Understanding & analyzing the compensation model
This is an note/information message.
This article assumes that you have reviewed your data in the Employee Overview page and the system has showcased the unadjusted pay gap corresponding to your data set, and that you have created your compensation model and ran your first pay equity analysis. If not, refer to Reviewing your data and Creating a compensation model, respectively.
The results section within the Pay Equity Analysis page has now been updated with data enabling the measurement of the adjusted pay gap. Both unadjusted and adjusted pay gaps are presented, indicating what demographic group is the lower paid.
You can now explore how your compensation model explains pay differences, and arrive at your adjusted pay gap, in the Compensation Model Result tab. Follow the steps below to arrive at your adjusted pay gap:
Interpreting the explanatory value of your model
Based on the analysis settings you’ve provided, the system has created an estimation of your compensation model and pay structure. The system will now run a regression analysis until it finds the most accurate relationship between each contributing factor (variable) and the actual pay of your employees.
Because it is a model, it will never be 100% accurate, but it can still explain the data in an effective in a proficient or deficient manner. The closer your compensation model is to predicting your actual pay variations, the healthier and more useful it is. Simply by adding large numbers of variables, you can get to a high level of explanation; however you also need to make sure that the variables are meaningfully contributing.
When you open the Compensation Model Results tab of the system, you're presented with insights into the explanatory value of your model on the left side of the Compensation Driver overview, as illustrated in the following figure:
Compensation Model Results Tab
The score displayed at the left top (R-squared) represents the extent to which your overall model explains compensation variability in your dataset. The score is on a scale of 0 to 100%, where the goal is to achieve a high explanatory value (over 85%).
The list of compensation drivers, displayed underneath your score, illustrates your compensation structure (based the variables you selected when running your pay equity model) and the individual explanatory value of each compensation driver within the overall model. The down pay is broken down into its main drivers to see how much each one contributes to the overall picture. For the calculation of the individual driver explanation, the system takes into account how common each factor is among employees and the relative importance of each factor.
For example, if job role and education are the two main factors, and you find that the job role explains more of the pay differences than education, you can say that job role plays a bigger part in the overall pay structure. You can also look at how well you model explains the differences in pay overall (the R-squared), and this helps you fine-tune exactly how much each factor is contributing.
Each compensation driver (or variable) is categorized based on its individual explanation to the overall compensation structure according to the following principles:
Core: a variables that explains over 30%
Central: a variable that explains between 15% to 30%,
Strong: a variable that explains between 5% and 15%,
Moderate: a variable that explains between 1% and 5%,
Marginal: a variable that explain less than 1%.
This is an note/information message.
Your selected demographic variable is also included in your compensation drivers. In an ideal world without pay gaps, your demographic variable would not contribute at all (or only marginally) to explaining pay differences. However, in practice you may find your demographic variable (e.g. gender) does have a moderate or even strong contribution to your pay structure.
Check your compensation health
This is a tip message.
Within a regression analysis there are two key indicators of model health:
Explanatory power R² value (0 to 100%) of your model - a technical way of expressing how much variation in pay your model is able to explain and therefore how healthy your model is. As a general guideline one should aim for a value of 85% or greater.
Significance (0 to 4 dots, 4 is best) that is based on p-value (0 to 1, where 0 is best) of each of your variables - a technical way of expressing how meaningful or significant each variable is at explaining pay variances. It represents the probability of pay variation being due to pure chance versus being explained by the variable. As a general guideline one should aim for a value of 0.001 or lower.
If your model has a high explanatory power (R² >85%) with highly significant variables (majority of the p-values <0.001), you are good to proceed with the review of the impact of variables and on your pay gaps.
If your model health is not yet at that level, you might want to consider replacing the compensation drivers in the category "marginal" with alternative compensation drivers to improve your compensation model, particularly if your compensation model explanatory value (R-squared) falls below the suggested threshold of 85%.
We also recommend heading back to run a new pay equity analysis using the steps described in the article Creating a compensation model. This time round, select different variables; removing any that were not significant, to see how it improves your compensation model.
Understanding the impact of each compensation driver category on compensation
Clicking on each compensation driver category in the left-hand section opens a center panel that displays the impact on compensation of each category within the selected compensation driver (or variable), as illustrated in the following figure:
Country compensation driver
For categorical variables, such as Global grade, Country or Job Family, the bar graph presents the estimated impact on compensation and how it varies across these categories when all the variables have been applied. For instance, if the model estimates the compensation for employees grouped under the HR department to be higher than the compensation for employees in Finance, these differences in estimated impact on compensation are displayed for each department bar, comparing to the department estimated to have the lowest pay.
This is a tip message.
For more information about categorical variables, refer to .
You can interpret the graph located in the top right part of the screen (for example, Country) as follows:
The first category at the top (example: Italy) is the so-called "base value", which means it is on average the lowest-paid category within the compensation driver.
The categories below are presented in order of increasing positive impact on compensation, based on your compensation model (for instance, "working in Denmark" leads to the greatest increasing impact on compensation, compared to the baseline "Italy"). This means that an employee in the second category (such as Portugal) according to your compensation model would, on average, receive a 9% higher compensation compared to the baseline category (Italy in our example), assuming that all other compensation drivers are equal.
-
The table below the graph displays for each compensation driver category:
The number of employees in the group. A general rule of thumb would be to have at least 5 to 7 employees in any given group.
The estimated impact in compensation, which is the same as the bars in the graph above.
The significance of the compensation driver category, on a scale of zero to four dots. The more colored dots, the higher the significance of the category and the greater the value it brings to the compensation model.
This is a tip message.
Click on your demographic variable (for example, Gender) in the list of compensation drivers on the left to view its impact in the right-hand section. You may notice that the impact of the demographic category is the same as your demographic pay gap. This is due to the fact that the regression model uses the impact (coefficient) to measure pay gaps.
This is an note/information message.
For numerical variables , such as the number of direct reports, the system calculates the estimated average percentage change in compensation for an employee who has one additional direct report, and the estimated percentage increase for the employee with the highest number of direct reports.
For more information about numerical variables, see .
The estimated impact on compensation for numerical variables shows the impact of each additional unit (year in the role or additional direct report, for example) on pay. The estimated impact for highest value shows the expected increase in salary when taking an employee from the lowest observed value of that variable to the highest. In other words, it showcases the maximum impact each variable could potentially have on the expected salary.
The following figure illustrates the Compensation drivers page with the estimated impact of the Time In Role variable:
Compensation drivers page
Additional insights into the compensation model
Base value
The compensation for base value represents the predicted salary for an employee who falls into the lowest-paid categories across all variables in your compensation model, including those with the lowest numerical values. Essentially, it's the theoretical minimum salary an employee would receive within your dataset, according to your model, as show in the following illustration:
This base value, along with the "estimated impact on compensation" from other factors, forms the bedrock for all predicted compensation calculations.
Multicollinearity
Multicollinearity occurs when variables in your dataset are highly, or even perfectly, correlated. In such cases, these variables essentially convey the same information to the model, making it difficult to isolate their individual impact on the outcome.
Consider an example: if every employee in a specific job role possesses the same high level of education, and no one else in the organization shares that educational attainment, it becomes impossible for the model to distinguish whether salary differences for this group are due to their education level or the inherent requirements of their job.
This strong correlation can lead to model instability, making it necessary to remove one of the highly correlated variables from the regression analysis. If your data exhibits perfectly correlated variables, you will find an additional table indicating which variables were removed due to multicollinearity, as shown in the following picture:
After an initial review of your compensation model, you may identify additional factors that need to be incorporated into your data. This iterative process of refinement is crucial for developing an optimal compensation model tailored to your organization's needs. For a list of commonly used factors in pay equity analysis, please refer to Commonly used factors in Pay Equity Analysis.
Good to know
Even though the results of a salary model are technical, they offer valuable insights into the factors driving compensation. Since this model will serve as the foundation for salary raises and other features, it's crucial that it's both robust and accurately reflects your organization's operational realities and values.
As you review the model, consider these questions:
Do the variables and their impacts make sense? For instance, do variables you expect to increase pay show a positive impact?
Do variables you anticipate might lower salaries have a negative impact? Keep in mind that the estimated impact of categorical variables is always relative to the base category listed in the table.
Once your compensation model is finalized and reviewed, you'll have also measured your initial Adjusted Pay Gap. To understand these gaps in more detail, see Interpreting pay gaps.
The next logical step is often: "How can we remediate these gaps?" To answer this, you can run a Raise Suggestion Analysis. Our article, Getting suggested remediation actions, provides guidance on how to proceed.