The Mismatch Mystery: Searching for the “Skills Gap” in North Carolina

Employers are posting job vacancies at a rate higher than we’ve seen in 15 years.  But if there are so many jobs available, why are measures of hiring and job-finding stuck below expansionary highs, while the unemployment rate and long-term joblessness remain elevated above levels seen during previous boom times?

There is widespread recognition that America’s labor market has gotten less efficient at matching unemployed job-seekers to job openings, but little agreement about the source of the problem.  Some employers and HR consultants blame the problem on a mismatch between the attributes of job seekers and the needs of employers (including a “skills gap”), while a large body of academic research casts doubt on this explanation.

This article reviews existing evidence from North Carolina for the purpose of quantifying the extent of labor market mismatch in our state.  Our findings broadly agree with the academic literature—there is currently no evidence that mismatch has worsened over the past nine years when looking at the occupational or geographic composition of our labor market.

However, we need more data on the specific skills demanded and supplied in the labor market in order to understand the full extent of mismatch (or the lack thereof.)  Moreover, despite the decline in mismatch we document here, we also show that some occupations continue to face high levels of excess unemployment while others experience relative shortages of labor.  Workforce professionals and economic developers can increase hiring and reduce joblessness by focusing on these particular areas of imbalance.

The Problem – An Inefficient Labor Market

Let’s start with some data points that illustrate the problem.  The job vacancy rate¹ and unemployment rate in North Carolina are both higher than prior to the 2007-2009 recession.  Typically, a higher degree of job vacancy leads to more hiring and, ergo, lower unemployment, meaning these indicators should be moving in opposite directions.  However, the unemployment rate remains elevated despite continued growth in job openings.

An increase in job vacancy without a corresponding decrease in unemployment is a sign of inefficiency in the labor market’s matching process.  Employers are now less likely to hire jobseekers in a given month than they were during the previous expansion, resulting in a relatively higher level of unemployment.  Joblessness that persists despite strong employer demand for labor is often described as “structural unemployment”.  

Economists use a couple of different tools in order to measure matching efficiency and structural unemployment.  One of the most common is the Beveridge curve, in which unemployment rates and job vacancy rates are plotted on an X-Y axis.  Here we examine data for the broader United States:

The light blue dots represent monthly observations between December 2000 and November 2009.  During this time, unemployment and vacancies in the U.S. followed a clear, curvilinear relationship.  When the economy was expanding, job vacancies increased and the unemployment rate decreased.  When the economy was in recession, job vacancies decreased and the unemployment rate increased. 

After 2009, this relationship shifted outward, as indicated by the dark blue dots.  The two highlighted months (January 2001 and July 2015) each saw job vacancy rates of 3.9%. The unemployment rate in July 2015 was 5.3%, over a point higher than the January 2001 unemployment rate of 4.2%.  This elevated level of unemployment in the face of strong labor demand signals a decline in matching efficiency.

We can also quantify the impact of matching efficiency in the U.S. with the help of a regression model.  Here we estimate the proportion of unemployed jobseekers finding work in a given month (the “hiring rate”) with and without the decline in matching efficiency.  Despite the strong recovery in employer demand, unemployed jobseekers are now 7.1 percentage points less likely to be hired than they would be if matching efficiency remained at pre-recession levels.

A mismatch between unemployed job seekers and available jobs may be one explanation for this slowdown in the labor market’s matching process.  For the remainder of this article we test this theory by quantifying several sources of mismatch in North Carolina.  (As we describe at the end of this article, there are many potential sources of inefficiency in the labor market other than mismatch.)

Occupational Mismatch

It might be that unemployed workers have the wrong job experience for the jobs currently available.  For example, after the recession induced large and persistent job losses in North Carolina’s manufacturing sector, displaced production workers in our state may find fewer opportunities for employment in their previous profession.

We calculate a commonly-used index to determine the extent of occupational mismatch in our state, based on the most recent occupation of employment reported by workers.  In plain English, this index can be interpreted as measuring the proportion of hires that were prevented due to a mismatch between jobseeker’s most recent work experience and the types of jobs actually available, with 0.0 representing a perfectly-matched market and 1.0 indicating a completely mismatched market. 

In our first example, we assume that jobseekers search for work within occupational labor markets defined at the 2-digit SOC level.  Here we find—contrary to expectations—that occupational mismatch in North Carolina has not increased over the past nine years.  Rather, our labor market appears to be substantially better-matched than it was prior to the 2007-2009 recession.

Although the index reported above is easy to compute and straightforward to explain, it provides a somewhat misleading picture of mismatch in the labor market.  For one, the measure of job vacancies we use is based on online job advertisements and as a result might under- or over-represent certain occupations which utilize internet job boards to different extents.  In addition, the hiring process is typically more efficient in certain occupations, such as construction; we might be willing to accept a higher number of unemployed in those markets at a given point in time if those jobseekers are able to find work relatively faster.

It might also be misleading to limit our comparison to individual occupations.  It is relatively common for unemployed jobseekers in the U.S. to find work in their previous occupation of employment; this was the case in 47% of hires during the period studied.  But rather than being limited to only one occupation (e.g. Production), workers often have transferable skills than enable them to seek work in similar occupations (such as Construction and Extraction); the unemployed found work within their broad, skill-based group 64% of the time during the period studied.

In the following graph, we report an alternative measure of mismatch that adjusts for sector-specific differences in online job ad coverage and matching efficiency, and we apply this measure to simplified labor markets consisting of broad, skill-based occupation groups (such as “manual/routine” jobs.)

The dashed line showing our alternative measure of mismatch across broad occupation groups provides the most conservative and, arguably, most realistic assessment of the degree of occupational mismatch in North Carolina.  This line shows that not only has occupational mismatch declined from its recessionary heights, it was not much of a factor even before the recession.

Geographic Mismatch

Another possibility is that unemployed workers are living too far away from where the available jobs are located.  For example, as employment grows ever-more concentrated in large urban counties such as Wake and Mecklenburg we might see a larger share of displaced workers living in outlying counties with fewer job opportunities.

We follow a similar approach as before in constructing an index of geographic mismatch in North Carolina, looking first at single-county labor markets and then at broader commuting areas (using the federal government’s Metropolitan Statistical Area definitions) and using our alternative calculation.  We again find that—contrary to expectations—geographic mismatch has declined since prior to the recession. 

This does not mean that the economic disconnect between thriving and distressed areas in our state has receded.  Rather, this tells us that the unemployed and the available jobs are increasingly more likely to be located in the same county, possibly reflecting the shifting of our population to large and prosperous localities.  The typical jobseeker in our state is relatively closer to where the job opportunities are than at any time during the past nine years.  (Future articles from LEAD will explore how labor market mismatch has shifted within and across North Carolina’s various regions.)

In sum: we currently have no evidence that occupational or geographic mismatch in North Carolina’s economy has worsened since prior to the recession. 

However, this is not the last word on the subject.  Although we have data quantifying labor supply and demand within occupations and geographic areas, we lack quality data on the specific technical or “soft” skills offered by job candidates and sought by employers.  This is a blind spot we are trying to remedy.  LEAD’s Employer Needs Survey aims to fill this gap by reporting on the challenges faced by businesses with hard-to-fill positions, and efforts are currently underway to leverage administrative data from the Common Follow-up System to better document the education and training completed by workers in North Carolina.

Workforce and economic development practitioners should also note that, regardless of the trend over time, some level of mismatch is a permanent feature of the labor market.  There is always a degree of structural unemployment in our economy, and some portion of this unemployment is always due to mismatch.  

Occupational Labor Markets

Practitioners can identify where the greatest needs are by consulting data on particular occupational labor markets.  Here we assess the degree of excess unemployment (or “labor surplus”), defined as the number of unemployed minus the number of job vacancies, for each occupation.

Occupations which are often categorized as routine and/or manual jobs—such as Production; Sales and Related; and Building and Grounds Cleaning and Maintenance—have the largest amounts of excess unemployment.  Meanwhile, the markets for cognitive/nonroutine jobs (or “knowledge jobs”)—such as Healthcare Practitioners and Technical; Architecture and Engineering; and Computer and Mathematical—have a relatively small labor surplus.  From a workforce planning perspective, the greatest need for assistance exists among jobseekers without recent experience in a cognitive/non-routine field.

Looking at rates rather than levels illustrates the labor supply and demand picture from a somewhat different perspective.  The “slack rates” depicted below show the relative amounts of excess unemployment within each market, defined as the number of unemployed-per-job vacancy.  The cognitive/non-routine labor market is tighter than other markets with only 0.7-1.8 unemployed competing for each available job on average, meaning that jobseekers in this field face relatively less competition from other jobless workers.

Readers should note that these data reflect a single point-in-time snapshot.  Workers don’t stay unemployed forever; there is generally some “churn” as the unemployed find jobs (or leave the labor force) and others take their place.  Moreover, labor supply and demand affect each other over time.  If the shortage in cognitive/non-routine occupations causes an influx of new jobseekers, employers in those fields might choose to post even more job vacancies as the talent pool deepens.

Despite these considerations, the data presented here depict in clear, simple terms the relative areas of need in our labor market.  Workforce professionals should focus their efforts on retraining displaced workers and helping students and trainees prepare for available jobs in the cognitive/non-routine job market where there are better prospects for employment.  Economic developers, on the other hand, might want to market our surplus in certain occupations (such as Production) as an underutilized pool of talent for employers seeking to set up shop or expand in North Carolina.   Either way, both groups can use the information presented here to improve the effectiveness of their efforts.

North Carolina’s workforce and economic development communities share a common goal of increasing hiring and reducing joblessness in our state.  NC Commerce offers a broad suite of support services for helping our partners in the field attain these goals.  Our division will continue to provide data, information, and analysis—such as presented here—to help these partners target their limited resources where they are needed most.

Other Causes of Labor Market Inefficiency

Before we finish this installment, we again note that mismatch is not the only source of inefficiency in the labor market.  Any factor that prevents job seekers and employers from forming a match can weigh on the efficiency of the matching process, delay hiring, and extend unemployment spells.  Decision-makers should understand that these factors, although challenging to quantify, may play an important role in hindering matching efficiency and, thus, hiring.  For example:

• Markets may fail to “clear” if buyers and sellers cannot agree on a price.  The unemployed will remain jobless and employers will struggle to find job candidates if wage offers are too low.  The fact that we are currently in an environment of relatively high unemployment, relatively low hiring, and relatively meager wage growth makes this theory intuitively credible.
• Markets can also fail to clear if buyers and sellers lack adequate information about the future and are unable to plan accordingly.   Post-recession uncertainty about future changes in economic policy and demand conditions might be causing employers to hesitate to hire and making applicants hesitate to accept job offers. 
• Certain jobseekers face unique disadvantages in the labor market, due either to their qualities as job candidates, ineffective job search practices, or bias on the part of employers.  Earlier work by LEAD demonstrated that declines in matching efficiency during the start of the recovery were accounted for by an elevated number of long-term unemployed, echoing previous findings from economists at the Federal Reserve Bank of Boston.
• Matching efficiency can also be impacted by changes over time in matching technology.  Online job boards have reduced the cost of advertising job vacancies and applying for jobs.  The effect of job posting on matching efficiency is ambiguous; while the availability of more information on jobs may result in faster hiring, the lower cost of job advertising and job search may also allow employers to leave openings up longer before hiring and jobseekers to search for longer before accepting an offer.

General disclaimers

Data sources cited in this article are derived from surveys and administrative records and are subject to sampling and non-sampling error.  Any mistakes in data management, analysis, or presentation are the author’s.

---

Technical appendix

I.Data

Throughout this article we use the number of unemployed (jobless individuals who are able, available, and actively seeking work) as a measure of labor supply, the number of job openings as a measure of labor demand, and the number of unemployed entering employment in a given month as a measure of hires (i.e. matches). 

Data on the number of unemployed by occupation are tabulated from public-use Current Population Survey (CPS) microdata from the U.S. Census Bureau², and the number of unemployed by county are taken directly from estimates published by the U.S. Bureau of Labor Statistics’ Local Area Unemployment Statistics program³.  The CPS interviews respondents for four months in a row, allowing us calculate the average flow from unemployed to employed status in a given month for these longitudinal observations.  We link CPS person-level records along two household identifiers—HRHHID1 and HRHHID2—and a person identifier—PULINENO—between consecutive months, as indicated by the date and “month in sample” identifiers.

Data on the number of job openings by occupation and county are from The Conference Board’s Help Wanted OnLine^© (HWOL) data series⁴.  Because HWOL data are generated by pulling information from online job boards, these data may over- or under-estimate the actual number of job openings in occupations that have greater (or lesser) utilization of web-based recruitment methods.  In our “alternative calculation” we adjust for these potential discrepancies using a method described in detail in section Ia below.   

Ia. Adjusting online job postings data

There is currently no official government source of data on the number of job openings posted by employers at the state- or local-level.  This is why HWOL is an indispensable resource for economists trying to understand local labor markets.  However, HWOL estimates are based on online job advertisement activity; as such, they may fail to accurately represent activity in the broader economy.  While this shortcoming is common to all statistics based on internet activity (see Kureková et al. [2015]), concerns about the representativeness of advertisement-based job vacancy estimates have received particularly close scrutiny in the literature.

Labor economists have developed methods to adjust The Conference Board’s previous measure of job vacancies—which was based on newspaper help-wanted ads—to approximate the demand for labor in the broader economy (see for example Abraham and Wachter [1985], Valetta [2005], and Barnichon [2010]).  In more recent work, Sahin et al. (2014) use the limited information on industry composition provided by HWOL to determine the differences in industry coverage between HWOL and the Bureau of Labor Statistics’ Job Openings and Labor Turnover⁵ (JOLTS) survey.  We propose an approach that uses a different benchmark for sectoral composition but leads to results that are consistent with those of Sahin et al.

We adjust job vacancy data from HWOL by benchmarking these data to the JOLTS survey.  JOLTS is a random sample survey of the entire population of establishments.  Although JOLTS has a much smaller sample size than HWOL—which covers all internet job posting activity—it captures aggregate and industry-specific job vacancy trends for all businesses, not just those that advertise online.

Our adjusted job vacancy measure is obtained by multiplying state-level HWOL occupational data by occupation-specific weights.  These weights are equal to national-level JOLTS job vacancies-by-occupation divided by national-level HWOL job vacancies-by-occupation.  For example, a weight of 2.0 means that JOLTS reports twice as many vacancies as HWOL in a given year (or, in other words, that HWOL undershoots JOLTS vacancies by 50%.) 

JOLTS does not publish job vacancy data at the occupation level.  To derive synthetic occupational estimates from JOLTS, we convert industry-level JOLTS data to the occupation level using occupation-by-industry employment shares from Occupational Employment Statistics (OES)⁶ and ownership type-by-industry employment shares from the Quarterly Census of Employment and Wages (QCEW)⁷.  (Abraham and Wachter [1985] report survey-based evidence that the occupational distribution of employment closely matches the occupational distribution of vacancies.)

The overall level and trend of HWOL job vacancies and our weighted series are roughly equivalent.  HWOL reported slightly fewer vacancies prior to 2008 and slightly more vacancies after 2008 relative to our weighted series, perhaps indicating an increase over time in the adoption of online job advertisement.  Notably, the two series demonstrate very different cyclical behavior.  While the weighted series took until 2012 to recover its 2007 level of job vacancies—roughly consistent with reported trends in employment—the HWOL series reached near-recovery as early as 2010.

The differences between the two series in occupational coverage are much starker.  For instance, Food Preparation and Serving Related Occupations had an average weight of 3.04 during the period studied, indicating that JOLTS reported a higher number of vacancies than HWOL in this sector by a factor of three-fold.  In contrast, Computer and Mathematical Occupations had an average weight of 0.23, indicating that JOLTS reported a lower number of vacancies in this sector by a factor of four-fold. 

As the last graph in the body of the article above shows, these weights materially alter the degree of labor surplus or shortage by occupation that we measure in North Carolina.  However, these weights have relatively little impact on the overall ranking of occupations; by either calculation, routine and manual markets are experiencing high levels of excess unemployment while some nonroutine/cognitive occupations are seeing labor shortages.

II. Matching function estimation

We estimate the matching function on national data in a similar manner to Sahin et al. (2014), imposing a Cobb-Douglass specification as follows:
 

log⁡(h_t/u_t) = α+ γ〖QTT〗_t + β log⁡(v_t/u_t ) + δ〖Month〗_t + ϵ_t ,  
 

where h_t/u_t is the ratio of hires to unemployed (the hiring rate) in a given month t, v_t/u_t is the ratio of job vacancies to unemployed (the rate of labor market tightness), and〖QTT〗_t is a vector of the four elements of a quartic time trend.  We add to Sahin et al.’s specification a vector of calendar month indicators (〖Month〗_t)  because much of the variation over time in hiring rates is seasonal in nature; we drop the March indicator as our base month.  The constant term (α), when combined with the quartic time trend, captures the level and change of aggregate matching efficiency—i.e. the hiring rate conditional on labor market tightness and seasonal effects.

For individual occupation groups, we again follow Sahin et al. and specify our model as follows:


log(h_it/u_it ) = γ〖QTT〗_t+ρ〖OccGrp〗_i+β log⁡(v_it/u_it )+δ〖Month〗_t+ ϵ_t ,
 

where i denotes individual occupation groups and〖OccGrp〗_i is a fully-saturated vector of occupation group indicators which capture occupation-specific matching efficiency.  We estimate both the aggregate and panel data models using ordinary least squares.

We compare our occupation-specific matching efficiency parameters to those estimated by Sahin et al. in the table below.  In addition to the differences in date range and specification, we estimate efficiency parameters for four broad, skill-based occupation groups instead of the five groups estimated by Sahin et al. in their baseline model.⁸  Moreover, we assume constant occupation-specific efficiency parameters during the period studied rather than estimating post-recession shifts; Sahin et al. find little change post-recession and use only the pre-recession parameter estimates in their baseline model. 

 

Our parameter estimates are uniformly lower than those of Sahin et al.  However, we note that the rank order and dispersion of our parameter estimates are roughly equivalent for the models using raw HWOL job vacancy data.  Following Sahin et al., we normalize occupation-specific efficiencies by an aggregator of efficiency across markets in our calculation of mismatch indexes.

III. Calculating mismatch indexes

We compute an index of labor market mismatch first proposed by Jackman and Roper (1987) and re-introduced by Sahin et al.  Although their calculations are equivalent, Jackman and Roper interpret their index as the share of unemployment caused by structural imbalance, while Sahin et al. interpret it as the proportion of potential hires from unemployment lost due to mismatch (Canon et al. [2013]), as follows:


M_t = 1 - h_t/(h_t^*) ,
 

where h_t/(h_t^*) is the number of observed hires divided by the number of hires in a labor market without mismatch.  We compute the mismatch index from observed data as follows:


M_t = 1 - ∑(v_it/v_i )^β (u_it/u_i )^1-β ,
 

using 0.5 as a value for the elasticity of hiring with respect to labor market tightness β which results in upper bound estimates of mismatch.  We also calculate an index that incorporates occupation-specific differences in matching efficiency as follows:


M_ϕt = 1 - ∑(ϕ_i/ϕ_t) (v_it/v_i)^β (u_it/u_i)^1-β ,
 

where ϕt represents occupation-specific efficiency parameters and aggregates occupation-specific efficiency weighted by vacancy shares in period t, as follows:
 

ϕ_t  = [∑ϕ_i^1/β (v_it/v_i )]^β .

IV. Calculating labor surplus and labor market slack

We also compute absolute measures of labor surplus and relative measures of labor market slack as descriptive indicators to guide workforce planning in North Carolina.  Labor surplus is simply the number of unemployed minus the number of job vacancies in a given market i:


Labor surplus_it = u_it - v_it ,
 

and the labor market slack rate is simply the number of unemployed-per-job vacancy:


Labor market slack rate_it = u_it/v_it   .
 

Following Burke (2015), we extend Sahin et al.’s methodology and weight vacancy amounts by (ϕ_i/ϕ_t)² in order to account for heterogeneous matching efficiency in our alternative calculations of labor surpluses and slack rates.

We have calculated matching efficiencies, mismatch indexes, and measures of surplus and relative slack using both raw job vacancy data from HWOL as well as data re-weighted to the JOLTS benchmark as described above.  In this article, we report results using detailed sectors, unweighted vacancies, and homogenous matching efficiency as our upper (baseline) estimates, and use broad sectors, weighted vacancies, and occupation-specific efficiency parameters in our lower (“alternative calculation”) estimates. 

Our alternative calculation uses HWOL-to-JOLTS job vacancy weights and matching efficiency parameters calculated at the national level and applies them to state (North Carolina) level data.  We assume that differences in matching technology between individual sectors are common across states within the U.S.

References

Abraham, Katharine G., and Michael Wachter. "Help-wanted advertising, job vacancies, and unemployment." Brookings Papers on Economic Activity 1987.1 (1987): 207-48.

Acemoglu, Daron, and David Autor. “Skills, Tasks and Technologies: Implications for Employment and Earnings.” In Handbook of Labor Economics. Vol. 4B, edited by Orley Ashenfelter and David E. Card, 1043–171. Amsterdam: Elsevier B.V (2011).

Barnichon, Regis. "Building a Composite Help Wanted Index," Economics Letters 109(3) (2010): 175-8.

Burke, Mary A. “The Rhode Island Labor Market in Recovery: Where is the Skills Gap?” Federal Reserve Bank of Boston Current Policy Perspectives 15.7 (2015): 1.

Canon, Maria E., Mingyu Chen, and Elise A. Marifian. “Labor Mismatch in the Great Recession: A Review of Indexes Using Recent U.S. Data.” Federal Reserve Bank of St. Louis Review 95.3 (2013): 237-71.

Jackman, Richard A. and Stephen Roper. “Structural Unemployment.” Oxford Bulletin of Economics and Statistics 49.1 (1987): 9-36.

Kureková, Lucia M., Miroslav Beblavý, and Anna Thum-Thysen. “Using online vacancies and web surveys to analyse the labour market: a methodological inquiry.” IZA Journal of Labor Economics 4.1 (2015): 1.

Şahin, Ayşegül, Joseph Song, Giorgio Topa, and Giovanni L. Violante. "Mismatch Unemployment." American Economic Review 104.11 (2014): 3529-64.

Valleta, Robert G. Why Has the U.S. Beveridge Curve Shifted Back? New Evidence Using Regional Data. No. 2005-25. Federal Reserve Bank of San Francisco (2005).