Self-employment experience and labor market integration of refugees: A field experiment in Sweden

This project focuses on how employers value the work history of job applicants. The purpose is to compare future labor market outcomes for people who have chosen to start their own business with those who are wage-employed and those who are unemployed and still seeking for employment.

We have designed a correspondence test that allows us to estimate a causal effect of self-employment experience, relative to continued wage employment and unemployment, on the likelihood of a positive call-back from employers.

2023-02-20

2023-12-31

The causal effect of the assigned work experience is estimated using both a non-parametric approach and a parametric approach. We will use a Linear probability model to estimate the probability of a positive callback depending on work experience. A Probit model and Average Marginal Effects (AME) are used for an alternative estimation of the probability of positive response on all specifications.

Heterogeneous Effects
In the non-parametric approach we will evaluate if there is heterogeneous effects by gender using a two tailed Chi2 test to evaluate if there is a significant difference in callback rates, within and across work experience types. In the parametric approach We will as a first step generate an interaction term between self-employment and gender, which is included in the regression specifications. We will also run separate regressions based on gender.

The database will have three main categories; 1) the candidate profiles, 2) the job ad characteristics
and employer information, 3) the type of response from the employer. The candidate profiles are
created to represent the fictitious individuals. The job ad characteristics and the employer information
are downloaded from Platsbanken. The type of response from employers are collected from the
correspondence test.

The main variables of interest from the candidate profile list are work experience and gender. The
variable work experience has three categories; Self-employed, Wage employed and unemployed, where Wage employed is the reference category. The variable gender is binary. The candidate profile also contains the name and country of origin of the
fictitious individual. The main variables of interest in the job ad characteristics and employer information category are: Position-type, Occupation-field, Contract-type, Experience-Required and “Must have” Language. Further, the variables describing the job ad and employers are Ad-id, Concept-id, Headline, Title, Todays-date, Publication-date, Publication-deadline, Employer (company name),
Email, url. The third category includes the response from employers. The four response variables are binary, mutually exclusive and exhaustive. Two types of positive callback measures are defined as follows: positive response positive that includes the observations
where the employer asks for more information and invitation for an interview and the sharper measure of positive response that only contains observations that received an invitation for an interview.

First we will study differences by gender because research shows that there are gender differences in self-employment e.g., (Brush et al., 2006; Parker, 2009; Simoes et al., 2015). Further, Alden et al., (2021) suggests that employers appear to value self-employment experience of women and men differently. Second, the work experience in the applications is from the Occupation field Restaurant branch. This implies a risk that applications sent to job ads in this field drives the result. We will test this by running separate regressions on observation within the restaurant branch and a pool of other fields. The randomisation is not performed on occupation-field so gender is added as control variable.

We will test if there is a difference by Occupation field.

We will define a dummy variable from the variable Occupation-field that equals one if Occupation-field is Restaurant Branch and zero otherwise. The subgroup analysis by OF does not allow causal inference since the number of observations will be inadequate.

The correspondence test is carried out by creating fictitious applicants who apply for jobs advertised on Platsbanken and document callbacks from employers. The candidate profiles comprise a set of names, gender, country of origin and work experience. The candidate profiles are used to create fictitious applicants with similar cover letters and CV’s.

Research Questions
The correspondence test is designed to answer the following questions:
1. How do employers value self-employment experience for a refugee immigrant of Middle Eastern
origin?
(a) Compared to a corresponding period as an employee?
(b) Compared to a corresponding period of unemployment?
2. In this setting, do employers value self-employment experience differently for women and men?

By using a field experiment we can address the issue of selection bias by randomising “group” affiliation. In this way, the applicants only differ in terms of the variable of interest, in this case type of labor market experience. As a result, any selection bias due to a correlation between individual characteristics and the push and pull factors of self-employment entry and exit, are eliminated. The design of the correspondence test allows us to study the causal effect of self-employment experience on the labor market prospects for refugee immigrants, relative unemployment or wage-employment. Thus, as we are constructing the applicant’s qualifications, we can control all information that the employer can observe about the applicant. This means that any difference in the employers assessment of the CV stem from our manipulation of the applicants work experience.

This study will be conducted on a population of employers who advertise a job opening on Platsbanken, the Swedish Public Employment Service online recruitment platform. Because the purpose is to investigate the labour market for refugee migrants with low education we will only apply for jobs in the low-skilled sector. In the standard for Swedish occupational classification (SSYK), professions with requirements for shorter training or introduction are classified with a four digit code staring with number 9. The format for the job ads is standardized such that the information on SSYK classification is available for all job ads posted on the platform. There are 93 occupation names that are included in the category SSYK9. The sample data contains all job ads on the platform classified as SSYK9 that match the occupation name list. The sample might not contain unique employers since each employer can post more than one vacancy during the time period of the data collection. The sample is representative for a subset of employers that are actively recruiting in the low skilled sector. The reason for recruiting can be higher turnover or expansion of the business. Further, employers/firms might use other channels for recruiting than Platsbanken, for example personal networks or other online platforms. This could be due to unobservable differential characteristics
of excluded employers/firms. Hence, the sampled employers might not reflect the population of employers in the low skilled sector in all aspects. However, the sampling method will provide a large sample of recruiting employers, that allows inference about employer’s valuation of self-employment experience.

The data collected in the experiment is call-backs from employers. The response from the employers will be documented for each of the applications. The employer can reply by sending an email or by leaving a voice message. The data collection is conducted in three steps. First, we will download the job ads from Platsbanken. The data from the job ad contains the following variables: Ad-id, Concept-id, Headline,
Title, Todays-date, Publication-date, Publication-deadline, Employer, Email, url, Position-type, Occupation-field, Contract-type, Experience-Required and “Must have” Language. Through the Ad-id we can at any point access a specific job ad as historical data is also provided. If the employer has specified must have language skills that do not match the language specified in resume of the fictitious individual, the job ad will be removed from the list since the applicant lacks education in the required language. The list of job ads will be merged to the candidate profile list. Each daily list of job ads will be appended to the candidate profile list consecutively, such that it always continues from the last
used candidate profile id. The main database is built by the collection of these merged lists. Second, individual resumes are created using information from the job ad and the candidate profile. The name and work experience are collected from the candidate profile. The headline or title will be collected from the job ad. The resume is included in the job application that is sent to the employer by email or by filling out an application form (url). Third, the responses from the employers will be sorted in accordance with the response category criteria and documented in the main database.

The randomisations is done on a computer using block randomisation.

Work experience (Self-employment, Wage employment, Unemployment) will be randomly assigned to the candidate profiles. The random assignment is performed such that there is no correlation between the assigned work experience and
gender, candidate name or country of origin. The units of analysis are work experience types and gender.
The names and country of origin are randomly assigned and will not be treated as units of analysis.

The treatment assignment is designed to ensure a balanced sample, each name country of origin and gender
will appear with the same frequency. I will not randomise on Position-type, Occupation-field,
Contract-type or Experience-Required and there might be correlation between each of these variables
and the probability of getting a positive response from an employer.

No

3882

3882

The sample size is 3882 applications. There are 3 treatment arms with 1294 applications in each arm. Within each arm there are two subgroups based on gender. Each gender subgroup will consist of 647 applications.

To ensure that we have enough observations to find an effect, given that there is one, we have performed power calculations. We will estimate differences in probability of receiving a positive callback from an employer. To calculate the expected callback frequency, the results from three previous studies have been taken into account (Ald´en et al., 2021; Ek et al., (2020); Koellinger et al., 2015). All studies use a field experiment with fictitious applications. Koellinger et al. al., (2015) estimate that the difference in feedback rate for applicants for qualified positions with experience of self-employment or experience as an employee is 10 percentage points. Ald´en et. al., (2021) shows that in a comparison based on applicants with Arab/Muslim origin or Swedish origin, there are relative differences in feedback depending on work experience and gender. The difference between self-employment and employment within the group with Arab/Muslim origin is greater for men with approx. 12 percentage points than for women with approx. 1 percentage point difference. Ek et al. (2020) have focused solely on the group of refugee immigrants with a low level of education and no work experience or experience in simple jobs. It lowers the feedback frequency relative to the above-mentioned studies and the estimated difference between unemployment and employment is approximately 0 for men and 1 percentage point for women. Given the results above, the expected callback rates for self-employed is 0.05, for unemployed 0.01 and for wage employed 0.1. For alpha-level α = 0.01, and the expected difference in callback rates of 0.04, the sample size required to estimate differences between 3 treatment groups and gender (6 groups in total) with 80% power is 3882 observations. The estimated sample size is based on the estimates from the power calculator provided by the statistics department at the university of British Columbia. https://www.stat.ubc.ca/ rollin/stats/ssize/b2 We will control for employer characteristics and job characteristics that may be correlated with employer response type and thus affect the estimated probability of a positive callback. The covariates that will be tested for inclusion in the regression analysis are: Occupation-field, Position-type, Contract-type and Experience Required. A priori we do not know the characteristics of the employers that will be in our sample and hence, predictions about the share of the variance that these control variables will explain are not meaningful. There is a risk that there is no difference in the call back rates or that the difference is smaller than the approximation that was used in the power calculations. To address this matter we will start with a test period of 1 month. Analyses of the test period data provides further information of the expected differences in callback rates and serves as a basis for adjusting the sample size.