We’d like to share the new draft of our paper titled “The Impact of COVID-19 on Small Business Dynamics and Employment: Real-time Estimates with Homebase Data”, which is available here: Measuring Small Business Dynamics and Employment with Private-Sector Real-Time Data by Andre Kurmann, Etienne Lalé, Lien Ta :: SSRN. In it, we use Safegraph data together with data from Homebase, Google and Facebook, to study small business dynamics and employment and the effects of emergency government loan programs during the pandemic. At a methodological level, we make a key contribution by showing how to address an important challenge when working with establishment-level data, which is namely to disentangle sample turnover from business openings and closings. Our approach is to combine information on business activity from all these data sources (Google, Facebook, Safegraph, matched to the Homebase data), and benchmark the data against (pre-pandemic) administrative statistics on business births and deaths. The distinction between sample turnover and business closings/openings turns out to be critical to estimate the effects of the pandemic on small business activity. In particular, we find four key results:
Employment of small businesses in four of the hardest hit service sectors contracted more severely in the beginning of the pandemic than employment of larger businesses, but small businesses also rebounded more strongly and have on average recovered a higher share of job losses than larger businesses;
Closings account for 70% of the initial decline in small business employment, but two thirds of closed businesses have reopened and the annual rate of closings is just slightly higher than prior to the pandemic;
New openings of small businesses constitute an important driver of the recovery but the annual rate of new openings is only about half the rate one year earlier ;
Small business employment was affected less negatively in counties with early access to loans from the Paycheck Protection Program (PPP) and in counties where Federal Pandemic Unemployment Compensation (FPUC) was more generous relative to pre-pandemic earnings of likely recipients, with business closings accounting for a large part of these two effects.
These results dispel the popular notion that small businesses continue to suffer more from the pandemic than larger businesses. At the same time, our analysis suggests that PPP and FPUC helped to significantly mitigate the negative effects of the pandemic for small businesses by, respectively, alleviating financial constraints and stimulating demand for local services.
Happy to hear your comments and to answer any questions about our work!!
This topic was automatically generated from Slack. You can find the original thread here.
I’m always trying to learn more about other non-SafeGraph datasets that we can refer our members to. I had never heard of Homebase, but would love to learn more since other members could find payroll data very useful in their research. What does the process look like to getting access to that data? Is that still an initiative they’re running? I wasn’t able to find too much except a few dashboards. I’d love to share this info next time it comes up in general-discussion.
Etienne and Andre - for context, a recent publication from these Community members was featured on the Berkeley Haas Newsroom. Their research looked at how big chain closures affected closure decisions of smaller businesses located nearby. Here’s a recent post highlighting their work. Would love to hear their thoughts on this newest draft!
I’m curious about your statement: “dispel the popular notion that small businesses continue to suffer more from the pandemic than larger businesses.” Do small businesses still have a lower level of employment than larger businesses compared to the pre-pandemic estimates? Are 1/3 of large businesses still closed? Just curious.
In the paper we discuss at length the sources of difference between our estimates and the BLS’s – beyond the fact that the BLS covers all business sizes while our estimates are for businesses with fewer than 50 employees. But yes, small businesses in L&S according to our estimates have more than recovered (and part of it is explained by the opening of new businesses).
As to your other question, we do look at the effects of policies in the last section of the paper, and find that the recovery is partly helped by PPP loans and FPUC:
For PPP loans, we find evidence that their temporary exhaustion in April 2020 occurred at a critical moment when many small business owners had to decide whether to continue operating and hope for loan relief from the government or cut their losses and close shop. This is evidence that PPP loans have been instrumental in alleviating the financial constraints faced by small businesses;
For FPUC, essentially we find that it stimulated local demand in a way that, in net, has helped small businesses cope with the crisis. We don’t rule out dinsincentive effects, i.e. it might be that bars and restaurants have had a hard time hiring workers due to the generosity of FPUC. But if these effects are present, they are dominated by the strong positive effect on consumer spending.
I really liked this paper, and think the point about sample churn is well-made and convincing. This seems like a very important thing to measure accurately, and your paper does a lot to convince me that the recovery is going (shockingly) well.I did have one major concern about measurement of demand shock, and a few clarification questions.
I wasn’t quite sure what the actual impact of other papers using Homebase exits as firm closings was (and think you should emphasize where you have improved on other work!). You mention how your results are different than Granja et al, but I wasn’t sure if that was the only paper that incorrectly codes firm exits.
Have you considered Roth 2021’s argument about tests for pre-trends often being underpowered? I might add some additional testing to show that even if there were violations of pre-trends, it is unlikely to affect the results (e.g. as suggested by Rambachan and Roth 2020).
You use UI received vs. prior income as a measure of the demand shock. I think this is a promising measure, but I was somewhat concerned about measurement error in how much money unemployed people actually had to spend. UI systems often had trouble paying out, particularly during the early part of the pandemic, and the % of people who qualified for UI that received it in a timely manner varied substantially by state and time period. You might end up with odd lumpy distributions, as UI systems that were back-logged processed previous payments. (Since payout varies over time within a county, fixed effects would not be sufficient to capture this.)
Would it be possible to perhaps weight your measure of county replacement rate by % of UI claims actually approved, or some similar measurement of how much of the UI money recipients would eventually receive was received in that country-week?