Probability Distributions: The Complete Guide With Decision Tree

A probability distribution is a mathematical function describing the probability of each possible outcome of a random variable. Discrete distributions (binomial, Poisson, geometric) describe outcomes that take countable values such as integers; continuous distributions (normal, t, chi-square, F, uniform, exponential) describe outcomes on a continuum and use probability density functions where probability is measured by area under the curve. Every distribution is defined by parameters (e.g., the binomial by n and p, the normal by μ and σ) and has known formulas for the mean, variance, and key probabilities. Choosing the right distribution comes down to four questions: discrete or continuous, how many possible outcomes per trial, are trials independent, and what is the underlying generative process. The Central Limit Theorem (CLT) provides a powerful unifying result: regardless of the underlying distribution, the sample mean of a sufficiently large sample is approximately normally distributed — which is why the normal distribution shows up everywhere in inferential statistics.

The binomial distribution models the number of successes in n independent Bernoulli trials, each with success probability p. Use it when each trial has exactly two outcomes (success/failure), the number of trials is fixed, and trials are independent. P(X = k) = C(n, k) × p^k × (1 − p)^(n − k) Mean = np; Variance = np(1 − p). Worked Example. A factory produces parts where 8% are defective. In a random sample of 20 parts, what is the probability that exactly 3 are defective? P(X = 3) = C(20, 3) × 0.08^3 × 0.92^17 C(20, 3) = 20! / (3! × 17!) = 1,140 P(X = 3) = 1,140 × 0.000512 × 0.2452 = 0.1431 Probability of exactly 3 defective parts is about 14.3%. Expected number = 20 × 0.08 = 1.6 defects. For "at least 3" or "at most 3" calculations, sum the relevant probability mass values. Modern practice uses statistical software (Python scipy.stats.binom, R dbinom, Excel BINOM.DIST); the formula is rarely computed by hand for n > 5.

The Poisson distribution models the number of events in a fixed interval of time or space when events occur independently and at a constant average rate λ. P(X = k) = (λ^k × e^(−λ)) / k! Mean = λ; Variance = λ. Worked Example. A call center receives an average of 6 calls per hour. What is the probability of receiving exactly 4 calls in a given hour? P(X = 4) = (6^4 × e^(−6)) / 4! = (1,296 × 0.002479) / 24 = 3.213 / 24 = 0.1339 Probability of exactly 4 calls is about 13.4%. P(X ≥ 4) requires summing P(X = 4) + P(X = 5) + ... — done with software in practice. The Poisson is also a limiting form of the binomial: when n is very large and p is very small, the binomial approaches Poisson with λ = np. This approximation is useful for rare-event modeling — defect rates in millions of parts, hospital readmission rates, equipment failure intervals, accident counts on highway segments. Key distinguishing feature: Poisson has Mean = Variance. If sample data shows variance much larger than mean (over-dispersion), Poisson is the wrong model — switch to negative binomial which allows variance > mean.

The normal distribution is the most important continuous distribution in statistics. It is symmetric, bell-shaped, and completely defined by its mean μ and standard deviation σ. The standard normal Z has μ = 0, σ = 1. Key properties (the empirical rule): - 68% of values within ±1σ - 95% of values within ±1.96σ (rounded to ±2σ) - 99.7% of values within ±3σ Probabilities are computed by converting to Z-scores: Z = (X − μ) / σ, then looking up P(Z < z) in a standard normal table or computing with software (Python scipy.stats.norm.cdf, R pnorm, Excel NORM.DIST). Worked Example. SAT scores are approximately normal with μ = 1,050 and σ = 200. What is the probability a randomly selected student scores above 1,300? Z = (1,300 − 1,050) / 200 = 250 / 200 = 1.25 P(Z > 1.25) = 1 − P(Z < 1.25) = 1 − 0.8944 = 0.1056 About 10.6% of students score above 1,300. The normal distribution is the asymptotic limit of many statistics by the Central Limit Theorem. Sample means, sample proportions, and many estimators become approximately normal as sample size grows, even when the underlying data is not normal. This is why z-tests and t-tests are robust at large sample sizes regardless of data distribution.

Three more continuous distributions, each tied to specific test families. Student's t-distribution. Similar to normal but with heavier tails; used when σ is unknown and estimated by s. Defined by degrees of freedom (df). As df grows, t approaches Z. For df = 24 (n = 25), the two-tailed critical value at alpha = 0.05 is t = 2.064 (vs Z = 1.96 for the normal). Chi-square distribution. Sum of squared standard normals; right-skewed; defined by df. Used in chi-square tests of independence, goodness-of-fit, and variance tests. For df = 5, the upper 5% critical value is χ² = 11.07. F-distribution. Ratio of two scaled chi-squares; right-skewed; defined by two df parameters (numerator and denominator). Used in ANOVA and regression overall-fit tests. For df = (3, 30), the upper 5% critical value is F = 2.92. These three are the workhorses of inferential statistics. Their critical values are tabulated in every statistics textbook (and computed instantly by software). Relationships among them: a t-statistic is approximately Z for large df; t² with df_t = F with df = (1, df_t); a chi-square divided by its df becomes an F-statistic with infinite denominator df. Other continuous distributions worth knowing: Uniform (equal probability over an interval; used in random-number generation and Monte Carlo); Exponential (time between Poisson events; right-skewed; memoryless property); Gamma (generalization of exponential; sum of exponentials); Beta (bounded between 0 and 1; used for proportions, prior distributions in Bayesian inference); Log-normal (right-skewed, takes only positive values; used for income, stock prices, response times).

A practical decision tree for selecting the right distribution from a problem description. Is the random variable discrete or continuous? Discrete: - Two outcomes per trial, fixed n, independent → Binomial - Two outcomes per trial, n until first success → Geometric - Two outcomes per trial, n until k-th success → Negative Binomial - Counts in fixed interval, constant rate → Poisson - Sampling without replacement from finite population → Hypergeometric - Equal probability over a fixed range of integers → Discrete Uniform Continuous: - Symmetric, bell-shaped, real-valued → Normal - Like normal but heavier tails (σ unknown) → t-distribution - Sum of squared normals, right-skewed, non-negative → Chi-square - Ratio of variances → F - Equal probability over an interval → Uniform - Time between Poisson events, right-skewed → Exponential - Generalization of exponential → Gamma - Bounded between 0 and 1 → Beta - Right-skewed positive (income, response times) → Log-normal Branch 1 (discrete vs continuous) is usually obvious from the problem context. Branch 2 (which discrete/continuous) requires identifying the generative process. Common rookie mistake: using normal for skewed or bounded variables (income, response times, proportions). Better: lognormal for income, gamma for response times, beta for proportions.

The Central Limit Theorem states that the sampling distribution of the sample mean approaches a normal distribution as sample size grows, regardless of the underlying population distribution. Specifically: for a population with mean μ and standard deviation σ, the sample mean x̄ from a sample of size n has approximate distribution N(μ, σ²/n) for sufficiently large n. The conventional rule of thumb is n ≥ 30, though strongly skewed populations may require larger samples. Worked Example. Daily call durations at a call center have mean μ = 6 minutes and σ = 4 minutes (right-skewed; not normally distributed). What is the probability that a sample of 50 calls has mean duration above 7 minutes? Sampling distribution of x̄: N(6, 4²/50) = N(6, 0.32) → SE = √0.32 = 0.566 Z = (7 − 6) / 0.566 = 1.768 P(x̄ > 7) = P(Z > 1.768) = 1 − 0.9615 = 0.0385 About 3.9% chance of observing a sample mean above 7 minutes, even though individual calls are not normally distributed. CLT is what enables most of inferential statistics. The t-test, z-test, and ANOVA all rely on the sampling distribution of estimators being approximately normal. CLT also justifies bootstrap methods (which resample and compute means) — the bootstrap distribution of the mean converges to normal even for non-normal data.

Probability distribution problems are core to introductory and intermediate statistics, the AP Statistics curriculum, and many engineering, finance, and biostatistics applications. The decision tree (which distribution to use) and parameter identification (which formula gets which numbers) trip up most students. Snap a photo of any probability problem and StatsIQ identifies the correct distribution, plugs in the parameters, computes the requested probability, and visualizes the relevant area under the curve. For Central Limit Theorem problems, StatsIQ produces both the underlying population distribution and the sampling distribution side by side. This content is for educational purposes only and does not constitute statistical advice.

Five errors recur. First, applying the normal distribution to skewed or bounded data — income, response times, proportions all need different distributions. Second, confusing P(X = k) with P(X ≤ k) or P(X ≥ k). Discrete distributions have specific values; continuous distributions only assign probabilities to ranges. Third, forgetting the n condition for Central Limit Theorem. Small samples from skewed populations remain skewed even at sample-mean level. Fourth, treating the sample standard deviation s as if it were the population σ when computing z-scores. Use t-distribution when σ is estimated by s. Fifth, applying Poisson when the variance differs sharply from the mean. Variance >> mean (over-dispersion) signals negative binomial; variance << mean signals different model entirely.