Who Creates Jobs? Small vs. Large vs. Young

Ron Jarmin of the U.S. Census Bureau noted the persistence of the debate over whether small businesses are responsible for job creation. There are two camps in the debate: those who contend that most new jobs are created by small businesses and those who argue that this is not true. Jarmin suggested there is truth to both positions and presented data from a longitudinal (1992–2005) database of private-sector, non-farm business establishments with firm identifiers.

The data included the size of businesses and their ages. Size was constructed by aggregating establishment employment numbers to firm totals, and firm age was defined as the age of the oldest establishment at the time of firm birth. Spin-offs from existing businesses were not treated as new or startup businesses.

If one looks at only firm size, most new jobs are in small businesses, but Jarmin’s work stresses the contribution of firm births, or “startups,” to job growth and the distinction between gross and net job creation. As he described it, there is a huge churn in firms and jobs and an “up or out” dynamic for young businesses. Many young businesses fail, but those that survive contribute to dynamic growth.

A key to Jarmin’s analysis is the relationship between firm size and age. New or startup firms tend to be small, while older firms tend to be larger. When one controls for the age of firms, there is a positive relationship between firm size and job growth.

Looking at data for 2005, Jarmin noted that among the largest firms, the biggest job gains were from the oldest firms, while among the smallest firms, most growth was among the youngest. This makes sense, as startups (businesses in their first year) cannot lose jobs (relative to the previous year). And because there are so many startups in the course of a year, small firms account for a large number of new jobs. Among small firms, job growth was greatest during the startup year, with the number of jobs added dropping sharply in years two and beyond. Again, there is an up or out dynamic, with the failure of many new businesses resulting in the destruction of many of the new jobs they contributed. One can contrast the gross versus net creation of new jobs.

Jarmin then described the challenge of “picking winners,” or establishing policies to promote job growth. While startups and surviving young firms contribute disproportionately to job growth, idiosyncratic factors seem to dominate in the determination of which ones survive—factors that are not observable or predictable for policy purposes.

Jarmin concluded by suggesting we need a more nuanced view of small businesses and their contribution to job creation. It is not just the size of businesses that matters, but their age. Another issue is the quality of jobs, and Jarmin noted that we need to look beyond the simple counting of new jobs to the kinds of jobs being produced by younger firms, the kinds of workers in these jobs, and the long-term labor market outcomes.

Local Employment Dynamics: Synthetic Data for OnTheMap Version 4

Jeremy Wu of the U.S. Census Bureau described OnTheMap as an online dynamic mapping and reporting tool for the bureau’s Local Employment Dynamics (LED) data. He also gave an overview of the integrated, synthetic data underlying the product.

The first OnTheMap release was in 2006, covering 14 states and data from 2002–2003. The product has grown through successive releases, covering 47 states and data for 2002–2008. A December 2010 release will cover 47+ states with data for 2002–2009.

OnTheMap allows one to select where workers live or work and report characteristics such as age, earnings, and cross-state flows. The base unit is the census block, and the product features innovative disclosure protection. Wu showed a screenshot of a Las Vegas–area map shaded where construction/manufacturing workers are employed and side-by-side maps, one showing workers in blocks near the Strip and the other expanding to show where those workers live. Data tables are presented with the maps, and while the examples illustrated census blocks, one could show data by other geographies, such as ZIP code or traffic analysis zones.

Turning to the data, Wu noted that censuses date to ancient Rome and China, but sampling was first discussed in 1895. Even then, the idea was not well received, even among statisticians, who clung to the notion that there was no substitute for a complete count. The debate went on for decades, and it was not until 1937 that the bureau developed sampling techniques to measure unemployment during the Depression. Sampling was then introduced to the decennial censuses and is now used in other surveys such as the Current Population Survey and American Community Survey.

With the introduction of sampling, it became clear that a 5% random sample is better than a 5% nonrandom sample and the field of mathematical statistics was born. But Wu noted that the field of sample surveys has not lived happily ever after, as computers and administrative records databases have released a flood of data and surveys are increasingly hampered by declining response rates, increasing labor costs, and confidentiality concerns. As recently as the 1990s, there was concern that we could either have access to microdata or confidentiality protection, but not both.

However, Wu described an LED approach that provides both with a design that involves record linkages, noise infusion, imputation, synthetic data modeling, and measures of goodness and quality. A slide diagramming how the synthetic data are prepared conveys its complexity. And with the workplace/residence data comprising an origin/destination matrix for 8 million census blocks (8 million times 8 million), the underlying database is huge.

Wu described OnTheMap and its data innovation as the latest development in sampling. It took decades for sampling to be accepted; Wu said he hopes it will not take so long for this innovation to become accepted and widely used.

Asked if these data have been accepted for academic research, Wu noted that academic researchers have been involved in their development, but the data are not yet widely used. There was agreement that more formal measures of goodness and quality are needed for the data to become more widely accepted in academic research.