Annex B. Econometric analysis
The European Company Survey (ECS) has substantial information on training and learning decisions made by firms and their background characteristics, such as sector, size, age, hierarchical structure, product market strategy, technology adoption and work practices. This rich dataset allows a detailed analysis of the relationship between firms’ characteristics and their training offer. To this end, the OECD has conducted regression analysis for different training and learning outcomes. Boxes throughout the chapters present summaries of the main findings. This annex provides a technical description of the methodology and the results.
The sample used for the econometric analyses includes enterprises from all countries covered in the ECS microdata (the EU-27 plus the UK), which have more than 50 employees and belong to the manufacturing and services sectors. The ECS has several categorical variables, which require managers to select one option among several categories, for instance:
In 2018, how many employees in this establishment have received on-the-job training or other forms of direct instruction in the workplace from more experienced colleagues?
A binary outcome variable was derived by merging multiple categories, using the sample average as the cut-off point. For example, for the question above, a dummy variable was created by assigning a value 1 if the company reported that more than 40% of employees had received on-the-job or direct instruction in the workplace. For each binary outcome variable, the OECD estimated a probit model. The list of independent variables includes dummy variables for country, number of employees, age, hierarchy levels, type of market strategy, adoption of several High Performance Workplace Practices (HPWPs), change in employment, profitability, share of permanent contracts and technological change. Standard errors are clustered by country and the industry-size strata for the sample. The OECD also explored the possibility of implementing a multinomial logit approach, but this was ultimately discarded, due to sample size considerations and difficulties in interpreting coefficients.
After estimating the probit model, average marginal effects were computed. Given that all variables are categorical, each marginal effect shows the change in probability of experiencing a particular outcome (e.g. having a comparatively high share of employees participating in training) for a category of enterprises (e.g. financial services sector enterprises) compared to the omitted category (e.g. manufacturing firms). The results should not be interpreted causally, but show how a particular firm characteristic is correlated with a training or learning outcome, while controlling for other factors.
The sections below present the results discussed across different chapters.