Inference from sample statistics and margin of error

This skill point is about using a number measured on a sample to say something defensible about a population that was never measured. The Digital SAT tests the reasoning, not the statistics: you will be given a sample statistic and a margin of error, and asked what may legitimately be concluded from them. No released item has required computing a margin of error from raw data — the value is always supplied, the standard-error formula is not on the reference sheet, and it is not needed.

Four shapes cover nearly every item. BUILD: given an estimate and a margin, identify the plausible interval, or a value inside it. INTERPRET: four sentences about the same study, of which three overreach, understate, or describe the wrong quantity. COMPARE: two studies, and a question about which has the smaller margin of error or what a wider interval implies. SCOPE: which population a result extends to, and what the study's design does or does not license.

College Board publishes the domain weight, not the skill weight: Problem-Solving and Data Analysis is about 15% of the scored Math questions — five to seven of them — split across seven skill points. Counts you will see for this particular skill, usually one item per form and sometimes none, are prep-industry inferences from released material rather than official figures. Treat them as a rough prior; the reason to learn this properly is not frequency but that the items are nearly all recoverable, because the arithmetic is trivial and the difficulty is entirely in the sentence.

That is the structural fact to carry into every item below. On most Math topics a wrong answer means you did the algebra wrong. Here, the arithmetic is one addition and one subtraction, and three of the four choices will contain the correct numbers. The question is which sentence they are wrapped in.

If "a sample mean of 12.5 with a margin of error of 0.8" already reads to you as "plausibly 11.7 to 13.3, for the population, not for any one member," skip to the next block. If it doesn't, this block builds the whole idea from nothing and it is the most valuable thing on this page.

A POPULATION is every member of the group you actually care about — all 1,200 employees at a plant, all adults in a city, all bottles produced in a day. A SAMPLE is the subset you actually measure. You take a sample for one reason: measuring everyone is impossible or unaffordable. A number computed from the population (its true mean) is called a parameter, and you almost never know it. A number computed from the sample (the sample mean) is called a statistic, and you always know it exactly. The whole topic is the bridge between the two.

Here is the problem the bridge has to solve. Draw 400 adults at random and you get a mean of 34.2 minutes. Send someone out to draw a different 400 the same way and they get 33.8. A third draw gives 34.9. Nobody made a mistake — different samples simply contain different people. That spread across hypothetical repeat samples is called sampling variability, and it is the only thing a margin of error measures.

A margin of error is reported so you can convert a single estimate into a range. The rule is one line: plausible interval = estimate − margin, up to estimate + margin. Written with the shorthand, estimate ± margin. For 34.2 ± 1.6, the interval runs 32.6 to 35.8. Write both endpoints out in full before you look at any answer choice — the most common careless loss on this topic is a subtraction slip on a decimal, and it is fully preventable.

The margin is the HALF-width of the interval, not the width. If a question hands you the interval instead of the margin — "plausibly between 61.4 and 68.2" — recover the two numbers by taking the midpoint and the half-gap: estimate = (61.4 + 68.2)/2 = 64.8, margin = (68.2 − 61.4)/2 = 3.4. Check it closes: 64.8 − 3.4 = 61.4 and 64.8 + 3.4 = 68.2. That check takes four seconds and catches the halving error every time.

"At the 95% confidence level" is a statement about the procedure, not about this one interval. It means that if you repeated the whole exercise many times — draw a fresh random sample, compute the estimate, build the interval — about 95% of the intervals you built would contain the population's true value. That is why every credited SAT answer says it is PLAUSIBLE that the parameter lies in the range, and never that it definitely does. You will hear people say "there is a 95% chance the true mean is in this interval"; that phrasing is loose, statisticians dispute it, and it is not how a credited answer is ever worded.

One notation trap before anything else. When the estimate is a percentage, the margin is given in PERCENTAGE POINTS, not as a percent of the estimate. "38%, margin of error 3 percentage points" means 35% to 41%. It does not mean 38 ± 3% of 38, which would be 36.86% to 39.14%. Percentage points are additive; percents of a number are multiplicative, and the SAT writes distractors from that confusion in more than one topic.

Finally, the design words. random sampling means members of the population were chosen at random to be measured; it is what licenses generalising the result back to that population, and nothing more. random assignment means subjects already in hand were assigned at random to a treatment or a control group; it is what licenses a causal claim. They are different operations, they license different conclusions, and confusing them is the most heavily tested idea in this corner of the syllabus.

Step 1 — name the population that was actually sampled, out loud, before touching a number. Not the population the writer of the study wants to talk about; the one the sampling procedure could physically reach. "150 students selected at random from those in the library on Tuesday afternoon" is a random sample of Tuesday-afternoon library visitors. Underline that clause. On scope items it is the answer.

Step 2 — name the parameter being estimated, with its unit. A mean number of hours. A percentage of riders. A mean fill volume in fluid ounces. Say whether it is a mean or a proportion, and say what it is a mean or proportion of. Half of the interpret-the-sentence items are decided here, because the tempting wrong choice is about a real quantity that is not this one.

Step 3 — build the interval and write both endpoints. Estimate − margin, estimate + margin. Do it before reading the choices, for the same reason you predict a word before reading the choices on Words in Context: once both numbers are on your page, three of the four sentences stop looking plausible.

Step 4 — test each choice against three filters, in this order. RIGHT POPULATION: is the sentence about the group the sample was drawn from? RIGHT QUANTITY: is it about the mean or proportion being estimated, rather than about individuals, or about the sample itself, or about a different variable? RIGHT STRENGTH: does it say plausible rather than certain, and does it avoid claiming a cause? A choice must pass all three. Most wrong choices fail exactly one, which is what makes them tempting.

The weak path on these items is to read the four sentences and pick whichever feels the least objectionable. That is recognition, and the choices are written to defeat it — they are usually all true-sounding, and often three of them are literally true statements about something. The strong path is to arrive at the choices already holding a population, a quantity, an interval, and a verb.

It DOES quantify sampling variability: roughly how far this estimate could sit from the population's true value purely because of which members of the sampling frame happened to be selected. It DOES let you convert a point estimate into a plausible range for a population parameter. It DOES shrink as the sample grows and widen as the population's underlying standard deviation grows.

It does NOT bound the maximum possible error — a value slightly outside the interval is not impossible, merely less plausible. It does NOT guarantee that the true value lies inside; at 95% confidence, about one interval in twenty built this way misses. It does NOT describe individuals: an interval for a mean says nothing about the range that any given person, bottle, or fish falls in. And it does NOT detect bias of any kind — not a sampling frame that excludes half the population, not non-response, not a leading question, not a survey answered only by people who felt strongly. Every one of those shifts the centre of the interval and leaves its width untouched.

That last point is the one the test rewards most. Ask which errors a margin of error could possibly have measured, and the answer is: only the ones that come from the draw. Everything upstream of the draw — who could be drawn at all, what they were asked, who bothered to answer — is outside its jurisdiction, so a small margin of error is never evidence that a study is sound.

What a WIDER interval implies, and what it does not. Three things widen it: a smaller sample, a more variable population, and a higher confidence level. Widening means less precision — a larger set of population values remains plausible — and, in practice, a weaker ability to distinguish two groups from each other. What widening does NOT mean is that the study is worse, that the estimate is biased, or that the estimate is wrong. A careful researcher who reports a 99% interval will publish a wider range than a colleague who reports a 95% interval from identical data, and the wider one is the more cautious claim, not the sloppier one.

The trade is worth stating explicitly because the SAT tests it directly: confidence and precision are bought from each other. Demand more confidence that your net contains the fish and you must use a bigger net. The only way to have both — high confidence and a narrow interval — is to collect more data, which is why sample size is the lever every comparison item turns on.

At 1500 the concepts are not the problem; the sentence-level discrimination and the execution are. Start with the one that costs the most: two answer choices carrying identical numbers, differing only in the verb. "The mean is between 7.07 and 7.63" versus "it is plausible that the mean is between 7.07 and 7.63." A student who has done the arithmetic correctly and stops reading at the numbers picks the first one, and the item was written by someone who knew that. Read every choice to the end of its verb phrase before selecting, and treat any unhedged claim about a population parameter as disqualified on sight.

The second is percentage points versus percent, which turns up in this topic more than anywhere else because so many estimates are proportions. A margin of 3 percentage points on an estimate of 38% gives 35% to 41%. Read as "3 percent of the estimate" it gives 36.86% to 39.14% — a plausible-looking interval, centred correctly, that will match a distractor. Whenever the estimate is a percentage, say the words "percentage points" to yourself as you build the endpoints.

Third, answering the right question about the wrong quantity, which at the top of the scale costs more than any misconception does. The item asks for the margin of error and you report the interval's width. It asks for the largest plausible value and you report the estimate. It asks how much LARGER one study's margin is and you report the second study's margin. It asks about the population and every number you wrote down describes the sample. Before selecting, re-read the final clause of the prompt and say what your answer is a measurement of, and in what unit.

Fourth, the sample-size relationship in both directions. To halve a margin you must quadruple the sample; to cut it to a third, ninefold. Read backwards, a study with a quarter of the data has twice the margin. The trap built on top of this is subtler: an item that asks how the margin changes when the sampling method and confidence level are held fixed can be answered from the ratio alone, because the confidence multiplier and the population spread cancel. A strong student who knows the full formula sometimes selects "cannot be determined without the standard deviation" — a sophisticated answer to a question that was never asked.

Fifth, the over-correction. Students who have properly learned to distrust surveys start selecting the most sceptical sentence available, and the SAT punishes that as reliably as it punishes overreach. A genuine random sample of a plant's employees DOES support a conclusion about that plant's employees. "No conclusion can be drawn," "the sample is too small," and "only the surveyed people can be described" are wrong answers, written specifically to catch a student who has learned half of this topic. Scepticism is a filter, not a default; apply it to the frame, not to inference itself.

Sixth, the design vocabulary, because one clause in the stem decides which conclusions are on the table. Random SAMPLING from a population licenses generalising to that population, and nothing else — in particular, it never licenses a causal claim. Random ASSIGNMENT to groups licenses a causal claim about the people in the study, and generalising that cause to a wider population additionally requires the sample to have been randomly drawn. A study with neither supports description only. Locate both words in the stem before reading any choice; on hard items, one of the distractors is a perfectly correct statement about the other design.

Finally, the slips that are not knowledge failures at all and therefore survive to test day. Decimal subtraction under time pressure: 24.6 − 1.4 is 23.2, not 23.4. A margin stated in minutes against choices in hours. A study reporting thousands of gallons against a choice reporting gallons. An interval given as endpoints when the question wanted the margin. None of these are caught by understanding statistics better; they are caught only by writing both endpoints down and re-reading the ask, which is the cheapest habit on this page and the one that separates a 1500 from a 1600.