On this concept: Formula

Not sure which lesson to do first? Find your level and get a plan →

Math · Sets & Logic · Grade 9-12

Robustness

Robustness is the property of a method, model, or statistic that keeps working when its assumptions are slightly off or the data is a little messy.

xˉ=1n∑xi is sensitive to outliers; median is not (robust statistic)

Quick Answer

Robustness is the property of a method, model, or statistic that keeps working when its assumptions are slightly off or the data is a little messy. It matters when data may contain outliers or typos, when a model's assumptions are only roughly true, or when a design must survive conditions nobody planned for. The median of a list is robust: change one value to something wild and it barely moves. The mean is not.

Why This Matters

Real data is never clean: a sensor glitches, someone types 1000 instead of 100, a survey reaches the wrong group. A robust method gives nearly the same answer anyway, so one bad value can't decide your conclusion. Without the idea you might report an average salary of $234,800 for a team where four of five people earn under $50,000.

Intuitive Explanation

Take five test scores: 72, 75, 78, 80, 85. The mean and the median are both 78. Now someone mistypes the 85 as 850. The median is still 78, because it only cares which value sits in the middle. The mean jumps to 11555=231, because every value pulls on it, the typo included.

That is robustness in miniature: a robust result changes only a little when its inputs or assumptions change a little, or when a few inputs go badly wrong. You test it by breaking things on purpose: add an outlier, delete a data point, loosen an assumption, and see how far the answer moves. The median survives even if nearly half the data is corrupted; a single value can drag the mean anywhere. The same idea applies beyond statistics: a shelf rated for 50 kg that holds a 55 kg load is robust to an overloaded day.

Robustness is not accuracy. The mean is the best estimate of the centre for clean, bell-shaped data, which is why it is so popular, yet it is fragile when the data is dirty. Robustness is not sensitivity either, although the two are linked. Sensitivity is a number, how far the output moves per unit change in an input; a robust method is one whose sensitivity to bad inputs is small. Ask "does it survive?" for robustness and "by how much does it move?" for sensitivity.

Think of it like this. A robust method is a four-legged table on an uneven floor: one short leg makes it rock a little, but it stays up. A fragile one is a plate balanced on a pencil, perfect only while everything is exactly right. Unlike a table, robustness is always against a particular kind of trouble: the median shrugs off outliers but not a biased sample.

Core idea. Robustness is whether a result stays approximately correct when its assumptions are slightly violated.


When to Use

Think about robustness when data may contain outliers, typos or missing values, when an assumption such as "the data is bell-shaped" is only roughly true, or when choosing between methods that agree on clean data. Test by deliberately breaking an assumption and seeing whether the answer holds. If the question asks how much the output changes per unit of input, compute a sensitivity instead; if it asks how close a result is to the true value under correct assumptions, that is accuracy.

Tip. Ask: If the assumptions are slightly violated, does the result stay approximately correct?

How to Recognize It

Before trusting a result, list what it assumes about the data or the setting, then picture the most likely way each of those assumptions could be off.

  1. Could the data contain an outlier, a typo or a value from a different group?

    If yes, prefer methods that one bad value cannot move far, such as the median and interquartile range over the mean and range. If the data are clean and checked, the mean is fine and uses more of the information.

  2. If you change or remove one data point, does the answer jump?

    If it jumps, the result is fragile; say so or switch methods. If it barely moves, the result is robust to that kind of change, though not necessarily to others, such as a biased sample.

  3. Is the question how much the output moves per unit change in an input?

    If yes, compute a sensitivity, a rate with a number attached. Robustness is a verdict on whether a result survives broken assumptions; a low sensitivity is one piece of evidence for it.

  4. Are the assumptions known to hold exactly, so that only closeness to the true value matters?

    If yes, the question is about accuracy. Robustness asks what happens when assumptions such as bell-shaped data are only roughly true; error growth from computer rounding is numerical stability.

Robustness compared with Sensitivity (meta), Accuracy, and Stability (numerical)

All four rows describe how a result behaves when something about its inputs is imperfect or changed. Ask what you are judging: whether it survives broken assumptions, how much it moves per small nudge, how close it is when assumptions hold, or whether rounding errors grow.

Robustness

Meaning
Use this when you must judge whether a method or statistic stays approximately right under noisy data, outliers, or slightly broken assumptions.
Key test
If the assumptions are slightly violated, does the result stay approximately correct?
Formula
xˉ=1n∑xi is sensitive to outliers; median is not (robust statistic)
Example
Salaries are 40,42,45,47,1000 (thousands). Which summary, mean or median, is robust to the outlier?

Sensitivity (meta)

Meaning
Measures HOW MUCH output changes per input change; robustness is the desirable LOW-sensitivity quality.
Key test
Use when quantifying the rate of response, not judging whether a result survives.
Formula
ΔoutΔin
Example
How much an answer shifts per unit input error

Accuracy

Meaning
How close a result is to the true value under correct assumptions, not its survival when they break.
Key test
Use when measuring correctness on clean, on-assumption data.
Example
An estimate within 1% of the true mean

Stability (numerical)

Meaning
Whether an algorithm avoids amplifying rounding errors, a related but computation-specific notion.
Key test
Use when worried about floating-point error growth in an algorithm.
Example
An iterative method that doesn't blow up

Formula & Notation

xˉ=1n∑xi is sensitive to outliers; median is not (robust statistic)
A statistic T has breakdown point ε∗=min⁡{m/n:T can be made arbitrarily large by changing m of n data points}; median has ε∗=0.5, mean has ε∗=1/n

How to read it: xˉ denotes the mean; a statistic is robust if changing a few data values, even by a lot (outliers), changes the result only a little

Worked Examples

Example 1 — Outlier in a salary list

Easy

Salaries are 40,42,45,47,1000 (thousands). Which summary, mean or median, is robust to the outlier?
Show the steps and the answer
  1. Compute both: mean =40+42+45+47+10005=11745=234.8; median =45, the middle of the sorted list.

    We need each summary's value with the outlier present before we can see how much it was affected.

  2. Remove the outlier and recompute: the mean of 40,42,45,47 is 1744=43.5, and the median is 42+452=43.5.

    Comparing with and without the extreme value shows how much each summary depended on it.

  3. Measure the shift: the mean moved from 43.5 to 234.8, up 191.3; the median moved from 43.5 to 45, up 1.5.

    A robust summary changes only a little when one value is wild; the median did, the mean did not.

  4. Note that the mean, 234.8, is larger than four of the five salaries.

    A centre that describes nobody in the group is a clear sign that one value has dragged it.

Answer: Median (45) is robust; mean (234.8) is not

Takeaway: To judge robustness, recompute with and without the suspicious value; here the median barely moves, so it is the robust choice.

Example 2 — Measuring sensitivity instead

Standard

You're asked how many dollars the mean salary shifts per $1 added to the top earner. Is that robustness?
Show the steps and the answer
  1. Read what is asked: a rate, dollars of change in the mean per dollar added to one salary.

    A per-unit rate is a sensitivity, a number; robustness is the judgement about whether a result survives bad data.

  2. Add Δ to the top salary: the total rises by Δ, so the mean rises by Δ5.

    The mean divides the total by 5, so each added dollar is shared five ways.

  3. Divide by the change: Δ/5Δ=15 dollar per dollar.

    The ratio is the same for every Δ, so the mean's sensitivity to that salary is exactly 15.

  4. Compare with the median: raising the top salary leaves the median at 45, a sensitivity of 0.

    That zero is the reason the median is robust; the mean's 15 per dollar keeps adding up however large the raise.

Answer: 15 dollar per dollar added

Takeaway: Sensitivity is the per-unit number (15 for the mean, 0 for the median); robustness is the conclusion you draw from it.

Example 3 — Find the mistake: clean data is not a test

Application

A room sensor logs seven hourly temperatures: 21,22,22,23,21,22,−40 °C, where the −40 is a glitch. Kai writes: "The mean worked perfectly on last week's clean data, so it's robust. The average temperature is 13.0 °C." Find the mistake and correct it.
Show the steps and the answer
  1. Check Kai's arithmetic: 21+22+22+23+21+22−40=91 and 91÷7=13.0.

    The calculation is right; the error is the claim that the mean is robust.

  2. Recompute without the glitch: 1316≈21.8 °C.

    One bad reading moved the mean by almost 9 degrees, so success on clean data said nothing about dirty data.

  3. Use the median: sorted, the values are −40,21,21,22,22,22,23, so the median is the 4th value, 22.

    The glitch only changes which value sits at the end of the list; the middle stays put.

  4. Confirm: without the glitch, the median of 21,21,22,22,22,23 is 22+222=22.

    The same answer with and without the bad value is exactly what robust means.

Answer: Report the median, 22 °C; the mean of 13.0 °C was dragged down by one glitch.

Takeaway: Accuracy on clean data is not robustness; add a bad value and see whether the answer survives.

Where You'll See It

  • Why the news reports median house prices

    In a town where 9 homes sell for about $300,000 and one mansion sells for $5 million, the mean price is 9(300,000)+5,000,00010=$770,000. The median stays at $300,000, which is what a typical buyer actually pays.

  • Judges' scores with the extremes dropped

    Diving and figure skating discard the most extreme judges' scores before averaging, so one biased judge can't decide a medal. With scores 8.5, 8.7, 8.8, 8.9 and a rogue 5.0, the plain mean is 7.98; dropping the highest and lowest leaves 8.5+8.7+8.83≈8.67.

Common Mistakes

  1. The slip: Calling a method robust because it's accurate on clean data

    Instead: test it with outliers and broken assumptions.

  2. The slip: Confusing robustness with sensitivity

    Instead: robustness is the property (survives), sensitivity is the measure (how much it moves).

  3. The slip: Defaulting to the mean for messy data

    Instead: prefer a robust statistic like the median when outliers are likely.


Mini Practice

Try these on your own. Tap Reveal when you want to check.

  1. Test scores are 72, 75, 78, 80 and 85. Then the 85 is mistyped as 850. How do the mean and the median change?

    Hint: Recompute both summaries with the typo included.

  2. Find the mean of 4, 5, 5, 6, 40, and the trimmed mean after dropping the lowest and highest values.

    Hint: Add all five for the mean; for the trimmed mean, keep only the middle three.

  3. For the data 2, 4, 4, 5, 6, 7, 8, 30, compare the range and the interquartile range (IQR) as measures of spread.

    Hint: Find the medians of the lower four and the upper four values.

  4. The rule of 72 estimates doubling time as 72÷r years at r% growth. True doubling times are 11.9 years at 6%, 8.0 at 9% and 1.7 at 50%. Where is the rule robust?

    Hint: Compute the rule's estimate at each rate and compare.

  5. Question A: if one cost estimate is off by 5%, is the project still profitable? Question B: how many dollars does profit change for each 1% change in that cost? Which asks about robustness?

    Hint: Which question asks whether a conclusion survives, and which asks for a rate?

  6. A spam filter scores 99% on a clean test set. Jen says, "So the filter is robust." Correct her.

    Hint: Robust to what? Was it tested on messy or unusual messages?

Want the full set?

50 practice questions for this concept — free to try, every one with a complete worked solution showing the why, not just the answer.

Frequently Asked Questions

How many bad values can the median survive?

Almost half. In a list of 11 numbers the median is the 6th. You can make the 5 largest values as large as you like, even a billion, and the 6th value stays where it is. Statisticians call the share of data that can go wrong before a statistic breaks its breakdown point: about 50% for the median, and 0% for the mean, since one value sent toward infinity drags the mean along with it.

Is the median always better than the mean, then?

No. The median ignores how far values sit from the middle, so it throws information away. With clean, roughly symmetric data, the mean estimates the centre more precisely from the same sample. Budgets need the mean too: if one insurance claim in a hundred costs $1 million, the median claim may be small, but the total paid out depends on the mean. Choose by what the number will be used for.

What is a trimmed mean?

A compromise between the mean and the median. You drop a fixed share of the highest and lowest values, often 10% at each end, and average the rest. Diving and gymnastics judging have used versions of this, discarding the top and bottom scores so that one biased or mistaken judge cannot swing the result. It keeps most of the data's information while resisting a few wild values.

Does robustness matter outside statistics?

Yes. Engineers design a bridge for loads well above any expected traffic, so it stays up if the steel is a little weaker or the trucks a little heavier than planned; the extra margin is called a safety factor. A navigation route that still works when one road is closed is more robust than one that needs every road open. Any plan that survives small surprises is robust in this sense.

Learning Path

← Before

Sensitivity
Robustness

You are here

Next →

Nothing after this yet
You need Sensitivity first: how much an output moves per unit change in an input. Robustness turns that measurement into a judgement: a result is robust when bad inputs and loosened assumptions move it only a little. With it you can choose statistics for messy data (median and interquartile range over mean and standard deviation), use trimmed means, and stress-test a model before you rely on it.

See Also