Statistical Methods — Differentiate descriptive, inferential, parametric, and nonparametric statistics - Eppp Knowledge

SECTION 1

Historical Context & Motivation

The evolution of statistical methods is deeply intertwined with the growth of the behavioral and social sciences. Early researchers in psychology and public health recognized that raw observations alone—case studies, anecdotal reports, and unstructured clinical impressions—were insufficient for drawing reliable conclusions about human behavior and mental health outcomes. The challenge was twofold: first, how to organize and summarize large bodies of data in meaningful ways, and second, how to generalize findings from small samples to broader populations. This dual need gave rise to the two great branches of statistics—descriptive and inferential—and the subsequent distinction between parametric and nonparametric approaches emerged as researchers confronted the reality that not all data conform to the neat assumptions of classical probability theory.

1805 • Legendre & Least Squares

Adrien-Marie Legendre publishes the method of least squares, laying the mathematical groundwork for regression analysis and descriptive curve fitting that would later become central to behavioral science.

1908 • Gosset's t-Test

William Sealy Gosset, writing under the pseudonym 'Student,' develops the t-distribution for small-sample inference—a parametric method that would become a cornerstone of psychological research.

1925 • Fisher's Statistical Methods for Research Workers

Ronald Fisher formalizes ANOVA and the logic of hypothesis testing, providing behavioral researchers with rigorous parametric tools for comparing group means in experimental designs.

1945 • Rise of Nonparametric Methods

Frank Wilcoxon introduces rank-based tests, and subsequent work by Mann, Whitney, and Kruskal provides distribution-free alternatives—critical for behavioral data that often violate normality assumptions.

1960s–Present • Integration into Clinical Science

With the growth of evidence-based practice, clinicians in behavioral health are expected to critically evaluate research using both parametric and nonparametric methods, making statistical literacy a professional competency tested on the EPPP.

The central question that this lesson addresses is one that every behavioral health practitioner and researcher must answer repeatedly: Given a particular research question, data type, and set of assumptions, which statistical method is appropriate? Choosing incorrectly can lead to invalid conclusions, Type I or Type II errors, and ultimately, clinical decisions that fail to serve clients. By understanding the distinctions among descriptive, inferential, parametric, and nonparametric statistics, you will be equipped to evaluate the quality of published research and to select appropriate analytic strategies in your own work.

SECTION 2

Core Principles & Definitions

Statistical methods can be classified along two independent dimensions. The first dimension concerns purpose—whether the goal is to describe the data at hand or to make inferences beyond them. The second dimension concerns assumptions—whether the method requires that data follow a specific probability distribution (parametric) or makes no such requirement (nonparametric). Understanding these two dimensions as independent axes is crucial because they combine to produce four categories, each with distinct applications in behavioral health research.

Descriptive Statistics

Methods that summarize, organize, and present data without attempting to generalize beyond the sample. Common measures include the mean, median, mode, standard deviation, range, and frequency distributions. These tools answer the question: What does this data look like?

Inferential Statistics

Methods that use sample data to draw conclusions about a population. Techniques include hypothesis testing (t-tests, ANOVA, chi-square) and confidence intervals. Inferential statistics answer: Can we generalize these findings to the broader population?

Parametric Statistics

Inferential methods that assume the data come from a known probability distribution (usually normal), require interval or ratio level measurement, and assume homogeneity of variance. Examples: independent-samples t-test, one-way ANOVA, Pearson's r.

Nonparametric Statistics

Inferential methods that do not assume a specific distribution shape. They work with ordinal or nominal data and are robust to outliers and non-normal distributions. Examples: Mann-Whitney U, Wilcoxon signed-rank test, Kruskal-Wallis H, chi-square test, Spearman's rho.

✦

KEY TAKEAWAY

Think of the descriptive vs. inferential distinction like the difference between taking a photograph of a room (describing exactly what is there) and using that photograph to make predictions about what other similar rooms look like (inferring beyond the observed). Parametric vs. nonparametric, meanwhile, is like choosing between a custom-tailored suit (parametric—more powerful but requires a precise fit to assumptions) and an adjustable garment (nonparametric—more flexible but sometimes less precise). In behavioral health research, the best statistical tool depends on both your question and the nature of your data.

SECTION 3

Visual Explanation — The Statistical Decision Framework

This flowchart illustrates the two-dimensional decision framework for selecting statistical methods. The first branch differentiates descriptive from inferential goals; the second branch subdivides inferential statistics into parametric and nonparametric methods based on whether distributional assumptions are met.

The flowchart above captures the essential logic of statistical method selection that you will encounter repeatedly on the EPPP and in clinical research. Notice that the descriptive–inferential distinction is about purpose (are you summarizing or generalizing?), whereas the parametric–nonparametric distinction is about assumptions and data characteristics (does your data meet the requirements of classical probability models?). A common error among students is to conflate these two dimensions—for instance, treating 'descriptive' as synonymous with 'nonparametric.' In reality, descriptive statistics precede any inferential analysis and are applied regardless of whether a parametric or nonparametric test will follow.

SECTION 4

Mathematical Framework — Key Formulas Across Categories

While the EPPP does not require extensive hand computation, understanding the formulas underlying each category clarifies what each statistic actually measures and why certain assumptions matter. Below are representative formulas from each category, with annotations explaining the behavioral health relevance of each.

Descriptive Statistics

SAMPLE MEAN

X̄ = (Σ Xᵢ) / n

Where X̄ is the sample mean, Σ Xᵢ is the sum of all observed values, and n is the number of observations. This is the most common measure of central tendency for interval/ratio data.

SAMPLE STANDARD DEVIATION

s = √[ Σ(Xᵢ − X̄)² / (n − 1) ]

The standard deviation quantifies the average distance of scores from the mean. Note the use of n − 1 (Bessel's correction) in the denominator, which provides an unbiased estimate of the population standard deviation from sample data.

Parametric Inferential Statistics

INDEPENDENT-SAMPLES t-TEST

t = (X̄₁ − X̄₂) / √[ (s₁²/n₁) + (s₂²/n₂) ]

This formula computes the t-statistic for comparing two independent group means. X̄₁ and X̄₂ are the sample means, s₁² and s₂² are the sample variances, and n₁ and n₂ are the sample sizes. Assumptions include normality, independence, and homogeneity of variance.

Nonparametric Inferential Statistics

CHI-SQUARE TEST OF INDEPENDENCE

χ² = Σ [ (O − E)² / E ]

Where O represents observed frequencies and E represents expected frequencies under the null hypothesis. The chi-square test evaluates whether two categorical variables are independent—for example, whether diagnosis type is associated with treatment completion. No distributional assumptions about the underlying variables are required.

💡

EPPP TIP

You are unlikely to be asked to compute a full t-test or chi-square on the exam, but you will be expected to identify which test is appropriate given a research scenario, interpret the meaning of results (e.g., 'a significant chi-square means the variables are not independent'), and recognize when assumptions have been violated.

SECTION 5

Detailed Classification — Matching Tests to Research Scenarios

A critical skill for the EPPP is the ability to match specific statistical tests to the research scenarios that call for them. The classification below organizes the most commonly tested parametric and nonparametric methods by the number of groups being compared and the level of measurement of the dependent variable. For each parametric test, there is typically a nonparametric equivalent that serves a similar purpose but relaxes distributional assumptions.

This reference chart pairs each common parametric test with its nonparametric equivalent. Note that chi-square has no parametric counterpart because it is inherently designed for nominal (categorical) data. This table is one of the highest-yield memorization tools for the EPPP.

When deciding between a parametric test and its nonparametric counterpart, the primary consideration is whether the assumptions of the parametric test are met. In behavioral health research, these assumptions are frequently violated—clinical outcome measures may be skewed (e.g., symptom severity scores that cluster near zero in a well-functioning population), samples may be small (e.g., rare diagnoses), and data may be ordinal rather than truly interval-level (e.g., Likert-scale ratings). In such circumstances, nonparametric tests provide valid alternatives, though they generally offer less statistical power—meaning they are less likely to detect a true effect when one exists.

SECTION 6

Worked Example — Selecting the Appropriate Test

Consider the following scenario, which represents a typical EPPP-style question: A clinical psychologist wants to compare depression scores (measured by the BDI-II, a continuous scale) between a CBT group and a psychodynamic therapy group. Each group has 15 participants, and preliminary analysis suggests the depression scores in the psychodynamic group are substantially skewed. Which statistical test should be used?

Choosing Between an Independent t-Test and the Mann-Whitney U

1
Step 1 — Identify the Research Question

The researcher wants to compare means (or central tendencies) between two independent groups on a continuous variable. This is an inferential question—the goal is to generalize findings beyond these 30 participants to the broader population of clients receiving CBT or psychodynamic therapy.

Goal: Inferential comparison of two independent groups
2
Step 2 — Determine the Level of Measurement

The BDI-II produces scores on a continuous scale (0–63), which is generally treated as interval-level data. This level of measurement is compatible with both parametric and nonparametric approaches.

Measurement level: Interval (continuous)
3
Step 3 — Check Parametric Assumptions

The independent-samples t-test requires: (a) the dependent variable is approximately normally distributed within each group, (b) the two groups have roughly equal variances (homogeneity of variance), and (c) observations are independent. Here, the psychodynamic group's scores are substantially skewed, violating the normality assumption. With only n = 15 per group, the Central Limit Theorem provides limited protection, and the skew is described as substantial.

Normality assumption VIOLATED
4
Step 4 — Select the Appropriate Test

Because the normality assumption is violated and the sample size is small, the parametric t-test may produce inaccurate p-values. The nonparametric alternative for comparing two independent groups is the Mann-Whitney U test, which ranks all scores across both groups and compares rank sums rather than means. It does not require normality.

Correct choice: Mann-Whitney U test
5
Step 5 — Interpret and Report

If the Mann-Whitney U yields a significant result (e.g., U = 52, p = .03), we conclude that the two treatment groups differ significantly in their depression score distributions. Unlike the t-test, the Mann-Whitney U does not directly compare means; it compares the rank orderings. A significant result indicates that one group tends to score higher than the other, which in clinical terms may suggest differential treatment effectiveness.

Report: U = 52, p = .03, indicating a significant difference in depression outcomes between groups.

SECTION 7

Strengths, Limitations, and Common Pitfalls

Category	Strengths	Limitations
Descriptive Statistics	Easy to compute and interpret; always applicable; provide essential data summaries; foundation for all subsequent analysis	Cannot test hypotheses or establish significance; can be misleading without context (e.g., mean of a bimodal distribution); do not support causal inference
Inferential Statistics (general)	Allow generalization from sample to population; provide probabilistic framework for hypothesis testing; support evidence-based clinical decisions	Susceptible to Type I and Type II errors; conclusions depend on sampling quality; require careful assumption checking
Parametric Tests	Greater statistical power when assumptions are met; more precise parameter estimates; wide range of complex models available (e.g., factorial ANOVA, multiple regression)	Require normality, interval/ratio data, and homogeneity of variance; results can be unreliable when assumptions are violated, particularly with small samples
Nonparametric Tests	No distributional assumptions; can handle ordinal and nominal data; robust to outliers; appropriate for small samples and skewed distributions	Lower statistical power than parametric equivalents when assumptions are actually met; fewer complex modeling options; may discard information by converting scores to ranks

Comparative strengths and limitations across the four statistical categories

✦

KEY TAKEAWAY

Choosing between parametric and nonparametric methods is not about one being 'better'—it is about fit. Think of it like choosing between a scalpel and a Swiss Army knife. The scalpel (parametric test) is more precise and efficient for the task it was designed for, but only if the conditions are right—you need the correct tissue, the right angle, and steady hands. If conditions are unpredictable (non-normal data, ordinal scales, outliers), the Swiss Army knife (nonparametric test) is the safer, more versatile choice. In behavioral health, where data often violate textbook assumptions, nonparametric methods are invaluable allies.

SECTION 8

Connection to Advanced Theory — Effect Sizes, Power, and Modern Approaches

While the EPPP primarily tests your ability to distinguish among the four categories and select appropriate tests, it is also important to understand how these categories connect to broader statistical concepts. Contemporary behavioral health research increasingly emphasizes effect sizes (e.g., Cohen's d, eta-squared, odds ratios) alongside p-values, recognizing that statistical significance does not equate to clinical significance. Similarly, statistical power analysis has become a standard component of research design, requiring investigators to estimate necessary sample sizes before data collection. Understanding parametric vs. nonparametric distinctions feeds directly into power considerations, since nonparametric tests typically require larger samples to achieve the same power as their parametric counterparts.

Concept	Basic Level (This Lesson)	Advanced Extension
Descriptive → Exploratory Data Analysis	Mean, SD, frequency tables	Boxplots, stem-and-leaf, kernel density estimation; identifying distributional shape before choosing inferential tests
Inferential → Bayesian Statistics	Null hypothesis significance testing (NHST) using p-values	Bayesian inference provides posterior probabilities rather than p-values; growing adoption in clinical trials and meta-analysis
Parametric → Generalized Linear Models	t-test, ANOVA, Pearson r as standalone tests	All parametric tests are special cases of the General Linear Model (GLM); logistic regression extends to binary outcomes
Nonparametric → Bootstrap Methods	Rank-based tests (Mann-Whitney, Kruskal-Wallis)	Bootstrap and permutation tests use resampling to build empirical sampling distributions without any parametric assumptions

From foundational categories to advanced extensions in behavioral health statistics

As you advance in your career, you will encounter sophisticated modeling techniques—multilevel models, structural equation modeling, and machine learning approaches—that blur the traditional parametric/nonparametric boundary. However, the foundational distinctions taught in this lesson remain essential for critically evaluating the vast majority of behavioral health literature and for answering EPPP items correctly. Mastery of when and why to apply each category is a prerequisite for understanding these more complex methods.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A researcher administers an anxiety questionnaire to 200 college students and reports that the mean anxiety score was 42.5 with a standard deviation of 8.3. She does not compare these scores to any other group or test any hypothesis. What type of statistical analysis has she conducted, and why?

Show answer

The researcher has conducted descriptive statistics. She has summarized the data by reporting the mean (a measure of central tendency) and standard deviation (a measure of variability) without attempting to generalize beyond the sample or test any hypothesis about a larger population. Descriptive statistics serve to organize and present data in a meaningful way, which is always the first step in any analysis. No inferential conclusions—about differences between groups, relationships between variables, or population parameters—have been drawn.

Question 2: BASIC CALCULATION

A clinical researcher wants to determine whether patients who receive mindfulness-based stress reduction (MBSR) show significantly different cortisol levels compared to a wait-list control group. Cortisol is measured on a continuous (ratio) scale, and preliminary testing confirms that cortisol levels are approximately normally distributed in both groups with equal variances. Which specific statistical test should the researcher use, and what category does it fall into?

Show answer

The researcher should use an independent-samples t-test. This is a parametric inferential test. It is inferential because the goal is to generalize from the sample to the population of all patients who might receive MBSR. It is parametric because all three key assumptions are met: the dependent variable (cortisol) is measured at the ratio level, the distributions are approximately normal in both groups, and the variances are equal (homogeneity of variance). The t-test will produce a t-statistic and associated p-value, allowing the researcher to determine whether the mean cortisol difference is statistically significant.

Question 3: INTERMEDIATE

A psychologist surveys 80 substance abuse clients and categorizes them by primary substance (alcohol, opioids, stimulants, cannabis) and by treatment completion status (completed, did not complete). She wants to know whether the type of substance is associated with treatment completion. What test should she use, and why would a parametric test be inappropriate here?

Show answer

The psychologist should use a chi-square (χ²) test of independence, which is a nonparametric inferential test. Both variables—primary substance and treatment completion—are categorical (nominal). Parametric tests such as the t-test or ANOVA require a continuous (interval or ratio) dependent variable and assume normal distributions of scores, neither of which applies here. The chi-square test compares observed frequencies in each cell of a contingency table to the frequencies that would be expected if the two variables were independent. A significant χ² result would indicate that treatment completion rates differ by substance type, which could inform targeted intervention strategies.

Question 4: APPLIED

A community mental health center compares client satisfaction ratings (measured on a 5-point Likert scale from 'very dissatisfied' to 'very satisfied') across three clinics. The ratings are markedly skewed, and the sample sizes are 12, 9, and 14 respectively. A new staff member proposes using a one-way ANOVA. Evaluate this proposal and recommend a more appropriate analysis.

Show answer

The one-way ANOVA would be inappropriate for at least two reasons. First, Likert-scale data are ordinal rather than truly interval, meaning the distances between response options (e.g., between 'dissatisfied' and 'neutral') may not be equal—a violation of the interval-level measurement assumption. Second, the distributions are markedly skewed and the sample sizes are small and unequal, violating the normality assumption and reducing the robustness that larger samples might provide. The appropriate alternative is the Kruskal-Wallis H test, which is the nonparametric equivalent of the one-way ANOVA. It ranks all scores across groups and compares the mean ranks, making no assumptions about distribution shape or interval-level measurement. If the Kruskal-Wallis test is significant, post hoc pairwise comparisons (e.g., Dunn's test) can identify which clinics differ.

Question 5: CRITICAL THINKING

A researcher argues that nonparametric tests should always be used in behavioral health research because clinical data rarely meet parametric assumptions perfectly. Provide a nuanced evaluation of this position, discussing the trade-offs involved and citing at least two specific statistical concepts to support your argument.

Show answer

While the researcher raises a valid concern—behavioral health data frequently involve ordinal scales, skewed distributions, and small samples—the blanket recommendation to always use nonparametric tests oversimplifies the issue. Two key concepts challenge this position. First, statistical power: parametric tests are more powerful than their nonparametric equivalents when assumptions are met, meaning they are more likely to detect true effects. Routinely choosing nonparametric methods increases the risk of Type II errors (failing to detect real treatment effects), which has direct clinical consequences. Second, the concept of robustness: many parametric tests (particularly the t-test and ANOVA) are fairly robust to moderate violations of normality, especially with larger samples, because the Central Limit Theorem ensures that sampling distributions of means approach normality. A balanced approach involves conducting exploratory descriptive analyses first (examining distributions, testing for normality with Shapiro-Wilk, checking variance equality with Levene's test), then selecting parametric tests when assumptions are reasonably met and nonparametric tests when they are clearly violated. This decision-making process represents best practice in evidence-based behavioral health research.

SUMMARY

Lesson Summary

Summary

Statistical methods in behavioral health research are organized along two independent dimensions. The first—descriptive versus inferential—distinguishes methods that summarize data (mean, median, standard deviation, frequency distributions) from those that generalize from samples to populations (hypothesis tests, confidence intervals). The second dimension—parametric versus nonparametric—classifies inferential methods by their assumptions. Parametric tests (t-test, ANOVA, Pearson r) assume normal distributions, interval/ratio data, and homogeneity of variance; nonparametric tests (Mann-Whitney U, Kruskal-Wallis H, chi-square, Spearman ρ) make no distributional assumptions and can handle ordinal or nominal data.

For the EPPP, the essential skills are: (1) recognizing which category a given test belongs to, (2) selecting the appropriate test based on research question, level of measurement, and assumption satisfaction, (3) understanding the trade-off between statistical power (favoring parametric when assumptions are met) and robustness (favoring nonparametric when assumptions are violated), and (4) pairing each parametric test with its nonparametric equivalent (e.g., independent t-test ↔ Mann-Whitney U; one-way ANOVA ↔ Kruskal-Wallis H; Pearson r ↔ Spearman ρ). Descriptive statistics always come first, providing the foundation on which inferential analyses are built.