Limited Impact of Business Development Programs on Profitability in the Presence of Ambiguity Aversion

This paper develops a theoretical framework to explain the limited effect of business development programs (BDPs) on entrepreneurs’ profits. We argue that a mismatch between a BDP’s narrow focus on business-promoting strategies and the wider context in which microentrepreneurs operate can limit the impact of business training. In our framework, entrepreneurs are ambiguity averse and have multiple sources of income (e.g., business and wage incomes). We show that for a sufficiently ambiguity-averse entrepreneur with multiple income sources, efficient training can result in a decline in expected profit. Notably, when the wider context (multiple income sources, ambiguity aversion) is considered, the business training impact is limited and can result in a posttraining expected profit decline. This limited impact is caused by the diversifying role that the business income plays in household finances.


INTRODUCTION
his "Banker to the Poor," argued that teaching microentrepreneurs is a waste (Yunus 1999). One cannot improve loan use since borrowers already use loans efficiently. Indeed, the fact that the poor are alive despite all the adversity they face is the best proof of their innate ability. Recent research, however, questions the scope of the "poor but rational" view. Karlan and Valdivia (2011) tested whether microentrepreneurs maximize their profit given constraints and found that "... [many microentrepreneurs'] activities prove to be generating an economic loss" (p. 510). De Mel, McKenzie, and Woodruff (2008) found that real returns on capital vary with borrowers' entrepreneurial ability, indicating that not everyone has the innate ability to do the best with what he or she has. Finally, there is no a priori reason why the "poor but rational" view would be true, as the poor lack the human capital and connections that help to build successful businesses (Banerjee 2013).
One well-recognized way to make loan use more efficient is the use of business training programs to improve microentrepreneurs' business knowledge (Prediger and Gut 2014). However, the effect of business training programs is mixed. Meta studies have shown that, while entrepreneurship programs do have a positive impact on business knowledge and practice, they have no impact on business expansion or income (Cho and Honorati 2014). To make matters worse, some studies have documented negative effects of business training on profits. Karlan and Valdivia (2011) reported that the training of female entrepreneurs in Peru led to a noticeable improvement in "bad months" and less noticeable improvement or even a decline in good months. Karlan, Knight and Udry (2012) studied the effect of training on a group of tailors in Ghana. While the business literacy of the tailors in their sample increased, their profits declined. Bruhn and Zia (2011) trained 445 clients in Bosnia and Herzegovina. They found that while basic financial knowledge improved, there was no improvement in the survival rate of business start-ups. Additionally, they find that profit declines, though insignificantly. Finally, Drexler, Fischer, and Schoar (2014) reported that only simplistic training-which consists mostly of basic rules of thumb-improves profits, while complex training does not.
An immediate explanation, which is that training programs are too complicated for microentrepreneurs to comprehend, is not supported by the evidence. Most papers report noticeable increases in business literacy after training. Giné and Mansuri (2014) specifically noted that "business training did lead to an increase in business knowledge, so lack of understanding is not the issue" (p. 19). This limited impact of training does not appear to be due to improved accounting. For example, Drexler, Fischer, and Schoar (2014) found that although there was a reduction in mistakes and more consistency across measures of how people calculate profits or sales, it did not affect the main results. Further, de Mel, McKenzie, and Woodruff (2014) compared self-reported profits to revenue and cost figures and controlled for detailed measures of accounting practices as a further robustness check. They found no significant evidence that training changes reporting. McKenzie and Woodruff (2014) argued that issues such as sample size and sample heterogeneity make it harder to detect the effect of training on profitability. In a follow-up paper, de Mel, McKenzie, and Woodruff (2014) addressed those issues using a large and homogeneous sample of 1,252 female entrepreneurs in Sri Lanka. The authors found no training's impact on the profitability of existing businesses but a positive impact on the profitability of new businesses and concluded that "the lack of impacts in most of the existing literature … may not be just due to power issues" (p. 200). They further conjectured that business training programs might be less effective than previously thought.
Finally, another explanation suggested in the empirical literature is that one reason for the weakness of BDPs and the business training they provide is related to their narrow focus on businesspromoting strategies, which ignores the wider economic context in which microfinance clients operate. First, many microfinance clients are neither interested nor "...particularly good at growing [their] businesses" (Banerjee 2013, p. 512). In a survey conducted in India, 80% of parents hoped their children would get government jobs, while 0% hoped their children would build successful businesses (Banerjee and Duflo 2011). Verrest (2013) argued that BDPs "are relevant to only a minority of entrepreneurs" due to variations in household vulnerability or a lack of business ambition (p. 58). Second, microentrepreneurs do not view their business activities solely as a way to bring in more money. Instead, they consider them as a valuable diversification tool for dealing with irregularities in income sources (Krishna 2004); as a way to reduce the household's vulnerability to negative shocks, such as job loss or illness (Ellis 2000); or as a strategy for consumption and income smoothing (Bateman andChang 2009, Banerjee andDuflo 2011).
The goal of this paper is to develop a theoretical model that shows how a mismatch between BDPs' narrow focus on business-promoting goals and the complex reality in which microentrepreneurs run their businesses can be responsible for a limited, or even negative, impact of business training on microentrepreneurs' profits.
To capture the wider context in which microentrepreneurs operate, we introduce two assumptions. Our first assumption is that the microentrepreneur has multiple sources of income. One source of income is profit from the business activity. This depends on the amount of capital invested and the state of nature. Other income sources can include farming, wage employment, temporary migration or income from informal risk-sharing arrangements. Nonbusiness income does not depend on the capital investment but can depend on the state of nature. It is well documented in the literature that households commonly rely on multiple income sources, 1 yet this assumption is rarely used in the theoretical microfinance literature, where nonbusiness incomes are typically normalized to 0. 2 As we show in this paper, disregarding multiple sources of income results in a loss of generality. The effectiveness of training differs depending on whether the microentrepreneur has one or multiple sources of income available. In particular, only in a setting with multiple income sources can the posttraining expected profit decline.
Our second assumption is that the microentrepreneur has two objectives. The first is to maximize expected income. We will refer to this as the business-oriented ambition. The second objective is to maximize the "rainy day" income or, more formally, the worst-case income. We will refer to this as the livelihoods-oriented ambition. We model the microentrepreneur's utility as a weighted average of the two objectives: business-oriented ambition (maximizing expected income) and livelihoods-oriented ambition (maximizing worst-case income). Mathematically, our setup follows the 1 For example, a survey of households in Masaka district, Uganda, showed that for an average household, 64% of its income came from farm income, 20% from business profits, and 10.6% from wages (Ellis 2000, Table 3.1). A survey of households in Mamone, a poor community in South Africa, showed that the primary income source was remittances and other transfers (63.4%), wages accounted for 9.1%, business activities for 6.3% and farming activities for 12.8% (Ellis 2000, Table 3.2). In Botswana, wage employment accounted for 21.5% of household income portfolio; crop and livestock farming for 45.8%; and other activities (beer brewing, basket weaving, carpentry) for 18.5% (Valentine 1993). 2 The focus on the business part of the household's income is a common assumption, starting from classical papers such as Besley and Coate (1995) and Ghosh and Ray (2001) and extending to more recent papers, including Chowdhury (2005); Ahlin and Waters (2014); de Quidt, Fetzer, and Ghatak (2016); and Shapiro (2015). framework of Gilboa and Schmeidler (1989) in that the microentrepreneur is modeled as an ambiguity-averse decision-maker who maximizes her minimum expected utility. 3 The microentrepreneur takes a business training course, which introduces a new business practice. The new business practice affects the profitability of the business activity only and does not affect nonbusiness income sources. The new practice is superior to the old one, and its impact is stronger in states where capital is more productive. The training's total impact on the expected profit is a sum of two effects. The first effect is the profit improvement effect. It is equal to the difference in posttraining and pretraining expected profits at the pretraining level of capital investment. It is always positive due to the superiority of the new practice. The second effect is the capital adjustment effect. It takes into account that the microentrepreneur will adjust her capital investment after the adoption of the new practice. It is equal to the difference in posttraining expected profit at posttraining and pretraining levels of capital investment. Unlike the profit improvement effect, it can be either positive or negative. If, after the training, the microentrepreneur invests more than before the training, it is positive. Otherwise, it is negative.
Having a negative capital adjustment effect undermines the efficiency of the business training and limits its impact. In our model, the business training has a stronger effect in higher states, so in order to fully reap the benefits of the new business practice, one should invest more than before the training. When the capital adjustment effect is negative, however, the microentrepreneur does the exact opposite and invests less, thereby limiting the training's impact.
Our main results are as follows. First, we show that for the capital adjustment effect to be negative it is necessary to consider the full context-multiple ambitions and multiple income sources-in which the microentrepreneur operates. If the microentrepreneur has only businessoriented ambition (i.e., she is an expected income maximizer) or her only income source is business profit, then the capital adjustment effect is positive. Second, we derive sufficient conditions for the capital adjustment effect to be negative. The livelihoods-oriented ambition (i.e., ambiguity aversion) should be sufficiently strong and the nonbusiness income sources should be sufficiently diversifying. Given the ambiguity-averse focus on the worst-case outcome, sufficiently diversifying means that nonbusiness income should be high enough in the worst state for the business activity. Finally, we show that not only can the capital adjustment effect be negative but it can also outweigh the alwayspositive profit improvement effect resulting in a posttraining decline in expected profit.
To see the intuition, consider a livelihoods-oriented entrepreneur whose objective is to maximize her worst-case income. The business training has stronger effect in states where the capital is more profitable and does not impact nonbusiness income sources. Thus, for any given capital investment, the posttraining worst-case state(s) is lower than the pretraining worst-case state(s). Since optimal capital investment is lower in states with lower capital profitability, a livelihoods-oriented microentrepreneur-instead of taking advantage of the improved profitability by investing more- 3 The literature has documented the role of ambiguity aversion on the willingness to adopt new technologies and practices, and the effect is distinct from risk aversion. Engle-Warnick, Escobal, and Laszlo (2007) documented that farmers in Peru use a traditional variety of potato with low expected yield, which, nonetheless, generates enough potatoes to feed the farmer's family. This is despite the availability of new varieties of potatoes, such as the Papa Caprio, which provide substantial yield improvement. They show that it was ambiguity aversion and not risk aversion that was responsible for the crop adoption decision. Similarly, Barham et al. (2014) examined the adoption of genetically modified corn and soya beans among 191 Midwestern grain farmers in the United States. Risk preference, measured using a coefficient of relative risk aversion, had no significant impact on adoption. Ambiguity aversion did have a significant effect and expedited the adoption of the less ambiguous genetically modified corn seeds.
invests less, which results in the negative capital adjustment effect. The role of the nonbusiness income sources is that the state with the lowest capital profitability is not necessarily the state with the lowest income. In particular, this means that a posttraining worst-case state can have strictly lower capital profitability than a pretraining worst-case state such that the capital adjustment effect is strictly negative.
Overall, the contribution of the paper is as follows. First, to the best of our knowledge, this is the first theoretical paper providing an explanation to a limited, including negative, effect of training on the profit. Second, differently from earlier theoretical literature, our model provides a holistic view of households by explicitly taking into account multiple sources of incomes, diversification needs, and the strength, or lack thereof, of the business-oriented ambition. Third, we show that the holistic modeling of the microentrepreneurial decision is crucial to understanding how efficient training can have a mixed-to-negative impact. Only when multiple income sources and multiple objectives are introduced can the training have a negative impact on expected profit.
The rest of the paper is organized as follows. Section II develops a basic setup and presents model's assumptions regarding microentrepreneurs' preferences and technology. Section III introduces training and studies its effect on the microentrepreneurs' profit. In subsection III.A we consider a simple functional form to model the training's effect. General results are presented in subsection III.B. The explicit example of posttraining profit decline is given in subsection III.C. Section IV provides the conclusion. All proofs are in the Appendix.

II. BASIC SETUP
Consider a microentrepreneur who has access to multiple sources of income that include profit from business activities as well as nonbusiness income sources. We assume that profit from business activities, ( , ), depends on the amount of capital invested, , and the state of nature, . Funds from nonbusiness income sources, ℎ( ), do not depend on but can depend on the state of nature. Given ( , ) and ℎ( ), the microentrepreneur's income is ( , ) = ( , ) + ℎ( ). States are the only source of uncertainty and are labeled by integer numbers from 1 to , ∈ {1, … , }. The probability of state is . We will use * ( ) to denote the capital level that maximizes income and profit in a given state : * ( ) = max ( , ) = max ( , ).
The microentrepreneur's utility is the weighted average of the expected income and the worstcase income, which she maximizes with respect to . Parameter is exogenously given, and we discuss its interpretation later in this section.
As follows from (1), the microentrepreneur is not an expected profit maximizer. Instead, her utility is a combination of two objectives that, following Verrest (2013), we will refer to as a businessoriented ambition and a livelihoods-oriented ambition. We define a business-oriented ambition as an expected income maximization. Since ℎ( ) does not depend on , maximizing expected income is equivalent to maximizing expected profit. We define a livelihoods-oriented ambition as maximizing the "rainy day" income, which is the income in the worst-case state, min ( , ). 4 We will use the terms "rainy day income" and "worst-case income" interchangeably throughout this paper.
Objective function (1) is a special case of the ambiguity-aversion preferences that were developed by Gilboa and Schmeidler (1989). 5 Consider an ambiguity-averse microentrepreneur who does not know the objective distribution of states, { }, and instead assumes that it belongs to a set of priors = { : ≥ (1 − ) , ≥ 0, ∑ = 1}. As in Gilboa and Schmeidler (1989), the microentrepreneur maximizes her minimum expected utility, that is, she chooses to maximize the expected income under the worst prior in : max min ∈ ( , ). (2) It is straightforward to verify that (2) is equivalent to maximizing (1): where { } is a prior from and { } is the objective distribution. 4 Entrepreneurs are categorized on being either business-or livelihoods-oriented, based on their answers to the in-depth interviews. People with business-oriented ambition are those who perceived entrepreneurship as the way out of poverty and whose dream is to have their own well-organized business. People with livelihoods-oriented ambition are those who view their entrepreneurship as a secondary income to secure their livelihoods, to create savings ("an apple for a rainy day") as well as to increase consumption or to have a hobby (Verrest 2013, p. 63, 64). 5 The theoretical literature that applies the ambiguity-aversion framework to problems in development economics is sparse. The two most commonly used ambiguity-aversion frameworks are those of Gilboa and Schmeidler (1989) and Klibanoff, Marinacci, and Mukerji (2005), which is a generalization of the Gilboa and Schmeidler (1989). For example, the Gilboa and Schmeidler (1989) framework is used by Bryan (2019) to model farmers' reaction to partial insurance products. The Klibanoff, Marinacci, and Mukerji (2015) framework is used by Elabed and Carter (2015) to model farmers' willingness to pay for microinsurance when they are ambiguity averse. 6 Let denote the worst-case state (any worst state if there are more than one) given : ( , ) ≤ ( , ) for any ≠ . For a given , the worst prior assigns the smallest probability to all states but the worst: ( ) = (1 − ) for ≠ ; and the worst state gets the remaining probability, Parameter has two mathematically equivalent interpretations. The first one, used to derive (3), is the multiple prior interpretation. It captures the ambiguity via the decision-makers' beliefs about a set or priors, , with a higher corresponding to a larger set. When = 0, the only element in is the objective distribution; when = 1, contains all the possible priors. The second interpretation is the decision weight interpretation. It captures the ambiguity via decision weight with a higher corresponding to a higher degree of ambiguity aversion. When = 0, the microentrepreneur is an expected income maximizer; when = 1, the microentrepreneur is the worst-case income maximizer. The economic difference between the two is that, in the former case, the microentrepreneur's preferences are maxmin and do not depend on ; in the latter case, they do. 7 Let * be the capital level that maximizes ( ), be the capital level that maximizes the worst-case income, = max min ( , ), and * be the capital level that maximizes expected income. Since {ℎ( )} does not depend on , * maximizes both expected income and expected profit. By definition, * = * and = * . Let ( ) denote a worst-case state (any if there is more than one), i.e., a state with the lowest ( , ). Finally, let ( ) : = ( ( ), ) denote the worst-case income for a given .
Propositions 1 and 2 characterize properties of the optimal capital choice. Proposition 1 looks at case of = 1 and shows that unless = * (1) there exist at least two worst-case states given . Proposition 2 shows * is a decreasing function of when < * . One can immediately adjust its proof to the case of > * when * is an increasing function of . Both propositions are proved in the Appendix and their proofs are straightforward applications of (A1)-(A3).
In terms of empirical evidence, on the one hand, it is well documented that many risks affecting the income sources available to poor households (e.g., own-farm production and agricultural wage labor) exhibit a high correlation (Ellis 2000, p. 60). Disastrous events, such as droughts, can 7 This discussion is based on Baillon et al. (2018) who develop a method to experimentally separate the two interpretations. Their analysis is based on a more general class of preferences axiomatized in Chateauneuf, Eichberger, and Grant (2007). Equation (3) in our model is a special case of equation (3)  adversely affect all income streams simultaneously. On the other hand, it is also well documented that, in much of the developing world, informal risk-sharing arrangements, which help poor households coping with income fluctuations, are widespread. (Ambrus, Mobius, and Szeidl 2014). Risk sharing is routinely mentioned as the most common way for households to deal with negative shocks, including death, sickness, crime and court cases, and shocks in income generating activities (De Weerdt and Dercon 2006;Mazzucato, 2009). Moreover, they are substantial enough to be successfully applied for smoothing household consumption. While household income in developing countries varies greatly, consumption is remarkably smooth (Fafchamps and Lund 2003). Also, in addition to informal risksharing arrangements, governmental programs can provide an income boost in a disastrous state. For example, in Botswana, the government drought relief program during 1985-1986 created wage employment opportunities, substituting for the decreased share of livestock in crops in income portfolios (Valentine 1993).
As an example of {ℎ( )} where ℎ(1) > min ℎ( ), consider a microentrepreneur whose only nonbusiness income sources are income from an informal risk-sharing arrangement with distant family members or other villagers and employment income. State 1 is a bad state for the microentrepreneur, so if state 1 is realized, she receives help through her risk-sharing arrangement, ℎ(1) > 0. State 2 is an intermediate state where the business is profitable enough that no help from risk sharing is needed, but there are no employment opportunities, ℎ(2) = 0. State 3 is a good state for the business, and the microentrepreneur also receives positive labor income, ℎ(3) > 0. Another example is when the only nonbusiness income source is a governmental subsidy via disaster relief programs, where state 1 is the disaster state. Then, one would have ℎ(1) > 0 and ℎ(2) = ⋯ = ℎ( ) = 0. Finally, it could be that the microentrepreneur receives net payments from a risk-sharing arrangement in her bad states, so that ℎ( ) > 0 for low states, and contributes net payments in good states, so that ℎ( ) < 0 in high states.

III. BUSINESS TRAINING
Assume that the microentrepreneur can take a training course that introduces her to a new business practice or a new technology. 8 We will use the superscript to refer to variables and functions related to the new practice. For example, ( , ) is the profit function under the new practice; * is the capital level that maximizes the microentrepreneur's posttraining objective function, and so on. We assume that the business training and the posttraining profit function satisfy the following assumptions (BT1)-(BT4).

8
The scope and level of training vary between different BDPs. In Karlan, Knight, and Udry (2012), for example, the training was on a small scale and involved targeted lessons, such as keeping time and transaction records, separating business and personal money. On the other hand, de Mel, McKenzie, and Woodruff (2014) used the global Start-and-Improve Your Business training program. This is a program with an outreach of more than 4.5 million people in more than 95 countries. It involves 3-5-day training courses and covers topics such as organization of staff, record keeping and stock control, marketing and financial planning.
Assumption (BT1) imposes the same regularity condition on ( , ) that assumption (A1) imposed on ( , ). Assumption (BT2) states that there is no cost associated with taking the training and no cost associated with implementing the new practice. In particular, it is (weakly) optimal for the microentrepreneur to undertake the training. Assumption (BT3) states that the business training affects the income from business activities only and does not affect incomes from other activities, {ℎ( )}. Finally, assumption (BT4) assumes that the new business practice is superior to the old one, and it improves profit in every state and for a sufficiently large domain of . The upper bound * ( ) is chosen so that whatever happens outside of [0, * ( )] does not affect the microentrepreneur's decision-making since * ∈ [ , * ] ⊆ [ * (1), * ( )]. It is imposed so as not to exclude a large group of possible profit improvements, such as the one used in section III.A and multiplicative improvements in section III.B. We assume that > 0 rather than ≥ 0 to allow for the possibility that ( , 0) = ( , 0).
Given our assumptions, it is straightforward to verify that the training increases the microentrepreneur's utility. Indeed, Here, the first inequality follows from (BT3) and (BT4). Indeed, ℎ ( ) = ℎ( ) by (BT3) and ( , * ) > ( , ) by (BT4) and the fact that * ≤ * ( ). Therefore, ( , * ) > ( , * ), which implies the first inequality. The second inequality is due to the fact that * is the optimal capital level for the posttraining utility function. We summarize the reasoning above in Proposition 3.

Proposition 3. If the training satisfies (BT1)-(BT4), then it strictly increases a microentrepreneurial utility.
From a welfare perspective, this indicates that the business training is valuable, as it has a positive effect on microentrepreneurs' well-being regardless of its effect on expected profit. Furthermore, a microentrepreneur will prefer the new practice and will always adopt it. Microentrepreneurs tend to follow, at least in the short run, the practices they learn during the training course, and this is well documented in the literature. Table 8 in McKenzie and Woodruff's (2014) survey summarizes the effects of training on business practice adoption with the conclusion that "almost all studies find a positive effect of business training on business practices" (p. 67).
Next, we look at the effect of training on the expected profit. Given that income from nonbusiness sources, {ℎ( )}, does not depend on , the effect of the training on expected income, ( , ), is exactly the same as on the expected profit, ( , ). The difference between the two is a constant, ℎ( ) and, in particular, if one increases (decreases) then so does another. The total effect of the training on expected profit can be disentangled into two effects: where * and * are utility-maximizing capital levels before and after training, respectively. The first effect is the profit improvement effect, and it measures how much the profit will change given the pretraining capital investment. The profit improvement assumption ensures that it is positive. The second effect is the capital adjustment effect. After the training, * is no longer optimal, and the microentrepreneur will change her capital investment to * . The capital adjustment effect measures how much the profit will change given the microentrepreneur's adjustment to the adoption of new business practices or a new technology. Unlike the profit improvement effect, the capital adjustment effect can be either positive or negative. If * > * , then it is positive. Otherwise, it is negative.
Whenever the capital adjustment effect is negative, it means that the microentrepreneur does not take advantage of improved profitability but instead adjusts her investment in such a way that it hurts her expected profit. In our model, the business training will have stronger effect in higher states; so, in order to take full advantage of it, one needs to invest more than before the training. However, because of the microentrepreneur's ambiguity aversion and availability of income from nonbusiness activities, the microentrepreneur might do the exact opposite and invest less, thereby limiting the training's effect.

A.
Effect of Business Training on Profit: A Special Case In this subsection, we consider a special case where the business training improves profitability by a fixed factor. We will consider a more general specification in the next subsection.
• (BT5-II) -improvement: Assumption (BT5-I) is imposed to ensure that (BT4) is satisfied. Assumption (BT5-II) states that the training increases the profit in state by a fixed factor, . (BT5-II) greatly simplifies the proofs due to its property that * ( ) = * ( ); however, as the next section shows, it is not necessary for the main message of the paper. Condition 1 < < ⋯ < , and its generalization in the next subsection, are imposed so that the effect of training is stronger in higher states, which are the states where capital is more profitable. An example of this would be if trainees learn how to find cheaper suppliers or become more efficient at inventory management, which will have a stronger effect during good states when sales are higher. 9 We begin the analysis of business training on expected profit with two benchmarks. The first is when the microentrepreneur has only business-oriented ambition. The second benchmark is when the microentrepreneur's only income source is the business income. As the next proposition shows, in both benchmarks, the capital adjustment effect is greater than or equal to 0. Given that the profit improvement effect is always positive, this means that the total effect is also positive.

Proposition 4. Assume that the training satisfies (BT5-I) and (BT5-II). The capital adjustment effect is nonnegative if either:
the microentrepreneur's only source of income is the business income. 9 Brooks, Donovan, and Johnson (2018) empirically studied the effect of training using two treatments. The first is a standard business training program used throughout Kenya. The second treatment is the so-called, mentor condition, where the entrepreneur is being mentored by a more experienced entrepreneur from the community. In the study, only the mentor treatment had a positive effect on the entrepreneurs' profit. As anecdotal evidence of why it worked, Brooks, Donovan, and Johnson (2018) mentioned Prudence, who was a participant of one of the mentor treatments, and who used to purchase inventory from suppliers at the entrance of a market area. After training, she started to purchase at stalls deeper into the market and only after comparing prices. Her cost dropped from 250 Kenyan shillings to 100 Kenyan shillings as a result, while she kept her sales price exactly the same. Relating this to our paper and (BT5-II), if training results in a posttraining reduction in marginal cost, it will have a stronger effect in states that are favorable to the business activity.

Corollary 1. If the conditions of Proposition 4 are satisfied, then the business training has a positive total effect on expected profit.
Intuitively, BDPs are designed to promote business-oriented strategies, such as business growth or production strengthening (Verrest 2013). When = 0, the microentrepreneur's only objective is to maximize her expected income, which is equivalent to maximizing her expected profit. The training's focus on improving profit from the business activities matches the microentrepreneur's objective, and posttraining expected profit goes up. When profit is the only source of income, there is nothing to supplement the microentrepreneur's income in the worst-case state when the profit is low. As Proposition 4 shows, the best available option to maximize the worst-case income is to increase capital investment, which will result in a higher expected profit.
Consider now the case of a microentrepreneur who has multiple income sources and whose objective differs from maximizing her expected income, ≠ 0. When ≠ 0, the microentrepreneur puts the positive weight on income in the worst-case state. When business income is the only income source, state 1 is the worst-case state for every . With multiple income sources, however, state 1 is not necessarily the worst-case state, as nonbusiness income sources (e.g., an informal risk-sharing arrangement) can supplement the low income from business activities in state 1. This changes how the microentrepreneur responds to the business training. Let ( * ) be the highest worst-case state before the training. If, because of the nonbusiness income sources, ( * ) > 1 then ( * ) < ( * ), that is, new worst-case state(s) are strictly lower. By the complementarity, lower states need lower capital investment, which puts a downward pressure on the optimal posttraining capital level and can result in the negative capital adjustment effect.
Proposition 5 formalizes the intuition above for the case of = 1. It imposes two conditions. First, * (1) < . By Proposition 1, * (1) ≠ means there is state ≠ 1 that is a worst-case state given . That is, as discussed above, nonbusiness income in state 1, ℎ(1), is high enough so that (1, * (1)) + ℎ(1) ≥ ( , * (1)) + ℎ( ) . The second condition is that < * (i.e., the microentrepreneur was underinvesting prior to the training). Proposition 5 shows that, on the one hand, * < * . To maximize expected income, one needs to invest more after the training. On the other hand, < . To maximize the worst-case income, one should invest less after the training. When = 1, the microentrepreneur maximizes her worst-case income, and so the capital adjustment effect is negative.
We conclude this section by looking at the effect of BDPs on expected profit when 0 < < 1. Proposition 4 has shown that when = 0, the capital adjustment effect is always positive. Proposition 5 has shown that if = 1 and * (1) < < * , then the capital adjustment effect is negative. A natural conjecture would be that, when * (1) < < * , the capital adjustment effect is a decreasing function of , and there exists ̂ such that it is positive when <̂ and negative when >. Similarly, one could conjecture that the total profit effect is also a decreasing function of . It turns out that neither is correct. Proposition 6 shows that the capital adjustment effect is not a decreasing function of , and the total effect is not necessarily a decreasing function of . A consequence is that, even for this simple form of profit improvement, not much can be said about the signs of the capital adjustment effect and the total effect for intermediate values of .
That is, when is sufficiently close to 0, the capital adjustment effect is an increasing function of . When is sufficiently close to 1, on the other hand, the capital adjustment effect is a decreasing function of . Therefore, it is a nonmonotone function of .
With the total effect, the situation is slightly different. First, similarly to the capital adjustment effect, it can be an increasing function of when is sufficiently close to 0. To see how, consider the limit case when = ⋯ = = 1 and = ∞. Then, the posttraining choice of capital is not sensitive to changes in , so * = * ( ) and the posttraining expected profit does not change. The pretraining expected profit, however, is a decreasing function of . Then, the total effect of the training, ( , * ) − ( , * ), is an increasing function of . Second, also similarly to the capital adjustment effect, the total effect is a decreasing function of when is sufficiently close to 1. Third, differently from the capital adjustment effect, the total effect can be a decreasing function of for every ∈ [0,1]. For example, as long as / is not too large, so that the extreme example above is not applicable, the total effect is a decreasing function of .
Proposition 6. Assume that the training satisfies (BT5-I) and (BT5-II). Assume also that < * . Then, the capital adjustment effect is an increasing function of for any sufficiently close to 0, (ii) a decreasing function of for any sufficiently close to 1.
The total effect of training on expected profit (iii) can be an increasing function of when is sufficiently close to 0; (iv) is a decreasing function of for any sufficiently close to 1; (v) there exists > 1 such that, if / < , the total effect is a decreasing function of .

B. Business Training: A General Case
In the previous subsection, we assumed a specific functional form of profit improvement, ( , ) = ( , ). In this subsection, we extend the analysis to more general functional forms of profit improvement. We will refer to them as additive and multiplicative improvements. Under an additive improvement, ( , ) > ( , ) for every and every , so it is stronger than what is required by (BT4). Also, notice that by definition, the training with an additive improvement has a stronger effect in higher states (the first inequality in [6]) and for higher levels of (the second inequality in [6]).

One example of the business training that satisfies (BTA) is
( , ) = + ( , ), where ≥ ⋯ ≥ ≥ 0. For another example, assume that the pretraining profit function is given by where ( ) ≥ 1 is a weakly increasing function of . By design, the business training is weakly more efficient in higher states because ( ) is an increasing function of . From equation (7), it follows that ( , )/ ( , ) is an increasing function of when ∈ [0, * ( )], so in that domain, the training has a stronger effect for higher . We impose (BTM-I) for the same reason as we imposed (BT5-I). The multiplicative improvement improves the profit function by multiplying it by some factor. To ensure that it is an improvement, the pretraining profit function needs to be positive. It is straightforward to verify that under multiplicative improvement * ( ) ≥ * ( ), which means that that multiplicative training satisfies (BT4).
An example of the multiplicative improvement is ( , ) = ( , ), where ≥ 1. The requirement that ( ) is a weakly increasing function of is satisfied if ≥ ⋯ ≥ .
It turns out that when the posttraining profit function satisfies either (BTA) or the two (BTM) assumptions, then the equivalent of Proposition 4 holds. When the microentrepreneur is risk neutral or has only one source of income, the training will have a positive effect on expected profit. The proof of Proposition 7, while more technical, follows the same steps as that of Proposition 4.

Proposition 7. Assume that the training satisfies either (BTA) or (BTM-I) and (BTM-II). The capital adjustment effect is nonnegative if one of the two conditions hold:
the microentrepreneur's only source of income is the business income.
Corollary 2. If conditions of Proposition 7 are satisfied, then the business training has a positive total effect on expected profit.
Finally, we derive sufficient conditions for the capital adjustment effect to be negative, which is a generalization of Proposition 5. Just like in Proposition 5, it is necessary that * (1) < . Nonbusiness income sources must provide sufficient cushion to the business income in state 1. However, in a more general setting of this subsection, an additional condition is needed. Since * ( ) is not necessarily equal to * ( ), it is possible to have a training so efficient at improving the marginal profitability of capital that it results in * ( ) > for every so that > . Then, the capital adjustment effect is positive. Thus, we need to impose a restriction on how much the training can improve capital's marginal profitability. Finally, the last condition of Proposition 8 is analogous to the assumption < ⋯ < from the previous subsection.

C. Example of Posttraining Profit Decline
As we discussed earlier, whenever the capital adjustment effect is negative, it undermines the effectiveness of the business training. Instead of taking advantage of improved profitability and expanding her business by investing more, the microentrepreneur finds it safer to invest less, thereby limiting the training's impact. In fact, the negativity of the capital adjustment effect can be large enough to outweigh the positive profit improvement effect and result in a lower posttraining expected income and expected profit.
Consider the following example. There are six states, each of which is equally likely, = 1/6. The microentrepreneur has three income sources: business profit, employment income, and income from informal risk-sharing arrangements. The microentrepreneur has the endowment of labor normalized to 1. Labor can be used for business activities and for employment. The microentrepreneur has no endowment of capital but can borrow it at rate . Capital can be used for business activities only. Income from risk-sharing arrangements does not require any inputs.
The timing is as follows. First, the microentrepreneur decides how much capital to invest into her business. Second, the state of nature, , is realized. Given , the microentrepreneur decides how to divide her labor endowment between business activities and employment. Finally, the microentrepreneur earns employment income and business profit according to her capital choice and labor allocation. She also receives income from her risk-sharing arrangements. Note that the choice of labor allocation is flexible and can be adjusted to the state of nature. The capital investment, on the other hand, cannot, as it is made before the uncertainty is realized.
If the microentrepreneur splits the labor between employment and business activities as (1 − , ) and makes capital investment, , then her profit in state is ( , ) − = (√ + √ ) − . Her employment income in state is (1 − ). We assumed that = such that wages are neither positively nor negatively correlated with the state of nature. 10 Her income from a risksharing arrangement in state is . We assume that the income from the risk-sharing arrangement 10 Depending on circumstances, employment and business incomes can be positively or negatively correlated. For example, own-farm production and agricultural wage labor will exhibit a high correlation. At the same time, Verrest (2013) showed that many households use business activities as a diversification tool against possible negative labor shocks: "They [homebased economic activities] may provide savings in the form of cash or kind as 'an apple for a rainy day' when other incomes disappear because jobs are lost or people fall ill." (p. 64). Thus, in the example section, we do not take either side and simply assume that there is no correlation between labor and business incomes. By continuity, our example will continue to hold for small values of positive and negative correlation between wages and business incomes.
has an expected payment of 0, = 0. Specifically, we assume that = > 0 when ≤ 3, and = − < 0 when ≥ 4. 11 Conditional on realized state , the microentrepreneur's objective is to maximize her total income max (√ + ) − + (1 − ) + , so that * = min{ , 1}. The objective is total income and not utility because this decision is made after the uncertainty (i.e., the state of nature) is realized so that the microentrepreneur no longer faces any risk. The state-income function ( , ), therefore, is Before the state of nature is realized, the microentrepreneur chooses capital to maximize her utility which is the same as (1).
Consider an additive improvement where the posttraining profit is given by (√ + √ ) − , where 1 ≤ … ≤ . We use the following numerical example. Let = 0.7, = 1, = 3, and the values of 's are such that = 1, = 1.02, = 1.05, = 1.15, = 1.16 and = 1.17. One can verify that when = 1, then not only is the capital adjustment effect negative, but, most importantly, the posttraining expected income and posttraining expected profit are also lower than pretraining expected income and pretraining expected profit, respectively.

11
An implicit assumption here is that informal insurance payments cannot be used for investment, which is generally consistent with the empirical evidence showing that the most common reason for accepting such payments is to meet immediate consumption needs rather than for investment purposes. Only 3.8% of all gifts and 18.4% of informal loans are used for investment purposes (Fafchamps and Lund 2003).    Figures 3 and 4 are plotted to study how robust the example above is to perturbations in parameters. In total, there are 13 parameters: { } and { } and . For the purpose of visualization, we parameterize { } and { } to make them functions of one-dimensional variables and , respectively. We set the probability of state 1 to be equal to and the probabilities of states 2 through 6 equal to (1 − )/5. When = 1/6, all states are equally likely. We define the effect in state as = 1 + ( − 1)( − 1), where { } are parameters used to build Figures 1 and 2. When = 1, then the training has no effect, = 1. When = 2, then = . Figure 3 shows that a high degree of ambiguity aversion is needed in order to have the negative capital adjustment effect: for the total effect to be negative, one needs to have a very high and an intermediate range of . When is set equal to 1, as on Figure 4, the capital adjustment effect is negative for almost all parameter values. When gets close to 1, then * → * (1) so that the condition < * in Proposition 8 is no longer satisfied, and the capital adjustment effect can be positive. The total effect is negative for low values of , and intermediate values of . When is high, then the profit improvement effect is also high and outweighs the capital adjustment effect. When is low, the capital adjustment effect is too small to outweigh the profit improvement effect. Thus, the total effect can be negative only for intermediate values of .

IV. CONCLUDING REMARKS
This paper provides a theoretical framework to understand the mixed impact of business training. We rely on a holistic view of a microentrepreneur as someone whose livelihood and goals are more complex than just being an entrepreneur. We model it using two assumptions. First, the microentrepreneur has several sources of income in addition to income from business activities. Second, the microentrepreneur has other ambitions in addition to maximizing business income.
The impact of the training varies depending on the microentrepreneurs' ambitions and the environment in which they operate. This is consistent with the observation that "BDPs have been more successful for some entrepreneurs than the others" (Verrest, 2013, p. 58). We further show that the reason behind the limited effect of business training is that BDPs' focus on growing microentrepreneurs' businesses ignores the nonbusiness aspects of microentrepreneurs' livelihoods. For microentrepreneurs who have a strong business-oriented ambition or whose only income source is profit from business activities, the training effect is always positive. When the microentrepreneur has other goals beyond profit maximization and other income sources beyond business activities, the training impact can be limited and even negative. There are several limitations of our approach that are left for future research. First, we focus on one factor that could be responsible for the posttraining expected profit decline. However, there are other factors that could lead to the same outcome (e.g., inefficiency of the training). Second, our framework cannot be used to explain the success of some of the training treatments studied in the literature. Drexler, Fischer, and Schoar (2014) showed that a simplistic rule-of-thumb training worked better than the more complex one that is commonly used by BDPs. Brooks, Donovan, and Johnson (2018) introduced the mentor treatment as an alternative to a standard business training program, where trainees were mentored by a more experienced entrepreneur from the community. They showed that the mentor treatment, and only the mentor treatment, had a positive effect on the profit. Finally, our model is static and thus cannot be used to explain some dynamic phenomena documented in the literature. For instance, Karlan, Knight, and Udry (2012) reported that Ghana tailors switched to a new practice after the training but then abandoned it 1 year after the training stopped. The static framework in our paper could not be used to explain this short-run switch and the medium-run reversal.
Let be the lowest and ′ be the highest states in ( ). One can apply the same reasoning as in (9) to show that ( , ) ≠ 0 and ( ′, ) ≠ 0. Furthermore, it cannot be the case that both state-profit functions are increasing (decreasing). Indeed, if, for example, ( , ) > 0 and ( ′, ) > 0 then < * ( ) < * ( ′). State is the smallest worst-case state given . Therefore, by the complementarity for every ″ ∈ ( ), the corresponding state-profit function is increasing at : ( ″, ) > 0. By continuity, in a sufficiently small neighborhood of , only states from can be the worst states. 13 But then, slightly above will result in a higher worst-case profit. Similarly, it cannot be the case that ( , ) < 0 and ( ′, ) < 0. Thus, ( , ) and ( ′, ) have different signs and neither is equal to 0. By the complementarity assumption, it has to be the case that ( , ) < 0 < ( ′, ), which completes the proof.
Here, the first inequality is due to the fact that the worst-case income given is less or equal than the income in state . The second inequality follows from the fact that > and ( , ) < 0. The final equality follows from the fact that is the worst-case state given . Therefore, if > then ( ) < ( ) so > cannot be optimal capital for the worst-case posttraining income. Moreover, is no longer the optimal for the worst-case posttraining income either. Since all worst-case states are less or equal than ′, it means that the corresponding stateincome functions are decreasing at . Then slightly below will give strictly higher worst-case income. It proves that < .
Next, we show that * < * . Recall that since ℎ( ) does not depend on , it means that * maximizes both expected income and expected profit. Similarly, * maximizes both posttraining expected income and posttraining expected profit. The posttraining expected profit is single peaked and concave. Therefore, to show * > * , it is sufficient to show that ( , * ) > 0, and we use here that derivatives with respect to of expected income and expected profit are equal. Let be the smallest state such that ( , * ) ≥ 0. Then ( , * ) = ( , * ) + ( , * ) > > ( , * ) + ( , * ) > 0 Here we used (BT5-II), the fact that the first sum is the summation of negative terms and the second sum is the summation of positive terms, and that * satisfies ( , * ) = 0.
Proof of Proposition 6: First, we establish two lemmas.
where the last equality follows from the fact that * ( ) maximizes ( , ). Thus, the first term in (15) converges to 0 and the second term does not depend on 's and is negative. Therefore, the limit of (15) is positive and, in particular, (15) is positive for sufficiently large but finite .
The second case, ≠ * (1). Let and be defined as in part i). Then * = for any ≥ , and ( , * ) is a constant function of for any > . Similarly * = for any ≥ , and ( , * ) is a constant function of for any > . Notice that < 1 as otherwise = * (1). One can use it to show that < . Indeed, The equality follows from the definition of , namely, (1 − ) ( , ) + (1, ) = 0. The inequality follows from the fact that < * and can be proved using the same technique as in the second part of Proposition 5. From (17), it follows that * > . In Proposition 5, we established that > . Therefore, > .
When ≥ , then * = and * = and they do not depend on . Both capital adjustment effect and total effects are constants. When ∈ [ , ], then * = and does not depend on , but * > and by Lemma 5.2 is a decreasing function of . Then both the capital adjustment effect and the total effect are strictly decreasing functions of . Combining the two results, we get that the capital adjustment effect and the total effects are weakly decreasing functions of when > .
That completes the proof.
Proof of Proposition 8: Because of the concavity of the utility function, to show that the capital adjustment effect is negative, it is sufficient to show that < and * < * .
From > * ( ), it follows that ( , ) < 0, which combined with the inequality above implies that ( , ) < 0 and, therefore, * ( ) < . In the case of a multiplicative improvement, we use the states-ranking assumption. If = then we are done. The case > is impossible. Indeed, if > then ( , * ( )) > ( , * ( )) > ( , * ( )). The first inequality is by the states-ranking assumption. The second inequality follows from the fact that * ( ) is optimal in . By definition of , ′ ( , ) < 0 and, therefore, > * ( ). One can then use complementarity to conclude that ( , ) > ( , ). But then cannot be the worst-case state given , which is a contradiction. From * ( ) < follows that > cannot be the new optimal worst-case capital: ( ) = min ( , ) ≤ ( , ) < ( , ) = min ( , ) = ( ) . The first inequality comes from the fact that the lowest income given is less or equal than the income at state . The second inequality comes from the fact that ( ,⋅) declines when > . Thus, ≤ .
Furthermore, is no longer the optimal worst-case capital either. By definition of , ( , ) ≤ ( , ) for every , including > . By the proposition's assumption, all inequalities in (6) are strict in the case of the additive improvement, and ( ) is strictly increasing in the case of the multiplicative improvement. Therefore, ( , ) > ( , ) for any > . Then, it must be the case that all worst states for are less than or equal to . 15 Take any posttraining worst-case state given , and denote it as ̂ . We know that ̂ ≤ . Then * (̂ ) < by the exact same reasoning as in the beginning of the proof of Proposition 8. Thus, all state functions that correspond to the worst-case states are strictly decreasing at . But then, slightly below will give strictly higher worst-case income, and is no longer optimal. That proves that < .
Now we need to show that * < * . In the proof of Proposition 7, it was established that * ≤ * for any < 1. When all inequalities in (6) are strict, in the case of the additive improvement, or ( ) is strictly increasing in the case of multiplicative improvement, it is trivial to show that * < * for every < 1. In particular, the inequality holds for = 0, which means that * < * . This completes the proof.