Essays on applied labour economics: personality traits, artificial intelligence and formalisation policies
Publication Date
July 27, 2026
Creator
Abstract
Chapter 1: The Effect of AI Adoption on Personality Traits Demand.
The first essay examines the demand for Big Five personality traits in UK job advertisements and how this demand relates to AI adoption. Using Boolean retrieval methods on 11.7 million online job vacancies from Adzuna between 2017 and 2022, I identify personality trait requirements through dictionaries compiled from established psychology and economics research (Goldberg, 1990; Costa Jr and McCrae, 1992; McCrae and John, 1992;Heckman and Kautz, 2012). Working with unstructured job advertisement text requires building custom dictionaries and implementing validation procedures to transform free text into job characteristics and requirements. This involved manual validation through multiple rounds of randomly selected advertisements and over 1,700 specific replacements to ensure accuracy.
The analysis reveals that personality traits are highly demanded in UK job advertisements. Seventy-two per cent of vacancies mention at least one personality trait, exceeding the proportion mentioning education requirements (11 per cent) or soft skills (62 per cent). Extraversion is the most frequently demanded trait, appearing in 44 per cent of advertisements, followed by Conscientiousness (34 per cent), Openness (30 per cent), Emotional Stability (14 per cent), and Agreeableness (13 per cent). These demands vary systematically across occupations, with customer-facing and administrative roles showing the highest requirements.
After accounting for occupational changes, personality demands remain mostly stable over time. I use two approaches to examine how AI relates to personality demands. At the vacancy level, fixed effects models show that jobs requiring AI skills are more likely to demand Openness (7.2 percentage points) but less likely to demand Extraversion (2.7 percentage points), Emotional Stability (2.2 percentage points), and Conscientiousness (1.7 percentage points). At the firm level, I implement a staggered difference-in-differences design following Callaway and Sant’Anna (2021). Using entropy balancing to ensure comparability between treated and control firms (Hainmueller, 2012), I find that firms increase their demand for Openness (3.6 percentage points) across all subsequent hiring after posting AI positions for the first time. These effects persist when excluding AI positions from the analysis, showing that the relationship extends across firms’ entire workforce.
The consistency of the Openness effect across both empirical approaches suggests that when firms adopt AI-related skills, they increasingly value traits related to creativity and adaptability. This contributes to research on how new technologies affect labour demand by showing that AI adoption relates not only to cognitive and technical skill requirements but also to personality characteristics.
Chapter 2: The Wage Penalty Paradox.
The second essay investigates a puzzling pattern: while individual-level studies show that personality traits increase wages, job advertisements mentioning these traits are associated with lower wage offers (Acemoglu et al., 2022b; Deming and Kahn, 2018; Brencic and McGee, 2023). Using the same dataset of 11.7 million UK job postings, I document this paradox and test four potential explanations.
The baseline analysis confirms that advertisements mentioning personality traits offer 4.6 per cent lower wages on average. The penalty is largest for Extraversion (3.2 per cent) and Emotional Stability (3.1 per cent), followed by Conscientiousness (2.0 per cent) and Agreeableness (1.4 per cent). Only Openness shows a small positive effect (0.7 per cent). First, I examine whether personality trait mentions correlate with occupation or job characteristics that independently determine wages. Occupation fixed effects produce the largest reduction in trait penalties, but significant penalties persist. Trait penalties are smaller in occupations with higher proportions of women and migrants but larger in occupations with more degree holders and higher mean wages. This pattern contradicts simple productivity-based explanations.
Second, following the theory of equalising differences (Rosen, 1986), I test whether workers accept lower wages in exchange for non-monetary benefits. Simple additive controls for amenities had minimal impact. However, interaction analysis reveals that when personality traits are combined with traditional benefits, workplace culture signals, and quality amenities, trait penalties are reduced or eliminated. Some amenity combinations amplified penalties, particularly remote work, showing strategic use of personality language depending on compensation context. Third, using TF-IDF text analysis, I examine whether trait mention intensity affects wage implications. The relationship is highly non-linear: low-intensity mentions carry minimal penalties, while high-intensity mentions show substantial penalties. This suggests employers use trait language strategically for different purposes depending on the emphasis level. Fourth, I train a BERT classifier to identify implicit gender targeting and examine whether trait penalties differ across gender categories. The evidence does not support demographic targeting as the primary driver. Most traits show larger penalties in male-targeted advertisements, and both categories show similar intensity patterns.
These findings indicate that the personality trait wage penalty reflects multiple overlapping employer strategies. The relationship between trait mentions and wages depends on context, linguistic presentation, and organisational characteristics rather than simply reflecting productivity differences.
Chapter 3: The Effectiveness of Formalisation Policies in Latin America.
The third essay, co-authored with Verónica Escudero and Elva López Mourelo, presents the first comprehensive meta-analysis of formalisation policies in Latin America. We analyse 79 impact evaluations with 527 estimates from interventions implemented between 1990 and 2020. Unlike previous meta-analyses (Card et al., 2010; Floridi et al., 2020), we include Spanish-language studies and unpublished government reports, examine interventions targeting both worker and firm formalisation, and calculate standardised mean differences to enable systematic comparison across studies.
Our analysis addresses publication bias using FAT-PET-PEESE methods. We find that bias correction substantially reduces effect sizes. Uncorrected pooled estimates overstate true effects, with formality estimates falling by approximately one-third and earnings estimates by half or more after correction. After correcting for publication bias, active labour market policies show consistent positive effects on formality outcomes (0.238 standard deviations), corresponding to increases of 12 to 14 percentage points in the probability of formal employment. Information campaigns (0.226 standard deviations) and labour inspections (0.320 standard deviations) also show positive effects. For earnings outcomes, social protection programmes with activation components show the largest effects (1.258 standard deviations) but a negligible formality impact. This suggests these programmes primarily affect income through direct transfers rather than facilitating transitions to formal employment.
Programme benefits strengthen over time, with long-term effects exceeding short-term impacts by approximately 0.10 standard deviations for formality and 0.32 standard deviations for earnings. Multi-component interventions outperform single-component policies, indicating potential complementarities between different policy tools.
Item Type
ethesis
Thesis Type
PhD
Supervisors
University of Nottingham
University of Nottingham
University of Nottingham
Subjects (LC)
Associated Schools / Departments
Author URL(s)
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
Matias_Golman_UoN_Thesis.pdf
Size
12.62 MB
Format
Adobe PDF
Checksum (MD5)
c00a0d4ac8b4f494431e878d098eed72