Conditional probability and independence

Objective

Given a word problem, table, or tree, you should be able to identify the event being conditioned on, calculate a conditional probability and a joint probability, use the multiplication rule, test the independence identity, and communicate the result without turning an observed association into a causal claim. The interview-ready sentence is: “I will restrict to the cases where B is true, use P(B) as the denominator, calculate P(A | B), and compare it with the marginal P(A).”

Start with a population filter

Let an event be a set of outcomes, such as “a session completes checkout.” The sample space is the collection of outcomes under consideration. A marginal, or unconditional, probability describes the whole defined population: P(A). A joint probability describes both events happening: P(A ∩ B).

A conditional probability changes the reference population. The expression P(A | B) is read “the probability of A given B.” The bar means “given”; it is not a division sign. Restrict first to outcomes where B happened, then measure how often A happens inside that smaller population:

P(A | B) = P(A ∩ B) / P(B),   provided P(B) > 0

In a count table, the same rule is even more concrete: P(A | B) = count(A ∩ B) / count(B). The numerator is the intersection cell. The denominator is the total for the conditioning row or column, not automatically the grand total. Words such as “among,” “for users who,” “given that,” and “if we know” usually tell you which event is the filter.

From conditional to joint: the multiplication rule

Rearrange the definition by multiplying both sides by P(B):

P(A ∩ B) = P(A | B)P(B)

This is what a tree makes visible. Take the branch probability for B, then the conditional branch probability for A after B: multiply along the path. If several disjoint paths end in A, add them to get the marginal. For a partition into cases B₁, B₂, …, Bₖ:

P(A) = Σᵢ P(A | Bᵢ)P(Bᵢ)

The addition rule here is not a shortcut for every probability question. We add only disjoint paths that cover the event. “And” follows a path and uses multiplication; disjoint “or” branches are combined by addition. Saying which operation you are using—and why the branches are disjoint—is a useful interview habit.

Worked calculation: audit every denominator

Consider 1,000 new-user sessions. Let B be “help search was used” and A be “checkout completed.” The sessions are counted once in this table, so every cell is a joint count:

Step 1 — name the marginal rates. Search occurred in 300 / 1,000 = 0.30, so P(B) = 0.30. Checkout occurred in 320 / 1,000 = 0.32, so P(A) = 0.32.

Step 2 — restrict to the search row. There are 300 search sessions, and 180 of them checked out. Therefore P(A | B) = 180 / 300 = 0.60. The denominator is 300 because the question says “among search sessions.”

Step 3 — calculate the complementary conditional rate. Among the 700 sessions that did not use search, P(A | B^c) = 140 / 700 = 0.20. These two subgroup rates, 60% and 20%, combine using their subgroup weights: 0.60(0.30) + 0.20(0.70) = 0.18 + 0.14 = 0.32, the overall checkout probability.

Step 4 — calculate the joint probability. Both search and checkout occurred in 180 of the 1,000 sessions, so P(A ∩ B) = 180 / 1,000 = 0.18. The multiplication rule checks it: P(A | B)P(B) = 0.60 × 0.30 = 0.18.

The conditional checkout rate among search users, 60%, is not the joint probability, 18%. Neither one is automatically a causal effect. The table says that search use and checkout are associated in these observed sessions; it does not say that forcing a user to search would raise checkout by 40 percentage points.

Independence: does knowing B change A?

Events A and B are independent when learning that one occurred does not change the probability assigned to the other, under the model. With a positive denominator, equivalent checks are:

The first identity says the conditional rate matches the marginal rate. The third follows directly from the multiplication rule: if P(A | B) equals P(A), substitute that equality into P(A ∩ B) = P(A | B)P(B). In the normal table, 0.60 ≠ 0.32, and 0.18 ≠ 0.32 × 0.30 = 0.096, so the events are dependent in this table.

Do not confuse independence with mutual exclusion. If two positive-probability events cannot happen together, then P(A ∩ B) = 0, while P(A)P(B) > 0; they cannot be independent. Independence allows joint cases. It means the rate does not change after conditioning, not that the events are physically unrelated and not that a hidden common cause is impossible.

Applied scenario: a checklist and a misleading overall rate

A product manager says, “People who open the onboarding checklist finish setup more often. Should we make it mandatory?” Let F be setup finished, O be checklist opened, and C be a cohort indicator. The descriptive table is:

Overall, opened users finish at 162 / 280 = 57.9%; users who did not open finish at 198 / 360 = 55.0%. The overall difference is only 2.9 percentage points. Within new users, the difference is 10 points; within returning users, it is also 10 points.

Why did the overall comparison shrink? The groups have different cohort mixes. Of the 280 openers, 180 are new users. Of the 360 non-openers, 180 are new users. Aggregation weights the same cohort-specific relationships differently. This is a conditional-probability problem before it is a modeling problem: state the population, preserve the conditioning set, and inspect the strata that could change the mix.

The decision question is also different from the descriptive question. “What fraction of openers finish?” asks for a conditional probability. “What would happen if we made the checklist mandatory?” asks for an intervention effect. Opening may be self-selected or may happen after a user is already engaged. Before changing the product, verify exposure and eligibility, define the time order, and randomize checklist exposure when feasible. Track setup completion alongside abandonment, time-to-complete, support contact, and any guardrail that would reveal a worse user experience.

Stratification and conditional independence

A third variable can change the relationship you see. We say A and B are conditionally independent given C when the events are independent within each relevant level of C. One form is:

P(A ∩ B | C) = P(A | C)P(B | C)

In the checklist example, cohort is a candidate conditioning variable. A single overall rate can hide a relationship that is consistent within cohorts, or it can create an apparent relationship when each cohort is internally balanced. This is the intuition behind Simpson-type reversals. The right response is not “always stratify on everything”; it is to name plausible pre-exposure variables, compare denominators within meaningful strata, and decide whether the product question requires a causal design or only a descriptive report.

Conditional probability also has boundaries. If P(B) = 0, the elementary expression P(A | B) is undefined, not zero. If only three cases satisfy B, the arithmetic may be defined but the estimate is fragile. Report the numerator and denominator, check filters and instrumentation, and avoid broad claims from a tiny conditioning group. Missing or selectively observed rows can change the population to which the result applies.

Nearest alternatives

The nearest mathematical alternative is the marginal probability P(A). Use it when the question concerns the entire defined population. Use P(A | B) when the question explicitly restricts the population. In a table, that is the difference between the grand-total denominator and the B-row denominator.

The nearest decision alternative is a causal effect. Conditional probability describes observed groups. If the question is “what fraction of people who did B experienced A?,” the conditional rate is direct. If the question is “what happens if we make B happen?,” use random assignment when possible, or a defensible quasi-experimental or adjustment strategy when it is not. Conditioning on variables does not automatically identify a causal effect.

A table or tree is also not the same as an independence test. The table expresses the relationship and supports a conceptual calculation. If the question is whether a finite-sample difference could be sampling noise under a specified design, add uncertainty or a formal test in the inference lessons. That test still would not turn observational association into causality.

Common interview trap

Interview practice bank

1. What does P(A | B) mean, and why is the denominator P(B)?

   Strong answer: It is the probability of A inside the subpopulation where B is known to be true: P(A ∩ B) / P(B), with P(B) > 0. Say “restrict first, then measure,” and name the row total in a table.

   Follow-up: How is this different from P(A ∩ B)? Signal: translates notation into a reference population.

2. A table has 40 successes among 100 treated users and 30 among 150 untreated users. Calculate three probabilities.

   Strong answer: P(S | T)=40/100=0.40; P(S ∩ T)=40/250=0.16; P(S)=70/250=0.28. Check the joint result with 0.40 × (100/250)=0.16.

   Follow-up: What changes for P(T | S)? Signal: keeps each denominator explicit and does not swap the condition.

3. Show why independence implies P(A ∩ B)=P(A)P(B).

   Strong answer: Start with P(A ∩ B)=P(A | B)P(B). Independence says P(A | B)=P(A), so substitute. The converse divides by positive P(B).

   Follow-up: Can mutually exclusive positive-probability events be independent? Signal: derives the identity and names its denominator condition.

4. A feature is used by 20% of users, and purchase given use is 12% while overall purchase is 8%. What can you conclude?

   Strong answer: Use and purchase are positively associated in the observed data: 12% versus 8%, a 4-point difference and 1.5× rate ratio. Do not call it a causal lift; ask about eligibility, timing, selection, and exposure.

   Follow-up: What evidence would support a causal claim? Signal: separates description, effect size, and intervention.

5. A large conditional lift disappears after splitting by device. What happened?

   Strong answer: Suspect composition, confounding, or selection. Recompute denominators within device, verify the split is pre-exposure, audit missingness and logging, and use correct subgroup weights for any standardized overall rate. The direction can even reverse.

   Follow-up: What would you do if a device stratum has only three observations? Signal: diagnoses aggregation rather than defending the dashboard number.

6. When would you use a probability tree instead of a two-way table?

   Strong answer: Use a tree for sequential evidence: multiply along paths and add disjoint paths. Use a table when joint counts and subgroup totals make the denominator comparison auditable. They encode the same relationships; choose the representation that matches the question.

   Follow-up: What is the alternative if the question is whether a product change causes an outcome? Signal: chooses a representation and recognizes the causal boundary.

7. A marketplace wants to know whether support contact predicts churn. How do you analyze it?

   Strong answer: Define an eligible-user population, contact window, and churn window so contact precedes outcome. Calculate churn conditional on contact and no contact; segment by tenure, issue severity, and channel; inspect selection and reverse timing. Use the result for prioritization, not a claim that preventing contact causes retention.

   Follow-up: What guardrail prevents the team from reducing contacts by hiding support? Signal: frames timing, denominators, segmentation, and product consequences together.

8. What does P(A | B) mean when no observed records satisfy B? What if only three do?

   Strong answer: With P(B)=0, the elementary conditional is undefined. With three cases, it is computable but unstable; report the count, check eligibility and instrumentation, and avoid broad generalization.

   Follow-up: Would smoothing fix the underlying problem? Signal: treats a denominator failure as a data and scope issue, not a number to invent.

Timed synthesis

Three quick checks

1. Why do P(A | B) and P(B | A) usually differ?

   They condition on different reference populations. The first divides the intersection by P(B); the second divides it by P(A). The shared cell does not make the denominators equal.

2. How do you derive the independence product rule?

   Start with P(A ∩ B)=P(A | B)P(B). If learning B does not change A, replace P(A | B) with P(A).

3. What should you report when the conditioning group is empty or tiny?

   For an empty group, say the conditional is undefined and investigate the population, filters, and logging. For a tiny group, show the numerator and denominator, describe the estimate as fragile, and limit generalization.

Interview-ready recap

• Translate first: “among” or “given” names the conditioning event.
• Restrict, then measure: P(A | B)=P(A ∩ B)/P(B); say the denominator aloud.
• Multiply a path: P(A ∩ B)=P(A | B)P(B); add disjoint paths for a marginal.
• Check independence: compare P(A | B) with P(A), or compare the joint with the product of marginals.
• Respect boundaries: zero denominator is undefined; a tiny denominator is fragile; association is not automatically causation.
• When the mix matters: stratify by plausible pre-exposure variables and match the next design to the decision.

Further reading: Penn State STAT 414 on conditional probability, independent events, and three-way tables. Technical sources support the probability identities; interview prompts are practice calibrated to the syllabus, not a claim about every employer’s loop.