Start with the question the table is supposed to answer
Before you touch any outlier, decide what the state salary table is for. That choice changes the cleanup rule.
If the table is meant for reporting, you want a summary that is readable and fair. If it is meant for relocation or job planning, you want the middle of the market, not the loudest record in the set. If it is meant for compensation review, the raw record trail matters more than a neat state ranking.
That is why a salary-by-state guide should not start by deleting extremes. It should start by asking whether the data mix is even comparable. A state can contain several pay markets at once, and those markets do not all belong in one average.
Use a two-step outlier rule
A practical way to handle salary outliers is to separate review from removal.
- Anything beyond 1.5×IQR from the median deserves a closer look.
- Anything beyond 3×IQR is a stronger signal that the point is either a real tail value or a bad record.
IQR, the distance between the 25th and 75th percentiles, is useful because salary data is usually skewed. The mean can jump around too much when a few high earners are in the sample. Median and percentile bands are steadier for state-level comparisons.
| Signal | What it often means | Best move |
|---|---|---|
| Beyond 1.5×IQR from the median | Needs review, but not an automatic error | Inspect role, level, location, and pay type |
| Beyond 3×IQR from the median | Strong tail value or bad entry | Split it out, correct it, or keep it with a note |
| Mean far above median | High-end skew | Lead with median and percentile bands |
| Fewer than 30 observations | Ranking is fragile | Combine years or avoid state-only ranking |
| One metro drives most records | State average hides local pay | Separate metro and non-metro views |
This kind of rule works because it keeps the data honest without pretending all extremes mean the same thing. A high value is not automatically wrong. A low value is not automatically a bargain. The job is to find out why the point sits far from the middle.
Sort the outlier by cause, not by emotion
Most salary outliers fall into a few buckets.
1. The number is simply wrong
Sometimes the problem is a bad unit, a misplaced decimal, or an annualized hourly wage that was entered incorrectly. Those records should not shape a state summary at all. If the number is clearly broken, move it out of the comparison set and fix the source entry.
2. The role does not belong in the same group
A state table often mixes occupations, and that is where confusion starts. Engineering, nursing, sales, operations, and public administration each have their own pay bands. Entry-level and senior-level records also live in different worlds. If the outlier belongs to a different occupation or level, it is not a true outlier in the same group. It is a mixed group problem.
3. The pay type is different
Base salary and total compensation should not be blended as if they are the same thing. Bonus-heavy or equity-heavy pay can make one row look far above the rest, but the comparison is no longer apples to apples. Keep those views separate if you want a state comparison that means anything.
4. The geography is different
Remote work, metro work, and non-metro work often follow different pay rules. One large city can push a state average upward, while remote roles can pull in pay that is tied to employer policy rather than local labor costs. If the point came from a different geography rule, separate it before you judge the state trend.
5. The point is a real tail value
Sometimes the outlier is legitimate. A rare high-pay role in the same occupation family may sit far above the center and still belong in the dataset. In that case, keep it in the raw file, but do not let it set the headline number for the state.
Decide whether to keep, trim, or split
A clean workflow usually needs three layers.
Keep the raw record set for audit work. That is where you can see whether a point is a bad unit, a pay mismatch, or a real high-end record.
Use trimmed or winsorized values for internal dashboards. Trimming removes the most extreme points. Winsorizing keeps them in the set but caps them so they do not dominate the summary. Both approaches can make a chart easier to read, but neither should replace the raw record set.
Use median and percentile bands for career planning and relocation decisions. Those numbers show the center of the market and the spread around it without giving one extreme row too much power.
A state salary table is strongest when it does not try to do everything with one number. The raw file answers, “What is in the data?” The trimmed summary answers, “What does the chart need to show?” The median answers, “What is the typical pay level?”
What to do in common state-salary situations
A single very high number in one state
Do not delete it just because it looks impressive. First ask whether it belongs to the same role family and level. If it does, keep it as a tail value and keep the state median where it belongs. A few top-end records can be real without being representative.
A low number caused by the wrong conversion
Fix the conversion first. An hourly wage turned into an annual salary the wrong way can create a fake low or high point. That is not a market signal. It is a data entry problem.
A state with a large public-sector share
Public-sector pay can sit in a different band from private-sector pay. If you combine them, the state average can look flatter or stranger than it really is. Separate the groups before you summarize.
A state dominated by one metro area
The state average may simply be the metro average in disguise. Split the metro from the rest of the state so your summary reflects both markets instead of pretending they are the same.
A small sample state
If the table has fewer than 30 observations, one record can change the story too much. In that case, combine years, widen the geography, or stop treating the state ranking as firm.
A simple workflow you can reuse
- Group records by occupation and level.
- Separate base pay from total compensation.
- Split remote, metro, and non-metro records.
- Calculate the median, 25th percentile, and 75th percentile.
- Flag anything beyond 1.5×IQR for review.
- Treat anything beyond 3×IQR as a serious review case.
- Keep true tail values in the raw set, but keep them out of the headline if they would distort the state view.
- Add a note whenever a record is moved, split, or removed.
This approach is plain, but it works. It stops the common mistake of overreacting to the loudest point in the table.
Mistakes that make state salary pages worse
Do not delete every high number. That wipes out real premium pay and makes the state look flatter than it is.
Do not keep every extreme point in the headline summary. That turns a state guide into a pile of unrelated records.
Do not rank states by mean alone. The mean moves too much when the sample is skewed.
Do not mix remote and on-site jobs without a split. That blurs the geography signal.
Do not mix base salary with total compensation. That is one of the fastest ways to create fake outliers.
Do not rely on one metro to speak for an entire state. That hides local variation and makes the summary less useful.
Bottom line
The right move with salary outliers is not to flatten them away. It is to figure out whether they are errors, mixed categories, or real tail values. For a state salary guide, the best summary usually comes from separating groups first, using median and percentile bands second, and treating outliers as clues instead of automatic mistakes.
If a point belongs to the wrong bucket, split it. If it is broken, remove it. If it is real but unusual, keep it visible without letting it define the state.
Frequently asked questions
Should I delete salary outliers from a state table?
Only when the record is clearly wrong, such as a bad unit, a conversion error, or a mismatch in role or level. If the record fits the group, keep it and rely on the median and percentile bands.
Is the median better than the mean for salary by state?
Yes, in most cases. The median stays closer to the center of the market when a few records sit far above or below the rest.
What if the state has too few observations?
Treat the result as unstable. Combine years, widen the geography, or stop using a state-only ranking as if it were a firm answer.
How should I handle remote jobs that look like outliers?
Separate them before judging the state. Remote pay often follows employer policy and job level rather than the worker’s local market.
When should I use trimmed or winsorized averages?
Use them for dashboards or internal summaries when you want a cleaner chart. Use the median for career planning, since it shows the center of the market more clearly.