Lesson 9 of 11 · 20 min

Investigating sleep: a correlational study and Pearson's r

Design a correlational study of normal hours of sleep and another variable, then read the scatterplot, find and interpret Pearson's r and its p-value, and explain why correlation does not show cause.

What you will be able to do

  • Explain why the relationship between hours of sleep and another variable is investigated with a correlational design, and operationalise both variables.
  • Write null and alternative hypotheses for a correlational study.
  • Describe a correlation from a scatterplot and from Pearson's r: direction, strength and significance (the p-value against 0.05).
  • Explain why a correlation does not show cause, and identify a possible third variable.

Does scrolling on your phone before bed cost you sleep? It is a reasonable question, but you cannot answer it with an experiment that deprives people of sleep: QCAA's guidelines for school research do not allow investigations that involve sleep deprivation, and changing someone's sleep on purpose could harm them. Instead, you can measure people's normal sleep and their device use as they happen, and see whether the two go together. That is a correlational study.

Definition

Correlational research

A correlational study measures two variables as they naturally occur, in the same participants, to see whether they are related. The researcher manipulates nothing, so there is no independent or dependent variable: the two measured variables are sometimes called co-variables.

Operationalise the variables. "Device use" and "sleep" are too vague to measure. Say exactly how each will be measured.

  • Device use before bed: the number of minutes spent on a phone, tablet or computer in the 60 minutes before lights out, recorded each night in a sleep diary (or from the phone's screen-time report) and averaged over 7 nights.
  • Hours of sleep: the time from falling asleep to waking, in hours, recorded each morning in the diary (or by a wearable device) and averaged over the same 7 nights.

Averaging over a week makes each measure more reliable than one night, and measuring normal hours of sleep keeps the study ethical.

Key idea

Hypotheses for a correlation

Null hypothesis (H₀): There is no correlation between the average minutes of device use in the hour before lights out and the average hours of sleep per night of Year 11 students.

Alternative hypothesis (H₁): There is a correlation between the average minutes of device use in the hour before lights out and the average hours of sleep per night of Year 11 students.

H₁ can name a direction when there is a reason to expect one: "…there is a negative correlation…".

Interactive

Who is in the sample?

A convenience sample of your own class is easy, but is it like the whole school? Draw convenience, random and stratified samples and compare them.

The data. Twelve Year 11 students kept a sleep diary for a week. Each point on the scatterplot is one student: their average device use before lights out (across) and their average hours of sleep (up).

Average device use in the hour before lights out, and average hours of sleep, for 12 students over 7 nights
StudentDevice use (min)Sleep (h)
A08.1
B108.6
C157.4
D208.3
E207.6
F308.0
G307.1
H407.5
I457.9
J506.8
K607.2
L606.6

Device use before bed and hours of sleep, with the least-squares trendline. r=−0.71r = -0.71, n=12n = 12, p=0.01p = 0.01.

Reading the scatterplot. The points run from top left to bottom right: students who used devices longer before bed tended to sleep less. That is a negative correlation. The points are scattered around the line rather than on it, so the relationship is far from perfect: student I used devices for 45 minutes and still slept 7.9 hours.

Definition

Pearson's correlation coefficient, r

A number from −1-1 to +1+1 that measures how closely the points fit a straight line.

  • The sign gives the direction: positive (both rise together) or negative (one rises as the other falls).
  • The size (ignoring the sign) gives the strength: ±1\pm 1 is a perfect straight line; 00 is no linear relationship.

Calculate it with technology: =CORREL(A2:A13, B2:B13) in a spreadsheet, or the statistics mode of a calculator. For these data, r=−0.71r = -0.71.

A common guide to the strength of r (the same scale as General Mathematics)
Size of r (ignoring the sign)Strength
0.75 to 1Strong
0.5 to less than 0.75Moderate
0.25 to less than 0.5Weak
Less than 0.25No correlation (no linear relationship)

For the 12 students, r=−0.71r = -0.71. Describe the correlation.

Choose one answer

Is it significant? Twelve students is a small sample, and even if device use and sleep were unrelated in the whole population, a sample of 12 could show some correlation by chance. The p-value tells you how likely a correlation at least this strong would be if the null hypothesis were true. Software works it out from rr and nn (using t=rn−21−r2t = \frac{r\sqrt{n-2}}{\sqrt{1-r^2}} with n−2n - 2 degrees of freedom).

For these data p=0.0096p = 0.0096, which is less than 0.05. A correlation this strong would happen by chance less than 1% of the time if there were no real relationship, so the result is statistically significant: reject the null hypothesis. The data support a negative correlation between device use before bed and hours of sleep.

Key idea

Significance depends on the sample size

The same rr can be significant in a large sample and not in a small one. r=0.35r = 0.35 gives p=0.32p = 0.32 with 10 participants, but p=0.03p = 0.03 with 40. A weak correlation can be real; it just takes more data to be confident of it.

Interactive

Drag the points, watch r and p

This is a different class's sample data. Move a point and watch rr, the trendline and pp change. One unusual student can change rr a lot.

Worked example

Interpret a correlation that is not significant

  • Complex familiar
  • Technology active

Ten students recorded their exercise and sleep for a week.

Exercise (min per day)0152030304560607590
Sleep (h per night)7.27.86.97.58.17.07.97.48.37.6

Software gives r=0.38r = 0.38 and p=0.28p = 0.28.

(a) Describe the correlation. (b) State whether the null hypothesis should be rejected, and conclude. (c) The students conclude that "exercise has no effect on sleep". Evaluate this conclusion.

3 steps

Try it yourself first, then check the working one step at a time.

Method

  1. Direction from the sign, strength from the size of rr.

  2. Compare pp with 0.05: below, reject H₀; above, retain it.

  3. Never conclude cause; a non-significant result is not proof of no relationship.

Watch out for

  • Saying "there is no correlation" when r=0.38r = 0.38. There is a weak one in the sample; it is just not significant.

  • Concluding cause from a correlation, in either direction.

Correlation does not show cause. The data fit "device use reduces sleep", but a correlation alone cannot show that, for three reasons.

  • Nothing was manipulated. Students chose their own device use; no one was assigned to it.
  • Direction. The cause could run the other way: students who are not sleepy (because their body clock is later) may pick up their phones because they cannot sleep.
  • A third variable. Something else could cause both: a heavy homework load, stress or anxiety could keep a student on their laptop late and shorten their sleep.

Write "is associated with" or "is related to", never "causes". To test cause you need an experiment, which here would raise ethical problems.

Another study finds a moderate positive correlation between the minutes of exercise students do in a day and their hours of sleep. Which conclusion is justified?

Choose one answer

Common mistake

Three ways to lose marks with r

  • Calling a negative correlation weak because it is negative: −0.71-0.71 is stronger than +0.45+0.45.
  • Saying "significant" without the pp-value and the 0.05 level.
  • Concluding cause. A correlation never shows cause, however strong it is.

Tip

Evaluating your own correlational study

  • Reliability: average several nights; use a wearable or screen-time report rather than memory; give everyone the same diary and instructions.
  • Validity: self-reported sleep is often overestimated; students who know the aim may change their habits (participant expectations); a convenience sample of one class may not represent all students.
  • Sample size: with only 12 participants, one unusual student can change rr a lot, and a real but weak correlation may not be significant.

Summary

  • A correlational study measures two co-variables as they occur and manipulates nothing; normal hours of sleep are studied this way because deprivation would be unethical.

  • H₀: no correlation between the operationalised variables in the population; H₁: a correlation (or a positive or negative one).

  • Pearson's r runs from −1 to +1: the sign is the direction, the size the strength (≥ 0.75 strong, 0.5–0.75 moderate, 0.25–0.5 weak).

  • If p < 0.05, the correlation is significant: reject H₀. Larger samples make smaller correlations significant.

  • Correlation does not show cause: nothing was manipulated, the direction is unknown, and a third variable may explain both.

Finished the lesson?

Try the 2 quick checks above and the lesson completes itself, or mark it complete now.

Lesson 9 of 110% read

Save your progress with a free account