On this page · 5 sections
Data Interpretation is Science Practice 3, one of the four science practices AP Psychology tests. The others are Concept Application (SP1), Research Methods and Design (SP2), and Argumentation (SP4), and that last one is assessed only on the free-response section. Data Interpretation, like Research Methods and Design, can often be scored straight from the stimulus without recalling specific theories or findings. The exam hands you a bar graph, a line graph, a scatterplot, a frequency distribution, or a table of means and standard deviations, and asks a narrow question: what does this show, and what does it not show? The wrong answers are almost never wrong about the numbers. They are wrong about what the numbers license you to say.
In Section I, 75 multiple-choice questions in 90 minutes, figures arrive as a stimulus with one or two questions attached. Data Interpretation is about 10 percent of that section, roughly 7 or 8 of the 75 questions. Concept Application carries about 65 percent and Research Methods and Design about 25 percent. You get about 72 seconds per question, so a graph you have to decode from scratch will eat two questions' worth of time unless you have a fixed routine.
In Section II, 70 minutes and 14 raw points, data reading is unavoidable. The Article Analysis Question is worth 7 points and the Evidence-Based Question is worth 7 points. The two are weighted equally, 16.65 percent of the exam each. The AAQ gives you a summary of one peer-reviewed study, including its figures and reported statistics. Parts A through E are worth 1 point each. Part C asks you to describe what a specific reported statistic means. Part F is worth up to 2 points and is scored under Argumentation. It asks you to explain how at least one reported finding supports or refutes the researchers' hypothesis. Budget about 25 minutes, the first 10 of which are reading. The EBQ gives you three summarized sources, each with its own results to cite. Budget about 45 minutes, the first 15 for reading.
Scoring is literal and point-by-point. Readers are not judging whether your paragraph is well written. They are checking whether you named the specific number, said what it means in that study's context, and drew a conclusion the design actually supports. A vague gesture at the right trend earns nothing. A correct claim wrapped around a causal overreach can cost the point you already had.
The steps
- 1
Read the frame before you read the data
Do not look at the bars first. Look at the labels. Pin down four things: what the x-axis measures, what the y-axis measures, the units, and where the y-axis starts. A y-axis beginning at 15 instead of 0 makes a three-point gap look like a landslide. Treat a truncated axis as a test of whether you read the scale. Next find n, the number of participants, and read the caption. The caption usually names the population — undergraduates at one university, children ages 4 to 6, adults recruited online — and that phrase is your evidence for any generalizability question later, including part E of the Article Analysis Question. Naming the sample is not the whole point, though. You still have to say what that sample does and does not let you generalize to. If there is a legend, decide what each bar color or line stands for before you interpret anything. This takes about fifteen seconds and prevents the most common data error on the exam, which is answering correctly about the wrong variable.
- 2
Name the graph type, because it fixes what claim is legal
The graph type is a rule about what you are allowed to conclude. A bar graph compares separate groups or categories, usually the mean score in each condition, so it supports statements like "the nap group recalled more." A line graph tracks a continuous variable, most often time or trials, so it supports statements about a trend or a rate of change. A scatterplot plots two measured variables for each participant, one dot per person, and supports only a statement about relationship. It can never support a statement about cause. A histogram looks like a bar graph but is a different animal: its x-axis is the score itself, chopped into intervals, and its height is how many people landed in that interval. Histograms answer questions about shape — normal, bimodal, positively skewed, negatively skewed. A pie chart shows parts of a whole and supports proportion claims only. Naming the type first eliminates whole answer choices before you read them.
- 3
Compress the figure into one plain sentence
Before you look at the answer choices, force the picture into a single sentence in your own words. For a bar graph the sentence is comparative: "Condition A scored about 3 points higher than condition B." For a line graph it is directional: "As trials increased, errors dropped sharply and then flattened out." For a scatterplot it is relational: "People who scored higher on X tended to score lower on Y." Keep the word tended. Use approximate values read off the axis, not exact ones, because the exam almost never depends on the last digit and the size of the gap matters more than its decimals. Write the sentence in the margin or in the digital notes if the figure has more than two series. If you cannot produce this sentence, you have misread an axis. Go back rather than guessing from the shape of the picture, because shape without labels is meaningless.
- 4
Decode the statistics printed beside the figure
CENTER. The mean (M) is the arithmetic average and gets dragged toward extreme scores. The median is the middle score and resists them. The mode is the score that occurs most often, and it is the only one of the three that works for categories like favorite color. A distribution with two clear peaks is bimodal, which usually means two different groups got mixed into one data set.
SKEW. In a positively skewed distribution the tail stretches right, a few very high scores pull the mean up, and the mean sits above the median. In a negatively skewed distribution the tail stretches left and the mean sits below the median. Skew is named for the tail, not for where the scores pile up.
SPREAD. The range is the highest score minus the lowest, which one outlier can wreck. The standard deviation (SD) is the typical distance from the mean, and it uses every score: small SD means tightly clustered, large SD means spread out. Always compare a gap between two means to their SDs.
THE NORMAL CURVE. In a normal distribution the mean, median, and mode sit at the same point, and the 68-95-99.7 rule fixes the rest. About 68 percent of scores fall within 1 SD of the mean, about 95 percent within 2 SD, about 99.7 percent within 3 SD. On an IQ test with M = 100 and SD = 15, that means about 68 percent of people score between 85 and 115 and about 95 percent score between 70 and 130.
PERCENTILE RANK. A score's percentile rank is the percentage of scores at or below it. Read it off the normal curve by adding up the area to the left: the mean is the 50th percentile, +1 SD is about the 84th, -1 SD is about the 16th, and +2 SD is about the 98th. An IQ of 115 is therefore roughly the 84th percentile.
SIGNIFICANCE AND SIZE. A p-value below .05 means statistically significant: a result this large would be unlikely if the manipulation truly had no effect. Effect size tells you how big the difference is, which significance does not. Cohen's d is the difference between two means divided by the standard deviation. The course uses set bands: 0.2 and below is small, 0.3 to 0.7 is medium, 0.8 or greater is large. Judge size with d against those bands. There is no rule that a gap must exceed a full standard deviation to count as large.
CORRELATION. The correlation coefficient r runs from -1.00 to +1.00. The sign gives direction only and the absolute value gives strength, so r = -.71 is a stronger relationship than r = +.34.
REGRESSION TOWARD THE MEAN. Extreme scores tend to drift back toward average on a retest, because part of any extreme score is luck and luck does not repeat. That is why a coach who screams after an athlete's worst game seems to get improvement, and why a pretest-posttest gain in a group chosen for scoring very low is not automatic evidence the program worked.
- 5
Check the design that produced the data
Data can never outrank the design behind it. Hunt for one phrase: random assignment. If participants were randomly assigned to conditions and the researcher manipulated something, it is an experiment and a causal claim is legal for that manipulation. If the researchers only measured variables that already differed among people — sleep, income, screen time, anxiety — it is correlational, and the strongest legal claim is "associated with" or "predicts." Random sampling is a separate idea: it concerns who got into the study at all, and it governs generalizability rather than causation. A study can have clean random assignment and an unrepresentative sample, or a beautiful sample and no random assignment. With correlational data, name the two escape routes in your head. The third-variable problem: if sleep correlates with grades, conscientiousness might drive both. Reverse causation: anxious students may sleep less because they are anxious, not the other way around.
- 6
Eliminate rather than choose
Read the multiple-choice options as things to delete. Four eliminations recur. First, any option using caused, leads to, results in, produces, or because when the data are correlational. Second, any option that flips direction — swapping which group scored higher, or calling a downward-sloping scatterplot a positive correlation. Third, any option that adds information the figure does not contain: a mental process, a motive, an emotion, or a variable that appears on neither axis. Fourth, any option that generalizes past the sample in the caption, especially with the words all, everyone, or people in general. What survives is usually a flat restatement of the sentence you wrote in step three. Flat is correct. The exam rewards the most cautious claim the data actually support, never the most interesting one, and distractors are built to be more interesting than the answer.
- 7
On free response, name the number and finish the logic
Describing a figure earns nothing on its own. A scoring answer has three parts in order: what the data show, the specific value with its units, and an explicit statement of what that means for the claim in the prompt. "The data support the hypothesis: the nap group recalled 21.4 word pairs on average compared with 17.9 for the reading group, and p = .01 indicates the difference is statistically significant." That structure is what part C and part F of the Article Analysis Question are checking. Parts A through E are worth 1 point each. Part F is worth 2, so give it roughly double the effort of any other part. Its bands are strict. Cite a result but never explain how it supports or refutes the hypothesis, and you earn 1 point. Explain without citing a specific result, and you also earn 1. Do both, with the result read correctly, and you earn 2. The Evidence-Based Question splits its 7 points the same lopsided way. Part A is 1 point for the claim. Part B(i) is 1 point for evidence from one source. Part B(ii) is worth 2: explain how that evidence relates to the claim and apply a psychological concept. Part C(i) is 1 point for evidence from a different source. Part C(ii) is worth 2 and must apply a psychological concept different from the one you used in B(ii). Use the source's own terminology. If it reports a correlation, write correlation. Answer only what is asked, label each part, and leave nothing blank, because every part is scored on its own.
Worked through
A study summary and two figures, worked start to finish.
THE STIMULUS. Researchers randomly assigned 140 high school students, 70 per condition, to either a 20-minute nap or 20 minutes of quiet reading. Every student then took a memory test scored as the number of word pairs correctly recalled out of 40. Figure 1 is a bar graph; the y-axis is labeled "Mean word pairs recalled (of 40)" and runs from 15 to 24. The nap bar sits at 21.4, the reading bar at 17.9. The text reports nap M = 21.4, SD = 8.0; reading M = 17.9, SD = 8.0; p = .01. The summary also states that the researchers obtained informed consent from the students and written permission from their guardians, and told every student they could stop at any time. Figure 2 is a scatterplot of each student's self-reported hours of sleep the previous night against that student's recall score, with r = +.28.
STEP 1, FRAME. Figure 1: x-axis is condition, nap or reading. Y-axis is mean word pairs recalled, out of 40. Sample is 140 high school students. Critically, the y-axis starts at 15, not 0. Measured from the bottom of that axis, the nap bar is 6.4 units tall and the reading bar is 2.9, so the nap bar looks more than twice as big. The real difference is 3.5 points out of 40, about a 20 percent edge. Correct for that before you judge anything as "large."
STEP 2, TYPE. Figure 1 is a bar graph comparing two groups, so it licenses a group comparison. Figure 2 is a scatterplot, so it licenses a relationship statement and nothing more.
STEP 3, ONE SENTENCE EACH. Figure 1: the nap group recalled about 3.5 more word pairs than the reading group. Figure 2: students who slept more the previous night tended to recall slightly more word pairs.
STEP 4, STATISTICS. Both SDs are 8.0, so Cohen's d = 3.5 ÷ 8.0 = 0.44. Against the course bands (0.2 and below small, 0.3 to 0.7 medium, 0.8 or greater large), that is a medium effect — real, but not large. The p-value of .01 means a gap this large would be unlikely if napping had no effect, so the difference is statistically significant. Significance and size are separate questions, and here the answers differ: significant, medium. The r of +.28 is a weak positive correlation: the dots trend upward but scatter widely, so sleep the night before predicts very little about any one student's score.
STEP 5, DESIGN. Figure 1 comes from random assignment plus a manipulation, so this is an experiment and a causal claim about the nap is legal for this sample. Figure 2 measured sleep the students chose on their own, so it is correlational. Students who sleep more may also study more, work fewer hours, or feel less anxious, and any of those could drive recall.
STEP 6, ELIMINATE. "Napping doubled recall performance" dies on the axis: the bars only look twice as tall because the scale starts at 15, and 21.4 is about 20 percent above 17.9, not double. "Napping improves memory in all adolescents" dies on the word all, since the sample is one group of high school students. "Sleeping longer the night before causes better recall" dies on causation from a scatterplot. "The difference between conditions was large relative to the variability within conditions" dies on the effect size: d = 0.44 is medium, not large. "Participants who napped recalled significantly more word pairs than participants who read" survives; it matches the sentence and the design, and it is the dullest option on the list.
STEP 7, FREE-RESPONSE LINES. These cover the parts of the Article Analysis Question that turn on the data; part D asks about ethical guidelines rather than about the figures, so it is listed here for completeness only. Research method (A): an experiment, because participants were randomly assigned to the nap or reading condition and the researchers manipulated that condition. Operational definition of memory (B): the number of word pairs correctly recalled out of 40 on the post-condition test. Meaning of the reported statistic (C): p = .01 means the difference between conditions is statistically significant, so a difference this large is unlikely to have occurred if the nap had no real effect. Ethical guideline (D): informed consent. The summary states that the researchers got consent from the students and written permission from their guardians before assigning conditions, and told students they could stop at any time. Part D scores only a guideline the summary itself describes, and it wants both the name and how the researchers applied it. If a summary reports no safeguard at all, there is nothing to cite, and a guideline you supply from memory earns zero. Generalizability (E): limited, because all 140 participants were high school students tested in a single study, so the result may not extend to adults or to older learners. Support or refute (F): the data support the hypothesis, since the nap group recalled 21.4 word pairs on average versus 17.9 for the reading group and that difference reached statistical significance at p = .01. That answer earns both points, because it names a specific result and explains what the result does to the hypothesis. Drop either half and it drops to 1 point.
Common mistakes
Reading a scatterplot or a correlation coefficient as evidence of cause. A relationship between two measured variables leaves the third-variable problem and reverse causation wide open, so the only legal wording is associated with or predicts.
Ranking correlation strength by sign instead of absolute value. An r of -.68 describes a stronger relationship than an r of +.31; the minus sign only means the variables move in opposite directions. Treating negative correlations as weak or as "no relationship" is a standing trap.
Hearing statistically significant and thinking large or important. Significance says the result is unlikely to be a fluke of sampling; it says nothing about size. With a big enough sample, a tiny difference reaches p < .05. Size lives in the effect size and in the gap relative to the standard deviations.
Judging a difference by how tall the bars look. A truncated or rescaled y-axis can make a small gap look decisive, so read the axis numbers and compare the gap to the SDs. Then judge the gap with Cohen's d against the course bands: 0.2 and below is small, 0.3 to 0.7 is medium, 0.8 or greater is large.
Mixing up mean and median in a skewed distribution, or reversing the skew direction. Skew is named for the tail, not the bulk of the scores, and the mean is pulled toward the tail. Positive skew means a right tail and mean above median; negative skew means a left tail and mean below median.
Misreading the normal curve. About 68 percent of scores fall within 1 SD of the mean and about 95 percent within 2 SD, so a score 1 SD above the mean sits near the 84th percentile, not the 68th. Percentile rank is the percentage of scores at or below yours, so add up the area to the left of the score, not the area around it.
On free response, describing the figure without citing a number or without connecting it to the claim. "The graph shows the nap group did better" earns nothing. Name the value, name the statistic, and state explicitly whether it supports or refutes the hypothesis. Adding a causal sentence you were not asked for can undo a point you already earned.
Checklist
- Can you name the x-axis variable, the y-axis variable, the units, and the y-axis starting value within fifteen seconds of seeing a figure?
- Can you look at a graph and say which type it is, and state what claim that type does and does not license?
- Can you write a one-sentence summary of any figure before reading a single answer choice?
- Can you interpret M, median, mode, range, SD, r, p, and Cohen's d in a sentence each, and say what each one does not tell you?
- Can you compute Cohen's d from two means and an SD, and label it small, medium, or large using the 0.2, 0.3-to-0.7, and 0.8 bands?
- Can you use the 68-95-99.7 rule to place a score on a normal curve and name its approximate percentile rank?
- Can you spot random assignment in a study summary and decide instantly whether a causal claim is legal?
- Can you name two alternative explanations — a plausible third variable and a reverse-causation story — for any correlation you are shown?
- Can you write a free-response sentence that states the claim, cites the specific value with its units, and says explicitly whether the data support or refute the hypothesis?