Discrimination in the university-to-work transition in Italy: a correspondence study

The integration of young adults in the labour market is a problem in many countries, and the youth unemployment rates are high almost everywhere (OECD 2021). Although finding a job is easier for people with higher educational attainment levels, the employment rates of recent graduates differ largely from country to country, with Italy as the EU Member State with the lowest employment rate in 2024 of recent tertiary graduates (ISCED 5–8): only 77.3% against more than 95% in Bulgaria and Estonia, and a European average of 86.7% (Eurostat 2025; table edat_ifse_24). Thus, at least in some countries, the university-to-work transition can be an extremely challenging process, even more so with the increasing uncertainty of the labour market and the changes in the socio-economic context that have followed the 2008 economic crisis (International Labour Office 2017) and the recent pandemic.

The employment rates of recent graduates also differ between men and women, with an average European gender gap steadily in favour of men, varying from 2.7% points in 2022 to 5.4 in 2016 and equal to 4.4% points in 2023; in Italy this gap is even larger, with a difference of 6.8% points in 2023 (72.5% for women, 79.3% for men, see Eurostat 2025; table edat_ifse_24).

Searching for a job is even more ticklish for graduates of foreign origin, who face the double challenge of a precarious job market, and of a discriminating society. Although about 80% of immigrants in Italy, especially those from non-EU countries, work as manual workers (Ministry of Labour and Social Policies 2019), foreign and second-generation graduates are more and more present in the job market, and have reached 4.7% of graduates in 2023 (Almalaurea 2024). Their presence is stronger at the bachelor’s level than at the master’s one, and their employment rate one year after graduation is slightly lower than for the generality of graduates (Almalaurea 2021).

In this study we aim at analysing the graduates’ characteristics that, when underlined in a curriculum vitae, might trigger discriminating mechanisms and a priori refusals or, conversely, can enhance their chances of grabbing attention and favour from the potential employers.

During the university-to-work transition process, in fact, a new graduate usually presents him/herself with his/her curriculum. At the very initial stage of the selection process, there is no direct contact previous to the receipt of the resume. Only at a later stage of the selection process may applicants land an interview and come to a possible agreement with the company. So, knowing employers’ attitudes will help new graduates in the thorny transition to work.

This is the justification for the experiment described in the following, that identifies aspects that either improve the candidate’s chances of recruitment or potentially compromise the application. The experiment belongs to the category of ‘correspondence studies’, statistical examinations based on fictitious resumes sent to employers but, instead of applying a typical “matched-pair” design, it relies on a factorial strategy that allows to control for several candidates’ characteristics. Its main aims include:

In correspondence studies, multiple applications are sent for each job opening, and the correlation among the responses/reactions of the same employer to multiple applications is usually fixed by clustering standard errors at the job-advertisement level. However, some authors claim the superiority of a multilevel regression approach to deal with non-independent data, and recommend a wider use of this method in empirical economic studies (see, e.g., Cheah 2009; Oshchepkov and Shirokanova 2022). Our paper attempts to fill this gap, by addressing the problem of correlated data by means of a multilevel approach with job applications as first level units, and job vacancies as second level units.

We carried out our experiment with reference to companies located in the Veneto region of Italy, and with fictitious graduates from the largest university in that region. The selected region has a strong labour market that includes various manufacturing and service industry opportunities and a large representation of immigrants in lower-level jobs. The graduates’ characteristic of coming from the same university aimed to avoid employers guessing about skills achieved in study programs having the same name but delivered by universities with differing reputations. The analysis of such a narrow segment of the Italian business population allowed us to rule out territorial differences and draw more cogent inferences from callbacks regarding the existence of discriminatory factors.

The rest of the paper is organised as follows: Sect. 2 is a short review of correspondence studies; Sect. 3 describes the selection criteria of job openings, the experimental methodology and the survey features; Sect. 4 presents the relevant results of the experiment; Sect. 5 is a discussion of the results in comparison with the relevant literature and the potential limitations of the research design, while Sect. 6 brings some conclusions.

Earlier experiments on the effect of applicant characteristics on callback response have been conducted in a variety of settings: among others, in the U.S., by Newman (1978) and Bertrand and Mullainathan (2004); in Australia, by Riach and Rich (1991); in the U.K., by Jowell and Prescott-Clarke (1970); in Belgium, by Smeesters and Nayer (1998); in the Netherlands, by Bovenkerk et al. (1995); in Sweden, by Carlsson and Rooth (2007); in France, by Duguet et al. (2010) and Adida et al. (2010); in Ireland, by McGinnity and Lunn (2011); in Canada, by Oreopoulos (2011); and in Italy by Busetta et al. (2018). As a whole, these studies signalled race, ethnicity, religion, gender, age and minority-belonging as possible discriminant factors.

In studies focused on race and ethnicity we find almost unanimous evidence of discrimination against minorities, nearly in all continents (see, e.g., the surveys of field experiments conducted by Lippens et al. 2023; Quillian and Midtbøen 2021; Riach and Rich 2002; Rich 2014; Bertrand and Duflo 2017; Neumark 2018). In the Italian job market, self-perceived workplace discrimination is reported significantly more frequently among immigrants (especially Africans) than among Italians (Salvatore et al. 2013; Palumbo and Manna 2019). Also the correspondence study carried out in Italy by Busetta et al. (2018) shows a significant amount of discrimination against first- and second-generation immigrants, but the number of correspondence studies in the Italian context is limited (Baert 2018). Overall, only few correspondence studies have focused on degree-level job roles and recent graduates (see, e.g., Nunley et al. 2015).

Regarding the discrimination against women, Firth (1982), Neumark et al. (1996), Weichselbaumer (2000), Booth and Leigh (2010), Carlsson (2011) all found that women were discriminated against when they applied for technical and higher-status/higher-paying jobs, but this did not happen for middle- or lower-profile and routine jobs, nor was it evident for traditionally female-dominated jobs. In contrast, Levinson (1975), Weichselbaumer (2000), Arai et al. (2018) and Fullin (2016) found substantial discrimination against men, particularly in women-dominated occupations. Schaerer et al. (2023) claim that hiring discrimination against women has declined significantly and is no longer observed in the last decade.

Even age can be a source of discrimination in an entry-level job (Bendick et al. 1999; Lahey 2008). People of different ages often possess different human capital and different potentials for productivity. Among recent graduates age differences are related to completely different study and life paths, hence employers’ discrimination as a direct effect of applicant’s age can be difficult to prove and will not be investigated.

Fry (1986) and Graham et al. (1990) tested discrimination on the basis of disability and found substantial discrimination regardless of the type of disability and the type of occupation. Though, this goes beyond the aim of this paper, and a wider set of experiments should be conducted — possibly differentiating by type and level of disability — in order to really understand how to overcome the employers’ resistance and favour employability of the disabled.

Riach and Rich (2002), Rooth (2009), Hamermesh (2011) and Galarza and Yamada (2017) have all found that having an attractive body increases the probability of finding a job, while obesity is negatively correlated to labour market outcomes (Busetta et al. 2020a). So, a captivating picture at the top of a curriculum may help at the initial hiring stage. However, since including a photo in a job application is hardly mandatory, examining the so-called “beauty premium” was beyond the scope of this experiment.

Some authors (e.g., Lundberg and Startz 1983; Lahey 2008) distinguish between “statistical” and “taste-based” discrimination. Both types of discrimination derive from preconceptions. Statistical discrimination presumes that certain characteristics of a candidate are negatively related to the quantity and/or quality of his/her human capital. For instance, people who are less educated, or less skilled in a language, or possess other objectively-documented impairments are less productive than a hypothesised worker. As a consequence, if there are enough candidates, applicants having one or more of these negatively perceived characteristics are eliminated from consideration. The taste-based, or animus-based, discrimination (Becker 1957) is instead related to recruiters’ values, for instance, of a political or religious or cultural kind. These are values that should not influence either employment or career chances according to constitutions and the human rights declaration.

As a whole, we obtained 55 (out of 600: 9.2%) callbacks (Table 1). By a large majority, companies called back through the telephone (36 cases: 6.0%), as an alternative, they used the email (19 cases: 3.2%). For the vast majority of vacancies (71.8%), none of the experimental resumes was called back, while 21.1% called back only one out of the five candidates, 4.2% called back two of them, and 2.1% called back three of them; only one of the 142 vacancies for which we sent fictitious resumes generated 4 callbacks. Out of the 55 callbacks, 45 were obtained within seven days from when the resume was sent. No callback exceeded 38 days. For the rest of the analysis, we will pool the obtained responses regardless of the method of communication.

The probability of getting a callback within 7 days (Table 1) was higher for Italians (16% for job applications declaring Veneto origins and 11% for candidates from Southern Italy), than for candidates from Eastern Europe (6%) North Africa (5%), Central Africa (5%), and the Middle East (2%). These differences are significant at a 1% level (F_5,494=3,82, p = 0,002). It appears that, with respect to people born in the same region as companies were located, residents from Southern Italy were not significantly discriminated against. This is partly unexpected, because since the internal migrations of the 1960s–70s the local population showed some level of discrimination against people from the South of Italy (see, e.g., Ginsborg 2003; Ballarino and Panichella 2015). 

As compared to local people, foreign-born people were instead very discriminated against. The pattern of discrimination was similar when looking at all the total callbacks, but in this case the differences are smaller (F_{5, 594} = 2,398, p = 0,036): responses were less frequent for people from Africa and Eastern Europe than for Italians, but the only significant pairwise comparison is between people from the Middle East and local graduates. We can observe that the rate of callbacks to graduates from Veneto is the same after seven days and after the entire observation period of more than one month, while callbacks of foreigners grow, especially those of candidates from Africa and Middle East, showing that delayed callbacks are mainly directed toward foreigners.

The multilevel logistic regression analysis confirms the same pattern (Tables 2 and 3, Model 1): foreigners are discriminated against, especially Africans and Middle Easterns, but the interaction with gender is non-significant.

Among the control variables, the academic major was significant. The graduates with a title in the humanities or in the social sciences were significantly preferred to engineers and hard scientists for the types of jobs at hand. This may mean that either the jobs retrieved from the Internet were prevalently low-profile, or at least non-technical, or that selectors tended to invite for a job interview graduates they assumed to have lower expectations. Maybe, both mechanisms operated to give these results, though their single relevance is food for thought.

In Model 2 (Table 2), that includes the second level variables, only the company size effect is marginally significant, showing that medium size companies have a slightly higher rate of callbacks. However, the AIC shows that Model 2 does not perform better than Model 1.

The comparison of Model 2 and Model 1 deviances also shows a non-significant contribution of the second level variables (χ²₁₀ = 255,1–244,7 = 10,4; p-value = 0,406). Moreover, neither the variability of the random intercept, nor the intra-class correlation coefficient, are appreciably reduced by introducing the second level variables⁴. This means that the characteristics of the vacancies and of the companies do not influence the decision to call back a candidate. The same holds true when we estimate models with all the callback occurrences (before and after 7 days) as the outcome variable (Model 3, Table 3).

Bild vergrößern

Figure 1 reports the predicted probabilities of callback with confidence intervals for all the significant variables. We can see once more that the number of callbacks for candidates of foreign origin is noticeably lower than for the Italian candidates, while candidates with a masters’ degree or a degree in social sciences or humanities get more callbacks. Although the wide confidence intervals, these results are consistent across different optimization algorithms (Dual Quasi Newton optimization algorithm and Bounded optimization by quadratic approximation), alternative reference categories for categorical variables, different sets of variables (with or without the second level variables, with or without the non-significant predictors) and even with alternative models (standard logistic model, logit multilevel regression or probit multilevel regression).

Bild vergrößern

Figure 2 shows the estimated random intercepts (and 95% CIs) from the full multilevel logistic model. The figure illustrates that, even after adjusting for individual-level and group-level covariates, substantial between-group variation remains. This is consistent with the fact that the second level covariates did not add much to the models. A possible explanation resides in the nature of the group-level covariates available in this study, since the decision to call back a candidate is likely to depend on the (unknown) traits of the recruiter more than on the company characteristics. Figure 2 also shows wide confidence intervals, as a consequence of the small size of resumes per vacancy; the CIs are particularly wide for those vacancies that did not produce callbacks for any of the fictional candidates in our experiment, that are the majority of vacancies and are displayed on the left side of the plot. The small group sizes prevented us from adding random slopes to the model.

The overall response rate obtained in our experiment was 9.2%. If we consider that this corresponds to less than one out of two job openings, the percentage is rather low. It remains lower than one per job offer even if we parameterize the callbacks to the largest response rate, that of Veneto applicants.

In the literature, response rates differ across populations and methods. Lahey (2008) and Bertrand and Mullainathan (2004) reported experiments similar to ours conducted in the U.S. that obtained an overall 9% and 8.4% callback rate, respectively. In France, Duguet et al. (2010) experimentally tested the potential discrimination of employers against young candidates whose resume made clear their ethnic origin and their residence in the underprivileged suburbs of Paris. The success rate in this experiment was 11.9%. The average response rates were similar to ours.

In Ireland, McGinnity and Lunn (2011) conducted an experiment similar to ours, but sent just two resumes per ad. They obtained 9.6% cases of interviews for both candidates and an additional 14.6% for just one of the candidates. The proportion of callbacks was higher than ours but not that high even in this study.

Studies conducted in other countries led to similar response rates as ours (for surveys of field experiments, see Riach and Rich 2002; Bøg and Kranendonk 2011; Krause et al. 2012; Quillian et al. 2017). Trying to imagine why this occurred, we can resort to the conjecture in McGinnity and Lunn (2011) that, typically, employers who recruit through job advertisements select only a small number from the candidates who respond, and apply a threshold that is toward the tail of the distribution of suitability across candidates. Thus, the possible responses are few. In a similar vein, Bertrand and Mullainathan (2004) conjecture that the low proportion of callbacks in correspondence studies is due to the fact that employers receive so many resumes to be induced to use quick heuristics in reading the resumes: the authors imagine that, as soon as they see a curriculum adequate to the vacancy, employers stop reading. It is difficult to say which hypothesis is more appropriate.

Preferences are at the core of the selection process, selection being a way to rank people fitting a vacancy. Discrimination may occur if individuals are ranked lower unrelated to their actual performance. The selection rationale may be considered discriminatory if the lower ranks of the list are systematically covered by people belonging to the same social groups. So, in social terms, there is a substantial difference between individual and group discrimination.⁵

The first screening may generate a consistent discrimination toward categories of candidates. An applicant’s resume is his/her self-introduction tool. It is expected to picture his/her education and training experience, his/her attitudes and expectations in absence of the possibility to show them to a selector. All experiments evidenced that the real selection, say the limitation of firms’ interest to a bunch of applicants, is usually made by employers at the access-to-interview stage. This makes discrimination silent and incontrovertible. Even a small amount of prejudice held by many employers could add up to a high rate of discrimination in the labour market (Cohn 2000). On the other hand, Bøg and Kranendonk (2011) do not exclude that a significant discrimination in the callback rates is consistent with no discrimination at the final offer stage. It depends, they say, on the complexity of the overall recruitment process. The issue of when discrimination may be introduced during the selection process deserves attention by scholars.

Our experiment confirmed the presence of strong discrimination against foreign candidates; if we consider that our study focused on graduates, and Quillian et al. (2019) found that there is less discrimination in jobs that require a college degree than in those that only require a high school degree, we can expect the general labour market to be even more unfair towards foreigners. Consistently with the meta-analysis performed by Lippens et al. (2023), Middle Eastern applicants face the most unfavourable treatment in hiring. Conversely, there is no evidence of gender discrimination, or rather, female candidates are called back slightly more frequently than male candidates, in agreement with Schaerer et al. (2023) who found that, over the last decade, discrimination against female applicants has diminished and sometimes slightly reversed.

A methodological question relates to our experiment: is it adequate to highlight the discrimination in the labour market of graduates? McIntyre et al. (1980), Heckman and Siegelman (1993), and Heckman (1998), commenting on discrimination research by employers, argued that other inferences could be deduced from the published conclusions. Since the expected worker’s skills cannot be directly observed but only imagined at the hiring stage, we are forced to guess if and to what extent the information contained in a resume is a proxy for relevant factors. Among others, Phelps (1972) states that even skin colour or gender can be taken as proxies for relevant data not sampled. Whatever tool is used to highlight discrimination, it is possible that an indicator of prejudice precedes something else, for instance, the expected productivity of candidates. Heckman asks if the inequalities observed in market outcomes are a proof of discrimination in the market or, instead, does the observed variability reflect the expected difference in candidates’ skills? The juxtaposition between taste-based discrimination, led by irrational animosity towards minorities, and statistical discrimination, motivated by a rational optimising behaviour in response to limited information, has informed the economics studies on labour market discrimination for decades (Guryan and Charles 2013). Each response would lead to a different policy: the former might imply the necessity to reduce employers’ discrimination, the latter the necessity of policies for the improvement of skills and the social environment of certain categories of candidates.

One common strategy to exclude statistical discrimination is to provide a great amount of information in the fictional resumes. However, other characteristics of the candidates remain unknown to the recruiter and have to be guessed. In fact, a worker’s productivity might depend not only on his/her educational path, but also on his or her social participation in a firm’s activities and positive attitudes towards a company’s success. In other words, the worker’s personality traits are to be considered a matter of assessment at the insertion stage. Should a company enrol an applicant who shows a leaving attitude even before recruitment, or who showed in previous work experiences negative attitudes toward collective life, or else who might enforce the dominance of a particular ethnic or linguistic group within the firm? Baert and De Pauw (2014) find that employers’ concern that co-workers and customers prefer collaborating with natives drives discrimination. Neumark (2012) suggests a way to address, under some assumptions, the Heckman-Siegelman critique. Alternatively, a research design might include varying levels of information provided in the CVs, in order to manipulate the strength of productivity signals. Anyway, there is no simple answer to this question. Lippens et al. (2022), in a systematic review of the recent literature, claim that ethnic discrimination in the labour market is in general better explained by taste-based discrimination, while Quillian and Midtbøen (2021) question the relevance of this distinction and affirm that “statistical discrimination is no less odious or harmful than other forms of discrimination”. In our experiment, we tried to provide a great amount of information in the fictional resumes to exclude statistical discrimination, at least the one based on academic achievements: we declared the final degree mark, the English level and the ICT skills, along with personal characteristics and interests. The high levels of discrimination observed despite this precaution suggest a preponderance of taste-based discrimination, but a definitive solution remains elusive.

Another methodological issue stems from the fact that, for randomization purposes, we created a bank of resumes of essentially equally qualified applicants. As an alternative, Midtbøen and Rogstadt (2012) suggest creating resumes for two subsequent studies, a correspondence study and an in-depth interview of employers exposed to the fictitious applications. With an analogous purpose, Neumark (2012) upholds that a correspondence study should include some degree of variability for certain characteristics that affect the probability of being hired. So, once the researcher detects that a set of productivity-related characteristics of applicants has affected hiring outcomes, it is possible both to test for bias in the estimates of the two groups (related to the null-discrimination hypothesis and that of the discrimination hypothesis) and also to detect the actual discrimination level.

One interesting result comes from the comparison of the models for the callbacks within seven days and overall: employers tend to call back first Italian candidates and those with a master’s degree, while delayed callbacks are concentrated on foreigners and bachelor’s degrees. We can hypothesise two alternative reasons for this behaviour: on the one hand, fast reactions are proven to be more governed by stereotyping and discrimination (see, e.g., Casey et al. 2013; Payne 2006), then the employers who call back a candidate a few days after receiving his/her resume might be less likely to discriminate against foreigners; on the other hand, foreigners (as much as bachelors’ degrees) might be considered as a second choice, suitable after the job has been declined by other candidates, and these delayed callbacks might be a further symptom of discrimination. A larger sample of vacancies, and consequently a larger number of callbacks, might allow to model the response time and get a deeper understanding of the reasons why, on average, foreigners are called back later than Italians.

The study’s design includes some aspects that could represent potential limitations. For example, the focus on graduates from the same university and firms based in only one Italian region has consequences. On the positive side, we can test the existence of regional discrimination, we rule out the possibility that the employers be influenced in their choice by the unlike reputations of equivalent degrees attained in different universities, and we control for differences in recruiting strategies that might be due to the characteristics of the various local economies. On the other hand, although their consistency with the results of analogous studies, we cannot be sure that our findings would be the same at a national level; Busetta et al. (2020b), for example, find that taste-based discrimination in Italy is higher in the Northern and Central regions than in the South.

The study implicitly assumes that all candidates, regardless of national origin, possess full command of Italian, as they are presented as graduates from an Italian university. This reduces the realism of the experimental design. In fact, many Italian universities now offer entire degree programs in English, which means that some foreign graduates may not have strong Italian language skills. In this sense, the study captures discrimination conditional on equal Italian proficiency, rather than reflecting the broader labour market situation where language differences may also play a role.

As previously highlighted, the study focuses on vacancies posted on the website subito.it. Strictly speaking, we had no sampling problem since openings were inserted in the experiment without sampling. Indeed, the type of jobs advertised through the Internet was peculiar, as evidenced in Sect. 3. Although all the selected ads were addressed to graduates, a considerable part of the examined ads concerned medium-to-low profile jobs. In order to limit the observed preference for graduates of social sciences and humanities, we should have sampled firms from a possible list that posted job ads representative of the whole spectrum of insertion wages and positions. In particular, we should have considered a larger quota of jobs from a hypothetical technical and higher-profile list. Would this possibility have given no evidence of discrimination, or possibly of a different kind of discrimination? This is difficult to say. Heckman himself, after suggesting a similar ‘purposive sorting’ of firms to solve the problem he saw in the audit-pair method, admitted that there may be “evil lurking in the hearts of firms that is never manifest in consummated market transactions”. On the other hand, Auspurg et al. (2020), in a study on racial discrimination in housing, compared results of field experiments drawn from different sources and found that results change very little, suggesting that field experiments be not sensitive to sampling methods. In this sense, the bias induced by this selection of vacancies might be less relevant than feared, especially if we consider that the focus is on recent graduates and entry-level jobs.

The two-level random intercept logistic regression model addressed the problem of the strong dependence of observations within vacancies, though the second-level covariates, referred to the company and the vacancy position, were scarcely effective in explaining the variability of decisions that depend on the recruiters more than on the companies. Finally, the study design itself, with five resumes sent for each vacancy, implies some limitations, as the limited groups size discouraged from introducing random slopes in the model. In this way the potential of multilevel modelling is to some extent weakened. Future studies might consider sending a larger number of CVs for each job vacancy, as in Firth (1981; 1982) or in Busetta et al. (2018), but this would raise the risk of suspicion and might ruin the experiment.

Our experiment concerned the detection of the variables that affect the variability of callback rates obtained in response to the sending of resumes to firms. The analysis showed that firms vary in propensity to contact candidates for a job interview according to nationality, level and major of university degrees. We found evidence of discrimination against foreign candidates and an amount of discrimination in which prejudice operated inversely than expected.

Discrimination against African candidates was higher for females than for males, while females from the Middle East are less discriminated against than males from the same area; this is consistent with the findings of Arai et al. (2018) and Dahl and Krog (2018), who found that Arabic men face stronger discrimination in the labour market than Arabic women. No substantial gender difference was found among candidates from Eastern Europe.

But why are there differences in discrimination? Bail (2008) and Fullin (2016) stress that religion, on top of somatic differences and cultural habits, may explain how applicants are perceived in the labour market. Sniderman et al. (2002) highlight the peculiarity of Italian immigration, which includes that from countries like Albania and Romania, linked to Italy by cultural and political ties that might give these immigrants a more privileged position in the labour market. Koopmans et al. (2019) and Quillian and Midtbøen (2021) claim that the perception of cultural distance increases discrimination. Though, all this explains just part of our results. Applicants from Eastern Europe suffer the same discrimination level as Africans, while Middle Eastern applicants, who are likely perceived as practising a different religion as Africans but are somatically closer to local people, and then should be preferred to black people, are far back in employers’ taste.

Following Sniderman et al. (2002), we conjecture that, in Italy as elsewhere (Booth et al. 2012; Quillian et al. 2017; Rinne 2018), there is a hidden hostility to immigrants, whatever their provenance. Insecurity is an additional feeling causing the low callback rate to both Eastern Europeans and people from the Middle East. Prejudices seem more influenced by the pictures of ethnic groups used in the mass media, than by somatic differences.

What is somewhat unexpected is the ‘inverse’ discrimination against graduate applicants, since we found that social sciences and humanities degrees were preferred to more scientific and technical ones. Although a bachelor’s or master’s degree was required, most of the selected vacancies concerned non-technical positions, for which social sciences and humanities graduates might be considered better aligned. Nevertheless, if a candidate with a more technical background shows interest for a nonspecific job, what prevents the employer from calling him back and make an interview? We can hypothesise that the higher proportion of callbacks to social sciences and humanities graduates also indicates a preference for lesser-demanding candidates, and the will to avoid losing time in job interviews to candidates that might refuse the job. Although the disadvantaged category be different in this case, this result is analogous to that found by Neumark et al. (1996), where the weaker category (in that case, women) was discriminated against in the higher-level workplaces (in that case, high-price restaurants), while it was favoured in the lower-level ones. The examined job offers considered in this experiment were addressed to less-demanding, less-entitled applicants, and can then be assimilated to the low-price restaurants in Neumark et al. (1996) experiment.

We found that the Internet conveys some entry-level, medium-to-high profile jobs, ideally suitable to graduates, but many more short-term, maternity-leave and the like substitutions, and entry-paid jobs. Based on the assumption that a more expensive ad through a newspaper indicates that employers put more emphasis on that job, we should sample ads separately from that channel and measure how this source differs. If the hypothesis of correlation between ad sources and expected wages holds, other conveyance channels should be sampled (public and private employment agencies, job offers conveyed through university channels, etc.). Indeed, a stratified-by-source sample of ads would merit additional research.

For future experiments, we list some questions that puzzle us. Which realistic characteristics should qualify a resume? We could distinguish part-time positions, season in which sampling was done, distance between home and work, candidate’s previous work experience or involvement in political/union/ethnic/environmental/other social activities, and like characteristics that could influence employers’ response. Intuitively, these variables are relevant alone but their relevance should be evaluated balancing them with other possibly favourable characteristics. Moreover, since unobservable error variance may mislead inference in correspondence studies, the richer a (fictitious) resume in terms of applicant’s characteristics, the more it is reasonable assuming that there are lesser differences in unobservable characteristics across groups of potential employees and across social environments for the estimation of discrimination.

Because a resume is a unique body that includes many aspects of the experience, personality and expectations of applicants, and the selection process of resumes is an employer’s choice, one could use conjoint analysis (Louviere 1988) as a method to highlight the aspects that increase and those that decrease the callback probability and the compensation capacity of the former ones to fight discrimination.

Moreover, how should we spread information to circumscribe the risk of suspicion without giving relevant hints that may ruin the experiment? Is that risk correlated to the number of candidates per job opening? In literature, the minimum is one and the maximum varies according to circumstances and researcher’s preferences. Among others, Weichselbaumer (2020) sent only one resume per ad; Newman (1978) and McGinnity and Lunn (2011) adopted a number of two resumes; Hubbuck and Carter (1980), Brown and Gay (1985) and Adida et al. (2010) used three; Firth (1981) sent seven resumes; Firth (1982) and Busetta et al. (2018) even more. Which, if any, is an optimum number of resumes for such an experiment? Finally, does a typo or other expressions of inadequacy in a resume have the same effect in a job application of a native than a foreign candidate?