408
Views
0
CrossRef citations to date
0
Altmetric
Research Article

Is business formation driven by sentiment or fundamentals?

ORCID Icon & ORCID Icon
Received 03 Dec 2020
Accepted 01 Feb 2022
Published online: 21 Feb 2022

The creation of a new business is an act of entrepreneurship. It is also a financial undertaking. Hence it is admissible to apply the apparatus of behavioral finance to study the determinants of business formation. Our results show that aggregate US business formation, nationally and regionally, is jointly predicted by economic fundamentals and sentiment. There is evidence of both ‘pull’ and ‘push’ motives for entrepreneurship. Yet this simple structure does not survive decomposition by payroll propensity. High-payroll-propensity entrepreneurs respond primarily to pull-motive fundamentals, with sentiment accounting for a small fraction of explained variance. Low-payroll-propensity entrepreneurs, on the other hand, respond to both sentiment and fundamentals, representing both pull and push motives, with sentiment accounting for a large fraction of explained variance. Low-payroll-propensity business formation is twice as volatile as high-payroll-propensity entrepreneurship, and similarly to noise-based decision making in behavioral finance, it is substantially driven by sentiment.

1. Introduction

The determinants of business formation have been studied from a variety of perspectives within the entrepreneurship literature. A large body of work has focused on the distinction between ‘opportunity-pull’ and ‘necessity-push’ motivations for entrepreneurship (Schoar 2010; Ardagna and Lusardi 2010; Fairlie and Fossen 2019). Another strand of work focuses on the psychological traits which predispose individuals to entrepreneurship (e.g. Baron 2000; Frese and Gielnik 2014). A third strand, international comparative entrepreneurship research, exploits Global Entrepreneurship Monitor data1 to discover factors underpinning differences between countries' entrepreneurial attitudes, activity, and aspirations, as well as the ways in which these can be influenced by government policies to enhance entrepreneurial activity and economic growth (e.g. Sternberg and Wennekers 2005; Bergmann, Mueller, and Schrettle 2014).

Yet, the creation of a new business is both an act of entrepreneurship as well as a financial undertaking: finance and entrepreneurship overlap. Here we exploit this overlap. We adapt concepts, tools, and data from behavioral finance to study the determinants of business formation. This approach leads to several layers of novelty.

The data we use to gauge business formation has only recently been made public: the US Census Bureau's weekly business formation statistics (Bayard et al. 2018). Aside from appearing in working papers of the Census Bureau and the Federal Reserve, it has not yet been utilized in the academic entrepreneurship literature. Through the lens of behavioral finance, which distinguishes between fundamentals-based traders and noise-based traders, we distinguish between two different classes of entrepreneurs: those whose information processing is fundamentals-focused, and those whose information processing does not reliably distinguish between fundamentals-based signal and ‘noise’ from non-fundamental sources. The latter class is susceptible to market sentiment – widespread mood or ‘affect’. Following its success as a predictor of market returns (Lemmon and Portniaguina 2006), we introduce the Michigan Index of Consumer Sentiment (MICS) into econometric tests to gauge its power to predict future business formation. And because of the special features of the Census Bureau's business-formation statistics, we are able to estimate the effect of sentiment separately for firms that have (i) a high likelihood of supporting a payroll within six months (HBA), and (ii) a low likelihood of supporting a payroll within six months (LBA). These elements are novel to the study of entrepreneurship: the key data and variables, the distinction between fundamentals and noise, and the consequent potential relevance of sentiment.

We find that aggregate business formation is jointly determined by both economic fundamentals and sentiment. Consumer sentiment (MICS) predicts month-ahead business formation positively and significantly, accounting for a majority (62.4%) of the explained variance. Other significant predictors of month-ahead business formation are fundamental variables: the composite Purchasing Managers' Index (PMI) and the NBER recession indicator (RECES). While MICS and PMI gauge opportunity-pull motives for entrepreneurship, RECES proxies the necessity-push motive for entrepreneurship.

But this picture changes substantially once high-propensity business formation is distinguished from low-propensity business formation. Sentiment is a significant predictor of month-ahead national low-propensity business formation, and accounts for 48% of explained variation. However sentiment only accounts for 7.1% of month-ahead national high-propensity business formation. This is the first indication that LBA and HBA entrepreneurs respond differently to information. The second indication is that although the coefficients on the short-term real interest rate variable (T30R) are significant in both the LBA and HBA models, they are nevertheless of opposite algebraic sign. For high-propensity entrepreneurs, T30R has a positive effect on business formation. In contrast for low-propensity entrepreneurs T30R has a negative effect on business formation. The third indication is that whereas PMI (‘pull’ motive) and RECES (‘push’ motive) continue to be significant for LBA, they are no longer significant for HBA. Thus, not only is business formation predicted by both fundamentals and sentiment, but there are two classes of entrepreneurs, differing by their propensity to support a payroll within six months, that respond differently to sentiment and particular fundamentals. Taken together, these results support the interpretation of HBA as being fundamentals-oriented entrepreneurs who primarily respond to pull motives, while LBA are not exclusively fundamentals-oriented (i.e. also noise- and sentiment-oriented) entrepreneurs who respond to both push and pull motives.

The sequel is organized as follows. Section 2 presents a primer on the terminology and concepts of entrepreneurship used in this paper. Section 3 roots the sentiment hypothesis in a formal model of entrepreneurship. Section 4 presents the data that is used in this paper. Section 5 presents the models and results, including an array of robustness checks. Section 6 concludes.

2. Entrepreneurship

Some of the terminology and concepts used in this paper are not common knowledge among finance specialists.

New business formation is a fundamental feature of entrepreneurship, and it has been viewed as such consistently over time. Joseph Schumpeter defined entrepreneurship as ‘the assumption of risk and responsibility in designing and implementing a business strategy or starting a business’ (Schumpeter 1911). Subsequently John W. Gough explained that the term entrepreneur ‘refers to a person who undertakes and operates a new enterprise or venture, and assumes some accountability for the inherent risks’ (Gough 1969). And in the current century, Klapper et al. (2010) have defined entrepreneurship as ‘The activities of an individual or a group aimed at initiating economic activities in the formal sector under a legal form of business.’

The concepts of ‘necessity-push’ and ‘opportunity-pull’ motivation for entrepreneurship were initially shaped by two influential studies. The first, by Gilad and Levine (1986), proceeds within the situational, contingent approach to entrepreneurship which emphasizes external environmental factors over internal psychological traits. They implement empirical tests to determine ‘which particular environmental factors elicit or hinder the entrepreneurial response’ – i.e. business formation. Under their contingent approach, ‘…people are pushed into entrepreneurship by negative situational factors such as dissatisfaction with existing employment, loss of employment, and career setback.’ Meanwhile the contingent pull hypothesis hinges upon early experiences (personal, or family), or early training, which encourage the search for profitable business opportunities. And of course attractive external opportunity can also present itself serendipitously, but the process of establishing metrics for where and when such opportunities arise is not straightforward.

The second influential study, by Amit and Muller (1995), proceeds within the internal-triggers approach to entrepreneurship which emphasizes psychological traits and motives. Accordingly they frame the push hypothesis in terms of an individual's personal dissatisfaction with current employment or lack of ability or motivation to thrive in a current position. And under the internal-trigger framing of the pull hypothesis, an individual is lured by a new venture idea, due to its attractiveness and its favorable personal implications.

The internal-triggers approach to the push and pull hypotheses can be viewed as an extension of the large literature which develops the psychological approach to entrepreneurship (e.g. Baron 2000; Frese and Gielnik 2014). On the level of individual psychology and decision making, this literature has shown that the distinguishing feature between entrepreneurs and non-entrepreneurs is not primarily rooted in risk tolerance or risk aversion, but in over- or under-assessment of risk (Licht and Siegel 2006). Hence it is a question of how individuals process information – which has trait components (trait optimism or pessimism) and transitory components (mood and affect). Behavioral finance also recognizes the role of mood and affect in individual decision making, but focuses on measures which gauge widespread, correlated mood and affect across investors, i.e. sentiment. We formalize this connection between business formation and sentiment in Section 3 below.

Contemporary formulations of ‘necessity’ and ‘opportunity’ entrepreneurship are aligned with the situational, contingent approach which emphasizes external environmental factors over internal psychological traits and states (Ardagna and Lusardi 2010; Schoar 2010; Hurst and Pugsley 2011; Decker et al. 2014; Fairlie and Fossen 2019). A wide range of operational definitions are in use. Fairlie and Fossen's (2019) proposal aims to capture consensus, while at the same time remaining consistent with the standard economic model of entrepreneurship:2 ‘individuals who are initially unemployed before starting businesses are defined as “necessity” entrepreneurs, and individuals who are wage/salary workers, enrolled in school or college, or are not actively seeking a job are defined as ‘opportunity’ entrepreneurs.’ Some authors use different labels for essentially the same distinction, e.g. subsistence or remedial entrepreneurship vs. transformational entrepreneurship (Schoar 2010; Ardagna and Lusardi 2010).

In the present paper, both necessity-push and opportunity-pull variables turn out to be empirically important, as do both internal-trigger and external-trigger variants of this distinction.

3. Sentiment hypothesis

The role of sentiment in business formation follows from the standard economic model of entrepreneurship when one of its variables is augmented with a behavioral interpretation. Building upon Fairlie and Fossen (2019), an individual's non-entrepreneurial (outside option) total annual income YW is (1) YW=wεw+rA,(1) where w is their annual market wage, εw is a wage shock, r is the annual interest rate, and A is the individual's assets. If the individual switches to entrepreneurial self-employment, their annual income YSE is (2) YSE=θf(k)εe+r(Ak),(2) where θ represents entrepreneurial ability, f() is the entrepreneurial production function for annual profits using capital k, and εe represents a production shock. The last term in (2) represents the annual interest earned from investing the residual of assets not deployed in the start-up. Entrepreneurial self-employment is chosen when (3) YSE|k=k>YW.(3) For instance, a downward shock to the wage εw<1 (e.g. partial or full unemployment) may cause (3) to hold, even though it would not be the case in the absence of the wage shock. This is the necessity-push hypothesis. On the other hand, inequality (3) may hold because of an upward shock εe>1 to the production term in (2). This is the opportunity-pull hypothesis. Here the εe term aggregates shocks from various sources, including not only contemporaneous demand shocks and production shocks, but also shocks to forward-looking beliefs (expectations) about future demand and production. This is the term that avails of a behavioral interpretation.

An individual whose positive mood skews their processing of information toward perceiving greater entrepreneurial opportunity experiences an upward shock in their personal εe term ceteris paribus. The component of mood that is correlated across individuals in the economy – i.e. sentiment – is also captured in the εe terms across the economy. Positive sentiment shocks – which can be detected and measured at the aggregate level – ceteris paribus increase YSE|k=k relative to YW, increasing the probability of an individual shifting from wage employment into entrepreneurial self-employment. Conversely negative sentiment shocks ceteris paribus decrease the probability of an individual forming a new business.3

To summarize, whereas the external pull and push hypotheses are rooted in standard economic interpretations of εw and εe, the internal pull hypothesis – i.e. the sentiment hypothesis – follows from a behavioral extension of εe.

4. Data

4.1. Business formation

US business formation statistics are available from the US Census Bureau.4 These statistics are compiled from information disclosed on Form SS-4, the IRS Application for obtaining an Employer Identification Number (EIN).5 We study national and regional (Northeast, Midwest, South and West respectively) series for Business Applications (BA) and High-propensity Business Applications (HBA). The BA series are broad measures of business formation that the Census Bureau characterizes as their ‘core business applications series’.6 The HBA series are subsets of the corresponding BA series, including only those applications that have a high likelihood of becoming businesses with a payroll.7 The difference between BA and HBA is recorded as Low-propensity Business Applications (LBA).

Beginning with the US Census Bureau's weekly frequency, not-seasonally-adjusted data, we first aggregate the series up to monthly frequency (recorded as thousands of EIN applications per month), and then remove seasonality from each monthly time series through seasonal-trend decomposition using LOESS (STL) as in Cleveland et al. (1990). These seasonally-adjusted time series are denoted as BAsa, HBAsa and LBAsa, respectively. Figure 1 illustrates the raw, seasonally adjusted, and seasonal components of national BA, HBA and LBA. The sample period spans from 2006M1 to 2018M12. Regional BA, HBA and LBA series have similar seasonal-decomposition patterns. Throughout the analysis in this paper, we focus on seasonally-adjusted measures of business formation, and omit the ‘-sa’ suffix.

Figure 1. STL decomposition of national BA, HBA and LBA (thousands). (a) BA. (b) HBA and (c) LBA.

Figure 2 plots the time series of seasonally-adjusted national and regional business formations. For BA, HBA and LBA, the South region accounts for around 41% of the national total, followed by the West region (around 23–24% of the national total). Both the Northeast and Midwest regions each account for 17.5% of the national total. HBA fell during the 2007–09 period, as a result of the financial crisis. Thereafter it grew gradually throughout the remainder of the sample period. Meanwhile LBA has been growing throughout the entire sample period, without dropping during the 2007–09 period.

Figure 2. Plots of national and regional series of BA, HBA and LBA (seasonally adjusted, thousands). (a) BA. (b) HBA and (c) LBA.

Table  reports descriptive statistics for the dependent variables. It is noteworthy that the standard deviation of national LBA is more than twice that of national HBA.8 National HBA is highly right-skewed (1.36>1) while LBA is moderately right-skewed (1>0.925>0.5). Similarly, compared to national LBA, the national HBA distribution has more mass in the tails relative to the rest of the distribution (kurtosis 4.22>3.34).

Table 1. Summary statistics for monthly business formation (thousands).

4.2. Sentiment

The most straightforward and direct indicators of sentiment are provided by survey data. Shiller (1999) suggests that the Yale School of Management Stock Market Confidence Indices can reflect the attitudes of institutional investors. In the behavioral asset pricing literature, Qiu and Welch (2006) show that data from the UBS/Gallup surveys can explain equity returns, particularly small-stock returns and returns of stocks held disproportionately by retail investors. Similar findings have also been obtained by Lemmon and Portniaguina (2006) with data from both the Index of Consumer Confidence and the University of Michigan Consumer Confidence Index. Brown and Cliff (2005) find significant long-horizon explanatory power in the Investors Intelligence survey to predict asset prices.

We adopt the monthly Michigan Index of Consumer Sentiment as our proxy for sentiment. It is calculated as a linear transformation of the percentages of positive and negative responses on five telephone-survey questions. The five questions cover (i) change in perceived household financial situation over the last year, (ii) expected year-ahead change in household financial situation, (iii) expected year-ahead national financial business conditions, (iv) expected national business conditions (continuous good times vs. periods of widespread unemployment or depression) over the coming 5 years, and (v) current purchasing conditions for major household durable items.

We standardize the indicator and denote the new series as MICS.9 For MICS, the sample also covers 2006M1 to 2018M12.

4.3. Fundamental variables

Augmenting the set of fundamental variables employed by behavioral finance sentiment studies to adequately capture the real economy, we assemble fundamental variables from six categories: monetary conditions, real-economy consumption conditions, real-economy production conditions, financial-market conditions, labor-market conditions, and GDP conditions. This set of fundamental variables combines those that are suggested by the Baker and Wurgler (2007) consumption-based capital asset pricing approach as well as those suggested by the Brown and Cliff (2005) conditional asset pricing approach.

  1. monetary conditions: CPI-based monthly inflation (INFL) and 1-month US Treasury bill return (T30R);

  2. real-economy consumption conditions: real growth rate in total consumption (CONS);

  3. real-economy production conditions: real growth rate in industrial production (PROD) and composite Purchasing Manager's Composite Index (PMI);10

  4. financial-market conditions: the spread between 3-month and 1-month US treasury bill returns (SPR3), the spread between 10-year US treasury notes and 3-month US treasury bill returns (SPR10), and the default spread between yields on Moody's Baa- and Aaa-rated 10-year corporate bonds (SPRD).

  5. labor-market conditions: growth rate in employment (EMPL);

  6. GDP conditions (growth vs. contraction): NBER recession dummy (RECES).

T30R, SPR3, SPR10 and SPRD data are obtained from the US Federal Reserve and the CRSP. PMI data are compiled by IHS Markit. Data for the rest fundamental variables are obtained from Jeffrey Wurgler's online data library.

4.4. Sample characteristics

Descriptive statistics for the explanatory variables introduced in Sections 4.2 and 4.3 are summarized in Table . With the exception of MICS all variables are more heavy-tailed than the normal distribution. The recession dummy (RECES), the 1-month short rate (T30R), and three variables representing financial market conditions (SPR3, SPR10 and SPRD) are right-skewed (probably due to the implicit left-truncation for each indicator) while all other indicators are left-skewed.

Table 2. Summary statistics of explanatory variables.

Table  reports pair-wise correlations across the full set of variables. Most correlations (34 out of 55) are statistically significant. MICS is significantly correlated with 8 out of 10 fundamental indicators.

Table 3. Correlation coefficients between indicators.

5. Models and results

5.1. Business formation and consumer sentiment

In this section we test for interactions between business formation measures and consumer sentiment, through a Vector Error Correction (VEC) Model. We first confirm that the national and regional business formation measures as well as the consumer sentiment proxy are all persistent time series, each following an I(1) processes. Furthermore, Johansen Cointegration Tests show that MICS is cointegrated with all fifteen business formation measures (BA, LBA, and HBA, for the national and four regional measures).11

Equation (4) presents the VEC model to be estimated, where Yt=[BAt,MICSt] for the BA series, Yt=[HBAt,MICSt] for the HBA series, and Yt=[LBAt,MICSt] for the LBA series. ECT is an error correction term that measures the deviation of Yt from its long-run cointegration. (4) ΔYt=c+ΘECTt+i=1kΦiΔYti+ϵt(4) where (5) ECTt=BusinessFormationt+βMICSt+α(5) Here the normalized fixed-unit coefficient on BusinessFormationt in Equation (5) allows more straightforward interpretation of the impact of business formation in the cointegrated long run.

We choose optimal lag order k for the VEC models according to the Schwarz Information Criterion (SIC) a.k.a. Bayesian Information Criterion (BIC). According to this criterion, k = 2 is optimal for all BA and LBA measures. For Northeast HBA it is k = 1 that is optimal, while for national, Midwest, South and West HBA it is k = 3 that is optimal.

Table  reports estimates for the ECT term for different business-formation measures. The t-statistic is recorded in square brackets below each β estimate. The coefficients suggest that in the long-run cointegration, business formation measures are all positively correlated with consumer sentiment (i.e. MICS). Moreover, the t-statistics show that such long-run correlation is often statistically significant, with three exceptions: national, Midwest, and South HBA.

Table 4. Long-run cointegration between business formation and consumer sentiment.

We also report in Table  the estimated Θ=[θ1,θ2] coefficients from Equation (4), where θ1 and θ2 represent the short-run sensitivities of the business-formation and sentiment measures, respectively, to the deviation from long-run cointegration. The θ1 estimates are negative and in most cases significant, confirming the presence of the error correction mechanism to long-run cointegration for business-formation measures. For nationalBA, southBA and all LBA measures, we also find positive and significant estimates of θ2, in line with the presence of the error correction mechanism to long-run cointegration for MICS. Notice that this is not the case for HBA measures.

In order to demonstrate the short-run dynamics between business formation and consumer sentiment that is captured by the VEC model, we plot in Figure 3 the impulse-response curves between consumer sentiment and national business-formation measures.12 Figures generated from regional-business formation measures behave similarly and are available upon request.

Figure 3. Impulse-response curves between consumer sentiment and national business formation measures. (a) Impulse response, national BA to MICS. (b) Impulse response, MICS to national BA. (c) Impulse response, national LBA to MICS. (d) Impulse response, MICS to national LBA. (e) Impulse response, national HBA to MICS and (f) Impulse response, MICS to national HBA.

To generate the impulse response curves, a unit shock is introduced in the impulse variable, and the response period is set at 12 months. For instance, panel (a) of Figure 3 plots the response of national BA to a one-unit shock in MICS over the subsequent 12 months. Business-formation measures are responsive to shocks in consumer sentiment, while consumer sentiment does not respond much to shocks in business formation. Low-propensity business formation and high-propensity business formation respond similarly to shocks in sentiment in the first three months. However from the fourth month onward the effect recedes for HBA while further strengthening for LBA. This is a first indication – which is consistently reinforced in subsequent empirical analysis – that consumer sentiment impacts upon low-propensity business formation more heavily than upon high-propensity business formation.

5.2. Controlling for fundamentals

In Section 5.1 we find evidence that business formation and consumer sentiment are cointegrated in the long-run, and that such correlation is largely statistically significant. In this section we investigate the incremental explanatory power of consumer sentiment in addition to that of a battery of fundamental variables. We base subsequent estimation on the following multivariate linear equation: (6) BusinessFormationt=c+βiFactort1i+ϵt(6) Table  reports results from estimating Equation (6) with BA as the business formation dependent variable.13 Several findings emerge.

Table 5. Coefficients of multivariate regressions and standard errors: BA.

First, the model explains a large proportion of overall business-formation variation. Adjusted R2 ranges from 0.425 to 0.618. The variance explained is largest for the West region, and somewhat smaller for the Midwest region.

Second, consumer sentiment is a strong predictor of business formation. MICS has positive coefficient estimates in all five BA-measure models, and these coefficients are all statistically significant at the 0.1% level. We interpret these result as evidence of a opportunity-pull motivation for entrepreneurship, consistent with the behavioral interpretation of the εe term in Section 3. Aone-standard-deviation increase in MICS on average leads to 22,717 new businesses being founded in the US, of which more than half (11,901) are located in the South region.

Third, PMI and the NBER recession indicator emerge as the key fundamental variables that explain business formation.

PMI has positive and significant coefficients for national, Northeast and Midwest BA series. PMI augments the opportunity-pull motivation driving business formation, where forward-looking improvements in the outlook across manufacturing and services strengthen entrepreneurs' propensity to establish a business, again via its impact upon εe. On average a one-unit increase in PMI will lead to an additional 2191 monthly business formations across the US.

The RECES indicator has positive and significant coefficients for national and all regional measures of BA, reflecting a clear necessity-push motivation behind business formation during recession periods. On average, 28.9k more businesses are newly founded nationally in each month during recession periods than non-recession periods. The effect is most prominent in the South region, and less so in the Northeast and Midwest regions.

Last but not least, sentiment proxied by MICS is the dominant predictor of BA series' variance. More than half of adjusted R2 can be attributed to consumer sentiment, showing that the broad measure of business formation is heavily responsive to people's sentiment and only partially to entrepreneurs' response to economic conditions. Following the Relative Weight methodology developed by Johnson (2000) and elaborated by Tonidandel and LeBreton (2011), we are able to decompose the adjusted R2 into proportions attributed to each explanatory variable. The relative importance of MICS, as measured by its contribution to adjusted R2, is reported for each model in the last row of Table . For all five BA measures, the contribution of MICS exceeds 50%. The relative importance of consumer sentiment is highest in the South, where 64.7% of the explained variance is attributable to MICS.

5.3. Decomposition by payroll propensity

Here we investigate high-propensity business applications (HBA) separately from low-propensity business applications (LBA). We estimate Equation (6) on HBA and LBA series, separately for the national and each regional level, and report the results in Table . It is clear that high-propensity and low-propensity business formation are driven by different sets of factors.

Table 6. Coefficients of multivariate regressions and standard errors: LBA and HBA.

For the HBA series, consumer sentiment and the real 30-day t-bill rate stand out as key predictors. Consistent with the findings in Table , positive coefficients are present for MICS, providing evidence of pull-effect motivation for entrepreneurship. Inflation, manufacturing and services outlook, labor market conditions, and the recession indicator show predictive power inconsistently across the regions.14

For the LBA series, consumer sentiment, the real 30-day t-bill rate, PMI, and the NBER recession dummy simultaneously show significant predictive power. Again positive coefficients are present for MICS, providing evidence of pull-effect motivation for entrepreneurship. Just as is the case in Table , PMI and RECES predict LBA with positive coefficients, supporting the interpretation that there is a pull effect from PMI and a push effect from RECES.

Interestingly, the coefficient on the 30-day short rate is negative for LBA but positive for HBA. We interpret this finding as evidence that HBA and LBA capture different compositions of entrepreneur types: the former comprised of entrepreneurs focusing more on fundamentals, the latter comprised of entrepreneurs who attend to information in a myopic or constrained manner.

In the most basic models of interest-rate determination (e.g. Lucas 1978), an increase in the growth rate of the economy increases the risk-free rate in equilibrium. Non-behavioral fundamentals-focused agents interpret changes in the short-term rate according to fundamentals, and therefore see an increase in the interest rate as a signal of an increase in the growth rate of the economy, leading to an increase in high-propensity business formation (HBA). Meanwhile more myopic agents see the short-term rate primarily as a cost increase (which it is, of course), but fail to associate increases in the short-term rate with changes in the economy's growth prospects, and therefore they are less likely to launch a business when short-term borrowing costs increase, leading to a decrease in low-propensity business formation (LBA). Hence the positive coefficient on T30R for HBA but negative coefficient on T30R for LBA.

The difference in the effect of sentiment (MICS) among the HBA and LBA cohorts indeed does support the interpretation that the former are more fundamentals-focused while the latter display non-normative behavioral information processing. But the difference in response to changes in the short-term interest rate could also be plausibly attributed to less business experience and more heavy reliance on short-term borrowing by the LBA cohort. This is consistent with the notion that the HBA cohort is comprised of more experienced, more well-resourced entrepreneurs who are indeed more likely to support a payroll within six months.

Many of the coefficients referred to here are quite large. As observed in connection with Table , the coefficient on MICS reflects the effect of a one standard-deviation increase in this sentiment variable, rather than the effect of a one-unit increase in the level of MICS. Meanwhile, the national-LBA measure's standard deviation is more than twice that of national HBA. Table  shows that sentiment primarily predicts the former, higher-variability measure. The coefficient on MICS is 10× larger in the national-LBA regression than in the national-HBA regression. Finally, several predictors are recorded as growth/return rates (INFL, T30R, CONS, PROD, SPR3, SPR10, SPRD, and EMPL). Hence the magnitude of the coefficient on e.g. T30R – which is −8413.977 for national LBA in Table  – reflects the effect of increasing T30R by 10,000 basis points. These features of the dependent and independent variables are crucial both for interpreting the regression coefficients as well as for evaluating their plausibility.

The adjusted R2 is consistently high across different models reported in Table . However the relative weight of MICS in accounting for the variance explained differs substantially between high- and low-propensity business formation. The HBA series are primarily predicted by fundamentals, and the effect of sentiment is much smaller in magnitude than for the LBA series (compare Figures 4(a,b)). The MICS coefficients for LBA series are 11× (for national), 19.5× (for Northeast), 11.4× (for Midwest), 11.3× (for South), and 4.4× (for West) larger than the corresponding HBA series. The relative weight of MICS in explaining HBA is consistently low and ranges from 1.3 to 14.5% (see Table  and Figure 4(c)). In contrast, the LBA series are jointly driven by sentiment and economic fundamentals, where the relative weight of MICS ranges from 35.7 to 50.6% (see Table  and Figure 4(d)).

Figure 4. MICS coefficients (top) and weights (bottom), separately for HBA (left) and LBA (right). (a) MICS HBA coefficients. (b) MICS LBA coefficients. (c) MICS HBA weight (%) and (d) MICS LBA weight (%).

5.4. Additional checks

5.4.1. Out-of-sample performance

We assess the out-of-sample (OOS) performance of three competing model specifications using two training samples, two estimation approaches, and three performance criteria. We run these analyzes separately for each combination of business-formation outcome (BA, LBA, HBA) and region (National, Northeast, Midwest, South, West).

The three model specifications included in this OOS horse race are: the Grand Mean model (GM), the Fundamentals-only model (F), and the Fundamentals + Sentiment model (F + S).

To ensure results do not hinge on a particular choice of within-sample estimation period, we work with two different training samples: a 10-year sample running from 2006M1 to 2015M12, and an 11-year sample running from 2006M1 to 2016M12. For each training sample, we estimate OOS performance both by fixing the estimation window to the training sample as well as by incrementally extending the estimation window by one month and then recalibrating. In the latter ‘rolling-and-recalibrating’ forecasting approach, each one-month-ahead forecast benefits from regression coefficients estimated on all previous data points. As performance criteria we report Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). These OOS performance criteria stand on their own without standard error information or the need for associated bootstrapping.

Several strong, consistent findings emerge from the OOS forecasting performance reported in Figures 5–8. First, for BA and LBA, the Fundamentals-only model improves OOS performance relative to the Mean-only model. Adding Sentiment to Fundamentals greatly improves OOS performance. This performance ranking is consistent across the RMSE, MAE, and MAPE criteria. Second, for HBA, differences between the OOS performance of the three different models are much less pronounced; neither Sentiment nor Fundamentals improve OOS performance relative to the Mean-only model. This finding is consistent across training samples, estimation approaches, and geographical regions on both RMSE and MAE performance criteria. On the MAPE performance criterion differences between models are somewhat more pronounced,15 though the ranking of models varies among training samples, estimation approaches, and geographical regions.

Figure 5. Out-of-sample forecasting performance: fixed 10-year training sample.

Figure 6. Out-of-sample forecasting performance: rolling-and-recalibrating from a 10-year training sample.

Figure 7. Out-of-sample forecasting performance: fixed 11-year training sample.

Figure 8. Out-of-sample forecasting performance: rolling-and-recalibrating from an 11-year training sample.

These two findings are consistent with the conclusions of the analysis conducted in Section 5.3: the non-Sentiment and Sentiment models have the same forecasting performance for HBA because these entrepreneurs do not confuse the ‘noise’ of sentiment for fundamental signal. Also note that the standard deviation of LBA is more than twice that of HBA – there is much more variation to be explained in the former than in the latter. In Table , the constant term and T30R16 are highly significant, large-coefficient predictors of HBA. The OOS findings reinforce an interpretation whereby the ‘innovation process’ that generates HBAs is relatively stable and not responsive to Sentiment.

5.4.2. Multicollinearity

Many of the variables used in this analysis are correlated – some of them highly and statistically significantly so (see Table ). These correlations may create multicollinearity problems when estimating Equation (6). Nevertheless multicollinearity between explanatory variables is possible even without high bivariate correlation (Salmerón, García, and García 2018). The analysis reported in this section employs Variance Inflation Factors (VIFs) – and in particular the VIF>10 rule of thumb – to identify multicollinearity.17 We apply a procedure whereby the predictor with the highest VIF above 10 is dropped, and the regression is re-estimated. This procedure is implemented recursively until VIF values fall below 10 for all remaining predictors. We find that removing SPRD from the bank of predictors will eliminate multicollinearity for all the business-formation measures in Tables  and  except national HBA and Northeast HBA. For these two regressions both SPRD and RECES should be removed. The results from the new regressions after eliminating multicollinearity are available from the authors upon request. These results show that the findings from Tables  and remain robust after controlling for multicollinearity.

5.4.3. Variable transforms

In this section we discuss two robustness checks: (i) replacing the recession indicator with a Global Financial Crisis (GFC) indicator, and (ii) substituting variables with their log transforms.

The NBER recession indicator used throughout the main analysis sections is a dummy variable for the 01/2008–07/2009 period. The spread between LIBOR and the Overnight Indexed Swap rate (OIS) is a measure of credit risk in the banking sector. The US Dollar 3-month LIBOR-OIS interest-rate spread spiked on 9 August 2007, and only returned to pre-crisis levels on 7 May 2009. Hence we define the GFC period in the credit crunch variable CRECRU as 08/2007–05/2009.

Results for the GFC dummy CRECRU are reported in columns 2–6 of Tables . Like the RECES indicator in the BA and LBA models, CRECRU is significant at the national level as well as in all sub-national regions. In these models, MICS remains significant across all regions at the 0.1% level, but with marginally higher coefficient estimates than with RECES. In the BA models with CRECRU, T30R is significant both at the national level as well as in the South, whereas with RECES it is only significant in the South. Meanwhile in the LBA models, T30R has large coefficients significant at the 0.1% level in all regions with both CRECRU and RECES. For the BA models, PMI is significant at the national level and in all four sub-regions, compared with only two out of four with RECES. For the LBA models, PMI is significant at the national level and in three out of four sub-regions – the same proportion as with RECES. For the BA and LBA models, substitution of RECES with CRECRU marginally increases the adjusted R2. Overall the BA- and LBA-model results are robust to substitution of RECES with CRECRU.

Table 7. Coefficients of multivariate regressions and standard errors with robustness checks: BA and log(BA).

Table 8. Coefficients of multivariate regressions and standard errors with robustness checks: LBA and log(LBA).

Table 9. Coefficients of multivariate regressions and standard errors with robustness checks: HBA and log(HBA).

For the HBA models, two of the regional coefficients for CRECRU are significant, compared with only one for RECES. With respect to MICS, T30R and the constant term as predictors of HBA, there is no difference between the RECES and CRECRU variants. The HBA-model results are robust to substitution of RECES with CRECRU.

Robustness is explored further by applying logarithmic transforms to MICS, PMI, and all regional variants of BA, LBA, and HBA. Results of re-estimating Equation (6) with these log transforms are reported in columns 7–11 of Tables .

For the BA models, the logarithmic transforms have no effect on the significance of MICS. However whereas PMI is a significant predictor of BA nationally, for the Northeast and for the Midwest, log(PMI) is a significant predictor of log(BA) only for the Midwest. The constant term is insignificant in all of the log-transformed BA models, wheres without log transforms, the constant term is significant only for the Northeast. The adjusted R2 and the relative weight of log(MICS) is marginally lower than in the models without log transforms.

For the LBA models, only two coefficients' significance is changed by log transforms, and all of the log(MICS) coefficients remain significant at the 0.1% level. The coefficient on log(PMI) becomes non-significant for national and West-region log(LBA). The log transforms reduce the adjusted R2 marginally from 0.576–0.665 to 0.556–0.646, whereas the relative weight of log(MICS) is reduced by approximately 10%, from 37.5–50.6% to 28.5–40.7%.

For the HBA models, only two coefficients' significance is changed by log transforms, and the constant terms remain significant at the 0.1% level across all regions. The coefficient on log(MICS) in the Midwest log(HBA) model drops into non-significance. Outside of any other broader pattern in the estimates, SPR10 pops up as significant in the Northeast log(HBA) model. Compared with the models in Table , the adjusted R2 figures are comparable, while the relative weight of log(MICS) is marginally greater than the relative weight of MICS.

None of the log-transform results undermine the main results and their interpretations set out in Sections 5.2 and 5.3.

5.4.4. Bootstrapped p-values

Up to this point, all regression-coefficient tests have been based upon Newey-West Heteroskedasticity and Autocorrelation-Consistent (HAC) standard errors.18 Nevertheless, the literature has long argued that such HAC corrections hold asymptotically, and therefore may be of questionable validity with finite samples, and especially with relatively small sample sizes. For instance, Goetzmann and Jorion (1993) show that the adoption of bootstrap and HAC corrections lead to opposite inferences regarding whether factors such as dividend yields can be used to forecast future stock prices. And as an anonymous reviewer has pointed out, although the sample size in the present study is technically acceptable, it can be treated as ‘small.’ Responding to these considerations, in this section we check the robustness of these Newey-West based inferences with bootstrapped p-values.

The bootstrap procedure is implemented as follows:

  1. For each estimation of Equation (6), coefficients, residuals, and test statistics are stored.

  2. Residuals are drawn with replacement to generate a bootstrapped artificial residual series.19

  3. For each targeted regressor (Factori), a pseudo series of the dependent variable is generated under the null hypothesis that a zero coefficient is associated with Factori. Estimated coefficients from step (a) are used for all other regressors, including the constant, in generating the pseudo dependent variable.

  4. The pseudo dependent variable is regressed on all regressors. Coefficients and the t-statistic of Factori are recorded.

  5. Steps noted above are repeated 9999 times. The bootstrap sample size is chosen so that α(1+B) becomes an integer, making the simulation closer to be exact, where α is the significance level and B is the bootstrap sample size (MacKinnon 2006).

  6. The empirical sampling distribution of the t-statistic under the null hypothesis that Factori shows no predictive power in business formation is then obtained by pooling together the 9999 t-statistic values from step (e).

  7. Reject the null hypothesis at the 5% level if its test statistic ti from step (a) falls outside the 95% quantile of the empirical sampling distribution.

Results are shown in Tables  and , with p-values reported in brackets.

Table 10. Coefficients of multivariate regressions and bootstrapped p-values: BA.

Table 11. Coefficients of multivariate regressions and bootstrapped p-values: LBA and HBA.

Some small differences are evident between the bootstrapped p-values in Table  and the Newey-West-standard-error based inferences reported in Table , but none that impinge upon the four principal findings discussed in Section 5.2. With bootstrapped p-values, PMI remains a significant predictor of BA at a national level as well as in the Northeast and the Midwest. But whereas with Newey-West standard errors PMI is not a significant predictor of BA in the South and West regions, with bootstrapped p-values it becomes a significant predictor of BA in these regions. With Newey-West standard errors, the RECES variable is significant nationally as well as in each individual region. With bootstrapped p-values, the RECES variable is also significant nationally, but only in three out of four individual regions. Finally, whereas the regression constant is non-significant for national-level BA under Newey-West standard errors, it becomes significant with bootstrapped p-values. Most importantly, the sentiment variable MICS is significant at the 0.1% level nationally and for each sub-national region under both Newey-West and bootstrapped p-values.

Bootstrapped p-values in the decomposition by payroll propensity (Table ) confirm the robustness of the Newey-West based findings in Table  and which are discussed in Section 5.3. The only difference is that according to bootstrapped p-values, the significance of RECES in predicting national-level LBA is driven entirely by LBA in the West region. All of the key distinctions between LBA and HBA survive the standard errors robustness check.

6. Conclusion

This study brings both new data and new conceptual apparatus to bear upon entrepreneurial business formation. The US Census Bureau's weekly business formation statistics, which we aggregate up to the monthly frequency, can be a rich resource for future entrepreneurship research. In this study we exploit the overlap between entrepreneurship and finance to justify the use of behavioral finance concepts such as fundamentals-focused information processing as well as its converse, which is susceptible to classifying noise as signal and vice versa. In this second category, widespread mood or affect – i.e. sentiment – can influence decision making.

Our results show that broad business formation is jointly determined by economic fundamentals and consumer sentiment. Sentiment, proxied by the Michigan Index of Consumer Sentiment, predicts month-ahead business formation positively and significantly. Thus, sentiment operates as an opportunity-pull factor explaining future business formation. Although two fundamentals variables (the composite purchasing managers' index and the 1-month real US Treasury bill return) show some predictive power in particular regions, one further variable stands out nationally and across all regions: the recession indicator. This variable operates as a necessity-push factor explaining future business formation.

The Census Bureau data allows business formation to be partitioned into high-payroll-propensity and low-payroll-propensity subsets. Separate analysis of these subsets reveals that the aggregate-level results conceal two very different response patterns. High-propensity business formation is mainly driven by fundamentals, with sentiment accounting for a small proportion of explained variation. Low-propensity business formation, on the other hand, is jointly driven by consumer sentiment and fundamentals, with sentiment accounting for close to half of explained variation. Moreover, the short-term interest rate variable (1-month real US Treasury bill return) predicts month-ahead low-propensity business formation negatively and month-ahead high-propensity business formation positively. The results for high-propensity entrepreneurs are consistent with fundamentals-oriented information processing responding primarily to pull motives. In contrast, the results for low-propensity entrepreneurs are consistent with not solely fundamentals-oriented – i.e. also noise- and sentiment oriented – information processing responding to both push and pull motives.

The results indicate not only that some entrepreneurs are responding to external-trigger factors while others are responding to internal-trigger factors, but that a subset of the entrepreneur population appears to be responding simultaneously to both internal- and external-trigger factors. An individual's mood or affect is a transient internal-trigger factor, but when individuals' mood or affect is correlated across the economy, it also satisfies the criteria for being an external-trigger factor. Thus sentiment is neither exclusively an internal-trigger factor nor exclusively an external-trigger factor. The question of whether the effective trigger is internal, external, or both, is left for future work to enrich the conception of sentiment employed in both behavioral finance and entrepreneurship.

Acknowledgments

We wish to thank the editors and reviewers for their painstaking and helpful input. The usual disclaimer applies. All remaining shortcomings are our own.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Notes on contributors

Kim Kaivanto

Kim Kaivanto holds the post of Senior Lecturer in Economics. He also holds the administrative role of Director of the MSc in Money, Banking and Finance. Prior to joining Lancaster, Kim held research fellowships at the Eitan Berglas School of Economics and Warwick Business School. He obtained his PhD from Warwick University. His research focuses on the interplay between normative and behavioral models of decision making under risk and uncertainty, with applications in financial and non-financial problem settings.

Peng Zhang

Peng Zhang is an Assistant Professor in Economics and Finance at College of Humanities and Sciences of Guizhou Minzu University and Business School of Guizhou Minzu University. Peng's main research interest falls into new theoretical and empirical findings within financial economics when conventional assumptions are challenged by informational, behavioural and/or methodological imperfections.

Notes

1 https://www.gemconsortium.org/

2 See Evans and Jovanovic (1989).

3 A real-options formulation of (3) features an additive, positive-value option term on the right-hand side, which introduces hysteresis in the transition from paid work to entrepreneurial self-employment. Discrete changes in εw can influence the magnitude of this option term, and thus the width of the hysteresis band. For instance if an individual loses current employment, εw=0, then the option value is extinguished and the width of the hysteresis band collapses to zero. However, neither the option term itself nor its dependence upon εw changes the signs of the marginal effects of εw and εe upon the probability of transition into entrepreneurial self-employment.

4 https://www.census.gov/econ/bfs/index.html

5 https://www.irs.gov/forms-pubs/about-form-ss-4

6 These series exclude

“applications outside of the 50 states and the District of Columbia or those with no state-county geocodes, applications with a NAICS sector code of 11 (agriculture, forestry, fishing and hunting) or 92 (public administration), and applications in certain industries (i.e. private households, certain financial services, civic and social organizations)” (U.S. Census Bureau 2020).

7 US Census Bureau defines high-propensity applications as those

“(a) for a corporate entity, (b) that indicate they are hiring employees, purchasing a business or changing organizational type, (c) that provide a first wages-paid date (planned wages); or (d) that have a NAICS industry code in manufacturing (31–33), retail stores (44), health care (62), or restaurants/food service (72)” (U.S. Census Bureau 2020).

For the relationship to the Longitudinal Business Database (LBD), see Jarmin and Miranda (2002).

8 LBA sd in the south region is the dominant driver of the national LBA's sd.

9 Here standardization means subtracting the mean and then dividing by the standard deviation.

10 A PMI value above 50 indicates that purchasing activity has improved month-on-month, whereas a PMI value below 50 indicates that purchasing activity has deteriorated month-on-month.

11 Model (3) of the Johansen Cointegration Test is adopted, i.e. both the cointegration equation and the Vector Autocorrelation Model have intercepts but no trend.

12 These curves capture the effect of multiple ϕi parameters, up to the optimal lag order k in each case.

13 We verify that residuals from regressing Equation (6) satisfy stationarity, validating the statistical inference.

14 Robustness tests confirm that the predictive power found in INFL, PROD, EMPL and RECES is not robust but rather subject to the choice of standard error adjustments.

15 MAPE treats over- and under-predictions asymmetrically, and is known to overstate prediction errors relative to other performance criteria.

16 Note that following the financial crisis, T30R remained close to zero with very low volatility throughout the sample period, including the OOS period.

17 Different levels of stringency are reflected in the choice of VIF cutoff. The cutoff of 10 is used widely, although some researchers advocate 5, or even 2.5.

18 Following Newey and West's original recommendation, the bandwidth parameter is set to 4(T/100)2/9.

19 In order to capture the time-series feature of business formation within a contiguous period, e.g. 12 months, we also conduct this step using a blocked-bootstrap approach, where residuals are drawn in blocks of length 12. The last draw is truncated to fit the required sample size. This parallel approach leads to no differences in our statistical-inference results.

References

  • Amit, R., and E. Muller. 1995. “‘Push’ and ‘Pull’ Entrepreneurship.” Journal of Small Business & Entrepreneurship 12 (4): 6480. [Taylor & Francis Online][Google Scholar]
  • Ardagna, S., and A. Lusardi. 2010. “Explaining International Differences in Entrepreneurship: The Role of Individual Characteristics and Regulatory Constraints.” In International Differences in Entrepreneurship, edited by J. Lerner, and A. Schoar, 17–62. Chicago: NBER and Chicago University Press. [Google Scholar]
  • Baker, M., and J. Wurgler. 2007. “Investor Sentiment in the Stock Market.” Journal of Economic Perspectives 21 (2): 129151. [Crossref], [Web of Science ®][Google Scholar]
  • Baron, R. A. 2000. “Psychological Perspectives on Entrepreneurship: Cognitive and Social Factors in Entrepreneurs' Success.” Current Directions in Psychological Science 9 (1): 1518. [Crossref], [Web of Science ®][Google Scholar]
  • Bayard, K., E. Dinlersoz, T. Dunne, H. Haltiwanger, J. Miranda, and J. Stevens. 2018. Early-Stage Business Formation: An Analysis of Applications for Employer Identification Numbers. Finance and Economics Discussion Series 2018-015. Washington, DC: Board of Governors of the Federal Reserve System. [Crossref][Google Scholar]
  • Bergmann, H., S. Mueller, and T. Schrettle. 2014. “The Use of Global Entrepreneurship Monitor Data in Academic Research: A Critical Inventory and Future Potentials.” International Journal of Entrepreneurial Venturing 6 (3): 242276. [Crossref][Google Scholar]
  • Brown, G. W., and M. T. Cliff. 2005. “Investor Sentiment and Asset Valuation.” Journal of Business 78: 405440. [Crossref], [Web of Science ®][Google Scholar]
  • Cleveland, R. B., W. S. Cleveland, J. E. McRae, and I. Terpenning. 1990. “STL: A Seasonal-trend Decomposition Procedure Based on Loess.” Journal of Official Statistics 6 (1): 333. [Google Scholar]
  • Decker, R., J. Haltiwanger, R. Jarmin, and J. Miranda. 2014. “The Role of Entrepreneurship in US Job Creation and Economic Dynamism.” Journal of Economic Perspectives 28 (3): 324. [Crossref], [Web of Science ®][Google Scholar]
  • Evans, D. S., and B. Jovanovic. 1989. “An Estimated Model of Entrepreneurial Choice Under Liquidity Constraints.” Journal of Political Economy 97 (4): 808827. [Crossref], [Web of Science ®][Google Scholar]
  • Fairlie, R. W., and F. M. Fossen. 2019. Defining Opportunity Versus Necessity Entrepreneurship: Two Components of Business Creation. NBER Working Paper 26377. [Crossref][Google Scholar]
  • Frese, M., and M. M. Gielnik. 2014. “The Psychology of Entrepreneurship.” Annual Review of Organizational Psychology and Organizational Behavior 1: 413438. [Crossref][Google Scholar]
  • Gilad, B., and P. Levine. 1986. “A Behavioral Model of Entrepreneurial Activity.” Journal of Small Business Management 24 (4): 4553. [Google Scholar]
  • Goetzmann, W. N., and P. Jorion. 1993. “Testing the Predictive Power of Dividend Yields.” The Journal of Finance 48: 663679. [Crossref], [Web of Science ®][Google Scholar]
  • Gough, J. W. 1969. The Rise of the Entrepreneur. New York: Schocken Books. [Google Scholar]
  • Hurst, E., and B. W. Pugsley. 2011. “What Do Small Businesses Do?Brookings Papers on Economic Activity 2: 73118. [Crossref][Google Scholar]
  • Jarmin, R. S., and J. Miranda. 2002. The Longitudinal Business Database. Working Paper Number CES-02-17. Washington, DC: Center for Economic Studies, U.S. Census Bureau. https://www.census.gov/library/working-papers/2002/adrm/ces-wp-02-17.html. [Google Scholar]
  • Johnson, J. W. 2000. “A Heuristic Method for Estimating the Relative Weight of Predictor Variables in Multiple Regression.” Multivariate Behavioral Research 35: 119. [Taylor & Francis Online], [Web of Science ®][Google Scholar]
  • Klapper, L., R. Amit, M. F. Guillén, and J. M. Quesada. 2010. “Entrepreneurship and Firm Formation Across Countries.” In International Differences in Entrepreneurship, edited by J. Lerner, and A. Schoar, 129–158. Chicago: NBER and University of Chicago Press. [Crossref][Google Scholar]
  • Lemmon, M., and E. Portniaguina. 2006. “Consumer Confidence and Asset Prices: Some Empirical Evidence.” Review of Financial Studies 19 (4): 14991529. [Crossref], [Web of Science ®][Google Scholar]
  • Licht, A. N., and J. I. Siegel. 2006. “The Social Dimensions of Entrepreneurship.” In Oxford Handbook of Entrepreneurship, edited by M Casson, B. Yeung, A. Basu, and N. Wadeson, 511–539. Oxford: Oxford University Press. [Google Scholar]
  • Lucas Jr., R. E. 1978. “Asset Prices in An Exchange Economy.” Econometrica 46 (6): 14291445. [Crossref], [Web of Science ®][Google Scholar]
  • MacKinnon, J. G. 2006. “Bootstrap Methods in Econometrics.” Economic Record 82: S2S18. [Crossref], [Web of Science ®][Google Scholar]
  • Qiu, L., and I. Welch. 2006. Investor Sentiment Measures. NBER Working Paper 10794. [Crossref][Google Scholar]
  • Salmerón, R., C. García, and J. García. 2018. “Variance Inflation Factor and Condition Number in Multiple Linear Regression.” Journal of Statistical Computation and Simulation 88 (12): 23652384. [Taylor & Francis Online], [Web of Science ®][Google Scholar]
  • Schoar, A. 2010. “The Divide Between Subsistence and Transformational Entrepreneurship.” In NBER Innovation Policy and the Economy, edited by J. Lerner, and S. Stern, Vol. 10, 57–81. Chicago: NBER and Chicago University Press. [Crossref][Google Scholar]
  • Schumpeter, J. A. 1911. Theorie Der Wirtschaftlichen Entwicklung. Leipzig: Duncker & Humblot. [Google Scholar]
  • Shiller, R. J. 1999. Measuring Bubble Expectations and Investor Confidence. NBER Working Paper 7008. [Crossref][Google Scholar]
  • Sternberg, R., and S. Wennekers. 2005. “Determinants and Effects of New Business Creation Using Global Entrepreneurship Monitor Data.” Small Business Economics 24 (3): 193203. [Crossref], [Web of Science ®][Google Scholar]
  • Tonidandel, S., and J. M. LeBreton. 2011. “Relative Importance Analysis: A Useful Supplement to Regression Analysis.” Journal of Business and Psychology 26 (1): 19. [Crossref], [Web of Science ®][Google Scholar]
  • U.S. Census Bureau. 2020. Business Formation Statistics, Second Quarter 2020. Release Number CB20-106. Washington, DC: U.S. Census Bureau. https://www.census.gov/econ/bfs/index.html. [Google Scholar]

Alternative formats