Flight to Safety: How Economic Downturns Affect Talent Flows to Startups

Abstract

Using proprietary data from AngelList Talent, we study how startup job seekers’ search and application behavior changed during the COVID-19 downturn. We find that workers shifted their searches and applications away from less-established startups and toward more-established ones, even within the same individual over time. At the firm level, this shift was not offset by an influx of new job seekers. Less-established startups experienced a relative decline in the quantity and quality of applications, ultimately affecting their hiring. Our findings uncover a flight-to-safety channel in the labor market that may amplify the procyclical nature of entrepreneurial activities.

Economists have long debated the role of entrepreneurship during economic downturns. Under the cleansing hypothesis, recessions are times of accelerated reallocation, where inefficient incumbents are replaced by new firms who seize market opportunities (Davis and Haltiwanger 1992; Foster, Haltiwanger, and Krizan 2001; Collard-Wexler and De Loecker 2015). However, an increasing body of evidence highlights that early-stage startups may be particularly vulnerable to economic downturns, and therefore less able to drive such cleansing effects (Parker 2009; Decker et al. 2016; Fabrizio and Tsolmon 2014). Existing explanations of startup vulnerability during recessions primarily focus on the financing constraints that early-stage firms face when attempting to raise capital during downturns (Barlevy 2003; Aghion et al. 2012; Townsend 2015; Nanda and Rhodes-Kropf 2016; Howell et al. 2020). In this paper, we explore a new channel: the ability of early-stage companies to attract human capital during economic downturns. It is theoretically unclear how downturns should affect the ability of early-stage startups to attract human capital. For at least two reasons, employee interest might shift toward more-established firms. First, workers may become more risk-averse during recessions, such that their desire for safety increases. Second, workers may perceive the safety offered by more-established firms relative to less-established firms as being greater during recessions. These two potential drivers are not mutually exclusive, so both could be operating simultaneously. On the other hand, many workers lose their jobs during downturns or face worse career trajectories at established firms, lowering their opportunity costs of joining early-stage startups (Gottlieb, Haltiwanger, and Krizan 2019). Thus, the overall increase in the supply of potential workers for early-stage startups may offset any changes in worker preferences away from them.

Empirically exploring whether and how the supply of talent available to startups changes during economic downturns is challenging due to the difficulty of distinguishing between supply and demand factors that drive labor market outcomes. For example, a decline in hiring by early-stage startups could reflect a change in the hiring policies of such firms (labor demand), a decline in worker interest in such firms (labor supply) or both. A handful of recent studies have used online job posting data to investigate various questions about labor demand (Campello, Gao, and Xu 2019; Campello, Kankanhalli, and Muthukrishnan 2020; Kahn, Lange, and Wiczer 2020); however, such data tell us little about labor supply.

To analyze labor supply, we make use of a novel data set that we obtained from AngelList Talent, the largest online recruitment platform for private and entrepreneurial companies. The data we use come from AngelList Talent’s back-end system, and therefore include not only publicly visible job postings but also the history of each user’s job searches on the platform, their application submissions, and whether employers responded to these submitted applications. Because we can observe the activities of job seekers in these data, we can learn about changes in labor supply. In particular, we are able to track changes in the search behavior of the same job seeker over time. This allows us to explore whether worker preferences shift during downturns, independently of changes in labor demand and, if so, what type of workers experience changes in preferences. In addition, we are also able to track changes in job applications for the same job posting over time. This allows us to examine whether, for firms, changes in worker preferences following a downturn are offset by changes in the number of workers seeking employment. Both analyses isolate supply-side factors much more cleanly than has been possible with standard data sets. Because of the nature of AngelList Talent, we primarily compare how labor supply changes during a recession for less-established startups relative to more-established startups. However, to the extent that job seeker interest shifts toward more-established startups, it seems plausible that their interest would also shift toward more-established nonstartups as well.

We focus on the economic downturn that followed the emergence of the COVID-19 pandemic. The pandemic caused massive economic disruptions, and its origins were external in nature, providing an ideal, exogenous setting to study the response of job seekers to adverse economic shocks. While the COVID-19 downturn was distinct from others in many ways, it shares some important similarities for our purposes. As during other recessions, the economic expectations of workers declined at the start of the COVID-19 downturn before eventually rebounding. This decline in expectations is what could drive a change in job seekers’ preferences and behavior. Survey evidence suggests that the magnitude of the decline in expectations during the COVID-19 downturn was in line with past recessions and, if anything, was smaller. However, economic expectations declined much more quickly during the COVID-19 downturn than in the past and also rebounded more quickly. The sharpness of this decline in expectations is advantageous for identification, as it allows us to focus on a short window in the months surrounding the start of the pandemic. Within this short window, it is unlikely that talent flows to startups coincidentally changed sharply at the same time as workers’ economic expectations. Moreover, the unusually quick rebound in expectations that occurred subsequently does not affect our interpretation of results. Rather, our results suggest that talent flows to startups are likely affected for a longer period of time during a more typical recession.

Exploring changes in the search parameters of AngelList Talent users, we find that job candidates searched for significantly larger companies after March 13, 2020, the date that a state of national emergency was first announced in the United States. Specifically, the average size of firms searched by candidates increased by 29%, and candidates became 20% more likely to search for firms with more than 500 employees. This result holds both across candidates and, importantly, within the same candidate over time. In other words, the COVID-19 downturn led job candidates to shift their search preferences toward more-established firms. Next, we examine whether changes in the search preferences of job seekers also translated into job applications. Consistent with the changes in job searches, we find a significant increase in the average size and financing stage of firms receiving job applications after the start of the downturn. Again, these effects not only hold in the cross-section across all candidates on the platform but also within candidates, suggesting that the crisis changed the type of firms that candidates chose to apply to.

Next, we explore whether the flight-to-safety effects that we document differ for high- and low-quality job seekers. In particular, we partition candidates according to two characteristics that we can observe in the data: their number of years of relevant work experience and an estimated score of their overall quality. The latter measure is created by AngelList Talent based on an algorithm that accounts for applicants’ experience, skills, and education. Interestingly, we find that higher-quality job seekers drive most of the flight to safety in job applications, shifting away from less-established startups. This could reflect the potentially better outside options such job candidates have, which may make them more averse to startup jobs when startups are perceived to be riskier during downturns. Relatedly, we also find that the flight-to-safety effect is stronger among candidates who search less intensely. These candidates are also more likely to have good outside options. Finally, we also find that the flight-to-safety effect is stronger among those living in Democratic-leaning counties, who likely perceived the pandemic as being more severe.

The results described above suggest a shift in worker interest away from less-established firms during the COVID-19 downturn. However, it is possible that despite this shift, less-established firms had no difficulties attracting human capital during the downturn, or even had an easier time. In particular, it could be that there was a large enough influx of new, high-quality job seekers that offset the change in worker interest. Thus, in the second part of the paper, we turn to estimating effects at the firm level. We find that, on average, the number of applications received per job posting did decline significantly after the onset of the pandemic. We also find again that the decline was concentrated within less-established firms and was driven by a decline in high-quality applicants. In principle, these results could reflect changes in the type of jobs posted by these firms. However, we find similar results within job postings as well. That is, holding the job posting fixed, high-quality applications declined after the crisis, and more so for jobs posted by less-established startups. We further find that the deterioration in the applicant pool for less-established startups led them to be less likely to respond to applications and ultimately to hire candidates. These results highlight the difficulty early-stage startups face when attempting to attract human capital during downturns.

We conduct a variety of robustness tests. First, we show that our main results are absent over the same time period in 2019, suggesting that our results are not driven by seasonality or unobserved trends. Second, we show through nonparametric graphs that our main results do not reflect a general trend in the labor market. Instead, reactions are steep and immediate, and coincide with the emergence of the pandemic in the United States. These graphical results also show that more- and less-established startups shared similar trends in the months before the crisis. Third, we show that candidates were not only searching for larger firms in an absolute sense during COVID-19 but also more likely to search for firms larger than their current employer. Fourth, we show that our results on changes in worker preferences hold within subsamples of searches that were not preceded by another recent search, suggesting that the results are not driven by individuals adjusting their search parameters in response to the job postings they see from previous searches. Fifth, we show that workers searching for in-office (remote-only) jobs after COVID-19 searched for larger firms than those searching for in-office (remote-only) jobs before COVID-19, suggesting that our results are not driven by a flight to remote work or job flexibility. Lastly, we show that our results are similar when we use the state-level number of COVID-19 cases or consumer economic expectations as continuous treatment variables.

We corroborate the flight-to-safety interpretation of our main results in several ways. If job seekers shifted away from less-established firms during COVID-19 due to a flight to safety, we should expect that such firms have indeed been less safe during recessions historically. Using U.S. Census data from 1979 to 2019, we confirm that smaller and younger firms are more likely to fail and shed jobs during recessions. Further, under the flight-to-safety hypothesis, job seekers also should be attracted to other signals of startup safety besides size and stage. Consistent with this idea, we show that job seeker interest also shifted toward startups funded by “top-tier” VCs as well as those who had recently raised financing.

What drives the flight to safety we document? There are two non-mutually- exclusive possibilities. First, workers may become more risk-averse during recessions, such that their desire for safety increases. Alternatively, workers may perceive the safety offered by more-established firms relative to less-established firms as being greater during recessions. While it is difficult to cleanly disentangle these two drivers, we explore which is likely to be more important in our data. Specifically, we examine how the shift in employee interest toward safer firms varies with startup quality. We find a significant shift from small low-quality firms to large low-quality firms but also an equally large shift from small high-quality firms to large high-quality firms. These results seem less consistent with the change in relative safety interpretation, as it is unlikely that the safety of high-quality startups declined by the same amount as that of low-quality startups. Rather, they point more toward the change in risk-aversion interpretation, as increased risk-aversion might drive job seekers away from less-established startups regardless of their quality. In any case, the primary contribution of this paper is to show that employee interest shifts toward safer firms during recessions, leading less-established firms to experience increased difficulty in recruiting. Our results thus suggest that labor market frictions can amplify the procyclical nature of entrepreneurial activities.

Our paper contributes to the literature on business cycles and entrepreneurship. Caballero and Hammour (1994), Davis, Haltiwanger, and Schuh (1996), Foster, Haltiwanger, and Krizan (2001), Koellinger and Roy Thurik (2012), and Collard-Wexler and De Loecker (2015) document accelerated reallocation and cleansing of inefficient incumbents during economic downturns. In contrast, Parker (2009), Decker et al. (2014), Decker, Haltiwanger, Jarmin, and Miranda (2016), and Fabrizio and Tsolmon (2014) show that entrepreneurship and R&D are procyclical rather than countercyclical. This procyclicality has been attributed to financing frictions (Aghion et al. 2012; Townsend 2015; Nanda and Rhodes-Kropf 2016), R&D externality (Barlevy 2007), and entrepreneurs’ human capital choice (Rampini 2004). Our paper introduces a new labor channel to explain startup vulnerability to economic downturns. Focusing on the performance of startups founded during recessions, Moreira (2016) shows that firms born in downturns are persistently smaller and that this effect is driven by both selection and demand-side channels. Bias and Ljungqvist (2022) find that, after purging selection, startups founded during the Great Recession have had better long-term performance, driven by lower turnover of founding inventors. In related work, Hacamo and Kleiner (2022) show that firms founded by students graduating during periods of high unemployment have better long-term performance, while Babina (2020) shows that incumbents’ financial distress spurs high-quality employees to enter entrepreneurship. Both Hacamo and Kleiner (2022) and Babina (2020) show that downturns affect the type of workers entering entrepreneurship. Our paper differs from those above in several ways. First, we study the decision to work at a startup rather than the decision to found a startup. Second, in terms of our individual-level results, we do not focus specifically on those who are forced into unemployment due to downturns. Rather, we study the overall effect of downturns on worker interest in startups, including workers who remain employed. Finally, in terms of our firm-level results, we do not focus on the effect of economic conditions at a startup’s founding date. Rather, we study how a decline in economic conditions affects all startups, including those that were founded during booms.

We also add to an emerging literature on the startup labor market. Babina and Howell (2018) and Babina, Ouimet, and Zarutskie (2020) study human capital flows between incumbents and startups. Moscarini and Postel-Vinay (2012) and Babina et al. (2019) examine employment and wage dynamics by young firms and their cyclicality. These papers study equilibrium employment outcomes, while we focus on individuals’ labor supply in the job search process. In that sense, our paper is related to a handful of papers that study job searches and applications (Brown and Matsa 2016,2020; Cortes et al. 2020; Gortmaker, Jeffers, and Lee 2019; Kuhnen 2017). Different from these papers, we focus on the startup labor market, which has received little attention relative to the broader labor market. In that regard, this paper relates to Bernstein et al. (2021), who study how venture capital backing affects the nature of human capital startups are able to attract.

Lastly, we add to a recent string of papers that study the labor market consequence of COVID-19. Using job posting and unemployment insurance data, Kahn, Lange, and Wiczer (2020) document a broad-based decline in job postings of 30% by the end of March 2020. Using household survey data, Coibion, Gorodnichenko, and Weber (2020) estimate a job loss of 20 million jobs and 7-percentage-point drop in labor participation rate by April 2020, both of which are greater than what happened over the entire Great Recession. Bartik et al. (2020) show that low-wage workers and business closures drive most of the decline in small business employment at the onset of COVID-19. Using job posting data, Campello, Kankanhalli, and Muthukrishnan (2020) show that, among public firms, small and credit-constrained firms cut back on job postings more during COVID-19; there is also a larger decline in high-skill jobs relative to low-skill ones. Our paper focuses on labor supply and the ability of startups to attract talent during the COVID-19 crisis. We also highlight the stark contrast between established and early-stage firms, as well as the disparate responses by high-quality and low-quality job candidates.

1 The AngelList Talent Platform

AngelList was originally founded in 2010 as a platform to connect startups with potential investors. In 2012, it expanded into startup recruiting. The original investment portion of the site, now called AngelList Venture, was separate from the recruiting portion of the site, AngelList Talent. One of the key features of AngelList Talent was that it did not allow third party recruiters. It also encouraged transparency about salary and equity upfront, before candidates applied.

Since its launch, AngelList Talent has rapidly grown in popularity, becoming an important part of the entrepreneurial ecosystem. Firms recruiting on the platform range from nascent startups with less than 10 employees, to mature technology companies, such as Google, Facebook, and Dropbox. Over its lifetime, more than 10 million job seekers have joined the platform, more than 100,000 companies have posted a job there, and more than 5 million connections have been made between job seekers and companies.

Panel A of Figure 1 illustrates the way that AngelList Talent works. Companies can post job openings, specifying their job’s location, role, description, type (i.e., full-time/part-time), salary range, equity range, and other details (Internet Appendix Figure A.1 shows an example). Job postings are also linked to AngelList company profiles that provide further firm-level information, including funding status, size, industry, and team members. After job postings are reviewed for spam they become live for search. Users can search live job postings, potentially specifying a variety of filters based on the job and company characteristics above (panel B of Figure 1 shows an example). Importantly for our purposes, a user must register on the site and provide basic resume information before performing a search. Thus, all searches can be linked to a user by AngelList, although user searches are not publicly visible to companies or other users.

Users can browse the search results and apply for the jobs they are interested in. After a user sends an application to a company, the company can “request an introduction” to the user, “reject” the user’s application, or do nothing, in which case the user’s application is automatically rejected in 14 days. Requesting an introduction to a user allows the two parties to communicate directly. After this connection is made, the rest of the hiring process occurs outside of the platform. Thus, AngelList Talent does not directly observe if a given candidate ends up being hired.

2 Data

2.1 AngelList Talent

The primary data we use in this paper were provided directly by AngelList and were collected by their back-end system. Our sample period runs from February 5 to May 14, 2020, and for comparison we also obtain data from the same period in 2019.¹ In the data, we can observe all user activities, including searches and applications by job candidates, and responses to those applications by firms. We also observe all jobs ever posted on AngelList Talent, with associated job- and firm-level characteristics, and the dates the jobs were live for search.

We use firm size and financing stage as our primary measures of the perceived safety of a firm during a recession. Both are tracked by AngelList. We measure a firm’s size as its number of employees. We categorize a firm as early-stage if its most recent financing round was a Series A round or earlier (i.e., Pre-Seed, Seed, Series A) and later-stage if the round was a Series B round or later (i.e., Series B, Series C, etc.). In addition to firm size and stage, we also use two alternative proxies for firm safety: an indicator for whether a startup was funded by a “top-tier” VC (as categorized by AngelList) and an indicator for whether a startup was funded recently (i.e., in the past 6 months).² The top-tier VC indicator could be viewed as a measure of startup quality and the recently funded indicator could be viewed as a measure of financial health.

All of our proxies for safety were visible to job seekers before they applied. However, only employment size could be specified as a search parameter during our sample period. Specifically, users could filter search results based on employment size by selecting any of seven size bins: 1–10, 11–50, 51–200, 201–500, 501–1,000, 1,001–5,000, and 5,000+. To measure a job seeker’s firm size preference based on a search, we take the midpoint of each bin, average it across all selected bins, and then log transform it.³

In some of our analysis, we are interested in understanding the quality of the candidates who are searching or applying for jobs. We use two primary measures of candidate quality. The first measure is the number of years of job experience a candidate has in their primary role (e.g., Software Engineer). The second is a quality score created by AngelList Talent based on a proprietary algorithm that scores candidates based on their experience, education, and skills. While the latter is somewhat opaque, all results are similar using both measures.

To measure talent flows to firms, we look at job application volume. Although not all job applicants are eventually hired, job applications allow us to measure the size of the talent pool available to firms. Specifically, we measure the number of job applications at the job posting level. This allows us to condition the supply of applications within each “unit” of labor demand, thus addressing concerns that changing talent flows to startups are driven by shifts in their labor demand or job requirements. We also study firms’ responses to job applications. As discussed earlier, we are able to observe whether a firm requests an introduction to the applicant, which indicates the initiation of further interactions. Although we do not observe the final hiring decision in the AngelList data, these introduction requests are precursors to eventual hiring.

2.2 LinkedIn

We augment the AngelList Talent data with data from LinkedIn, which we obtained from Revelio Labs. We use the LinkedIn data for three purposes. First, these data give us visibility into where the individuals in our sample were employed at the time of their searches.⁴ While AngelList users do often upload their resume data directly to AngelList, we do not observe any firm characteristics for the employers they list on their resumes. In contrast, LinkedIn has employer profiles with firm characteristics, including employment size. This allows us to investigate, for example, whether AngelList users are searching for jobs at firms that are larger or smaller than their current firm. Second, the LinkedIn data give us visibility into whether individuals in our sample were ultimately hired by the firms they applied to on AngelList Talent. We obtain a snapshot of LinkedIn profiles in June 2022, and can thus check whether individuals list a firm they applied to on their resume within 12 months of their application. Finally, the LinkedIn data also allow us to observe firm-level work force dynamics, such as changes in a firm’s overall hiring rate over time. This allows us to confirm that firms that seem to be garnering less interest on AngelList Talent indeed are hiring less overall.

2.3 Sample restrictions

We limit our sample to include only the activities of users and firms located in the United States in order to ensure that our findings do not reflect a mix of countries with very different startup ecosystems or labor markets. We also exclude the top 1% of users in terms of their number of searches during the sample period so as to limit the influence of “bots” (i.e., fake users) that might be scraping the AngelList website. Consistent with the idea that these users are bots, their search activity does not fluctuate between weekdays and weekends in the same way as that of other users. Our final sample includes 178,793 users and 83,921 job applicants that were active during our sample period, and 113,382 jobs that were live for search during that period.

3 Empirical Strategy

Our goal is to explore whether worker interest in less-established startups changed following the start of the COVID-19 downturn and whether any such changes affected the ability of early-stage startups to hire high-quality employees. We use the online search and application activities of job candidates on AngelList to identify changes in their preferences and labor supply. Our data have several advantages relative to existing data used in the literature. First, our search parameter data allow us to capture job seekers’ preferences independent of the job vacancies posted by firms, thus separating the labor supply from labor demand. This is not feasible with job posting data that have been used thus far (Campello, Gao, and Xu 2019; Campello, Kankanhalli, and Muthukrishnan. 2020; Kahn, Lange, and Wiczer 2020). Second, compared with surveys of job seekers (Coibion, Gorodnichenko, and Weber 2020; Mui and Schoefer 2020), our data also allow us to measure job seekers’ preferences at a higher frequency and without potential self-reporting biases. Lastly, our granular job application data contain complete information on candidates, jobs, and firms. This allows us to conduct within-candidate and within-job analyses, which are critical in controlling for compositional changes among job seekers and changes in labor demand by firms.

3.1 External validity

Before discussing our empirical specifications, we first begin by considering how generalizable any results that we find are likely to be with respect to other downturns. While the COVID-19 downturn was distinct from others in many ways, it does share some important similarities for our purposes. As during other recessions, the economic expectations of workers declined at the start of the COVID-19 downturn. This decline in expectations is what could drive a flight-to-safety effect. One may worry that the decline in economic expectations may have been more severe during the COVID-19 recession, making it a poor setting for understanding the effects of typical recessions on talent flows to startups. However, if anything, survey evidence suggests that economic expectations actually declined less from peak to trough during the COVID-19 recession than in previous ones. Figure 2 shows the consumer confidence time series over the past two decades, with panel A using the University of Michigan’s Consumer Sentiment Index and panel B using the Conference Board’s Consumer Confidence Index.⁵ The shaded regions represent NBER-defined recessions. The figure shows that during the COVID-19 recession consumer confidence declined by 29%–35%. By comparison, consumer confidence declined by 40%–77% during the financial crisis and by 29%–41% during the dotcom crash.

In terms of changes in economic expectations, a more distinctive feature of the COVID-19 downturn was that expectations declined much more quickly than in past recessions. During the COVID-19 downturn consumer confidence went from its pre-recession high to its recession low over the course of 2-3 months. In contrast, it took 19 months to go from peak to trough during the financial crisis and 18 months during the dotcom crash. The sharpness of the decline in economic expectations is advantageous for us in terms of identification, as it allows us to focus on a short window in the months (February 2020 to May 2020) surrounding the start of the pandemic. Within this short window, it is unlikely that talent flows to startups coincidentally changed sharply at the same time as workers’ economic expectations.

Another distinctive feature of the COVID-19 downturn was that economic expectations rebounded more quickly than usual. Again, because we focus on a short window around the start of the pandemic, this subsequent rebound does not affect our findings.⁶ It is also worth emphasizing that our goal is not to evaluate the full impact of the COVID-19 downturn on the startup labor market. Rather, our goal is to test whether there is a potentially generalizable flight-to-safety effect among workers during recessions. We do this using a short window where there was a sharp shift in workers’ economic expectations. To the extent that expectations usually take longer to recover in other recessions, that would suggest that talent flows to startups may be affected for a longer period of time in other recessions.

Finally, the COVID-19 recession was also somewhat unique in that the government responded with more extreme economic interventions than in past recessions. It could be argued that almost all of these interventions served to economically protect workers to some extent, either directly or indirectly. Therefore, if anything, the government’s policy response would have muted a flight-to-safety effect among workers. Overall, since the decline in economic expectations during the COVID-19 downturn was smaller in magnitude and shorter-lived than in past recessions—and the government policy response was also more robust—our results could be argued to represent a lower bound on the effect of a more typical recession on talent flows to startups.

3.2 Effect on worker preferences

3.2.1 Search parameters

We first explore changes in the search parameters of job seekers on AngelList Talent around the start of the COVID-19 downturn. Specifically, we estimate the following specification at the search level:

where y_it is the searched firm size specified by candidate i searching at time t and $PostCOVIDt$ is an indicator equal to one on dates after March 13, 2020, the date that a state of national emergency was first announced in the United States.⁷ Our main specification includes job seeker fixed effects α_i, which means that we estimate how the preferences of the same individual change in response to the downturn. In some specifications we replace these individual fixed effects by user state fixed effects to allow for compositional changes in the types of individuals seeking jobs around the crisis. In terms of standard errors, at a minimum, it seems sensible to cluster at the user level, as a users’ searches are likely to be serially correlated. To be more conservative, we cluster at the user’s state level, which nests the user level. It is also possible that there is cross-correlation in errors at the day level as well. To account for this, we therefore double cluster all standard errors by state and day.

3.2.2 Applications

We also use a similar specification to explore changes in the types of firms job seekers apply to. Specifically, we explore whether individuals tended to submit applications to larger or later-stage startups after the downturn. To do so, we estimate the following specification at the job application level:

where y_it represents either the size or the financing stage of the firm candidate i applied to at time t, and $Xt$ is a vector of day-level controls that include the average number of employees of firms hiring on AngelList and the total number of job postings on AngelList. Similar to Equation (1), we include candidate fixed effects α_i in the full specification to examine within-candidate changes in application preferences. Standard errors are again clustered by user state and application date.

3.3 Effect on firms

The estimation strategies described above allow us to learn about how worker preferences shifted after the start of the downturn. However, it is possible that the effect of such a shift in preferences on firms could be offset or even reversed by a large enough influx of new job seekers after the crisis. In other words, even though existing workers on the platform may be less interested in working for less-established startups, there may be enough additional workers seeking jobs due to the crisis that these firms actually find it easier to attract human capital. To explore this possibility, we also estimate effects on job applications at the job posting × day level.

Our baseline specification here examines whether the number of applications received by jobs declined following the onset of the crisis. In addition, we examine whether applications declined more for less-established startups than for more-established ones. We estimate the following equation at the job posting × day level:

where $Applicationsfjt$ is the number of new applications to job j at startup f on day t; LessEstablishedStartup_ft is either an indicator for whether a startup has 50 employees or fewer, or an indicator for whether its last financing round was a series A round or earlier at the time of application;⁸θ_jt are fixed effects for the number of days since the job was posted, which account for temporal patterns in application volumes over the lifecycle of a job posting; $Xft$ is a vector of controls that includes the total number of live job postings associated with a startup on a given day and the average size (i.e., number of employees) of all startups hiring on AngelList on a given day. In some specifications, we include firm fixed effects, α_f, thus exploring changes in application volumes within firms. However, changes in application volumes under this specification may reflect changes in the number or type of job vacancies posted by a firm, thus picking up both supply- and demand-side factors. Therefore, in our main specification we include job posting fixed effect, α_j. By examining within-job changes in applications, we are able to hold labor demand factors constant. This allows us to isolate changes in labor supply. We cluster standard errors by a firm’s state and application date.

Lastly, we also examine how the downturn affected the average quality of talent flowing to startups. To do this, we estimate the following specification at the application level:

where ApplicantQuality_ifjt is the number of years of relevant work experience or the estimated quality score for candidate i applying to job j at startup f at time t; LessEstablishedStartup_ft and $Xft$ are defined the same way as those in Equation (3). Standard errors are clustered by a firm’s state and application date. Similar to Equation (3), we control for job fixed effects α_j in the main specification, which ensures that any identified changes in applicant quality are not driven by firms adjusting the types of jobs posted with different job requirements.

4 Results

4.1 Summary statistics

Table 1 provides basic summary statistics. Panel A presents statistics on search parameters entered by job seekers when the unit of observation is at the search level. The average startup size searched by job seekers is 162 employees, with 30% of searches looking for companies with more than 500 employees and 70% of searches looking for companies with more than 50 employees.

Panels B and C present statistics on job applications at the job posting × day level and the application level, respectively. On an average day, a job posting receives 0.19 applications.⁹ The average startup has about 7 live job postings on a given day. The average applicant has 4.2 years of work experience and a candidate quality score of 13.2. About 76% of the applications go to startups with fewer than 50 employees and 73% go to startups that are pre-Series B. The average startup receiving applications has 26 employees.

Figure 3 shows the firm size distribution at the firm level, search level, and application level. The firm-level analysis has one observation for each firm that had a live job posting on AngelList during our sample period, the search-level analysis has one observation for each search, and the application-level analysis has one observation for each application. As can be seen, the size distributions for firms posting jobs on AngelList and for firms that users apply to are highly skewed toward small firms. However, this skewness is less pronounced for searched firm size. Internet Appendix Figure A.2 analogously shows the financing stage distribution and industry distribution. Most of the startups are early stage (Pre-Seed, Seed, or Series A) and are concentrated in the IT, e-commerce, media, business service, and healthcare industries. Finally, Internet Appendix Figure A.3 uses the LinkedIn data to show the size distribution of candidates’ current employers (i.e., at the time of their search), for both the search and the application samples. As can be seen, the modal employer size bin is the largest one (>5,000 employees). Approximately 35% of users who perform a search and 42% of users who apply for a job come from firms in this largest size bin. It should be noted, however, that the firm-level LinkedIn data may be biased toward larger firms. Additionally, users without LinkedIn profiles may tend to work for smaller firms.

4.2 Effect on worker preferences

4.2.1 Job search parameters

We start by analyzing whether job seekers changed their job search criteria following the start of the downturn. Table 2 presents the results estimated from Equation (1) with dependent variables related to the size of firms users search for, as measured by the number of employees. The dependent variable in panel A is the logarithm of the firm size searched for. Observations are at the individual search level. In column 1, we find that following the start of the COVID-19 downturn, users increased the firm size they were searching for. The coefficient of 0.223 is highly statistically significant and indicates a 25%(=exp(0.223)) increase in the size of firms searched for after the crisis began. In column 2, we add job candidate fixed effects, which ensures that the results are not driven by compositional changes in the type of users seeking jobs on AngelList Talent. We find a similar result, with a coefficient of 0.254, reflecting a 29% increase in the size of firms searched for by the same user. Panel B reveals similar findings when examining the likelihood of searching for companies above various size thresholds. For example, based on the coefficient in column 2, with candidate fixed effects, users are 20% more likely to search for large firms with above 500 employees after the crisis compared with the precrisis mean. Overall, the results from Table 2 are consistent with a flight-to-safety channel, in which the preferences of job seekers shift toward more-established firms.¹⁰

We check the validity of the above results in several ways. First, we plot the nonparametric relationship between searched firm size and the date of search in Figure 4. We see a sharp jump in searched firm size that coincides with the outbreak of COVID-19 in the United States (panel A). We also see a sharp decrease in the probability of searching for firms with less than 50 employees and an increase in the probability of searching for firms with more than 500 employees following the COVID-19 outbreak (panels B to E). These sharp changes, together with the lack of a pre-trend, help to alleviate concerns that other non-COVID-19-related events may explain our results. To further alleviate such concerns, we examine whether similar changes were present in 2019. Panel A of Internet Appendix Table A.4 presents the result of this placebo test. We find no statistically significant changes in searched employment size around March 13, 2019. The coefficients are not only insignificant but also economically small. Consistent with this, Internet Appendix Figure A.5 shows a largely flat nonparametric relationship between searched firm size and search date around March of 2019. These results suggest that the flight-to-safety finding documented around the start of the COVID-19 downturn is unlikely to be driven by seasonality or unobserved trends in the data.

Another potential concern is that users may adjust their filters in response to the job postings they see from previous searches. In this case, within-user changes in search parameters may reflect learning about demand rather than changes in preferences. However, it is not clear that a user would actively filter out their preferred type of job from their search results simply because that type of job is less common. In addition, in Internet Appendix Table A.5, we show that our findings remain similar when we restrict our sample to “fresh” searches that are the first search by a user in a day, a week, or a month. The above results show that, during the COVID-19 downturn, users searched for larger firms than they searched for previously. Next, we investigate whether users also searched for larger firms than their current firm. To measure the size of a user’s current firm, we use the LinkedIn data described in Section 2.2. Table 3, panel A, shows that the majority of searches on AngelList Talent are “downsize” searches for jobs at firms that are smaller than the candidate’s current firm. However, after the start of the COVID-19 downturn, there was a significant decrease in the percentage of such downsize searches and a corresponding increase in the percentage of upsize searches. Panel B shows that this pattern remains in a regression framework. First, columns 1 and 2 show that our baseline results continue to hold within the LinkedIn-matched sample. That is, following the COVID-19 downturn, candidates searched for larger firms. Columns 3 and 4 show that they not only searched for larger firms in absolute terms but also for larger firms relative to their current firm. Finally, columns 5–8 show that their probability of searching for a larger firm than their current firm also increased, while their probability of searching for a smaller firm than their current firm decreased. All of these results hold both with and without candidate fixed effects. Overall, we view these results as being consistent with a flight-to-safety channel.

4.2.2 Job applications

Do changes in search preferences translate into job applications? In Table 4, we investigate this question using the specification in Equation (2). The analysis is at the job application level and the dependent variables are the logarithm of size (number of employees) of the firm applied to and an indicator for whether the firm applied to has raised a Series B round or later. Note, we can examine changes in the financing stage of the firms users apply to but we could not do the same for the firms users search for, as users could not search based on financing stage during our sample period. Consistent with our findings on changes in search parameters, in panel A we find that job candidates applied to larger and later-stage firms after the onset of the crisis. Again, these changes hold even within the same candidate over time (panel B). In particular, columns 1 and 4 of panel B show that job seekers applied to firms that were 8% larger and that were 16% more likely to be later stage after the start of the downturn.¹¹ Similar results are not found in a placebo test using 2019 data (panel B of Internet Appendix Table A.4). Further, in Internet Appendix Table A.6, we show that our results are similar even when we include job role fixed effects and startup industry fixed effects, suggesting that changing preferences for certain positions or industries do not drive our results: even within the same role and industry, candidates shift applications toward more-established firms. Thus, flight-to-safety appears to persist from search activities to job applications.

Next, we explore whether the flight-to-safety effects that we document differ for high- and low-quality job seekers. High-quality workers may tend to already be employed at more-established firms and therefore have greater access to an economically resilient outside option. As a result, when the perceived riskiness of less-established startups increases during a downturn, the interest of high-quality candidates may shift away from less-established startups more so than that of low-quality candidates. To examine whether this is the case, we partition candidates according to two characteristics that we can observe in the data: their number of years of experience and an estimated score of their overall quality. Consistent with the resilient outside option hypothesis, we find stronger effects among high-quality applicants. In particular, in panel B of Table 4 we find that high-quality applicants shifted their applications to firms that are 14% larger and 19% more likely to be late-stage.¹²

4.2.3 Heterogeneity

Opportunity cost of joining a startup

The effect of recessions may differ between employed and unemployed individuals. Specifically, recessions may increase the opportunity cost of joining a startup for employed workers, but may have no effect on that opportunity cost for unemployed workers. Unfortunately, we do not observe employment status in our data, as workers generally avoid reporting unemployment spells on their resumes.

While we cannot observe employment status, we can proxy for a worker’s opportunity cost of joining a startup more generally. An unemployed worker may have a low opportunity cost of joining a startup, but an employed worker with a bad job (e.g., high risk of being laid off or low job satisfaction) may also have a low opportunity cost as well. Both would likely search for jobs more intensely due to their low opportunity cost. Thus, search intensity could be viewed as a continuous measure of the opportunity cost of joining a startup. Based on this logic, we would predict that workers with high search intensity would shift less toward established firms during a recession than workers with low search intensity. This is indeed what we find. Table 5, panel A, repeats our baseline specifications, but interacting the PostCOVID indicator with a high search intensity indicator.¹³ Estimate a negative and significant coefficient on the interaction term, meaning that workers shift less toward established firms the higher their search intensity. These findings are consistent with the idea that flight to safety should be strongest among those who already have good jobs.

Perceived downturn severity

Under the flight-to-safety interpretation, the shift toward more-established firms should be more pronounced where the downturn is perceived to be more severe. Some evidence indicates that Democrats were generally more pessimistic about the economy during the COVID-19 downturn than were Republicans (Fetzer et al. 2021; Sheng, Sun, and Wang 2021). Therefore, we use an area’s political orientation as a proxy for the economic expectations of its residents during the COVID-19 downturn. Specifically, we measure a county’s political orientation based on its Democratic vote share in the 2016 presidential election, which was the last presidential election before the pandemic. We then check whether the shift toward more-established firms is stronger in more Democratic-leaning counties. This is generally what we find in Table 5, panel B. Specifically, we find that individuals from more Democratic-leaning counties shift their searches more strongly toward larger firms. Similarly, these individuals shift their applications more strongly toward larger (and later-stage) firms as well. The economic magnitudes are sizable: a one-standard-deviation increase in Democratic vote share increases flight-to-safety by 10% to 50% relative to the mean. While the results are a bit noisy, this is likely because we only observe political orientation at the county level, rather than at the individual level. Among the group of tech workers in our sample, variation in political leaning is likely smaller than that among typical local voters. Nonetheless, these findings are consistent with the idea that there is a greater flight to safety when a downturn is perceived to be more severe.

4.3 Effect on firms

So far, we have documented a significant shift in worker preferences away from less-established startups during the downturn, an effect driven mostly by high-quality workers. How do these changes affect startups? In this section, we examine the effect of the downturn on the quantity and quality of talent flows to startups.

4.3.1 Job applications

We first examine how the downturn affected the volume of job applications to startups. If flight-to-safety is prevalent, we should see a drop in job applications to all startups, as job seekers who would otherwise work for startups turn instead to nonstartup employers. Further, such flight to safety should also drive a wedge between less-established and more-established startups.

Panel A of Table 6 presents the results. The specification is based on Equation (3) and the dependent variable is the number of new applications to a job posting in a given day. We find that, within a firm, the average number of applications to a job posting declined by 14.1% overall during the COVID-19 downturn, when compared with pre-COVID-19 means (column 1).¹⁴ Given that AngelList Talent focuses primarily on entrepreneurial firms, this result in itself is consistent with a flight-to-safety across platforms, where job candidates leave AngelList to search for jobs at more-established firms. We then examine whether this decline is homogeneous across firms in columns 2 and 3. We find that the decline is stronger for less-established startups. For example, startups with fewer than 50 employees saw a 21.4% decline in applications compared with no decline for startups with above 50 employees (column 2). Similarly, job applications to early-stage startups declined by 23.8%, while those going to later-stage startups declined by only 4.7% (column 3). In columns 4–6, we further include job posting fixed effects, therefore exploring the shift in the number of applications within the same job posting. We find similar results with slightly smaller magnitudes. This within job posting analysis rules out the possibility that our results are driven by changes in the types of jobs posted by firms during the downturn.

Next, we explore what type of candidates drive the decline in applications to less-established startups. Specifically, we split the number of applications by candidate experience or quality score at the median. Panels B and C of Table 6 show the results, controlling for firm fixed effects and job fixed effects, respectively. In both panels, we find that the stronger declines in applications to early-stage startups are entirely driven by high-quality candidates (columns 1–2 and 5–6), while low-quality candidates did not apply differentially to early-stage startups (columns 3–4 and 7–8), as indicated by the insignificant interaction terms. These results hold whether we measure candidate quality by experience or AngelList’s proprietary quality score. Moreover, the results are absent in a placebo test using 2019 data (panel C of Internet Appendix Table A.4), suggesting they are not driven by general time trends over these particular months of the year.

How do these application patterns affect the average quality of talent available to startups? Table 7 investigates this, focusing on applicant quality at the application level. Columns 1 and 4 of panel A show that, within a firm, the average applicant experience declined by 3.1% and applicant quality by 6.5% during the COVID-19 downturn. However, such an average decline is entirely driven by less-established startups, as shown in columns 2–3 and 5–6. In particular, startups with fewer than 50 employees experienced a 4.2% decline in applicant experience and an 8.1% decline in applicant quality. Similarly, average applicant experience dropped by 3.8% for pre-Series-B startups and applicant quality dropped by 6.4%. In contrast, more-established startups saw no significant declines in applicant quality and, if anything, experienced slight increases. Panel B shows that these results hold not only within firms but also within jobs, suggesting declining applicant quality is not driven by firms lowering job requirements or canceling higher-skilled jobs (i.e., downskilling in labor demand).¹⁵

Internet Appendix Table A.8 shows that similar results hold using two alternative measures of quality as well: an indicator for candidates with a degree from a world top-100 university, and an indicator for candidates who received an intro request in response to their application.¹⁶ The latter employer-response-based measure captures quality that may be unobservable to us but observable to employers. It also captures firm-specific or position-specific match quality. Using the education-based measure, we find a 6.2%–6.9% decline in the average quality of the applicant pool for less-established startups (columns 1–3). Using the employer-response-based measure, we find a 14.0%–20.8% decline (columns 4–9). Thus, our analysis using these alternative measures points toward an even larger decline in the average quality of the applicant pool.

Figures 5 and 6 show changes in applications and applicant quality graphically. In Figure 5, we see that large and small firms, as well as late-stage and early-stage firms had very similar trends in the number of applications received per job before mid-March. Yet they started to diverge significantly after mid-March, when the COVID-19 crisis started. In particular, less-established startups saw a larger drop in the number of applications per job than more-established startups. Further, all firms saw a precipitous drop in applications around mid-March, suggesting the result is not simply a continuation of a previous downward trend. In contrast, panels A and B of Internet Appendix Figure A.6 show that applications per job in 2019 were largely flat around mid-March 2019 and did not differ between less- and more-established startups.

Figure 6 shows that the average quality of job applicants to small or early-stage startups dropped sharply around mid-March. This holds whether we measure quality by job experience (panels A and C) or AngelList’s quality score (panels B and D). In contrast, applicant experience and quality score did not decline significantly for large or late-stage startups and, if anything, somewhat increased. Further, small and large startups trended similarly in applicant quality measures before COVID-19, and so did early- and late-stage startups. Panels C and D in Internet Appendix Figure A.6 show the placebo graphs for average applicant quality over the same months in 2019. We see no differential trends between the quality of the candidates applying to less and more-established firms. These patterns suggest that our results are not driven by a general downward trend in applicant quality for less-established startups, or these startups being on a differential trend than more-established ones.

Taken together, our results show that workers’ desire to join safer firms during economic downturns has real adverse consequences for early-stage startups in terms of their ability to attract talent. During downturns, job candidates, especially high-quality ones, shift toward more-established firms.

4.3.2 Response to job applications

Does the decline in talent flows to startups affect their actual hiring during downturns? If there was an oversupply of candidates prior to COVID-19 and the post-COVID-19 decline was concentrated among candidates that startups would not hire anyway, then actual startup hiring may not have been affected. Although we do not observe eventual hirings in the AngelList data, we can proxy for them using positive responses by startups to submitted applications—requests for introductions. As discussed earlier, intro requests are required to facilitate further Interactions with candidates, such as interviews or job offers, and act as precursors to eventual hiring.¹⁷ We can therefore examine whether the number of candidates receiving intro requests declined after COVID-19, as well as changes in their average quality. Importantly, this approach allows us to focus on the candidates who startups are most likely to hire, without taking a stance on the characteristics of these candidates (i.e., a revealed preference approach).

We estimate the following equations that are analogous to Equations (3) and (4):

where IntroRequests_fjt is the number of intro requests on job j by startup f on day t, QualityOfApplicantReceivingIntro_ifjt is the years of experience or quality score of candidate i who received an intro request from startup f on job j on day t, and all other variables are the same as those in Equations (3) and (4). We estimate Equation (5) at the job posting × day level and Equation (6) at the intro level.

Because intro requests are firm actions, we take extra care to control for firm demand and to make sure we only capture supply-driven changes. In both equations, we examine within-job-posting changes. We further mitigate concerns about changing demand within an open job posting (i.e., stale job postings) by focusing on jobs that were actively monitored by startups. Specifically, in both equations, we restrict to jobs that firms took some action on, either in the form of a rejection or an intro request. In Equation (5), we additionally restrict to days from a job’s posting date to the last day the firm took action on the job. These restrictions make sure that changes in intro requests are driven by changing talent flows within jobs, rather than firms’ weaker labor demand.

Table 8 presents the results. Panel A shows that within a job, the daily number of intro requests declined substantially after COVID-19 by 29%; such a decline was concentrated among smaller and earlier-stage firms (45% and 38% respectively), and was largely absent among larger and later-stage firms. In contrast, panel B shows no differential change in the experience/quality of the applicants that less-established firms made intro requests to.

Overall, the large decline in intro requests by less-established startups suggests that the decline in applications to these firms is likely consequential for their hiring. To further investigate the effects on hiring, we estimate Equation (3) changing the outcome variable to the number of applications received from eventually hired candidates, identified as those who list the firm they applied to on their LinkedIn resume within 12 months of their application. Internet Appendix Table A.10 shows the results. Consistent with our results on total applications and intro requests, we find that there was a within-job-posting decline in the number of applications received from eventually hired candidates during the COVID-19 downturn. This decline was again mostly concentrated among smaller and earlier-stage firms.¹⁸

4.3.3 Magnitudes

Are the effects on firms large enough that they are likely to actually matter? Overall, we find a substantial reduction in the size of the applicant pool, which is accompanied by a smaller but still meaningful decline in the average observable quality of the pool. Specifically, panel A of Table 6 shows a 16.5%–23.8% decline in the number of applications received by less-established startups during the COVID-19 downturn. Separately, Table 7 and Internet Appendix Table A.8 show that the average observable quality of the applicant pool for less-established startups declined by 3.3%–3.5% when quality is measured based on experience, by 5.2%–7.4% when it is measured based on AngelList’s score, by 6.2%–6.9% when it is measured based on education, and by 14%–20.8% when it is measured based on the response of employers.

Both the quantity and quality margins are likely to matter for startups. In terms of the size of the applicant pool, a startup would likely be adversely affected simply by a decline in applications, even if the average quality of the pool remained constant. First, startups are likely forced to hire more slowly when they have a smaller applicant pool. This is important because long recruiting cycles can hinder growth and consume valuable resources. Second, even ignoring the pace of recruiting, a smaller applicant pool may result in a worse candidate being hired. For example, if the size of the applicant pool declined by 20% uniformly across the quality distribution, that would imply a 20% chance of the startup missing out on the candidate who would have otherwise been its preferred choice. If the decline in applications were skewed toward the higher end of the quality distribution, that would likely imply a greater than 20% chance of the startup missing out on its preferred candidate. Of course, one could argue that the startup might still be able to hire a high-quality candidate, even if that candidate were not as good as the counterfactual candidate it would have hired. However, the measures of quality that we have are very coarse (e.g., years of experience). Thus, even though the two candidates may appear similar on these observables, it seems likely that the counterfactual candidate would still have been a better match for the firm/position. Overall, the magnitudes in Table 6, Table 7, and Internet Appendix Table A.8 jointly suggest that the COVID-19 downturn had a substantial impact on the ability of less-established startups to attract talent.¹⁹

4.4 Other potential concerns

4.4.1 COVID-19 was not a discrete event

Concerns about COVID-19 did not arrive all at once. Indeed, they were rising even before a state of national emergency was declared on March 13, 2020. Consistent with this idea, Figures 4, 5, and 6 show that trends were already changing in late February. Recognizing this, we also examine the correlation between continuous time-varying treatment measures and worker interest in more-established firms. First, we show in Internet Appendix Table A.12 that our main results are similar if we use the cumulative number of local COVID-19 cases in a job seeker’s state as an alternative treatment variable. The number of local cases captures not only the staggered arrival of COVID-19 across states but also the differential escalations of the pandemic that may have shaped job candidates’ economic expectations.²⁰ Second, we also directly examine economic expectations as proxied for by a daily survey-based index from Civiqs.²¹Internet Appendix Table A.13 shows that our main results are similar when we use this index as a time-varying continuous treatment variable. Internet Appendix Figure A.4 further shows that changes in searched firm size and application volume coincide very closely with the drop in economic expectations. These results again suggest that lower economic expectations correspond to greater interest in established firms.

4.4.2 Sample selection associated with within-candidate analysis

Another potential concern is that our within-candidate estimates are based on individuals who searched/applied for jobs both before and after the start of the downturn. This could introduce potential sample selection issues, as such workers may be the ones who have a difficult time getting hired. For example, it is possible that workers with a preference for a job at an early-stage startup eventually shift to searching and/or applying for jobs at more-established startups because of a lack of success in getting hired, rather than because of an economic downturn. However, we do not find similar results in 2019, when similar sample selection issues exist but without a downturn. In addition, such sample selection concerns would not explain the similar results we find without individual fixed effects, the heterogeneity we find between high-quality and low-quality candidates, or the effects that we find within job postings, which allow for compositional changes in job candidates.

5 Short Blip or Long-Lasting Effects?

Our data from AngelList only allow for analysis through mid-2020 (May 2020 for searches, June 2020 for applications); however, Figure 2 shows that consumer confidence remained low until almost March 2021. Ultimately, the goal of our paper is not to evaluate the full impact of the COVID-19 downturn on the startup labor market, but rather to test the flight-to-safety hypothesis. We do so using a short window where there was a drastic shift in job seekers’ expectations and for which we have data. Nonetheless, showing that the effects persisted for as long as the shift in expectations persisted, would provide nice confirmatory evidence.

To try and address this issue as well as possible without additional data from AngelList, we use aggregate hiring data for the firms in our sample. These data, which are derived from LinkedIn records, are discussed in Section 2.2. We are able to match 82% of the firms in our AngelList sample to LinkedIn. The downside is that, unlike in the AngelList data, we can only observe firm hirings in the LinkedIn data and not candidate searches or applications.

Following the U.S. Census, we define a firm’s hiring rate in month t as $inflow countt0.5(empt+empt−1).$Figure 7 shows the impact of COVID-19 on the monthly hiring rates of the small and large firms in our sample (i.e., employment$≤$50 or employment >50 as of February 2020). We find that the small and large firms in our sample exhibit parallel trends in hirings before COVID-19, but the small firms experienced a greater decline in hirings during the COVID-19 downturn, with the effect persisting until March 2021, a year into the pandemic.²²Internet Appendix Table A.14 confirms these findings in a regression framework and shows that they are not sensitive to the sample period start month or additional fixed effects.²³ Of course, unlike in our main analysis, a shortcoming here is that hiring rates could reflect both changes in labor supply and labor demand around the COVID-19 downturn. However, these results are at least suggestive of long-lasting effects and provide a sort of “out of sample” test. Moreover, these results allow us to confirm that firms that seem to be garnering less interest on AngelList Talent are indeed hiring less overall.

6 Flight to Safety?

6.1 Less-established firms less safe during recessions historically?

Having established that job seeker interest shifted away from less-established firms during the COVID-19 downturn, we check whether such firms have indeed been less safe during recessions historically. Table 9 shows the results of this analysis from 1979 to 2019 using BDS (Business Dynamics Statistics) data from the U.S. Census Bureau.²⁴ Panel A shows that historically, both smaller firms (employees$≤20$ or employees$≤100)$ and younger firms (age$≤5)$ have been more sensitive to NBER recessions in terms of their death rates and job destruction rates.²⁵ Panel B further shows that, within young firms (i.e., startups), smaller firms have been more sensitive to recessions than larger firms. Thus, the historical evidence suggests that the job seekers in our sample would be correct in viewing larger and later-stage firms as being safer during a recession. These results are also consistent with findings in the literature (see Gertler and Gilchrist 1994; Fort et al. 2013; Sedláček and Sterk 2017; Pugsley and Sahin 2019; Crouzet and Mehrotra 2020).

6.2 Alternative proxies for safety

Our hypothesis is that job seekers use a startup’s size and stage as proxies for its safety. If that is the case, job seekers also should be attracted to startups with other signals of safety. If they are only attracted to size and stage, but not other signals of safety, that would call into question the flight-to-safety interpretation.

Therefore, we also explore two additional proxies for safety: (1) an indicator for whether a startup was funded by a “top-tier” VC (as categorized by AngelList) and (2) an indicator for whether a startup was funded recently (i.e., in the past 6 months). These measures were discussed in more detail in Section 2.1. The former could be viewed as a measure of startup quality and the latter could be viewed as a measure of financial health. Table 10 shows that job seeker interest also shifted toward firms with both of these characteristics during the COVID-19 downturn as well, and that such a shift was stronger among higher-quality candidates. These results are consistent with a flight-to-safety interpretation.

A related question is whether our two main proxies for safety—size and stage—capture independent dimensions of safety, or whether one subsumes the other. We show in Internet Appendix Table A.9 that all our application results go through if we control for firm size when examining financing stage, or when we control for financing stage when examining firm size.²⁶ Hence, both size and financing stage independently matter. In other words, during a downturn, there is a preference for larger firms regardless of financing stage, and a preference for later-stage firms regardless of firms size. Regardless of these findings, however, the contribution of our paper does not depend upon each proxy capturing an independent dimension of safety.

6.3 Potential confounds

We view non-safety-related factors that may correlate with our proxies for safety as potential confounds to the flight-to-safety interpretation of our results. Below we discuss such potential confounds.

6.3.1 COVID-19-related potential confounds

Given that we study the period surrounding the start of the COVID-19 pandemic, there may be COVID-19-related confounds that are not related to financial safety but that correlate with our proxies for financial safety. For example, it is possible that more-established firms were better able to offer COVID-19-specific benefits during the downturn, such as remote work, flexible hours, or strict health protocols. Employees may be attracted by these benefits rather than the greater financial safety offered by established firms.

Beginning with remote work, it is worth noting that at the start of the pandemic essentially all tech firms went remote. Thus, in the short run, differential availability of remote work would not provide an impetus to search for jobs at large firms. In addition, over the longer run, it has actually been the less-established firms that have been most open to embracing remote/flexible work arrangements.²⁷

Nonetheless, to ensure that our results are not driven by a flight to remote work, we examine whether the shift in interest toward larger firms occurred even among those searching specifically for in-office jobs. If workers searching for in-office jobs after COVID-19 searched for larger firms than those searching for in-office jobs before COVID-19, that would suggest that the reason for their shift did not have to do with greater perceived availability of remote work at larger firms. Consistent with this idea, Internet Appendix Table A.15 (columns 1 and 2) shows that in-office job searches did indeed shift toward larger firms.²⁸ Similarly, if workers searching for remote-only jobs after COVID-19 searched for larger firms than those searching for remote-only jobs before COVID-19, that would also suggest that the reason for their shift did not have to do with greater perceived availability of remote work at larger firms. We find this as well in Internet Appendix Table A.15 (columns 3 and 4).

The results for the remote-only search subsample also point away from a flight to other COVID-19-specific benefits, such as flexible hours or better health protocols. For remote jobs, the flexibility and health aspects were likely similar, regardless of firm size (e.g., ability to take breaks for household errands; no coworkers physically present). Thus, the increased interest in remote large-firm jobs relative to remote small-firm jobs, is more likely attributable to a greater perceived financial safety for large firms, rather than a greater perceived ability for them to offer flexibility or strict health protocols.

In addition, much of the demand for flexibility among workers during COVID-19 was likely driven by the increased childcare burden on households. It is generally thought that this increased childcare burden fell disproportionally on women (Del Boca et al. 2020; Alon et al. 2022). Therefore, if our main results were driven by a flight to flexibility, we might expect the shift in interest toward larger firms to be stronger among women. Yet, we find in Internet Appendix Table A.16 that there was actually a similar shift in interest toward large firms among women and men.

6.3.2 Non-COVID-19-related potential confounds

One potential confound not related to COVID-19 could be compensation. In particular, large firms may perform no better through recessions, but they may tend to offer higher pay during such times. In particular, it is possible that large firms tend to offer a greater proportion of their compensation in the form of cash, while small firms rely more on equity compensation, the value of which may be lower during recessions. To test whether changes in compensation preferences drive our results, we rerun our baseline regressions including detailed compensation controls, focusing on searches that specified compensation and applications to jobs with compensation information. Internet Appendix Table A.17 shows the results. The odd-numbered columns first confirm our baseline results in these subsamples without compensation controls. Further, holding fixed the salary and equity ranges specified by users in their searches, we continue to find an increase in the size of the firms searched for following the start of the COVID-19 downturn (column 2). In addition, holding fixed the salary and equity ranges advertised by firms, we continue to find an increase in the size (and stage) of the firms applied to during the downturn (columns 4 and 6). In both cases, the magnitudes remain similar with and without compensation controls. Thus, our baseline results do not appear to be driven by more-established firms offering better or more cash-based compensation during downturns.

Another potential concern is that job postings may have revised their advertised compensation during our sample period, as we only observe advertised compensation as of the date the data were extracted. To make sure that our application results are not driven by downward revisions to the compensation advertised by less-established firms in their job postings, we conduct two additional robustness tests in Internet Appendix Table A.18. First, we restrict to the 84% of job postings that were not refreshed during our sample period. A job posting refresh is associated with some content revision, though not necessary about compensation. Second, we restrict to job postings that provide no information on compensation. These job postings are not subject to concerns about compensation revisions since they do not disclose compensation. In both subsamples, we find results similar to our main results.

7 What Drives the Flight to Safety?

Our primary goal in this paper is to investigate whether employee interest shifts toward safer firms during recessions, leading less-established firms to experience increased difficulty in recruiting. The results presented thus far are consistent with such effects. Next, we consider why employee interest might shift toward safer firms during recessions. While this is harder to nail down, it is nonetheless interesting to explore. There are at least two reasons that employee interest might shift toward safer firms: a change in worker risk aversion or a change in the relative safety of firms. Specifically, workers may become more risk averse during recessions, such that their desire for safety increases. Alternatively, workers may perceive the safety offered by more-established firms relative to less-established firms as being greater during such times. These two potential drivers are not mutually exclusive, so both could be operating simultaneously.

Each potential driver would also have somewhat different implications. Under a pure “change in risk aversion” interpretation, less-established firms would have a harder time recruiting during a recession, even if their prospects remained exactly the same. These increased recruiting challenges would create a decline in prospects where none existed before. In contrast, under the “change in relative safety” interpretation, less-established firms would have a harder time recruiting because their prospects were already declining. In this case, the increased recruiting challenges would amplify an existing decline in firm prospects. Both interpretations are interesting and would generate a negative causal effect of recessions on recruiting for less-established firms, consistent with what we find.

To learn more about what drives the flight to safety we have documented, we examine whether the shift in employee interest toward more-established firms varies with startup quality. If job seekers primarily avoid less-established firms of low quality during recessions, that would be more consistent with the “change in relative safety” interpretation, as these firms likely suffer the largest decline in their prospects. However, if job seekers also avoid less-established firms of high quality during recessions, that would be more consistent with the “change in risk-aversion” interpretation, as these startups likely experience a smaller decline in their prospects.

Of course, measuring startup quality is challenging. To do so, we again make use of the top-tier VC measure discussed in Section 2.1. Startups funded by top-tier VCs are likely to be high quality because skilled VCs with access to many potential deals chose them. They are also likely to be high quality insofar as they benefit from the advice and connections of these reputable investors. Further, top VCs have deeper pockets to support their portfolio companies through downturns. As an alternative measure of quality, we also categorize startups by whether or not they raised funding recently (i.e., in the past 6 months).

Table 11 examines how the shift in applications toward more-established firms depends on startup quality. High-quality startups are the omitted category. The coefficient for the uninteracted $PostCOVID$ variable therefore represents the effect of the COVID-19 downturn on the size of the high-quality firms users applied to. As can be seen, even for applications to high-quality firms, we find a significant shift from small high-quality firms to large high-quality firms. Moreover, we estimate a statistically insignificant coefficient on the interaction term $PostCOVID×LowQualityStartup$⁠. This means that for applications to low-quality firms, we find a similar shift from small low-quality firms to large low-quality firms. If anything, based on the point estimates, the shift toward large firms is greater for high-quality firms than for low-quality firms, although the difference is not statistically significant. We find similar results when examining stage instead of size. These results seem less consistent with the “change in relative safety” interpretation, as it is unlikely that the prospects of high-quality startups declined by the same amount as those of low-quality startups. Rather, they point more toward the “change in risk-aversion” interpretation, as increased risk-aversion might drive job seekers away from less-established startups regardless of their quality. While this may be an imperfect test, at a minimum, these results show that the shift in interest away from small firms was not confined to the types of firms that typically fail anyway.

8 Conclusion

Young firms are central to innovation and productivity growth. Yet, their ability to grow and innovate depends crucially on their ability to attract high-quality talent, potentially from established firms. Before achieving standardization, human capital is fundamentally intertwined with the success of startups. In this paper, we show that young firms’ ability to attract talent suffered during the most recent economic downturn, namely, the COVID-19 crisis. Using unique job search data as well as within-candidate and within-job analysis, we show that job seekers pivot to larger and more mature firms when a downturn hits. This leads to a decline in talent flows to startups, especially to less-established ones, and ultimately affects their hiring. Importantly, such flight-to-safety is stronger among higher-quality candidates, leading to a deterioration in the quality of human capital available to small, young startups. Our results provide a novel mechanism through which economic downturns negatively affect entrepreneurship. More broadly, our study highlights the importance of labor market frictions in understanding the procyclicality of entrepreneurship.

Acknowledgement

We thank Kunal Mehta for data assistance. We are grateful for helpful comments from Holger Mueller (the Editor), two anonymous referees, Paul Gompers, Issac Hacamo, Jessica Jeffers, Song Ma, David Matsa, Ramana Nanda, Paige Ouimet, Sergios Salgado, Antoinette Schoar, Elena Simintzi, and Chris Stanton and seminar participants at NBER Entrepreneurship, RCFS Winter Conference, SFS Cavalcade, MFA, FOM Virtual Corporate Finance Seminar, Junior Entrepreneurial Finance/Innovation Lunch Group, Junior Corporate Finance Workshop, CMU, Tsinghua, CUHK Shenzhen, LSE, Rochester, University of Illinois Chicago, Indiana (Kelley), HBS, Harvard Law School, UVA Darden, and Virginia Tech. Supplementary data can be found on The Review of Financial Studies web site.

Footnotes

Author notes

Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online.

References