Why Variables and Data Types Matter for Your EPPP Success
Understanding variables and data types isn't just about passing a test question. It's the foundation for reading research, evaluating treatment effectiveness, and making evidence-based decisions in your future practice. When you read that a new therapy approach "significantly reduces depression," you need to know what variables were measured, how they were measured, and whether the conclusions make sense. This lesson will break down these concepts so they stick with you not just for exam day, but for your entire career.
The Cast of Characters: Types of Variables
Every research study tells a story with different players. Let's meet the main characters and understand their roles.
Independent and Dependent Variables: The "What Affects What" Relationship
The independent variable is what researchers manipulate or compare (it's the "cause" side of the equation. The dependent variable is what they measure to see if anything changed) it's the "effect" side.
In plain words: the independent variable is the possible cause you test. The dependent variable is the outcome you score afterward to see whether the cause did anything.
Think of planning a dinner party where you're testing whether background music affects how much your guests enjoy the meal. The type of music (jazz, classical, or silence) is your independent variable (you control it. Your guests' enjoyment ratings are the dependent variable) you measure it to see if the music made a difference.
Here's a practical trick that works every time: Convert any study description into this question format: "What are the effects of [independent variable] on [dependent variable]?"
Let's practice:
- Study: Researchers compare cognitive-behavioral therapy, medication, and combined treatment for reducing panic attacks.
- Question: "What are the effects of treatment type on panic attack frequency?"
- Answer: Treatment type (with three levels: CBT, medication, combined) is the independent variable. Panic attack frequency is the dependent variable.
Remember: The independent variable always has at least two levels. You're always comparing something to something else. Treatment versus no treatment, high dose versus low dose, Group A versus Group B.
Moderator Variables: The "It Depends" Factor
A moderator variable changes the strength or direction of the relationship between your independent and dependent variables. It's like how the effectiveness of your morning coffee (independent variable) on your alertness (dependent variable) might depend on whether you're a regular coffee drinker or someone who rarely has caffeine (moderator variable). The same cup affects different people differently.
Put simply: a moderator is an "it depends on" variable. The treatment has an effect, but how big that effect is (or even which way it goes) depends on who gets it or under what conditions. In statistics, you test a moderator as an interaction, which just means the effect of one variable changes across levels of another.
Clinical example: You're studying whether exposure therapy reduces social anxiety. You find it works great for some clients but not others. When you dig deeper, you discover that clients with strong social support networks improve more than those without support. Social support is a moderator variable. It changes how effective the treatment is.
Mediator Variables: The "Why It Works" Explanation
A mediator variable explains why or how an independent variable affects a dependent variable. It's the mechanism, the middle step in the process.
If you notice that people who start meal planning (independent variable) lose weight (dependent variable), the mediator might be that meal planning leads to fewer impulsive fast-food purchases, which then leads to weight loss. The reduced fast-food consumption is the mediator. It explains the connection.
In psychotherapy research, cognitive therapy for depression assumes this pathway:
- Therapy (independent variable) → Changes in negative thinking patterns (mediator) → Reduced depressive symptoms (dependent variable)
The mediator answers the question: "What's happening between the treatment and the outcome?"
Worked Example: Sorting the Roles in One Study
Worked example: A study compares an 8-week mindfulness program with a waitlist for lowering anxiety scores in graduate students. The program helps students who sleep at least 7 hours a night more than it helps students who sleep less. Students in the program also ruminate less, and the drop in rumination accounts for much of the drop in anxiety.
- Turn the study into the question: "What are the effects of the mindfulness program on anxiety scores?"
- The first blank is the independent variable: group (mindfulness or waitlist). It has two levels.
- The second blank is the dependent variable: anxiety score.
- Ask "for whom, or when, does the effect change size?" Sleep. The program works better for good sleepers, so sleep is a moderator.
- Ask "through what step does the program work?" Rumination. The program lowers rumination, and lower rumination lowers anxiety, so rumination is a mediator.
- Ask "what else differs between the groups?" Suppose most program students were in their final year and most waitlist students were in their first year. Year in school would then be an extraneous variable that confounds the result.
So on the exam: "for whom" or "when" points to a moderator. "How" or "why" points to a mediator.
Extraneous Variables: The Unwanted Guests
Extraneous variables (also called confounding variables) are factors you didn't plan for but might mess up your results. They make it hard to know if your independent variable really caused the changes you observed.
A fine point: many books use "extraneous" and "confounding" as if they mean the same thing. Strictly, an extraneous variable is any outside factor that could affect the outcome. It becomes a confound when it also differs between your groups, so you can't pull its effect apart from the effect of the independent variable. In the example below, time of day is a confound because it differs between the two programs.
Imagine you're comparing two different stress-reduction programs. Program A meets only at 6 AM and Program B meets only at 6 PM. If Program B shows better results, is it because of the program itself, or because of the time of day? Time of day differs along with the program, so it confounds your results.
Two related cases: if the programs were identical and time of day were what you deliberately varied, time of day would be the independent variable. And if people pick their own program, preexisting differences between the groups are a separate problem called selection.
Good research design tries to control or eliminate extraneous variables. That's why randomization matters. It helps spread out potential confounds evenly across groups.
Scales of Measurement: How We Quantify Experience
Not all data is created equal. The way we measure something determines what we can say about it and which statistical tests we can use. Let's work through the four scales from simplest to most informative.
In plain words: a scale of measurement tells you what your numbers are allowed to mean. Some numbers are only name tags. Some show order. Some let you add and subtract. Only some let you say "twice as much."
Nominal Scale: Categories Without Order
Nominal scales simply sort people into categories with no inherent ranking. These are labels, nothing more.
Examples:
- Gender identity
- Eye color
- Diagnosis (depression, anxiety, PTSD)
- Treatment group assignment (Group A, Group B)
You can count how many people fall into each category, but you can't say one category is "more" or "better" than another in a quantitative sense. It's like organizing your streaming services. Netflix, Hulu, and Disney+ are different categories, but none is mathematically "greater" than the others.
A special type of nominal variable is the dichotomous variable, which has only two categories (yes/no, treated/untreated, improved/not improved).
Ordinal Scale: Ranked but Unequal Steps
Ordinal scales put people in order, but the distances between rankings aren't necessarily equal.
Think about how you rate your stress level: "not stressed," "somewhat stressed," "very stressed," and "extremely stressed." You know "very stressed" is worse than "somewhat stressed," but the jump between categories might not be the same size each time. The difference between "not" and "somewhat" might be smaller than the difference between "very" and "extremely."
Common examples:
- Likert scales (strongly agree to strongly disagree)
- Class rankings (1st, 2nd, 3rd place)
- Education level (high school, bachelor's, master's, doctorate)
With ordinal data, you know the order but not the exact distance between points.
Interval Scale: Equal Distances, No True Zero
Interval scales have equal distances between adjacent points, but no absolute zero point where the quality being measured completely disappears.
Put simply: every step on the ruler is the same size, but the zero is just a spot someone picked. It does not mean "none."
IQ scores are the classic example. The difference between IQ scores of 100 and 101 equals the difference between 130 and 131. However, an IQ of 0 doesn't mean "no intelligence". There's no true zero point on the scale.
Most standardized psychological tests produce interval data. The key feature: equal intervals mean you can add and subtract scores meaningfully, but you can't make ratio statements.
Ratio Scale: The Full Package
Ratio scales have everything: ordered categories, equal intervals, and an absolute zero point that means "none of this quality exists."
Examples:
- Weight in pounds
- Age in years
- Number of therapy sessions attended
- Reaction time in milliseconds
- Annual income in dollars
The absolute zero point lets you make ratio statements. Someone who weighs 200 pounds is exactly twice as heavy as someone who weighs 100 pounds. Someone who attended 10 sessions attended twice as many as someone who attended 5 sessions.
Here's the key distinction between interval and ratio: You cannot say someone with an IQ of 150 is 1.5 times as intelligent as someone with an IQ of 100 (interval scale), but you can say someone who's 60 years old is twice as old as someone who's 30 (ratio scale).
| Scale | Order? | Equal Intervals? | True Zero? | Example |
|---|---|---|---|---|
| Nominal | No | No | No | Diagnosis categories |
| Ordinal | Yes | No | No | Likert ratings |
| Interval | Yes | Yes | No | IQ scores |
| Ratio | Yes | Yes | Yes | Number of symptoms |
Worked Example: Naming the Scale
Worked example: Name the scale for each variable. Ask three questions in order: Is there an order? Are the steps equal? Does zero mean "none of it"? Stop at the first "no."
- Primary diagnosis (MDD, GAD, PTSD). Order? No. Stop. This is nominal.
- Finishing place in a race (1st, 2nd, 3rd). Order? Yes. Equal steps? No. The winner might beat 2nd place by one second, while 2nd beats 3rd by a full minute. This is ordinal.
- Temperature in degrees Celsius. Order? Yes. Equal steps? Yes. Does 0 mean "no heat"? No. 0 °C is just the point where water freezes. This is interval, so 20 °C is not "twice as hot" as 10 °C.
- Number of panic attacks last week. Order? Yes. Equal steps? Yes. Does 0 mean "none"? Yes. This is ratio, so 6 attacks really is twice as many as 3.
- Now pick the graph. Diagnosis and race place are categories, so they get bar graphs. Temperature and panic attack counts are numeric scores, so they get histograms or frequency polygons.
So on the exam: if zero means "none of it" and "twice as much" makes sense, it's ratio. If steps are equal but "twice as much" makes no sense, it's interval.
Visualizing Data: Choosing the Right Graph
Your measurement scale determines which graph type is appropriate. Using the wrong graph for your data type is like trying to eat soup with a fork. Technically possible but definitely not right.
Bar Graphs: For Categories
Bar graphs work with nominal and ordinal data. Each category gets its own bar, and the bars are separated by spaces to show they're distinct categories, not a continuous scale.
If you're displaying how many clients at your clinic have different primary diagnoses (depression, anxiety, OCD, PTSD), you'd use a bar graph. Each diagnosis is a separate category, and the space between bars reinforces that these aren't points on a continuum.
Histograms: For Continuous Data
Histograms are used with interval and ratio data. The bars touch each other because the data represents a continuous range of scores, not separate categories.
If you're showing the distribution of depression severity scores (measured on a scale from 0-100) among your clients, you'd use a histogram. The touching bars show this is a continuous scale where someone could theoretically score anywhere along the range.
Frequency Polygons (Line Graphs): Another Option for Continuous Data
Frequency polygons also work with interval and ratio data. Instead of bars, you place dots above each score or score interval and connect them with lines, creating a smooth shape that makes patterns easy to see.
Understanding Distribution Shapes: What the Data's Shape Tells You
When you graph data, the shape of the distribution reveals important information about your sample and helps you choose appropriate statistical analyses.
The Normal Distribution: The Ideal Bell Curve
A quick term first. The standard deviation (SD) is roughly the typical distance between a score and the mean. A small SD means scores huddle close to the average. A large SD means they spread out.
The normal distribution is symmetrical and bell-shaped. It has special properties that make it beloved by statisticians:
- The mean, median, and mode all equal the same value (they're all at the center peak)
- About 68% of scores fall within one standard deviation of the mean
- About 95% fall within two standard deviations
- About 99.7% fall within three standard deviations (this is why people call it the 68-95-99.7 rule)
If you measure the wait times for appointments at a large clinic and get a normal distribution with a mean of 15 minutes and standard deviation of 5 minutes, you know that about 68% of clients wait between 10 and 20 minutes, about 95% wait between 5 and 25 minutes, and virtually everyone (about 99.7%) waits between 0 and 30 minutes.
Many statistical tests assume your data follows a normal distribution, so understanding this pattern matters for analysis decisions.
Worked example: IQ scores have a mean of 100 and an SD of 15. About what share of people score above 130?
- Find how far 130 is from the mean: 130 minus 100 = 30 points.
- Turn points into SDs: 30 / 15 = 2 SDs above the mean.
- The rule says about 95% of scores fall within 2 SDs of the mean, so between 70 and 130.
- That leaves about 5% outside that band. The curve is symmetric, so half of that 5% is above 130 and half is below 70.
- So about 2.5% score above 130. (A z table gives about 2.3%, because the true share within 2 SDs is a bit over 95%.)
The same steps work for 1 SD. About 68% score between 85 and 115, which leaves 32% outside. Half of that, about 16%, score above 115.
So on the exam: subtract the mean, divide by the SD, then split whatever falls outside the 68, 95, or 99.7 band in half.
Skewed Distributions: When Data Piles to One Side
Skewed distributions are asymmetrical, with most scores bunched on one side and a few extreme scores creating a tail on the other side.
Negatively skewed distribution: Most scores are high, with a few low scores stretching out the left tail. Think of scores on an easy exam where most students get As and Bs, but a few students who didn't study get Ds and Fs.
Positively skewed distribution: Most scores are low, with a few high scores stretching out the right tail. Annual income in the general population is positively skewed. Most people earn moderate amounts, but a few high earners create a long tail on the high end.
Here's the memory trick: The tail tells the tale. The tail points toward the skew direction. Negative skew = tail points left (toward negative numbers). Positive skew = tail points right (toward positive numbers).
For many familiar, single-peaked skewed distributions, this pattern is a useful sketch, not a rule for every data set:
- The mean gets pulled toward the tail (toward the extreme scores)
- The mode stays with the hump (where most scores cluster)
- The median sits in between
| Distribution Type | Mean Position | Median Position | Mode Position |
|---|---|---|---|
| Negatively Skewed | Lowest value | Middle | Highest value |
| Normal | All three equal | All three equal | All three equal |
| Positively Skewed | Highest value | Middle | Lowest value |
Read this table as the usual pattern for a skewed curve with one hump, not a law. When you have the actual numbers, compute them (see "Calculating the center" below).
Worked example: Five scores on an easy quiz: 40, 85, 90, 95, 95.
- Mean: 40 + 85 + 90 + 95 + 95 = 405. Then 405 / 5 = 81.
- Median: the scores are already in order, so the middle one (the 3rd of 5) is 90.
- Mode: 95 shows up twice, so the mode is 95.
- Order: mean 81, median 90, mode 95. The one low score (40) makes a tail on the left and drags the mean down toward it. This is a negative skew.
- Which number best shows a typical student? The median (90). The mean (81) is lower than four of the five scores because of one student.
So on the exam: the mean is the measure most pulled by extreme scores, so in skewed data the median is usually the safer "typical" score.
When data is skewed, the median is usually better than the mean for representing a typical score because the mean gets distorted by those extreme values.
Leptokurtic and Platykurtic: Peak Height Variations
These describe how pointy or flat a distribution looks compared to a normal curve.
The technical word for this is kurtosis. Despite the "pointy versus flat" labels, kurtosis mostly tracks the tails, meaning how often you get extreme scores far from the mean. That is why each definition below mentions the tails, not just the peak.
Leptokurtic distribution: Sharper peak AND fatter tails than normal. Scores cluster tightly in the middle, but there are also more extreme high and low scores than a normal curve would have. Picture most scores bunched tightly around the average, plus a handful of people who scored way higher or way lower than everyone else.
Platykurtic distribution: Flatter peak AND thinner tails than normal. Scores spread out more evenly with no strong clustering and fewer extreme outliers. Imagine a test where scores spread out across the range with no strong bunching in the middle.
These terms come up less frequently on the EPPP than skewness, but they're worth knowing for recognition purposes.
Common Misconceptions and Exam Traps
Misconception 1: "The independent variable is always manipulated by the researcher."
Not quite. While researchers often assign participants to different treatment conditions (true manipulation), comparing pre-existing groups also involves an independent variable. When you compare therapy outcomes for clients with different attachment styles, attachment style is your independent variable even though you didn't create or assign it.
Misconception 2: "Moderator and mediator are basically the same thing."
They're fundamentally different. A moderator changes whether or how much something works (it modifies the effect). A mediator explains why something works (it's part of the causal chain). Moderators answer "for whom?" while mediators answer "how?"
Misconception 3: "Higher level scales are always better."
While ratio scales provide more information than nominal scales, the appropriate scale depends on what you're measuring. You can't measure someone's diagnosis category on a ratio scale. Some variables are inherently categorical. Use the scale that matches your construct.
Misconception 4: "If bars touch, it's always a histogram."
The touching bars in a histogram reflect continuous data, but make sure the data is actually interval or ratio before calling it a histogram. The graph type should match your measurement scale.
Misconception 5: "In a positively skewed distribution, most scores are positive."
No! The direction of skew refers to where the tail points, not the sign of the scores. In a positively skewed distribution, the tail points toward higher (more positive) values, but most scores are actually on the lower end. The few extreme high scores create the positive tail.
Memory Aids and Study Strategies
For Independent vs. Dependent Variables: Create the question "What are the effects of ___ on ___?" The first blank is always your independent variable, the second is your dependent variable.
For Moderator vs. Mediator:
- Moderator: Sounds like "moderate" or "modify". It modifies the strength of the effect
- Mediator: Think of the word "medium" or "middle". It's the middle step that explains the connection
For Scales of Measurement (NOIR): Use the acronym NOIR (French for "black"): Nominal, Ordinal, Interval, Ratio. They go from least to most information.
For Skewness: Visual memory: Draw a quick distribution and notice where the tail points. That tells you the direction of the skew. Then remember: mean follows tail, mode follows hump, median often in the middle. Check actual values rather than treating this sketch as a universal rule.
For Bar Graph vs. Histogram:
- Bar graph = Breaks between bars (for categories)
- Histogram = Hugging bars (for continuous data)
Practice Exercise: Take any research abstract from a psychology journal and identify all five variable types: What's the independent variable? What's dependent? Can you spot any moderators or mediators mentioned? What about potential extraneous variables they tried to control?
Calculating the center
In plain words: the mean, median, and mode are three ways to answer "what's a typical score?" The mean is the average. The median is the middle score once you line them up. The mode is the most common score.
The mean is the sum of the values divided by their number. The median is the middle ordered value; with an even number of values, average the two middle ones. The mode is the most frequent value, and a data set can have more than one mode. Extreme values can shift the mean strongly. For 20, 20, 20, 30, and 110, the mean is 40 while the median and mode are both 20. Compute these values instead of assuming that a skew label guarantees their order (Manikandan, 2011a, 2011b).
Worked example: Six clients attended this many sessions: 3, 8, 2, 5, 5, 1.
- Sort them first: 1, 2, 3, 5, 5, 8.
- Mean: 1 + 2 + 3 + 5 + 5 + 8 = 24. Then 24 / 6 = 4.
- Median: there are six values, an even number, so average the two middle ones (the 3rd and 4th): (3 + 5) / 2 = 4.
- Mode: 5 shows up twice, so the mode is 5.
- Compare with the skew table. The longer side of this set is the high end (8 sits 4 above the mean, while 1 sits only 3 below), and a skewness statistic comes out slightly positive. Yet the mode (5) is the highest of the three, which is the opposite of what the table predicts for positive skew.
So on the exam: if a question gives you the numbers, compute the mean, median, and mode. Use the "mean follows the tail" sketch only when all you have is a picture or a label.
Key Takeaways
- Independent variables are what researchers manipulate or compare; dependent variables are what they measure as outcomes
- Moderator variables change the strength or direction of effects (they modify relationships)
- Mediator variables explain why or how effects occur (they're the mechanism)
- Extraneous variables are unintended factors that confound your results
- Nominal scales provide categories without order (diagnosis, gender)
- Ordinal scales provide order without equal intervals (rankings, Likert scales)
- Interval scales have equal intervals but no true zero (IQ, most standardized tests)
- Ratio scales have equal intervals and a true zero (age, weight, count data)
- Bar graphs (with spaces between bars) display nominal and ordinal data
- Histograms and frequency polygons (continuous) display interval and ratio data
- In a normal distribution, mean = median = mode, and scores follow the approximate 68-95-99.7 rule
- In skewed distributions, the tail tells the direction, and the mean gets pulled toward the tail
- Leptokurtic = pointy peak and fat tails (more extreme scores); platykurtic = flatter peak and thin tails (fewer extreme scores)
Understanding these fundamentals will help you analyze research questions correctly on the EPPP and evaluate scientific evidence throughout your career. When you can identify what type of variable you're dealing with and what its measurement properties are, you're equipped to think critically about research claims and statistical conclusions.
Random assignment balances in expectation
In plain words: random assignment is fair on average, but not perfect every time. "In expectation" means "on average, if you could redo the random split many times."
It's like shuffling a deck and dealing two hands. Over many deals, each hand gets about the same number of aces. In any one deal, one hand can still get three aces.
Random assignment balances measured and unmeasured characteristics in expectation across repeated assignments. It does not guarantee equal groups in a particular trial, equal group sizes, or a representative sample. Random assignment addresses treatment-group comparability; sampling addresses who enters the study (Handley et al., 2018).
What this means in practice: if the two groups in one randomized study differ at the start (say, one group is more depressed at intake), that can happen by chance even when randomization was done right. Researchers check for these baseline differences and can adjust for them in the analysis. Also keep the two ideas apart. Random assignment decides which group you land in, which protects internal validity. Random sampling decides who gets into the study at all, which supports external validity (how well results generalize).
Flag-resolution sources
- S Manikandan. Measures of central tendency: The mean. Journal of pharmacology & pharmacotherapeutics. 2011 Apr;2(2):140-2. doi:10.4103/0976-500X.81920. PMID:21772786. https://pubmed.ncbi.nlm.nih.gov/21772786/
- S Manikandan. Measures of central tendency: Median and mode. Journal of pharmacology & pharmacotherapeutics. 2011 Jul;2(3):214-5. doi:10.4103/0976-500X.83300. PMID:21897729. https://pubmed.ncbi.nlm.nih.gov/21897729/
