Sampling, data quality and outliers
In a minute
Statistics is only as good as its data. You learn how a stands in for a , how the way it is chosen can bias it, and how to handle missing or wrongly recorded values. You also find an with the rule and decide whether to keep it.
You'll be able to
- Use the words population, sample and , and tell from .
- Judge a data source, and say where , missing values and recording errors come from.
- Find outliers with the rule and decide whether each is a real value or an error.
- Describe simple random, systematic, stratified, quota and , and judge how well each represents the population.
- Find the interval for a systematic sample and the group sizes for a stratified sample.
Before you start3 quick questions on what this builds on
Populations, samples and types of data
A city wants the average commute time of its office workers. Timing all of them is not practical. Timing well-chosen workers gives a good estimate.
Rounding doesn't change the type. A wait recorded as minutes is still continuous. A dress size is discrete: it jumps from to .
Try it first, then reveal one step at a time.
- Ask of each one: is it counted or measured?
- (a) Visitors and (c) nights are counted, so they are discrete.
- (b) Temperature and (d) time are measured, so they are continuous, even if time is recorded to the nearest minute.
- (a) discrete; (b) continuous; (c) discrete; (d) continuous
Try it first, then reveal one step at a time.
- (a) The population is all bags that week. The sample is the bags opened.
- (b) The sample is random, so apply its fraction to the whole week: bags.
- Interpret: about bags contain such items. It is only an estimate, because another sample of would give a slightly different fraction.
Reliable data and sampling bias
Before you trust a data set, ask: who collected it, and do they gain from a particular result? How were the members chosen? Are any values missing or impossible?
Common causes: part of the population can't be reached, volunteers choose themselves, and people who don't reply may differ from those who do. A bigger sample chosen the same way is just as biased.
Missing data: never record a blank as . Leave it out and say how many values are missing. A genuine is real data and stays.
Recording errors: a slipped decimal point, a wrong unit or a typing slip. Look for impossible values. Correct one only if you know the true value; otherwise remove it and say so.
Try it first, then reveal one step at a time.
- (a) Leave out the two missing values. The is real (that bus was on time), so it stays.
- from buses, so the mean is minutes.
- (b) Entering s gives minutes. That claims two buses were on time when nobody knows, so the mean is too low.
Try it first, then reveal one step at a time.
- (a) Apply to all subscribers: .
- (b) Only subscribers who chose to reply are in the sample. Those annoyed by the price rise are more likely to reply than those who don't mind.
- So unhappy subscribers are over-represented, and is probably too high. Emailing more subscribers would not fix this bias.
Outliers
Some values sit far from the rest. To decide which count as unusual, everyone uses one rule, built from the .
In ordered data, about a quarter of the values lie below the lower quartile , and about three quarters below the upper quartile . Read them from your GDC's one-variable statistics.
The , , is the spread of the middle half of the data. It is in the formula booklet.
Learn this rule: it is not in the booklet. The two boundaries are the lower and upper fences. The inequalities are strict, so a value exactly on a fence is not an outlier.
Try it first, then reveal one step at a time.
- , so hours.
- Upper fence: . Lower fence: .
- , so hours is an outlier. , so hours is not an outlier.
- hours a day is unusual but possible, for example during a holiday. It is a valid value, so keep it.
- Fences and hours: is an outlier but a possible value, so keep it; is not an outlier.
The rule only flags a value; it doesn't make it wrong. A value that is possible in context is valid: keep it. An impossible value or clear slip is an error: correct it if you know the true value, or remove it and say so.
Try it first, then reveal one step at a time.
- (a) Run one-variable statistics on the times and write down and .
- (b) , so the fences are and .
- and , so both are outliers. Nobody swims m in s (the fastest take over s), so it is an error, probably a slipped decimal point. s is slow but possible, so it is valid.
- (c) The true value of the error can't be checked, so remove it and say so. Keep . The mean of the other times is s.
Sampling methods
Effectiveness: free of sampling bias, but it needs a full list, and by chance it can miss a small group.
Effectiveness: quickest and cheapest, but the most likely to be biased.
Effectiveness: quick and evenly spread, but biased if the list has a pattern that repeats every members.
Effectiveness: quick and needs no list, but not random, so the collector's choices can bias it.
Effectiveness: usually the most representative, but it needs each stratum's size and a list of its members.
Try it first, then reveal one step at a time.
- (a) The easiest people to reach: convenience sampling.
- (b) A random start, then every th in order: systematic sampling.
- (c) Fixed group sizes, but she chooses the people: quota sampling.
- (d) Random numbers from a full list: simple random sampling.
- (a) convenience; (b) systematic; (c) quota; (d) simple random
Try it first, then reveal one step at a time.
- (a) . Pick a whole number from to at random as the first ticket, then take every th ticket after it.
- (b) The tickets chosen are , , , From to is , a whole number of steps, so ticket is chosen.
- It is steps after the first, so it is the th ticket: .
Try it first, then reveal one step at a time.
- Shares: Science , Arts , Business .
- Rounding each to the nearest whole number gives , one short.
- Instead, round every share down, then give each missing place to the largest decimal part. Science has , so . Check: .
- If the faculties' students answer differently, each faculty is fairly represented. A simple random sample could under-represent one by chance.
Common mistakes
- Measuring the fences from the median. Start from for the lower fence and for the upper one.
- Calling a value on a fence an outlier. The inequalities are strict.
- Deleting an outlier just because it is one. Remove or correct only an impossible value or a known error.
- Giving a bias reason about the survey, not the sampling method. Say who the method leaves out or favours.
- Calling the first to arrive, or everyone who turns up one day, a systematic sample. That is convenience sampling.
- Recording a missing value as . Leave it out and divide by the number of real values.
Exam tips
- State or write down a sampling method: the name is enough. Describe one: say how members are chosen; for systematic, give and a random start in the first .
- Explain why a sample may be biased: tie the reason to the method and the context, such as who could never be chosen.
- Verify that a value is an outlier: write the fence with numbers in, then compare using or .
- Comment on an outlier: say whether it is possible in context, then whether to keep it.
Key takeaways
Practise this
10 past-paper questions on Sampling and data collection in the AI HL bank, each with a mark scheme and a worked explanation.
Also: Outliers and data interpretation (8)