Data

In our empirical analysis, we have relied on two primary data sources. The first, the Moroccan Labour Force Survey for the years 2004 and 2018, provides a comprehensive dataset with individual demographic and socio-professional characteristics of employed population. To ensure consistency, our sample excludes occupations related to politics, diplomacy, religion, military/security and unpaid workers.

Our second data source is the U.S. Department of Labour’s Occupational Information Network (O*NET) database, offering detailed task content information for over 900 occupations. The use of O*NET data is prompted by the absence of national data-capturing task characteristics in Morocco. This aligns with a common practice in the empirical literature in studies on emerging or developing countries (Apella and Zunino Reference Apella and Zunino2017; Chang et al Reference Chang, Rynhart and Huynh2016; Das and Hilgenstock Reference Das and Hilgenstock2022; Adamczyk et al Reference Adamczyk, Ehrl and Monasterio2023).

While acknowledging the limitations of this choice, we do not assume a similarity between the U.S. and Moroccan labour markets. Rather, our emphasis lies in the ranking of task importance within occupations that is supposed to be more or less comparable across countries (Poole et al Reference Poole, Santos-Paulino, Sokolova and DiCaprio2017). In this respect, we preferred adopting the O*NET database over the World Bank’s Skills Measurement Program database (STEP).Footnote ⁴ Despite the latter being the first to measure task importance in certain low- and middle-income countriesFootnote ⁵ , it is appropriate not to adopt it due to its exclusion of rural workers. This omission would have resulted in overlooking a substantial part of the agricultural employment, a critical element in analysing the structural transformation process in developing countries.

Measuring routine task intensity

To measure RTI, we opted for the methodology proposed by Autor and Dorn (Reference Autor and Dorn2013) which was applied in several more recent studies (Acemoglu et al Reference Acemoglu, Koster and Ozgen2023; Koster and Ozgen Reference Koster and Ozgen2021; Lewandowski et al Reference Lewandowski, Keister, Hardy and Górka2020a). Despite criticisms directed at Autor and Dorn’s RTI measure (Haslberger Reference Haslberger2022), identifying an alternative in the context of developing countries poses a considerable challenge. Recent methods, such as Lewandowski et al’s (Reference Lewandowski, Park and Schotte2020b) parametric approach, also have inherent limitations. Their regression-based approach, utilising a restricted set of explanatory variables, raises concerns about the optimality of the model’s specification and the risk of omitted variable bias, which could potentially impact the accuracy of the proposed measure. Furthermore, this method would be challenging to apply to detailed occupational classifications beyond 2-digits. Similarly, regarding the approaches of Haslberger (Reference Haslberger2022) and Lewandowski et al (Reference Lewandowski, Park, Hardy, Du and Wu2022), they remain unfeasible for Morocco, given the lack of national surveys incorporating workers’ individual task assessments. Ultimately, the constrained application of Autor and Dorn’s (2013) method still enables a meticulous categorisation of occupations based on two primary criteria: routine versus non-routine and manual versus abstract tasks. In line with Acemoglu and Autor (Reference Acemoglu and Autor2011), we initially matched sixteen task descriptors (t _o ) from O*NET to the microdata of the Moroccan labour survey (Table 1). For consistency, we subsequently homogenised the occupational nomenclature between 2004 and 2018 labour surveys by creating a correspondence table between NAP2014 (2018 survey) and NAP2001 (2004 survey) classifications.

Table 1. Composition of aggregated indices of task importance ^a

Source: Acemoglu and Autor (Reference Acemoglu and Autor2011).

^a Reverse score = 6 – original score, knowing that the original score varies between 1 and 5.

Next and due to the absence of official crosswalks between the Moroccan NAP2001 and the detailed 4-digit O*NET-SOC nomenclature, we elaborated an adapted mapping procedure that involves an intermediate phase using the ISCO-88 nomenclature from the International Labour Organisation. This process resulted in a harmonised occupational structure, encompassing 479 occupations.Footnote ^{6 ,} Footnote ⁷

Having assigned the task items to the labour force survey data, we followed Acemoglu and Autor (Reference Acemoglu and Autor2011) and Lewandowski et al (Reference Lewandowski, Park, Hardy, Du and Wu2022), standardising the worker-level values of each task descriptor (t _i ) using weights from the aforementioned surveys to ensure data comparability:

(1) $${{\rm{V}}_{{i}}}{{\rm{V}}_{{{j}} \in {{J}}}}{{\rm{t}}_{{{i,j}}}} = {{{{{t}}_{{i}}} - {\mu _{{j}}}} \over {{\sigma _{{j}}}}}$$

where J is the set of sixteen task descriptors (j) listed in Table 1, i is the i ^th observation in the labour force survey, and ${\mu _j}$ and ${\sigma _j}$ are, respectively, the weighted average and standard deviation of j ^th task descriptor.

Following Acemoglu and Autor’s (Reference Acemoglu and Autor2011) methodology, we aggregated these descriptors into five task indices (T _k ): ‘non-routine cognitive analytical’ (NRCA), ‘non-routine cognitive interpersonal’ (NRCI), ‘routine cognitive tasks’ (RC), ‘RM’, and ‘non-routine manual’ (NRM), as illustrated in Table 1. We calculated the five aggregated tasks by adding up the appropriate task descriptors as depicted in the following formula. Subsequently, we normalised the five task indices to the range [0,1] to ensure comparability.

(2) $$\boldsymbol{{T_{i,k}} = \;\mathop \sum \nolimits_{i = {{\bf 0}}}^n t_{i,j}^k}$$

where ${T_{i,k}}$ is the value of aggregated task indice k for the observation i and $k\; \in \;\left\{ {NRCA,\;NRCI,\;RC,\;RM,\;NRM} \right\}$

The RTI index results from combining the five task indices, aligning with methodologies outlined by Acemoglu et al (Reference Acemoglu, Koster and Ozgen2023), Autor and Dorn (Reference Autor and Dorn2013), and Lewandowski et al (2020). In recognition of the distinct employment composition in developing countries, we adapted the original RTI formula by incorporating the aggregated index of NRM tasks. This adjustment underscores the pivotal role of this task category in shaping labour dynamics in the context of a developing country. For consistency, we normalised the RTI measure to the range [0,1]. On the basis of the formula (3), occupations with higher RTI values are more intensive in routine tasks and may face increased vulnerability to automation.

(3) $${\rm{RTI}} = \;\ln \left( {{{{\rm{RC}} + {\rm{RM}}} \over 2}} \right) - \ln \left( {{{{\rm{NRCA}} + {\rm{NRCI}} + {\rm{NRM}}} \over 3}} \right)$$

The findings indicate that over the period from 2004 to 2018 (Table 2), the average routine task intensity index remained relatively stable. The Moroccan labour market exhibited high intensities in both RM and NRM tasks, as well as RC tasks. In contrast, NRCA tasks recorded the lowest scores, consistent with expectations for a lower-middle-income economy. This feature aligns with existing literature indicating weaker intensity in terms of non-routine cognitive tasks in developing countries compared to advanced economies (Lewandowski et al Reference Lewandowski, Park and Schotte2020b; Lo Bello et al Reference Lo Bello, Sanchez Puerta and Winkler2019). This pattern can be attributed to several factors such as the prevalence of low-skilled workers in the labour supply and an industrial structure dominated by small and micro-firms with limited technological absorptive capacity, resulting in lower demand for non-routine analytical tasks and a higher emphasis on manual tasks.

Table 2. Routine task intensity and aggregated task indices for Morocco ^a

^a All the indicators in the table are calculated on the basis O*Net 23.1 database and labour force surveys’ weights that correspond to the year indicated. For purpose of readability and comparability, RTI, Offsh and aggregated task indicators are 0–1 normalised.

Sectoral employment patterns (Table 3) align with expectations.Footnote ⁸ The manufacturing sector, including roles such as assemblers and machine operators, shows the highest concentration of routine tasks, rendering its workforce highly susceptible to automation. Public administration services also show significant routine task intensity as do mining and construction sectors. Conversely, agriculture, fisheries and services, rank lower in the occupation hierarchy based on RTI level, indicating comparatively lower routine task intensity in these sectors.Footnote ⁹ The differences observed across industries in terms of their intensity in routine tasks affirm the importance of examining the impact of the structural change process in Morocco, specifically the phenomenon of premature deindustrialisation, on the evolution of occupational employment composition in terms of routinisation.

Table 3. Routine task intensity by main industry in Morocco (Employment structure 2018) ^a

^a Indices are weighted based on the weights from the 2018 Moroccan labour survey. For purpose of readability RTI are 0–1 normalised.

Differences in routine task intensity are also observable between formal and informal employment. Figure 9 illustrates that formal workers are, on average, involved in occupations with a relatively higher RTI than informal workers, suggesting that, at an aggregate level, formal workers may be more exposed to automation than their informal counterparts. These results are in line with our expectations, as informal workers in Moroccan labour market are commonly concentrated in occupations with moderate routine task intensity (low-skilled services and sales, etc.). This concentration tends to lower the average RTI of informal workers relative to their formal counterparts.

Figure 9. RTI by job’s formality status in Morocco*. *Average RTI for formal and informal jobs, weighted according to 2018 labour survey structure.

Measuring offshorability index

Considering the potential role of offshorability in explaining the lack of de-routinisation in the employment composition of specific developing economies within the literature, we have adopted Firpo et al’s (Reference Firpo, Fortin and Lemieux2011) methodology that allows measuring the predisposition of a given occupation to be offshored. This approach, building on Jensen and Kletzer (Reference Jensen and Kletzer2010) pioneering work, addresses criticisms raised by Blinder (Reference Blinder2009). The framework outlined by Firpo et al (Reference Firpo, Fortin and Lemieux2011) utilises 26 work descriptors from the O*NET database, which are classified into five categories of work characteristics, namely ‘Information content’ (IC), ‘Automation/Routinisation’ (AR), ‘Face-to-Face’ (FtF), ‘On-Site Job’ (OSJ), and ‘Decision making’ (DM). However, the approach we have adopted incorporates several modifications compared to the original version. Specifically, the category of descriptors labelled ‘Automation/Routinisation’ has been excluded from the calculation of the offshorability indicator (Table 4). This decision is justified by our concern to avoid any suspected overlap with the RTI indicator as these two indicators are interrelated (Blinder Reference Blinder2009, Blinder and Krueger Reference Blinder and Krueger2013, Goos et al Reference Goos, Manning and Salomons2014) and thus mitigate the risk of multicollinearity in subsequent regression analyses. Additionally, within the scope of this study, we have opted to combine the four selected descriptor categories into a single composite index (Offsh) that encompasses various dimensions associated with offshorability. The formula utilised for this purpose is as follows:

(4) $$\rm Offsh = IC - FtF - OSJ - DM $$

Table 4. Descriptors used in measuring the composite index of occupational offshorability

Source: Elaborated on the basis of the approach of Firpo et al (Reference Firpo, Fortin and Lemieux2011). The category of tasks labelled ‘Automation/routinisation’ by Firpo et al (Reference Firpo, Fortin and Lemieux2011) is omitted from our calculation of the offshorability index. The positive sign (+) indicates that the corresponding category is positively related to offshorability, while the negative one (−) indicates that the corresponding group is negatively related.

The score for each of the four categories corresponds to the average of the values of the descriptors composing it. The four scores and Offsh index have been normalised within the range of [0, 1].

Our results (Table 5) indicate that, despite methodological adjustments, the offshorability index exhibits a significant positive correlation with the RTI index, consistent with Goos et al (Reference Goos, Manning and Salomons2014), while supporting the hypothesis of Blinder and Krueger (Reference Blinder and Krueger2013), which states that offshorability is conceptually distinct from, though related to, an occupation’s routineness. Additionally, the Offsh index shows a negative correlation with non-routine cognitive task intensity (NRCA and NRCI) and a positive correlation with routine task intensity (RM and RC), further supporting the assumption that jobs that can be broken down into simple routine tasks are easier to offshore than those requiring complex thinking, judgement, and human interaction. This, however, does not exclude the fact that numerous non-routine cognitive tasks are increasingly offshorable due to advancements in telecommunications and information technology (Blinder and Krueger Reference Blinder and Krueger2013).

Table 5. Offshorability index in Morocco (Employment structure 2018)^a

^a Indices are weighted based on the weights from the corresponding Moroccan labour survey.

*** Significant at 0.01.

Is the employment de-routinisation process confirmed in the context of a lower-middle-income country like Morocco?

To verify the existence of a process of de-routinisation of employment structure, we adopted a two-step approach. The initial phase, predominantly descriptive, aims to highlight the presence or absence of a process of employment de-routinisation through a tertile analysis. The rationale behind selecting this initial step lies in its simplicity, offering a straightforward way to categorise jobs into three major groups based on routine-task intensity. It also provides an initial insight into the evolving profile of occupational employment structure over the studied period. The second phase is based on regression estimations with the objective of confirming the results of the first step and checking their robustness through various validation tests.

First, we grouped harmonised occupations into tertiles of RTI in 2004 to get three categories: low (Ter1), medium (Ter2), and high (Ter3) routine-intensive occupations. This categorisation enables an assessment of whether employment shares in the most routine-intensive occupations (Ter3) declined in favour of the least routine-intensive ones (Ter1) over the 2004–2018 period.

In the second phase, the approach employed entails regression analysis, adapted from prior research that links the RTI indicator to variables measuring employment volume, either in levels or in growth rates (Goos et al Reference Goos, Manning and Salomons2014; Sebastian Reference Sebastian2018; Cirillo et al Reference Cirillo, Rinaldini, Staccioli and Virgillito2021; Consoli et al Reference Consoli, Marin, Rentocchini and Vona2023; Brambilla et al Reference Brambilla, César, Falcone, Gasparini and Lombardo2023). We regressed average annual growth in occupational employment from 2004 to 2018 (g) on the corresponding RTI level in 2004. To flexibly control for industry-level factors impacting employment dynamics, we organised data by grouping workers into cells based on harmonised 4-digit occupation and 1-digit industry combinations. Sociodemographic control variables were included to minimise omitted variable bias. The Offshorability index (Offsh) served as a time-invariant determinant of occupational employment change, capturing the influence of global value chains and foreign direct investments (FDI) on (de-)routinisation dynamics.

Several specifications are then adopted: the first one includes the RTI as independent variable, while the second specification introduces the squared RTI to take into account a suspected non-monotonic relationship between RTI and change in employment. In the third specification, we introduced several control variables ${X_{oj,2004}}$ including the offshorability index. To check the robustness of our estimates, we performed the same estimations while excluding the agricultural sector.

As for the last specification, we have replaced the RTI with the five aggregated indexes of task importance $\;{T_{o,2004}}$ at the starting period, to detect potential differentiated effects by main task category. The objective of this latest specification is to discern the divergent impacts on employment dynamics between tasks requiring high qualifications, specifically non-routine cognitive analytical and interpersonal tasks (NRCA, NRCI), and those necessitating modest to intermediate qualifications, encompassing routine cognitive and manual tasks alongside non-routine manual tasks (RC, RM, NRM).

(5) $${g_{o,j,2004 - 2018}} = \alpha + \beta {\rm{\;RT}}{{\rm{I}}_{o,2004}} + {\varepsilon _{o,j}}$$

(6) $${g_{{\rm{o}},{\rm{j}},2004 - 2018}} = \alpha + {\rm{\beta }}\,{\rm{RT}}{{\rm{I}}_{o,2004}} + \partial \,{\rm{RT}}{{\rm{I}}^2}_{o,2004} + {\varepsilon _{o,j}}$$

(7) $${g_{o,j,2004 - 2018}} = \alpha + {\rm{\beta \;RT}}{{\rm{I}}_{{\rm{o}},2004}} + \partial {\rm{\;RT}}{{\rm{I}}^2}_{o,2004} + {\rm{\;}}\mu {\rm{\;}}{X_{o,j,2004}} + {\varepsilon _{o,j}}$$

(8) $${g_{o,j,2004 - 2018}} = \alpha + \beta {\rm{\;}}{T_{o,2004}} + {\rm{\;}}\mu {\rm{\;}}{X_{o,j,2004}} + {{\rm{\varepsilon }}_{o,j}}$$

Where

${g_{o,j,2004 - 2018}}\!$ = average annual growth of employment in the occupation o and industry j from 2004 to 2018

RTI = Routine Task Intensity Index

X = Vector of control variables by occupation–industry cell (Offshorability index, Share of workers in the corresponding occupation-industry cell working in firms with less than 5 employees (micro firms), Share of workers in the corresponding cell with middle skills, share of workers in the corresponding cell with low skills.

‘o’ refers to the occupation (harmonised 4-digits)

‘j’ refers to sector or industry (1-digit)

T = (NRCA, NRCI, RC, RM, NRM)

To confirm the hypothesis of employment structure de-routinisation, we need a negatively significant coefficient for the RTI variable. Concerning offshorability, we anticipate a positive impact of the variable (Offsh) on employment changes. The growing integration of developing countries into global value chains and the inflow of routine jobs through offshoring are expected to contribute to an increase in routine task-intensive jobs (Das and Hilgenstock Reference Das and Hilgenstock2022; Martins-Neto et al Reference Martins-Neto, Cirera and Coad2023).

How does the process of structural transformation and employment (de-) routinisation interact within a premature deindustrialisation context in a middle-income country?

Regarding this second question, we propose a shift-share decomposition of employment share changes for each of the three RTI tertile-based occupational groups (Ter1, Ter2, and Ter3). This method, acknowledged for its ability to distinguish between within effect and between effect (Bárány and Siegel Reference Bárány and Siegel2019; Intraligi et al Reference Intraligi, Vittori and Ricci2021), allows for a nuanced analysis of contributions from the structural transformation process, including the impact of premature deindustrialisation, and those arising from intra-sectoral changes.

Formally, the change in the share of each occupational group in total employment comprises two components: the between-sector component, representing the effect of structural transformation, and the within-sector component. The former captures variations in the occupational group’s share due to shifts in the sectoral composition, while the latter reflects changes in the distribution of different occupations within each major industry.

(9) $${\Delta _{2004 - 2018}}\left( {{{{E_{{\rm{Ter\;}}i}}} \over E}} \right) = \mathop \sum \limits_{j = 1}^N {{{E_{{\rm{Ter\;}}i,j}}} \over {{E_j}}}\Delta \left( {{{{E_j}} \over E}} \right) + \mathop \sum \limits_{j = 1}^N {{{E_j}} \over E}\Delta \left( {{{{E_{{\rm{Ter\;}}i,{\rm{\;}}j}}} \over {{E_j}}}} \right)$$

With ${E_{{\rm{Ter\;}}i,{\rm{\;}}j}}$ being the employment in the occupational category Ter i and industry j

and E = total employment

In the initial phase of traditional structural transformation process, as the economy is shifting from agriculture to manufacturing, we anticipate a positive contribution of the between-sector effect to the change in the share of the most routine-intensive occupational group in total employment (Ter3). As emphasised by Das and Hilgenstock (Reference Das and Hilgenstock2022), structural transformation in developing nations tends to mitigate or counteract the employment de-routinisation effect associated with the adoption of labour-saving technologies.

However, as noted earlier, premature deindustrialisation in some developing countries like Morocco challenges this scenario. In nations where there has been an early move of labour from manufacturing to services with low routine task intensity, the expected process of routinisation of employment structure may be hindered. To test this assumption, we decomposed the between effect (structural change) to highlight the contributions of each major industry (Agriculture and fisheries; Mining; Manufacturing; Utilities; Construction; Low-skill services; Medium-to-high skill services; Public administration). A negative contribution from manufacturing would confirm the role of premature deindustrialisation in reducing the share of high-routine task intensity jobs.

In parallel with investigations related to the role of premature deindustrialisation, in this section, we will also conduct specific-sector tertile analysis to examine the internal processes of (de-)routinisation within each major industry to highlight potential heterogeneity among them. Confirming such heterogeneity could also contribute to influencing the overall process of employment (de-)routinisation.

To what extent does the prevalence of informal labour in various industries influence the process of employment (de-)routinisation?

To explore this issue, we employ a three-way interaction model to uncover the nuanced variations in the response of changes in occupational employment composition to routine task intensity across industries and according to formal/informal worker status. The inclusion of informality status represents a novel contribution to the existing literature. As public administration comprises almost solely formal employment, this sector was excluded from the model.

The dataset utilised is structured across three-dimensional cells: occupation, industry, and worker informality status. The regression model includes a three-way interaction term between three explanatory variables: occupational RTI index, industry dummy (ref = agriculture and fisheries), and job informality status dummy (informal = 1), along with control variables. The dependent variable represents the change in employment within a specific cell. This approach allows for flexible estimation of distinct slopes for RTI based on both industry and workers’ informality status.