1.2 Data, Sampling, and Variation in Data and Sampling

The taxonomy of data and the methods for collecting it: qualitative data that name a category against quantitative data that are always numbers, and within quantitative the split between discrete data that are counted and continuous data that are measured; pie charts and bar graphs for qualitative data and the two situations where a pie chart is not merely ugly but wrong; the four random sampling methods — simple random, stratified, cluster and systematic — with convenience sampling as the non-random alternative; and the difference between sampling error, which every sample has and a larger sample reduces, and sampling bias, which no sample size can cure.

Subject: Statistics · 72 slides · symbolic lesson

Open the interactive version of this deck

What this lesson covers

The lesson, slide by slide

1. Data, Sampling, and Variation

Title

Statistics · Section 1.2

What kind of numbers you have, how they were chosen, and how far to trust them

2. What this lesson gives you

Objectives

Section 1.1 gave you the words: population, sample, parameter, statistic, variable, data. This section asks the two questions that come before any calculation. What kind of data is this, and how was the sample chosen?

Those questions sound like bookkeeping, and they are where most wrong conclusions in statistics are born. A number that should not have been averaged, a pie chart that adds up to 150 percent, or a sample of two million that still gets the answer wrong: each of these is a failure of this section, not of arithmetic.

Keep one habit running through every slide: whenever a number appears, ask where it came from. Almost every idea in this lesson is an answer to that question.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 10 — the section opening

3. A statistic from five glasses of milk

Warm-up

Start with the book's own collaborative exercise from the end of Section 1.1, because it uses every word you need and hides a problem this lesson will name.

You want the average number of glasses of milk college students drink per day. Yesterday, in your English class, you asked five students how many glasses they drank the day before. The answers were 1, 0, 1, 3 and 4.

Discussion prompt

Name the population, the sample, the parameter, the statistic and the variable, and compute the statistic. Then say whether you would trust it to describe all college students, and why.

Notice also what kind of number the answers were. You counted glasses, so every answer was a whole number, and yet the statistic came out as 1.8. Hold that thought: it is the first trap of Part 1.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.1, p. 10 — Collaborative Exercise

4. What kind of data is it?

Section

Part 1

5. Qualitative data name a category

Concept

The first split is between data that name something and data that measure an amount. Data that name are called qualitative, or categorical: hair colour, blood type, ethnic group, the car a person drives, the street a person lives on.

qualitative data — Data that result from categorising or describing an attribute of a population. They are generally described by words or letters, and are also called categorical data.

The book's Example 1.8 records the colours of five students' backpacks: red, black, black, green and gray. Those five words are the data.

Figure (svg): Five backpacks drawn as coloured squares grouped by colour: one red, two black, one green and one gray, with the count under each group.

The colours themselves are the data. The counts underneath are something you make from them, and they are the only arithmetic this variable allows.

You can still do arithmetic, but only on the counts you build from the words: two blacks, one of everything else. What you cannot do is average the data themselves. There is no average hair colour and no average blood type, and that is the working test for this category.

Try It 1.8 is the same situation with houses: white, yellow, white, red and white. Colours, so qualitative, whatever the number of houses happens to be.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 10-12 — the definition, Example 1.8 and Try It 1.8

6. Quantitative data are counts or measurements

Concept

Quantitative data are always numbers, and they are numbers of a particular kind: the result of counting or measuring something. Amount of money, pulse rate, weight, the number of people in your town.

Inside that branch there is a second split, and it depends on how the number was produced.

quantitative discrete data — Data that are the result of counting. They can take only certain values, such as 0, 1, 2 and 3 phone calls in a day.

quantitative continuous data — Data that are not restricted to counting numbers and may include fractions, decimals or irrational values. They usually come from measuring lengths, weights or times.

Figure (svg): Two number lines. The top one, for the number of phone calls in a day, has open circles only at 0, 1, 2, 3, 4 and 5 with nothing possible in between. The bottom one, for call lengths in minutes, is a solid bar with the values 2.4, 7.5 and 11.0 marked on it.

Look at the space between the marks. A count can only land on the circles; a measurement can land anywhere on the bar, and a better instrument simply records more decimal places.

The picture is the whole distinction. When you count phone calls, 2.4 calls is impossible, so the possible values have gaps between them. When you time the calls, 2.4 minutes is perfectly ordinary, and so is every value in between, limited only by how precise your stopwatch is.

The book's hint for spotting discrete data is worth keeping: they often start with the words the number of. The number of calls, the number of books, the number of machines.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 10 — quantitative discrete and continuous data

7. Reading the letters for a variable and its values

Notation

Statistics uses capital and lowercase letters for two different things, and the book introduces the lowercase ones in this section. Take the backpack example apart before using the letters on your own.

Annotate

On: \( X = \text{books carried}, \quad x_1 = 3, \; x_2 = 3, \; x_3 = 4 \)

  • The variable: the characteristic you record for each person. It is a description of a question, not a number, and it is the same for every student in the sample.
  • One data value, the answer one student gave. The book says lowercase letters such as x or y are generally used for data values.
  • Subscripts that say which student. The first student carried 3 books, the second also 3, the third 4, so different people can have equal values.
  • Books, so every data value will be a whole number. That is how you can tell, from the definition of X alone, that these data will be quantitative discrete.

Keep the capital and the lowercase apart in your own writing. The variable is the question you ask; the data are the answers you get back, and there is one datum per person.

You will see the lowercase letters again in the very next chapter, inside formulas for the mean and the standard deviation, where each x stands for one value from the list.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 10 — lowercase letters for data values

8. Examples 1.5 and 1.6: the same five backpacks, two kinds of number

Worked example

The book samples five students twice over: once counting the books in each backpack and once weighing the backpacks. Same students, same backpacks, and the two variables land on different branches.

\[ \text{books: } 3, 3, 4, 2, 1 \qquad \text{weights (lb): } 6.2, 7, 6.8, 9.1, 4.3 \]

Ask how each number was produced

Why: the books were counted one by one, so only whole numbers can come out.

\[ \text{books} \;\Rightarrow\; \text{counted} \;\Rightarrow\; \text{discrete} \]

Do the same for the weights

Why: a scale produced them, and it could have printed 6.25 or 6.253 with a finer mechanism.

\[ \text{weights} \;\Rightarrow\; \text{measured} \;\Rightarrow\; \text{continuous} \]

Figure (svg): Two dot plots for the same five students. The top plot, number of books, has dots only at whole numbers: one at 1, one at 2, two stacked at 3 and one at 4, with the mean 2.6 marked by a yellow triangle. The bottom plot, backpack weight in pounds, has dots at 4.3, 6.2, 6.8, 7 and 9.1 with the mean 6.68 marked.

The books pile up on whole numbers because they were counted; the weights sit wherever the scale put them. The mean of the counts is 2.6 all the same, which is a summary, not a book count.

The book adds a telling remark: backpacks carrying three books can have different weights. Two students carry three books each, and nothing forces their bags to weigh the same, which is one more sign that the two variables are genuinely different questions.

Compute the mean of the counts

Why: add the five counts, then divide by five.

\[ \dfrac{3 + 3 + 4 + 2 + 1}{5} = \dfrac{13}{5} = 2.6 \text{ books} \]

Compute the mean of the weights

Why: same two moves.

\[ \dfrac{6.2 + 7 + 6.8 + 9.1 + 4.3}{5} = \dfrac{33.4}{5} = 6.68 \text{ lb} \]

Check the classification against the mean

Why: a mean of 2.6 books does not make the books continuous, because no single backpack can hold 2.6 books; the data are the five counts, and every one of them is whole.

\[ 3, 3, 4, 2, 1 \in \{0, 1, 2, \dots\} \]

That is the test to trust: look at the individual values and the process that made them, never at a summary computed afterwards. Summaries of discrete data are routinely fractions.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 11 — Examples 1.5 and 1.6

9. Two questions, asked in order

Picture it

Everything in Part 1 fits into a tree with two questions. Put them in this order every time, because the second one only makes sense after the first.

Figure (svg): A tree diagram. The top box reads what was recorded. Two branches lead to a box labelled a name or category, qualitative, and a box labelled a number, quantitative. From the quantitative box two further branches lead to counted, discrete, and measured, continuous. Examples are written under each leaf.

The second question is only ever asked on the right-hand branch. Discrete and continuous are two kinds of quantitative data, not two rivals of qualitative data.

The first question is about what the recorded value is. If it is a word or a label, you are done: qualitative, and there is no further split. If it is a number that stands for an amount, it is quantitative and you ask the second question.

The second question is about the process. Counting gives discrete data; measuring gives continuous data. It is about how the number was produced, so you can often answer it before you have seen a single value.

Two warnings follow straight from the shape of the tree. Asking whether qualitative data are discrete or continuous is a question with no answer. And a number is not automatically an amount: a student ID or a postcode is a label written in digits, which puts it on the left-hand branch.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 10-12 — the categories of data

10. Example 1.7: one shopping trip, three kinds of data

Worked example

Real situations rarely hand you one variable. The book's supermarket trip contains all three kinds at once, and the skill is to pull them apart.

\[ \text{3 soups (19 oz, 14.1 oz, 19 oz), 2 nut packs, 4 vegetables, 2 desserts} \]

Find what was counted

Why: three cans, two packages, four kinds of vegetable and two desserts are all answers to how many.

\[ 3, \; 2, \; 4, \; 2 \;\Rightarrow\; \text{quantitative discrete} \]

Find what was measured

Why: the soup weights came off a scale, and 14.1 ounces is not a number anyone counts to.

\[ 19, \; 14.1, \; 19 \text{ oz} \;\Rightarrow\; \text{quantitative continuous} \]

Find what was named

Why: tomato bisque, lentil and Italian wedding are kinds of soup; walnuts and peanuts are kinds of nut.

\[ \text{bisque, lentil, walnuts, broccoli} \;\Rightarrow\; \text{qualitative} \]

The book calls this one possible solution and invites you to find more. The dessert sizes, 16 ounces of ice cream and 32 ounces of cookies, are another continuous data set; the names of the four vegetables are another qualitative one.

Check each set against the tree

Why: every value in the first set is a whole count, the second set contains a decimal a count could never produce, and the third set contains no numbers at all, so each sits on exactly one branch.

\[ \text{words} \to \text{qualitative}, \quad \text{counts} \to \text{discrete}, \quad \text{weights} \to \text{continuous} \]

One shopping trip, three data types. The type belongs to the variable you choose to record, never to the situation or the objects.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 11 — Example 1.7

11. Try It 1.7: the purchase manager's order

Worked example

The matching Try It works the same way on a less friendly list. A purchase manager buys two types of nails, one type of oil and four types of screws, each with a size.

\[ \text{nails: 2 kg box, 3 kg roofing; \; oil: 4 L machine} \]

\[ \text{screws: 3 kg wood, 5 kg machine, 1 kg set, 2 kg socket} \]

The counts of types

Why: two types of nails, one of oil, four of screws: each is an answer to how many types.

\[ 2, \; 1, \; 4 \;\Rightarrow\; \text{quantitative discrete} \]

The amounts bought

Why: kilograms and litres are read off a scale or a measuring vessel.

\[ 2, 3, 4, 3, 5, 1, 2 \text{ (kg or L)} \;\Rightarrow\; \text{continuous} \]

The amounts happen to be whole numbers here, and that is the trap in this Try It. A box weighed at exactly 2 kilograms is still a measurement; a finer scale could have said 2.03. Whole-looking values do not make measured data discrete.

The names

Why: box nails, roofing nails, machine oil, wood screws: labels for kinds of item.

\[ \text{box, roofing, machine, wood, set, socket} \;\Rightarrow\; \text{qualitative} \]

Check the whole-number weights

Why: ask the process question, not the appearance question: were the kilograms counted or measured? Measured, so continuous, even though every printed value is whole.

\[ 2 \text{ kg} = 2.0 \text{ kg} = 2.00 \text{ kg} \]

A measurement is always rounded to the precision someone chose. Writing it as 2.0 or 2.00 changes nothing about the object and shows the hidden decimal that makes it continuous.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 11-12 — Try It 1.7

12. Example 1.9: sort the eight variables

Sorting

The book's Example 1.9 is a sorting exercise, so do it as one. For each variable, ask the two questions in order: a name or an amount, and then counted or measured.

Sort into buckets

Sort each variable into its data type.

Qualitative
the type of car you drive; the type of calculator you use
Quantitative discrete
the number of pairs of shoes you own; the number of classes you take per school year; number of correct answers on a quiz
Quantitative continuous
the distance from your home to the nearest grocery store; weights of dogs at an animal shelter; IQ scores
ql
The type of car and the type of calculator are names of makes or models. There is no amount here and no sensible average.
qd
Pairs of shoes, classes and correct answers are all counted, and all three begin with the words the number of, which is the book's own hint.
qc
A distance and a weight are measured and could be recorded to any precision. The book files IQ scores here too, while warning that this one may cause discussion.

The book's answer is items a, d and g discrete, c, f and h continuous, and b and e qualitative. If you disagreed only on IQ, you are in good company, and the next slide argues it out.

Notice how the hint does most of the work for the discrete bucket. Three of the eight items begin with the number of, and exactly those three are discrete. It is a hint, though, not a rule: the number of kilograms of nails is a measurement dressed up in counting words.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 12 — Example 1.9

13. Why the book says IQ scores may cause discussion

Socratic

The book files IQ scores as continuous and adds, in brackets, that this may cause some discussion. A classification that needs defending deserves an argument rather than a memory.

Discussion prompt

IQ scores are reported as whole numbers such as 98, 104 or 117. Give the argument for calling them discrete, the argument for calling them continuous, and say which you find stronger and why.

This is the useful lesson from a messy example: classification is a judgement about the variable, and sometimes reasonable people make it differently. What matters is that you can state the question you asked and why the answer follows.

In practice the stakes are low for this particular pair. Discrete data with many possible values, like IQ scores or incomes in dollars, are usually analysed with the same tools as continuous data, which is why the book does not fight hard for either side.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 12 — Example 1.9(h)

14. Trap: it is written in digits, so it must be quantitative

Trap

The trap

A tempting shortcut is to classify by appearance: if the column is full of digits, call it quantitative.

\[ \text{IDs: } 20417, \; 20982, \; 21150 \]

By that rule, student ID numbers are quantitative discrete, and you could report the average ID of the class.

\[ \bar{x} = 20849.67 \]

The arithmetic is flawless and the answer means nothing at all.

The fix

Ask what the digits stand for. An ID names a person; it measures nothing and counts nothing.

\[ \text{ID} \;\Rightarrow\; \text{a label} \]

\[ \text{a label} \;\Rightarrow\; \text{qualitative} \]

The same goes for postcodes, phone numbers and jersey numbers. The test is whether an average would mean anything, and for a label it never does.

The reverse slip exists too: 2.6 books is a mean, not a datum, so it cannot turn counts into continuous data.

15. Collected as numbers, reported as categories

Concept

The book adds a note that catches people out: you may collect data as numbers and report them categorically. Its example is quiz scores, recorded as numbers throughout the term and reported at the end as A, B, C, D or F.

Figure (svg): A number line of term averages from 50 to 100 cut into five coloured bands labelled F, D, C, B and A. Four students' averages, 58, 74, 81 and 89, are dotted on the line with arrows down to the letter each one becomes.

Once the number is replaced by a letter, 81 and 89 are the same datum. The type of data is decided by what is reported, not by how it was first collected.

Follow one student through. An average of 81 goes in as a number, quantitative and measured to whatever precision the teacher kept. It comes out as the letter B, which is a category, and it now sits in the same category as a student who averaged 89.

\[ 81 \;\mapsto\; B, \qquad 89 \;\mapsto\; B \]

So the data type belongs to the data as reported, not to their history. The letter grades keep an order, since an A is better than a B, but they have thrown away the distances, and you can no longer say by how much one student beat another.

The ordering is a hint of something the next section makes precise. Some qualitative data can be ranked and some cannot, and Section 1.3 gives those two kinds the names ordinal and nominal.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 12 — the Note on collecting and reporting data

16. Example 1.10 and Try It 1.10: what kind of data does the graph show?

Worked example

Sometimes the data arrive already drawn. The book asks you to classify the data behind two graphs, and the graph's own layout is the strongest clue.

Look at the pie chart of class standing

Why: its wedges are labelled first-year, sophomore, junior and senior, which are names of categories.

\[ \text{first-year, sophomore, junior, senior} \;\Rightarrow\; \text{qualitative} \]

Figure (svg): A pie chart with four wedges: first-year, about two thirds of the circle; sophomore, about a sixth; junior, about a ninth; senior, a thin sliver.

Each wedge names a category. There is no number line anywhere in this graph, which is the quickest sign that the data are qualitative.

Now the histogram of credit hours

Why: its horizontal axis is a number line from 10 to 25, and the bars sit over intervals such as 10 to less than 13.

\[ \text{credit hours completed} \;\Rightarrow\; \text{an amount} \;\Rightarrow\; \text{quantitative} \]

Figure (svg): A histogram of credit hours completed per student with five touching bars over the intervals 10 to 13, 13 to 16, 16 to 19, 19 to 22 and 22 to 25. The heights are roughly 250, 575, 730, 630 and 250 students.

Here the horizontal axis is a number line. The bars sit over intervals of an amount, and the amount is a count of credit hours.

Ask counted or measured

Why: the registrar records how many credit hours each student completed, which is a count of whole hours.

\[ \text{the number of credit hours} \;\Rightarrow\; \text{quantitative discrete} \]

The class boundaries are written as 10 to less than 13 and so on, which looks like continuous-data language. That is a convention for building histograms, covered in Chapter 2, so that no value can fall on the edge of two bars; it does not change what the underlying values are.

Check with the average test

Why: an average class standing means nothing, while an average number of credit hours is a sensible figure a registrar would report, which confirms qualitative for the first and quantitative for the second.

\[ \text{mean standing: meaningless} \qquad \text{mean hours: sensible} \]

The graph type is a hint, not a proof: a bar graph could show either kind of data. The test is always what the recorded value is.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 13 — Example 1.10 and Try It 1.10

17. Try Its 1.5, 1.6, 1.8 and 1.9: fill in the type

Comparison

Four of the book's short Try Its in one table. The data column shows what was actually recorded; decide the type, and for quantitative data say whether it is discrete or continuous.

Comparison matrix

what was recordedsample valuesqualitative or quantitative?discrete or continuous?
machines in each of five gyms12, 15, 10, 22, 20quantitativediscrete
lawn areas of five houses144, 160, 190, 180, 210 sq ftquantitativecontinuous
colours of five houseswhite, yellow, white, red, whitequalitativeneither
cars in a parking lotone count per lotquantitativediscrete

The lawn areas are the row to slow down on. All five values are whole numbers of square feet, yet the areas were measured, and a more careful survey could have reported 144.3. The process decides, not the digits.

The colours row has neither in its last column, and that is not a blank you forgot. The discrete or continuous question does not apply to qualitative data at all, exactly as the tree showed.

The two counting rows both answer a question beginning with how many, which is the book's hint doing its job. Machines and cars come in whole units, so the only possible values are whole numbers.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 11-12 — Try Its 1.5, 1.6, 1.8 and 1.9

18. Showing qualitative data

Section

Part 2

19. Table 1.3: turning counts into percents

Worked example

The book compares full-time and part-time students at two colleges, De Anza and Foothill, for one spring quarter. The counts come first; the percentages are what make a comparison possible.

De AnzaFoothill
full-time9,2004,059
part-time13,29610,124
total22,49614,183

Check the totals

Why: each college's two counts should add to its total.

\[ 9200 + 13296 = 22496, \qquad 4059 + 10124 = 14183 \]

Divide each count by its own college's total

Why: De Anza first; this proportion is the relative frequency.

\[ \dfrac{9200}{22496} = 0.40896 \approx 40.9\%, \qquad \dfrac{13296}{22496} = 0.59104 \approx 59.1\% \]

Now Foothill, over Foothill's total

Why: never over De Anza's, and never over the combined total.

\[ \dfrac{4059}{14183} = 0.28619 \approx 28.6\%, \qquad \dfrac{10124}{14183} = 0.71381 \approx 71.4\% \]

Figure (svg): Two horizontal bars, each the full width representing 100 percent. De Anza's bar is split 40.9 percent full-time and 59.1 percent part-time; Foothill's is split 28.6 percent full-time and 71.4 percent part-time. A dashed line marks the 50 percent point.

Stretching both colleges to the same length is what the percent column does. Now the orange part can be compared directly, and Foothill's is plainly the longer.

In counts, De Anza has more part-time students, 13,296 against 10,124. In percents, Foothill is the more part-time college by a wide margin, 71.4 against 59.1. Both statements are true; only the second compares like with like, because the colleges are different sizes.

Check that each college's percents add to 100

Why: each student is either full-time or part-time, never both, so the shares must fill the whole.

\[ 40.9 + 59.1 = 100, \qquad 28.6 + 71.4 = 100 \]

That check will matter in a few slides. Percentages that add to exactly 100 are what permit a pie chart, and two of the book's tables fail it.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 13-14 — Table 1.3

20. Error analysis: four readings of the two-college table

Error analysis

Four students read Table 1.3 and each writes one sentence. Two of the sentences are sound and two contain a classic misreading. Find the faults before you open the notes.

Annotate

On: \( 13296 > 10124, \quad 59.1\% < 71.4\%, \quad 40.9\% + 28.6\% = 69.5\% \)

  • True as a count, and the writer concluded that De Anza is the more part-time college. That is the error: De Anza simply has more students of every kind. Comparing counts across groups of different sizes answers a question about size, not about share.
  • Sound. Each percentage is a share of its own college, so the comparison is fair, and it says a larger fraction of Foothill's students are part-time.
  • A student added the two colleges' full-time percentages and called it the full-time rate for both colleges together. The two percents have different denominators, so adding them is meaningless; the true combined rate is 13,259 out of 36,679, about 36.1 percent.
  • Another student wrote that a bigger college gives more trustworthy percents. For a complete count of each college there is no sampling at all, so both percents are exact; size would matter only if these were samples.

\[ \dfrac{9200 + 4059}{22496 + 14183} = \dfrac{13259}{36679} \approx 36.1\% \]

The line above is the honest way to pool the colleges: add the counts, then divide by the combined total. Adding two percentages that describe different wholes produces a number that describes no whole at all.

Both errors come from the same place. A percentage only means something alongside its denominator, so before comparing or combining two percents, say out loud what each one is a percent of.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 13-14 — Table 1.3 and the discussion of different totals

21. A pie chart turns each share into an angle

Concept

The book names two graphs for qualitative data. In a pie chart, each category is a wedge of a circle whose size is proportional to the percent of individuals in that category.

\[ \text{wedge angle} = 360^\circ \times \text{proportion in the category} \]

Proportional to the percent means the angle is the share of 360 degrees. Compute the four wedges for Table 1.3, keeping the unrounded proportions.

\[ 0.40896 \times 360^\circ \approx 147^\circ, \qquad 0.59104 \times 360^\circ \approx 213^\circ \]

\[ 0.28619 \times 360^\circ \approx 103^\circ, \qquad 0.71381 \times 360^\circ \approx 257^\circ \]

Figure (svg): Two pie charts side by side. De Anza's has a full-time wedge of about 147 degrees and a part-time wedge of about 213 degrees; Foothill's has a full-time wedge of about 103 degrees and a part-time wedge of about 257 degrees.

The angles are the percentages in disguise: 360 degrees times the share. The difference between the colleges is there, but you have to compare two angles in two different circles to see it.

Now you can see what a pie chart assumes. The whole circle stands for one whole group, every individual lands in exactly one wedge, and the wedges together use up all 360 degrees. Two conditions follow: the categories must not overlap, and none may be missing.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 14 — pie charts, Figure 1.5

22. The same shares as bars

Picture it

The other display the book names is the bar graph, in which each bar's length is proportional to the number or percent in its category. Bars may be vertical or horizontal.

Figure (svg): A bar graph with two groups, full-time and part-time. In each group a blue bar for De Anza and an orange bar for Foothill stand side by side on a percent scale from 0 to 80: full-time 40.9 and 28.6, part-time 59.1 and 71.4.

With the bars on one shared scale, the comparison is a glance: the orange part-time bar towers over the blue one, and the full-time pair runs the other way.

The book asks you to look at its two versions of the college data and decide which displays the comparison better. With bars on one shared axis, your eye compares two heights measured against the same gridlines; with pies, it compares two angles in two separate circles, which people judge far less accurately.

Predict first

A reader wants to know which college has the larger share of part-time students, and by roughly how much. Which display answers fastest?

  • The grouped bar graph, because both part-time bars stand on one scale
  • The two pie charts, because each shows a college's complete picture
  • The count table, because it has the exact numbers
  • They are equally fast, since they carry the same information

Correct: The grouped bar graph, because both part-time bars stand on one scale

Why: The bars put the 59.1 and the 71.4 side by side against the same gridlines, so the gap of about 12 points is visible at once. The pies carry the same shares but hide them in angles, and the count table invites the misleading comparison of 13,296 against 10,124.

There are no strict rules here, and the book says so. The best graph depends on the data and on the question you are asking, which is why it recommends drawing several and choosing the one that makes your comparison easiest to see.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 14-15 — bar graphs, Figures 1.5 and 1.6

23. Table 1.4: percentages that add to 150.5

Worked example

The first case where a pie chart is not just less clear but wrong. The book lists three characteristics of De Anza students in the same quarter.

characteristicpercent
full-time students40.9%
intend to transfer to a 4-year institution48.6%
under age 2561.0%

Add the percentages

Why: a pie chart needs them to total exactly 100.

\[ 40.9 + 48.6 + 61.0 = 150.5 \]

Compare with 100

Why: the total overshoots by half as much again.

\[ 150.5 - 100 = 50.5 \]

Find the reason

Why: one student can be full-time, intend to transfer and be under 25 all at once, so the same person is counted in up to three rows.

\[ \text{categories overlap} \;\Rightarrow\; \text{one person, several rows} \]

Figure (svg): A horizontal bar graph of three De Anza characteristics: full-time students 40.9 percent, intend to transfer 48.6 percent, under age 25 61.0 percent. A dashed line at 100 percent is marked, and a stacked strip below shows the three bars laid end to end reaching 150.5, well past the 100 line.

Each bar is fine on its own. Laid end to end they run half as far again past the 100 line, because one student can be full-time, planning to transfer and under 25 all at once.

Choose the legal display

Why: a bar graph has no whole to fill, so each bar can show its own percent independently.

\[ \text{total} > 100\% \;\Rightarrow\; \text{bar graph, not a pie} \]

Check that the diagnosis is overlap, not arithmetic

Why: each percent is correct on its own; what fails is the assumption that every student sits in exactly one category, and no correction to the numbers can repair that.

\[ \text{each row: a correct share of all students} \]

The same verdict holds in the other direction. The book adds that a pie could not be used if the percentages added to less than 100, which is the next table.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 15 — Table 1.4 and Figure 1.7

24. Table 1.5: percentages that add to only 90.4

Worked example

The second failure. The book's table of De Anza students by ethnicity lists seven groups but leaves out the Other/Unknown category: students who fit no listed category or declined to answer.

groupfrequencypercent
Asian8,79436.1%
Black1,4125.8%
Filipino1,2985.3%
Hispanic/Latino4,18017.1%
Native American1460.6%
Pacific Islander2361.0%
White5,97824.5%

Add the frequencies

Why: the seven listed counts.

\[ 8794 + 1412 + 1298 + 4180 + 146 + 236 + 5978 = 22044 \]

Compare with the number of students

Why: the college had 24,382 students that term.

\[ 24382 - 22044 = 2338 \text{ students unlisted} \]

Turn the gap into a percent

Why: divide by the full total, not by the listed total.

\[ \dfrac{2338}{24382} = 0.0959 \approx 9.6\% \]

Confirm the listed share

Why: the same division for the students who were listed.

\[ \dfrac{22044}{24382} = 0.9041 \approx 90.4\% \]

Figure (svg): A horizontal bar graph of De Anza students by ethnicity in alphabetical order: Asian 36.1 percent, Black 5.8, Filipino 5.3, Hispanic/Latino 17.1, Native American 0.6, Pacific Islander 1.0, White 24.5, and a red bar for Other/Unknown at 9.6 percent.

The red bar is the category the table left out. It is bigger than three of the named groups put together, which is exactly why leaving it out distorts the picture.

The missing category is not small. At 9.6 percent it is bigger than the Black, Filipino, Native American and Pacific Islander groups taken separately, so a chart that silently leaves it out changes the story the data tell.

Check the percents against the counts

Why: each printed percent is its count over 24,382; recomputing them in node gives 36.07, 5.79, 5.32, 17.14, 0.60, 0.97 and 24.52, which round to the book's figures.

\[ 36.1 + 5.8 + 5.3 + 17.1 + 0.6 + 1.0 + 24.5 = 90.4 \]

So the verdict is a bar graph while the category is missing. Unlike the overlap in Table 1.4, this failure is repairable: put Other/Unknown back and the percents total 100, and a pie becomes legal again.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 15-16 — Table 1.5 and Figures 1.8 and 1.9

25. Order the categories for a Pareto chart

Ranking

The book calls its alphabetical bar graph difficult to understand visually and fixes it with a Pareto chart, in which the bars are sorted from largest to smallest. Build the order yourself before seeing it drawn.

Put in order

Put the eight ethnicity categories in Pareto order, largest first.

  1. Asian, 36.1%
  2. White, 24.5%
  3. Hispanic/Latino, 17.1%
  4. Other/Unknown, 9.6%
  5. Black, 5.8%
  6. Filipino, 5.3%
  7. Pacific Islander, 1.0%
  8. Native American, 0.6%

Why: Sorting by percent gives Asian at 36.1, White at 24.5, Hispanic or Latino at 17.1, Other or Unknown at 9.6, Black at 5.8, Filipino at 5.3, Pacific Islander at 1.0 and Native American at 0.6. Because every percent shares the same denominator of 24,382 students, sorting by count gives exactly the same order.

The restored Other/Unknown category lands fourth, ahead of four named groups. That placement is the book's point about why the missing category matters: in sorted order you can see at once how much of the picture it holds.

If two of your bars were in the wrong order, check which column you sorted on. The percent column and the frequency column must agree, since each percent is its frequency over the same total, and if they seem not to, one of the numbers has been misread.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 16 — the Pareto chart

26. The Pareto chart: sort the bars

Concept

Pareto chart — A bar graph whose bars are sorted into order by category size, from largest to smallest.

Figure (svg): The same eight ethnicity bars sorted from longest to shortest: Asian 36.1, White 24.5, Hispanic/Latino 17.1, Other/Unknown 9.6, Black 5.8, Filipino 5.3, Pacific Islander 1.0, Native American 0.6.

Sorted, the bars form a staircase. The three biggest groups make up more than three quarters of the students, a fact the alphabetical version hides.

Sorting changes nothing about the data and a great deal about the reading. In alphabetical order your eye has to hunt for the big categories; in sorted order the bars descend like a staircase, and the ranking is the first thing you see.

\[ 36.1 + 24.5 + 17.1 = 77.7 \]

That line is the kind of statement the sorted chart makes easy. The three largest groups make up 77.7 percent of De Anza's students, more than three quarters, a fact you could only assemble from the alphabetical chart by searching for the three longest bars.

Alphabetical order is not wrong, just unhelpful. It is the right choice when readers need to look up a particular category quickly, as in a reference table, and the wrong choice when the question is which categories dominate.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 16-17 — Figures 1.9 and 1.10

27. A pie chart with nothing missing, sorted by size

Concept

With Other/Unknown restored, the ethnicity percents add to 100, so the pie chart is finally legal. The book draws it twice, once with the wedges in alphabetical order and once sorted by size.

\[ 90.4\% + 9.6\% = 100\% \]

Figure (svg): A pie chart of all eight ethnicity categories, sorted by size clockwise from the top: Asian, White, Hispanic/Latino, Other/Unknown, Black, Filipino, Pacific Islander and Native American, the last two as thin slivers.

With Other/Unknown restored the wedges total 100 percent, so the circle is honest. Sorting the wedges by size, as the book's Figure 1.11(b) does, makes the order readable at a glance.

The book calls the sorted version the more visually informative graph, for the same reason the Pareto chart beats the alphabetical bars. Going clockwise from the top, each wedge is smaller than the one before, so size and position tell the same story.

Notice what the smallest wedges look like. Native American students at 0.6 percent get about 2 degrees of the circle, too thin to label inside the pie, which is why the figure labels them beside it. When a category is that small, a bar graph shows it more honestly than a pie.

\[ 0.006 \times 360^\circ \approx 2.2^\circ \]

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 17 — Figure 1.11

28. Estimate the Asian wedge before computing it

Estimation

A quick sense of how percents become angles is worth more than the formula. The Asian group is 36.1 percent of De Anza's students. Estimate the angle of its wedge in the sorted pie before you multiply.

Predict first

Roughly how many degrees does a 36.1 percent wedge take?

  • about 36 degrees
  • about 90 degrees
  • about 130 degrees
  • about 180 degrees

Correct: about 130 degrees

Why: A quarter of the circle is 90 degrees and a third is 120 degrees, and 36.1 percent is a little more than a third, so the wedge should be a little more than 120. The exact figure is 0.361 times 360, which is 130.0 degrees.

\[ 0.361 \times 360^\circ = 129.96^\circ \approx 130^\circ \]

The landmarks make estimation fast. Half the circle is 180 degrees, a quarter is 90, a third is 120, and a tenth is 36. Any share can be placed between two of these before you touch a calculator.

The option of 36 degrees is the classic slip: it reads the percent as if it were already an angle. It would be right only if the circle had 100 degrees, and a wedge that thin would look like a tenth of the pie, not a third.

29. A streaming survey whose percentages total 214

Elimination

A new situation with the same logic. A survey asks people which of five streaming services they subscribe to, and many subscribe to several, so the five percentages add to 214.

Eliminate the wrong options

Which display is appropriate for the five subscription percentages?

  • a. A pie chart of the five services
  • b. A pie chart after dividing every percent by 2.14
  • c. A bar graph with one bar per service
  • d. A pie chart with an extra wedge for people with several services

Survives elimination: c

Why: Each bar can show its own service's percent without any whole to fill, which is exactly what overlapping categories need. The rescaled pie is the tempting wrong answer: it produces a legal-looking circle, but a wedge of 20 percent would claim that a fifth of people use only that service, which nobody measured.

The rescaling option deserves a moment, because it is what people do when a software package refuses to draw their pie. Dividing by 2.14 makes the numbers fit the shape without making them true.

The general rule from Tables 1.4 and 1.5 covers this case with no new thinking: if the percentages do not add to exactly 100, use a bar graph. Over 100 means overlap, under 100 means something is missing, and neither can be drawn as one honest circle.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 15 — percentages that add to more or less than 100 percent

30. Four ways to sample at random

Section

Part 3

31. Why sample, and what random means here

Concept

Gathering information about an entire population often costs too much or is virtually impossible. So you study a sample instead, and the sample is only useful if it has the same characteristics as the population it stands for.

You cannot check that directly, since you do not know the population's characteristics; that is why you are sampling. What you can control is the method, and the book's answer is to let chance do the choosing.

random sampling — A method in which each member of the population initially has an equal chance of being selected for the sample.

\[ P(\text{a given member is chosen}) \text{ is the same for everyone} \]

The word initially matters. Once some people have been drawn, the remaining chances change slightly, and that is allowed. What is not allowed is a method under which some members of the population never had a chance at all, or had a worse one from the start.

The book describes four random methods, each with pros and cons, and one non-random method to contrast them with. The next slides take them one at a time, starting with the simplest to describe.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 17-18 — Sampling

32. A simple random sample: every group equally likely

Concept

The simple random sample asks for more than equal chances for individuals. It asks that every possible group of the chosen size be equally likely to be the sample.

simple random sample — A sample of size n chosen so that any group of n individuals is equally likely to be chosen as any other group of n individuals.

The book's example is Lisa, who wants three study partners from the 31 other members of her pre-calculus class. Count how many different groups of three she could end up with: choose the first, then the second, then the third, and divide out the six orders in which the same three could be drawn.

\[ \dfrac{31 \times 30 \times 29}{3 \times 2 \times 1} = \dfrac{26970}{6} \]

\[ = 4495 \text{ possible groups} \]

A simple random sample gives every one of those 4,495 groups the same chance, one in 4,495. Names in a hat, shaken well and drawn blind, does exactly that, and so does a random number generator used correctly.

\[ P(\text{any particular group}) = \dfrac{1}{4495} \approx 0.00022 \]

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 18 — the simple random sample and Lisa's study group

33. Lisa's study group: reading random digits two at a time

Worked example

Lisa numbers her classmates 00 to 30 and generates random numbers on a calculator. She reads each one as overlapping two-digit groups, keeps the first group that names a classmate, and lets each random number contribute at most one person.

\[ 0.94360, \; 0.99832, \; 0.14669, \; 0.51470, \; 0.40581, \; 0.73381, \; 0.04399 \]

Read the first number

Why: the groups are 94, 43, 36 and 60, and none of them is 30 or below, so nobody is chosen.

\[ 0.94360 \to 94, 43, 36, 60 \;\Rightarrow\; \text{none} \]

Read the second

Why: 99, 98, 83 and 32 are all above 30.

\[ 0.99832 \to 99, 98, 83, 32 \;\Rightarrow\; \text{none} \]

Read the third

Why: the first group, 14, names Macierz.

\[ 0.14669 \to 14 \;\Rightarrow\; \text{Macierz} \]

Read the fourth

Why: its groups are 51, 14, 47 and 70; the 14 is already in the group, so this number contributes nobody.

\[ 0.51470 \to 51, 14, 47, 70 \;\Rightarrow\; \text{repeat, skip} \]

Read the fifth

Why: 40 is too big, then 05 names Cuningham.

\[ 0.40581 \to 40, 05 \;\Rightarrow\; \text{Cuningham} \]

Read the sixth and seventh

Why: 73, 33, 38 and 81 are all too big; then 04 names Cuarismo, and the group is complete.

\[ 0.73381 \to \text{none}, \qquad 0.04399 \to 04 \;\Rightarrow\; \text{Cuarismo} \]

Figure (svg): Lisa's class roster of 31 names numbered 00 to 30 in three columns. The rows for 04 Cuarismo, 05 Cuningham and 14 Macierz are highlighted in yellow.

Every name has a two-digit label, so a string of random digits can be read straight off as names. Only the labels 00 to 30 point at anyone.

Check the draw

Why: three different labels, each between 00 and 30, each from a different random number, so the result is a genuine simple random sample of size three.

\[ \{14, 05, 04\} \subset \{00, 01, \dots, 30\} \]

The book's final sentence spells the first name Marcierz; the roster in its own Table 1.6 has Macierz, which is the spelling used here. Notice also why 31 to 99 are skipped rather than wrapped round: wrapping would give some classmates two labels and so two chances.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 18 — Table 1.6 and Lisa's random numbers

34. Stratified sampling: some from every group

Concept

Now the methods that use groups. In a stratified sample you first divide the population into groups called strata, then take a proportionate simple random sample from each stratum.

stratified sample — A sample formed by dividing the population into groups (strata) and taking a proportionate simple random sample from every stratum.

Figure (svg): Five boxes for five college departments, art with 8 members, biology 12, business 16, English 10 and maths 14. In every box a quarter of the circles are highlighted: 2, 3, 4, 3 and 4 members chosen.

Every department is represented, and a bigger department sends more people. That proportionality is what makes the sample a scale model of the college.

The book's example stratifies a college by department. You number the members of each department separately and use simple random sampling inside each one, taking more people from a bigger department so that the proportions match the college.

\[ \text{a quarter of each: } 8 \to 2, \quad 12 \to 3, \quad 16 \to 4 \]

The payoff is a guarantee that simple random sampling cannot give. Every stratum is represented, and in the right proportions, so a small department can never be missed by bad luck. That is why stratifying makes sense when you expect the groups to differ in the thing you are measuring.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 19 — the stratified sample

35. Cluster sampling: everyone in some groups

Concept

A cluster sample uses groups the other way round. Divide the population into clusters, choose some of the clusters at random, and take every member of the chosen clusters.

cluster sample — A sample formed by dividing the population into clusters, randomly selecting some clusters, and including all members of the selected clusters.

Figure (svg): The same five department boxes. Two whole boxes, biology with 12 members and English with 10, are outlined in yellow with every circle highlighted; the other three departments have no one chosen.

Here the randomness picks departments, not people. Three departments have nobody in the sample at all, which is the price of the convenience.

The book numbers the college's departments and uses simple random sampling to pick four of them; all the faculty in those four departments form the sample. Here two departments are chosen, biology and English, and all 22 of their members are in.

\[ 12 + 10 = 22 \text{ members, from 2 of 5 clusters} \]

The appeal is practical: visiting two whole departments is far cheaper than finding a few people in every department. The cost is that the unchosen clusters contribute no one, so if departments differ, the sample can easily lean one way.

Put the two definitions next to each other and they differ in exactly two words. Stratified takes some members from every group; cluster takes every member from some groups.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 19 — the cluster sample

36. Systematic sampling: a random start, then every kth

Concept

A systematic sample needs a list of the population. Choose a starting point at random, then take every kth member from the list.

systematic sample — A sample formed by randomly selecting a starting point in a listing of the population and then taking every kth member thereafter.

The book's example is a phone survey from a phone book of 20,000 residences, needing 400 names. The spacing comes from dividing the list by the sample size.

\[ k = \dfrac{20000}{400} = 50 \]

Figure (svg): A long strip representing the 20,000 phone-book listings. A random start at listing 37 is marked, then yellow ticks every 50 listings at 87, 137, 187 and so on; a zoomed inset shows the first few chosen positions.

Only the first pick uses chance. Everything after it is fixed by the interval of 50, which is what makes the method quick and also what makes it vulnerable to a list with a rhythm.

Pick a random start among the first 50 listings, say listing 37, and then take listings 87, 137, 187 and so on. The book numbers from 1 to 20,000 and picks the start from the whole list, which is why it adds that you might have to go back to the beginning; choosing the start from the first 50 avoids the wrap entirely and gives exactly 400 names.

\[ 37 + 50 \times 399 = 37 + 19950 = 19987 \]

The book notes that systematic sampling is often chosen because it is simple, and it is. The one chance event is the start, which is also the method's weak point, as a later slide shows.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 19 — the systematic sample

37. Decoding the sampling interval

Notation

The book never writes the spacing as a formula, but every systematic sample uses one. Take it apart so that you can build any systematic sample, not just the phone-book one.

Annotate

On: \( k = \dfrac{N}{n} \qquad \text{picks: } s, \; s + k, \; s + 2k, \; \dots \)

  • The size of the population list: 20,000 residences in the phone book, or 60 quiz scores in Try It 1.11. A capital N is the usual symbol for a population's size.
  • The sample size you want: 400 names, or 12 quiz scores. Lowercase n is the sample's size throughout the course.
  • How far apart the picks are. Dividing the list into n equal stretches and taking one member from each stretch is what makes the sample spread evenly.
  • A random whole number from 1 to k, and the only use of chance in the whole method. Every later pick is s plus a whole number of intervals.

If N over n is not a whole number, round the interval down so that the list is not exhausted before you have n picks. Rounding up could run you off the end of the list and force a wrap-around.

Notice that the formula has N in it. A systematic sample needs the list and its length, which is exactly what a phone book, a student roster or a factory's serial numbers provide, and what a street crowd does not.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 19 — the systematic sample, every fiftieth name

38. Try It 1.11: four samples from one table of quiz scores

Worked example

The book's Try It gives six columns of ten quiz scores, 60 scores in all, and asks for four different samples from the same table. Number the rows 1 to 10 within a column, and number the scores 1 to 60 down the columns, so that column 1 holds 1 to 10 and column 6 holds 51 to 60.

Stratified: three rows from every column

Why: each column is a stratum; with rows 3, 1, 7 in column 1, rows 1, 6, 10 in column 2, and so on, the 18 scores add to 140.

\[ \bar{x}_{\text{strat}} = \dfrac{140}{18} \approx 7.78 \]

Cluster: two whole columns

Why: draw two column numbers from 1 to 6, say 3 and 6; column 3 adds to 92 and column 6 to 69.

\[ \bar{x}_{\text{cluster}} = \dfrac{92 + 69}{20} = \dfrac{161}{20} = 8.05 \]

Simple random: fifteen numbers from 1 to 60

Why: the generator gave 4, 20, 27, 16, 59, 39, 35, 23, 11, 45, 34, 1, 5, 42 and 45, whose scores add to 113.

\[ \bar{x}_{\text{SRS}} = \dfrac{113}{15} \approx 7.53 \]

The number 45 came up twice. With 15 draws from only 60 numbers and repeats allowed, a repeat is the likely outcome, not bad luck: the chance of at least one is about 85 percent, a calculation Part 4 makes. The book's instructions keep the repeats, so this sample counts score 45 twice.

Systematic: twelve scores, every fifth

Why: the interval is 60 over 12, which is 5; a random start of 5 gives numbers 5, 10, 15, up to 60, whose scores add to 90.

\[ \bar{x}_{\text{sys}} = \dfrac{90}{12} = 7.5 \]

Figure (svg): Four copies of the 10 by 6 table of quiz scores. In the stratified copy three cells in every column are highlighted; in the cluster copy the whole of columns 3 and 6 is highlighted; in the simple random copy fifteen scattered cells are highlighted; in the systematic copy the highlighted cells fall every fifth cell counting down the columns.

Look at the shape each method leaves on the table: a few from every column, two whole columns, a random scatter, and an evenly spaced march down the columns.

Check against the whole table

Why: all 60 scores add to 479, and every sample mean lands within half a point of the population mean, with none of them equal to it.

\[ \mu = \dfrac{479}{60} \approx 7.98 \]

The book tells you to count ten scores between picks for the systematic sample and to wrap around if needed. With 60 scores that interval only reaches 6 different scores before returning to the first, so each would be counted twice; the interval 60 over 12, which is 5, gives 12 different scores, and it is used here.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 21-22 — Try It 1.11 and Table 1.7

39. The interval for a systematic sample of 12 from 60

Fill the middle

The correction in the last example is worth doing yourself, because it is a mistake the book's own instructions make. You have 60 quiz scores in a numbered list and want a systematic sample of 12 different scores.

Fill in the blanks

k = \dfrac5___ = \dfrac______ = ___

Why: The interval is 60 divided by 12, which is 5. Starting from a random number between 1 and 5, the twelve picks run up the list in steps of 5 and end exactly at the last stretch, so every pick is a different score and no wrap-around is needed.

Compare the book's interval of ten. From a start of 3, counting ten at a time gives 3, 13, 23, 33, 43 and 53, and the seventh pick wraps round to 3 again, so a sample of twelve is only six different scores taken twice.

\[ 3, 13, 23, 33, 43, 53, \; 3, 13, \dots \]

The general lesson: when the interval times the sample size exceeds the list length, a systematic sample repeats itself. Divide the list length by the sample size first, and the method can never lap the list.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 22 — Try It 1.11, step 4

40. Five answers to one question

Picture it

Put the four sample means from Try It 1.11 on one number line with the mean of all 60 scores. This picture is a preview of an idea the whole course is built on.

Figure (svg): A number line of average quiz score from 7.2 to 8.4. The population mean 7.98 is a yellow vertical line. Dots mark the four sample means: systematic 7.50, simple random 7.53, stratified 7.78 and cluster 8.05.

Four honest random samples, four different answers, none of them exactly the population's 7.98. That spread is sampling variation, and it is the subject of the whole of Part 5.

Every method was random and every sample was honest, and yet no two answers agree and none is exactly 7.98. The book says this plainly later in the section: two random samples from the same population will not be exactly the same, and neither is wrong.

\[ 7.50, \; 7.53, \; 7.78, \; 8.05 \quad \text{vs} \quad 7.98 \]

The cluster mean sits above the truth because column 3 happens to be the highest-scoring column in the table, at a mean of 9.2. That is the cluster method's weakness in miniature: with only two clusters, one unusual cluster moves the whole answer.

Do not over-read one run, though. With other random numbers the order of these four dots would shuffle. Judging which method is better needs many repetitions, and that is exactly what the sampling distributions of Chapter 7 are for.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 21-22 — Try It 1.11

41. Example 1.11: five ways to ask about tuition

Worked example

A study wants the average tuition San Jose State undergraduates pay per semester. The book describes five samples, and each one is identified by the single feature that makes it that method.

Sample (a): 25 from each of four classifications

Why: the students are grouped by year and some are taken from every group.

\[ 4 \times 25 = 100 \;\Rightarrow\; \text{stratified} \]

Sample (b): every 50th student after a random start

Why: a list, a random starting point, and a fixed step.

\[ s, \; s + 50, \; s + 100, \dots \text{ (75 students)} \;\Rightarrow\; \text{systematic} \]

Sample (c): 75 students by a completely random method

Why: each student has the same probability of being chosen at every stage.

\[ \text{every group of 75 equally likely} \;\Rightarrow\; \text{simple random} \]

Sample (d): two of the four years chosen, everyone in them

Why: chance picks the groups, then the whole of each chosen group is taken.

\[ 2 \text{ of } 4 \text{ years, all students} \;\Rightarrow\; \text{cluster} \]

Sample (e): the first 100 students met outside the library

Why: no chance at any point, only whoever came by that Wednesday.

\[ \text{whoever is nearby} \;\Rightarrow\; \text{convenience} \]

Samples (a) and (d) both start by splitting the students by year, which is why they are the pair to watch. The difference is what happens next: (a) samples inside every year, (d) samples the years themselves.

Check against the book's answer key

Why: the book gives stratified, systematic, simple random, cluster and convenience, matching all five.

\[ \text{(a) strat, (b) sys, (c) SRS, (d) cluster, (e) conv.} \]

Sample (e) is the only one that is not random, and it is also the only one whose answer you should expect to be off. Students near the library on one Wednesday may differ in tuition from students who never go there.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 21 — Example 1.11

42. Example 1.12 and Try It 1.12: name the method

Sorting

Seven short study descriptions from the book. For each, find the one feature that identifies the method: a list with a step, groups sampled inside, whole groups taken, pure chance over everyone, or no chance at all.

Sort into buckets

Sort each study by the sampling method it uses.

Simple random
A counselor picks students whose names match 50 computer-generated random numbers
Stratified
A coach picks 6 boys aged 8-10, 7 aged 11-12 and 3 aged 13-14 for a team; A researcher interviews 50 public and 50 private high school teachers; A principal polls 50 students from each of the four year groups
Cluster
A pollster interviews all human-resources staff in five high-tech companies
Systematic
A researcher interviews every third patient on a hospital's cancer list
Convenience
A student asks classmates in their algebra class how many pairs of jeans they own
srs
Random numbers matched to the whole list of students gives every group of 50 the same chance.
st
The coach, the researcher and the principal each split people into groups by age, school type or year and then took some from every group.
cl
Five whole companies were chosen and everyone in human resources there was interviewed, so the companies are the clusters.
sy
Every third patient on a list is a fixed step through a listing of the population.
cv
Classmates in one algebra class are simply the people the surveyor could reach, with no chance involved.

Three of the seven are stratified, and they look different on the surface: ages, school types and year groups. What they share is the structure, some people from every group, and the structure is what you are classifying.

The coach's team takes different numbers from each age group, 6, 7 and 3. The book still calls it stratified, and a sensible reading is that the numbers follow the sizes of the three groups, which is what proportionate means.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 22-23 — Example 1.12 and Try It 1.12

43. Trap: stratified and cluster swapped

Trap

The trap

A study splits a college into 20 departments, picks 4 departments at random and surveys every faculty member in them. A student writes:

\[ \text{groups used} \Rightarrow \text{stratified} \]

The reasoning is that stratified means using groups, and groups were used.

Both methods use groups, so that cannot be what separates them.

The fix

Ask what chance selected: people inside every group, or whole groups?

chance picksmethod
people inside every groupstratified
whole groupscluster

Here chance picked 4 whole departments and 16 departments contributed nobody, so this is a cluster sample.

A quick test: in a stratified sample, can a group end up with no one in it? No. In a cluster sample most groups do.

44. What each method guarantees

Comparison

Fill in the table to fix what each method promises. Each cell asks a yes or no question about the method in its row; answer yes, no, or first only.

Comparison matrix

methoddoes chance choose?is every group represented?does it need a full list?
simple randomyesnot guaranteedyes
stratifiedyesyesyes, within each stratum
clusteryes, the clustersnoonly a list of clusters
systematicfirst onlynot guaranteedyes
conveniencenonono

The last column explains why cluster sampling is popular despite its weakness. You need a list of the clusters, schools or city blocks, not a list of every person in them, and for large populations the list of people often does not exist.

The middle column is stratified sampling's selling point: the only method here that guarantees every group appears. The systematic row's answer, first only, is the reason systematic sampling counts as random at all. The start is random, so every member initially has an equal chance, even though everything after the start is fixed.

And the convenience row is all noes, which is the honest summary of why the book treats it separately. Nothing about it is designed to make the sample resemble the population.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 19 — the four random sampling methods

45. When systematic sampling goes wrong

Counterexample

Systematic sampling is simple, and the simplicity has a price. Build a list on which it fails badly, then say why the other methods would not fail the same way.

Discussion prompt

A block of flats lists its units floor by floor, four per floor, and on every floor the fourth unit is the large corner flat. A survey of rent takes every fourth unit on the list. What goes wrong, depending on the random start, and which method would avoid it?

The danger is any list whose pattern repeats with the same period as your interval: every seventh day of a daily list is the same weekday, every tenth item off a machine with ten moulds comes from the same mould.

Before using a systematic sample, look at how the list was ordered. Alphabetical lists are usually safe, because surnames have no rhythm tied to what you are measuring; lists ordered by position, time or batch deserve suspicion.

46. Convenience, and names back in the hat

Section

Part 4

47. Convenience sampling: whoever is easy to reach

Concept

One method in the book is not random at all. Convenience sampling uses results that are readily available, which means the people who happen to be at hand.

convenience sampling — A non-random method that uses results that are readily available, such as the people who happen to be present.

The book's example is a software store doing a marketing study by interviewing customers who happen to be browsing in the store. They are easy to reach, and they are exactly the people most interested in buying software.

Figure (svg): Ninety people scattered across a town, drawn as circles, most of them blue and some orange. A rectangle in the bottom right corner, labelled the store, contains a cluster in which most are orange.

The people in the store are not a random slice of the town; they are mostly the people who were already shopping for software. Easy to reach and typical are different things.

The book is careful here: convenience samples may be very good in some cases and highly biased in others. The trouble is that you cannot tell which from the sample itself, because the people you never reached are the ones who would have told you.

The same warning covers surveys mailed to households and returned by whoever feels like it. The book says it is better for the person conducting the survey to select the respondents than to let them select themselves.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 19 — convenience sampling and mailed surveys

48. Try It 1.13: the radio station's concert survey

Worked example

A radio station with 20,000 listeners wants to know whether its audience would prefer more music or more talk. It surveys the first 200 people it meets at one of its music concerts: 24 prefer more talk and 176 prefer more music.

Compute the sample proportions

Why: each count over the 200 people asked.

\[ \dfrac{176}{200} = 0.88, \qquad \dfrac{24}{200} = 0.12 \]

Check that the display is legal

Why: each person gave one of two answers, so the shares are exclusive and add to 1.

\[ 0.88 + 0.12 = 1 \]

Figure (svg): A pie chart of the 200 concert-goers' answers: more music 176, or 88 percent, a wedge of about 317 degrees, and more talk 24, or 12 percent, a wedge of about 43 degrees.

The pie is a legal display here, because every respondent gave exactly one of two answers. What is wrong is not the graph but who got asked.

Identify the method

Why: the first 200 people met at an event, with no chance involved.

\[ \text{first 200 at a concert} \;\Rightarrow\; \text{convenience} \]

Ask who could not be in the sample

Why: listeners who never go to concerts, and especially those who tune in for the talk shows, had no chance of being among the 200.

\[ \text{talk-show fans at a music concert} \;\Rightarrow\; \text{under-represented} \]

The station would probably report that 88 percent of listeners want more music. The honest statement is narrower: 88 percent of the first 200 people met at a music concert want more music, which is close to saying that people who like music like music.

Check the conclusion against the method

Why: a sample chosen at a music event favours music lovers, so the 0.88 is very likely an overestimate of the share among all 20,000 listeners; the sample is not representative, and a random sample from the station's listener list is the fix.

\[ 0.88 \text{ describes concert-goers, not all 20000 listeners} \]

Notice that nothing in the arithmetic was wrong and no amount of arithmetic could rescue it. The flaw was fixed at the moment the 200 people were chosen.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 24 — Try It 1.13

49. With replacement and without

Concept

One more distinction about how names are drawn. In sampling with replacement, a member who is picked goes back into the population and may be chosen again; in sampling without replacement, each member can be chosen only once.

sampling with replacement — Once a member is picked, that member goes back into the population and may be chosen more than once.

sampling without replacement — A member of the population may be chosen only once. Surveys are typically done this way.

The book calls sampling with replacement truly random, since every draw faces the same population. In practice nobody wants to interview the same person twice, so surveys are done without replacement.

\[ \text{draw 1: } \dfrac{1}{N}, \quad \text{draw 2: } \dfrac{1}{N} \text{ (with) or } \dfrac{1}{N-1} \text{ (without)} \]

The line above is the whole difference. Without replacement the pool shrinks by one each draw, so each remaining member's chance creeps up slightly. When N is in the thousands, one over N and one over N minus 1 are practically the same number, and the two methods behave alike.

The book's conclusion is that the choice becomes a mathematical issue only when the population is small. The next slide measures exactly how small.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 19-20 — sampling with and without replacement

50. How likely is a repeat? A class of 31 against a college of 10,000

Worked example

If you draw with replacement, the risk is picking the same person twice. Compute that risk for a small population and a large one, and the book's claim about when it matters becomes a number.

Set up Lisa's class, drawing 3 of 31 with replacement

Why: the chance that all three draws differ is the first draw anything, the second avoiding one name, the third avoiding two.

\[ P(\text{no repeat}) = \dfrac{31}{31} \times \dfrac{30}{31} \times \dfrac{29}{31} \]

Multiply out

Why: the top is 31 times 30 times 29 and the bottom is 31 cubed.

\[ = \dfrac{26970}{29791} \approx 0.9053 \]

Take the complement

Why: a repeat is everything that is not no repeat.

\[ P(\text{repeat}) = 1 - 0.9053 = 0.0947 \]

Almost a one in ten chance, in a tiny population with a tiny sample. That is why Lisa draws without replacement and simply redraws on a repeat, as the calculator instructions say.

Now the college, 10 students of 10,000

Why: the same product, ten factors long, computed in node.

\[ P(\text{no repeat}) = \dfrac{10000}{10000} \times \dfrac{9999}{10000} \times \cdots \times \dfrac{9991}{10000} \approx 0.99551 \]

\[ P(\text{repeat}) \approx 0.0045 \]

Check against the book's claim

Why: about 1 chance in 220 for the college, against nearly 1 in 10 for the class: with a small sample from a large population the two methods practically coincide, and the difference matters only when the population is small.

\[ 0.0045 \ll 0.0947 \]

The same arithmetic explains the repeated 45 in Try It 1.11. Fifteen draws from only 60 give a repeat with probability about 0.85, so there a repeat was nearly guaranteed.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 20 — sampling without replacement approximates sampling with replacement

51. Watch the chance of a repeat fall as the population grows

Tweak it

The previous slide computed two points. The slider draws the whole curve: the horizontal axis is the population size N, and the slider sets how many draws you make with replacement.

Parameter explorer

Set the sample size, then read off how large the population must be before the chance of drawing someone twice falls below 0.1.

\[ P(\text{repeat}) = 1 - \dfrac{N(N-1)\cdots(N-{k}+1)}{N^{{k}}} \]

  • k — from 2 to 30: draws with replacement

Two things move the curve, and they pull in opposite directions. A bigger population pushes the chance of a repeat towards zero; a bigger sample pushes it up, faster than you might expect, because every pair of draws is a chance for a clash.

What matters is the sample size compared with the population, not either one alone. With 10 draws a population of a few thousand makes repeats rare; with 30 draws you need several thousand to get the same comfort.

Real surveys sit far to the right of this window: a thousand people from a population of millions. Out there the curve is flat against zero, which is why the book can say that without replacement is approximately the same as with replacement.

52. Does a bigger population need a bigger sample?

Prediction

The book adds a note that surprises almost everyone. Commit to an answer before reading it.

Predict first

You want the same precision for a population of 100,000 and for one of 1,000,000. How should the sample sizes compare?

  • About the same sample size will do for both
  • The larger population needs a sample ten times as big
  • The larger population needs a sample about three times as big
  • The smaller population needs the bigger sample

Correct: About the same sample size will do for both

Why: The book states that sample size is a function of the desired precision and not of the population size, so a sample of 1,000 can be adequate for a population of 100,000 or of 1,000,000 if the same precision is wanted. What limits precision is how many people you actually measured, not what fraction of the population they are.

An analogy helps. To judge the saltiness of a pot of soup you stir it and taste one spoonful, and a big pot needs no bigger spoon than a small one, provided both are well stirred. The stirring is the randomness; the spoonful is the sample.

The formulas that make this precise arrive in the confidence-interval chapter, where you will notice that the population size simply does not appear. For now, keep the consequence: a well-drawn national poll of about 1,200 people can be as good as a local poll of the same size.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 23 — the Note on sample size and population size

53. Variation, error and bias

Section

Part 5

54. Variation in data: eight sixteen-ounce cans

Concept

Variation is present in any set of data. The book's example measures the beverage in eight cans, each labelled 16 ounces.

\[ 15.8, \; 16.1, \; 15.2, \; 14.8, \; 15.8, \; 15.9, \; 16.0, \; 15.5 \]

Figure (svg): A dot plot of the amounts in eight 16-ounce cans, from 14.8 to 16.1 ounces, with 15.8 appearing twice. A dashed line marks the labelled 16 ounces and a yellow line marks the mean, 15.64 ounces.

No two cans agree exactly, and only one of the eight holds more than the label promises. Variation is the normal state of data, not a sign that something went wrong.

Compute the average to see where the cans cluster. Add the eight amounts, then divide by eight.

\[ \dfrac{125.1}{8} = 15.6375 \approx 15.64 \text{ oz} \]

The book gives two reasons the amounts differ: different people make the measurements, and the exact 16 ounces was not put into every can. The first is variation in the measuring, the second is variation in the thing measured, and a real data set carries both at once.

The book's advice about your own data follows directly. Expect your measurements to differ somewhat from someone else's for the same purpose; that is natural. If two people's results differ a great deal, though, it is time to re-examine your methods and your accuracy.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 24 — Variation in Data

55. Variation in samples: Doreen, Jung and a die

Concept

Samples vary too. Doreen and Jung each sample 500 students to estimate how long students at their college sleep, Doreen systematically and Jung by clusters. Their samples will differ, and even with the same method they would differ. Neither would be wrong.

You can watch this happen with the book's collaborative exercise: roll one fair die 20 times, record the faces, then do it again. Here are two such runs.

Figure (svg): Two bar charts of die faces, each from 20 rolls of the same fair die. The first run gave 6 ones, 7 twos, 2 threes, no fours, 2 fives and 3 sixes; the second gave 2, 8, 1, 3, 3 and 3. A dashed yellow line at about 3.3 marks the expected count for every face.

Same die, same person, same number of rolls, and the two charts disagree in every column. Neither run is wrong; twenty rolls is simply not many.

\[ \text{run 1: } 6, 7, 2, 0, 2, 3 \qquad \text{run 2: } 2, 8, 1, 3, 3, 3 \]

Each run adds to 20, and in the long run each face should come up about 20 divided by 6, or 3.3, times. Yet run 1 has no fours at all and seven twos, and run 2 has only one three. Which experiment had the correct results? The book's answer: both did.

Larger samples bring results closer to the population's, and the book says so, but two large samples will still differ from each other. It adds that this variability cannot be stressed enough, and that the job of the statistician is to see through the variability and draw appropriate conclusions.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 24-26 — Variation in Samples and the die exercise

56. Sampling error and nonsampling error

Concept

When a sample result misses the population value, the book sorts the cause into two kinds.

sampling error — Error caused by the actual process of sampling: the sample is only part of the population. It is always present, and as a rule a larger sample makes it smaller.

nonsampling error — Error caused by factors unrelated to the sampling process, such as a defective counting device, a mistyped value or a badly worded question.

\[ \text{sample result} = \text{population value} + \text{sampling error} + \text{nonsampling error} \]

The book is blunt about the first one: a sample will never be exactly representative of the population, so there will always be some sampling error. That is not a flaw in the method. It is the price of looking at part of the population instead of all of it, and it is the spread you saw in the four quiz-score means.

Nonsampling error is different in kind. A broken scale reads wrong for the whole sample, and a bigger sample weighed on the same scale is simply wrong more often. The cure is fixing the process, not adding people.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 20 — sampling and nonsampling errors

57. Sampling bias: when some members are less likely to be chosen

Concept

Sampling error is the random miss every sample has. Sampling bias is a systematic tilt built into how the sample was chosen.

sampling bias — Created when a sample is collected from a population and some members of the population are not as likely to be chosen as others.

Compare the definition with the definition of random sampling from Part 3. Random sampling is exactly the promise that every member initially has an equal chance; bias is that promise broken. The concert survey broke it for talk-show fans, the software store for people who never shop for software.

\[ \text{equal chances} \;\Rightarrow\; \text{error only} \qquad \text{unequal chances} \;\Rightarrow\; \text{bias} \]

The book's warning is that when a sampling bias happens, incorrect conclusions can be drawn about the population. The incorrectness does not shrink as the sample grows, because every extra person is drawn by the same tilted method.

Keep the two words apart in your own writing. Error is a scatter around the truth that you can reduce by taking more; bias is a displacement from the truth that you can only remove by changing how you choose.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 20 — sampling bias

58. Error is scatter, bias is aim

Picture it

A target makes the distinction easy to hold. Imagine repeating a study many times; each repetition's result is one shot, and the population value is the centre.

Figure (svg): Four targets with the true population value at each centre, each hit by twelve sample results. Top left, small unbiased samples: shots scattered widely around the centre. Top right, large unbiased samples: shots packed tightly on the centre. Bottom left, small biased samples: shots scattered widely around a point up and to the right of the centre. Bottom right, large biased samples: shots packed tightly around that same off-centre point.

Going down a column, a bigger sample tightens the cluster. Going across, bias moves the whole cluster, and no amount of tightening brings it back to the yellow centre.

Read down the left column: a fair method with small samples scatters widely around the centre, and the same method with large samples packs its shots tightly on it. That tightening is sampling error shrinking as n grows, exactly as the book's rule says.

Now read down the right column. The biased method scatters around the wrong point, and large samples pack the shots tightly around that same wrong point. A bigger sample made the result more precise and not one bit more accurate.

The bottom right panel is the most dangerous in statistics, because it looks the most convincing. A huge biased sample gives results that agree beautifully with each other, and that agreement is easily mistaken for being right.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 20 and 25 — sampling error, sampling bias and the size of a sample

59. Example 1.13(a) and (b): two samples that miss in opposite directions

Worked example

ABC College has 10,000 part-time students, and the question is the average amount a part-time student spends on books in the fall term. Two convenience samples of ten are taken.

\[ \text{chemistry: } 128, 87, 173, 116, 130, 204, 147, 189, 93, 153 \]

\[ \text{seniors: } 50, 40, 36, 15, 50, 100, 40, 53, 22, 22 \]

Average the chemistry sample

Why: the ten amounts add to 1420 dollars.

\[ \bar{x}_1 = \dfrac{1420}{10} = 142 \text{ dollars} \]

Average the senior-citizen sample

Why: the ten amounts add to 428 dollars.

\[ \bar{x}_2 = \dfrac{428}{10} = 42.8 \text{ dollars} \]

Figure (svg): Two dot plots on one axis of dollars spent on books, from 0 to 300. The organic chemistry sample, in orange, runs from 87 to 204 with mean 142. The senior-citizen sample, in blue, runs from 15 to 100 with mean 42.8. The two clouds barely overlap.

Two samples of the same population, and the averages are more than three times apart. When two methods disagree this badly, at least one of them is not measuring the population.

Ask part (a): is either representative?

Why: the chemistry students are science-oriented and many also take first-term calculus, whose books are expensive; the seniors take classes for health and interest and buy few books.

\[ 142 \text{ probably too high}, \qquad 42.8 \text{ probably too low} \]

Ask part (b): may either describe the population?

Why: no, because not every part-time student had an equal chance of being chosen; students outside those two classes had no chance at all.

\[ \text{unequal chances} \;\Rightarrow\; \text{biased} \]

Check the diagnosis with the gap between the samples

Why: two samples of the same population differing by a factor of more than three is far beyond what random variation among ten students would produce, which confirms that the methods, not the luck, are at fault.

\[ \dfrac{142}{42.8} \approx 3.3 \]

The book's words are that both samples are biased. It is also true that ten is small, but taking a hundred chemistry students would only give a more precise estimate of what chemistry students spend.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 23-24 — Example 1.13, parts (a) and (b)

60. Example 1.13(c): a fair sample that is still small

Worked example

A third sample takes one part-time student from each of ten disciplines, chemistry, maths, English, psychology, sociology, history, nursing, physical education, art and early childhood development, each chosen by simple random sampling. The book assumes these are the only disciplines and that each has the same number of part-time students.

\[ 180, \; 50, \; 150, \; 85, \; 260, \; 75, \; 180, \; 200, \; 200, \; 150 \]

Name the method

Why: every discipline contributes, and chance chooses within each one, so this is a stratified sample by discipline; with equal-sized disciplines, one from each is proportionate.

\[ 1 \text{ from each of } 10 \text{ strata} \;\Rightarrow\; \text{stratified} \]

Average the sample

Why: the ten amounts add to 1530 dollars.

\[ \bar{x}_3 = \dfrac{1530}{10} = 153 \text{ dollars} \]

Figure (svg): A dot plot of the third sample's ten amounts on an axis from 0 to 300 dollars: 50, 75, 85, two at 150, two at 180, two at 200 and 260, with mean 153 marked. Faint bands show where the two biased samples lay.

The fair sample spreads across both earlier ranges and beyond, from 50 dollars to 260. Only ten students, though, so its mean of 153 could easily move by tens of dollars in a repeat.

Answer part (c): is it biased?

Why: no: every part-time student had an equal chance, because each discipline has the same size and one student is drawn at random from each.

\[ \text{equal chances} \;\Rightarrow\; \text{unbiased} \]

Unbiased is not the same as close. The ten values run from 50 to 260 dollars, a range of 210, so one different student in the sample could move the mean by more than 20 dollars. That is sampling error, and the book's remedy is a larger sample.

Check the book's recommendation against the target picture

Why: the method is fair, so the sample sits in the left column of the targets; more students would tighten its scatter around the true mean, whereas for the biased samples even a large sample runs the risk of not being representative.

\[ \text{fair method} + \text{larger } n \;\Rightarrow\; \text{closer to } \mu \]

The book's closing line answers the question students always ask. Is a sample good enough, instead of surveying the entire population? If the survey is done well, yes.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 24 — Example 1.13, part (c)

61. Bigger samples shrink error, not bias

Tweak it

The slider turns the target picture into a curve. The horizontal axis is the sample size n; the vertical axis is the typical distance, in percentage points, between a sample percentage and the true one for a yes or no question. The slider adds a bias built into the method.

Parameter explorer

Set the bias to 0 and watch the typical miss fall towards zero as n grows. Then set a bias of 5 points and find what the typical miss approaches for very large samples.

\[ \text{typical miss} = \sqrt{\dfrac{2500}{n} + {b}^2} \]

  • b — from 0 to 8: bias in the method (points)

The first term under the square root is sampling error: for a question near a fifty-fifty split, a sample of n people typically misses by about 50 divided by the square root of n points, a fact the confidence-interval chapter derives. The second term is the bias, and n does not appear in it anywhere.

So with a bias of 5, the curve flattens onto a floor of 5 points. Quadrupling the sample halves the sampling-error part and leaves the floor exactly where it was. That is the book's sentence about biased techniques in one picture: even a large sample runs the risk of not being representative.

This is also why the book says samples of 1,200 to 1,500 are considered good enough in polling if the survey is random and well done. Beyond that size the sampling error is already small, and the remaining danger is the bias that more people cannot fix.

62. Trap: a sample this big cannot be wrong

Trap

The trap

A call-in television poll gets 80,000 responses, and a presenter says:

\[ n = 80000 \Rightarrow \text{reliable} \]

The reasoning is the sampling-error rule turned into a slogan: the larger the sample, the smaller the error.

With 80,000 people, sampling error is indeed tiny.

The fix

The book says it directly: many large samples are biased, and call-in surveys are invariably biased, because people choose whether to respond.

\[ \text{self-selected} \Rightarrow \text{bias} \]

\[ \text{bias does not shrink with } n \]

The 80,000 are the people who felt strongly enough to call. A random sample of 1,200 would be better evidence about the viewers as a whole.

Size answers how precise; only the method answers aimed at what.

63. Critical evaluation: name the problem

Matching

The book lists common problems to look for before accepting any study's results. Match each short case to the problem it shows.

Match the pairs

  • a. A website poll counts only the visitors who chose to click an answer
  • b. A survey asks: do you agree that the unfair new parking fee should be scrapped?
  • c. Half the people mailed a survey never return it, mostly those with no strong view
  • d. Towns with more ice-cream sales have more drownings, so ice cream causes drowning
  • e. A drinks company funds and runs the study showing its drink improves memory
  • w. self-selected sample
  • x. undue influence
  • y. non-response
  • z. causality claimed from a relationship
  • v. self-funded or self-interest study

Why: The website poll hears only from people who chose to respond, which is self-selection. The word unfair in the question pushes respondents towards one answer, which is undue influence. The missing returns leave only the strongly opinionated, which is non-response. Ice-cream sales and drownings both rise in hot weather, so a third variable links them, which is why a relationship does not show causation. The drinks study has an interest in its own result, which calls for careful reading, though not automatic rejection.

The book adds three more to the list: problems with the sample itself, sample-size issues, and misleading use of data such as badly drawn graphs, which Part 2 of this lesson guarded against. It also names confounding, when the effects of several factors cannot be separated, as in the ice-cream case where heat is tangled with both variables.

On self-funded studies the book is even-handed. Do not assume such a study is good, and do not assume it is bad; evaluate it on its merits and the work done. The funding is a reason to read closely, not a verdict.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, p. 20 — Critical Evaluation

64. The book's collaborative exercise: representative or not?

Sorting

Five proposed samples from the book's class exercise. For each, ask whether every member of the stated population had a fair chance of being included and whether the sample is big enough to mean anything.

Sort into buckets

Sort each proposed sample by whether it is representative of its population.

Representative
Average cost of a two-day hospital stay in Massachusetts, surveying 100 hospitals chosen by simple random sampling
Not representative
Average GPA of all students, using all honours students as the sample; Favourite cereal of children under ten, asking every 20th child entering one large supermarket; Average income of all U.S. adults, sampling U.S. Representatives from randomly chosen states; Share of people taking public transport to work, asking 20 people who sit next to you in Central Park
yes
The hospital survey draws 100 hospitals at random from the state, so every hospital had an equal chance and the sample is a reasonable size.
no
Honours students have high GPAs by definition. One supermarket over three hours reaches only its own customers, even with a systematic step. Representatives earn far more than typical adults, however the states are chosen. Twenty people on one park bench in one city is a tiny convenience sample.

The Representatives case is instructive because it sounds rigorous: states, strata, clusters, random selection. The book even muddles the words, calling the states strata and then building a cluster sample from them. None of that matters, because the people sampled are drawn from the wrong population, and no random mechanism can fix sampling from the wrong list.

The cereal case shows the same thing on a smaller scale. Every 20th child is a systematic step, but it is a step through the customers of one shop at one time, so the population it really represents is that shop's customers.

OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2, pp. 20-21 — Collaborative Exercise

65. Putting it together

Section

Part 6

66. Two chains of questions for any study

Pattern

Everything in this lesson reduces to two short chains of questions, one about the data and one about the sample. Work through them in order whenever you meet a study, in a textbook or a newspaper.

Figure (svg): A two-row flow chart. Top row, about the data: what was recorded, a name or a number, then counted or measured, ending at qualitative, discrete or continuous. Bottom row, about the sample: did chance choose, if not it is convenience or self-selected; if so, how were groups used, ending at simple random, stratified, cluster or systematic; and finally, could a larger sample fix the problem, only if it is error rather than bias.

Two short chains of questions, asked in order, classify nearly every study you will read. The last box is the one people skip, and it is the one that decides whether the conclusion can be trusted.
  1. What was recorded? A word or label is qualitative, and there is no second question
  2. If a number, counted or measured? Counted is discrete, measured is continuous, decided by the process and never by a summary
  3. How is it displayed? Percentages that add to exactly 100 may be a pie; more than 100 means overlap and less means something is missing, so use bars
  4. Did chance choose the sample? If not, it is convenience or self-selected and bias is likely
  5. If so, how were groups used? None: simple random. Some from every group: stratified. All from some groups: cluster. A start and a step: systematic
  6. Error or bias? Error shrinks with a larger sample; bias needs a better method

The last question is the one that decides whether a conclusion can be trusted, and it is the one most often skipped. A sample's size is easy to report and easy to admire; its method is where the truth of the result is decided.

67. Check: classify the data

Check

One variable, four candidate classifications. Ask the two questions in order before reading the choices.

Check your understanding

A café records how long, in minutes, each customer waits for their order. What type of data is this?

  • A. quantitative continuous (correct)
  • B. quantitative discrete
  • C. qualitative
  • D. discrete if the times are rounded to whole minutes, continuous otherwise

Answer: A

Why: A waiting time is an amount, so it is quantitative, and it is measured with a clock rather than counted, so it is continuous. Rounding the recorded times to whole minutes changes the precision of the measurement, not the fact that time was measured.

Why B tempts people
Discrete data come from counting, and nobody counts a waiting time; the stopwatch could always show more decimal places.
Why C tempts people
Qualitative data name a category, while a waiting time is a number whose average means something, such as an average wait of 4.2 minutes.
Why D tempts people
This is the whole-number trap. The type is decided by how the number was produced, and rounding a measurement to whole minutes leaves it a measurement.

The last option is the one worth studying. It sounds sophisticated and it confuses how precisely something was recorded with what kind of thing it is. The kilograms of nails in Try It 1.7 and the lawn areas in Try It 1.6 were the same trap.

If you hesitated between the first two choices, use the book's hint in reverse. You would say the number of customers, but you would not say the number of waiting time, and that phrase test separates counting from measuring almost every time.

68. Check: name the sampling method

Check

A description in the style of Example 1.12. Find the feature that identifies the method rather than matching surface words.

Check your understanding

A city has 40 schools. An official picks 5 schools at random and surveys every teacher in those 5 schools. Which method is this?

  • A. cluster sample (correct)
  • B. stratified sample
  • C. systematic sample
  • D. simple random sample

Answer: A

Why: Chance chose whole schools, and every teacher in each chosen school was surveyed, so the schools are clusters. Thirty-five schools contribute no teachers at all, which is the hallmark of a cluster sample and impossible in a stratified one.

Why B tempts people
A stratified sample would take some teachers from every one of the 40 schools. Here most schools contribute nobody, so the groups were not used as strata.
Why C tempts people
There is no list with a fixed step through it; the schools were drawn at random rather than every eighth school from a list.
Why D tempts people
A simple random sample of teachers would make every group of teachers equally likely, but here any group containing teachers from more than five schools is impossible.

The distractor about simple random samples tests the definition from Part 3 precisely. It is not enough that chance was involved; every possible group of the sample size must be equally likely, and a cluster sample rules most groups out.

If you chose stratified, you latched onto the fact that groups were used. Both methods use groups; the question is whether chance picked people inside every group or picked whole groups.

69. Check: error or bias?

Check

The last distinction, and the one with the most practical weight. Decide what kind of problem this is before choosing a remedy.

Check your understanding

A gym asks the members who are working out at 7 a.m. on Mondays how many days a week they exercise, and gets an average of 5.1. Which remedy addresses the main problem?

  • A. Choose members at random from the full membership list (correct)
  • B. Ask three times as many members at 7 a.m. on Mondays
  • C. Report the average to two decimal places
  • D. Repeat the same survey next Monday and average the two results

Answer: A

Why: Members who are at the gym early on a Monday are exactly the keenest exercisers, so the sample is biased towards high values, and 5.1 probably overstates the members as a whole. Only changing the method, by giving every member an equal chance, removes that bias.

Why B tempts people
More people chosen the same way reduces sampling error but keeps the bias, so the answer would become more precise around the same inflated value.
Why C tempts people
Extra decimal places change how the answer is written, not how the members were chosen, and they suggest a precision the survey does not have.
Why D tempts people
Two Mondays at 7 a.m. are two draws from the same tilted method, so the average of the two is biased in the same direction.

Two of the wrong options are the same mistake in different clothes: more data collected the same way. The target picture is the reason they fail, since the biased column tightens without ever moving to the centre.

Notice that the question asked for the main problem. The sample might also be small, but a bigger biased sample is not better evidence, so the method has to be fixed first.

70. Explain why size cannot rescue a biased sample

Explain it to yourself

One idea sits underneath all of Part 5. Putting it into your own words is what turns it from a sentence you can repeat into a judgement you can make about a study you have never seen.

Discussion prompt

In three or four sentences, explain to someone who has not studied statistics why a self-selected sample of 100,000 people can be weaker evidence than a random sample of 1,000. Try to do it without using the words bias or error.

If your explanation used the soup-tasting picture from Part 4, it probably went well. Stirring is what makes one spoonful trustworthy, and a huge spoonful from an unstirred pot still tastes only of the top.

If it leaned on the phrase bigger is better, go back to the target figure and read the right-hand column once more. More shots from a misaimed method land more tightly on the wrong place.

71. Exit ticket

Exit ticket

One study, all three questions of the lesson. Answer in full sentences, naming the evidence for each answer.

A school wants the mean number of hours per week its 1,200 students spend on homework. It lists the students alphabetically, picks a random starting number from 1 to 20, and takes every 20th student after that. It records each student's answer to the nearest half hour.

Discussion prompt

Classify the data, name the sampling method and give the sample size, and say whether a larger sample drawn the same way would fix any bias present.

The most common slip here is the sample size. The interval is 20, not the sample size; the sample size is the list length divided by the interval, 1,200 over 20, which is 60.

If you called the data discrete because of the half hours, reread the classification check. Rounding to a half hour is still rounding a measurement.

72. What you can do now

Recap

This section built the questions that come before every calculation in the course. Each row of the table is one of them, with the mistake it prevents.

the questionwhat to dothe trap
what kind of data?name, or amount; then counted or measureddigits that are labels; a mean of 2.6 books
may this be a pie chart?only if the percents total exactly 100overlapping or missing categories
which random method?some from every group, all from some groups, a start and a step, or pure chancecalling any use of groups stratified
with or without replacement?without, in practice; it matters only for small populationsrepeats in small pools, like 15 draws from 60
error or bias?error shrinks with n; bias needs a better methodtrusting a huge self-selected sample

\[ \text{bigger } n \;\Rightarrow\; \text{less sampling error} \qquad \text{better method} \;\Rightarrow\; \text{less bias} \]

That line is the lesson in miniature. Everything later in the course, from confidence intervals to hypothesis tests, measures the first kind of problem with great care, and every one of those calculations assumes you have already dealt with the second.

Section 1.3 picks up straight from here: frequency tables for organising data, and levels of measurement, which refine the qualitative and quantitative split into nominal, ordinal, interval and ratio.

OpenStax Introductory Business Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling §1.2 — the same definitions of data types and sampling methods, in business settings

Sources

  1. OpenStax Introductory Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling — Illowsky & Dean, OpenStax / Rice University, CC BY 4.0, pp. 9-26
  2. OpenStax Introductory Business Statistics 2e, §1.2 Data, Sampling, and Variation in Data and Sampling — Illowsky & Dean, OpenStax / Rice University, CC BY 4.0

Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.

Book on Wyzant · Text (657) 465-8108