Discrete vs Continuous Distributions: How to Choose

A discrete random variable takes specific separable values — counts, integers, categories. Examples: number of customer arrivals per hour (0, 1, 2, ...), number of defective products in a batch, number of heads in 10 coin flips. The variable can only assume specific values; there is no continuum. A continuous random variable takes any value within a range. Examples: height in cm (170.5, 170.51, 170.515, ...), time to complete a task, temperature, financial returns. Between any two values, infinitely many intermediate values exist. This distinction drives the probability function form. Discrete distributions have a probability mass function (PMF) — P(X = k) gives the probability that the variable equals each specific value. Continuous distributions have a probability density function (PDF) — f(x) gives density, and probabilities are computed as areas under the curve over intervals: P(a < X < b) = integral of f(x) from a to b. P(X = exactly some specific value) = 0 for continuous distributions (the area at a single point is zero). Mixed distributions exist (e.g., a variable that is sometimes zero and sometimes continuous) but are less common in introductory statistics. The discrete/continuous distinction handles the vast majority of practical cases.

Binomial. Number of successes in n independent trials with constant probability p. Parameters: n, p. Mean = np. Variance = np(1-p). Use for fixed-sample-size scenarios with binary outcomes: defects in a batch, conversions in a sample, heads in coin flips. Poisson. Number of events in a fixed interval at constant rate. Parameter: lambda. Mean = Variance = lambda. Use for arrival, defect, or count data where events occur independently at a constant rate: customer arrivals per hour, defects per page, calls per minute. Approximates binomial when n is large and p is small (set lambda = np). Geometric. Number of trials until the first success. Parameter: p. Mean = 1/p. Use for "time to first success" scenarios with constant probability: number of customer calls before the first sale, number of submissions before acceptance. Negative Binomial. Number of trials until r successes (or number of failures before r successes). Parameters: r, p. Use when extending geometric to multiple required successes, OR as overdispersion-tolerant alternative to Poisson when variance >> mean.

Normal (Gaussian). Symmetric bell curve. Parameters: mean (mu), SD (sigma). Use for heights, weights, test scores, measurement errors. Most natural phenomena approximate normal, and Central Limit Theorem produces normal sampling distributions for sample means. Exponential. Time between events in a Poisson process. Parameter: rate (lambda). Mean = 1/lambda. Variance = 1/lambda^2. Use for time-between-arrivals at a service counter, time between failures of components. Memoryless property: P(X > t + s | X > s) = P(X > t) — the wait time does not depend on how long you have already waited. Log-normal. The log of the variable is normally distributed. Use for variables that are always positive and right-skewed: income, stock prices, file sizes, sound intensity. If Y is log-normal, log(Y) is normal — many transformations exploit this. Uniform. Equal probability across a range [a, b]. Parameter: a, b. Mean = (a+b)/2. Variance = (b-a)^2 / 12. Use for random number generation, scenarios with no prior preference for any value in a range. Limited real-world use but appears in simulations and Bayesian priors.

Step 1: Is the variable a count or a measurement? Counts = discrete. Measurements = continuous (or treated as continuous for large enough scales). Step 2 (discrete): How are events generated? Fixed trials with binary outcomes → Binomial. Events in fixed interval at constant rate → Poisson. Time to first success → Geometric. Time to r successes OR overdispersed counts → Negative Binomial. Step 3 (continuous): What is the shape and support? Symmetric, unbounded → Normal. Positive, right-skewed → Log-normal. Memoryless time-between-events → Exponential. Equal probability over a range → Uniform. Step 4: Run goodness-of-fit checks. Q-Q plots compare observed to theoretical quantiles. Statistical tests (Kolmogorov-Smirnov, Anderson-Darling, Chi-square) test fit. Visual inspection often reveals deviations the eye sees better than tests. Step 5: If no standard distribution fits well, consider transformations (log, sqrt, Box-Cox) or non-parametric methods that do not require distributional assumptions. Permutation tests and bootstrapping are robust alternatives.

Snap a photo of any data sample and StatsIQ identifies the best-fitting distribution from a library of 15+ common distributions, runs goodness-of-fit tests, produces Q-Q plots, and recommends transformations for poor fits. For exam prep, the app generates problems at all complexity levels including distribution-identification questions. StatsIQ also handles applied problems: given the chosen distribution, compute probabilities, percentiles, and sample-size requirements for various inferential goals. This content is for educational purposes only.