In the analysis, we use individual-level data from the PIAAC database. The data is collected using personal interviews. Basically, 200,588 persons in 31 countries are included in the data set. Four of the 31 countries (Austria, Canada, Germany, the US) are excluded from the sample since all values are missing for at least one of the explanatory variables. This reduces the sample to 158,300 persons. All the remaining 27 countries are included in the sample.

We only include employees and not self-employed in the sample. The reasons are that dummy variables for firm size are included as control variables in the regressions, and firm size only refers to employees. Persons with missing values for at least one of the explanatory variables are excluded from the sample, with two exceptions: persons with unknown educational level or firm size are included in the sample. The sample is also limited to the age group 20–59 years. However, most countries have a formal retirement age that is higher than 59 years.

The final sample consists of 75,777 employees living in 27 European and non-European countries. The sample size (N) by country is shown in Tables 1 and 2. The data for the Russian Federation only covers the Moscow municipal area.

The dependent variable measures IPP. IPP is measured by the following five individual characteristics: (1) how often a person learns new work-related things from co-workers or supervisors in his or her own job (abbreviated formulation ‘learning new work-related things’), (2) how often a person’s job involves learning-by-doing from the tasks he or she performs (‘learning-by-doing from tasks’), (3) how often a person’s job involves keeping up to date with new products (‘keeping up to date with new products’), (4) how often a person’s job usually involves instructing, training or teaching people, individually or in groups (‘instructing, training or teaching people’), and (5) how often a person’s job usually involves advising people (‘advising people’).

For each of the characteristics (1)–(5), the response categories are: 1 Never, 2 Less than once a month, 3 Less than once a week but at least once a month, 4 At least once a week but not every day, and 5 Every day. Since the Cronbach’s alpha for the five characteristics is within the acceptable range for the internal consistency of these characteristics (i.e. alpha > 0.7) for all countries as a whole, we include them in the same scale as an overall measure of IPP and treat this measure as one-dimensional.

We have also carried out a (principal-component) factor analysis for all countries as a whole. After orthogonal rotation (varimax), this factor analysis gives us a two-variable factor-solution, but prior to rotation the factor-solution shows that each of the five items in the measure of IPP has a higher factor loading on the first factor variable than on the second factor variable, which in addition to the alpha-score supports our choice of including all five items into one single variable. The dependent variable is therefore set equal to the sum of the characteristics (1)–(5), and thus varies from 5 to 25.

Endogeneity is a common problem in regression analyses, which refers here to the possibility that the values of at least one of the explanatory variables is affected by the dependent variable. Instrumental variables (IV) techniques are commonly used to address this problem. According to Chang and Kang (2018), the instruments must be both relevant and exogenous. They suggest strategies to test both the relevance and exogeneity requirements. It is likely that at least one of several potential instruments will be related to an explanatory variable that we suspect to be endogenous. Therefore, we will probably be able to find an instrument that is associated with the variable that may be endogenous. However, the problem is to find exogenous instruments; that is, they should not be affected by the employees’ IPP, but correlated with the explanatory variable suspected to be endogenous. Based on the PIAAC data set, we have no suggestions for such instruments. Furthermore, the cross-sectional nature of the PIAAC data prevents us from lagging variables that we suspect to be endogenous, and using the lagged variables as instruments to reduce the potential endogeneity of the variables that are lagged. Because it is difficult to find suitable instruments, we cannot use IV techniques in the analysis. As a substitute for this, we try to reduce the potential bias from endogeneity by not controlling for individual characteristics that we expect will be affected by the measure of IPP (for example, characteristics like ‘confronting more complex problems’, and ‘how often the job usually involves planning own activities’).

We have also selected the explanatory variables to account for the potential multicollinearity problem. This selection is based on calculations of the Variance Inflation Factor (VIF). These calculations indicate that we have no serious multicollinearity problems.

In accordance with the purpose of the article, age is the only key regressor in the analysis. Since we lack information about exact age for some of the countries, we use four age dummy variables in the regressions: 20–29 years, 30–39 years (the reference category), 40–49 years, and 50–59 years.

The following control variables are included in the estimations: gender (1=female, 0=male), educational level (four dummy variables), skills level, seniority, participation in formal and non-formal learning, weekly working hours, firm size (six dummy variables), occupation (11 dummy variables), and industrial sector affiliation (12 dummy variables). Controlling for occupation accounts to some degree for characteristics that are more relevant for specific jobs such as physical strength and innovative learning.

There are three educational levels: lower secondary or less, upper secondary or post-secondary (the reference category), and higher education. We also include a category for those with unknown educational level.

We have chosen to measure skills level as literacy skills, since the correlation between literacy skills and numeracy skills is very high.Footnote 2 Literacy (and numeracy) skills in the PIAAC database consist of a set of ten plausible values estimated for each person.

Seniority measures approximately how many years a person has had paid work in total (only years where 6 months or more was spent in either full-time or part-time work). We control for seniority since age is highly correlated with seniority for a particular occupation,Footnote 3 and since employees in different occupations may begin their active careers at different ages. The age variables are thus net of seniority effects.

We differentiate between participation in formal and non-formal learning. Participants in formal learning are studying for any kind of formal qualification at the time of the interview, or have studied for any formal qualification during the last 12 months. Participants in non-formal learning have participated in at least one of the following activities during the last 12 months: open or distance education, on-the-job training or training by supervisors or co-workers, seminars or workshops, and other courses or private lessons.

Firm size (i.e. number of employees) is a variable with five categories: 1–10 employees, 11–50 employees, 51–250 employees (the reference category), 251–1000 employees, and more than 1000 employees. We also include a category for those with unknown firm size.

Occupation is based on ISCO codes (1-digit), where ‘professionals’ is used as the reference category. Persons with a manual occupation are either employed as ‘clerical support workers’, ‘service and sales workers’, ‘skilled agricultural, forestry and fishery workers’, ‘craft and related trades workers’, ‘plant and machine operators, and assemblers’, or employed in ‘elementary occupations’. Industrial sector is based on ISIC codes (alphabetical level), where ‘trade, transportation and service’ is used as the reference category.

Skills level, seniority and weekly working hours are continuous variables. All other explanatory variables are dummy variables.

All results in Table 2 and Fig. 1 are weighted by using the full sample (final) weight and the 80 replicate weights in the PIAAC database. The weighting procedure ensures representative data. The ‘repest’ command in Stata, which is used when weighting the results, ensures correct estimates for standard errors for the skills level.

Composition of sample of employees, by country and age

This procedure is also used when weighting the results in Table 3, but in this case the weights are corrected to provide that the number of observations for all the 27 countries is the same. We set this number to 5200 in each country (see Støren 2015, Appendix 1). This weighting procedure secures that all the country samples will have the same influence on the results when using all countries as a whole.

In Table 4, the results are weighted by using the ‘mixed’ command in Stata, where we only use the full sample (final) weight (based on sampling weighting). Sampling weights are specified at the first level in the multilevel model. The type of standard error reported is derived from asymptotic theory. A likelihood-ratio test is carried out after the regression (in M6). Since this test is invalid in the case of sampling weighted estimators, the test is based on frequency weighting. In this case, Stata produces exactly the same results (in terms of estimated coefficients and standard errors, including estimates of random-effects parameters), regardless of whether we use sampling or frequency weighting.

All calculations of the VIFs are weighted, where we only use the full sample (final) weight (based on sampling weighting). The results from the factor analysis in “The Dependent Variable” section, and the calculations of the pairwise correlation coefficients in “The Explanatory Variables” section, are also weighted by using only the full sample weight (based on analytic weighting). The Cronbach’s alpha calculation in “The Dependent Variable” section, and the calculations of the pairwise correlation coefficients in “Descriptive Statistics” and “The Effects of National Averages” sections, are done without using any weights.

Several econometric methods are used in the analysis. When examining whether the effects of the explanatory variables on the IPP score differ between the countries, we run regressions for the 27 countries separately. One possibility is to use ordered logistic regression in this case. The problem is that the assumptions of the ordered logit model are often violated (Williams 2016). We therefore prefer to use linear regression, for two further reasons. First, since the dependent variable can be seen as a coarsened version of a continuous variable, and not a truly ordered variable, we can derive a meaningful scale of the dependent variable. Second, the results from linear regression are in most cases easier to interpret than the corresponding results from ordered logistic regression.Footnote 4 Based on linear regression, larger values of the dependent variable correspond to higher IPP scores.

We also run a regression for all countries as a whole using multilevel mixed-effects linear regression. In this case, we examine the effects of the explanatory variables on the IPP score at the individual level and the effects of some of these variables (i.e. the ‘higher education’ variable and the skills level) at the national level.

Two further regressions are performed for all countries as a whole. One is based on linear regression, where we examine the effects of explanatory variables on the country-specific estimated coefficient for the oldest age group. The other is based on logistic regression, where we examine the effects of the same explanatory variables on the probability that this estimated coefficient is significant (at the 5% level).