7.4 Samples, Scatterplots & Association
Key Takeaways
Define the target population before choosing a sample. Check the sampling frame and which selected people responded.
Random samples of the same size can produce different estimates. Compare center and spread in context.
A scatterplot point gives two measurements on the same case. Describe shape, direction, clusters and unusual observations.
For an informal linear fit, judge the vertical gaps. Give slope in axis units and interpret the intercept only when zero is meaningful.
Association alone cannot establish the effect of changing one measurement on another.
7.4 Samples, Scatterplots & Association
Quick Answer: Choose a sample that can represent the population you want to describe, and expect estimates to vary across samples. In a scatterplot each point shows two measurements on the same case. Describe direction, shape, clusters and unusual points before fitting a line. An association describes a pattern in the data, and a causal conclusion requires additional evidence about what produced it.
Samples and the Population
The population is the full group the question concerns. A sample is the subset you observe. State what belongs to the population before deciding whom to measure. Suppose a library wants to estimate what share of all 1,000 registered borrowers prefer weekend hours. Interviewing the first 50 people who happen to be there on Saturday may overrepresent people who can already visit then. The target is all registered borrowers, including people who rarely visit.
A sampling frame is the list or procedure used to reach that population. With a current list of all 1,000 borrowers, assign each a unique number. Select 50 different numbers at random by making every set of 50 equally eligible. This is a simple random sample without replacement. A complete frame and random selection reduce selection problems, but one random sample can still differ from the population. Check whether selected people actually reply. If replies come primarily from borrowers who already visit on weekends, random selection at the start does not resolve that nonresponse problem.
Comparing Estimates From Equal-Size Samples
Consider three possible simple random samples of 50 from that list. The separately selected samples contain 26, 30 and 24 people who say they prefer weekend hours. Their estimates are 52%, 60% and 48%. Applied separately to 1,000 borrowers, those would suggest 520, 600 and 480 people respectively. The 12-percentage-point span illustrates that different samples of the same size can produce different estimates. Treat these as illustrations of sampling variation. With only three samples, the smallest and largest estimates are not guaranteed bounds on the population share.
You can simulate the same method using a list of 100 model records, 60 marked Yes and 40 marked No. Shuffle or randomly reorder the records and draw 10 without replacement. Count the Yes records and divide by 10. Return all 10 records before drawing the next sample of size 10. Results with 5 or 7 Yes records estimate 50% or 70%, although the model population share is fixed at 60%. The simulation illustrates variation under its own assumed population.
Compare Groups Using Both Center and Spread
Suppose two library branches independently take random samples of visits from the same month and record wait times the same way. The first has mean wait 8 minutes and mean absolute deviation from its mean of 2 minutes. The second has mean wait 14 minutes and mean absolute deviation of 5 minutes. The second sample is 6 minutes higher on average and also shows more variability by that measure. These summaries support an informal comparison of the branches for the sampled month. They do not show that every visitor at one waited longer or establish the same pattern for another month.
Construct and Read a Scatterplot
A scatterplot shows two numerical measurements for each case. For example, each dot might pair the hours a volunteer team worked with the number of boxes it packed on that shift. Put hours on the horizontal axis and boxes on the vertical axis. Label both axes and choose steady increments. Plot each pair without connecting the points in sequence.
| Hours | Boxes |
|---|---|
| 1 | 6 |
| 2 | 7 |
| 3 | 10 |
| 4 | 11 |
| 5 | 12 |
| 5 | 6 |
Plot these at the following locations. There are two shifts with 5 hours, which can have different outputs because these are observations, not a rule assigning one output to an input.
Boxes
12 | •
11 | •
10 | •
9 |
8 |
7 | •
6 | • •
+----------------
1 2 3 4 5 Hours
Most points follow an upward, roughly straight pattern. The point is far below that pattern and may merit checking. A downward pattern would show a negative association. A set of points that climbs and then falls can show a nonlinear association even when a straight line reveals little. Look for clusters separated by gaps, too. A cluster can reflect a subgroup, but the plot alone does not identify what created it.
Fit and Use an Approximate Line
For a roughly straight pattern, sketch a line through the center of most points and inspect the vertical gaps. A line through and is one informal fit for the shift data. Its slope is boxes per hour and its vertical intercept is boxes. The equation for predicted boxes is . For the first five shifts, predictions are 5, 7, 9, 11 and 13 boxes, all within one box of the recorded values. At the line predicts 13, so the observed value is 7 boxes below the line. This is a reason to investigate the shift or assess how strongly this line represents all shifts. Keep the observed point visible.
At 4.5 hours the line predicts boxes. Slope describes a change of about two predicted boxes for each additional hour in this model. The intercept 3 is the predicted value at zero hours in the equation. Zero is outside the observed 1–5-hour range, so the intercept does not establish what anyone would pack in zero hours. A fit can help predict within its observed range while remaining a model of the data.
Explain What the Pattern Supports
An association says the measurements tend to vary together. In the volunteer example, larger teams might both stay longer and pack more. Team size could help explain the upward pattern, so the graph alone cannot tell how much the extra hours cause output to change. Likewise, a sample drawn at random helps generalize to the population it represents. It does not randomly assign different working hours. If a causal question matters, a design that compares similar teams and randomly assigns schedules could reduce systematic differences between groups. Even then, the measurements, adherence to the schedule and sampling scope affect what can be concluded.
A training program wants to estimate the share of all 400 enrolled learners who can attend a morning class. Which method best addresses selection bias?
Ask the first 100 learners to arrive for an existing morning class.
Randomly select 60 learners from the complete enrollment list. Contact all selected learners and check which ones respond.
Post a voluntary poll on the morning class sign-up page.
Survey every learner already attending a morning class and use the larger sample to represent the entire program.
Two separate random samples of 40 from the same 800 households found 22 and 26 households that recycle weekly. Which interpretation follows?
The estimates are 22% and 26%, and the difference indicates incorrect data.
All later samples must produce an estimate between 55% and 65%.
The two counts prove that exactly 480 households recycle weekly.
The estimates are 55% and 65%, corresponding to 440 and 520 households. They differ because random samples can vary.
For visits lasting 2 through 8 hours, a fitted line predicts y = 12 − 0.5x unused supply kits, where x is time in hours. One visit at x = 6 finished with 11 kits. Which interpretation is correct?
Each additional hour is associated with 0.5 fewer predicted kits. At 6 hours the line predicts 9, and the recorded value is 2 kits above the line.
Each additional hour predicts 0.5 more kits. At 6 hours the model predicts 15 kits.
Each additional hour predicts 0.5 fewer kits. At 6 hours the line predicts 9, and the recorded value is 2 below the line.
The model predicts 6 kits at 6 hours. It proves that a zero-hour visit always starts with 12 kits.
A scatterplot of weekly ice sales and outdoor pool visits shows that both tend to be high in the same weeks. Which conclusion does that information support?
Higher ice sales caused more pool visits.
More pool visits caused all the increase in ice purchases.
The variables are positively associated. Warmer weather could raise both, and the graph alone does not establish causation.
The variables must be independent because different organizations recorded them.
Sections you finish are checked off in the contents.