Beyond Schmoozing
Many believe that networking is key for job market opportunities, with some claiming that large parts of the job market are hidden from non-networkers. On the contrary, I show that UK job adverts data effectively represent the job market. Moreover—building on US survey data—I find networking efforts are a mild predictor of job offers.
The Economist’s management column is called ‘Bartleby’. A recent piece published in the Bartleby column—named ‘Hire, liar’—depicts hiring practices as a game of pitiful lies. In order to play the game, our career persona now requires its own go-to-market strategy and a social network profile. The result is insincere, unconvincing content that is overwhelming LinkedIn.
The name of the Bartleby column originated from a 1853 short story masterpiece by Herman Melville. The plot follows an unhappy Wall Street clerk named Bartleby, who refuses to do any work. “I would prefer not to,” he responds to any request. Bartleby stays day and night in the office—even after his boss relocates—and ends up dying from starvation in The Tombs of downtown Manhattan.
15 years ago, The Economist called LinkedIn ‘Facebook for suits’ and pointed to the rise of the professional social network. That was true. Nowadays the role of networking in career advancement is widely accepted. Resume applications and job posts scrolling are considered passé—at least for white-collar jobs. To secure a desirable job—one that will allow you to pump your next LinkedIn post—you will need a strong network, THEY say. By ‘THEY’, I mostly mean career-advice content creators, who have primary interest in promoting the benefits of networking.
A common internet gospel says the majority of jobs are filled via networking (e.g. 1 2 3 4 5). Explicitly, that 80-85% of hirings are done from a professional network referral. It is also frequent to read on the web about a large hidden pool of jobs that are not published online or anywhere else. This hidden job market is considered to be available only to the well-connected and well-networked.
Jobs filled through networking are not necessarily hidden. It could be that networking is a common way to match job seekers and employers, while the right candidate could have an equal opportunity when applying online. Nonetheless it is common to come across the claim that 70–80% of the job market is hidden.
These two theses coupled together—that most jobs are filled via networking, and that most jobs are hidden—tell a story of an unjust, inequitable job market. According to this narrative, you are either a schmoozer or a loser (or just unemployed for a long time). It is quite comparable to the New-York rental market, where so many of the incoming tenants are ‘take-overs’. They inherit the lease from their friends who leased it before them. Those leases are filled fast and never advertised.
Networking-heavy markets exist. Stories about those markets—true or false—are sticky. When it comes to the job market, I think the proclaimed role of networking is broadly a myth. I aim to cut this myth down to size.
Hide-and-seek-a-job
The sources that disseminated the ‘networking thesis’ refer—circularly—to one or two specific surveys. A great LinkedIn article—written by an honest career expert—shows with simple groundwork how one of the original surveys is merely a simplistic, inexpert poll—that has generated bloated conclusions about networking. Even the plain results of the poll are not exactly what has been massaged into claiming that ‘85% of jobs are filled via networking’. If you ever dealt with ‘newspaper statistics’ before, you are probably not surprised.
The claim that the majority of open positions in the job market are hidden is the more peculiar one; to my judgment. These allegations are the ones to create the ‘career FOMO’ and ‘LinkedIn culture’ among white-collar professionals; the feeling that you are losing access to the truly treasured opportunities because you have not been schmoozing enough.
In recent years online job adverts data have become available and highly granular. Adopted by economics research, there is a growing body of analysis and research utilising job adverts data. Some of which can indirectly answer the question: What share of total labor demand is covered by job adverts data? Hence what is the proportion of positions advertised in online job boards?
Vacancy surveys—done by official statistics bureaus—are a trusted source for labor demand estimates. Examples are the JOLTS report for the US and the Office for National Statistics’ (ONS) Vacancy Survey for the UK. When comparing online job adverts data to the official survey estimates of vacancies—one can reliably estimate the share of openings that are hidden. A number of recent papers have done that.
In 2014, a technical report by the Center on Education and the Workforce in Georgetown University found that north of 80% of openings for candidates with higher education are posted online. Authors used a data source that crawls more than 15,000 online job portals in the US.
An NBER working paper from 2019 found that data from a single UK portal (reed.co.uk) is enough to cover 40% of the vacancies from the ONS survey.
A recent IZA discussion paper by LSE economists—including leading labor economist Alan Manning—found the stock of online job adverts to be on average 93% of the ONS survey vacancies stock. Authors used Adzuna job-search engine data; scraped from thousands of sources in the UK.
The work done in those papers is serious and rigorous; authors provide lengthy discussion of data challenges and solutions regarding cleaning, de-duplication, flow vs. stock and more. Clearly, it suggests that the share of jobs hidden from job seekers is insignificant.
Textkernel collect job adverts data web-scraped from roughly 90,000 sources. I used their raw data for monthly new job adverts in the UK and the raw—not seasonality adjusted—vacancies data from the ONS official survey. In order to match the ONS survey, I have excluded from the Textkernel dataset all adverts from Northern Ireland—simply turning it into the Great Britain—and all Agriculture adverts; both of which are not surveyed by the ONS. The results plotted below show that the two sources follow each other remarkably well. The adverts data come from Textkernel cleaned and de-duplicated using NLP.
I have run a regression with the two series—the ONS vacancy survey explained by the Textkernel new online job adverts. The estimation show that we cannot reject a full parity i.e. that the two sources equal each other within the 95% confidence intervals.
The data does not tell a story of a hidden job market. Rather, it tells a story about the domination of online job-search over traditional job-search methods. Even some degree of labor demand un-represented by online job adverts data is acceptable—yet there is a way to go to prove that those “hidden” openings are filled via networking. Logically some job openings will be filled through word of mouth or door-to-door search.
Does it pay to network?
To further analyse the role of networking in labor markets, I will explore the job-search extension from The Survey of Consumer Expectations by the Federal Reserve Bank of New York (SCE FRBNY; disclaimer at the end of this article). Via this extension, respondents are being asked a comprehensive set of questions regarding their job-search intentions and activity. Using SCE FRBNY we can tackle the remaining half of the ‘networking thesis’—that most jobs are filled through networking. My target is more fundamental—and practical—given the data at hand. I will focus on estimating the return of networking activity for job seekers. Does it pay to network?
To set the stage, we will only be looking at respondents with higher education (with a bachelor’s degree or more). This is because usually—nor it is claimed to be otherwise—blue-collar workers do not gain much from networking.
Among job seekers in the SCE FRBNY, on average 53% did any networking in the last 4 weeks (per the day of the survey). Job seekers are respondents who did anything at all to look for work within the last 4 weeks. I label networking activity as one the following: Contacted friends or relatives; Contacted former co-workers, supervisors, teachers, business associates; Contacted current employees at other companies. I find the results below to be a sign of how well-perceived are the returns on networking in the eyes of job seekers. The sharp drop in networking in 2021 could be attributed to Covid-19—as in lockdowns and the rise of working from home.
What does a successful networking look like? Perhaps as being contacted by an employer you did not apply to; and who was referred to you by your professional network. If so, the SCE FRBNY data can help with that too. Looking at all highly educated respondents—who were contacted at least once by a potential employer—between 16 to 19% of US employees reported a ‘successful networking’ contact (depending on the time period in the question).
Interesting to notice a higher prevalence of overall unsolicited contact when respondents being asked about the ‘last 4 weeks’ (75%) vs. ‘before accepting your current job’ (29%). This is due to the fact that respondents who were contacted recently—regardless of a job change—are in high-demand and enjoy higher probability of being contacted unsolicitedly. However, when asked about how were they contacted before accepting their current position, it is a given that respondents were contacted. This was verified by filtering for employees that recently started their current job; and can also be seen by a substantially smaller sample size of respondents being contacted in the last 4 weeks (bottom panel below).
A straight-forward generalisation can tell a great deal. While more than half of US job seekers engage in networking—less than a fifth of hiring contacts are sourced from an unsolicited referral. This suggests that the anticipated returns on networking are inflated. Still, this is an abstraction. One must consider the many interactions and fixed effects when analysing job-search effort and reward.
For that I will use gradient-boosted decision trees (GBDT). The decision to use GBDT is driven by two main reasons: First, because of null records treatment. The SCE FRBNY includes a fair amount of follow-up questions that were introduced conditionally on one’s response to a previous question. For example the question ‘How long have you been looking for a job?’ is only introduced to respondents who looked for a job in the last 4 weeks. GBDT naturally accept nulls. Second, the extent of interactions within the explanatory variables. For example, not only that age has an impact on job-search outcomes, it can also impact other channels like e.g. the length of a job-search. A day spent job-searching at age 45 is more efficient than at age 25—perhaps. GBDT allow for some degree of causal—more accurate, predictive—inference; while sorting through the interactions unaided.
The target variable equals 1 when the respondent has received any job offers in the last 6 months and 0 otherwise. Again, we are only keeping records for respondents with a bachelor’s degree as a minimum. The model includes more than 30 variables, most of which are job-search related. Additionally, there are socio-economic and demographic controls, a time-trend, and an indicator of labor market shocks: the ratio of unemployed persons to job opening per year and state (from the U.S. Bureau of Labor Statistics).
In line with normal GBDT processing—I tune the model for the maximum depth of a tree; and using ‘early-stopping’ to choose for the number of trees in the system. The training dataset is imbalanced, showing 77% of rows with target value of 0 (i.e. no job offers in the last 6 months). Since the imbalanced GBDT showed no predictive power—I downsampled the data. As seen below, the final model is able to moderately improve the null model—with 70% out-of-sample accuracy.
Networking activity is also part of the GBDT model input. Same as earlier, networking variables are in the form of three fields indicating a contact with one’s professional network. Using a rank of features’ importance, we can examine the predictive power of networking on job offers. The results show that networking efforts are behind other traditional job-search indicators.
The best predictor of job offers is the high frequency of job changes—implying that one’s repetitive intentions, efforts and experience in accepting and getting offers is an early indicator of job offers. Following ‘job changes’ in the rank are a list of vanilla job-search activities like applications, using head-hunters, and just ‘looking for work’. Among the three variables labeled as ‘networking’—contacting an employee in another company has non-existent contribution; while contacting associates or friends are in the middle to bottom-half of the importance ranking.
Closing remarks
The analysis outlined here does not imply that networking is pure fluff. Bonding with your professional community has many “first-order” benefits—it enhances learning and is genuinely fun.
I argue that the idea of networking as a job-search strategy is exaggerated, and that it is an empty idea that networking is inevitable in order to access a “hidden” job market. Debunking such ideas is often best done with careful empirical work.
SCE disclaimer:
Source: Survey of Consumer Expectations, © 2013–2023 Federal Reserve Bank of New York (FRBNY). The SCE data are available without charge at http://www.newyorkfed.org/microeconomics/sce and may be used subject to license terms posted there. FRBNY disclaims any responsibility or legal liability for this analysis and interpretation of Survey of Consumer Expectations data.
