Probability and conditional probability

Problem-Solving and Data Analysis is about 15% of the Math section — that is College Board's published domain weight, and it works out to roughly six or seven of the 44 Math questions. Probability is one of the domain's seven skill points. How the questions divide among those seven is not published, so the "one or two probability questions per test" figure that circulates in prep materials is an estimate drawn from released item pools, not an official statement. What is not in dispute is the format: the overwhelming majority of probability items on the digital SAT are built on a two-way frequency table.

A two-way frequency table classifies one group of people or things by two traits at once. The rows are the categories of one trait, the columns are the categories of the other, each interior cell holds the count of members with that row's trait and that column's trait, and the margins hold the totals. Every count you need is already printed. Nothing has to be multiplied, no formula from a statistics course is required, and the arithmetic is one division.

So the item cannot be testing whether you can count. It is testing one decision: which total goes underneath. A single interior cell can legitimately sit over three different denominators — its row total, its column total, or the grand total — and each of the three answers a genuinely different question about the same people. The test knows this, which is why a well-built probability item puts all three numbers in front of you as choices.

It is worth saying plainly how narrow the scope is, because probability is a large subject and almost none of it is here. Conditional probability on the digital SAT means reading a denominator off a table. There is no multiplication rule to memorise, no independence formula to apply, no permutations or combinations, and no probability distributions. You will never be asked to apply Bayes' theorem by name — although the reversed-conditional questions in this lesson are exactly what it computes, done by counting instead.

If the phrase "of the 50 employees who took the bus, 30 were late" already reads to you as 30/50, skip ahead; the rest of the page is one idea applied repeatedly and you have it. If probability still feels like a topic with rules in it, this is the highest-value block here, because there are no rules — there is one fraction.

A probability is a fraction: how many outcomes count, over how many outcomes there are to draw from. That is the whole definition. It runs from 0 (never happens) to 1 (always happens), so a value outside that range is an arithmetic error rather than a strange result. A number above 1 nearly always means two overlapping groups were added; a negative one means a subtraction went the wrong way.

"Selected at random" is doing one specific job: it means every member of the pool is equally likely, which is exactly what allows the answer to be one count divided by another rather than a weighted average. It is not a hint about which pool.

Fraction, decimal, and percent are the same number in three costumes: 30/50, 0.6, and 60%. Divide to get the decimal (30 ÷ 50 = 0.6), multiply by 100 to get the percent. The test specifies which it wants — in the choices, or in a phrase like "to the nearest hundredth" — and the right number in the wrong form still loses the point.

The complement: P(not A) = 1 − P(A). If 0.6 of a group were late, 0.4 were not. This is the fast route whenever the trait you want spans more cells than the trait you do not want — count the small side and subtract.

Now the table itself, using the one from the prequestion. Two hundred employees, each classified twice: by how they commute (drives, or takes transit) and by age (under 40, or 40 or older). Four interior cells hold four counts — 45 drive and are under 40, 105 drive and are 40 or older, 30 take transit and are under 40, 20 take transit and are 40 or older. Add across a row and you get how many drive (150) and how many take transit (50). Add down a column and you get how many are under 40 (75) and how many are 40 or older (125). Both directions must land on the same grand total of 200, and checking that they do is the cheapest error-detector in the skill: if a row does not sum to its printed total, a cell has been misread.

Three kinds of question live in those four cells, and they differ only in the denominator. A joint probability — "the probability that a randomly selected employee takes transit and is under 40" — is one cell over the grand total: 30/200 = 0.15. A marginal probability — "the probability that a randomly selected employee takes transit" — is one row total over the grand total: 50/200 = 0.25. A conditional probability — "given that the employee takes transit, the probability that the employee is under 40" — is one cell over that row's total: 30/50 = 0.6. Same table, three answers, and two of them are built from the very same 30 people.

The word "conditional" carries no extra machinery. It does not name a different kind of probability with different rules; it names a probability taken after part of the table has been discarded. That is why conditional probability is a denominator skill and not a formula skill.

One protocol, run in this order, every time. It costs about eight seconds and it is the entire skill.

Step 1 — find the conditioning phrase and underline it. The vocabulary is small and it repeats: "given that," "among," "of the," "from those who," "if the selected student is." If one is present, the group it names becomes the new population. If none is present, the population is everyone in the table.

Step 2 — write the denominator down before you go looking for the numerator. This inverts the order almost everyone uses, and inverting it is the point. The numerator is easy and it is a magnet: once a cell is in your hand you will divide it by whichever total your eye lands on, and on a printed table the grand total is the one that catches the eye. Commit to the pool first and the numerator has nowhere wrong to go.

Step 3 — the numerator is the count of members inside that denominator who also have the trait the question asks about. Inside is the operative word. If the denominator is a row total, the numerator can only be assembled from cells in that row.

Step 4 — check two things before you write anything down. The numerator must be less than or equal to the denominator, and the value must lie between 0 and 1. A numerator drawn from outside its own pool fails the first check almost every time it happens.

Step 5 — reread the last line of the question. It may want a percent when you produced a decimal, a count of people when you produced a fraction, or the other conditional. This step recovers more points near the top of the score range than any amount of additional technique.

The phrase-to-denominator map, stated once so it never has to be re-derived under time pressure. "A randomly selected member is A and B" → the A-and-B cell, over the grand total. "A randomly selected member is A" → A's row (or column) total, over the grand total. "Given that the member is B, the probability that the member is A" → the A-and-B cell, over B's total. "Of the members who are B, what fraction are A" → identical to the previous one. The SAT alternates those last two wordings freely and they mean exactly the same thing.

One more pattern, because it uses a different rule: "A or B" always means A, or B, or both — the inclusive reading, on every item this test writes. Count it as (A's total) + (B's total) − (the cell where they cross), over the grand total. Subtracting the shared cell is not a refinement to apply when you have time; skipping it counts those members twice and can push the answer above 1.

Once the denominator rule is automatic, the points left in this skill sit in six places: tables you have to finish building, conditionals stated as percentages of a subgroup, counts asked for instead of probabilities, relative-frequency tables, rate-versus-count comparisons, and a short list of execution errors that survive knowing every rule above.

Incomplete tables. The hard version of this item prints half the entries and hides a constraint in the surrounding sentence, and the reconstruction is the question. Two rules recover everything: every row and every column sums to its printed total, and the row totals and the column totals must both sum to the same grand total. Fill from whichever line has exactly one blank left, then verify each filled cell against the other direction — subtracting once and trusting it is how a single slip propagates into a wrong answer that looks entirely reasonable. When the constraint arrives as a percentage ("40% of the students who chose the morning session were seniors"), convert it to a count immediately and write the count into the table. Percentages left as percentages get applied to the wrong base a few lines later.

Conditionals stated as percentages of a subgroup. "Seventy percent of the 400 volunteers completed the training; 30% of those who completed it and 55% of those who did not took the follow-up survey" is a two-way table written in prose. Build it: 280 completed and 120 did not; 0.30 × 280 = 84 and 0.55 × 120 = 66 took the survey; 150 took the survey in total. Every percentage applies to its own subgroup and never to the whole. Once the counts are on paper the item is an ordinary denominator question again — "given that a volunteer took the survey, what is the probability the volunteer completed the training" is 84/150 = 0.56. The characteristic error here is not arithmetic. It is answering 0.30, which is the reversed conditional, and which is sitting in the choices precisely because building the table correctly and then reading the wrong direction off it is what a strong student does under time pressure.

Counts, not probabilities. A large share of the hardest items in this skill end with "how many" rather than "what is the probability," and every step before the last line is identical. Two failure modes, both of which produce a number that looks like an answer: reporting the fraction when a count was asked for, and stopping at a partial count — the subgroup you happened to compute rather than the total across subgroups. The direction is worth stating explicitly: a probability multiplied by the size of the population it was measured on gives the count. Multiply by any other population and the answer is wrong by exactly the ratio between the two.

Relative-frequency tables. Sometimes the table holds proportions or percentages instead of counts, which means a denominator has already been applied — and the question is which one. If each row sums to 1 (or to 100%), the entries in a row are already conditional probabilities given that row, and the answer can be read straight off; the column conditionals are not recoverable without the underlying counts. If the whole table sums to 1, every entry is a joint probability and the conditionals still have to be built by dividing by a row or column sum. Check what the table sums to before reading a single value out of it, because the same printed number means different things in the two cases.

Rates versus counts. When two groups are different sizes, the group with more members having a trait can be the group with the lower rate of it. Two clinics recording 414 successes out of 460 and 132 out of 150 have the first ahead on both count and rate — but change 132 to 141 and the counts still favour the first clinic (414 to 141) while the rates reverse (0.90 to 0.94). Any comparison across groups of unequal size has to be made on the conditional probabilities, never on the cell counts, and items are built specifically to make the raw count the tempting answer to a question about likelihood.

Independence, when it is asked in words. If a conditional equals the corresponding unconditional — a row's internal split matches the table's overall split — the condition carried no information and the two traits are independent. The test rarely uses the word. It asks whether the data "suggest an association," and the answer is decided by comparing one row's conditional probability against another row's, not by comparing counts across rows of different sizes. Note also the ceiling on what any two-way table can support: an association between two survey columns is not evidence that one caused the other, and no arrangement of counts becomes a causal claim without random assignment.

Execution errors, which is where the last few points actually live. Answering the right question about the wrong quantity — the reversed conditional, the count when the probability was wanted, the complement when the trait itself was wanted. Rounding before the final step: keep 8/15 as a fraction and convert once at the end, because 0.53 entered where 0.533 was required is marked wrong. Reading a cell from the wrong row when a table has three or more rows and the question lists them in a different order than the table does. And the student-produced response format rules, which are College Board's own published directions and not prep-industry folklore: a fraction is accepted, an unreduced fraction is accepted, symbols such as the percent sign are not, and a mixed number is not — five halves is entered as 5/2 or as 2.5, never as 2 1/2, which the scoring reads as twenty-one halves. If the question asks for a percent, the grid takes the number of percent (60), not the decimal (0.6), so the last line decides the format as well as the quantity.