The Treatment Effect of Business Education on the Supply of High School Entrepreneurs in Atlanta and New Orleans

Abstract

Strengthening the pathway to entrepreneurship for high school students could be important in regions of the United States where economic mobility is low. We examine the impact of high school business education on the decision to be a self-employed entrepreneur in two southeastern urban U.S. high schools. We appeal to a potential-outcomes framework to estimate the treatment effect of having taken a business and coding/programming course in high school on actually being a self-employed entrepreneur, and planning to do so in the future. We find evidence that having taken a business course in high school increases the likelihood of actually being a self-employed entrepreneur, and on planning to be one in the future. Our results suggest that, at least in Atlanta and New Orleans, urban high school business education can be effective in increasing the supply of entrepreneurs, which could improve economic mobility in these urban regions.

JEL Classification : C14, C21, E10, I26, J01, J20, J40, M13

Keywords

supply of entrepreneurs high school business education treatment effects

Introduction

There is evidence that entrepreneurship and self-employment enhance economic and social mobility (Boudreaux, 2014; Quadrini, 1999, 2000), particularly for disadvantaged groups such as Black males (Fairlie, 2005), who, along with Black Americans in general, have some of the lowest rates of economic mobility (Chetty et al., 2020). Economic mobility is particularly low in southern U.S. states (Economic Mobility Project, 2012), where entrepreneurial potential has been found to be low relative to the United States as a whole (Low et al., 2005). Southeast U.S. states have high school drop-out rates that exceed the national average (National Center for Education Statistics, 2018). Given the positive correlation between entrepreneurship and human capital (Bishop & Brand, 2014; Block et al., 2012, 2013; Doms et al., 2010; Guo et al., 2016; Madriz et al., 2018), the high school human capital environment in southern states could condition entrepreneurial outcomes.

This article examines the impact of high school business education on the decision to be a self-employed entrepreneur.¹ If entrepreneurship is a pathway to economic mobility, strengthening the pathway to entrepreneurship for high school students could be important in regions of the United States where economic mobility is low. As part of the nation’s labor supply, high school graduates and drop-outs constitute a source of entrepreneurs, and the extent to which they choose to become self-employed entrepreneurs could plausibly be conditioned on their exposure to entrepreneurship in the high school educational curriculum. With survey data on students in urban high schools in both New Orleans and Atlanta, we appeal to a potential-outcomes framework to estimate the treatment effects of having taken a business course in high school on actually being self-employed entrepreneur, and planning to do so in the future.

Our inquiry makes at least three contributions. First, as we consider the treatment effects of high school business classes on entrepreneurial outcomes, our results add to the literature on the causal effects of entrepreneurship training/instruction on individual decisions to become an entrepreneur (Fairlie et al., 2015; Huber et al., 2014; Walter & Block, 2016), particularly as it relates to high school students (Elert et al., 2015). Second, our exploration of treatment effects also considers the effects of coding/programming classes. Evidence of a positive relationship between computer ownership and entrepreneurship (Fairlie, 2006), and the positive impact of information and communication technology development on venture creation (Lee & Lio, 2018), suggests that taking computer coding courses may have a treatment effect on entrepreneurship, as computer ownership, information and communication technology, and computer coding skills are complements. As such, our findings will inform the extent to which a high school course in computer coding/programming a so-called Science, Technology, Engineering and Mathematics (STEM) discipline can induce entrepreneurship (Henrekson & Rosenberg, 2001; Winkler et al., 2015). Finally, as we consider entrepreneurial outcomes in urban high schools that enroll students from minority groups who have relatively low rates of entrepreneurship (Bates et al., 2018), our results will inform the extent to which business education can not only induce business opportunity recognition among populations underrepresented as entrepreneurs, but also business formation (Kourilsky & Esfandiari, 1997).

The remainder of this article is organized as follows. In the “Data” section, we describe the data. The “Method” section discusses the econometric methodology that aims to estimate the causal treatment effects of high school courses in business and coding/programming on the decision to start and run a business currently, and in the future. We report treatment effect parameter estimates in the “Results” section. The “Conclusion” section concludes the article.

Data

The source of our data is a survey conducted by the University of New Orleans Urban Entrepreneurship and Policy Institute (UEPI) of 16- and 17-year-old high school students (mostly sophomores and juniors, but also some seniors) in Atlanta and New Orleans. Starting in August 2019 and ending in January 2020, the UEPI administered 1,150 surveys of 16 and 17-year-old high school students in New Orleans, Louisiana and Atlanta, Georgia, as part of a research effort funded by the John Templeton Foundation. Survey questions were designed to capture the individual sociodemographic and high school curriculum characteristics that can plausibly condition the underlying factors that affect decisions and perceptions associated with entrepreneurship. The measured responses of high school students aged 16 to 17 enables observations over a large treatment possibility window to the extent that business and coding/programming classes are not available to freshman and, and perhaps sophomores, who are typically 14 to15 years old.

A 38-question one-page survey instrument was administered to respondents. The coded responses were a mixture of numeric values measuring ratio outcome measures, and coded as binary or ordinal categorical responses. Both the binary and ordinal coded responses enable the assignment of numeric values to particular responses, which can serve as inputs into treatment effect specifications that can estimate the effect of a binary treatment on a binary outcome. All surveys were either administered in-person by UEPI research associates, or by school officials provided with the survey instrument. Each participating high school was offered a cash incentive proportional to the number of respondents that completed the survey.

Four of the UEPI questions inform the basic core of our inquiry. Survey respondents were queried on whether or not they currently run a business and whether or not they plan to run a business in the future. We view these questions as two particular measures of entrepreneurial outcomes. Respondents were identified as “running a business” if they generated a steady stream of income without the help of an employer. So, a student who frequently cut grass, babysat children, resold clothing or shoes, sold cooked meals, or participated in similar activities (or some combination of things)—no matter whether they had the appropriate businesses licenses, were incorporated, paid taxes, or were otherwise operated legally (and almost none were)—is viewed as running a business. Survey respondents were also queried as to whether or not they have ever completed at least one business class or one class on coding or programming in high school. The last two questions are viewed as “treatments” with potential causal effects on entrepreneurial outcomes. In particular, we explore the extent to which taking business and coding/programming courses possibly cause students to start a business while in high school and/or plan to start one in the future.

The UEPI questions on taking at least one business and/or coding/programming course do not explicitly identify whether they occurred before or after a respondent started a business, or formed an aspiration to run a business in the future. As such, it is possible that there is “reverse causality”—survey respondents decided to take courses in business and/or coding/programming because they were already running a business, or already aspiring to run a business in the future. Put differently, there could be selection into what we envision as the treatments—those predisposed toward, or actually running or planning to run a business in the future are plausibly more likely to take business and/or coding programming classes. Given such possibilities, our methodology applies a treatment effect estimation methodology that controls for as many observables as possible within the UEPI survey to mitigate or eliminate selection into the treatment which renders the implied causal relationship more credible. Still, as we note at the end of the article, it remains possible for all we say that an unobservable, prior interest in business which does not correlate with our observables is determining selection into treatment.

Method

Our theoretical approach for parameterizing and estimating treatment effects is the potential outcomes causal framework of Rubin (2005). As in Imbens (2004), suppose for a sample characterized by ( $Y_{i}$ , $X_{i}$ , $T_{i}$ ), where the $Y_{i}$ are outcomes for the treated and untreated states of Y(1) and Y(0), the $X_{i}$ are covariates measuring individual characteristics, and the $T_{i}$ are treatment indicators. For $M$ potential matches for treated observations, the imputed potential outcomes are ${\hat{Y}}_{i}$ (0) = $Y_{i}$ if $T_{i}$ = 0, ${\hat{Y}}_{i} (0) = \frac{1}{M} \sum_{j ε l_{m (i)}} Y_{j}$ if $T_{i}$ = 1, ${\hat{Y}}_{i} (1) = \frac{1}{M} \sum_{j ε l_{m (i)}} Y_{j}$ if $T_{i}$ = 0, and ${\hat{Y}}_{i}$ (1) = $Y_{i}$ if $T_{i}$ = 1, where $l_{m} i$ is an index $l$ for T_l ≠ T_ithat satisfies $\sum_{j | T_{j} \neq T_{i}} 1 [| | X_{j} - X_{i} | | \leq | | X_{l} - X_{i} | |]$ = m ∈ M. In general, the indicator function $l (\cdot)$ selects and matches observations in the control group that are the mth closest with respect to the distance $norm ‖ \cdot ‖$ .

In a sample of N observations with N₁ treated and N₀ controls, a matching estimator for the population average treatment effect (ATE) (Abadie et al., 2004) is:

τ^{P} = \frac{1}{N} \sum_{i = 1}^{N} [{\hat{Y}}_{i} (1) - {\hat{Y}}_{i} (0)] = A T E,

where $τ^{P}$ is the ATE for a randomly assigned member of the population. To the extent that individuals can select in or out of treatment, the effect of the treatment on the subpopulation of those actually treated is perhaps a better measure of the true causal effects of the treatment. As such, the population treatment effect for those actually treated (ATET) is:

τ_{T}^{P} = \frac{1}{N_{1}} \sum_{i : T_{i} = 1} [{\hat{Y}}_{i} (1) - {\hat{Y}}_{i} (0)] = A T E T .

If assignment to the treatment is independent of the potential outcomes, then conditional on the $X_{i}$ , $τ^{P}$ , and $τ^{T}$ , are identified. In this context, matching estimates of treatment effects can approximate or replicate those from a randomized experiment (Heckman et al., 1997; Rubin, 2007), as treated and control observations are similar in observable characteristics. If observable characteristics are correlated with unobservable characteristics, matching estimators also enable causal estimates of treatment effects, as treatment and control observations differ only in their assignment to treatment (Stuart, 2010).

We utilize a treatment effect estimator that matches on covariates to enable replication of a true randomized experiment whereby covariates between treatment and control groups that presumably don’t affect treatment outcomes, are similar (Imbens & Rubin, 2010). Relative to matching on a propensity score—a scalar indicating the probability of receiving treatment—matching on covariates can enable more precision in treatment effect estimates (Elze et al., 2017) and provides estimates that are a better approximation to a fully blocked randomized experimental design (King & Nielsen, 2019). This enables treatment effect parameter estimates that are less model dependent, and with less bias (Imai et al., 2008, 2009; Nguyen et al., 2017), especially when the matching is based on the Mahalanobis distance metric (Amusa et al., 2019).

A primary virtue of estimating treatment effects in the potential outcomes is that it enables causal interpretations of the ATE and ATET for the presumed direction of causality. This follows from the conceptualization of ${\hat{Y}}_{i} (0)$ , which is the “counterfactual” observation for being untreated, including outcomes alternative to the outcome of interest. The causal effect ${\hat{Y}}_{i} (1)$ – ${\hat{Y}}_{i} (0)$ is essentially a quasi-experimental and contrived/hypothetical manipulation (Imbens & Rubin, 2010). Both ATE and ATET estimate the causal effect of the treatment by comparing the outcome that would be observed if the manipulated treatment were applied relative to the outcome that would be observed for the same counterfactual observation—which is determined by the matching—if the manipulated treatment were not applied.

Results

The definition and coding of the UEPI survey questions and covariates which are utilized in our treatment effect specifications are reported in Table 1. From the survey, we selected those questions which permitted a binary and/or ordinal quantitative measurement. This enables the construction of matching covariates to fit naturally within the distance metric we use—the Mahalanobis distance—to impute for treatment observations the counterfactual nearest neighbor untreated observation.² Table 1 statistically summarizes and classifies the covariates as the treatment, control and outcome covariates. Table 2 reports a statistical summary of the coded numeric values of the covariates. With respect to the treatments under consideration, approximately 56% and 38% of students in our sample had taken a business and coding/programming course respectively, in high school. As for the outcomes under consideration, approximately 6% and 63% of students in our sample are currently or planning to run a business in the future respectively. Individually, the number of observations for each covariate ranges from a low of 949 observations to a high of 1,151 observations.

Table 1.

Description of Covariates.

Covariate	Deﬁnition	Survey question	Survey coding
Treatment covariates
Business class	Respondent has taken at least one business class	Have you taken at least one business class in school?	Yes = 1 No = 0
Coding/programming class	Respondent has taken at least one coding/programming class	Have you taken at least	Yes = 1
		one class on programming or writing code in school?	No = 0
Control/matching covariates
Age	Age of respondent	Respondent’s self-reported age	Numeric value of age
Grade	Grade level of respondent	Respondent’s self-reported grade level	Numeric value of grade level (range: 9–12)
Race	Respondent’s race/ethnicity	What is your race/ethnicity?	White = 1Black/African American = 2 Hispanic/Latino = 3Asian = 4Hawaiian/Paciﬁc Islander = 5 Middle Eastern = 6Native American = 7Other = 8Multi-racial = 9
Moved in life	Number of times respondent reports having moved in life	In your life, how how many times have you moved?	Numeric value of number of times moved
Moved this year	Number of times respondent reports having moved this year	How many times have you moved this year?	Numeric value of number of times moved
Grade point average	Respondent’s grade point average	What is your grade point average?	<2.0 = 1
			2.0–2.5 = 22.5–3.0 = 33.0–3.5 = 43.5–4.0 = 5>4.0 = 6
Arrested	Respondent has been arrested	Have you ever been arrested?	Yes = 1No = 0
Suspended	Respondent has been suspended from school	Have you ever been suspended from school?	Yes = 1No = 0
Bank	Respondent has a bank account	Do you have a bank account?	Yes = 1No = 0
Parent’s education	Either of respondent’s	Did either of your	Yes = 1
	Parent graduated from college	Parents graduate from college?	No = 0
Zip code	Respondent’s zip code	Respondent’s self-reported zip code	Numeric value of zip code (5 digits)
Close to	Someone close to respondent owns a business	Does anyone close to you own a business?	Yes = 1No = 2
Failure	Percentage of businesses respondent thinks fails within 2 years	What percentage of businesses do you think close within the ﬁrst 2 years?	Numeric value (0%–100%)
Community	Respondent has ideas for making community better	I have ideas for how to make things better in my community	Strongly Disagree = 1Weakly Disagree = 2Weakly Agree = 3Strongly Agree = 4
Better oﬀ	Respondent expects to be better oﬀ ﬁnancially than parents	I expect to be better oﬀ ﬁnancially than my parents	Strongly Disagree = 1Weakly Disagree = 2Weakly Agree = 3Strongly Agree = 4
Safe	Respondent likes to play it safe	I like to play it safe	Strongly Disagree = 1Weakly Disagree = 2Weakly Agree = 3Strongly Agree = 4
Trust	Respondent trusts the government to do the right thing	I trust the government to do the right thing	Strongly Disagree = 1Weakly Disagree = 2Weakly Agree = 3Strongly Agree = 4
Barriers	Respondent believes people like him/her face barriers to starting a business	People like me face higher barriers to starting a successful business	Strongly Disagree = 1Weakly Disagree = 2Weakly Agree = 3Strongly Agree = 4
Rules	Respondent often breaks rules if it beneﬁts him/her	I often break the rules if it beneﬁts me	Strongly Disagree = 1Weakly Disagree = 2Weakly Agree = 3Strongly Agree = 4
Change	Respondent will do something to change the world	I will do something that changes the world	Strongly Disagree = 1Weakly Disagree = 2Weakly Agree = 3Strongly Agree = 4
Money	Respondent’s primary goal is to earn money in career	My primary goal is to earn money in my career	Strongly Disagree = 1Weakly Disagree = 2Weakly Agree = 3Strongly Agree = 4
Male	Respondent is a male	Respondent’s self-reported gender	Numeric Male = 1Female/Other = 0
Outcome covariates
Business current	Respondent is currently running a business	Are you currently running a business?	Yes = 1No = 0
Business planned	Respondent plans to run a business in the future	I want to run my own business someday	Yes = 1No = 0

Source. Urban Entrepreneurship and Policy Institute 2019–2020 Southeastern U.S. High School Survey.

Table 2.

Covariate Summary.

Covariate	M	SD	No. of observations
Business class	0.5595	0.4967	1,151
Coding/programming class	0.3831	0.4864	1,151
Age	15.88	1.16	1,135
Grade	10.60	0.9684	1,135
Race	2.93	2.58	1,116
Moved in life	3.03	2.79	1,085
Moved this year	0.2427	0.6298	1,092
Grade point average	4.12	1.24	1,067
Arrested	0.0309	0.1732	1,131
Suspended	0.2805	0.4495	1,105
Bank	0.6530	0.4762	1,101
Parent’s education	0.7604	0.4270	1,077
Zip code	63,176.94	15,201.91	1,047
Close to	0.5859	0.4928	1,099
Failure	55.98	21.67	949
Community	1.82	0.7952	1,108
Better oﬀ	1.71	0.8609	1,098
Safe	2.09	0.9634	1,111
Trust	2.94	0.9456	1,096
Barriers	2.09	1.03	1,075
Rules	2.63	0.9729	1,099
Change	1.93	0.9145	1,100
Money	1.71	0.9208	1,107
Male	0.4657	0.4990	1,151
Business current	0.0595	0.2366	1,126
Business planned	0.6364	0.4812	1,081

Source. Urban Entrepreneurship and Policy Institute 2019–2020 Southeastern U.S. High School Survey.

Given the variance in the number of control/matching observations due to missing observations—the outcomes of interest have no missing observations—parameter estimates of the treatment effects could be subject to bias if the pattern of missing observations is not completely random. Given this possibility, Table 3 reports the results of testing whether or not the control/matching covariates are missing completely at random (MCAR), and if not, is there covariate-dependent missingness (CDM). If MCAR is rejected, but CDM is not, estimating the treatment effects as a function of the covariates will not result in biased estimates of the treatment effect. In Table 3, The respective MCAR and CDM test, of Li (2013) based upon that of Little (1988) rejects the null hypothesis that the control/matching covariates are MCAR, but cannot reject CDM.³ This suggests that estimating our treatment effect specifications with the control/matching covariates across differential sample sizes due to missingness will not result in any bias in the estimated treatment parameters.

Table 3.

Missing Completely at Random (MCAR) and Covariate-Dependent Missingness (CDM) Tests.

	MCAR test	CDM test
No. of observations	1,151	740
	χ2 = 3,011.06	χ2 = 21.22
	2	46
p-value	.0001	.9993

Note. The tests for MCAR and CDM are that of Li (2013) based upon the test proposed by Little (1988). For the covariates of interest yi, it is assumed that y ∼ N ( µ , Σ). Let y¯_o
j (p_j × 1) be the observed sample average for the jth missing value data pattern for j ∈ J, and B~oj be a p_j × q submatrix of B—the coeﬃcient vector for a vector of regressors x_i and let _µo
j and Σ_o
j be the p_j × 1 dimensional mean vector and the pj × p_j covariance matrix of the observed components of the jth missing value pattern respectively. Finally, let r _i = (r_i1, . . ., r_ip)^T be the p-dimensional binary indicator vector of whether each observation in y_i is observed, (e.g., r_ik = 1 if y_ik is observed, and zero otherwise), and let I_j ⊆ [1, 2, . . ., n] be the index set of pattern j in j−1 the sample where n_j =|I_j,| and ΣJ_j-1nj = n. Little’s χ2 test statistics with Σ ^J _j=1 pj − p degrees of freedom is for MCAR and CDM respectively are:

$d^{2} = Σ_{j = 1}^{J} n_{j} {({\bar{y}}_{o j} - μ_{o j})}^{T} Σ_{o j}^{- 1} ({\bar{y}}_{o j} - μ_{o j})$

$d^{2} = Σ_{j = 1}^{J} \sum_{i = 1}^{I} x_{i} {({\tilde{B}}_{o j} - B_{o j})}^{T} Σ_{o j}^{- 1} ({\tilde{B}}_{o j} - B_{o j}) x_{i})$

The respective MCAR and CDM null hypotheses are:

$H_{o 1} : y o, i | r i \sim N (μ_{o j}, Σ_{o j})$

$H_{o 2} : y o, i | r_{i}, x_{i} \sim N (B_{o j} x_{i,} Σ_{o j})$

If H_o1 is rejected, the y_i cannot be viewed as MCAR. If H_o2 is rejected, the y_i cannot be viewed as CDM.

To determine the treatment effects of taking high school business and coding/programming classes in the absence of selection into the treatment, Tables 4 and 5 report simple Probit parameter estimates, with the control covariates. For brevity, the estimated coefficients on the control covariates are omitted, and only the parameter on the binary treatment is reported. The ATE and ATET are estimated as the marginal effect of the treatment, for the entire sample and for those who actually received the treatment respectively.⁴

Table 4.

Simple Probit Parameter Estimates: The Treatment Eﬀects of Completing a High School Business and Coding/Programming Course on Being a Current High School Entrepreneur.

Treatment	High school business class	High school coding/ programming class
Treatment effect
ATE	.0606	−.0088
	.0012^a	.5651
No. of observations	722	722
Pseudo-R²	.167	.168
ATET	.0821	−.0082
	.0072^a	.5213
No. of observations	722	722
Pseudo-R²	.167	.168

Note. Approximate p-values are in parentheses. The simple Probit treatment eﬀects were estimated in a speciﬁcation that included the control/matching covariates and the treatment covariate as the relevant marginal eﬀect of the treatment for the entire sample (ATE) and for observations that actually received the treatment (ATET).

Significant at the .01 level.

Table 5.

Simple Probit Parameter Estimates: The Treatment Eﬀects of Completing a High School Business and Coding/Programming Course on Planning to be an Entrepreneur in the Future.

Treatment	High school business class	High school coding/ programming class
Treatment effect
ATE	.0843	−.0147
	.0082^a	.6652
No. of observations	724	724
Pseudo-R²	.184	.184
ATET	.0827	–.0047
	.0071^a	.663
No. of observations	724	722
Pseudo-R²	.184	.184

Significant at the .01 level.

The results in Table 4 reveal that for currently having a business, both the ATE and ATET are positive and statistically significant for having taken a high school business class. In the case of having taken a coding/programming class, both the ATE and the ATET are negative but statistically insignificant. For planning to be an entrepreneur in the future, the effects of the treatment in Table 5 are similar to the results in Table 4. The ATE and ATET are positive and statistically significant for having taken a high school business class. When the treatment is having taken a coding/programming class, both the ATE and the ATET are negative but statistically insignificant.

In general, the simple Probit treatment effect estimates in Tables 4 and 5 suggest that taking a business class in high school has an effect on whether or not a student is currently running a business, and planning to run a business in the future. Practically, the effects are nontrivial. The treatment for both a randomly selected student (ATE), and for those actually receiving the treatment (ATET), the likelihood of a high school student currently running a business increases from a low of approximately 6% to a high of approximately 8.4%. This suggests that high school business courses are a consequential factor that can explain the supply of high school entrepreneurs.

To the extent that there is selection into treatment, the estimated simple Probit treatment effects in Tables 4 and 5 could be biased. Moreover, the direction of causality could be from the outcome to the treatment, as the simple Probit treatment estimates do not allow a contrast between an individual potential outcome of interest relative to the individual’s potential counterfactual outcome if there were no treatment. Such a contrast is important in our data, as we do not know if the observed outcomes for currently running a business, and planning to run a business in the future, occurred prior to the treatments under consideration. It is the comparison of potential outcomes across the treated and control observations that allows a matching estimator to determine causal effects when selection into the treatment is on observables. Table 4 to 7 report nearest neighbor matching parameter estimates of treatment effects.⁵ As there is evidence that matching parameter estimates are robust when selecting between 1 and 4 matches with replacement (Imbens, 2004), the treatment effect parameter estimates are based on 4 Mahalanobis distance nearest neighbor matches with replacement.

The matching treatment effect parameter estimates in Table 6 reveal that for currently having a business, both the ATE and ATET are positive and statistically significant for having taken a high school business class. In the case of having taken a coding/programming class, both the ATE and the ATET are negative but statistically insignificant. For planning to be an entrepreneur in the future, the matching treatment effect parameter estimates in Table 7 are similar to the results in Table 6. The ATE and ATET are positive and statistically significant for having taken a high school business class. When the treatment is having taken a coding/programming class, both the ATE and the ATET are negative but statistically insignificant.

Table 6.

Matching Parameter Estimates: The Treatment Eﬀects of Completing a High School Business and Coding/Programming Course on Being a Current High School Entrepreneur.

Treatment	High school business class	High school coding/ programming class
Treatment effect
ATE	.0548	−.0198
	.0001^a	.2731
No. of observations	739	739
No. of matches (with replacement)	4	4
ATET	.0591	−.0138
	.0001^a	.5213
No. of observations	739	739
No. of matches (with replacement)	4	4

Note. Approximate p-values are in parentheses.

Significant at the .01 level.

Table 7.

Matching Parameter Estimates: The Treatment Eﬀects of Completing a High School Business and Coding/Programming Course on Planning to be an Entrepreneur in the Future.

Treatment	High school business class	High school coding/ programming class
Treatment effect
ATE	.0387	–.0514
	.0062^a	.2091
No. of observations	724	724
No. of matches (with replacement)	4	4
ATET	.0424	–.0112
	.0021^a	.7893
No. of observations	724	724
No. of matches (with replacement)	4	4

Note. Approximate p-values are in parentheses.

Significant at the .01 level.

Relative to the simple Probit treatment effect estimates in Tables 4 and 5, the matching estimates of the treatment effect are smaller in magnitude. Nonetheless, the treatment effect parameters for having taken a business class remain positive and significant. While smaller in magnitude, the effects are nontrivial suggesting that even after accounting for selection on observables, the estimated ATE and ATET suggest that taking a business class in high school increases the likelihood of a high school student currently running a business and planning to run a business in the future within a range of approximately 3.8% to approximately 5.9%.

The practical significance of the estimated treatment effects can be illustrated by considering what the estimated ATET for having taken a business class in high school implies for the population of high school students in Atlanta and New Orleans. High school enrollment data for the 2018–2019 academic year reported by the National Center for Education Statistics (NCES) indicates that Atlanta and New Orleans had 34,930 and 17,924 students, respectively.⁶ As the estimated ATET’s are the increase in probabilities which approximate percentages, this has implications for the effect of providing business courses to all high school students in the high schools of Atlanta and New Orleans.

The treatment effect parameter estimates in Tables 6 and 7 cohere with existing findings that high school business education has a beneficial effect on entrepreneurial outcomes (Elert et al., 2015; Huber et al., 2014; Sanchez, 2013). However, the estimated treatment effects do not provide any support for the hypothesis of Kourilsky and Walstad (2002) that high school courses in computer science are beneficial for entrepreneurial outcomes. For both currently running a business and planning to run a business in the future, the treatment effect parameter estimates in Tables 6 and 7 suggest that having taken a coding/programming class has no effect.

Conclusion

This article considered the impact of high school business education on the decision of high school students in Atlanta and New Orleans to operate their own businesses. Appealing to a potential-outcomes framework, we found evidence that having taken a business course in high school does increase the likelihood of actual entrepreneurship, and on planning to be an entrepreneur in the future. No such effects were found for taking a coding/programming course. Our results suggest that, at least among high school students in Atlanta and New Orleans, business education can be effective in increasing the supply of entrepreneurs, which could improve economic mobility in urban areas and cities with similar demographics as Atlanta and New Orleans.

Our findings suggest that high school business education can be an effective strategy for increasing the supply of entrepreneurs. As such, policies aimed at increasing economic and social mobility in urban areas of the United States should consider strengthening and expanding business education in high schools. Of course inducing entrepreneurship in urban areas need not be a panacea for enhancing economic and social mobility, as there is evidence that entrepreneurship may not mitigate racial or ethnic wealth inequality (Bradford, 2014; Darity et al., 2018), which could undermine individual incentives to complete high school (Kearney & Levine, 2016). However, to the extent that income and social mobility is a convergence process (Avery & Rendall, 2002; Bhattacharya & Mazumder, 2011) with a variety of determinants, our results suggest that strengthening and expanding business education courses in high schools might be included in a set of policy interventions designed to improve economic and social mobility by inducing more entrepreneurship in urban areas with population demographics similar to New Orleans and Atlanta.

There are at least two limitations of our results that merit attention. First, our treatment effect parameter estimates are identified if assignment to the treatment is only a function of observable individual characteristics. If there are unobservables that matter for having taken a business or coding/programming course (for example, a prior interest in business), our treatment effect estimates might be inaccurate. Just as it would be premature to conclude that movie-going causes a greater interest in film from the fact that those who have recently seen a movie are more likely to report an interest in film, so too would it be premature on the basis of our own study to conclude that business classes cause entrepreneurial intent. However, as Imbens (2004) shows, this need not cause bias if the unobservables are unrelated to the outcome of interest. In our analysis, it is important that some of the covariates we match plausibly capture some of the unobservable marginal costs of taking business and coding/programming courses in high school, which could matter for the decision to do so. For example, one of our matching covariates is whether respondents have someone close to them who owns a business—such as a parent. To the extent that children of parents who are entrepreneurs are more likely to be entrepreneurs (Lindquist et al., 2015), this covariate may increase or decrease the marginal cost of electing into certain high school classes (business or coding, for example). Still, in the end, nothing we have said rules out that an unobservable prior interest in business predicts both selection into business courses and higher levels of entrepreneurial intent. Such an interest is not only highly relevant to the outcome that we identify—it is partly constitutive of it. As such, further research is required to more soundly ground any claim that high school business courses cause higher rates of entrepreneurship.

Second and finally, our sample may not be representative of all urban high school students, nor reflect adequately the heterogeneity in state testing regimes, curriculum, and education mandates in Atlanta or New Orleans, or the United States generally. As such, even if we have genuinely identified treatment effects, our treatment effect parameter estimates might not generalize to high schools in different geographies, with different student populations. Thus further research is warranted to determine whether there are measurable features unique to our surveyed population that may make high business and coding/programming classes more or less effective for it, than for other populations.

Footnotes

Acknowledgements

The UEPI data utilized in this study are proprietary, and are not available in any public depository. The authors are willing to make an anonymized copy of the data utilized in this article to interested parties upon request. The authors acknowledge and appreciate the generous research support of the John Templeton Foundation.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The authors are grateful for generous research support from the John Templeton Foundation.

ORCID iD

Gregory N. Price

Notes

Author Biographies

Gregory N. Price is professor of economics and research scholar with the Urban Entrepreneurship and Policy Institute at the University of New Orleans. His scholarly work focuses on economic development, racial/ethnic inequality, and entrepreneurship.

Chris W. Surprenant is professor of ethics, strategy, and public policy, and director of the Urban Entrepreneurship and Policy Institute at the University of New Orleans. His scholarly work focuses on ethics, public policy, criminal justice reform, and entrepreneurship.

References

Abadie

Drukker

Herr

J. L.

Imbens

G. W.

(2004). Implementing matching estimators for average treatment effects in Stata. The Stata Journal, 4, 290–311.

Abadie

Imbens

(2006). Large sample properties of matching estimators for average treatment effects. Econometrica, 74, 235–267.

Amusa

Zewotir

North

(2019). Evaluation of subset matching methods: Evidence from a Monte Carlo simulation study. American Journal of Applied Sciences, 16, 92–100.

Avery

R. B.

Rendall

M. S.

(2002). Lifetime inheritances of three generations of whites and blacks. American Journal of Sociology, 107, 1300–1346.

Bates

Bradford

W. D.

Seamans

(2018). Minority entrepreneurship in twenty-first century America. Small Business Economics, 50, 415–427.

Bhattacharya

Mazumder

(2011). A nonparametric analysis of Black–White differences in intergenerational income mobility in the United States. Quantitative Economics, 2, 335–379.

Bishop

Brand

(2014). Human capital, diversity, and new firm formation. The Service Industries Journal, 34, 567–583.

Block

J. H.

Hoogerheide

Thurik

(2012). Are education and entrepreneurial income endogenous? A Bayesian analysis. Entrepreneurship Research Journal, 2, 1–27.

Block

J. H.

Hoogerheide

Thurik

(2013). Education and entrepreneurial choice: An instrumental variables analysis. International Small Business Journal, 31, 23–33.

10.

Boudreaux

C. J.

(2014). Jumping off of the Great Gatsby curve: How institutions facilitate entrepreneurship and intergenerational mobility. Journal of Institutional Economics, 10, 231–255.

11.

Bradford

W. D.

(2014). The “myth” that black entrepreneurship can reduce the gap in wealth between black and white families. Economic Development Quarterly, 28, 254–269.

12.

Chetty

Hendren

Jones

M. R.

Porter

S. R.

(2020). Race and economic opportunity in the United States: An intergenerational perspective. The Quarterly Journal of Economics, 135, 711–783.

13.

Darity

Jr. Hamilton

Paul

Aja

Price

Moore

Chiopris

(2018). What we get wrong about closing the racial wealth gap (pp. 1–67). Samuel DuBois Cook Center on Social Equity and Insight Center for Community Economic Development, Duke University.

14.

Doms

Lewis

Robb

(2010). Local labor force education, new business characteristics, and firm performance. Journal of Urban Economics, 67, 61–77.

15.

Economic Mobility Project. (2012). Economic mobility of the states. Pew Center on the States, Washington DC.

16.

Elert

Andersson

F. W.

Wennberg

(2015). The impact of entrepreneurship education in high school on long-term entrepreneurial performance. Journal of Economic Behavior & Organization, 111, 209–223.

17.

Elze

M. C.

Gregson

Baber

Williamson

Sartori

Mehran

R.. . .

Pocock

S. J.

(2017). Comparison of propensity score methods and covariate adjustment: Evaluation in 4 cardiovascular studies. Journal of the American College of Cardiology, 69, 345–357.

18.

Fairlie

R. W.

(2005). Entrepreneurship and earnings among young adults from disadvantaged families. Small Business Economics, 25, 223–236.

19.

Fairlie

R. W.

(2006). The personal computer and entrepreneurship. Management Science, 5, 187–203.

20.

Fairlie

R. W.

Karlan

Zinman

(2015). Behind the GATE experiment: Evidence on effects of and rationales for subsidized entrepreneurship training. American Economic Journal: Economic Policy, 7, 125–161.

21.

Guo

Chen

(2016). Is college education worth it? Evidence from its impacts on entrepreneurship in the United States. Journal of Small Business & Entrepreneurship, 28, 1–26.

22.

Heckman

J. J.

Ichimura

Todd

P. E.

(1997). Matching as an econometric evaluation estimator: Evidence from evaluating a job training programme. The Review of Economic Studies, 64, 605–654.

23.

Henrekson

Rosenberg

(2001). Designing efficient institutions for science-based entrepreneurship: Lesson from the US and Sweden. The Journal of Technology Transfer, 26, 207–231.

24.

Huber

L. R.

Sloof

Van Praag

(2014). The effect of early entrepreneurship education: Evidence from a field experiment. European Economic Review, 72, 76–97.

25.

Imai

King

Nall

(2009). The essential role of pair matching in cluster-randomized experiments, with application to the Mexican universal health insurance evaluation. Statistical Science, 24, 29–53.

26.

Imai

King

Stuart

E. A.

(2008). Misunderstandings between experimentalists and observationalists about causal inference. Journal of the Royal Statistical Society: Series A, 171, 481–502.

27.

Imbens

G. W.

(2004). Nonparametric estimation of average treatment effects under exogeneity: A review. Review of Economics and Statistics, 86, 4–29.

28.

Imbens

G. W.

Rubin

D. B.

(2010). Rubin causal model. In Durlauf

S. N.

Blume

L. E.

(Eds.), Microeconometrics (pp. 229–241). Palgrave Macmillan.

29.

Kearney

M. S.

Levine

P. B.

(2016). Income inequality, social mobility, and the decision to drop out of high school. Brookings Papers on Economic Activity, Spring, 2015, 333–396.

30.

King

Nielsen

(2019). Why propensity scores should not be used for matching. Political Analysis, 27, 435–454.

31.

Kourilsky

M. L.

Esfandiari

(1997). Entrepreneurship education and lower socioeconomic black youth: An empirical investigation. The Urban Review, 29, 205–215.

32.

Kourilsky

M. L.

Walstad

W. B.

(2002). The early environment and schooling experiences of high-technology entrepreneurs: Insights for entrepreneurship education. International Journal of Entrepreneurship Education, 1, 1–20.

33.

Lee

M. H.

Lio

M. C.

(2018). Can information and communication technology promote venture creation? A cross-country study using an instrument variable approach. Information Development, 34, 321–333.

34.

(2013). Little’s test of missing completely at random. The Stata Journal, 13, 795–809.

35.

Lindquist

M. J.

Sol

Van Praag

(2015). Why do entrepreneurial parents have entrepreneurial children? Journal of Labor Economics, 33, 269–296.

36.

Little

R. J.

(1988). A test of missing completely at random for multivariate data with missing values. Journal of the American Statistical Association, 83, 1198–1202.

37.

Low

Henderson

Weiler

(2005). Gauging a region's entrepreneurial potential. Economic Review-Federal Reserve Bank of Kansas City, 90, 61–89.

38.

Madriz

Leiva

J. C.

Henn

(2018). Human and social capital as drivers of entrepreneurship. Small Business International Review, 2, 29–42.

39.

National Center for Education Statistics. (2018). Trends in high school dropout and completion rates in the United States: 2018. U.S. Department of Education.

40.

Nguyen

T. L.

Collins

G. S.

Spence

Devereaux

P. J.

Daurès

J. P.

Landais

Le Manach

(2017). Comparison of the ability of double-robust estimators to correct bias in propensity score matching analysis. A Monte Carlo simulation study. Pharmacoepidemiology and Drug Safety, 26, 1513–1519.

41.

Quadrini

(1999). The importance of entrepreneurship for wealth concentration and mobility. Review of Income and Wealth, 45, 1–19.

42.

Quadrini

(2000). Entrepreneurship, saving, and social mobility. Review of Economic Dynamics, 3, 1–40.

43.

Rubin

D. B.

(1979). Using multivariate matched sampling and regression adjustment to control bias in observational studies. Journal of the American Statistical Association, 74, 318–328.

44.

Rubin

D. B.

(2005). Causal inference using potential outcomes: Design, modeling, decisions. Journal of the American Statistical Association, 100, 322–331.

45.

Rubin

D. B.

(2007). The design versus the analysis of observational studies for causal effects: Parallels with the design of randomized trials. Statistics in Medicine, 2, 20–36.

46.

Sanchez

J. C.

(2013). The impact of an entrepreneurship education program on entrepreneurial competencies and intention. Journal of Small Business Management, 51, 447–465.

47.

Stuart

E. A.

(2010). Matching methods for causal inference: A review and a look forward. Statistical Science: A Review Journal of the Institute of Mathematical Statistics, 25, 1–21.

48.

Van der Zwan

Thurik

Grilo

(2010). The entrepreneurial ladder and its determinants. Applied Economics, 42, 2183–2191.

49.

Walter

S. G.

Block

J. H.

(2016). Outcomes of entrepreneurship education: An institutional perspective. Journal of Business Venturing, 31, 216–233.

50.

Winkler

Troudt

E. E.

Schweikert

Schulman

S. A.

(2015). Infusing business and entrepreneurship education into a computer science curriculum—A case study of the STEM virtual enterprise. Journal of Business and Entrepreneurship, 27, 1–21.