does forced solidarity hamper entrepreneurial activity? Evidence from seven West-African , countries” Mimeo. International Institute of Social Studies, Eramus University Rotterdam. Harrison, Glenn, Morten I. Lau, and E. Elisabeth Rutstrom (2007). "Estimating Risk Atti- tudes in Denmark: A Field Experiment", Scandinavian Journal of Economics, 109(2): 341-68. Jakiela, Pamela and Owen Ozier (2011) "Does Africa need a rotten kin theorem? Experi- mental evidence from village economies", Mimeo. University of Washington, St Louis. Karlan, Dean and Jonathan Zinman (2010) “Expanding Microenterprise Credit Access: Us- , ing Randomized Supply Decisions to Estimate the Impacts in Manila” Mimeo. Yale University. Laibson, David (1997) "Golden Eggs and Hyperbolic discounting", Quarterly Journal of Economics 112(2): 443-77. Lee, David (2009) “Training, Wages, and Sample Selection: Estimating Sharp Bounds on 40 , Treatment E¤ects” Review of Economic Studies 76(3): 1071-1102. McKenzie, David (2011) "Beyond baseline and follow-up: The case for more T in experi- ments", World Bank Policy Research Working Paper no. 5639. McKenzie, David and Christopher Woodru¤ (2008) “Experimental Evidence on Returns to , Capital and Access to Finance in Mexico” World Bank Economic Review 22(3): 457-82 Platteau, Jean-Philippe (2000) Institutions, Social Norms and Economic Development. Har- wood Academic Publishers: Amsterdam. , Ravallion, Martin (2009) “Should the Randomistas Rule?” The Economists Voice 6(2). Schündeln, Mathias (2006) “Modeling Firm Dynamics to Identify the Cost of Financing , Constraints in Ghanaian Manufacturing” Mimeo. Harvard University. Somville, Vincent (2011). "Daily Collectors, Public Good Provision, and Private Consump- tion: Theory and Evidence from Urban Benin", Department of Economics, Namur University (mimeograph) Spears, Dean (2009) “Dosas by the Dozen: Theory and Evidence of Present Bias in Microen- , trepreneurs” Institute for Financial Management and Research Working Paper no. 27 Udry, Christopher (1996). "Gender, Agricultural Production and the Theory of the House- hold", Journal of Political Economy, 104(5): 1010-1046, October , Udry, Christopher and Santosh Anagol (2006) “The Return to Capital in Ghana” American Economic Review Papers and Proceedings 96(2): 388-93. 41 Appendix 1: Steady state …rm size with time inconsistency Let denote the one period-ahead discount rate:17 1 : 1+ Let k s be the level of capital that satis…es: 0 (k s ; ) = : Is k s the steady state capital of a time inconsistent entrepreneur? It depends on whether the decision maker is sophisticate or myopic, that is, whether he or she realizes that future decisions were taken according to (3) or not. Suppose the decision maker is sophisticate and sets kt = k s . Is this a steady state? The Euler equation between t and t + 1 is: 0 1 u0 (ct ) t 1+ (kt+1 ; ) = (11) u0 (cP ) t+1 t+1 where cP denotes the household’ predicted future decision about ct+1 . If the household is t+1 s myopic, cP is expected to coincide with the decision made at time t, i.e., as given by (2). If t+1 the household is sophisticate, it is the correctly anticipated decision taken at time t + 1 as given by the solution to (3). First note that if cP = ct , then u0 (ct ) = u0 (cP ) and setting kt = k s satis…es the t+1 t t+1 t+1 above Euler equation. If the entrepreneur is sophisticate and sets kt = k s , she realizes that the decision problem and Euler equation at t + 1 will be identical to those at t. Hence she correctly anticipates that cP = ct . It follows that k s is the steady state level of …rm capital for t+1 a sophisticate entrepreneur. If the entrepreneur is myopic and sets kt = k s , she incorrectly believes that she will be more patient next period. Let cM denote the consumption level she sets for t + 1, not realizing that t+1 at t + 1 she will want to increase consumption beyond cM . At kt = k s the entrepreneur expects t+1 cM < ct , which implies that u0 (cP ) > u0 (ct ). Hence k s does not satisfy the Euler equation t+1 t+1 t+1 t (11) and is not a steady state. For a myopic decision maker, the steady state capital k m is such u0 (ct ) that ct = ct+1 and cM = cM . Since cM < ct+1 , it follows that t+1 t+2 t+1 0 t ut+1 (cM ) < 1, which in turn t+1 implies that k s < k m and 0 (k m ; ) > : 17 It is clear that > . If, as is likely, > r, the household will never want to set w > 0. So we ignore savings here. 42 Appendix 2: Robustness to Attrition Attrition in the panel comes from …rms closing, refusing to answer the survey, or answering the survey but not providing pro…ts data. Appendix Table A1 provides attrition rates per round for the experimental sample. Recall that we eliminated …rms which closed or refused to answer the round 2 survey before undertaking the randomization. As a result, attrition from the survey is zero by de…nition for the experimental group in rounds 1 and 2, although there is some item non-response on pro…ts. Over the course of our experiment we observe 6 percent of the …rms closing, with this rate not varying between treatment and control. We were able to keep attrition fairly low over waves 3 through 6 of the survey, and exerted additional e¤ort in round 6 to try and track and induce responses by …rms that had attrited in previous waves. As a result, only 8 percent of the sample is not present in wave 6, although 11 percent do not report pro…ts data. Nevertheless, overall attrition rates are higher for the control group than either treatment group, likely re‡ecting either an implicit obligation felt by those receiving grants to continue in t the survey, or discouragement of those who weren’ randomly selected for the grants. Whilst statistically signi…cant, the di¤erence in attrition magnitudes are not that large, which should limit the impact of this di¤erential attrition on our results. To examine how robust our results are to attrition, we use the bounding approach of Lee (2009) to construct upper and lower bounds for the treatment e¤ect. The key identifying assump- tion for implementing these bounds is a monotonicity assumption that treatment assignment a¤ects sample selection only in one direction. In our context, this requires assuming that there are some …rms who would have attrited if they had not been assigned to treatment, but that no …rm attrits because of getting assigned to treatment. This seems plausible in our context. We then construct the bounds by trimming either the top or the bottom of the distribution of pro…ts for the treatment groups by the relative di¤erence in attrition rates between treatment and control. This is done on a wave by wave basis, and involves trimming up to 6 percent from the top or bottom of the distribution of the treatment group. Table A2 shows the results of estimating these Lee bounds. Columns 1 and 2 repeat the main trimmed estimates from Table 3 for comparison. These lie between the bounds estimated in columns 3 and 4 using OLS, and in columns 5 and 6 using …xed e¤ects. We see that our parameter estimates are much closer to the upper bounds than the lower bounds, which re‡ects the skewed distribution of pro…ts. The lower bounds occur only if it is the most pro…table control …rms that attrit. However, 43 a panel regression predicting attrition in the control group (in the form of missing pro…ts) as s s a function of the previous period’ pro…ts …nds that having the previous period’ pro…ts in the top 10 percent or in the bottom 10 percent, or below the median has no signi…cant e¤ect on attrition. Similarly, we …rms which experience large changes in pro…ts over two waves are no more likely to attrit in the subsequent wave. As a result, it seems attrition in the control group is not associated with previous levels or previous changes in pro…ts. Given this, it seems reasonable to assume that pro…ts are either missing at random, or missing in …rms which su¤er negative shocks that cause the …rm to shut down or the owner to be sick in the survey period. That is, there seems reason to believe either the panel estimates in columns (1) or (2), or the upper bound estimates which are based on the least successful control …rms attriting. There seems to be no evidence to support the most successful control …rms attriting, which is what the lower bound estimates assume. We therefore conclude the main results do not seem to be driven by attrition. Appendix 3: Is it reasonable to pool e¤ects over time? To test for pooling of treatment e¤ects we allow the coe¢ cients on treatment in equation (1) to vary with time since treatment. In doing this, one should note that we only observe e¤ects 12 months after treatment for the …rms treated after round 2, which is half of the treated sample. In contrast, we observe e¤ects at 3 months and 6 months for the entire treated sample, and e¤ects at 9 months for almost all the sample (excepting the 18 …rms treated after round 4). Appendix Table A3 then shows the results. We cannot reject that the impact of treatment does not vary with time since treatment for the pooled sample, and for the male sample, or for the female sample using OLS. For the female sample using …xed e¤ects, the p-value for equality of in-kind treatment e¤ects over time is 0.057, o¤ering some suggestion that the impact is greater with more time since treatment. 44 Figure 1: Post-treatment CDFs of Monthly Profits for Males by Treatment Group 1 .8 Cumulative Distribution .2 .4 0.6 0 500 1000 1500 Monthly Profits in Waves 5 and 6 Control Cash Equipment Figure 2: Post-treatment CDFs of Monthly Profits for Females by Treatment Group 1 .8 Cumulative Distribution .2 .4 0 .6 0 500 1000 1500 Monthly Real Profits in Waves 5 and 6 Control Cash Equipment   Figure 3: Post-treatment CDFs of Capital Stock for Males by Treatment Group 1 .8 Cumulative Distribution .2 .4 0 .6 0 1000 2000 3000 4000 5000 Capital Stock in Waves 5 and 6 (Cedi) Control Cash Equipment Figure 4: Post-treatment CDFs of Capital Stock for Females by Treatment Group 1 .8 .6 .4 .2 0 0 1000 2000 3000 4000 5000 Capital Stock in Waves 5 and 6 Control Cash Equipment   Figure 5. Evolution of Capital Stock After Treatment 250 Cash - High Control 200 Equipment - High Control Cash- Low Control 150 Equipment - Low Control Treatment Effe on Capital Stock 100 ect 50 0 1 2 3 4 -50 -100 Quarters Since Treatment Table 1: Characteristics of Microenterprises and Verification of Randomization Full Sample Trimmed Sample N Control Cash In‐kind N Control Cash In‐kind Variables Using to Stratify or Match Monthly profits in January 2009  781 128 132 131 753 103 99 115 Female 793 0.60 0.60 0.61 765 0.62 0.62 0.62 High Capture 793 0.58 0.58 0.57 765 0.58 0.58 0.57 High Baseline Capital Stock 793 0.49 0.49 0.49 765 0.48 0.48 0.48 Male in Male dominated industry 793 0.18 0.19 0.18 765 0.18 0.18 0.18 Male in Mixed industry 793 0.21 0.21 0.21 765 0.20 0.20 0.20 Female in Female dominated industry 793 0.29 0.29 0.29 765 0.30 0.29 0.30 Female in Mixed industry 793 0.31 0.31 0.31 765 0.32 0.32 0.32 Other Variables Monthly profits in October/November 2008 729 124 133 104 704 93 129 99 Monthly sales in January 2009 790 724 463 630 762 412 402 595 Number of hours worked in last week 785 58.82 60.55 57.13 757 59.03 60.64 56.64 Total Capital Stock in January 2009 784 468 454 418 757 446 438 410 Inventories at end of January 2009 791 258 213 201 763 239 203 198 Uses a Susu Collector 791 0.49 0.46 0.49 763 0.49 0.46 0.51 Business operated out of home 793 0.76 0.78 0.82 765 0.77 0.78 0.83 Age of Firm 788 7.87 7.13 7.22 761 7.88 7.11 7.14 Ever had bank or microfinance loan 793 0.11 0.10 0.07 765 0.10 0.09 0.07 Business has a tax number 786 0.15 0.14 0.13 758 0.14 0.14 0.13 Owner is Married 791 0.65 0.64 0.67 763 0.65 0.63 0.68 Owner's Years of Education 775 8.87 8.75 9.05 749 8.81 8.70 9.00 Owner's Digitspan Recall 768 5.11 5.07 5.03 740 5.07 5.10 4.99 Owner is Akan Speaker 793 0.45 0.41 0.43 765 0.46 0.41 0.43 Owner is Ga/Dangme Speaker 793 0.28 0.27 0.31 765 0.29 0.27 0.32 Owner's Age 791 36.39 35.43 35.74 763 36.36 35.37 35.79 Note: Trimmed Sample eliminates matched groups in which baseline profits for at least one firm  in group exceed 1500 cedis per month The only differences between groups which are statistically significant at conventional levels are January 2009 sales  and October/November 2009 profits in the trimmed sample. Table 2: Correlates of Baseline Capital Stock (1) (2) (3) Hyperbolic Discounter ‐196.1** (82.13) Low Discount Rate  94.53 (73.46) Lack of self‐control index ‐60.34* (32.48) Feels pressure to share in household ‐201.3** (87.38) Digitspan Recall 34.02* 35.51* 44.34** (18.24) (18.32) (18.10) Raven test score 13.20 15.85 21.77 (13.49) (13.51) (13.32) Education Years 13.95 13.42 6.674 (11.15) (11.19) (11.09) Female ‐193.9** ‐198.5*** ‐137.8* (75.45) (75.87) (75.17) Age 13.63*** 14.02*** 12.57*** (4.162) (4.179) (4.228) Father's schooling 1.880 2.144 0.168 (7.408) (7.445) (7.283) Mother's schooling 4.200 3.842 2.374 (8.064) (8.106) (8.001) Constant ‐244.5 ‐263.3 ‐207.3 (220.8) (220.9) (220.0) Observations 652 650 621 R‐squared 0.053 0.049 0.043 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 3: Main Treatment Effects Dependent Variable: Real Monthly Profits (Cedi) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) OLS OLS FE FE OLS OLS FE FE OLS OLS Cash Treatment 14.50* 9.59 3.96 0.48 (8.68) (7.32) (13.89) (8.23) In‐kind Treatment 38.60*** 36.75*** 43.23*** 30.87*** (11.21) (10.67) (12.31) (10.73) Cash Treatment*Female 5.21 5.17 1.22 ‐2.30 5.74 5.59 (8.47) (8.54) (9.35) (8.77) (11.57) (11.62) In‐kind Treatment*Female 35.75** 37.65** 35.61*** 32.87** 47.35** 49.92** (14.94) (14.94) (13.56) (13.21) (21.35) (21.44) Cash Treatment*Male 28.99 16.81 8.74 5.13 44.79** 34.17** (17.68) (13.25) (31.58) (16.10) (19.42) (15.51) In‐kind Treatment*Male 43.38** 35.45** 55.15** 27.83 60.33*** 50.61*** (16.80) (14.04) (23.06) (18.15) (19.76) (17.66) Constant 119.69*** 102.19*** 120.34*** 103.05*** 119.70*** 102.20*** 120.37*** 103.05*** 99.47*** 94.92*** (8.84) (4.40) (7.37) (3.71) (8.85) (4.39) (7.38) (3.70) (5.95) (5.50) Baseline trimming No Yes No Yes No Yes No Yes No Yes Waves All All All All All All All All 5 and 6 5 and 6 Observations 4354 4203 4354 4203 4354 4203 4354 4203 1392 1344 Number of firms 792 764 792 764 792 764 792 764 736 710 P‐values for testing:   Cash = In‐kind 0.0668 0.0306 0.0128 0.0156   Cash = In‐kind for Females 0.0725 0.0565 0.0205 0.0187 0.0736 0.058   Cash = In‐kind for Males 0.4873 0.2998 0.1486 0.3051 0.5164 0.4207   Cash Male = Cash Female 0.2254 0.4604 0.8196 0.6854 0.0845 0.1406   In‐kind Male = In‐kind Female 0.7346 0.9145 0.4653 0.8224 0.6555 0.9804 Notes: All estimation includes wave effects, which vary by gender in columns 5 on. Standard errors in parentheses, clustered at the firm level.  Trimmed specifications trim out matched quadruplets which have at least one firm with profits above 1500 cedis per month in wave 1 or  2 OLS estimation includes dummies for the matched quadruplets. *, ** and *** denote significant at the 10%, 5% and 1% levels. Table 4: Treatment Heterogeneity by Randomization Strata Dependent Variable: Real Monthly Profits (Cedi) (1) (2) (3) (4) (5) (6) (7) (8) OLS FE OLS FE OLS FE OLS FE Interaction Category A Single‐Sex Industry Low Capture Low Capital Low Profits Interaction Category B Mixed Industry High Capture High Capital High Profits Panel A:  Females Cash Treatment*Category A 9.62 ‐6.87 2.12 ‐8.53 3.13 ‐11.25 3.29 ‐8.58 (10.08) (10.57) (12.40) (13.55) (10.62) (11.75) (7.15) (9.65) Cash Treatment*Category B 1.44 1.78 7.89 4.49 8.29 8.98 6.83 6.81 (13.37) (13.47) (12.00) (11.35) (14.05) (13.06) (20.59) (17.01) In‐kind Treatment*Category A 26.37* 25.39 28.30 35.41 15.96 14.25 2.21 4.58 (14.31) (17.03) (23.00) (24.07) (10.77) (10.41) (6.97) (7.52) In‐kind Treatment*Category B 48.26* 39.77** 46.66*** 31.06** 65.06** 55.67** 96.18*** 76.53** (25.60) (19.94) (14.15) (12.50) (30.21) (26.19) (36.95) (30.69) Number of Observations 2604 2604 2604 2604 2604 2604 2604 2604 Number of Firms 474 474 474 474 474 474 474 474 P‐values for testing:   Cash Treatments equal 0.625 0.614 0.740 0.462 0.771 0.250 0.871 0.432   In‐kind Treatments equal 0.456 0.584 0.457 0.873 0.124 0.142 0.013 0.023   Cash=In‐kind 0.156 0.058 0.056 0.061 0.155 0.051 0.119 0.056 Panel B: Males Cash Treatment*Category A ‐2.82 ‐5.75 ‐0.06 10.72 0.68 ‐0.72 17.23 ‐1.50 (16.42) (21.54) (19.55) (23.92) (18.06) (20.14) (12.99) (12.76) Cash Treatment*Category B 36.60* 17.00 25.13 0.77 30.16 8.66 15.43 9.50 (20.25) (23.63) (17.36) (21.00) (18.83) (24.00) (22.96) (27.99) In‐kind Treatment*Category A 44.85** 23.47 43.56 58.33 46.55** 26.33 35.08* 32.20 (21.72) (31.46) (27.06) (35.74) (19.24) (25.52) (18.00) (23.07) In‐kind Treatment*Category B 28.55 33.69 30.49* 8.94 25.78 28.51 34.88 21.99 (18.54) (20.66) (15.76) (19.42) (20.31) (25.59) (21.57) (27.48) Observations 1599 1599 1599 1599 1599 1599 1599 1599 Number of Firms 290 290 290 290 290 290 290 290 P‐values for testing:   Cash Treatments equal 0.132 0.477 0.337 0.755 0.260 0.765 0.946 0.721   In‐kind Treatments equal 0.569 0.786 0.677 0.226 0.458 0.952 0.994 0.776   Cash=In‐kind 0.151 0.596 0.312 0.349 0.171 0.509 0.563 0.417 Notes: All estimation includes wave effects which vary by category. Standard errors in parentheses, clustered at the firm level.  Trimmed sample used. OLS estimation includes dummies for the matched quadruplets. *, ** and *** denote significant at the 10%, 5% and 1% levels. Table 5: Where do the grants go? Quarterly Truncated Made a Amount  Weekly Quarterly Health & Quarterly Total Log Capital Capital Transfer Transferred Food Clothing Education Ceremonies Quarterly Quarterly Stock Stock Out Out Spending Spending Spending Spending Spending Spending FE FE OLS OLS FE FE FE FE FE FE Panel A: Males and Females Cash Treatment*Female 82.61 49.17 0.05* 8.05** 3.81 3.38 ‐1.05 1.39 120.54*** 0.08* (72.01) (37.27) (0.03) (3.46) (2.44) (3.90) (13.42) (3.17) (45.61) (0.04) In‐kind Treatment*Female 135.34** 120.24*** 0.02 1.76 ‐0.07 ‐0.50 ‐6.08 2.33 45.36 ‐0.02 (65.55) (34.51) (0.03) (2.92) (2.60) (4.39) (13.03) (3.46) (44.36) (0.04) Cash Treatment*Male 31.36 2.21 0.03 ‐4.06 3.93 9.52* 0.98 3.27 63.94 0.03 (70.33) (61.10) (0.04) (3.93) (3.12) (5.08) (11.26) (3.92) (50.82) (0.04) In‐kind Treatment*Male 157.71 83.74 0.01 ‐6.01 ‐2.82 3.63 ‐0.85 4.36 20.95 ‐0.01 (102.12) (69.85) (0.04) (3.95) (3.42) (5.83) (23.28) (5.20) (65.12) (0.05) Number of Observations 4256 4256 2033 2203 4268 3911 3713 4286 4495 4299 Number of Firms 765 765 722 722 765 761 753 765 765 765 P‐values testing:    Cash = In‐kind Females 0.573 0.107 0.294 0.137 0.198 0.478 0.776 0.817 0.172 0.054    Cash = In‐kind Males 0.212 0.291 0.693 0.630 0.111 0.428 0.942 0.856 0.573 0.611 Panel B: Female Sub‐sample Cash Treatment*Low Profits ‐6.77 ‐6.78 0.07** 6.13** 7.26** 4.66 15.39 2.94 197.84*** 0.16** (29.67) (29.69) (0.03) (2.80) (3.32) (4.24) (18.93) (4.11) (58.16) (0.06) Cash Treatment*High Profits 238.00 145.84* 0.02 11.54 ‐2.13 8.29 ‐25.71 ‐8.05 ‐53.38 ‐0.07 (185.23) (85.70) (0.04) (8.35) (4.27) (8.08) (18.79) (5.48) (81.92) (0.06) In‐kind Treatment*Low Profits 59.17** 59.17** 0.01 ‐0.40 1.11 4.10 3.83 ‐2.38 32.92 ‐0.02 (28.45) (28.46) (0.03) (2.02) (3.93) (5.20) (18.81) (3.09) (63.98) (0.06) In‐kind Treatment*High Profits 262.60 223.24*** 0.03 5.12 ‐1.99 ‐2.43 ‐18.48 3.11 18.07 ‐0.04 (166.25) (77.66) (0.05) (6.76) (3.94) (8.01) (17.24) (7.79) (68.53) (0.06) Number of Observations 2654 2654 1260 1260 2657 2440 2323 2666 2790 2670 Number of Firms 475 475 446 446 475 475 468 475 475 475 P‐values testing:      Cash Treatments Equal 0.193 0.093 0.351 0.540 0.083 0.691 0.124 0.109 0.013 0.007      In‐kind Treatments Equal 0.228 0.048 0.769 0.435 0.578 0.494 0.382 0.513 0.874 0.827 Notes: All expenditure data truncated at the 99.5th percentile of the data. All estimation includes wave effects which vary by gender, and by category in panel B. Standard errors in parentheses, clustered at the firm level.  High and Low profits refers to groups defined on pre‐treatment profits. Trimmed sample used. OLS estimation includes dummies for the matched quadruplets. *, ** and *** denote significant at the 10%, 5% and 1% levels. Table 6: Comparison of Characteristics of High and Low Profit Women Low High Initial Profit Initial Profits Sri Lankan Men Women Women Women a  Monthly profits in January 2009      Mean 130 38 173*** 28      Median 91 37 137*** 20 Monthly sales in January 2009      Mean 502 187 822*** 87      Median 240 120 500*** 50 Total Capital Stock in January 2009       Mean 611 251 456*** 207       Median 255 102 162*** 100 Age of Owner 35.4 35.9 37.0 41.1 Age of Firm 9.1 6.0 7.4** 9.5 Ever had a formal loan 0.07 0.08 0.15** 0.23 Keeps accounts 0.45 0.31 0.44** 0.29 Years of Education 10.04 7.80 8.63** 9.44 Digitspan Recall 5.70 4.59 4.80 5.68 Chose sector as it had low capital requirements 0.17 0.40 0.32* n.a. Chose sector for profit potential 0.18 0.11 0.18** n.a. Willingness to Take Risks 5.64 4.28 4.40 6.08 Save regularly 0.71 0.62 0.73** 0.67 Household Asset index 0.29 ‐0.40 0.14*** n.a. Household has a Cellphone 0.94 0.88 0.91 0.22 Sample Size 290 296 179 190 Notes: Means shown unless indicated otherwise. Trimmed subsample used. *, **, and *** indicate high profit women statistically different from the low profit women at the 10%, 5% and 1% levels respectively. a. Figures for Sri Lanka are reported as of March 2005 Sri Lankan baseline, converted at  an approximate exchange rate of 100 Sri Lankan rupees to 1 cedi. n.a. indicates not available in Sri Lankan data. Table 7: Heterogeneity according to self‐control  Dependent variable: Real profits Interaction Category: Used a  Said they Discount Hyperbolic Lacks Says there is Can spend Household Number of Susu at Save rate above Discounter Self‐control pressure to  freely without Size Siblings Baseline regularly median share with hh spouse in Area Panel A: Pooling Treatment Effects across Gender Cash Treatment ‐5.117 ‐34.97** 13.26 6.219 2.768 1.187 ‐8.618 2.191 0.414 (11.05) (14.58) (13.92) (10.33) (8.579) (9.886) (18.29) (8.350) (8.624) In‐kind Treatment 25.24 ‐4.260 13.97 40.35*** 29.80*** 31.54** 8.972 28.79*** 25.61** (15.34) (9.341) (11.20) (14.49) (10.81) (12.49) (13.33) (10.62) (10.40) Cash Treatment * Interaction 14.54 50.95*** ‐21.97 ‐18.70 ‐16.13** ‐0.705 13.44 5.937 2.166 (16.70) (17.69) (17.15) (16.82) (8.102) (17.82) (22.54) (4.238) (3.284) In‐kind Treatment *Interaction 10.35 49.41*** 31.00 ‐38.88** ‐6.587 ‐23.61 20.68 0.451 6.469* (20.80) (17.91) (20.66) (18.09) (6.273) (20.36) (20.09) (5.022) (3.835) p‐value for testing interactions jointly zero 0.6355 0.001 0.1024 0.0708 0.0983 0.501 0.535 0.371 0.198 Observations 4,170 4,157 4,169 4,157 4,145 3,880 2,702 4,169 3,750 Number of firms 758 755 758 755 753 675 490 757 652 Panel B: Females Sub‐sample only Cash Treatment ‐8.313 ‐20.90 9.207 0.955 ‐0.608 ‐3.369 0.710 ‐3.461 0.0197 (12.94) (17.44) (15.23) (10.69) (9.100) (10.76) (19.53) (8.711) (9.261) In‐kind Treatment 24.79 ‐2.700 12.62 41.12** 31.20** 35.56** ‐4.982 30.52** 33.84** (21.96) (8.508) (11.47) (17.56) (13.04) (17.04) (12.88) (13.40) (13.80) Cash Treatment * Interaction 12.37 28.15 ‐19.78 ‐11.43 ‐11.65 16.55 2.671 10.49** 5.174** (17.51) (20.15) (18.47) (18.97) (7.963) (16.63) (24.14) (4.602) (2.092) In‐kind Treatment *Interaction 14.44 50.76** 38.26 ‐34.74 ‐3.526 ‐10.42 42.57** ‐1.392 2.144 (27.10) (20.86) (25.00) (23.17) (7.445) (23.13) (17.84) (6.528) (2.694) p‐value for testing interactions jointly zero 0.704 0.033 0.139 0.297 0.318 0.479 0.051 0.064 0.045 Observations 2,588 2,586 2,587 2,580 2,574 2,398 1,730 2,592 2,375 Number of firms 471 470 471 469 468 418 312 471 414 Notes: Results from Fixed effects estimation on trimmed sample. *, ** and *** denote significant at the 10%, 5% and 1% levels. All regressions include wave effects which vary with the interaction. Table 8. Dependent Variable: Real Profits Interaction is: Whether they are Share of assets Difference in education Difference in age Spend income Baseline profits Baseline profits married brought to marriage between spouses between spouses without consent as share of income as share of consumption (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) Female Male Female Male Female Male Female Male Female Male Female Male Female Male Cash Treatment ‐9.640 16.78 20.54 1.823 18.69 2.177 ‐0.0262 ‐3.651 0.710 ‐47.59 ‐8.692 ‐25.43 0.0859 7.697 (12.84) (24.73) (17.50) (60.36) (17.00) (30.95) (25.64) (33.45) (19.53) (47.88) (20.83) (27.64) (13.84) (21.89) In‐kind Treatment 40.89 34.99 49.01* 23.79 40.52** 20.61 43.21 26.00 ‐4.982 29.59 32.56 147.1 64.50* 24.56 (33.71) (23.63) (27.46) (63.60) (17.86) (60.32) (27.16) (28.63) (12.88) (25.65) (23.58) (104.1) (33.17) (23.21) Cash*Interaction 11.56 ‐17.11 ‐6.289 ‐14.47 0.752 1.820 ‐0.162 0.0687 2.671 55.74 8.387 34.63 ‐2.005 16.82 (17.18) (33.14) (27.52) (77.95) (2.474) (5.634) (2.339) (6.121) (24.14) (54.47) (24.37) (39.97) (17.24) (26.29) In‐kind*Interaction ‐11.91 ‐12.19 ‐16.09 ‐6.332 0.101 6.441 0.0207 ‐2.986 42.57** ‐13.18 ‐0.605 ‐135.2 ‐54.76 12.98 (35.28) (34.08) (39.86) (84.32) (3.249) (9.761) (2.550) (4.113) (17.84) (43.76) (32.46) (108.0) (34.21) (37.70) Observations 2,599 1,593 1,273 864 956 799 1,520 932 1,730 972 2,408 1,501 2,604 1,599 Number of firms 473 289 222 151 172 147 264 163 312 178 431 267 474 290 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 9: Heterogeneity according to external pressure Dependent variable: Real profits Interaction Category: Household Says there is Number of Size pressure to  Siblings share with hh in Area Panel A: Pooling Treatment Effects across Gender Cash Treatment 2.191 1.187 0.414 (8.350) (9.886) (8.624) In‐kind Treatment 28.79*** 31.54** 25.61** (10.62) (12.49) (10.40) Cash Treatment * Interaction 5.937 ‐0.705 2.166 (4.238) (17.82) (3.284) In‐kind Treatment *Interaction 0.451 ‐23.61 6.469* (5.022) (20.36) (3.835) p‐value for testing interactions jointly zero 0.371 0.501 0.198 Observations 4,169 3,880 3,750 Number of firms 757 675 652 Panel B: Females Sub‐sample only Cash Treatment ‐3.461 ‐3.369 0.0197 (8.711) (10.76) (9.261) In‐kind Treatment 30.52** 35.56** 33.84** (13.40) (17.04) (13.80) Cash Treatment * Interaction 10.49** 16.55 5.174** (4.602) (16.63) (2.092) In‐kind Treatment *Interaction ‐1.392 ‐10.42 2.144 (6.528) (23.13) (2.694) p‐value for testing interactions jointly zero 0.064 0.479 0.045 Observations 2,592 2,398 2,375 Number of firms 471 418 414 Notes: Results from Fixed effects estimation on trimmed sample. *, ** and *** denote significant at the 10%, 5% and 1% levels. All regressions include wave effects which vary with the interaction. Table 10: Is self‐control just proxying for external pressure Dependent Variable: "Lack of Self‐control" Index Males & Married Married   Females Males & Females Females  Females (1) (2) (3) (4) Female ‐0.0860 ‐0.0969 (0.0885) (0.110) Baseline profits below the median 0.0954 ‐0.00624 0.101 ‐0.0329 (0.0887) (0.113) (0.118) (0.148) Says there is pressure to share extra profits 0.0505 ‐0.0609 0.0359 ‐0.111   with other household members (0.103) (0.122) (0.138) (0.162) Baseline household Size 0.0144 0.00290 ‐0.000428 ‐0.0191 (0.0226) (0.0305) (0.0307) (0.0404) Number of Siblings in Accra/Tema 0.0239 0.0120 0.0378** 0.0250 (0.0157) (0.0208) (0.0190) (0.0238) Agrees that whenever they have money on hand, their  0.0583 0.115 ‐0.00133 0.0620    spouse or other family members always end up requesting some. (0.0989) (0.125) (0.128) (0.154) Agrees that people who do well in their business are likely to receive  0.0393 ‐0.100 0.0916 ‐0.0605   additional requests from family and friends for money to help out  (0.115) (0.146) (0.145) (0.184) Agrees that machines and equipment held in their business are a good  0.0303 0.0507 ‐0.00774 ‐0.0186    way of saving money so that others don’t take it.  (0.0906) (0.113) (0.115) (0.143) At baseline spouse had compelled them to give money that they ‐0.0699 0.220   didn't want to during last 3 months (0.158) (0.210) Can spend their income without consulting their spouse ‐0.148 ‐0.219 (0.122) (0.156) Spouse is supportive of them running a business ‐0.218 ‐0.215 (0.160) (0.204) Constant 0.00766 0.487** ‐0.0548 0.493* (0.130) (0.240) (0.157) (0.284) Number of Observations 667 427 403 262 R‐squared 0.009 0.017 0.010 0.025 P‐value for testing joint insignificance of all variables 0.581 0.897 0.682 0.705 Notes: Coefficients are from an OLS regression of an index formed as the first principal component of using a susu, saving regularly, being a hyperbolic discounter, and having above the median discount rate on the variables listed in the table. Robust standard errors in parentheses; *, **, and *** indicate significance at the 10%, 5% and 1% levels respectively. Appendix Table A1: Attrition Rates by Round All firms Control Cash In‐kind P‐value test  of equality Didn't Answer Survey    Wave 1 0 0 0 0 1    Wave 2 0 0 0 0 1    Wave 3 0.029 0.031 0.010 0.042 0.106    Wave 4 0.073 0.086 0.068 0.052 0.303    Wave 5 0.112 0.131 0.099 0.089 0.262    Wave 6 0.080 0.102 0.047 0.068 0.050    Any Wave 0.166 0.196 0.131 0.141 0.070 Missing profits data    Wave 1 0.080 0.091 0.071 0.071 0.615    Wave 2 0.016 0.013 0.025 0.010 0.477    Wave 3 0.069 0.076 0.061 0.071 0.740    Wave 4 0.098 0.123 0.076 0.071 0.064    Wave 5 0.129 0.149 0.121 0.106 0.207    Wave 6 0.114 0.141 0.086 0.086 0.059    Any Wave 0.285 0.329 0.236 0.246 0.019 Ever close business 0.064 0.073 0.063 0.047 0.463 Note: Test of equality if based on regression of attrition on treatment group with controls for stratification groups and robust standard errors. Appendix Table A2: Robustness of Treatment Effect to Lee Bounds  Dependent Variable: Real Monthly Profits (Cedis) (1) (2) (3) (4) (5) (6) OLS FE OLS OLS FE FE Cash Treatment*Female 5.167 ‐2.298 6.093 1.148 ‐1.441 ‐3.297 (8.545) (8.768) (8.767) (7.106) (8.927) (7.226) In‐kind Treatment*Female 37.65** 32.87** 40.88*** 9.378 35.34*** 11.06 (14.94) (13.21) (15.41) (7.066) (13.59) (7.661) Cash Treatment*Male 16.81 5.132 21.82 6.218 9.154 ‐5.718 (13.25) (16.10) (13.28) (11.28) (16.02) (13.87) In‐kind Treatment*Male 35.45** 27.83 37.26*** 14.71 28.11 8.421 (14.04) (18.15) (14.07) (10.14) (18.21) (14.07) Lee Bounding No No Upper Lower Upper Lower Number of Observations 4203 4203 4165 4167 4165 4167 Number of Firms 764 764 764 764 764 764 Notes: All estimation includes wave effects. Standard errors in parentheses, clustered at the firm level.  Trimmed Sample used for all columns OLS estimation includes dummies for the matched quadruplets. *, ** and *** denote significant at the 10%, 5% and 1% levels. Appendix Table A3: How does Treatment Effect Vary with Time Since Treatment? Dependent Variable: Real Monthly Profits Males and Females Pooled Males Females (1) (2) (3) (4) (5) (6) (7) (8) OLS OLS FE FE OLS FE OLS FE Cash Treatment at 3 months 14.27 9.12 5.89 1.13 5.05 ‐2.58 11.52 3.25 (10.26) (8.01) (11.23) (8.31) (15.03) (15.86) (9.15) (9.24) Cash Treatment at 6 months 7.18 6.30 ‐1.36 ‐2.75 16.11 5.90 ‐0.18 ‐8.42 (9.86) (9.16) (17.27) (10.34) (17.21) (20.45) (10.32) (10.91) Cash Treatment at 9 months 12.97 5.99 9.60 3.74 12.64 11.01 2.37 ‐0.30 (12.23) (10.96) (15.97) (11.07) (20.12) (21.36) (12.47) (11.93) Cash Treatment at 12 months 38.09*** 27.98** 17.73 17.01 57.54*** 30.41 10.01 8.82 (13.55) (12.81) (23.52) (13.42) (20.87) (25.94) (16.15) (14.69) In‐kind Treatment at 3 months 26.37** 26.65** 30.20** 18.86* 33.59 25.34 22.25* 14.81 (12.10) (11.42) (12.64) (11.36) (22.86) (24.82) (11.89) (10.06) In‐kind Treatment at 6 months 34.62*** 32.61*** 38.34*** 25.49** 19.12 9.98 41.03*** 35.16*** (11.68) (11.19) (12.75) (10.93) (15.11) (18.99) (15.44) (13.10) In‐kind Treatment at 9 months 48.33** 48.90** 54.91*** 45.24** 39.49** 36.59* 54.76* 50.66* (20.63) (19.96) (20.25) (18.50) (17.33) (19.41) (30.35) (27.33) In‐kind Treatment at 12 months 58.35*** 46.91*** 78.17*** 58.00*** 69.76* 75.71** 32.76* 47.10*** (19.42) (17.52) (19.23) (17.02) (35.62) (36.58) (17.47) (15.33) Constant 119.70*** 102.20*** 120.34*** 103.05*** 127.88*** 128.69*** 86.43*** 87.33*** (8.85) (4.40) (7.38) (3.71) (7.52) (6.47) (5.40) (4.49) Baseline trimming No Yes No Yes Yes Yes Yes Yes Number of Observations 4354 4203 4354 4203 1599 1599 2604 2604 Number of Firms 792 764 792 764 290 290 474 474 P‐value for testing constant effect:    of Cash Treatments 0.166 0.435 0.262 0.389 0.170 0.534 0.579 0.353   of In‐kind Treatments 0.492 0.577 0.121 0.163 0.458 0.249 0.189 0.057 Notes: All estimation includes wave effects. Standard errors in parentheses, clustered at the firm level.  *, ** and *** denote significant at the 10%, 5% and 1% levels. Trimmed specifications trim out matched quadruplets which have at least one firm with profits above 1500 cedis per month in wave 1 or 2. OLS estimation includes dummies for the matched quadruplets.