Skip to content
SAT essentials · Math

SAT sample statistics and margin of error

A survey of some people estimates something about everyone. The margin of error says how far off that estimate might reasonably be.

These are the concepts this skill rests on, the same pages as in the lesson.

Samples and Populations

A Sample Stands In for a Population

The population is the whole group a question is about, such as all 4,000 members of a club. Asking every member usually takes too long or costs too much, so a researcher asks a sample instead: a smaller group chosen from the population.

A sample can stand in for a population when it's chosen at random: every member has the same chance of being picked, so chance decides who's in it, not anyone's choice. The usual way is to draw names at random from a complete list.

A population of 4,000 club members, drawn as a box of 60 dots. Eight dots, spread through the box, are blue: they stand for the members chosen at random. The random sample of 40 members gives a sample estimate: 30% say yes. The percentage for the whole population is not known. population: 4,000 members → random sample of 40 members sample estimate: 30% say yes blue dots: chosen at random

Every question in this skill has two numbers in it:

  • the population value, such as the percentage of all 4,000 members who say yes. You almost never know it.
  • the sample estimate, the same percentage measured on the sample. You know it exactly.

If 12 of the 40 sampled members say yes, the sample estimate is 1240=30%\frac{12}{40} = 30\%. That 30% is your best guess at the population value. It isn't the population value itself.

Every Sample Gives a Different Estimate

Chance decides who ends up in a random sample, so a second sample gives a different estimate, and a third gives yet another. Below, fifteen random samples of 100 members were taken from the same club.

A dot plot of the estimates from 15 random samples of 100 members each, on a scale from 15% to 45%. The estimates run from 22% to 39%, most of them between 26% and 35%. A dashed line marks the population value, 30%, near the middle of the estimates. 15 20 25 30 35 40 45 sample estimate (%) population value each dot is the estimate from one sample 15 random samples of 100 members each

Most estimates land near the population value, and a few land farther away. No single estimate is guaranteed to hit it.

That's why an estimate comes with a margin of error: a number that tells you how far a sample's estimate could reasonably be from the population value. The question always gives you the margin of error. You never have to calculate one from the data.

Estimate, Margin, Interval

Estimate Plus or Minus the Margin

Put an estimate and its margin of error together and you get an interval of plausible values for the population value:

estimate−margintoestimate+margin\text{estimate} - \text{margin} \quad \text{to} \quad \text{estimate} + \text{margin}

An interval on a scale of percentages from 36 to 50. The estimate, 43%, is a dot in the middle. The interval runs 3 points down to the lower end, 40%, and 3 points up to the upper end, 46%. Its width is 2 times 3, or 6 points. 36 38 40 42 44 46 48 50 percentage of all members estimate 43% − 3 + 3 lower end upper end 43 − 3 = 40 and 43 + 3 = 46 width: 2 × 3 = 6 points

Say a sample of club members gives an estimate of 43%, with a margin of error of 3 percentage points. Subtract 3 and add 3: the plausible values for the percentage of all members run from 40% to 46%. The interval's width is twice the margin, 2×3=62 \times 3 = 6 points.

The margin uses the same units as the estimate. A mean wait of 24 minutes with a margin of 2 minutes gives 22 to 26 minutes. For a percentage, the margin is in percentage points: a margin printed as "3%" beside an estimate of 43% still means 3 points each way, not 3% of 43%.

Endpoints Give Back the Estimate

Sometimes a question gives you the interval and asks for the estimate or the margin. The estimate sits right in the middle of the interval, and the margin is half its width.

An interval from 28 to 36 minutes on a scale from 24 to 40. The two endpoints are given. The estimate is the midpoint, (28 + 36) divided by 2, which is 32; the margin of error is half the width, (36 minus 28) divided by 2, which is 4. 24 28 32 36 40 mean wait (minutes) given: 28 given: 36 estimate: 32 estimate = (28 + 36) ÷ 2 = 32 margin = (36 − 28) ÷ 2 = 4

A report says the mean wait for all customers is plausibly between 28 and 36 minutes.

  • Estimate: the midpoint, 28+362=32\frac{28 + 36}{2} = 32 minutes.
  • Margin of error: half the width, 36−282=4\frac{36 - 28}{2} = 4 minutes.

Check: 32−4=2832 - 4 = 28 and 32+4=3632 + 4 = 36. The full width, 8, is not the margin. Halving it is the step people forget.

Which Values Are Plausible

A value inside the interval is a plausible population value. A value outside it isn't supported by the sample.

An interval from 55% to 63% around an estimate of 59%, on a scale from 50% to 70%. Three values are tested: 58% is inside the interval and is plausible; 52% and 66% are outside it and are not supported. 50 55 60 65 70 percentage of the population 52% not supported 58% plausible 66% not supported estimate 59%, margin 4 points: 55% to 63%

An estimate of 59% with a margin of 4 points gives 55% to 63%. So 58% is plausible, but 52% and 66% aren't supported.

Be careful with both of those words:

  • Plausible doesn't mean proven. 58% might be the population value, but so might 61%. The sample can't tell them apart.
  • Not supported doesn't mean impossible. 66% is unlikely given this sample, but another sample could tell a different story.

An interval written "from 55% to 63%" includes its endpoints, so 55% and 63% are plausible too.

When an Interval Crosses a Boundary

Some claims are about a range, not a single value: "at least 60%," "more than half," "under 30 minutes." For these, compare the whole interval with the claim's boundary, not just the estimate.

An interval from 58% to 66% around an estimate of 62%, on a scale from 50% to 70%. A dashed line marks the goal, 60%, and crosses the interval. The part below 60% is orange and the part at 60% or more is blue: values on both sides of the goal are plausible. 50 55 60 65 70 percentage of all adults goal: 60% below 60% plausible 60% or more plausible 62% ± 4 points: 58% to 66% Both sides of 60% are plausible.

An estimate of 62% with a margin of 4 points gives 58% to 66%, and the goal was at least 60%.

  • The interval crosses the boundary. Values below 60% and values of 60% or more are both plausible, so the sample can't tell you whether the goal was met.
  • The whole interval is on one side. If the interval had been 61% to 69%, every plausible value would meet the goal, and the sample would support the claim.

The estimate alone, 62%, is above 60%. That's not enough: the margin decides.

Comparing Two Intervals

Two groups, two random samples, two estimates. Do the data show that the groups really differ? Build both intervals first, then look for a value they share.

Two pairs of intervals. Top: Group A, 47% plus or minus 4, runs from 43% to 51%; Group B, 53% plus or minus 3, runs from 50% to 56%. They overlap, so no difference is shown. Bottom: Group C, 22% plus or minus 2, runs from 20% to 24%; Group D, 29% plus or minus 3, runs from 26% to 32%. They do not overlap, so D is higher. 40 45 50 55 60 A: 47% ± 4 B: 53% ± 3 overlap: no difference shown 15 20 25 30 35 C: 22% ± 2 D: 29% ± 3 no overlap: D is higher
  • They overlap. Group A runs 43% to 51%, and Group B runs 50% to 56%. Some values, such as 50%, are plausible for both groups, so these intervals don't show that the groups differ. They don't show that the groups are equal, either.
  • They don't overlap. Group C runs 20% to 24%, and Group D runs 26% to 32%. Every plausible value for D is above every plausible value for C, so the data support that D's percentage is higher.

If you compared only the estimates, 47% and 53%, you'd claim a difference the data don't actually show.

What a Margin of Error Means

The Margin Is About the Population Value

A margin of error is about one number only: the population value, such as the mean or percentage for the whole population.

An interval from 22 to 26 minutes, in green, on a scale of mean commute times from 18 to 30. The sample mean, 24 minutes, is a blue dot: it is known exactly. The population mean is unknown, shown by a question mark; it is plausibly anywhere from 22 to 26 minutes. 18 20 22 24 26 28 30 mean commute time (minutes) ? sample mean: 24 minutes (known) population mean: unknown, plausibly from 22 to 26 minutes

A random sample of a city's bus riders had a mean commute of 24 minutes, with a margin of error of 2 minutes. That backs up exactly one sentence:

The mean commute time for all the city's bus riders is plausibly between 22 and 26 minutes.

Both bold parts matter. The sample's own mean doesn't need a margin: it's exactly 24 minutes. And the interval is a range of plausible values, not a promise.

Not the Spread of the Individual Values

The interval of 22 to 26 minutes is about the mean, not about individual riders. The thirty riders in that sample had commutes anywhere from 12 to 36 minutes.

A dot plot of the commute times of the 30 riders in the sample, from 12 to 36 minutes; their mean is 24. Above it, in blue, the interval of plausible values for the mean commute time of all riders runs only from 22 to 26 minutes, and most riders in the sample are outside it. 10 15 20 25 30 35 40 commute time (minutes) mean: plausibly 22 to 26 each dot: one rider in the sample
  • Spread of the data: how much individual values differ from one another. A standard deviation measures it.
  • Margin of error: how far the sample's mean could reasonably be from the population mean.

So "most riders commute between 22 and 26 minutes" isn't supported. In fact, most riders in the sample are outside that range. A larger sample shrinks the margin of error, but it doesn't make individual commutes any less spread out.

Plausible, Not Certain

Picture taking many random samples from one population and building an interval from each one. Margins of error are set up so that most of those intervals catch the population value, but not all of them.

Ten intervals, one from each of ten random samples, stacked above a scale from 16% to 44%. A vertical line marks the population value, 30%. Nine of the intervals cross the line. One, in red, from 31% to 39%, misses it. 16 20 24 28 32 36 40 44 population value misses

So an interval is never a guarantee. Watch for three readings that go too far:

  • "Certainly between." The population value is plausibly in the interval, not certainly.
  • "A 3% chance." A margin of 3 points is a distance. It isn't the chance that the estimate is right or wrong.
  • "Another sample would land in it." The interval is about the population value. A second sample's estimate could fall outside it.

Sample Size

A Larger Sample, a Narrower Interval

When random samples come from the same population and answer the same question, a larger sample is expected to give a smaller margin of error: a narrower interval and a more precise estimate. A smaller sample gives a larger margin.

Two intervals on a scale from 25% to 55%. A sample of 100 people gives 40% plus or minus 10, from 30% to 50%, in orange. A sample of 400 people gives 38% plus or minus 5, from 33% to 43%, in blue: narrower, with its center a little lower. 25 30 35 40 45 50 55 estimate (%) 100 people: 40% ± 10 400 people: 38% ± 5 more people: a narrower interval

Sample size changes the width, not the center. A larger sample's estimate won't necessarily land near the first sample's estimate. It may come out a little higher or lower, like the 38% here.

A margin also depends on how varied the population is. For a mean, more spread-out individual values give a larger margin. For a percentage, margins are largest for estimates near 50%. So you can compare two samples by size alone only when that variety is about the same for both: for a mean, individual values that vary about as much; for a percentage, estimates about equally far from 50%.

More data narrows the interval only when all of it comes from the population you're estimating. Adding last year's answers to this year's doesn't sharpen your estimate of this year's opinion. And making a sample bigger does nothing to fix one that wasn't chosen at random.

Four Times the Sample, About Half the Margin

How much does a larger sample help? When everything else stays the same, the margin of error is about proportional to 1n\frac{1}{\sqrt{n}}, where nn is the sample size. Because of that square root, the margin shrinks more slowly than the sample grows.

Sample sizeMargin of error
multiplied by 4divided by 4=2\sqrt{4} = 2
multiplied by 9divided by 9=3\sqrt{9} = 3
divided by 4multiplied by 2

So doubling a sample doesn't halve its margin. It divides the margin by only about 2≈1.4\sqrt{2} \approx 1.4. To cut a margin in half, you need about four times the sample, all else equal. For example, a margin of 10 points from 100 people becomes about 5 points with 400 people and about 2.5 points with 1,600.

Use the slider below to connect the margin to what it tells you. This illustration holds the estimate at 50% and the other survey conditions constant. With 100 people and a 10-point margin, plausible population percentages run from 40% to 60%. With 400, the 5-point margin gives 45% to 55%; with 1,600, the 2.5-point margin gives 47.5% to 52.5%.

A graph of margin of error, in percentage points, against sample size, from 0 to 1,600. The curve falls steeply, then levels off. Marked points: 100 people, margin 10; 400 people, margin 5; 1,600 people, margin 2.5. Each time the sample is multiplied by 4, the margin is halved. 0 400 800 1,200 1,600 0 2 4 6 8 10 12 sample size margin of error (percentage points) 100 people: 10 400 people: 5 1,600 people: 2.5

The narrower band means a more precise estimate of the population percentage, not less variation among individual people. A real new sample may move the estimate too; holding it at 50% here lets you see the width change on its own.

Scaling to a Population

Sample Proportion Times Population Size

To estimate how many people (or things) in a population have some trait, multiply the sample proportion by the population size:

estimated count=count in samplesample size×population size\text{estimated count} = \frac{\text{count in sample}}{\text{sample size}} \times \text{population size}

Two bars. The top bar is a sample of 200 households, with 22% shaded: 44 of 200. The bottom bar is the population of 7,500 households, with the same 22% shaded: 22% of 7,500 is 1,650 households. 22% sample: 44 of 200 households = 22% 22% population: 7,500 households 22% of 7,500 = 1,650 households

In a random sample of 200 of a town's 7,500 households, 44 recycle glass. The sample proportion is 44200=0.22\frac{44}{200} = 0.22, so about 0.22×7,500=1,6500.22 \times 7{,}500 = 1{,}650 of the town's households recycle glass.

Two checks keep you on track:

  • Multiply by the population the sample came from. A sample of the town's households scales up to the town's households, not the county's.
  • Use the whole population size. Don't subtract the sampled households: they're part of the population too.

A Mean Scales by the Number of Units

Sometimes a sample gives a mean instead of a proportion: pets per household, books per student, liters per day. A mean is an amount per unit, so multiply it by the number of units in the population to estimate a total.

Three boxes: a sample mean of 1.8 pets per household, times 1,500 households in the town, equals about 2,700 pets in the town. sample mean: 1.8 pets per household × 1,500 households in the town = about 2,700 pets in the town A mean is per unit: multiply by the number of units.

A random sample of a town's 1,500 households found a mean of 1.8 pets per household. The estimated number of pets in the town is

1.8×1,500=2,7001.8 \times 1{,}500 = 2{,}700

The same check applies: the 1,500 has to be the households the sample came from. If you multiplied by the number of people in the town, you'd count each household several times.

Scale Both Ends of an Interval

When the estimate comes with a margin of error, the population count is a range too. Build the interval, then multiply both ends by the population size.

Two lined-up scales. On the top scale of percentages, the interval 28% plus or minus 3 points runs from 25% to 31%. On the bottom scale of numbers of households, each end is multiplied by 4,000: the interval runs from 1,000 to 1,240 households, around 1,120. 22 24 26 28 30 32 34 percentage of households 28% ± 3 points: 25% to 31% 880 1,000 1,120 1,240 1,360 number of households × 4,000: 1,000 to 1,240 households

A random sample of a city's 4,000 households estimated that 28% have a vegetable garden, with a margin of error of 3 points. The plausible percentages run from 25% to 31%, so the plausible numbers of households run from

0.25×4,000=1,000to0.31×4,000=1,2400.25 \times 4{,}000 = 1{,}000 \quad \text{to} \quad 0.31 \times 4{,}000 = 1{,}240

Read what the question asks for. A plain "how many" with no margin wants the single estimate, 0.28×4,000=1,1200.28 \times 4{,}000 = 1{,}120. "The least plausible number" wants the lower end.

Who Was Sampled

The Estimate Covers the List, No Wider

A random sample tells you about the group it was drawn from, and no wider. Ask yourself who could have been selected. That group is what the estimate describes.

Three nested boxes. The outer box is all adults in the city. Inside it, in blue, is the library's list of members, the group the estimate covers. Inside that, in green, is the random sample of 300 drawn from the list. all adults in the city library members: the list the estimate covers this group random sample of 300

A random sample of 300 people drawn from a library's list of members describes the library's members. It says nothing about city residents without a library card, because none of them could have been chosen.

So when a conclusion names a group, check it against the list:

  • The same group: the conclusion is supported (with the interval, not a single value).
  • A wider group, such as all the city's adults: not supported, no matter how good the sample is.
  • Only the people sampled: too narrow. The whole point of a random sample is to describe the larger group it came from.

A Bigger Sample Cannot Fix a Biased One

When people decide for themselves whether they're in a sample, the sample is self-selected: volunteers, a website poll, people who choose to reply to a survey. The people who choose to answer can be different from the people who don't, in exactly the thing being measured.

Two intervals on a scale from 30% to 80%, with a dashed line at a population value of 48%. A random sample of 400 gives 46% plus or minus 5, from 41% to 51%, which covers 48%. A self-selected sample of 9,000 website volunteers gives 71% plus or minus 1: narrow, and far from 48%. 30 40 50 60 70 80 percentage population value: 48% random sample of 400: 46% ± 5 9,000 volunteers: 71% ± 1

A large self-selected sample can have a small margin of error and still be far from the population value. The margin of error only covers the luck of the draw in random sampling. It says nothing about how the sample was chosen. The same goes for a convenience sample, such as the people at one bus stop or the first customers through the door.

A sample that isn't random can't give you an estimate for the population, no matter how large it is. A bigger sample makes an estimate more precise; only random selection makes it fair.

Keep going

Check Your Skills, the strategy for each kind of question, worked examples and practice from Easy to Hard are in the Quaero SAT Course.

More Problem-Solving and Data Analysis