Employment Mismatches Drive Expectational Earnings Errors among Mozambican Graduates

Abstract

Biased beliefs about future labor-market earnings are commonplace. Based on a longitudinal survey of graduate work transitions in Mozambique, this study assesses the contribution of employment mismatches to a large positive gap between expected (ex ante) and realized (ex post) earnings. Accounting for the simultaneous determination of pecuniary and non-pecuniary work characteristics, employment mismatches are found to be material and associated with large earnings penalties. A decomposition of these expectational errors shows that around two-thirds are attributable to employment mismatches, suggesting job seekers systematically overestimate the ease of securing “good jobs.”

1. Introduction

Prior to COVID-19, many developing countries recorded steady progress in expanding access to education, putting the average attainment of working-age adults on an upward trend (Porzio et al. 2022). Nevertheless, access to quality employment and youth un(der)employment remain ongoing challenges, particularly in Sub-Saharan Africa (Pauw et al. 2008; Banerjee and Sequeira 2020; Bandiera et al. 2022). Even in countries where unemployment has been modest, changes in workforce composition often outpaced shifts in the structure of employment—i.e., available jobs are not always well aligned to workforce skills (e.g., Sevilla and Farías 2020).

This paper investigates the extent to which systematic divergences between expected and realized labor-market earnings might be accounted for by employment mismatches. In the context of developing countries, such expectational earnings errors are often large and positive, especially among young people. Abebe et al. (2021) run a beliefs-elicitation experiment among job seekers in Ethiopia. They find that while beliefs about the distribution of wages are largely accurate, the expected probability of finding a “good job” (e.g., a permanent or professional appointment) is highly skewed toward over-optimism. This type of biased forward-looking belief constitutes one of the various kinds of matching friction in developing country labor markets (see also Banerjee and Chiplunkar 2018; Abel et al. 2020; Abebe et al. 2021; Jones and Santos 2022).

Our working hypothesis is that matching frictions, as well as possible rationing of high-quality jobs, are not fully internalized by job seekers, which in turn drives optimistic earnings expectations. To investigate this, we draw on a representative longitudinal survey of Mozambican university graduates as they transition into the labor market. As we show, although this group represents a small share of the working-age population, they do not easily find high-quality jobs. Our survey collected information on both expected earnings and anticipated occupational outcomes as elicited prior to graduation, as well as realized labor-market outcomes collected on a quarterly basis over a period of 18 months after concluding their studies. We observe that expectational errors are positive, large, and persistent (over the short term). On average, while around three-quarters of the sample undertook some paid work within 18 months of finishing their studies, their starting salary was less than half of what they had originally expected. At the same time, job mismatch is significant, with many individuals only finding precarious forms of work—e.g., just one-third of workers obtain fixed positions, just over half have a written contract, and fewer than half work in the sector they had originally anticipated.

We analyze the link between mismatch and expectational earnings errors in three steps. Guided by a conceptual framework , we first create a composite measure of the non-pecuniary quality of an individual’s labor-market position, covering a spectrum from complete detachment to revealed satisfaction in stable formal employment aligned with their skills and preferences. This index confirms recent graduates satisfy fewer than half of all quality domains. Further econometric analysis shows observed employment (match) quality is systematically associated with individual characteristics (e.g., gender and study field), as well as stated future expected earnings and recent labor-market experiences.

Second, following studies on the costs of mismatch (e.g., McGuinness et al. 2018; Brunello and Wruuck 2021), we investigate the pecuniary implications of employment match quality. Accounting for both selection into wage work and the likely endogeneity of employment match quality, we find that for the average graduate in paid employment, realized mismatches account for a discount of 50 percent in earnings relative to a scenario of no mismatch. Third, evaluated against stated earnings expectations prior to labor-market entry, a decomposition of the sources of expectational errors shows that employment mismatches are the dominant component, accounting for two-thirds of the error on average.

Our analysis contributes to existing literature in four main ways. First, we extend the scope of prior work on employment quality and mismatch to a relevant new context, encompassing multiple facets of employment. In doing so we connect to previous studies of specific aspects of job mismatch (see below), but place them in a wider context. Second, recognizing that wages are just one attribute of a job (Cassar and Meier 2018), we go beyond a conventional focus on the binary extensive margin of wage employment; rather, we use a conditional mixed-process estimator to address the simultaneous determination of pecuniary and non-pecuniary job attributes. Third, we use panel rather than single cross-section data. This has the advantage of allowing us to assess the extent to which job quality evolves over time and permits controls for individual-specific unobserved effects. Fourth, complementing Abebe et al. (2021), we link our estimates of the pecuniary costs of mismatch to expectational errors, providing a proximate explanation as to why these errors are systematically positive and large.

This study also connects to Jones and Santos (2022), who study the same context. Their analysis considers the evolution of future earnings expectations in response to new information, showing a strong degree of inertia or weak responses to negative news. Nonetheless, the authors argue this can be rational in a dynamic setting where participants believe they have a non-zero probability of obtaining a better job in future. The present study complements this analysis, focusing directly on the monetary costs of mismatch in their current job—i.e., here we demonstrate that if all participants did indeed land good jobs, the gap between an individual’s prior expectations and realized earnings would fall dramatically.

2. Background

An enormous literature addresses questions of employment quality and mismatch. Nevertheless, two general observations merit note. First, employment mismatches appear pervasive, both in developed and developing countries, but take various forms. Adopting a broad notion of mismatch (see further below), encompassing both quantity and quality dimensions, conventional unemployment trivially constitutes an extreme form of mismatch. However, since some individuals can be discouraged (not searching) or opt to pursue precarious (non-wage) activities as a substitute for a preferred type of work, standard (un)employment classifications may not fully capture the extent of quantity-type mismatches, especially in contexts where social security nets are limited (for discussion see Baah-Boateng 2016; Betcherman and Khan 2018).

With respect to job-quality-type mismatches, there is ample evidence that objectively poor employment conditions are commonplace. For instance, the “OECD Framework for Measuring and Assessing Job Quality” defines job quality under three main dimensions—earnings quality, labor-market security, and the working environment. Applying the framework to a range of countries (OECD and non-OECD), Cazes et al. (2015) find significant cross-country variation, but also generally much lower job quality in middle-income countries, noting that “informal workers have lower earnings quality, face a higher risk of extremely low-paying jobs and a higher probability of working very long hours” (p. 34).

A related strand of scholarship considers the match between what workers do and their capabilities or preferences. A common focus here has been on educational (vertical) mismatches—namely, being over- or under-educated for a given role (e.g., Kampelmann and Rycx 2012; Kiersztyn 2013). Synthesizing numerous studies, Leuven and Oosterbeek (2011) point out that the aggregate incidence of both over- and under-schooling appears material (at around 30 percent in each case), but warn that such measures are prone to error. McGuinness et al. (2018) find a similar average rate of incidence of over-education across various contexts, but a lower rate of under-education. In the African region, Bandara (2019) finds a broadly similar pattern (see also contributions in Herrera and Merceron 2013), but Morsy and Mukasa (2020) find the opposite, pointing to higher rates of under- as opposed to over-education, estimated at 57 and 8 percent of workers respectively.

Other dimensions of job (mis)match have been proposed (for one typology see McKee-Ryan and Harvey 2011). Expanding the notion of educational mismatch to the task content of jobs, mismatches with respect to professional skills and field of study have gained emphasis (Altonji et al. 2016; Montt 2017; Guvenen et al. 2020). Following the Great Recession, attention has also been given to occupational and sectoral mismatches, which can arise when there is a structural imbalance between labor-demand conditions (vacancy rates) and job seekers’ interests (e.g., Shimer 2007). For example, an excess supply of individuals in particular occupations may place downward pressure on wages in these activities, inducing job seekers to remain unemployed or accept less-preferred positions. Similarly, mismatches may arise from involuntary underemployment (working fewer hours than desired; Brown and Pintaldi 2006; Frei and Grund 2022) or distaste for specific employment conditions (e.g., commuting distance; Andersson et al. 2018).

Second, job mismatches are often associated with significant earnings penalties (Nordin et al. 2010; Bender and Roche 2013; Somers et al. 2019). In Canada, for instance, Lemieux (2014) estimates that a measure of occupation–field-of-study relatedness accounts for a material fraction of returns to (graduate) education—i.e., where this mismatch is substantial, wage returns to higher education appear lower. Similar results have been found in developing country contexts. Carmichael et al. (2021) find an approximate 18 percent discount to earnings among over-educated workers in Ghana and Kenya (also Sam et al. 2018; Bahl and Sharma 2021). And, relatedly, Carmichael et al. (2022) show that workers matched to jobs with a strong fit to their skills and interests tend to be more satisfied, independent of their level of earnings.

Notwithstanding this literature, gaps remain in our understanding of mismatch. On the one hand, the vast bulk of evidence pertains to advanced or (to a lesser extent) middle-income countries. On the other hand, and despite the acknowledged multi-dimensional nature of job quality, most studies of the pecuniary costs of mismatch limit attention to specific aspects, such as educational matches, providing only a partial analysis of the phenomenon. Furthermore, existing studies are primarily cross-sectional in nature and, therefore, have not investigated mismatch as a proximate source of expectational error.

3. Conceptual Framework

At the core of our analysis is the relationship between earnings and employment outcomes. Following earlier discussion, we integrate both job quantity and quality considerations, taking a view that employment quality is usefully viewed on a continuum ranging from complete detachment to full integration in meaningful work. So, in contrast to most existing approaches, we do not make a sharp distinction between quantity versus quality attributes. As Herrera and Merceron (2013) put it, “Classic indicators of unemployment and time-related underemployment [are] inappropriate, if not misleading, for understanding labor markets in Sub-Saharan Africa....[Rather] mismatch should be considered an extension of the underemployment notion” (pp. 103–4).

To fix ideas, we focus on the link between non-pecuniary and pecuniary employment attributes. Consider a simple setup in which job seekers face three potential labor-market states: unemployment, a temporary paid apprenticeship, or a salaried permanent position. The non-pecuniary attributes of each state are summarized by an index q, which is increasing in all its arguments. In this basic environment two such arguments are sufficient: j, which takes a value of 1 if the individual is working, and p, which is 1 if they have a permanent position. Thus, q(j = 1, p = 0) is the non-pecuniary employment quality index associated with temporary paid apprenticeships. And we assume higher values of q, reflecting better non-wage amenities, are preferred by workers (see Sockin 2022; Lavetti and Schmutte 2023).

In addition to non-pecuniary employment attributes, workers also value pecuniary attributes, given by their wage w, which varies according to their labor-market status and other characteristics. Following Cassar and Meier (2018) among others (e.g., Cremer et al. 1995), the expected utility associated with a specific employment state is |$U^{*}(w_i^{*},q_i)$|⁠, where asterisk superscripts denote latent (partially observed) variables. Presuming individual i is willing to remain unemployed if their expected utility associated with available alternative states falls below a reservation threshold |$\underline{u}_i$|⁠, this structural process can be described as

The expression on the left describes the process governing selection into paid activity, and the other says we observe (non-zero) labor-market earnings only when selection into work is positive. This constitutes a standard Heckman (Tobin-2) sample selection framework, but does not explicitly account for choice over a broader range of alternative non-pecuniary employment attributes. Hence, a more general formulation is

This clarifies that an individual’s chosen employment status is observed in all outcome states, but a wage is observed in just some. Moreover, it can be seen that selection into (paid) work can be expressed in terms of a threshold condition on q. By assumption, since q(j = 1, p = p_i) > q(j = 0, p = p_i), then there exists a threshold |$\underline{q}_i = \max (q_i \mid j=0)$| such that |$U_i^{*} \lt \underline{u}_i \Rightarrow q_i \le \underline{q}_i$|⁠.¹ In other words, the conditions under which we observe an individual’s wage also can be expressed as a function of their non-pecuniary employment status.

To place empirical meat on these bones, we propose the following equations to describe the wage and employment quality offers:

where x is a vector of (given) individual characteristics, q₀ represents anticipated employment quality as stated at baseline (t = 0), y is a vector of time-varying factors that reflect determinants of their unemployment reservation utility |$\underline{u}_i$|⁠, and the final line indicates that we assume the error terms follow a bivariate normal distribution.²

This framework, which nests the Heckman selection model as a special case (for q_i = j_i), structures our thinking about the joint relationship between non-pecuniary and pecuniary work attributes. It clarifies that employment status (q) may be related to labor-market earnings (w) via two channels: indirectly through sample selection, captured by the error covariance term ρ, and directly in terms of the pecuniary value attached to non-pecuniary attributes among individuals in paid work, as per parameter δ. This framework guides our empirical analysis, which we start with a brief review of the context and data at our disposal.

4. Context and Data

Our context is Mozambique. Like other (low-income) countries, it has witnessed rapid growth in access to education at all levels over recent decades (Jones et al. 2023). In the tertiary sector, the number of students graduating from higher education each year has risen dramatically, from under 700 in 2003 to over 18,000 in 2016 (Jones et al. 2018), implying an annual growth rate of around 30 percent. However, educational expansion has occurred from a very low base and stocks of tertiary-educated workers remain some of the lowest in the world. Recent statistics from the 2017 population census indicate that fewer than 2 percent of Mozambicans aged 15 and over have completed studies at the bachelor level or above (in a country of over 25 million inhabitants).

Given their scarcity, one might think university graduates find good jobs easily. However, the jobs environment in Mozambique is challenging. Formal employment remains limited— e.g., less than 12 percent of all workers report receiving a wage income and this proportion has increased only slowly over time, even in urban areas (Jones and Tarp 2016a,b). In turn, competition for such jobs is extremely high. Over 400,000 young people enter the job market each year, while opportunities for non-agricultural employment remain thin and are found largely in the (informal) services sector. Since around the mid-2000s, economic growth has become increasingly driven by extractive industries, which are capital intensive and reliant on foreign workers to fill key technical and managerial positions. Thus, neither rapid nor sustained growth in demand for workers with a university education has been forthcoming.³

The above motivates our interest in the experience of university students as they transition into the labor market. In the absence of existing systems to track graduate career trajectories, in 2017 we implemented a face-to-face survey of over 2,000 students studying in their final year of an undergraduate degree programme. This survey, hereafter referred to as the baseline and described in Jones et al. (2018), was designed to be representative of the population of Mozambican university final year undergraduates enrolled at the six largest public and private universities in the country, stratified by gender and study area (viz., Education, Humanities, Social Sciences (including law), Natural Sciences, Engineering, Agriculture, and Health).⁴

Our baseline survey collected extensive information on personal characteristics, educational and professional histories, cognitive abilities, and future labor-market expectations. Starting from early 2018, after their studies should have been completed, we re-contacted the same individuals six times by telephone on a quarterly basis. On each occasion we collected data on their employment situation, including realized earnings, type of work undertaken, and employment outlook. We refer to these later rounds as the follow-up or telephone surveys, numbered consecutively (one through six).

As per fig. 1, of the 2,175 finalists surveyed in the baseline (1,024 women and 1,151 men), the vast majority expressed a concrete intention to (seek) work after finishing their studies, of which 2,081 consented to participate in the follow-up telephone rounds. Of these, we were later able to track 2,050 individuals at least once during the six follow-up rounds. This group constitutes our core analytical sample—i.e., individuals for whom we have meaningful baseline information on employment expectations (before entering the labor market) as well as at least one observation on later labor-market outcomes.⁵ Naturally, not all individuals who intended to seek work eventually did so. Over the course of the six follow-up rounds, 1,577 found some kind of paid work, implying around 25 percent of the core sample did no work.

Table 1 summarizes further information from the survey, distinguishing between different rounds. Panel (a) reports the number of unique individuals in each such subsample—namely, moving from left to right, our core analytical sample and the phone surveys (in consecutive pairs). This confirms 92 percent of the core sample were successfully re-contacted in at least one of the last two follow-up rounds, nearly two years after the baseline. Panels (b) and (c) report survey-weighted averages of individual characteristics as collected at baseline, where (c) denotes their aggregate study fields. Since all the variables in these two panels are time invariant (they were not recaptured during the follow-up rounds), comparisons across the columns indicate the extent to which the composition of the sample altered over time due to attrition. Visual inspection indicates no substantive changes. To test this, we regress each variable against dummy variables for each follow-up round, reporting the joint probability these are not different from zero. The results confirm no statistically significant compositional changes over the follow-up rounds.⁶

Panel (d) compares data on baseline labor-market expectations (elicited while still at university) against later realizations (having completed their course). Specifically, in the baseline we asked participants to state their expectations as regards how long it would take them to find work and their preferred type of job (employer and sector), as well as their anticipated net monthly earnings. Thus, averages reported in column (I) of the table indicate these expected outcomes. The remaining columns of panel (d) report corresponding realized outcomes. For instance, individuals in the core sample expected to find a job within six months on average; however, only about half of all (contacted) individuals had undertaken some paid work in the first six months of the follow-up rounds and, by the final two rounds, the average individual had spent 10 months searching. Additionally, the vast majority of individuals both intended to and eventually worked in a service-related job, but with more individuals in educational-related jobs than originally expected. This gives a preliminary indication of the existence of mismatches with respect to non-pecuniary attributes of employment (see further below).

With respect to earnings, the baseline survey asked individuals to estimate their “expected monthly earnings (after tax) in the first month of work after graduation.” Since we surveyed individuals in their final year of study, we assume all individuals intended to complete their course. Even so, in Mozambique, formal graduation only occurs after completing a final thesis, which (as became apparent) is often delayed. Consequently, not only did many individuals take some time to find a job, but among those that did, they often started work before having formally graduated. As the table shows, in the first two follow-up rounds, only 14 percent of individuals had formally graduated, but nearly half had started some work (see also supplementary online appendix fig. S1.1).

At the baseline, the average expected monthly starting salary was about US$430 per month (after tax), which compares to a minimum wage of around US$100 per month.⁷ However, throughout the follow-up surveys, realized net monthly earnings were substantially lower than anticipated—on average, individuals reported to be earning less than US$250 per month or about half their baseline expectation. Focusing on the subsample of individuals observed in their first job after graduation, fig. 2 plots (a) the cross-sectional distributions of their expected and realized earnings, and (b) the individual-specific differences (in US$). The latter show that fewer than 20 percent of respondents received a salary that met or exceeded their earlier expectations, and close to 50 percent reported receiving at least US$200 less than originally expected. And while the presence of a positive expectational error is not necessarily surprising, the magnitude of the error is large compared to studies elsewhere (see Jones et al. 2020).

5. Non-pecuniary Employment Quality

5.1. Index Construction

To implement our analytical framework , we build an index of non-pecuniary employment quality. We start by enumerating a detailed set of the revealed (ex post) non-monetary attributes of work captured in our survey. As shown in table 2, these cover five main dimensions with a total of 16 underlying components, all of which take a value of 1 if satisfied and 0 otherwise (see supplementary online appendix S2 for further notes on variable construction). The selected components encompass aspects of labor-market attachment (being in work; working full time), job satisfaction, perceived employment readiness (completion of degree; not pursuing other outside training), and more specific attributes of their employment position (if any).

We include some features normally associated with employment quantity for two specific reasons. First, recall that our survey purposefully elicited future labor-market expectations and we restrict our analysis to individuals with an explicit interest in finding work.⁸ Thus, in our case, not being in work can be considered a pertinent form of mismatch given stated prior expectations. Second, as already noted, there is substantial overlap between concepts of labor-market attachment, under-employment, and job quality, particularly in contexts where unemployment is something of a luxury. So, rather than imposing a sharp distinction at some definitional boundary of employment, we locate all individuals on a continuum from that of complete mismatch to full alignment with their preferences as stated in the baseline survey.

Column (I) of table 2 reports means for each component for the full sample in all follow-up rounds, where specific employment attributes (e.g., vertical match) are set to “worst-case” zeros for individuals out of work, reflecting their detachment from the labor market and thereby non-alignment with prior expectations. Column (II) restricts the sample to those reporting to be in work. Regardless, there is clear evidence that average employment quality is poor in many dimensions. For example, only 60 percent report to be working, and, of those in work, around one in seven (12 percent) are unpaid (e.g., interns), 33 percent have a fixed or permanent position, just 58 percent state their job requires a university education (vertically matched), and 70 percent work in a different field to that of their university degree.

Figure 3(a) displays the same information by follow-up round, scaling each item by the inverse count of components, meaning the height of each bar would equal 1 if all participants satisfied all quality criteria. The figure confirms the low general quality of employment, with an average score across dimensions and periods (bar height) of 0.40. However, there is evidence of improvement over time—the height of the bar rises by around 50 percent, or 16 percentage points, between the first and last rounds, driven primarily by improvements in job stability, as well as skills and match factors. For instance, in the first round fewer than 5 percent of all participants had formally graduated, rising to about two-thirds in the final round.

Aggregation into a single composite quality index is helpful, not least to give a global summary. To do so, we must decide how to weight (combine) different underlying dimensions. Among various options, item response theory (IRT) has been widely used to estimate the value of an underlying factor based on observations of different items across units (e.g., Sandefur 2018). For each item (here, the components), the probability associated with a specific outcome is related to the value of a latent trait, with estimated model parameters reflecting the difficulty and discrimination of each item. To give a sense of this, column (III) of table 2 reports the anti-logs of the estimated difficulty parameters associated with each component from a one-parameter IRT model. These show a negative correlation with the mean proportions reported for each component (column I), implying components achieved by fewer (more) individuals are associated with higher (lower) difficulty.

Our preferred approach to construct a composite employment quality index uses a two-parameter IRT model, where discrimination parameters are fixed within but not across domains.⁹ These results are summarized in fig. 3(b), which plots means of the standardized index (latent trait), by follow-up round and by gender.¹⁰ The solid lines show our preferred measure, while the dashed lines provide a comparison based on the predicted score from the first principal component of the same underlying variables (PCA factor loadings are reported in column (IV) of table 2). Results from these two approaches are qualitatively almost identical, with an overall Pearson correlation of over 0.95. They indicate significant gains in non-pecuniary employment quality (especially in the last two rounds), but also large and persistent gender gaps which we discuss below. Last, supplementary online appendix fig. S1.3 plots the empirical distributions of our preferred index, distinguishing between individuals in and out of work. As might be expected, the mass of the latter distribution is far to the left (below zero). And while there is substantial heterogeneity in employment quality among the employed, there also is some overlap with non-workers—around 1 percent of observations are encountered in this region, which can emerge because non-workers can score positive values on certain dimensions of the index (e.g., rejected offers; see table 2).

5.2. Proximate Determinants

In the next step we assess the proximate determinants of non-pecuniary employment quality. Supplementary online appendix table S1.1 summarizes results of regressions following equation (1a), reporting selected coefficients only. As shown in supplementary online appendix S2, which reports the full list of variables employed, we rely on a very rich set of individual characteristics, including objective scores for ability (academic and fluid). Anticipated job attributes (q₀) are also included in the model, including indicators for an individual’s preferred sector and type of employment, how long they expect to take to find a job, and, to capture omitted individual effects, their expected level of future earnings (denoted by w₀). Columns (Ia)–(Id) are OLS estimates, treating the quality index as continuous; columns (IIa)–(IId) apply an ordered logit estimator, relaxing the assumption that the index is interval level (reported coefficients here are estimated odds ratios).

To get a sense of the relevance of selection effects into paid work, we start in subcolumns (a) by restricting the sample to paid workers only and use just baseline information on individual characteristics and employment preferences, as well as period fixed effects, as explanatory variables. In columns (b) we expand the sample to all participants, and in columns (c) add variables in vector y, which capture accumulated recent labor-market experiences. Concretely, these are a metric of lost experience, being the difference between the number of years of professional experience they had hoped to have acquired since completing their studies by the interview date (taken from the baseline) minus their actual experience;¹¹ the count of prior periods, not including the present round, in which the individual was unemployed; and the count of prior periods in which the individual was inactive (neither in work nor looking for work). Assuming the cost of being out of work increases with time, we expect larger values for these variables to be associated with downward pressure on the reservation threshold above which they accept a job offer. Last, columns (d) add individual-specific random effects.

The results point to various systematic correlates of employment quality. Measures of fluid and academic ability, as well as self-reported academic level and English language proficiency, are strongly associated with obtaining a better quality labor-market position. Prior work experience also predicts better quality employment, either through previous internships or past work; and individuals anticipating work in commercial firms or the education sector later tend to find themselves better placed in the labor market. The natural logarithm of future earnings expectations enters as an additional positive explanatory factor, plausibly capturing other (unobserved) individual-specific information regarding their skills and/or market potential. However, this term is only significant among paid workers, not in the sample as a whole. Also, as per fig. 3, there is a clear female disadvantage—holding other variables fixed, women score almost 0.2 standard deviations lower in employment quality then men on average, or have around 30 percent lower odds of being in a higher quality category (column IIb). But this relationship switches sign if we restrict attention to paid workers, suggestive of a marked gendered selection effect related to labor-market attachment or entry.

Reflecting the time-invariant nature of baseline covariates, survey round effects are positive and significant, particularly for the full sample. The additional time-varying variables are also material, substantially increasing the explanatory power of the model (see column Ic). In particular, loss of potential professional experience, consistent with not finding work within an anticipated duration, is associated with a large reduction in predicted employment quality. Accumulated prior labor-market experiences are also relevant, but flip sign once individual random effects are introduced, suggesting the presence of some material unobserved preferences or individual attributes associated with recent experiences.

6. Earning Implications of Employment Quality

We now ask how much variation in employment quality matters for subsequent earnings. To connect to issues of mismatch, in the analysis of earnings we apply the following approximation:

The first line trivially states that in the absence of mismatch, realized employment quality will align perfectly with earlier expectations. However, as only incomplete information regarding these expectations is available, the second line makes two assumptions. First, we approximate q₀ using a set of variables encompassing anticipated job attributes (z₀) and future earnings (w₀) taken from the baseline survey; second, we presume that individuals report future earnings expectations in the anticipation of finding good or well-matched work—i.e., frictions that may lead them to take a poorly matched job are discounted. Thus, we define |$q_{it} - \bar{q} = \tilde{q}_{it}$|⁠, which takes a value of 0 when there is no mismatch, representing a measure of employment match quality. And, since this term constitutes a linear transform, we use this latter measure throughout.¹²

Our main results, focusing on the earnings equation (1b), are summarized in table 3, again showing selected coefficients. Supplementary online appendix table  S1.2 reports results for the corresponding selection equations, which are qualitatively similar to previous results. For reference, we start by estimating two naïve models that do not account either for possible selection effects or for the likely endogeneity of realized employment quality. Column (I) regresses the natural log of labor-market earnings against baseline control variables plus round effects, while column (II) adds the match quality index, as per the approximation in equation (2), treating this term “as if” exogenous.

The remaining columns build up to our full model. Column (III) replicates a conventional full-information Heckman sample selection model, using a binary indicator for having a salaried job (one component of our employment quality index) as the dependent variable in the selection model (cf. equation (1a)). Here we exclude both q₀ and y_t from the selection equation, thereby only relying on functional form assumptions for identification. Column (IV) adds the vector of proxies for q₀ to the selection equation, including earnings expectations from the baseline survey, as well as vector y_t. Column (V) extends the selection equation, replacing the binary selection indicator with a simplified version of q, in which all individuals without a salaried job are allocated to the lowest value of the index. Column (VI)—our full model—employs q directly as the dependent variable in the selection equation. Last, as per the previous section, column (VII) adds individual random effects to the selection equation, constituting an extra component of q₀.

The models in columns (I) and (II) are estimated by OLS. The remaining models are estimated using a conditional mixed-process estimator (Roodman 2011), which accounts for the partially observed nature of the main outcome of interest and where the estimated correlation between the selection and wage equations is captured by |$\hat{\rho }$|⁠.¹³ When specific variables enter the selection but not the earnings equation, these operate as excluded instruments to support identification—i.e., estimates in columns (IV)–(VIII) address concerns around both selection into remunerated work and the endogeneity of the employment match quality index.

Four main insights stand out. First, looking across the coefficients, various factors previously noted as being associated with better quality employment are also associated with high earnings—e.g., English language proficiency, academic ability, and prior earnings expectations. The basic model reported in column (I) accounts for around 25 percent of the variation in earnings, which is broadly consistent with the explanatory power of earnings regressions in various other contexts. Second, employment match quality is a very strong predictor of earnings. The explanatory power (R²) of column (II) jumps to around 38 percent,¹⁴ and, regardless of the specification, we find a one-standard-deviation increase in match quality is associated with at least a 0.47 log point (approx. 60 percent) increase in predicted earnings. Third, the latter estimates rise in magnitude once bias associated with selection and endogeneity are addressed. Notably, estimates in columns (III) and (IV) are hardly distinguishable from those of column (II). But allowing for selection on our granular metric of employment quality as per column (V) onwards, the coefficient on |$\tilde{q}$| increases by around one-quarter.

Fourth, the estimated correlations between the selection and wage equation residuals are consistently negative, implying individuals with higher-than-predicted employment match quality tend to attract lower-than-predicted earnings (and vice versa). On the one hand, this may simply reflect a technical feature of the data. As Ermisch and Wright (1994) discuss, a negative correlation can emerge when, given observed characteristics, the variance of wage offers is smaller than the covariance between these offers and the reservation wage. On the other hand, this may well be capturing the effects of compensating differentials—e.g., at the top end of employment quality, which often coincides with work in the public sector, employers may offer a slightly lower wage in exchange for higher employment stability; but at the lower-quality end, some employers offer higher wages to induce individuals into work.¹⁵ Notably, this negative correlation is much larger once the individual effects are included in the selection equation. While not straightforward to interpret, this suggests an important role for unobserved individual factors, which once accounted for may only inflate the (possible) influence of compensating differentials.

7. Implications for Expectational Errors

We now turn to consider what these results mean for expectational errors, defined as the difference between each participant’s (earlier) estimate of their expected future earnings and (later) realizations of the same. As already shown, these errors are large, averaging 0.87 log points across all observations, implying expected earnings were about double the magnitude of realized earnings. Given the high incidence of employment mismatches, as well as their material pecuniary costs, we derive an adjusted or counterfactual expectational error. This is defined as the difference between expected and realized earnings under the assumption of no employment mismatch. Using our previous regression estimates, we thus calculate |$\Delta _{it}(\tilde{q}_{it} = 0) = w_{i0} - (w_{it} - \tilde{q}_{it}\hat{\delta } )$|⁠, which is just the raw expectational error less the cost of mismatch.

The footer of table 3 provides a summary of these calculations for each model, showing both the average cost of mismatch (⁠|$\tilde{q}_{it}\times \hat{\delta }$|⁠) and the corresponding counterfactual expectational error, where column (I) implies no adjustment relative to the observed error. Reflecting previous results, we see that once we model selection into paid work using the granular measure of employment match quality, the counterfactual expectational error falls dramatically—to 0.32 points in column (VI) and to around zero in column (VII). In other words, conditional on finding paid work, a substantial share of the expectational error can be attributed to low realized employment quality.

We complement this analysis in two ways. First, with respect to the system comprised of equations (1a)–(1c) we add a further equation for expected earnings as elicited at baseline, which we specify as |$w_{i0} = {x}_i^{\prime }\beta _{0} + {z}_{i0}^{\prime }\gamma _{0} + \psi _{i0}$|⁠.¹⁶ Recalling the current wage is specified in similar fashion (see equations 1b and 2), the expectational error thus can be decomposed as

where the first and second terms in parentheses reflect differences in the valuations of individual and job attributes between baseline and the current period; δ, as before, refers to the contribution of employment match quality relative to baseline expectations, and ϕ captures any unexplained correlation between baseline and current wages, such as due to other unobservable factors.

Column (I) of table 4 reports results for this expanded system of three equations (for q, w₀, w_t), showing selected coefficient estimates for current and baseline earnings respectively, while column (II) incorporates individual effects in the selection equation. The footer of the table provides a decomposition of the components of the mean expectational error, as per equation (3), where we aggregate groups of variables by summing their individual contributions. Despite some differences with respect to estimates of how certain individual and anticipated job attributes are rewarded (e.g., gender), these factors do not contribute materially to the aggregate expectational error on average. In contrast, and in line with the magnitude of previous estimates, we find that employment mismatches represent the dominant component of the expectational error.

As a second exercise we replace the composite match quality index with the set of variables constituting its subcomponents. Echoing the definition of |$\tilde{q}$|⁠, each enters as the difference in value between the observed outcome (e.g., having a written contract) and the best case (generally valued as 1). As reported in columns (III) and (IV) of table 4, these results not only confirm the strong positive contribution of non-pecuniary employment quality to earnings, but they also show no one single feature drives our earlier results. For instance, the earnings increment associated with the combination of being in full-time employment, having a written contract, and being horizontally and vertically matched is around 0.80 log points. Aggregating across all dimensions, the model in column (III) (without individual effects) yields an unexplained residual expectational error not different from zero. And, as before, inclusion of individual random effects only serves to increase the magnitude of the (aggregate) mismatch component.

8. Robustness and Sensitivity

As a final exercise we probe the robustness of our results as well as variation across subsamples. We start by running two-step versions of our previous results, which—although not likely to be as efficient—may be less sensitive to functional form assumptions underlying the maximum-likelihood estimators (see Puhani 2000). Thus, row (1) of table 5 reestimates the model from column (IV) of table 3, equivalent to applying a two-step Heckman sample selection model in which the inverse Mill’s ratio enters as a control function. Row (2) reestimates our full model from table 3, column (VI), where the control function term is now defined as the residual from a first-stage regression of the employment match quality index against baseline regressors and vector y, based on the full sample, and row (3) reestimates the model from table 3, column (VII) that adds individual random effects. Focusing primarily on estimates for the coefficient on |$\tilde{q}$| in the earnings equation (as reported in column I of table 5), as well as the corresponding mismatch cost, these results are highly consistent with those before.¹⁷ However, when individual random effects are included in estimates for the first stage, the second-stage results for δ are moderately smaller, not larger in magnitude as previously. This suggests the estimates from table 3, column (VII) may be somewhat unreliable or extreme.

Next, we rerun estimates for our full two-equation model using alternative restricted instrument sets. Specifically, we sequentially redefine y by selecting just one of the (excluded) instruments deployed in column (VII) of table 3. The relevant results are summarized in rows (4)–(7) of table 5. They show that estimates for δ do vary somewhat with the definition of y, perhaps due to differences in instrument strength, and that inclusion of random effects alone is sufficient to yield much larger estimates for the cost of mismatch than when they are excluded. Even so, the qualitative message, namely of large pecuniary costs driven by lower employment quality, remains consistent throughout. And, as shown in column (IV), these costs translate to large differences between observed and counterfactual expectational errors. For instance, in precise agreement with table 3, the expectational error in row (2) falls by over half, from 0.87 to 0.33, under the counterfactual of no employment quality mismatch.

In row (8) of the same table we further investigate the role of individual random effects. To do so, we exclude individual effects from the definition of y but add them to the earnings specification. The results closely correspond to the lower bound of estimates from table 3. The remaining rows of table 5 apply different sample restrictions, in all cases running the model from table 3, column (VI). Note that row (9), which excludes non-salaried participants from all equations, constitutes a (limited information maximum likelihood) instrumental variables estimator—i.e., it exclusively addresses possible endogeneity of |$\tilde{q}$|⁠, but not (bias due to) sample selection. Notably, these estimates for δ are substantially larger than the OLS estimates from table 3, column (II), suggesting the latter do indeed under-estimate the pecuniary costs of mismatch. However, the same costs reported in row (9) also are larger than those that also account for sample selection, suggesting the two sources of bias may work in opposite directions. Additionally, a Hansen J test of over-identifying restrictions from the same LIML estimator is insignificant at conventional levels (probability = 0.15), giving no cause to doubt the validity of the excluded instruments.

In rows (10)–(14) we restrict the sample by gender and consecutive pairs of follow-up rounds. While these do not suggest major new insights, we note the costs of low employment quality are moderately higher among men, but also increase over time despite the trend improvement in employment quality on average (see fig. 3b). This may be driven by differences in earnings growth associated with employment quality. For instance, if growth in earnings only occurred among workers above some threshold of non-pecuniary employment quality, then a stronger positive association between pecuniary and non-pecuniary factors would emerge. However, the negative selection effect also becomes stronger over time, perhaps driven by a more prominent role for compensating differentials as time goes on—e.g., individuals who take on jobs later in time are more sensitive to such differentials.

Last, in rows (15)–(17) we run our estimates for selective cross-sections. Recall, we elicited future earnings expectations in the baseline survey on a conditional basis—based on their “first job after graduation.” Consequently, we look at these two margins separately and in conjunction, highlighting that in these estimates we model q from the full sample but model w from the restricted sample only. The results again suggest some variation in parameter estimates, including a slightly lower estimate of the cost of employment mismatch in the final row when compared to previous estimates using all earnings observations (−0.45 versus −0.55). This indicates final completion of all studies may be associated with a transition to a better job. Even so, employment mismatch clearly remains a material factor both in quantity and cost terms throughout the sample period.

9. Conclusion

This paper set out to investigate the proximate sources of the large gap between expected and realized labor-market earnings. Using longitudinal data following the transition of a representative sample of Mozambican university graduates into the labor market, we documented significant employment mismatches. At the extensive margin, many individuals simply do not find work within the time they had anticipated (if at all), and at the intensive margin, they are often employed in poor quality jobs—e.g., without written contracts or working part time.

To quantify the importance of mismatches for earnings, we developed a composite index of the quality of non-pecuniary employment characteristics, ranging from complete labor-market detachment to revealed satisfaction in a “good job” aligned with their skills and training. Generalizing a conventional Heckman sample selection model, we simultaneously estimated the relationship between pecuniary and non-pecuniary employment outcomes, finding that mismatches—defined as deviations from the highest-quality jobs—are associated with large discounts to earnings. Our preferred model indicated that a one-standard-deviation increase in the employment quality index was associated with a 0.65 log point increase in earnings. So, for the average participant in work, observed mismatches implied at least a 50 percent discount to earnings compared to being in high-quality work, also implying lower returns to their education.

We used these estimates of the costs of mismatch to calculate the counterfactual expectational error, being the difference between anticipated and realized salaries under a scenario of no mismatches. These indicate that expectational errors fall by over half, from 0.87 to 0.32 log points. This is underlined by a decomposition exercise, which shows that while individuals do not systematically mis-estimate how individual and job characteristics are rewarded, mismatch accounts for around two-thirds of the expectational error. Extensions to this model indicate that mismatch costs pertain to virtually all dimensions of employment, and that our results are robust to different estimation choices and subsamples.

In summary, and consistent with an emerging literature, we find prospective workers tend to overestimate their prospects of securing high-quality jobs, making ex ante earnings expectations look optimistic in retrospect. The deeper sources of these errors are less clear. Evidence from the same survey suggests that individuals may be subject to behavioral biases, such as a representativeness heuristic, whereby information about specific individuals (e.g. high earners) is perceived as more relevant than generic salary information (Jones et al. 2020). Further understanding the sources of optimism with respect to labor-market matching, and possible policy responses, would be a valuable object of future research.

Data Availability Statement

The data underlying this article will be shared on reasonable request to the corresponding author.

Footnotes

Notes

Sam Jones, corresponding author, is a Research Fellow with UNU-WIDER, Mozambique; his email address is [email protected]. Ricardo Santos is a Research Fellow with UNU-WIDER, Mozambique; his email address is [email protected]. Gimelgo Xirinda is a Research Economist with the Manhiça Health Research Center (CISM: Centro de Investigação em Saúde de Manhiça); his email address is [email protected]. The study could not have been conducted without the support of the Ministry of Higher Education, Science and Technology and the Ministry of Labour of the Government of Mozambique, as well as the active collaboration of the University of Eduardo Mondlane, Catholic University of Mozambique, Pedagogic University, Polytechnic University, St. Thomas Aquinas University, and UniZambeze. We acknowledge the invaluable research assistance of Felix Mambo, Edson Mazive, Thomas Sohnesen, Vincenzo Salvucci, and Yonesse Paris. We further thank the editor, two anonymous referees, and participants at various conferences and seminars (UEM, Maputo; IESE, Maputo; UNU-WIDER, Helsinki; UNU-WIDER 2019 annual conference) for encouragement and helpful feedback. This is a completely revised version of WIDER Working Paper 2020/47 (‘Misinformed, Mismatched, or Misled?: Explaining the Gap between Expected and Realized Graduate Earnings in Mozambique’). All errors and omissions are our own. A supplementary online appendix is available with this article at The World Bank Economic Review website.

References