What Makes a Count Binomial?
A count of successes is not automatically binomial just because it comes from repeating an activity. The model describes a particular trial scheme: the number of trials is fixed, each trial has a success or failure outcome, trials are independent, and the success probability is the same on every trial. To recognize the model, check those conditions in the context of the process—not just the form of the final count.
For example, counting defective items in a fixed set of independently tested products may fit the scheme. Counting defectives in a sample drawn without replacement generally does not: learning that one item is defective changes the chances for the next draw. The count alone does not tell us which situation applies; the way the data were generated matters.
The indicators \(I_i\) are Bernoulli variables: each takes the value 1 with probability \(p\) and 0 with probability \(1-p\). Thus, the defining structure is a fixed number of independent, identically distributed Bernoulli trials. In the definition, “mutually independent” means that the joint probabilities for any collection of trial outcomes factor into the product of their individual probabilities. Pairwise independence—checking only pairs—is a weaker condition and need not suffice.
A Three-Condition Recognition Test
Use the following questions in order. First, state what counts as success. Then check the number of trials, independence, and the success probability. Each question concerns the process that generates the observations.
There must be a specified number \(n\) of trial positions, not a number determined by the outcomes already observed. A fixed sample size can satisfy this condition even if the trial outcomes are random.
Knowing outcomes on some trials must not change the joint probabilities in a way that violates independence. A physical mechanism that links outcomes, or sampling without replacement from a finite population, can create dependence.
There must be one value \(p\) such that \(P(I_i=1)=p\) for every trial \(i\). Trials can be independent yet have different success probabilities, in which case the scheme is not binomial.
All three checks matter. Fixed trial number does not imply independence; independence does not imply equal probabilities; and equal marginal probabilities do not imply independence. If any condition fails, the usual binomial trial scheme is not justified, even if the observed count happens to look plausible under one.
Worked Example: Counting Failures in a Fixed Quality Check
A facility tests 10 items. Assume each item independently fails its test with probability \(0.08\). Let success mean “the item fails,” and let \(X\) be the number of failures. Does the count have a binomial model?
Fixed number: The plan tests exactly 10 items, so \(n=10\) is fixed before the outcomes are observed.
Two outcomes: Each item either fails (success under this chosen coding) or does not fail. Each trial has a binary outcome.
Independence: The model assumption says the item outcomes are independent. In particular, a failure on one tested item does not change the probability model for another.
Common probability: Every item has failure probability \(0.08\), so the same \(p=0.08\) applies to all 10 trials.
All defining conditions hold. Therefore the number of failures has a binomial model with parameters \(n=10\) and \(p=0.08\). Calling failure “success” is only a coding choice: the event being counted must be clearly specified, but it need not be a desirable outcome.
What the Conditions Imply
The checks are more than a convenient checklist: they specify the joint behavior of every ordered pattern of outcomes. The following result makes that implication precise without requiring a formula for the distribution of the total count.
Proof. Fix one particular pattern, such as success on some specified trial positions and failure on all the others. For each success position \(i\), the required event has probability \(P(I_i=1)=p\). For each failure position \(j\), it has probability \(P(I_j=0)=1-p\). The events across all positions are mutually independent, so the probability of their intersection is the product of these probabilities. There are \(k\) factors of \(p\) and \(n-k\) factors of \(1-p\), giving \(p^k(1-p)^{n-k}\). \(\square\)
This result explains why independence and a common probability are separate requirements. Independence lets us multiply probabilities for a specified pattern; a common \(p\) ensures that the factors do not vary by trial. A total count combines multiple ordered patterns, but diagnosing the scheme does not require grouping or adding those pattern probabilities.
Worked Example: Why Sampling Without Replacement Is Different
A lot contains 12 items, of which 3 are defective. Four items are selected uniformly at random without replacement. Let success mean that a selected item is defective. Check whether the number of defective items selected follows a binomial trial scheme.
Fixed number and binary outcomes: Four items are selected, and each selected item is either defective or not defective. These two conditions are satisfied.
Common marginal probability: At any one specified draw position, each of the 12 original items is equally likely to appear. Thus the probability of a defective item at that position is \(3/12=1/4\). The marginal success probabilities are the same at each position.
Independence: Compare the second draw’s success probability under two conditions. If the first item was defective, 2 defective items remain among 11, so
Without conditioning on the first result, the second position is equally likely to contain any of the original 12 items, so
These probabilities differ: \(2/11\ne 1/4\). The first and second outcomes are therefore dependent. Although the number of draws is fixed and each position has the same marginal probability of success, the independence condition fails. This count is not generated by a binomial trial scheme.
Equal Probabilities Do Not Replace Independence
The previous example had equal marginal probabilities but dependent outcomes. The converse issue also arises: outcomes may be pairwise independent and have the same success probability, yet fail to be mutually independent. Since the binomial scheme requires mutual independence, checking only pairs can give the wrong conclusion.
Worked Example: Pairwise Independence Is Not Enough
Choose uniformly at random one of the four three-bit strings \(000\), \(011\), \(101\), and \(110\). Let \(I_1,I_2,I_3\) be the bits, with a 1 counted as success. Are these three trials a binomial trial scheme?
Each coordinate is 1 in exactly two of the four strings. Hence \(P(I_i=1)=2/4=1/2\) for each \(i\). For any chosen pair of coordinates, the four possible pairs \(00,01,10,11\) each occur exactly once among the four strings. Each pair therefore has probability \(1/4\) for each of its four outcomes, so the coordinates are pairwise independent.
But all three coordinates are not mutually independent. The pattern \(111\) never occurs, so \(P(I_1=1,I_2=1,I_3=1)=0\), whereas mutual independence with marginal probabilities \(1/2\) would require this probability to be \((1/2)^3=1/8\). Thus the mutual independence condition fails. The number of successes is only 0 or 2 in this construction, also reflecting that it is not the count from three independent trials with success probability \(1/2\).
Unequal Probabilities and Outcome-Dependent Stopping
A fixed number of independent trials still does not guarantee a binomial scheme if the success probabilities differ. For example, the number of patients who respond to treatment in a fixed group is not automatically binomial merely because response is coded as success. If the patients’ response probabilities differ, the common-probability condition needs separate justification.
Worked Example: Independent Trials With Different Success Probabilities
Four independent laboratory tests have positive-result probabilities \(0.7\), \(0.7\), \(0.4\), and \(0.9\), respectively. Let \(X\) count the positive results. Is \(X\) a binomial count?
There are exactly four tests, and each result is positive or negative, so the fixed-number and binary-outcome conditions hold. The problem states that the tests are independent, so that condition also holds.
However, the success probabilities are not all equal: for example, the third test has probability \(0.4\), while the first has probability \(0.7\). There is no single common \(p\) for all four trials. Therefore \(X\) does not arise from a binomial trial scheme, despite the fixed number of independent binary outcomes.
A stopping rule can invalidate the fixed-number condition. Suppose a fair coin is tossed until the third head appears, and \(X\) is the number of tails observed before stopping. The number of tosses is not fixed in advance; it depends on the outcomes. The count \(X\) is therefore not the success count from a fixed-\(n\) binomial trial scheme. By contrast, if the question is how many heads occur in the first 10 tosses, the number of tosses is fixed. With independent fair tosses, that count does meet the binomial scheme’s conditions.
Common Mistakes
- Assuming repetition means independence. Repeated draws from a finite population without replacement are dependent. Check how the sampling process works rather than relying on the word “repeated.”
- Checking only the average chance of success. Having the same marginal probability on each trial does not establish independence, as the without-replacement example shows.
- Checking only pairs. Pairwise independence does not imply mutual independence for three or more trials. The full joint condition is required.
- Ignoring unequal trial probabilities. Independent trials with probabilities \(0.7,0.7,0.4,0.9\) do not have a common \(p\).
- Letting the outcomes determine the number of trials. “Continue until” rules generally do not give a fixed number of trials. Identify the count and the planned trial positions before applying the checklist.
- Confusing the label “success” with a favorable result. Success is the outcome chosen for counting. It can mean failure, disease, a negative result, or any other designated binary outcome.
The binomial scheme is useful because it replaces a detailed repeated-trial description with a model indexed by a fixed trial count and a common success probability. But that simplification is justified only when the trial design supports the assumptions. If a condition fails, a different model may be needed; relabeling outcomes or looking only at the resulting total cannot repair the mismatch.
Check Your Understanding
For each situation, identify which binomial-scheme conditions hold and explain your conclusion.
- A fixed group of 20 devices is tested independently, and each device has the same probability of passing. What additional outcome description is needed to define a success count?
- Three cards are drawn from a deck without replacement. Why is a fixed number of draws and a common marginal probability not enough to establish a binomial scheme?
- Five independent trials have success probabilities \(0.2,0.2,0.2,0.2,0.6\). Which condition fails?
- In the three-bit example, why do pairwise independence and equal marginal probabilities not establish mutual independence?
- A researcher records the number of successes before the fifth failure occurs. Is the number of trials fixed in advance? Explain.