Why Trial Count Changes the Count Distribution
A binomial model counts successes across a fixed number of independent trials, each with the same success probability. Holding that probability constant while increasing the number of trials changes the scale of the count: there are more possible successes, so the distribution generally shifts toward larger counts. Its probabilities also change, and an event such as “exactly two successes” can become more or less likely depending on the count and the parameter values.
This comparison is useful whenever the same kind of opportunity is observed a different number of times. For example, a hypothetical insurer might review more policies in one audit cohort than another, with the same probability that any reviewed policy has a specified feature. The model describes the count of such policies; it does not, by itself, describe any financial interest or mortality assumption.
The support is the set of values that can occur with positive probability. When \(0<p<1\), every count from zero through \(n\) is possible. At the boundary \(p=0\), success never occurs, while at \(p=1\), every trial is a success. Those cases must be treated separately when describing support or spread.
What Changes When \(n\) Increases?
For a binomial random variable, the mean and variance are \(E[X]=np\) and \(\operatorname{Var}(X)=np(1-p)\). Write \(q=1-p\) when it is convenient. The standard deviation is \(\sqrt{npq}\). With \(p\) fixed, these formulas show how the numerical summaries respond to a change in the number of trials.
The increase in standard deviation is slower than the increase in the mean when \(0<p<1\). In particular, for \(p>0\), the coefficient of variation (standard deviation divided by mean) is $$ \frac{\sqrt{np(1-p)}}{np} =\sqrt{\frac{1-p}{np}}. $$ For fixed \(0<p<1\), this relative spread decreases as \(n\) increases. If \(p=1\), the relative spread is zero for every \(n\). If \(p=0\), the mean is zero and the coefficient of variation is undefined.
A decreasing relative spread does not mean that the standard deviation decreases. For interior \(p\), the standard deviation increases, but the mean increases faster in relative terms. Also, increasing \(n\) does not guarantee that every particular probability increases. The location of the selected count matters: a fixed small count may become less likely as the distribution shifts toward larger counts.
Why the Mean and Variance Scale This Way
A short indicator-variable argument makes the dependence on \(n\) clear. Let \(I_j=1\) if trial \(j\) is a success and \(I_j=0\) otherwise, for \(j=1,\ldots,n\). Then \(X=I_1+\cdots+I_n\). Each indicator has mean \(p\) and variance \(p(1-p)\). Independence makes the variance of the sum equal to the sum of the variances.
Thus, adding one independent trial with the same success probability adds \(p\) to the mean and \(p(1-p)\) to the variance. This explains the linear changes in the two quantities. Taking the square root of the variance gives the standard deviation, so its scale is proportional to \(\sqrt n\), not \(n\).
Comparing Selected Probabilities
To compare probabilities at different trial counts, use the binomial formula separately for each \(n\). The parameters must be clear: the success probability remains fixed, but the possible count range and the probability weights change. A useful comparison may include a fixed count, such as exactly two successes, or an event whose threshold is stated in words, such as at least one success.
Worked Example: Comparing a Fixed Count and “At Least One”
A hypothetical quality team checks independent items. Each item has probability \(p=0.20\) of having a specified defect. Compare cohorts of \(n=5\) and \(n=10\) items. Find the means and variances, then compare the probability of exactly two defects and the probability of at least one defect. There is no interest or mortality assumption in this binomial count model.
Method and parameters. The defect indicators are modeled as independent Bernoulli trials with common probability \(p=0.20\). Thus \(X_5\sim\operatorname{Bin}(5,0.20)\) and \(X_{10}\sim\operatorname{Bin}(10,0.20)\). In both cases, \(q=1-p=0.80\).
For \(n=5\), the mean is \(5(0.20)=1\), and the variance is \(5(0.20)(0.80)=0.80\). For \(n=10\), the mean is \(10(0.20)=2\), and the variance is \(10(0.20)(0.80)=1.60\). The supports are \(\{0,1,2,3,4,5\}\) and \(\{0,1,\ldots,10\}\), respectively.
For exactly two defects in five items, $$ P(X_5=2)=\binom52(0.20)^2(0.80)^3 =10(0.04)(0.512)=0.2048. $$ For exactly two in ten items, $$ P(X_{10}=2)=\binom{10}{2}(0.20)^2(0.80)^8 =45(0.04)(0.16777216)=0.301989888. $$ Rounded to four decimal places, these probabilities are \(0.2048\) and \(0.3020\).
For at least one defect, use the complement of no defects. For five items, \(P(X_5\ge1)=1-(0.80)^5=1-0.32768=0.67232\), or \(0.6723\) to four decimals. For ten items, \(P(X_{10}\ge1)=1-(0.80)^{10}=1-0.1073741824=0.8926258176\), or \(0.8926\) to four decimals.
Check. Both the mean and the probability of at least one defect are larger in the larger cohort. The probability of exactly two also happens to be larger here; that outcome is not guaranteed for every fixed count. A common mistake is to assume that every exact-count probability must rise when \(n\) rises. The support and probability weights change together, so each selected event must be calculated. The model has no interest or mortality component.
Worked Example: A Fixed Small Count Can Become Less Likely
For a hypothetical independent screening process, each screened record has probability \(p=0.50\) of meeting a criterion. Compare \(n=4\) records with \(n=8\). Find the mean and variance in each cohort, then compare the probability of exactly two qualifying records and the probability of at most two. There is no interest or mortality assumption.
Method and parameters. Use \(X_4\sim\operatorname{Bin}(4,0.50)\) and \(X_8\sim\operatorname{Bin}(8,0.50)\). For both distributions, \(q=0.50\). The means are \(4(0.50)=2\) and \(8(0.50)=4\); the variances are \(4(0.50)(0.50)=1\) and \(8(0.50)(0.50)=2\). The supports are \(\{0,1,2,3,4\}\) and \(\{0,1,\ldots,8\}\).
For exactly two, the probabilities are $$ P(X_4=2)=\binom42(0.50)^2(0.50)^2 =\frac{6}{16}=0.375, $$ and $$ P(X_8=2)=\binom82(0.50)^2(0.50)^6 =\frac{28}{256}=0.109375. $$ To four decimals these are \(0.3750\) and \(0.1094\).
For at most two in the four-record cohort, add the probabilities for counts 0, 1, and 2: $$ P(X_4\le2)=\frac{\binom40+\binom41+\binom42}{2^4} =\frac{1+4+6}{16} =\frac{11}{16}=0.6875. $$ For the eight-record cohort, $$ P(X_8\le2)=\frac{\binom80+\binom81+\binom82}{2^8} =\frac{1+8+28}{256} =\frac{37}{256}=0.14453125. $$ Rounded to four decimals, this is \(0.1445\).
Check. The mean moves from 2 to 4, so a count of at most two is much less likely in the larger cohort. The calculation also verifies that “exactly two” is less likely here despite the larger support. A common mistake is to interpret a larger support as evidence that the probability of every count rises. The probabilities across the full support must sum to one, and the distribution’s location changes as the trial count increases.
Boundary Cases and Interpretation
The formulas \(np\) and \(np(1-p)\) remain valid at \(p=0\) and \(p=1\), but their interpretation is especially simple. If \(p=0\), then \(X=0\) with probability one for every \(n\). If \(p=1\), then \(X=n\) with probability one, so increasing \(n\) moves the certain count upward. In both cases, the variance is zero because there is no randomness in the count.
For \(0<p<1\), the support for a larger trial count contains all the counts from the smaller support and additional larger counts. The expected count rises by \(p\) for each added trial. The standard deviation rises as the square root of \(n\), while the mean rises in proportion to \(n\). These are statements about summaries, not a promise that every specific probability changes in the same direction.
When comparing cohorts, keep the event’s meaning in view. “Exactly two” is a fixed count even though the cohort size changes. “At most 20% of the items” is a changing count threshold when \(n\) changes, and must be translated into an integer bound separately for each cohort. Comparing such events without adjusting the threshold can answer a different question than intended.
Check Your Understanding
For each problem, identify the binomial parameters before comparing the requested quantities. Probabilities in the answers are rounded to four decimal places where needed.
- A hypothetical claims team classifies each independently reviewed file as requiring follow-up with probability \(0.10\). Compare \(n=6\) and \(n=12\). Give the mean and variance for each model and the probability of no follow-up files in each cohort. (Answer: Means \(0.6,1.2\); variances \(0.54,1.08\); no-follow-up probabilities \(0.5314,0.2824\).)
- A process has independent trials with success probability \(0.25\). For \(n=4\) and \(n=8\), compare the mean and variance, then calculate the probability of exactly one success. (Answer: Means \(1,2\); variances \(0.75,1.50\); probabilities \(0.4219,0.2670\).)
- A hypothetical data audit marks each record with probability \(0.40\), independently. Compare \(n=5\) and \(n=10\), giving the support, mean, variance, and probability of at least one marked record. (Answer: Supports \(\{0,\ldots,5\}\) and \(\{0,\ldots,10\}\); means \(2,4\); variances \(1.20,2.40\); probabilities \(0.9222,0.9939533824\), or \(0.9222,0.9940\) to four decimals.)
- For \(n=7\), determine the support, mean, variance, and standard deviation when \(p=0\), and then when \(p=1\). Explain what happens to the count if \(n\) is increased to 9. (Answer: At \(p=0\), support \(\{0\}\), mean \(0\), variance and standard deviation \(0\), and the count stays \(0\). At \(p=1\), support \(\{7\}\), mean \(7\), variance and standard deviation \(0\); with \(n=9\), the count is certainly \(9\).)
- For fixed \(0<p<1\), \(n\) changes from 9 to 36. By what factor do the mean, variance, and standard deviation change? State what happens to the coefficient of variation. (Answer: The mean and variance each multiply by \(4\); the standard deviation multiplies by \(2\); the coefficient of variation is halved.)