Beyond the Averages: Understanding Predictive Power
Standard charts show you what is happening—for example, the median salary for a Senior Designer. But standard charts often fail to show you why those numbers are the way they are.
To help you build a better career strategy, we use advanced data science techniques to identify the real drivers of salary and satisfaction in the service design industry.
🎯 What is "Predictive Power"?
In our report, you will see badges labeled "Strong Driver" or "Moderate Driver." These are based on a metric called the Predictive Power Score (PPS).
While a traditional correlation (like Pearson's r) only looks for simple linear relationships, the PPS is much smarter:
- It handles non-linear patterns: It can find relationships that aren't just straight lines. For example, salary often follows a curve of "diminishing returns"—you might see a massive pay jump between year 1 and year 5 of your career, but a much smaller increase between year 20 and year 25. A traditional straight-line model would miss this nuance.
- It is asymmetric: It knows that while your Seniority Level might predict your Salary, your Salary doesn't necessarily predict your Seniority Level.
- It handles all data types: It can compare categorical data (like "Industry") with numerical data (like "USD Salary") in a way that traditional statistics cannot.
How to read the scores:
| Score | Label | Interpretation |
|---|---|---|
| > 15% | Very Strong | This variable is a primary driver. Changing this (e.g., getting promoted) has a high probability of changing your outcome. |
| 5% - 15% | Strong | This variable has a significant impact, but it's likely working in combination with other factors. |
| 1% - 5% | Moderate | This variable has a measurable effect, but it's not the main story. |
| < 1% | Weak / Noise | This variable has little to no predictive power over the outcome. |
🌲 The "Forest of Experts" (Random Forest)
To calculate overall "Feature Importance," we use an algorithm called Random Forest.
Think of it as asking 100 different experts (Decision Trees) to look at the data. Each expert looks at a random subset of the variables and tries to predict a salary. By averaging the findings of all 100 experts, we get a much more robust and reliable picture of what actually matters.
Why this is better than simple averages:
- It captures interactions: It understands that Years of Experience might matter more in the Finance industry than in Non-profits.
- It is resistant to outliers: A few extreme salaries won't trick the "Forest" as easily as they might trick a simple average.
😊 The Satisfaction Paradox
One of our most consistent findings is that while many things predict Salary (Level, Experience, Industry), almost nothing in our survey strongly predicts Compensation Satisfaction.
This is what we call the Satisfaction Paradox.
Our ML models show that even if you have a high salary and a senior title, your satisfaction score might still be low. This suggests that happiness in service design is driven by factors we don't (yet) measure numerically: Autonomy, Purpose, Team Culture, and Management Quality.
📈 The Paradox of Trends (Correlation vs. Prediction)
You might notice a chart showing a clear trend—for example, Company Size where median salaries increase as companies get larger—yet the badge says "Weak Driver." This seems like a contradiction, but it's actually a key insight from the PPS:
- Medians hide the "Noise": A bar chart shows the median (the middle person). But the PPS looks at everyone. In large companies, the range of salaries is massive (from juniors to directors). Knowing someone is at a large company doesn't help the model "predict" their specific salary because the overlap with small companies is too high.
- Hidden Drivers (Redundancy): Often, a trend is actually a different variable in disguise. Large companies have more Seniority Levels. The model "sees" that Seniority is the real driver. Once it knows your level, knowing your company size adds almost zero new information.
- Correlation ≠ Prediction: A trend shows that two things move together. Prediction proves that one thing determines the other.
⚠️ Important Limitations
Machine learning is a powerful tool, but it's not magic.
- Correlation is not Causation: Just because Level of Impact predicts salary doesn't mean that "faking impact" will get you a raise. It usually means that high-impact roles are compensated better.
- Survey Noise: Our models are only as good as the data provided. By combining 2025 and 2026 data, we analyze ~1,700 records for strong signals, but fine-grained details can still be noisy.
- Unmeasured Factors: Our models currently explain about 35-40% of salary variance. The remaining 60% is driven by things we don't see: your negotiation skills, your specific company's budget, your individual performance, and luck.
- Redundant Variables: If two variables are highly related (like "Age" and "Years of Experience"), the model will often pick the "strongest" one and label the other as weak, even if both show a trend in a chart.