Methodology

1. Data Sources

Rolehue salary data comes from two primary sources:

• User submissions (primary): verified professionals submit their real compensation anonymously through our submission form. Each submission includes job title, city, base salary, bonus (optional), and years of experience. All submissions are pseudonymized — no identity is stored alongside salary data.
• Public data (supplementary): when user submissions are sparse for a role/city combination, we supplement with publicly available data from: H1B visa salary databases (US), ONS/ASHE (UK), Statistics Canada, ABS (Australia), and Glassdoor/LinkedIn public ranges. Supplementary data is clearly marked and weighted lower than verified submissions.

Minimum cluster size: we only publish benchmarks when a role/city combination has at least n ≥ 25 data points. Below this threshold, the page shows “insufficient data” and invites submissions.

2. IQR-Based Outlier Detection

Before computing benchmarks, we remove outliers using the Interquartile Range (IQR) method:

1. Calculate Q1 (25th percentile) and Q3 (75th percentile) for the cluster.
2. Compute IQR = Q3 − Q1.
3. Define bounds: Lower = Q1 − 1.5 × IQR, Upper = Q3 + 1.5 × IQR.
4. Data points outside these bounds are flagged as outliers and excluded from public benchmarks.

Outliers are not discarded — they're retained in the database and can be re-included automatically if the cluster grows and their value falls within the new IQR bounds. This handles edge cases like legitimate ultra-high compensation at top-tier companies.

3. Clustering Methodology

Salary data is clustered across 6 dimensions:

When a sub-cluster (e.g., specific role × city) has n < 25, we roll up by removing the most specific dimension first: City → Country → broader Role category. Users always see the dimensional scope of their results.

4. Time Decay Weighting

Newer salary data is weighted more heavily than older data using exponential decay:

Where λ = ln(2) / 2.0 yields a half-life of 2 years. A salary submitted today carries full weight (1.0); a salary from 2 years ago carries half weight (0.5); from 4 years ago, quarter weight (0.25). Data older than 5 years is excluded entirely.

Weight tiers used in computation:

5. Benchmark Computation

After outlier removal and time decay weighting, we compute:

• Percentiles: P10, P25, P50 (median), P75, P90 — computed on weighted data.
• Mean: weighted arithmetic mean (informational, not the primary metric).
• Range: P10 to P90 span.
• Sample size (n): always displayed to give confidence context.

All monetary values are in USD unless explicitly noted. For non-US cities, we also show a PPP-adjusted equivalent.

6. Data Quality & Verification

• Pseudonym consistency: each submission generates a stable pseudonym, letting us detect duplicate submissions from the same person without knowing their identity.
• Statistical flagging: submissions that are extreme outliers (>3x IQR) are automatically held for review.
• Source labeling: each data point is tagged with its source (verified user, H1B, public survey), and source weights affect the benchmark computation.
• Freshness monitoring: clusters without new data in 6+ months are flagged as “stale” to users.

7. Limitations

• Self-selection bias: users who submit salaries may not represent the full population. We mitigate this by including public data sources.
• Small sample sizes: benchmarks with low n carry more uncertainty. We enforce n ≥ 25 but still encourage users to consider the sample size.
• Total comp vs base: Rolehue currently focuses on base salary + bonus. Equity (RSUs, options) is on the roadmap but not yet included.
• Geographic coverage: US cities have the most data. Non-US coverage is growing and may rely more heavily on supplementary sources.