The data used in this research were derived from a multi-country volunteer web survey: the WageIndicator dataset (Tijdens et al., 2010). This survey is posted continuously on national WageIndicator websites (www.wageindicator.org). It is operational in 85 countries and attracts millions of visitors (25 million in 2014). The websites provide job-related content, labour law and minimum wage information, VIP wages, and a free ‘salary check’ presenting average occupational wages based on the web survey data. In return for free information, web visitors are asked to complete a questionnaire with a lottery prize incentive. Approximately 70,000 questionnaires are completed each year.Footnote 9
The web survey contains detailed questions about demographic variables together with work-related ones such as occupation, industry, wage and firm size. Occupation is asked for not through an open text field with post-survey coding, but by means of self-identification using a database of approximately 1,700 occupational titles. Respondents can search the database by text string matching or with the help of a 3-step search tree. Respondents who are unable to identify their occupation are instructed to choose an occupational title that comes close to theirs. The 1,700 titles are classified according to the International Standard Classification of Occupations (ISCO–08), which is maintained by the International Labour Organisation (ILO) and is increasingly considered the global standard. The 1,700 titles are clustered into 433 occupational units at the ISCO’s 4-digit level.
When respondents self-identify their occupation title, they are presented with a list of between 4 and 14 occupation-specific tasks related to their 4-digit occupational unit (designed by the ILO when defining the present ISCO–08; see Hunter 2009). They can then state how often they perform each of the tasks on the basis of five answer categories: never, yearly, monthly, weekly and daily. To provide an example of the measurement of task intensity, Fig. 1 shows a screenshot for the occupation ‘Business services and administration managers’, which distinguishes between 10 occupational tasks. As the tasks are occupation specific, they vary across occupations. A full list of occupations and corresponding tasks is available on the ILO website.Footnote 10
Given our focus on occupations and occupation-specific tasks, an online survey seemed to be the best way to collect data on task implementation. In all other modes of data collection, the handling of a list of tasks for 433 occupations would have been too complicated and too time-consuming. Moreover, a tight fieldwork budget and other priorities very often hinder the inclusion of rather difficult and complex items in a questionnaire. Against this background, it is not surprising that task data are rarely available.
To our knowledge, only a few other datasets provide information about the task implementation of specific jobs; the best known are the North American O*NET and the German Qualification and Career Survey. The O*NET dataset provides, among other information, a measure of the relevance of (thousands of) tasks distributed across (hundreds of) occupations. However, this assessment is based on mixed data collection methods involving job incumbents and occupation experts, while the online survey approach, such as that of the WageIndicator is based on a bottom-up approach. The latter produces an intensity measure for each task listed within each occupation and captures individual variation within an occupation. This advantage is exploited in the present article by relating task intensities variation with wage variation. We do not exclude that similar exercises can be performed on the O*NET task rating data; however, the possibility of triangulation/cross-checking between the WageIndicator task data and the O*NET task ratings seems, at best, unlikely since they make use of different classifications for both occupations and tasks within occupations. The Qualification and Career Survey carried out by the German Federal Institute for Vocational Training and the Research Institute of the Federal Employment Service, provides information directly provided by employees over time. Nevertheless, tasks are not occupation specific but generic, and no information about the time spent on each activity is included.
Given the abovementioned characteristics, data collection through a web survey is the most suitable method when aiming to collect job-specific task information in the case of a large number of occupational titles (Tijdens et al., 2012).Footnote 11 The advantages are obvious: respondents can read the tasks themselves, thus reducing their time burden, and web surveys are often cheaper, as researchers do not have to consider fieldwork budgets. The authors regard the respondents’ reported task frequencies as trustworthy because the tasks refer to activities that are considered to be the core of their occupation. Hence, respondents are familiar with a task because they perform it regularly, or they are not familiar with it because they never or hardly ever perform the task. It is our belief that respondents do not generalise on frequency of tasks from memory. In the case of the present analysis, the main advantage of the dataset was clearly the availability of comparable task intensities data at the 4-digit occupation level. The combination of such a specificity, the individuals’ information provided and the large sample size allowed us to identify similar workers and, at the same time, to observe heterogeneities in salaries and task implementation. Moreover, it contributed to the improvement of empirical wage models by including what had been a rather unobserved variable that contributes to a better understanding of sources of wage heterogeneity.
Given the structural differences in national labour markets, we confined our analysis to the Dutch labour market, which is generally considered to be well-performing, as evidenced by a prolonged period of low unemployment, high participation and stable wages (Gerritsen and Høj, 2013). These conditions reduced the possibility of any observed effect being due to some nation-specific shock in the market. In order to avoid a possible impact on the wage structure of changes in job tasks over time (see e.g. Acemoglu and Autor, 2011), we used pooled data for a limited time span (2013–14). Moreover, to guarantee the representativeness of the occupations considered, we selected the 100 most common 4-digit occupations, leading to a total number of 745 tasks for 5,230 observations (see Table 1 for a list and statistical descriptives of the 100 occupations). Finally, as self-reported wage data, which we used to estimate salary heterogeneity, are often noisy,Footnote 12 we relied on the following two-step filtering to improve the data: first, we excluded implausible data (values below €0.5 or above €2,000 per hour), and second, we concentrated on the 5th through 95th percentiles of the resulting distribution. Despite these restrictions, our range of values is in line with other works using survey wage data (e.g. Acemoglu and Autor, 2011, working on Current Population Survey data).
As the survey is voluntary, however, the question of representativeness needs to be addressed. Volunteer web surveys are an extreme case of a convenience sample which suffer from a combination of coverage, self-selection and non-response biases. Moreover, one person might be able to fill in the questionnaire several times. Previous research has shown that generally people who self-select into a (web) survey differ from those who do not with respect to socio-demographics but also in terms of time availability, web skills, or altruism to contribute to the project (Bethlehem 2010; Couper et al. 2007; Malhotra & Krosnick 2007). In case of the WageIndicator survey, which is used in the present study, previous findings show that high educated, younger and male respondents as well as people working predominantly in non-manual occupations are overrepresented in various country samples such as the Netherlands, Germany, Spain and Brazil (Bispinck et al. 2010; De Pedraza et al. 2010; Steinmetz et al. 2012). Moreover, it seems that the survey topic is an important criterion for the self-selection process. This leads to an increase of people with a particular interest in the topic, and those people might differ from those who have no strong interest in the topic. For instance, the study by Steinmetz et al. (2014) showed that respondents of the Dutch WageIndicator are more interested in wages and career advances. The authors, however, could also show that even if the web sample deviates to some extent from representative reference sample, the obtained conclusions are very often comparable. Against the background that literacy skills and internet penetration in the 100 most common Dutch occupations can be assumed to be at a medium or high educational level, the above described problems are less severe. The results for low-educated workers should be considered cautiously and be taken into account in approaches that are intended to go further than this exploratory study.
As our aim was to find out whether and, if so, to what extent a task approach could explain part of the unrevealed wage variation (a) across a priori similar individuals and (b) across occupations. Our empirical strategy is as follows. In the first step, we adjust the wage heterogeneity due to individuals’ productive attributes, discrimination and compensating differentials. We therefore regress a common set of wage predictors on the log hourly wages and considered the exponential of the residuals for the rest of our analysis. The common wage predictors used here are a cubic of work experience,Footnote 13 years of education, and binary and categorical dummies for gender, migrant status, firm size and industry (according to the 1-digit NACE classification).Footnote 14 The exponential of the residuals of this first regression are individual wage measures that are orthogonal to the individual and firm productive characteristics. The dispersion of these residual wages is therefore independent with respect to the main individual and firm observable characteristics.
In the second step, we estimate within-occupational wage and task-implementation heterogeneities and test whether these measures are significantly related. We therefore correlate three task-heterogeneity indexes of each of the 100 selected occupations with two corresponding within-occupation wage dispersion indexes (see Section 3.3 for the full description of the measures implemented). Observing the relation between these measures, and in particular the share of explained wage heterogeneity variation, provides insights into whether a task approach can explain part of the unrevealed wage variation across a priori similar individuals. Using wage measures orthogonal to the individual and firm productive characteristics and considering workers of the same ISCO-08 4-digit disclosure allows us to address similar workers.
In the final step, we shift the analysis to a cross-occupation level and examine whether the observed heterogeneity in task implementation of each occupational group is related to its salary level. This is done by examining the relation between the three task heterogeneity indexes and the corresponding median wages received (i) by all workers in each occupation and (ii) by subgroups of workers differentiated by skill levels. This helps us better understand whether a task approach is capable of explaining some of the differences across occupations that determine wage diversity. In particular, it shows whether the role of standardisation in jobs is related to the salary structure.
The main measures we use in the framework presented in this paper are related to wage heterogeneity and task implementation heterogeneity. With respect to wage heterogeneity, we define two measures: the Gini index and the frictional wage dispersion (FWD) index. The Gini, which is the most widely accepted single-measure of wage inequality in economic literature, measures the inequality of a distribution. It ranges from 0 (= a continuous uniform distribution where all individuals receive the same income) to 1 (= the maximum inequality where a single individual obtains all available income and the rest of the population’s income is 0).Footnote 15 The FWD is based on the difference between the average wage and the lowest (the reservation) wage paid in each group. This measure is similar to the mean–min ratio of Hornstein et al. (2011), which measures wage differentials induced by labour market search frictions.Footnote 16 Although slightly different, our measure shares the same elements, and we therefore refer to it as the FWD. The choice of the two measures is justified by the fact that they are differently respondent to the salary levels, and we know that this aspect could be of interest in our analysis.
We also include measures that would indicate the extent to which tasks are performed differently in each occupation. To proxy this within-occupational task heterogeneity, we rely on the relevance of the typical worker in each occupation, where the ‘typical worker’ is defined as a worker who employs the most common way of performing an occupation. In order to measure this dispersion consistently, we make use of three indexes: the variation ratio, the entropy and the Gibbs index, all computed on the self-assessed task intensities as explained below. The choice of the three measures is justified by the need to avoid possible index-specific bias. All three are measures of qualitative variation suited to assess dispersions in nominal distributions, which is the case of the self-assessed task intensities where respondents are asked how often they perform specific tasks and are given five frequency category options (see the previous section for details). Moreover, they are based on the weight of the most common class, which makes them suited to assess the idea of typical workers in an occupation. More concretely, the variation ratio is one of the most widely used measures of statistical dispersion in nominal distributions. It is defined as the proportion of cases that are not the mode (Wilcox, 1973). Entropy is a measure of the unpredictability of information content (based on the seminal work of Shannon, 1948) and is an established measure of dispersion for ordinal and nominal data (Vanhoy, 2008). It is a measure of uncertainty; in discrete distribution this is minimised when all cases belong to a single category and maximised in a uniform distribution (for further details, see Harris, 1982). Entropy makes use of the information provided by the probability distribution of the discrete variable and is elaborated as the negative of the sum of the category probabilities times the (base two) logarithms of the category probabilities. The Gibbs index was developed by Gibbs and Poston (1975)Footnote 17 and can be interpreted as one minus the likelihood that a random pair of samples will belong to the same category. It is therefore higher when the distribution across categories is uniform.
With regard to the interpretation of the indexes, for each task, all three measures produce low values when the proportion in the largest class of response is high; in other words, when the proportion of respondents performing a specific task with the same intensity is high. We first compute a task-specific index for each of the 745 tasks in our dataset. We then average the task heterogeneities corresponding to each occupation and obtain an occupation-specific value of tasks heterogeneity. By doing so, we create an occupation-specific index measuring the proportion of workers who perform their occupation in the same manner; that is, the index reflects how much workers are like the typical worker in each occupation. Low values of the indexes correspond to occupations with a large proportion of workers acting like the typical worker, whereas high values identify occupations where the weight of the typical worker is low, and therefore the task implementation is more heterogeneous.