Multiple Linear Regression (MLR)
Pros
- The model is easy to implement as it is accessible with standard tools like Excel, R, or Python.
- It is a widely accepted and understood statistical technique.
Cons
- The resulting equal pay gap does mathematically not align with the reporting definition of the gender pay gap, which makes it more difficult to communicate the results.
– Spoiler alert, we will give you more details on this in the next blog post. –
- Restrictions apply due to the assumptions of regression. Besides causing the issue mentioned above, this also limits the set up as it cannot handle variables with nonlinear relationships with pay or that are correlated with other variables in the model.
- Transformations are done to the data to comply with the assumptions of regression. These transformations also change the way the model should be interpreted. When working with the output of the model: the (predicted) values of the dependent variable pay, this is often not much of a problem. However, when working with the coefficients – which is the case with this particular method – much more care needs to be taken when interpreting the results.
Oaxaca-Blinder Decomposition (OBD)
Pros
- Due to its different set up, the equal pay gap and the reported gender pay gap can be aligned mathematically.
Cons
- Two separate models are required; one for each gender group. This can be problematic as statistical models need “enough” data points (read, employees) for reliable results. However, within HR datasets, the number of data points – the number of employees – is often already too small for one model, let alone for two models, cause then you roughly need twice as many employees. Therefore, this model is less suitable for smaller companies or companies that would like to perform an analysis on smaller subsets of the workforce (e.g. Business Units, entities, or Countries).
- Similar to MLR, restrictions apply due to the assumptions of regression.
Pay Prediction approach
Pros
- This method allows you to slice results across the entire workforce with ease. For each individual employee, the predicted salary is compared to the actual salary, and this information is used to calculate an equal pay gap. This means that once the predictions are made, you do not have to rerun this prediction model per part of the workforce you like to explore; you can simply just filter. This is different for the other methods as they require a separate analysis of each part of the workforce for which you would like to calculate the equal pay gap.
- It exceeds single group comparisons; as gender is not part of the prediction model, the results of the model can also easily be used to compare other groups, like Western versus non-Western employees. Thus, you do not need to invest time and money in multiple separate analyses.
Cons
- There are ethical implications when using predicted salaries as it is implied that the model produces a fair result while that does not have to be the case. Most likely there will be some bias in the data. This bias will also be included in the predicted salaries.
- Like MLR, the resulting equal pay gap does mathematically not align with the reporting definition of the gender pay gap, which makes it more difficult to communicate the results. Next blog this will be explained in more detail.
- Again, restrictions apply due to the assumptions of regression.
As with any analytical tool, understanding the trade-offs is just as important as running the numbers. Hence, in this blog we have explored the pros and cons of the regression-based Equal Pay methods.
Later we will explore the pros and cons of our preferred method: the weighing-based Rake method. But first, in our next blog post, we will dive deeper into the issue of the misalignment between the (uncorrected) gender pay gap and the (corrected) equal pay gap that we have seen among the cons of MLR and the Pay Prediction approach.
