When Is a Binomial Result Surprising?
In Shape of Binomial Distributions, you used \(n\) and \(p\) to describe how a binomial distribution is shaped. A new question is whether an observed count would be surprising if that binomial model were true. For example, if a process usually produces only a few successes, would observing many successes be unusual?
There are two useful ways to assess a result. The first is to calculate a probability for the result or a relevant group of results. The second is to compare the count with the distribution’s mean and standard deviation. These methods are related, but they are not identical: the mean-plus-or-minus-two-standard-deviations rule is a quick screening tool, while probabilities provide a more direct assessment.
The direction matters. If an observed count is high, the upper tail represents outcomes at least that high. If it is low, the lower tail represents outcomes at least that low. When a question asks whether a result is unusual in either direction, include both relevant tails. State which event you are assessing so the probability has a clear meaning.
As in Defining \(n\), \(p\), and \(X\) in Context, define \(X\) as the count of successes before calculating. The methods here assume the binomial model is appropriate. For a real chance process, check the BINS conditions as in The BINS Checklist for Binomial Conditions.
The Two-Standard-Deviation Guideline
For \(X\sim B(n,p)\), the mean is \(\mu_X=np\), and the standard deviation is \(\sigma_X=\sqrt{np(1-p)}\), as established in the earlier tutorials on the mean and standard deviation of a binomial distribution. The interval from two standard deviations below the mean to two above it is:
A count more than two standard deviations below or above the mean is a potential unusual result. Since a binomial count can only be a whole number from 0 through \(n\), interpret the interval using the possible integer counts. Do not round the interval endpoints to the nearest integer and then accidentally include a count that is actually outside the interval. Instead, determine which integers are strictly below the lower endpoint or strictly above the upper endpoint.
The guideline is useful partly because, in many distributions, values within about two standard deviations of the mean cover most of the probability. But a binomial distribution is discrete and can be skewed. That means the probability in its lower and upper tails need not be equal, and the guideline should not be treated as an exact probability rule.
A Practical Decision Process
State what \(X\) counts and identify the binomial parameters \(n\) and \(p\). Check that the binomial conditions are reasonable.
Decide whether the concern is a high count, a low count, or a result in either direction. Write the event using \(X\).
Calculate \(\mu_X=np\), \(\sigma_X=\sqrt{np(1-p)}\), and the endpoints \(\mu_X\pm2\sigma_X\). Translate those endpoints into possible counts.
Use an exact or cumulative binomial probability for the event. Compare it with 0.05 and explain what it means in context.
The order helps keep the two methods in perspective. The interval can provide a quick warning that an outcome is far from the mean. The probability answers how often the specified event would occur under the model. For a probability calculation, use the correct inclusive endpoint: “at least \(k\)” means \(X\geq k\), and “at most \(k\)” means \(X\leq k\), as in the earlier tutorials on binomial cumulative probabilities.
Worked Example: Unusual Counts in a Symmetric Model
Worked Example: Unusual Counts in a Symmetric Model
A student is practicing a skill in 12 independent rounds. Suppose the probability of success in each round is 0.5. Let \(X\) be the number of rounds with a success. Would a result of 2 or fewer successes, or 10 or more successes, be unusual?
State. The model is \(X\sim B(12,0.5)\). One trial is one round, success is completing the skill successfully in that round, and \(X\) counts successful rounds.
Plan. The model assumes two outcomes per round, a fixed number of 12 rounds, independence, and the same success probability of 0.5 in every round. These are the BINS conditions. The event includes both tails: \(X\leq2\) or \(X\geq10\). We will use the two-standard-deviation interval as a screen and calculate the probability of the stated event.
Do. The mean and standard deviation are:
The interval is \(6-2(1.7321)\) to \(6+2(1.7321)\), or approximately 2.536 to 9.464. The possible counts outside this interval are 0, 1, and 2 on the low side, and 10, 11, and 12 on the high side. Thus the observed event is outside the two-standard-deviation interval.
Because \(p=0.5\), this binomial distribution is symmetric. The probability of the two tails is:
The numerator for one tail is \(1+12+66=79\), so the probability is \(2(79/4096)=158/4096\approx0.0386\), rounded to four decimal places. This agrees with a binomial cumulative probability calculation.
Conclude. Under the model, the probability of getting 2 or fewer successes or 10 or more successes is about 0.0386. Since this is less than 0.05, such a result would be considered unusual. The result is also outside the two-standard-deviation interval.
Worked Example: A High Count in a Right-Skewed Model
Worked Example: A High Count in a Right-Skewed Model
A greenhouse uses a seed variety for which each seed has probability 0.1 of sprouting under a particular set of conditions. Assume 20 seeds are planted independently. Let \(X\) be the number that sprout. Is observing 5 or more sprouts unusual?
State. The model is \(X\sim B(20,0.1)\), where a trial is one planted seed and success is that seed sprouting.
Plan. The setup has two outcomes per seed, a fixed 20 seeds, independent outcomes, and the same probability of sprouting, 0.1, for each seed. These satisfy BINS under the stated assumptions. Since 5 or more is a high count, calculate the upper-tail probability \(P(X\geq5)\). Also compare 5 with the interval \(\mu_X\pm2\sigma_X\).
Do. The mean and standard deviation are:
The interval is \(2-2(1.3416)\) to \(2+2(1.3416)\), approximately \(-0.6832\) to 4.6832. Since a count cannot be negative, the possible counts in the interval are 0 through 4. A count of 5 or more is outside it.
To find the probability, subtract the probability of four or fewer sprouts from 1:
For a check using individual probabilities, \(P(X\leq4)\) is the sum of the probabilities for 0 through 4 sprouts, approximately \(0.1216+0.2702+0.2852+0.1901+0.0898=0.9569\). Using the unrounded probabilities, the exact cumulative probability is about \(0.9568255\), which rounds to \(0.9568\). Thus \(1-0.9568\approx0.0432\), rounded to four decimal places.
Conclude. If each seed independently has a 0.1 probability of sprouting, the chance of at least 5 sprouts among 20 seeds is about 0.0432. Because this is below 0.05, observing 5 or more sprouts would be unusual under the model. It is also beyond the upper endpoint of the two-standard-deviation interval.
Worked Example: When the Guideline and Probability Differ
Worked Example: When the Guideline and Probability Differ
A quality-control process checks 10 items. Suppose each item independently has probability 0.1 of having a particular defect. Let \(X\) count defective items. A batch contains 3 defective items. Is this result unusual?
State. The model is \(X\sim B(10,0.1)\). A trial is checking one item, success means that item has the defect, and \(X\) counts defective items.
Plan. The binomial model assumes two outcomes, a fixed 10 items, independent checks, and a constant defect probability of 0.1. The observed count is high relative to the mean, so we will examine \(P(X\geq3)\). We will also see what the two-standard-deviation guideline indicates.
Do. The mean is \(10(0.1)=1\), and the standard deviation is \(\sqrt{10(0.1)(0.9)}=\sqrt{0.9}\approx0.9487\). The interval is approximately \(1-2(0.9487)\) to \(1+2(0.9487)\), or \(-0.8974\) to 2.8974. Since 3 is above the upper endpoint, the guideline flags it as a potential unusual result.
Now calculate the probability of 3 or more defects:
For a hand check, the probabilities of 0, 1, and 2 defects are approximately 0.3487, 0.3874, and 0.1937. Their sum is 0.9298, so the upper-tail probability is \(1-0.9298=0.0702\), rounded to four decimal places.
Conclude. The count of 3 is outside the two-standard-deviation interval, but the probability of 3 or more defects is about 0.0702, which is greater than 0.05. By the probability guideline, this result is not unusual. This example shows why the two-standard-deviation rule is a screening tool rather than an exact substitute for calculating probability.
What Counts as “More Extreme”?
For a high observed count, “more extreme” generally means the observed count or any larger count. For a low observed count, it means the observed count or any smaller count. For example, if \(X=7\) is a high result, assess \(P(X\geq7)\), not just \(P(X=7)\). If the question asks whether either a very low or very high result would be unusual, specify both tails. In a symmetric distribution, the two tails can be equal; in a skewed distribution, they may not be.
A probability for one exact value and a tail probability answer different questions. The exact probability \(P(X=7)\) is the chance of exactly 7 successes. The upper-tail probability \(P(X\geq7)\) includes 7 and every larger possible count. When judging whether an observed high count is surprising, the tail usually better captures the chance of getting that result or something more extreme in the same direction.
Common Mistakes and AP Exam Tips
- Using only the probability of the exact count. For an observed high result, include the observed count and all larger counts. A full-credit response identifies the event, such as \(P(X\geq5)\), and interprets that probability in context.
- Using the wrong endpoint. “At least 5” includes 5, so it is \(X\geq5\). With binomcdf, find \(1-P(X\leq4)\), not \(1-P(X\leq5)\).
- Calling every count outside the interval unusual without checking. The interval is a quick flag, not a guarantee that the relevant probability is at most 0.05. If the question asks for a probability-based decision, calculate the binomial probability.
- Rounding the endpoints too early. Keep enough decimal places in the endpoints and compare integer counts with the unrounded values. For example, an interval from about 2.536 to 9.464 excludes 2 and 10; it does not include them because they are close to an endpoint.
- Ignoring the direction of the event. A large count calls for an upper-tail probability; a small count calls for a lower-tail probability. If both directions matter, say so and include both tails.
- Skipping the model assumptions. State why a binomial model is reasonable by addressing binary outcomes, independence, fixed \(n\), and constant \(p\). A surprising-looking result does not establish that the model conditions hold.
For a complete AP response, define \(X\), state \(n\) and \(p\), identify the event in probability notation, show the calculation, and compare the result with the stated guideline. Finish with a sentence in context that explains what the probability means and whether the outcome is considered unusual under the model.
Key Takeaway
A binomial count can be assessed using its distance from the mean and its probability under the model. The interval \(\mu_X\pm2\sigma_X\) quickly identifies counts far from the center, but the relevant tail probability gives the more direct basis for deciding whether an event is unusual. Always define the event carefully and interpret the probability in context.
Check Your Understanding
Use the binomial model and the meaning of the event to decide which calculation would assess whether each result is unusual.
- For \(X\sim B(15,0.4)\), find the mean and standard deviation, then write the two-standard-deviation interval.
- A binomial count has an observed value of 8, and the question asks whether this is a high result. Which tail probability should be calculated?
- In a model \(X\sim B(10,0.2)\), what event should be used to assess whether an observed count of 4 is unusually high?
- Why is \(P(X=4)\) not generally the best probability for judging whether an observed high count of 4 is surprising?
- A count falls outside \(\mu_X\pm2\sigma_X\), but the probability of the relevant tail is 0.08. Is it unusual by the 0.05 probability guideline? Explain why the two methods can differ.