Types of Data Distributions

Why This Matters

When you're analyzing data, the distribution tells you everything about what statistical tools you can use and what conclusions you can draw. You're being tested on your ability to recognize distribution shapes, understand their parameters, and—most critically—know when to apply each one. A normal distribution lets you use z-scores and standard confidence intervals; a Poisson distribution handles count data; a t-distribution saves you when sample sizes are small. Choosing the wrong distribution leads to invalid results.

Think of distributions as the foundation of statistical inference. Every hypothesis test, every confidence interval, and every predictive model assumes some underlying distribution. The concepts you need to master include probability density functions, parameters and their meanings, when distributions approximate each other, and real-world applications. Don't just memorize shapes—know what each distribution models and why you'd reach for it in a given scenario.

---

Symmetric Distributions: The Workhorses of Statistics

These distributions are balanced around their center, making them mathematically convenient and widely applicable. Symmetry means the mean, median, and mode coincide, which simplifies calculations and interpretation.

Normal Distribution

• Bell-shaped and symmetric around the mean ($\mu$)—approximately 68% of data falls within one standard deviation ($\sigma$), 95% within two
• Defined by just two parameters: $\mu$ (center) and $\sigma$ (spread), making it easy to standardize using z-scores
• Central Limit Theorem connection—the sampling distribution of means approaches normal as sample size increases, regardless of the original distribution

Uniform Distribution

• All outcomes equally likely within a defined range—probability is constant across the interval
• Two forms exist: discrete (rolling a fair die) and continuous (random number generators between 0 and 1)
• Characterized by minimum and maximum values—the probability density function is flat, making expected value simply the midpoint

Student's t-Distribution

• Heavier tails than normal—accounts for extra uncertainty when sample sizes are small and population $\sigma$ is unknown
• Degrees of freedom (df) control the shape; lower df means fatter tails and more conservative estimates
• Converges to normal as df increases—with $n > 30$, the difference becomes negligible for most purposes

---

Count and Event Distributions: Modeling Discrete Outcomes

These distributions handle scenarios where you're counting occurrences—successes in trials, events over time, or items in categories. The key is recognizing whether you have a fixed number of trials or an open-ended count.

Binomial Distribution

• Models successes in $n$ independent trials—each trial has only two outcomes (success/failure) with constant probability $p$
• Parameters are $n$ (trials) and $p$ (success probability)—mean is $np$ and variance is $np(1-p)$
• Applications include quality control and surveys—any scenario asking "how many out of n?" is likely binomial

Poisson Distribution

• Counts events in a fixed interval of time or space—characterized by rate parameter $\lambda$ (average occurrences)
• Assumes independence and constant rate—events don't cluster or influence each other
• Best for rare events: call center volumes, website clicks per minute, defects per unit—when $n$ is large and $p$ is small, binomial approximates Poisson

---

Time and Continuous Processes: Modeling Duration and Rates

When you're measuring how long until something happens rather than how many times it happens, these continuous distributions apply. They're essential for reliability engineering, survival analysis, and queuing systems.

Exponential Distribution

• Models time between events in a Poisson process—if arrivals follow Poisson, wait times follow exponential
• Single parameter $\lambda$ (rate), with mean wait time equal to $1/\lambda$
• Memoryless property—the probability of an event in the next interval doesn't depend on elapsed time; past waiting doesn't change future odds

---

Hypothesis Testing Distributions: The Statistical Inference Toolkit

These specialized distributions emerge when you're testing claims about data. They're derived from combinations of other distributions and are essential for formal inference.

Chi-Square Distribution

• Sum of squared standard normal variables—used when analyzing categorical data and testing goodness-of-fit
• Degrees of freedom ($k$) determine the shape; distribution is right-skewed but approaches normal as $k$ increases
• Key applications: testing independence in contingency tables, comparing observed vs. expected frequencies

F-Distribution

• Ratio of two chi-square distributions—specifically designed for comparing variances between groups
• Two degrees of freedom parameters: one for numerator, one for denominator—always positive and right-skewed
• ANOVA workhorse—whenever you're testing whether group means differ, the F-statistic follows this distribution

---

Non-Symmetric Distributions: When Data Doesn't Play Nice

Real-world data often violates the assumption of symmetry. Recognizing asymmetry is crucial because it affects which measures of center and statistical tests are appropriate.

Skewed Distributions

• Asymmetric with a longer tail on one side—positive skew (right tail) or negative skew (left tail)
• Mean gets pulled toward the tail—in right-skewed data, mean > median > mode; this ordering reverses for left skew
• Impacts statistical validity—many tests assume normality, so skewed data may require transformation or non-parametric methods

Multimodal Distributions

• Two or more distinct peaks (modes)—signals multiple underlying groups or processes in your data
• Often indicates mixed populations—heights of adults (male/female peaks) or customer segments with different behaviors
• Standard techniques may fail—mean becomes meaningless; requires identifying and analyzing subgroups separately

---

Quick Reference Table

---

Self-Check Questions

1. Which two distributions are mathematically related as "two sides of the same coin," with one modeling counts and the other modeling wait times?

2. You have survey data from 200 respondents answering yes/no questions. Which distribution models the number of "yes" responses, and what parameters would you need?

3. Compare and contrast the normal distribution and t-distribution: when would you use each, and what happens to the t-distribution as sample size increases?

4. A dataset of household incomes shows mean significantly higher than median. What type of distribution is this, and why would using the mean be misleading?

5. An FRQ asks you to test whether three treatment groups have different average outcomes. Which distribution would your test statistic follow, and why is it appropriate for this comparison?