Secondary 4 Statistics and Probability Revision | How to Read Data Under Exam Pressure

Secondary 4 statistics and probability revision becomes useful when a student can explain what a number means before deciding how to calculate it. Mean, median, standard deviation, cumulative frequency, box-and-whisker plots and probability trees are not separate calculator routines. They are ways to answer different questions about a collection of data or a chance process. This guide connects those questions to worked examples, independent practice and the decisions students must make under examination pressure.

For students and parents in Bukit Timah, the practical concern is often familiar: the student understands a solution when someone explains it, but loses marks when a graph changes, a frequency column appears, or a probability question says “without replacement”. The repair is not necessarily another complete paper. It may be learning to identify the correct total, distinguish a value from its position, or recognise that the second selection no longer has the same probabilities as the first.

This Secondary 4 Mathematics guide covers data interpretation, averages, measures of spread, grouped data, cumulative frequency, sample spaces and combined probability. Its central rule is read the representation, identify the quantity, choose the model, calculate, then interpret and check. The explanations and practice data below are original teaching examples. They are not school results, exam predictions or claims about any named school or tuition programme.

Start here: a 50-second revision router

Choose the symptom closest to your own work. Read that section, attempt its example with the answer covered, and use a different question later to check whether the improvement survives without help. You do not need to read the entire guide in one sitting.

Five routes through this guide

What is happening?	Start with	What you should produce
I misread tables or graphs.	Reading data and representations	A correctly identified quantity, unit and total.
I confuse averages or calculator outputs.	Averages, frequencies and standard deviation	A calculation whose denominator you can explain.
I struggle with cumulative frequency or box plots.	Positions, quartiles and distributions	A graph reading followed by a defensible comparison.
I do not know when to add or multiply.	Sample spaces and probability models	Labelled outcomes and an expression justified by the model.
I understand the chapter but lose marks in tests.	Exam control, diagnostics and revision plans	A short, specific error-repair plan and a delayed retest.

Open the complete contents

1. Your cohort and syllabus · 2. Read before calculating · 3. Choose and interpret graphs · 4. Stem-and-leaf diagrams · 5. Mean, median and mode · 6. Frequency tables and weighted means · 7. Missing and corrected observations · 8. Range, quartiles and consistency · 9. Standard deviation · 10. Grouped estimates · 11. Cumulative frequency · 12. Box plots · 13. Misleading claims and sampling · 14. Probability foundations · 15. Complete sample spaces · 16. Event language · 17. Trees and replacement · 18. Complements and repeated trials · 19. Conditional reasoning · 20. Unknown quantities · 21. Calculator verification · 22. Examination control · 23. Diagnostic practice · 24. Mixed teaching laboratory · 25. Revision plans · 26. Parent and teacher use · 27. Questions and glossary · 28. Sources and further routes.

1. Match the revision to your actual Mathematics paper

Cohort guidance checked on 20 September 2026. A student sitting the 2026 O-Level Mathematics examination should use syllabus 4052. The 2027 SEC school-candidate listings identify G3 Mathematics as K310 and G2 Mathematics as K210. These are Mathematics papers, not the separate Additional Mathematics syllabuses. The links in the source section lead to the official listings and syllabus documents.

The statistics-and-probability content in the 2026 4052 and 2027 K310 documents includes statistical representations, averages, grouped-data mean, quartiles, percentiles, spread including standard deviation, and simple combined probability. Both explicitly list histograms with equal class intervals. The optional discussion of unequal widths later in this guide explains the mathematics; it is not a claim that unequal-width histograms are a required topic in those papers. See 4052 for 2026 and K310 for 2027.

For K210, an important distinction is visible in the actual PDF, not just a plain-text topic list: some content is underlined and reserved for Section B of Paper 2. That section offers a choice between geometry/measurement and statistics/probability questions. Cumulative frequency, box plots, quartiles, spread and combined probability include underlined entries. Read the K210 assessment scheme and content pages with your teacher before copying a G3 revision strategy.

That distinction does not mean a G2 learner must avoid this guide. It means the learning route and examination route are not identical. You can understand standard deviation deeply while making a school-guided decision about which optional examination area to prepare. Conversely, do not abandon all statistics because some advanced content belongs to an optional section: the un-underlined entries still need attention.

For IP, IGCSE, another national curriculum or a different examination year, use your own school and examination-board documents as the controlling map. The mathematics below remains useful, but a useful explanation is not proof that every topic is examinable in every programme. In particular, formal conditional-probability notation and unequal-width histogram examples are labelled as extensions where introduced.

Keep one simple reference at the top of your revision folder: examination year, subject name, syllabus code and the school’s current coverage list. This prevents the avoidable problem of doing difficult work from a different course while neglecting the ordinary questions on the paper you will actually sit. The first act of efficient revision is deciding what the target is.

2. Read the data before reaching for the calculator

A graph can be read incorrectly in less than a second. A student sees bars, recognises the chapter and starts calculating. Yet the vertical axis might show percentages rather than counts, the horizontal intervals might represent minutes rather than marks, or the total might refer only to respondents rather than the whole school. Once that first interpretation is wrong, accurate arithmetic simply produces a tidy answer to the wrong question.

Before calculating, finish this sentence: “Each observation is one ___, the measured variable is ___, and it is recorded in ___.” For a table of journey times, each observation might be one school-day trip, the variable journey duration, and the unit minutes. For a table of favourite clubs, the variable is a category, so a numerical mean of the club names makes no sense. Knowing the kind of variable limits the statistics that can reasonably be used.

The three numbers that students confuse

A data value is the quantity observed. A frequency tells you how many observations have that value or belong to that interval. A cumulative frequency tells you how many observations have been accumulated up to a stated boundary. If a table shows a score of 6 with frequency 9, it does not say there were six students with nine marks. If the cumulative frequency at 30 minutes is 28, it does not say 28 students took exactly 30 minutes.

Try a small example. A reading survey records the number of books completed by 20 students: 0 books for 2 students, 1 for 5, 2 for 8, 3 for 4, and 4 for 1. The number 8 is a frequency, not the largest book count. The number 4 is a possible book count, not the sample size. The total number of books is neither of those numbers. It requires weighting each count by its frequency.

The quickest safeguard is to annotate the column headings before touching any numbers. Write “value” over the book-count column and “number of students” over the frequency column. Then check that the frequencies sum to 20. This looks elementary, but it stops several different errors at once: choosing the wrong denominator for the mean, finding the median of the frequencies, and reporting the highest frequency as the mode.

Read the question’s boundary words literally

“More than 20”, “at least 20” and “no more than 20” describe different events. For a discrete score, more than 20 means 21 upwards; at least 20 includes 20; no more than 20 includes everything through 20. “Between” can be ambiguous in ordinary speech, so the inequality signs or the examination wording should control the interpretation. Never silently add or remove endpoints.

If the observations are measurements grouped into intervals, an exact threshold inside an interval may not be recoverable from the table. You can know precisely that seven observations lie between 10 and 20 minutes while not knowing how many are above 16 minutes. An interpolation provides an estimate under an assumption; it does not recover the lost raw observations.

Separate the sample from the population

If 30 students complete a questionnaire, your immediate data describe those 30 responses. A claim about every Secondary 4 student in Bukit Timah needs more support: who was invited, who responded, which schools were included and whether the selection excluded some kinds of student. You do not need an advanced inference formula to recognise that a convenient sample can be an imperfect picture of a larger population.

In an exam answer, make the conclusion match the data. “Among these respondents, 40% selected option A” is narrower and more defensible than “40% of teenagers prefer A”. That small wording difference is part of mathematical accuracy. The percentage may be correct while the claim attached to it is too broad.

A twenty-second reading routine

Identify the people or objects, variable, units, sample size, scale and requested quantity. Then ask what information is missing. Only after those questions are settled should you decide whether the task needs addition, an average, a cumulative reading or a probability model. With practice, this becomes a quick habit rather than an additional essay written before every calculation.

For a parent observing revision, ask one question: “What does that 12 stand for?” A student who can answer “the number of observations in the third interval” has a much stronger start than one who says “the number I use in the formula”. The formula should follow meaning, not replace it.

3. Know what each statistical representation preserves and hides

Different diagrams preserve different information. A stem-and-leaf diagram can retain every individual value. A grouped histogram usually cannot. A box plot preserves five summary positions but not every score. A pie chart makes proportions visible but can hide the sample size. Revision becomes more efficient when you ask not only how to draw a diagram, but what questions that diagram can and cannot answer.

Bar graphs, pictograms and pie charts

A bar graph is useful for comparing frequencies across categories. Category order may be arbitrary, and gaps between bars help show that one category is not the next stretch of a continuous numerical interval. Inspect the axis before comparing apparent heights. A bar at 48 is not “twice as much” as a bar at 44 merely because a cropped axis starts at 40 and one visible column is twice as tall.

A pictogram requires its key. If one full symbol represents four students, three full symbols and a half-symbol represent fourteen students, not three and a half. A half-symbol is only meaningful because the key defines a quantity. If the problem uses three-dimensional-looking pictures of different sizes, identify whether area or height is changing; visual enlargement can exaggerate differences.

For a pie chart, a category’s fraction of the total equals its sector angle divided by 360 degrees. If a category contains 18 of 60 students, its angle is (18/60) × 360 = 108 degrees. Conversely, a 72-degree sector from a sample of 50 represents (72/360) × 50 = 10 students. Write the fraction before calculating; it keeps the reference whole visible.

A larger sector in one pie does not necessarily represent more people than a smaller sector in another. Forty percent of 30 is 12; thirty percent of 100 is 30. The second category has the smaller percentage and the larger count. Without the totals, the comparison of absolute numbers cannot be made. This is a common example of why “percentage of what?” is a mathematical question, not a stylistic detail.

Line graphs and dot diagrams

A line graph often represents change across ordered time points. Read both the vertical change and the horizontal interval. An increase of 20 over two months is not the same rate as an increase of 20 over one month. Also check whether joining points implies measurements between them. A straight line segment is a display choice; it does not prove that the real quantity changed at a perfectly constant rate throughout the interval.

A dot diagram places one dot for each observation, usually above a number line. It makes clusters, repeated values and gaps visible while preserving individual numerical values. If two dots are stacked above 7, that usually represents two observations of 7. Count all dots to verify the sample size. Do not read the height of the stack as a separate measurement.

Histograms with equal class intervals

A histogram groups numerical values into intervals. With equal widths and a vertical axis labelled frequency, heights can be compared directly as frequencies. The intervals, rather than arbitrary category names, give the horizontal scale its meaning. Contiguous classes normally meet because the measurement axis is continuous, even if a particular class has zero frequency.

Suppose task times are grouped into intervals 0–10, 10–20 and 20–30 minutes, with frequencies 4, 9 and 7. A taller middle bar means more observations lie in that interval. It does not identify the most common exact task time. Calling 15 minutes the mode would invent a raw value that the graph never supplied. The defensible statement is that 10–20 minutes is the modal class.

Optional extension: unequal widths. Where a question explicitly uses frequency density, area represents frequency and density equals frequency divided by class width. A width-10 class containing 20 observations has density 2; a width-20 class containing 30 has density 1.5. The wider class contains more observations despite its lower bar. This explains why the vertical-axis label and interval widths matter. It is not a reason to replace the equal-interval syllabus scope stated earlier with a different course’s histogram requirements.

A diagram-selection decision

Use a bar chart for category frequencies, a line graph for change across ordered time points, a stem-and-leaf diagram when a small numerical dataset should retain individual values, a histogram for grouped numerical distribution, and a box plot for a compact comparison of medians and spread. These are purposes, not absolute rules that make every alternative invalid.

When asked to explain a poor graph, identify the defect and its consequence. “The vertical axis begins at 40 rather than zero, making a four-unit difference appear much larger relative to the visible bars” is stronger than “the scale is wrong”. Similarly, “the sample totals differ, so the pie-sector percentages cannot establish which school has more students in that category” states the exact limit on inference.

4. Stem-and-leaf diagrams: preserve the observations

Consider these twelve invented scores out of 40: 12, 14, 14, 17, 20, 21, 21, 21, 25, 28, 31 and 36. A stem-and-leaf representation can be written as stem 1 with leaves 2, 4, 4, 7; stem 2 with leaves 0, 1, 1, 1, 5, 8; and stem 3 with leaves 1, 6. The key is “2 | 5 means 25 marks”. Without the key, readers cannot be certain what a stem and leaf represent.

The individual scores are still recoverable. There are twelve leaves, so there are twelve observations. The smallest is 12 and the largest 36. The range is 24 marks. The score 21 appears three times, so it is the mode. After ordering, the sixth and seventh observations are both 21, so the median is 21. Each answer uses a different feature of the same representation.

The key can change the entire answer

If the key instead says “2 | 5 means 2.5 minutes”, the same printed digits represent a different scale. The median is now 2.1 minutes and the range 2.4 minutes. The frequency and sample size do not change. This is a useful transfer exercise: alter the key and ask which features are preserved and which are scaled.

Leaves should be ordered within each stem when you are constructing a diagram for analysis. Unordered leaves can still record the values, but they make median positions and comparisons harder to read. Repeated leaves must remain repeated; removing duplicates changes the dataset. A score appearing three times contributes three observations, not one unique category.

Split stems and back-to-back diagrams

A split stem can use one row for leaves 0–4 and another for 5–9. The key must make that convention clear, and rows with no observations should not be confused with omitted parts of the scale. Splitting is a display choice that can reveal a crowded distribution more clearly. It does not change the underlying values.

In a back-to-back diagram, two groups share the stems. One group’s leaves may extend to the left, so read their ordering carefully instead of treating the printed left-to-right order as an ordinary list. Reconstruct several values using the key before calculating. Also count each group’s leaves separately; two groups sharing a diagram need not have the same sample size.

A comparison should use numerical evidence and context. If group A has a median completion time of 21 minutes and group B 25, group A has the lower typical completion time by that measure. That does not mean every person in A finished before every person in B. To compare consistency, you need an appropriate spread measure, not merely the location of the median.

Do not lose the unit during calculation

For marks, the mean, median, mode and range are expressed in marks. For time, they are expressed in time units. A frequency is a count of observations. These distinctions are simple enough to check quickly and powerful enough to expose many errors. If your “median” has units of students in a dataset of journey times, you have probably found a position or frequency rather than the required time.

A useful independent check is to reconstruct the sorted list from the diagram, hide the original, and calculate the summary statistics from that reconstructed list. If your answers differ from a diagram-based calculation, locate the first reading error rather than repeating all the arithmetic. Often the problem is one omitted leaf, one reversed stem or one forgotten repeated observation.

5. Mean, median and mode answer different questions

The mean is the total of the values divided by the number of observations. It can be understood as a fair-share value: if the total were redistributed equally, each observation would receive the mean. The median is the middle position after ordering. The mode is a value or category with the greatest frequency. They are not interchangeable labels for “the average”, because each preserves a different feature of the data.

Take the five values 4, 5, 5, 6 and 20. Their total is 40, so their mean is 8. Their median is 5 and their mode is 5. The mean is pulled upward by the large value 20. That does not make 8 an incorrect statistic; it means 8 answers a different question from the middle observation. An appropriate answer depends on whether the task concerns total allocation, a typical position or the most frequent outcome.

Why the mean need not be an observed value

No observation in that dataset equals 8. A mean is a summary, not necessarily a real member of the list. The same is true of an even-sized dataset’s median when the two central values differ. For 3, 5, 8 and 12, the median is (5 + 8)/2 = 6.5, although nobody recorded 6.5. Do not round that median to an observed value unless the problem explicitly requires a rounding convention.

If the values are whole-number counts, the mean can still be non-integer. A mean of 1.85 books per student does not claim that a particular child completed 0.85 of a listed book. It summarises total completed books divided by students. Whether a context permits fractional counts concerns individual observations, not the mathematical possibility of their mean.

Median positions are not median values

For an odd number n of ordered observations, the central position is (n + 1)/2. For an even n, use the observations in positions n/2 and n/2 + 1 and average their values. If n is 20, the two central positions are 10 and 11. The median is not automatically 10.5: that number describes a location between positions, not the measured values sitting there.

In the twelve-score example, positions six and seven both contain 21. In a frequency table, you may need cumulative counts to locate those positions. This is why ordering, frequency and median are one connected skill. Memorising a position formula without understanding what comes next does not finish the question.

Choosing a useful average

If a school is budgeting the total number of exercise books used, the mean can help link per-student use to a total. If one unusual journey took much longer than the others and the question asks about a typical journey, the median may be more representative of the central experience. If a shop needs to identify its most frequently sold size, the mode answers the category-frequency question directly.

Do not learn “median is always better with an outlier” as an unqualified slogan. An unusually large cost may be exactly the information a total-budget calculation must retain. The statistic should serve the decision. Explain which property matters in context: sensitivity to all values, resistance to a single extreme observation, or identifying the most frequent category.

A change-of-data thought experiment

Replace 20 in the list 4, 5, 5, 6, 20 with 50. The new total is 70 and mean 14. The median and mode remain 5. The range increases from 16 to 46. This shows precisely which summaries react to the altered observation. It is more useful than merely calling the mean “affected by outliers” because you can see the mechanism: the numerator in the mean changes by 30 while the denominator stays five.

Now add a new observation of 8 to the original list. Because 8 equals the original mean, the mean remains 8: total 48 divided by six. The median becomes (5 + 6)/2 = 5.5. One statistic can remain fixed while another changes. These questions train structural understanding and frequently expose overgeneralised rules about what adding a value must do.

6. Frequency tables and weighted means: respect the denominator

Return to the book-count table: x = 0, 1, 2, 3, 4 and frequencies f = 2, 5, 8, 4, 1. The total frequency is 20. The contributions to the total number of books are 0 × 2, 1 × 5, 2 × 8, 3 × 4 and 4 × 1. They sum to 37. Therefore the mean is 37/20 = 1.85 books per student.

Original practice data: books completed by 20 students

Books x	Students f	Total contribution fx	Cumulative frequency
0	2	0	2
1	5	5	7
2	8	16	15
3	4	12	19
4	1	4	20

The mode is 2 books, because that value has the largest frequency, 8. The median is also 2 books because the tenth and eleventh observations both lie among the eight students with two books. The modal frequency is 8 students; it is not the mode of the original book counts. Distinguishing those phrases is worth more than memorising another formula.

Why averaging the displayed x-values is wrong

Adding 0 + 1 + 2 + 3 + 4 and dividing by five gives 2. That calculation gives each distinct displayed value equal weight. The actual survey gives 2 books eight times the weight of 4 books because eight students recorded two and only one recorded four. Unless the frequencies are equal, an unweighted average of the row labels answers a different question.

The standard compact notation is mean = Σfx/Σf. Sigma means sum the indicated column. The numerator is the total of all observations after allowing for their frequencies; the denominator is the number of observations. Before using the formula, say those meanings aloud. If the denominator is the number of rows, the mistake is conceptual, not merely arithmetic.

Combining groups correctly

Suppose one revision group has 12 students with mean score 18, and another has 8 students with mean 23 on the same assessment scale. Their totals are 216 and 184 marks. The combined mean is (216 + 184)/(12 + 8) = 400/20 = 20. Averaging 18 and 23 gives 20.5 and would incorrectly give the smaller group the same weight as the larger one.

A useful check is that the combined mean must lie between the two group means when both groups have positive sizes. Here 20 lies between 18 and 23 and closer to 18, as expected because the 18-mean group has more students. This check cannot prove every weighted calculation correct, but it can reject a result outside the interval immediately.

Do not combine unlike quantities casually

Combining scores out of 20 with scores out of 50 is not the same task as pooling two groups on the same test. You first need to know what the combined statistic is intended to represent. A weighted assessment grade may attach prescribed weights to percentages, whereas a pooled mean of comparable raw scores weights by group size. Both involve weighting, but the weights serve different purposes.

For example, if a grade uses 30% from a quiz and 70% from an examination, convert each result to the appropriate percentage and then apply those policy weights. Do not silently substitute the number of students or the maximum marks as the weights. A student who asks “what does each weight represent?” is less likely to use a familiar formula in the wrong setting.

The same caution applies to comparing percentages across different totals. Twelve successes from fifteen attempts is 80%; eighteen from thirty is 60%. The first has a higher success rate, while the second has more successes. Both statements can be true. Decide whether the question asks for count, rate, total or average before choosing the arithmetic.

7. Missing observations, corrections and reverse questions

A mean is a relationship among three quantities: total, number of observations and mean. A reverse question gives you two of them and asks about the third. Start from total = number × mean. This is usually safer than trying to manipulate a remembered average formula while several unknowns are moving around.

Suppose five observations have mean 12, and four are 8, 11, 13 and 15. The required total is 5 × 12 = 60. The four known values sum to 47, so the missing observation is 13. Check by reconstructing the full list: 8 + 11 + 13 + 15 + 13 = 60. The duplicate 13 is allowed; a missing value does not have to be different from the others.

A wrongly recorded observation

Eight observations were reported to have a mean of 15. One value was entered as 16 when it should have been 24. The recorded total was 8 × 15 = 120. Remove the incorrect contribution and insert the correct one: 120 − 16 + 24 = 128. The corrected mean is 128/8 = 16. The number of observations stays eight because an entry was corrected, not added.

The sign of the correction provides a quick check. Replacing 16 with 24 increases the total by 8. Shared across eight observations, that increases the mean by 1. A corrected mean below 15 would contradict that direction. Students who routinely perform this check catch subtraction-order mistakes before a long recalculation is needed.

Compare this with a genuinely omitted observation. If eight recorded values have mean 15 and an additional value of 24 was omitted, the new total is 144 and the new number is nine. The new mean is 16, coincidentally the same answer in this particular example. The two mechanisms are nevertheless different. Numerical coincidence is not evidence that the denominator may be ignored.

Removing a group from a combined total

A group of 30 participants has mean score 18. Ten of them have mean 15. What is the mean for the remaining 20? The full total is 540, and the removed group’s total is 150. The remaining total is 390, giving mean 390/20 = 19.5. Subtracting the two means, 18 − 15, would have no sensible interpretation as the remaining group’s mean.

This is the same weighted-mean idea in reverse. A mean belongs to a particular collection of observations. Before combining or separating groups, recover their totals. You must also establish that the groups do not overlap and that their scores use the same scale. Otherwise subtracting totals may remove some observations twice or combine incomparable quantities.

An unknown frequency

A table has values 1, 2 and 3 with frequencies 2, k and 4. Its mean is 2.25. The total frequency is 6 + k and the weighted total is 14 + 2k. Therefore (14 + 2k)/(6 + k) = 2.25. Multiplying gives 14 + 2k = 13.5 + 2.25k, so 0.5 = 0.25k and k = 2. Check: the total frequency is eight, the weighted total is eighteen and 18/8 = 2.25.

A frequency must be a non-negative integer. That condition is part of the model, not a cosmetic final check. If an algebraic solution produces a negative or fractional number of students, either the given information is inconsistent or an earlier equation is wrong. Do not round a frequency to make the answer look realistic; revisit the mathematics.

Why the median behaves differently

A missing-value median question is about ordering, not simply totals. If the five values are 2, 4, x, 9 and 12 and their median is 9, several values of x may work: x must be at least 9. The list as printed need not already be ordered. By contrast, a specified mean normally gives a total equation that can identify one missing value.

This difference explains why some questions ask for a range of possible values. Do not force every reverse-statistics question into one equation. Ask which summary is being constrained. Mean constrains the total; median constrains central positions; mode constrains repetition; range constrains endpoints. The form of the information should determine the form of your reasoning.

8. Spread: why the same average can hide different experiences

Consider two invented sets of completion times. Set A is 18, 19, 20, 21, 22 minutes. Set B is 10, 15, 20, 25, 30 minutes. Both have mean 20 and median 20. Yet A is tightly clustered, while B includes both much shorter and much longer times. A statement that the groups are “the same because the averages are equal” loses the most important difference.

Range, interquartile range and standard deviation describe different aspects of dispersion. The range uses the endpoints. The interquartile range focuses on the central half of the ordered data. Standard deviation uses every observation’s distance from the mean. No one measure tells the complete story, and the most useful comparison depends on the question.

Range is fast, but it uses only two observations

The range of A is 22 − 18 = 4 minutes. The range of B is 30 − 10 = 20 minutes. This supports the conclusion that B has a wider overall span. However, the range alone cannot tell whether most observations are tightly clustered with one unusual endpoint or spread throughout the interval.

For example, the lists 10, 20, 20, 20, 30 and 10, 11, 12, 29, 30 both have range 20. Their internal shapes differ. The range discards everything except the smallest and largest values. When a question asks about the middle of the distribution, a different spread measure may be more informative.

Quartiles divide positions, not the numerical scale

Quartiles locate positions in ordered data. They are not found by dividing the maximum value by four. If journey times run from 12 to 36 minutes, the lower quartile is not automatically 9 minutes. That would even lie below the minimum in this example. The values must first be ordered and their positions identified.

For the twelve scores used earlier—12, 14, 14, 17, 20, 21, 21, 21, 25, 28, 31, 36—use the median-of-halves convention for this particular exercise. The lower six observations have central values 14 and 17, giving Q1 = 15.5. The upper six have central values 25 and 28, giving Q3 = 26.5. The interquartile range is 26.5 − 15.5 = 11 marks.

Small raw datasets can produce different quartile answers under different legitimate conventions, especially when there is an odd number of observations. Follow the convention specified in your question or taught for your course. Do not mix a calculator’s interpolation rule with a teacher’s median-of-halves rule halfway through a solution. Cumulative-frequency graph questions use positions on the cumulative scale and should be handled as that different representation requires.

What “more consistent” must mean

In a comparison of journey times, a smaller interquartile range means the central half occupies a narrower interval. A smaller standard deviation means less dispersion around the mean by that measure. Write the measure you used. “A is more consistent because its IQR is 6 minutes rather than 12” is more precise than “A is better”. Whether smaller times are desirable is a separate question about the context.

Consistency is not the same as high performance. A set of uniformly low scores can be very consistent. A runner with consistently slow times may be reliable but not the fastest. A good comparison normally separates centre from spread: which group has the higher or lower typical value, and which group varies less?

Changes that preserve spread

Add 5 to every observation in A. The new values are 23, 24, 25, 26, 27. The mean increases to 25, but every pairwise difference stays unchanged. The range remains 4 and standard deviation remains the same. Adding a constant shifts the whole distribution; it does not stretch it.

Multiply each original value by 2 instead. The values become 36, 38, 40, 42, 44. The mean doubles and the range doubles to 8. Standard deviation also doubles because all distances from the mean double. These transformation questions are best solved by understanding what changed, not by entering a complete dataset again.

When converting minutes to seconds, multiply by 60. The mean and standard deviation both acquire a factor of 60 and are expressed in seconds. Variance, if used in working, acquires a factor of 60 squared and has squared units. Unit reasoning can expose an incorrect transformation even when the arithmetic looks plausible.

9. Standard deviation without mysterious calculator buttons

Standard deviation measures the typical scale of deviation from the mean using squared distances. Subtract the mean from each observation, square those deviations, average them using the relevant denominator, then take the square root. Squaring prevents positive and negative deviations cancelling and gives greater influence to larger distances. Taking the square root returns the answer to the original units.

For the school-level population standard deviation used in the examples here, the denominator is the number of observations N. With a frequency table, N = Σf. The computational form is σ = √[Σfx²/Σf − (Σfx/Σf)²]. The official K310 syllabus formula list provides the relevant mean and standard-deviation expressions. Use the statistic your actual question requires rather than selecting a calculator output because it appears first.

A small example you can check by hand

For A = 18, 19, 20, 21, 22, the mean is 20. Deviations are −2, −1, 0, 1, 2. Their squares are 4, 1, 0, 1, 4, with total 10. Divide by five to obtain variance 2, then take the square root: σ = √2, approximately 1.41 minutes. Notice that the unsquared deviations sum to zero, which is why merely averaging signed distances would not measure spread.

For B = 10, 15, 20, 25, 30, the mean is also 20. Squared deviations are 100, 25, 0, 25, 100. Their mean is 50, so σ = √50, approximately 7.07 minutes. The larger value reflects the greater spread already visible in the list. This numerical calculation supports, rather than replaces, the initial inspection.

Use frequencies correctly

For the twenty-student book table, Σf = 20 and Σfx = 37. The additional column fx² is 0, 5, 32, 36, 16, giving Σfx² = 89. Therefore the variance is 89/20 − (37/20)² = 4.45 − 3.4225 = 1.0275. The standard deviation is √1.0275, approximately 1.01 books.

Weighted squares for the book-count example

x	f	x²	fx²
0	2	0	0
1	5	1	5
2	8	4	32
3	4	9	36
4	1	16	16

The notation fx² means f multiplied by x squared. It does not mean (fx)². For x = 2 and f = 8, the correct contribution is 8 × 4 = 32, representing eight observations each with square 4. Squaring the weighted contribution would give 16² = 256, which represents a completely different operation. Writing the x² column first is a useful temporary scaffold if this error repeats.

Population and sample outputs are not interchangeable

Many calculators provide both a population standard deviation and a sample standard deviation. The latter commonly uses N − 1 in its variance denominator. In the five-value example A, the population standard deviation is √2, while the N − 1 version is √2.5. Both are legitimate statistics in appropriate settings, but they answer different conventions. For these exercises, select the population version.

A calculator cannot know which convention the question intends. Check your model’s manual for the names or symbols of the outputs. Do not rely on a universal button sequence: different approved calculators organise their menus differently. The reliable part of the method is knowing which quantity you need and verifying N, the mean and the relevant total before reading the output.

Interpret the result with the mean

If two sets have the same mean and one has a smaller standard deviation, its values are less dispersed around that common mean. If the means differ, compare both measures and keep the context in view. A higher mean score may be desirable; a higher mean waiting time usually is not. Standard deviation does not decide what “better” means.

Standard deviation is never negative. It equals zero when every observation is identical. It has the same units as the observations, not squared units. A very large result for a tightly clustered dataset is a reason to inspect entry, frequency and brackets. A tiny negative variance produced only by rounding nearly equal quantities is not a reason to take a square root of a negative number; retain more precision and recompute.

10. Grouped data: calculate estimates honestly

When individual observations are grouped into intervals, information is lost. If a table says twelve journeys took more than 20 minutes and no more than 30, it does not identify each journey’s exact duration. We commonly use the class midpoint as a representative value to estimate a mean or standard deviation. The answer should be described as an estimate because several different raw datasets could produce the same grouped table.

Original grouped practice data: 30 journey times

Time t in minutes	Frequency f	Midpoint m	fm	fm²
0 < t ≤ 10	3	5	15	75
10 < t ≤ 20	7	15	105	1575
20 < t ≤ 30	12	25	300	7500
30 < t ≤ 40	6	35	210	7350
40 < t ≤ 50	2	45	90	4050

The frequencies sum to 30. The weighted midpoint total is 720, so the estimated mean is 720/30 = 24 minutes. The weighted squared-midpoint total is 20,550. The estimated population variance is 20,550/30 − 24² = 685 − 576 = 109, giving estimated standard deviation √109, approximately 10.4 minutes.

The midpoint calculation does not prove that all twelve journeys in the third class took 25 minutes. It substitutes one representative value for unknown values in that interval. Avoid writing an explanation that quietly turns the approximation into exact observations. This distinction becomes especially important when comparing a raw-data answer with a grouped-data answer from the same survey.

Read interval boundaries carefully

In this table, a journey of exactly 20 minutes belongs to the second interval, 10 < t ≤ 20, not the third. A journey of exactly 30 belongs to the third. The inequalities avoid double counting at the boundaries. When classifying a value, read the printed inequality instead of using a vague phrase such as “the twenties group”.

The modal class is 20 < t ≤ 30 because it has the highest frequency, 12. The exact mode of the raw data cannot be recovered. Even its midpoint, 25, is not necessarily an observed value. A question asking for the modal class wants an interval; a question asking for a mode from raw observations wants a value or values with greatest frequency.

Why you cannot average the midpoints alone

The five midpoints have unweighted mean 25, but the estimated mean of the thirty journeys is 24. The difference arises because interval frequencies are unequal. Averaging midpoints alone would treat each interval as containing the same number of observations. The table does not say that.

Similarly, the median is not found by taking the middle midpoint simply because there are five classes. The sample contains thirty journeys, not five journeys. Locate central observation positions using cumulative frequency. Their values lie in a class, and obtaining a more precise estimate requires a graph or an explicit interpolation assumption.

Reasonableness checks for grouped answers

The estimated mean must lie between the smallest and largest midpoints when frequencies are non-negative. It should be influenced most strongly by the classes with many observations. Here the large central frequency makes a mean near the middle plausible. A mean of 48 or −2 would conflict with the table’s structure.

For continuous intervals, an exact minimum or maximum is not normally recoverable from class boundaries. The first class beginning above zero does not prove that a journey took zero minutes, and the final upper boundary of 50 does not prove that the longest journey was exactly 50. You can report interval bounds and estimates with appropriate language, but should not invent raw observations.

A useful extension is to imagine all observations clustered near each class’s lower boundary, then near its upper boundary. The grouped frequencies stay unchanged while the true mean changes. That thought experiment explains why the midpoint method is an approximation more clearly than memorising the word “estimated” at the end of an answer.

11. Cumulative frequency: move from positions to values

Cumulative frequency answers a counting question: how many observations are at or below a stated boundary, under the convention used in the table? It is a running total. Therefore it should never decrease as the boundary increases, and its final value should equal the sample size. Those two properties are excellent checks before interpreting any graph.

Use a new invented set of forty journey times. Its cumulative counts at upper boundaries 10, 20, 30, 40 and 50 minutes are 4, 12, 28, 36 and 40. Include the starting point (0, 0) for this example. The corresponding class frequencies are 4, 8, 16, 8 and 4, found by subtracting consecutive cumulative totals.

Cumulative-frequency coordinates for the forty-journey example

Upper boundary, minutes	Cumulative frequency	Frequency added in interval
0	0	—
10	4	4
20	12	8
30	28	16
40	36	8
50	40	4

The median is a value read at a position

For the cumulative graph, the median is estimated at cumulative frequency N/2 = 20. Begin at 20 on the frequency axis, move horizontally to the graph, then down to the time axis. You are looking for a time, not reporting “20 students” as the median.

If the plotted points are joined with straight segments for this exercise, cumulative frequency rises from 12 to 28 between 20 and 30 minutes. Position 20 is halfway through that frequency increase, so the estimated median is 25 minutes. This is interpolation within grouped information. The exact middle raw journey times remain unknown.

Quartiles use the same route

Q1 is read at N/4 = 10. This lies between cumulative frequencies 4 and 12. It is 6/8 of the way through that increase, so the straight-segment estimate is 10 + (6/8) × 10 = 17.5 minutes. Q3 is read at 3N/4 = 30, which is 2/8 of the way from cumulative frequency 28 to 36. Its estimated time is 30 + (2/8) × 10 = 32.5 minutes.

The estimated IQR is 32.5 − 17.5 = 15 minutes. A common error is to subtract the positions 30 − 10 = 20 and call that the IQR. That subtraction measures twenty observations, not the time spread. The frequency positions help you find Q1 and Q3; the final subtraction uses the values on the time axis.

Follow your actual question’s graph instructions. A supplied smooth curve may yield slightly different visual readings from straight-line interpolation. Do not force the numbers above onto a different graph. Read with the scale’s realistic precision and show construction lines where useful. Reporting several decimal places from a coarse graph suggests more precision than the representation supports.

Count above a threshold by subtracting from the total

At 30 minutes the cumulative frequency is 28, so twelve of the forty journeys took more than 30 minutes. The relative frequency is 12/40 = 0.3. At 35 minutes, the straight-segment estimate gives cumulative frequency 32, leaving an estimated eight journeys above 35 and an estimated proportion 0.2.

The distinction between the two answers matters. Thirty is a supplied boundary with a recorded cumulative count. Thirty-five lies inside a grouped interval and requires estimation. Do not describe the latter count as directly observed. For “between” questions, subtract cumulative counts at the two boundaries and check whether the inequality conventions include the relevant endpoints.

Percentile language

A percentile is a position within the ordered distribution. The 80th percentile corresponds to cumulative position 0.8N, which is 32 here. Under the same interpolation model, it occurs at 35 minutes. It does not mean 80 minutes, 80 students or an 80% mark. The measured variable may not be a score at all.

When comparing cumulative curves for groups of different sizes, raw vertical counts can mislead. A cumulative count of 20 represents half of a forty-person group but only a quarter of an eighty-person group. Compare equivalent percentiles or proportions rather than equal cumulative counts when the purpose is to compare distributions.

12. Box-and-whisker plots: make a comparison the data supports

A box plot compresses a distribution into a small number of position summaries. In the simple minimum-to-maximum version used here, the five-number summary is minimum, lower quartile, median, upper quartile and maximum. The box spans Q1 to Q3, and a line inside marks the median. Whiskers extend to the minimum and maximum. Some other statistical conventions plot outliers separately and use different whisker rules; read the question’s definition rather than importing a different software convention.

Consider two invented completion-time summaries. Group A has minimum 18, Q1 21, median 24, Q3 27 and maximum 36 minutes. Group B has minimum 12, Q1 19, median 25, Q3 31 and maximum 43. The summaries are ordered correctly in both cases, which is the first construction check.

Original box-plot comparison data, in minutes

Group	Minimum	Q1	Median	Q3	Maximum
A	18	21	24	27	36
B	12	19	25	31	43

Two comparisons, not one vague judgement

A has the lower median time, 24 compared with 25 minutes. Its IQR is 6 minutes, compared with B’s 12. Therefore A has a slightly lower central completion time by the median and a tighter central-half spread. Its overall range is also smaller: 18 compared with 31 minutes. These are specific conclusions with numerical support.

It would be wrong to say every person in A is faster than every person in B. B’s minimum is 12, below A’s minimum. It would also be wrong to conclude A has a higher mean score: the variable is time, and the mean is not given by these five summaries. Keep both the variable and the available statistic explicit.

The median need not sit in the centre of the box

The distances Q1 to median and median to Q3 describe different numerical spans containing corresponding positional portions of the data. Equal numbers of observations do not require equal interval lengths. If the median is near one side of the box, the adjacent quarter of observations occupies a narrower numerical interval.

However, a box plot does not reveal every internal feature. You cannot identify all modes, recover individual observations or count exact ties from the five-number summary alone. A long upper whisker can suggest a longer upper tail in the displayed range, but it does not identify the cause. Do not invent a story about one lazy student or one traffic accident from the shape of a statistical display.

Sample sizes and percentages

Two boxes drawn with the same width do not imply equal sample sizes. The plot summarises positions, not the number of people represented by the thickness of the box. If one group has twenty observations and another eighty, their central halves correspond to different numbers of observations even when their IQRs are identical.

Quartile statements also need care with ties and small datasets. Informal language such as “about a quarter of the data lies below Q1” describes positional division, not an automatic exact count strictly below that numerical value. If several observations equal Q1, the strict inequality count can differ. Use exact counts only when the raw data or the question’s convention supports them.

A reliable comparison sentence

Write: “Group A has a lower median completion time, 24 minutes versus 25 minutes, and a smaller interquartile range, 6 minutes versus 12 minutes. Its central completion times are therefore slightly quicker and less dispersed.” This answers centre, spread and meaning without claiming that the entire distributions are ordered.

For test marks, the direction of preference may reverse: a higher median may be desirable. For manufacturing lengths, neither higher nor lower is automatically better; closeness to a target matters. The statistic describes the data, while the context determines the decision. Practising that distinction makes interpretation questions less dependent on memorised adjectives.

13. Misleading conclusions: when correct arithmetic is not enough

A calculation can be exact while the conclusion is unjustified. A survey may cover only a particular group, a percentage may have a hidden denominator, or two groups may differ before the event being compared. Statistics revision should therefore include a final question: “What does this information allow me to conclude, and what additional information would I need?”

Suppose a class survey finds that 18 of 30 respondents prefer option A. The observed proportion is 60%. It does not prove that 60% of every student in the school prefers A, because the respondents may not represent the school. If only enthusiasts answered an optional survey, the selection process could influence the result. A larger number of such respondents would not automatically remove that particular problem.

Sample size is not the same as representativeness

A sample of 500 people recruited from one club can still be poorly suited to describing all residents’ interests. A smaller sample selected appropriately for the question may provide a better basis for that specific inference. For school questions, identify the population the claim concerns and whether the sampling method gives relevant members a reasonable chance to be represented.

Do not replace one overclaim with another by writing “small samples are always wrong”. Small samples contain real information; they simply support limited conclusions and may vary more from sample to sample. The appropriate response is to state the limitation, not dismiss every observed value. An individual measurement can be useful without being a complete national survey.

Percentages need their base

Suppose group A improves from 40 to 50 points and group B from 80 to 90. Both gain ten points. A’s relative increase is 10/40 = 25%; B’s is 10/80 = 12.5%. Different statements describe different comparisons. Saying A improved “twice as much” without specifying relative percentage could mislead a reader who assumes the comparison concerns points gained.

Similarly, an increase from a 20% rate to a 25% rate is five percentage points but a 25% relative increase in the rate. The difference between percentage and percentage points is not merely wording. It changes the denominator and therefore the numerical claim. In an examination explanation, name both quantities when needed.

Association is not an explanation of cause

If students who practise more have higher marks in a table, the table shows an association in that dataset. It does not isolate the effect of practice from prior attainment, topic difficulty, attendance or other differences. The reasonable statement is narrower than “practice caused every improvement”. This guide’s suggested routines are teaching plans, not measured intervention results.

The same caution applies to comparing tutoring groups or schools. Different starting points, selection and assessments can make raw averages incomparable. A parent should ask what was measured, on which scale, for whom, and against what baseline. A percentage without that context is not a complete account of learning.

Optional extension: changing group weights can reverse a comparison

Here is a fully invented example to make the issue visible. In an easier task, method A succeeds 90 times out of 100, while B succeeds 19 times out of 20. In a harder task, A succeeds once out of 10 and B succeeds four times out of 20. B has the higher success rate within both task types: 95% versus 90%, and 20% versus 10%.

Yet A’s combined rate is 91/110, approximately 82.7%, while B’s is 23/40 = 57.5%. The aggregate comparison reverses because A was used much more often on the easier task. Nothing mathematically contradictory happened. The weights changed the overall average. This extension is not an additional syllabus requirement; it is a demonstration of why comparing totals without considering group composition can produce a misleading story.

The practical lesson is modest: when a conclusion seems surprisingly strong, inspect the denominators and subgroup structure before accepting it. You do not need advanced statistical software to ask that question. The same weighted-mean reasoning introduced earlier already provides the tool.

14. Probability begins with a model, not an operation

Probability describes uncertainty under a specified model. Before calculating, identify the experiment, the outcomes, the event of interest and the assumptions about how outcomes occur. A bag, spinner, die or survey does not provide those assumptions automatically. The phrase “chosen at random” needs to connect to physical objects or people, not merely the labels you find convenient to count.

In a finite sample space whose outcomes are equally likely, the probability of an event is the number of favourable outcomes divided by the total number of outcomes. The equally likely condition is essential. Counting three colour names and assigning each probability one third would be wrong if the bag contains four red tokens, two blue tokens and one green token and every physical token is equally likely to be selected.

Count objects before counting categories

In that seven-token bag, P(red) = 4/7, P(blue) = 2/7 and P(green) = 1/7. The probabilities sum to one because the colours are exhaustive and mutually exclusive for a single draw. You can verify the model by labelling the physical tokens R1, R2, R3, R4, B1, B2 and G1. Those seven labelled outcomes are equally likely; the three colour categories are not.

A spinner requires similar care. Suppose sectors A and B each occupy 90 degrees and C occupies 180 degrees. Under an ideal model in which stopping angle is uniformly distributed, the probabilities are 1/4, 1/4 and 1/2. Three sector labels do not imply equal probabilities. The geometric model, rather than the number of letters, supplies the weights.

For a school problem describing a fair die, each face has probability 1/6. For two independent fair dice, each ordered pair has probability 1/36. Both fairness and independence matter. If the second displayed number were forced to copy the first, each display could still have a uniform marginal distribution, but only six matching pairs would occur. The ordinary thirty-six-pair model would no longer describe the experiment.

Experimental relative frequency

If a spinner lands on A 27 times in 100 trials, its observed relative frequency is 27/100. That is a description of those trials, not proof that its theoretical probability is exactly 0.27. It also does not prove an ideal quarter-circle spinner is unfair simply because 27 differs from 25. Random samples need not reproduce model proportions exactly in a finite number of trials.

If a question asks for an estimated number of A outcomes in another 200 similar trials using the observed relative frequency, 200 × 0.27 = 54 is an estimate. It is not a guarantee. If the question instead supplies a theoretical probability of 1/4 and asks for an expected count under that model, the answer is 50. Read which source of probability the question authorises.

Interpret zero, one and intermediate values carefully

For the finite models in this guide, probability zero means no listed outcome belongs to the event, and probability one means all listed outcomes do. An intermediate probability expresses uncertainty. A probability of 0.8 does not mean the event must happen exactly eight times in every ten trials, and it does not mean it is scheduled for the next attempt after two failures.

Always check that a calculated probability lies between zero and one. That check is necessary but not sufficient: a wrong answer of 1/2 can still satisfy the bounds. The stronger check is to revisit the event and sample space. Ask which outcomes are included, which are excluded and whether each is weighted correctly.

15. Build a complete sample space without double counting

A sample space is a structured description of all possible outcomes. Its purpose is to prevent omission and duplication. An organised table, list or tree is often more reliable than mentally guessing which possibilities “seem different”. The representation should match whether order matters in the experiment and whether outcomes have equal weights.

Two dice: sums are not equally likely

Roll two independent fair six-sided dice and record the first and second results. There are 36 equally likely ordered pairs. A total of 8 arises from (2,6), (3,5), (4,4), (5,3) and (6,2), so P(total 8) = 5/36. The possible totals 2 through 12 are not eleven equally likely outcomes. Total 2 has one contributing pair, while total 7 has six.

The word “ordered” means (2,6) and (6,2) are different elementary outcomes because the first die and second die have switched results. The event “total 8” includes both. You can also solve using unordered pairs, but then their probabilities are not all equal: a non-matching pair has two orderings while a double has one. Switching representation without adjusting weights is a common source of error.

A six-by-six table makes completeness visible. Rows represent the first die, columns the second. For total 8, mark the five appropriate cells. For a product that is even, mark every cell where at least one die is even. The latter event is easier by its complement: the product is odd only when both dice are odd. There are 3 × 3 = 9 such cells, leaving 27 even-product cells and probability 27/36 = 3/4.

Three coin tosses: describe sequences, not just totals

For three independent fair tosses, the eight sequences are HHH, HHT, HTH, HTT, THH, THT, TTH and TTT. Exactly one head occurs in HTT, THT and TTH, giving probability 3/8. The counts of heads—0, 1, 2, 3—are not four equally likely outcomes. Counts one and two each have three sequences, while zero and three each have one.

When writing a list, use a fixed ordering rule. Start with all sequences beginning H, then all beginning T, and organise the second and third positions similarly. A casual list may repeat HTH and omit THH without the student noticing. A tree or systematic binary ordering provides a completeness check: after each fair toss, the number of sequences doubles.

Selections: named people, categories and order

Suppose two students are selected in sequence from three named students, A, B and C, without selecting the same person twice. The ordered outcomes are AB, AC, BA, BC, CA and CB. There are six. If the question asks whether A is selected at either stage, favourable outcomes are AB, AC, BA and CA, giving probability 4/6 = 2/3 under uniform sequential selection.

If instead the experiment records only the final two-person committee, the unordered outcomes are {A,B}, {A,C} and {B,C}. Under the same fair selection mechanism these three committees are equally likely, and A belongs to two, again giving 2/3. Both methods work because the representation and weighting stay consistent. Mixing six possible ordered outcomes with only two favourable unordered committees would not.

A sample-space audit

Check that every possible outcome appears exactly once. Check whether impossible repetitions have been excluded when replacement is not allowed. Check whether outcome labels represent equally likely elementary events or grouped categories. Finally, check that the event wording matches the marked outcomes. “At least one”, “exactly one” and “both” are three different selections from the same sample space.

For two dice, exactly one six consists of ten pairs: five with the first die six and second not six, plus five in the opposite order. At least one six includes those ten plus (6,6), giving eleven. Thus the probabilities are 10/36 and 11/36 respectively. One extra cell captures the entire difference between “exactly” and “at least”.

16. Event language: and, or, exclusive and independent

The words in a probability question determine the set of outcomes required. “A and B” means both conditions hold. “A or B” normally includes outcomes satisfying either condition or both, unless the question explicitly says exactly one. “Not A” means the complement within the stated sample space. Translate the words into outcomes before choosing addition or multiplication.

Mutually exclusive means no shared outcome

On a single fair die, event A = “roll a 1” and event B = “roll a 6” are mutually exclusive. One roll cannot satisfy both. Therefore P(A or B) = 1/6 + 1/6 = 1/3. Addition works here because no outcome is counted twice.

Now let A = “roll an even number” and B = “roll a number greater than 4”. A contains {2,4,6}; B contains {5,6}. They overlap at 6. Adding 3/6 + 2/6 counts that outcome twice, so subtract 1/6. The union is {2,4,5,6} and its probability is 4/6 = 2/3.

The general union relationship is P(A or B) = P(A) + P(B) − P(A and B). This is counting with a correction for overlap. It does not depend on events being independent. A Venn diagram or a two-way table can make the overlap easier to identify than a formula alone.

Independent means one event does not change the other’s probability

For two independent fair coin tosses, the first being heads does not change the chance that the second is heads. P(first H and second H) = 1/2 × 1/2 = 1/4. Both events can occur together, so independence does not mean mutual exclusivity.

In fact, two mutually exclusive events with positive probabilities cannot be independent: their intersection has probability zero, but multiplying their positive probabilities gives a positive number. That contradiction is a direct way to remember that the two terms describe different relationships. Exclusive concerns overlap; independent concerns whether conditioning changes probability.

Why “and means multiply” is incomplete

Multiplication for a two-stage event uses the probability of the first stage and the appropriate probability of the second stage after that first stage. Under independence, the second probability equals its original value. Without independence, it may change. The multiplication remains valid when the changed probability is used correctly.

Drawing two red tokens without replacement illustrates this. The second red probability depends on what was removed first. You still multiply along the two-red path, but you do not multiply the original red fraction by itself. The next section makes the changing denominator and numerator explicit.

Why “or means add” is incomplete

Add probabilities of disjoint alternatives. If the alternatives overlap, account for the overlap or rewrite the event as disjoint cases. “First draw red or second draw red” includes the two-red outcome in both descriptions. Simply adding the marginal probabilities double counts that shared event.

A safe method is to state the alternatives in full sentences. For exactly one red in two draws, the alternatives are “red then not red” and “not red then red”. Those paths cannot both occur in the same two-draw sequence. They can therefore be added. Naming the paths removes the ambiguity hidden in a short mnemonic.

Use definitions to repair an error

When an answer is wrong, do not label it merely “used addition instead of multiplication”. Ask why the event was misrepresented. Did the student combine stages as alternatives? Did they forget overlap? Did they assume independence? A useful correction changes the model, not only the arithmetic symbol.

The same habit helps parents who do not remember probability notation. Ask the student to describe one successful outcome and one unsuccessful outcome. If those examples are unclear, the event itself has not been understood. Calculating before fixing that uncertainty is unlikely to help.

17. Tree diagrams and replacement: follow the experiment as it changes

A probability tree represents a sequence of stages. A branch label is the probability of the next outcome given the path already taken. Outgoing branches from a node must cover all available possibilities without overlap, so their probabilities sum to one. Multiply along a complete path; add the probabilities of disjoint complete paths belonging to the target event.

Use a bag containing four red tokens and three blue tokens. Select two tokens at random without replacement. The first-stage probabilities are 4/7 for red and 3/7 for blue. After red, the bag contains three red and three blue, so the second-stage probabilities are 3/6 and 3/6. After blue, it contains four red and two blue, so the second-stage probabilities are 4/6 and 2/6.

Two draws without replacement from four red and three blue tokens

Path	Calculation	Probability	Remaining condition used
RR	(4/7)(3/6)	2/7	One red removed first
RB	(4/7)(3/6)	2/7	Blue count unchanged after red
BR	(3/7)(4/6)	2/7	Red count unchanged after blue
BB	(3/7)(2/6)	1/7	One blue removed first

The four path probabilities sum to one. The probability of different colours is P(RB) + P(BR) = 4/7. The probability of the same colour is P(RR) + P(BB) = 3/7. These two events are complements, giving another independent check on the calculations.

What replacement changes

Now replace the first token and mix the bag before the second draw. The composition returns to four red and three blue. Under the stated random-draw model, the second probabilities are again 4/7 and 3/7 regardless of the first colour. RR has probability 16/49, BB has probability 9/49, and the two different-colour paths each have probability 12/49.

Thus same colour has probability 25/49 and different colours 24/49. These differ from the without-replacement answers. The difference is not a minor wording detail: replacing the first token resets the experiment. “Put back” without remixing or a specified random selection mechanism may raise practical modelling questions, but ordinary school exercises normally state the ideal process to use.

A second way to verify without replacement

Label all seven physical tokens. There are 7 × 6 = 42 equally likely ordered pairs of distinct tokens. Different colours can occur in 4 × 3 = 12 red-blue pairs and 3 × 4 = 12 blue-red pairs. Therefore the probability is 24/42 = 4/7, agreeing with the tree.

This counting check is especially useful when a tree contains fractions that look suspicious. It uses the same experiment but a different representation. Agreement is not a guarantee against every possible error, yet it is stronger than repeating the same multiplication with the same mistaken branch labels.

Unequal first outcomes and symmetric final paths

Notice that red is more likely than blue on the first draw, but RB and BR have equal probabilities. Both final events choose one of four red tokens and one of three blue tokens in opposite orders. This equality is not a universal rule for every sequential process; it follows from this particular random selection without replacement.

If the first result causes a rule change—for example, adding a token, switching bags or using a different spinner—the two mixed paths may no longer be equal. Read the process after each branch. A probability tree is a record of the experiment, not a decorative template whose second level always copies the first.

Three stages without replacement

Select three tokens from the same four-red, three-blue bag without replacement. The probability of no red is P(BBB) = (3/7)(2/6)(1/5) = 1/35. Consequently, the probability of at least one red is 34/35. The complement saves the work of adding seven red-containing colour paths.

For exactly one red, the disjoint colour paths are RBB, BRB and BBR. Each has probability 4/35: for instance, RBB gives (4/7)(3/6)(2/5) = 4/35. Their sum is 12/35. Write at least one path in full before using equality among the three. The justification is the same set of one red and two blue selections arranged in different orders under this experiment.

Tree checks that cost little time

At every node, count what remains. Check that outgoing branches sum to one. After constructing all complete paths, their probabilities should sum to one. Check whether your target paths overlap; complete sequences are disjoint, but vague descriptions may not be. Finally, compare with an intuitive bound: the probability of both draws being red cannot exceed the probability that the first is red.

Do not rely only on the final probability being between zero and one. Using 4/7 for the second red without replacement gives 16/49, which lies within the bounds but models replacement instead. The branch-count explanation is the decisive check.

18. At least, exactly and none: choose the simpler event

The complement of an event consists of all outcomes not in it. Since an event and its complement partition the sample space, their probabilities sum to one. This becomes especially useful when the event has many successful paths but the failure event has only one or a few.

For two independent fair dice, at least one six has probability 1 − P(no six) = 1 − (5/6)² = 11/36. Exactly one six instead requires one six and one non-six, in either order: 2(1/6)(5/6) = 10/36. The two answers differ by P(two sixes) = 1/36. The complement of “exactly one six” is not merely “no six”; it also includes two sixes.

Three independent attempts with unequal success and failure probabilities

Consider a hypothetical device trial with success probability 0.8 on each attempt. Assume attempts are independent and the probability stays constant. The probability of no success in three attempts is 0.2³ = 0.008, so at least one success has probability 0.992. These numbers describe an invented model, not a claim about a real product.

Exactly two successes has three disjoint sequences: SSF, SFS and FSS. Each has probability 0.8² × 0.2 = 0.128, giving total 0.384. There is no need to introduce a new named distribution to solve this small experiment. List the paths, justify their probabilities and add the disjoint alternatives.

The probability of exactly one success is 3 × 0.8 × 0.2² = 0.096. All three successes has probability 0.8³ = 0.512. The four count probabilities—0.008, 0.096, 0.384, 0.512—sum to one. This count-level check is a useful way to verify several answers from the same tree.

Expected counts are not promised counts

Under the same constant-probability model, the expected number of successes in ten attempts is 10 × 0.8 = 8. That does not mean exactly eight must occur. A count can be above or below its expected value. Expected value is an average over repeated versions of the experiment, not a schedule controlling an individual sequence.

Likewise, failing repeatedly does not automatically make success due on the next independent trial. If the probability is fixed at 0.8 and trials are independent, the next success probability is still 0.8. If real conditions change, that is a different model and needs to be stated. Do not use a belief about balancing past results as a substitute for the independence assumption.

Without replacement is not the same repeated-trial model

The three-token bag example cannot use the same constant blue probability at every stage. P(BBB) is (3/7)(2/6)(1/5), not (3/7)³. Both calculations describe coherent but different experiments. The first removes tokens, the second effectively restores the original proportions at each independent stage.

Complement methods are powerful only when the complement is modelled correctly. “One minus something” is not automatically safer than direct enumeration. Write the failure event in words, check that it truly is the complement, then calculate it with the appropriate changing or constant probabilities.

When the direct method is shorter

For exactly two heads in three fair tosses, three simple paths give 3/8. Using the complement would require counting zero, one and three heads, which is unnecessary. Efficiency comes from comparing the complexity of the event and its complement, not from always preferring one technique.

A good pre-calculation question is: “Which side has fewer disjoint cases?” A second is: “Can I describe those cases without overlap?” Those two questions provide a principled choice of method under time pressure and reduce dependence on spotting a familiar worksheet pattern.

19. Conditional reasoning: the denominator follows the information

This section extends the core discussion. Simple sequential probabilities already involve updating information, but formal conditional-probability notation and broader table analysis should not be assumed examinable merely because they appear here. Use your actual syllabus and teacher’s coverage. The purpose is to strengthen the idea that a probability is always relative to a specified population or information set.

In an invented survey of forty students, eighteen attend chess club, sixteen attend coding club and seven attend both. The overlap is counted within each club total. Therefore chess only has eleven students, coding only nine, both seven and neither thirteen. These four disjoint categories sum to forty.

Original two-way table for a club survey

	Coding	Not coding	Total
Chess	7	11	18
Not chess	9	13	22
Total	16	24	40

One table, three different denominators

If one of all forty students is selected uniformly, the probability that the student attends both clubs is 7/40. If the selected student is known to attend coding, the relevant population is the sixteen coding students. The probability that this student also attends chess is then 7/16. If the student is known to attend chess instead, the probability of coding is 7/18.

The numerator is seven in all three calculations, but the question changes the denominator. This is why reversing the conditional statement can change the answer. “Among coding students, what fraction attend chess?” is not the same question as “Among chess students, what fraction attend coding?” The table makes the difference visible without requiring advanced notation.

Do not count the overlap twice

The probability of attending at least one of the clubs is (18 + 16 − 7)/40 = 27/40. Alternatively, add the disjoint cells 11 + 7 + 9. The probability of neither is 13/40, providing a complement check. The probability of exactly one club is (11 + 9)/40 = 1/2, which differs from at least one because it excludes the seven students in both.

This example brings together event language, Venn-diagram overlap and conditional denominators. A wrong answer can therefore have several different causes. If the student reports 34/40 for at least one club, overlap was counted twice. If they report 7/40 when coding is already known, the information set was not updated. If they report 27/40 for exactly one club, the event was mistranslated.

Independence can be tested within a specified model

For a student selected uniformly from these forty, P(chess) = 18/40 and P(coding) = 16/40. Their product is 0.18, whereas P(both) = 7/40 = 0.175. These events are therefore not exactly independent in this finite empirical model. That observation does not establish a causal relationship between the clubs or a general claim about students elsewhere.

The distinction between a mathematical property of this table and an inference about a wider population matters. School probability can ask whether an equality holds within a given model. Real data analysis would need further reasoning before generalising beyond the sample. Keep the conclusion as narrow as the information permits.

A practical denominator question

Before any conditional calculation, ask: “Who or what is still eligible after the new information?” If you know the chosen object is blue, red objects leave the relevant set. If you know the first draw was red without replacement, one red object leaves the bag. In both cases information changes the population or state used for the next probability.

This is also a useful reading habit outside examinations. A statement about “the percentage of successful applicants who chose subject X” is not the same as “the percentage of subject-X students who succeeded”. Even when both statements involve the same overlap, their denominators answer different questions.

20. Unknown quantities: turn the probability model into an equation

Some questions give a probability and ask for an unknown number of objects. The first task is still modelling. Define the unknown with its meaning and restrictions, write the event probability, and only then solve the equation. The final answer must satisfy both the algebra and the physical conditions of the experiment.

A bag contains three red tokens and n blue tokens. Two are drawn at random without replacement. The probability that both are red is 1/12. The first red probability is 3/(n + 3). After a red token is removed, two red tokens remain among n + 2 total tokens. Therefore [3/(n + 3)] × [2/(n + 2)] = 1/12.

Multiplying gives (n + 3)(n + 2) = 72. Expanding and rearranging gives n² + 5n − 66 = 0, which factorises as (n + 11)(n − 6) = 0. The possible algebraic roots are −11 and 6. Since n counts blue tokens, reject −11 and take n = 6.

Substitute into the original experiment

With six blue tokens there are nine tokens initially. The two-red probability is (3/9)(2/8) = 6/72 = 1/12, matching the given information. This is a stronger check than substituting into the expanded quadratic alone, because it also tests whether the probability model and count restrictions were understood.

A common mistake is to reduce both colour counts after every draw. Along the RR path, the blue count never changes. Another is to keep the total denominator n + 3 for both stages, accidentally modelling replacement. The equation may still be solvable, but it will answer the wrong experiment.

Single-stage reverse probability

For a simpler version, suppose the same bag has three red tokens and n blue tokens, and P(red) = 1/4 on one draw. Then 3/(n + 3) = 1/4, so n + 3 = 12 and n = 9. No quadratic is needed because there is only one stage. Do not introduce a tree or advanced formula when one fraction already represents the experiment.

If a calculation gives n = 2.5, do not round to three tokens. A count must be integral. Check whether you confused probability of red with probability of blue, forgot existing tokens in the denominator, or misread a percentage. In some deliberately constructed problems, the conclusion may be that the information cannot describe such a bag.

An unknown success probability

Suppose two independent trials have the same success probability p, and the probability of at least one success is 0.91. The complement has probability 0.09, so (1 − p)² = 0.09. Because 0 ≤ 1 − p ≤ 1, take 1 − p = 0.3 and obtain p = 0.7. The negative square-root branch is rejected because a probability cannot be negative.

Check directly: P(no success) = 0.3 × 0.3 = 0.09, so P(at least one) = 0.91. The shared probability and independence assumptions are essential. If trials have different success probabilities, the same one-equation problem would generally not identify both values. Do not claim more information than the model supplies.

When equations are underdetermined

If a question says only that a bag contains red and blue tokens and P(red) = 1/3, the red-to-total ratio is known, but the total number of tokens is not. One red and two blue, two red and four blue, and many larger multiples fit. A probability ratio does not identify an absolute count unless additional information is provided.

Recognising insufficient information is a mathematical skill. It is not failure to find a clever technique. State what is determined, what remains free and which extra fact would fix it. This disciplined reading prevents students from inventing convenient sample sizes to force a numerical answer.

21. Calculator use: verify the input before trusting the output

A calculator is valuable when it reduces routine arithmetic without hiding the mathematical structure. It becomes dangerous when the student cannot explain what was entered, which statistic was selected or whether the result fits the data. The most useful calculator habit is not speed alone. It is being able to detect a plausible-looking wrong output.

For a frequency table, enter the data values and their corresponding frequencies in the appropriate fields. A table with five distinct values can represent forty observations. If your calculator reports N = 5 rather than N = 40, it has probably treated the five values as equally weighted observations or has not used the frequency column. Do not proceed to standard deviation until the count is correct.

A four-stage verification routine

First check N, the number of observations. Second check the mean against a hand-calculated weighted total for a manageable dataset. Third check whether the question needs population or sample standard deviation. Fourth check the units and rough size of the result. These stages inspect different possible failures rather than repeating one calculation several times.

For the book table, the correct count is twenty and the mean is 1.85. If a calculator gives a mean of 2 from the values 0, 1, 2, 3, 4, that is the unweighted mean of the category labels. The output is mathematically correct for a different dataset. This is an input-model error, not a calculator arithmetic error.

Grouped data use midpoints and frequencies

For the thirty-journey table, enter midpoints 5, 15, 25, 35 and 45 with frequencies 3, 7, 12, 6 and 2. The expected count is thirty and the estimated mean is 24. Do not enter upper boundaries 10, 20, 30, 40 and 50 as though they were the observed values. Doing so would shift the estimate upward by five minutes because every entered representative value would be five too large.

That particular mistake leaves standard deviation unchanged: all representative values were shifted by the same constant. Therefore a plausible standard deviation does not prove the mean was entered correctly. It is an instructive example of why checking several outputs is useful. Each statistic responds differently to input changes.

Keep enough precision inside a calculation

When calculating standard deviation from Σx²/N − mean², do not round the mean aggressively before squaring it. Subtracting two relatively close quantities can make early rounding consequential. Use stored precision or the original fraction where possible, then round the final result according to the question.

For the twenty-student example, mean 1.85 is exact from 37/20, so no approximation has yet occurred. The standard deviation is the square root of 1.0275 and requires approximation for a decimal answer. Distinguish these stages. Writing 1.01 books is a final rounded summary, not a value to reuse as though exact in an unrelated derivation.

Probability calculations need complete brackets

Write a probability expression on paper before entering it. For at least one success in three independent trials with success probability 0.8, the expression is 1 − (0.2)³. It is not (1 − 0.2)³, which calculates all three successes. Brackets encode the event structure. A calculator faithfully evaluating the wrong brackets cannot fix the interpretation.

For a sum of paths, keep each product visibly grouped: (4/7)(3/6) + (3/7)(4/6). Where exact fractions are simple, simplify or retain them until the final step. A recurring decimal is not automatically more informative than 4/7. Exact arithmetic also makes complement and total-probability checks easier.

Prepare with the actual permitted calculator

The official syllabus documents state the relevant calculator provisions, and SEAB’s current examination information and school instructions should control equipment decisions. Practise with the model you will actually use. Learn how to clear old statistical data, edit a frequency, distinguish standard-deviation outputs and return to ordinary calculation mode.

Do not memorise a button sequence from a different model without checking its meaning. A useful practice test is to enter four values whose mean and standard deviation you already know, deliberately change one frequency, then explain how the count and mean should move. This makes the calculator part of a verified process rather than an unexplained answer machine.

22. Examination control: protect interpretation under time pressure

A student may understand a topic in an untimed lesson yet make different errors in a paper. Time introduces choices: how long to spend reading, when to leave a difficult item, whether to check an answer and how to preserve enough working to return later. The remedy is not to rush every calculation. It is to make the high-value reading and checking actions short enough to use reliably.

Use a three-pass question routine. On the first pass, identify the variable, representation, total and event. On the second, perform the calculation with visible structure. On the third, check the answer type, bounds and meaning. The passes need not be long or formal; they are a sequence of decisions. Their purpose is to stop a familiar-looking diagram from triggering the wrong method automatically.

Read the final demand before doing every possible calculation

A table may support mean, median, mode, standard deviation and a probability, but the question may ask for only one. Calculating all available summaries wastes time and creates opportunities for unnecessary errors. Underline or briefly paraphrase the target: “estimate mean time”, “compare consistency using spread”, or “probability of exactly one blue”.

If a later part asks for an explanation, preserve useful intermediate results. A labelled mean and standard deviation can support a comparison without re-entering the data. A clear tree can support several related probability events. Good working serves both the examiner and your own navigation through the problem.

Show the structure that earns confidence in your method

For a weighted mean, show the weighted total and total frequency. For standard deviation, identify the statistic or formula and relevant inputs. For a cumulative graph, indicate the target cumulative position and graph reading. For a tree, label branches and show the complete path products being added. You do not need to narrate every calculator keypress.

The official 4052 examination syllabus notes the importance of essential working. Do not interpret calculator permission as permission to replace reasoning with an unexplained number. At the same time, avoid writing long prose where one labelled expression communicates the mathematics clearly.

Use local checks rather than redoing the whole question

A mean must lie within the data’s minimum and maximum. Cumulative frequencies must not decrease. Q1 ≤ median ≤ Q3. A range and standard deviation cannot be negative. A probability must lie between zero and one. A two-stage “both” event cannot be more likely than either necessary first-stage condition. These checks are cheap and often catch large errors.

Use a second representation selectively. If a tree is confusing but the sample space is small, enumerate outcomes in a table. If a grouped mean seems wrong, estimate which interval contains most observations. If a box-plot comparison sounds too strong, imagine one observation near each extreme. You are checking the reasoning at its weakest point, not repeating every line.

What to do when stuck

Write the last fact you trust. For cumulative frequency, that might be “N = 40, median position 20”. For probability, it might be “without replacement: total falls from 7 to 6”. For an unknown frequency, it might be “total frequency = 6 + k”. This preserves a valid starting point and may make the next step visible.

If no useful progress follows and the question is consuming disproportionate time, leave a brief re-entry note and move on. A note such as “need count above threshold, subtract CF from total” is more helpful than an abandoned page of arithmetic. When you return, you should not have to reconstruct the entire interpretation from scratch.

Practise timing on learned material

Do not use a stopwatch to teach a method the student has never understood. Begin with untimed accurate work, then shorter mixed sets, then sections under realistic constraints. Use the actual paper’s duration and instructions from the relevant syllabus and school. This guide does not impose one timing plan on G2, G3 and every other curriculum.

After a timed set, record both accuracy and time. Faster work with more interpretation mistakes is not necessarily improvement. Equally, perfect work that reaches only half the available questions indicates a performance issue worth investigating. Separate slow calculator entry, long method hesitation, repeated rereading and overchecking; each needs a different repair.

The final review should be personal

Use a short list based on your own recent errors. One student may need “frequency denominator, population SD, graph units”. Another may need “without replacement, exactly versus at least, overlap”. A twenty-item generic checklist is difficult to use in the final minutes. Three recurring risks that you actually check can be more useful.

Change an answer only when you find a specific reason. Anxiety alone is not evidence that the first answer was wrong. If your model is valid, the arithmetic checks and the interpretation fits, keep it. If a branch count or denominator is wrong, repair that precise point and carry the correction through the affected parts.

23. An independent diagnostic: eighteen questions with explanations

These are original teaching questions, not an official paper or a prediction of examination difficulty. Work without the solutions visible. Use your calculator where appropriate and show enough working to identify your method. The conditional question is labelled as an extension. Do not convert the score into a school grade; use the pattern of errors to choose what to practise next.

For each response, record one of three states: solved independently, solved after a small cue, or not yet solved. A correct answer after reading the explanation is useful learning, but it is not the same evidence as a correct first attempt. Retest with a different question after a delay.

Questions

D1. Frequency versus value. Values 0, 1, 2 and 3 have frequencies 3, 4, 2 and 1. Find the number of observations, mean, median and mode. State the unit of each if the values represent books completed.

D2. Combined mean. Six observations have mean 11 and another four have mean 16, on the same scale. Find the mean of all ten observations.

D3. Corrected mean. Nine observations were recorded with mean 18. One entry of 27 should have been 18. Find the corrected mean.

D4. Positions and spread. For 3, 4, 4, 7, 8, 9, 12, 17, find the median, range and IQR using the median-of-halves convention specified for this exercise.

D5. Grouped estimate. Intervals 0 < t ≤ 10, 10 < t ≤ 20 and 20 < t ≤ 30 have frequencies 4, 6 and 10. Estimate the mean and state why it is not necessarily exact.

D6. Cumulative counts. A fifty-observation cumulative table has counts 6, 18, 38 and 50 at upper boundaries 10, 20, 30 and 40. How many observations exceed 30? Estimate the median using straight-line interpolation between the supplied coordinates.

D7. Standard deviation. Values 0, 2 and 4 have frequencies 1, 2 and 1. Find the population standard deviation, showing the total frequency and mean.

D8. Box-plot comparison. Group A’s five-number summary is 8, 12, 15, 18, 28. Group B’s is 7, 10, 16, 22, 30. The variable is completion time in minutes. Compare centre and central spread. Can you conclude every person in A is faster?

D9. Unequal sectors. An ideal spinner has three sectors with angles 54°, 126° and 180°. Find the probability of landing in the 126° sector.

D10. Joint outcomes. Toss a fair coin and roll a fair die independently. Find the probability of heads and an even die result.

D11. Overlapping events. On one fair die, A is “even” and B is “prime”. Find P(A or B), and decide whether A and B are mutually exclusive.

D12. No replacement. A bag contains two green and three yellow tokens. Two are selected uniformly without replacement. Find the probability of one of each colour.

D13. With replacement. Repeat D12, but replace and remix the first token before the second draw. Find the probability of one of each colour.

D14. At least one. Three independent trials each have success probability 1/3. Find the probability of at least one success, then exactly one success.

D15. Conditional extension. In a thirty-person group, twelve swim, ten play music and four do both. Among those who play music, what proportion swim? What proportion of the whole group does neither?

D16. Unknown count. A bag has five red tokens and an unknown number of blue tokens. Each token is equally likely to be selected, and P(red) = 5/8. How many blue tokens are there?

D17. Compare proportions. Seven of twenty respondents in A and nine of thirty in B choose an option. Which group has the higher observed proportion? Does the comparison alone establish a cause?

D18. Estimated count. A model assigns an event probability 0.7 on each of forty trials. What is the expected count? Explain why that is not a promise of the exact observed count.

Open the diagnostic answers and error explanations

D1–D4: totals and positions

D1: N = 10 and the total of the values is 0 × 3 + 1 × 4 + 2 × 2 + 3 × 1 = 11. Mean = 1.1 books. The fifth and sixth observations are both 1, so median = 1 book. The mode is also 1 book, with frequency four students. Reporting 4 as the mode confuses the most frequent value with its frequency.

D2: Recover totals: 6 × 11 = 66 and 4 × 16 = 64. The combined mean is 130/10 = 13. An answer of 13.5 averages the two means equally despite unequal group sizes. The correct value is closer to 11 because the group with mean 11 contains more observations.

D3: The recorded total is 9 × 18 = 162. Correct it to 162 − 27 + 18 = 153. The number of observations remains nine, giving mean 17. Adding a tenth observation would model an omitted value, not the replacement of an incorrect entry.

D4: Median = (7 + 8)/2 = 7.5. Range = 17 − 3 = 14. The lower half is 3, 4, 4, 7, giving Q1 = 4. The upper half is 8, 9, 12, 17, giving Q3 = 10.5. Thus IQR = 6.5. The specified quartile convention matters; do not silently switch algorithms.

D5–D8: grouped information and dispersion

D5: Midpoints are 5, 15, 25. The estimated total is 4 × 5 + 6 × 15 + 10 × 25 = 360, giving estimated mean 18. The individual values within each interval are unknown, so midpoint substitution gives an estimate. The five-unit or ten-unit interval labels are not frequencies.

D6: Fifty minus thirty-eight gives twelve observations above 30. The median cumulative position is 25. Between boundaries 20 and 30, the cumulative count rises from 18 to 38; position 25 is 7/20 of that rise. The straight-line estimate is 20 + (7/20) × 10 = 23.5. State it as an estimate, not a recovered raw observation.

D7: N = 4, total = 8 and mean = 2. The squared-value total is 0 + 2 × 4 + 16 = 24. Variance = 24/4 − 2² = 2; population standard deviation = √2, approximately 1.41. Using N = 3 would mistake the number of table rows for the number of observations.

D8: A has median 15 minutes versus B’s 16, and IQR 18 − 12 = 6 versus 22 − 10 = 12. A therefore has lower central time and less central-half spread by these measures. You cannot conclude every A time is lower: A’s maximum is 28 while B’s minimum is 7. A comparison of summaries is not a complete ordering of individuals.

D9–D14: model and event structure

D9: The probability is 126/360 = 7/20 = 0.35 under the uniform-angle model. One third would incorrectly assume the three sectors are equal. The printed angles sum to 360, providing a useful completeness check.

D10: P(H and even) = (1/2)(3/6) = 1/4. Alternatively, three of twelve equally likely coin-die pairs qualify. Independence is supplied in the question, so multiplying the separate probabilities is justified.

D11: A = {2,4,6}, B = {2,3,5}; their union is {2,3,4,5,6}. Probability = 5/6. They are not mutually exclusive because 2 belongs to both. Adding 1/2 + 1/2 without subtracting the overlap would incorrectly give certainty.

D12: Add the disjoint GY and YG paths: (2/5)(3/4) + (3/5)(2/4) = 3/5. The second denominator is four because one token has been removed. Both colour orders satisfy “one of each”.

D13: With replacement, use (2/5)(3/5) + (3/5)(2/5) = 12/25. The original composition is restored before the second draw. The difference from D12 tests whether the process was read rather than whether the multiplication was quick.

D14: At least one success has probability 1 − (2/3)³ = 19/27. Exactly one success has three paths, each (1/3)(2/3)², giving 4/9. The complement of exactly one includes zero, two and three successes, so it is not merely no success.

D15–D18: information and interpretation

D15: Among the ten music participants, four swim, giving 4/10 = 2/5. The union has 12 + 10 − 4 = 18 people, so twelve do neither and the whole-group proportion is 12/30 = 2/5. These answers happen to be equal, but they have different denominators and meanings.

D16: The total must be eight because 5/(5 + b) = 5/8. Hence b = 3. Eight is the total, not the number of blue tokens. Substitute the answer into the original ratio to check.

D17: A has observed proportion 7/20 = 0.35, higher than B’s 9/30 = 0.30. B has more respondents selecting the option but a lower proportion. The counts alone do not identify a cause or establish a wider population conclusion.

D18: The expected count is 40 × 0.7 = 28 under the stated model. An individual set of forty trials can give another count. Expected value is a model-based average, not a guarantee that randomness will produce its average in every finite experiment.

Use the diagnostic to choose one repair

Errors in D1–D3 usually call for closer attention to totals, frequencies and denominators. D4–D8 test positions, grouped estimates and spread. D9–D14 test equally likely outcomes, replacement and event wording. D15–D18 test the information set and the limits of interpretation. These are learning categories, not diagnoses of a child’s intelligence.

Choose the earliest recurring failure. If a student uses the wrong total frequency in both a mean and standard-deviation question, repair frequency reading before drilling two separate formulas. If probability errors appear only when replacement changes, compare those experiments side by side. A targeted repair should produce a better independent attempt, not merely a longer completed worksheet.

24. Mixed teaching laboratory: connect the whole strand

The following investigations combine several representations and decisions. They are designed to prevent a student from succeeding only when the question announces a chapter. Each investigation begins with a small dataset or experiment, then changes the demand. Attempt the questions before reading the discussion, and explain the transitions between methods.

Investigation A: reading counts become a probability question

A fictional group of forty students reports the number of books completed during a fixed period. Counts x = 0, 1, 2, 3, 4, 5 have frequencies 2, 6, 12, 10, 8, 2. Find the mean, median, mode and population standard deviation. Then find the probability that a uniformly selected student reported at least four books. Finally, select two different students without replacement and find the probability that exactly one reported at least four.

Begin by verifying the frequencies sum to forty. The weighted contributions are 0, 6, 24, 30, 32 and 10, totalling 102. Thus mean = 102/40 = 2.55 books. Cumulative frequencies are 2, 8, 20, 30, 38, 40. The twentieth observation is 2 and the twenty-first is 3, so median = 2.5. The mode is 2 because it has frequency twelve.

This table deliberately makes the mean, median and mode different. A student who assumes a symmetric pattern or chooses the highest frequency as the mode will be exposed by the distinct answers. A median of 2.5 is valid even though each reported count is a whole number. It averages two central observed values rather than inventing a student with a fractional report.

For standard deviation, Σfx² = 0 + 6 + 48 + 90 + 128 + 50 = 322. The population variance is 322/40 − 2.55² = 1.5475. The standard deviation is approximately 1.24 books. Keep the unrounded variance during calculation and report the final precision appropriate to the task.

Ten students reported at least four books, so the single-selection probability is 10/40 = 1/4. For two selections, treat those ten as category H and the remaining thirty as category L. Exactly one H has paths HL and LH: (10/40)(30/39) + (30/40)(10/39) = 5/13. The underlying dataset has not changed; the question has moved from summarising observations to randomly selecting people from them.

The probability both are H is (10/40)(9/39) = 3/52. These answers must not be replaced by calculations using two independent 1/4 probabilities, because selecting distinct students without replacement changes the second-stage counts. A short tree gives the exact update.

Now ask what the survey allows you to conclude. It describes reports from these forty students over the stated period. It does not establish every student’s reading ability, the difficulty of the books, or a causal effect of any programme. The numerical summaries are valid within the dataset; broader educational claims require information not supplied here.

Investigation B: equal means, unequal reliability

Two fictional delivery processes produce times in minutes. A records 8, 9, 10, 10, 11, 12. B records 4, 8, 10, 10, 12, 16. Find mean, median, range and population standard deviation for each. Compare the processes without claiming every delivery in one process is faster.

Both totals are sixty and both sample sizes are six, so both means are ten. Their third and fourth values are ten, giving the same median. A has range four and B range twelve. Squared deviations from ten total ten for A and eighty for B. Thus their population standard deviations are √(10/6), approximately 1.29, and √(80/6), approximately 3.65.

Under the median-of-halves convention for this six-value exercise, A has Q1 = 9 and Q3 = 11, so IQR = 2. B has Q1 = 8 and Q3 = 12, so IQR = 4. Every selected spread measure indicates that A varies less in this sample, while centre is equal. That is a stronger explanation than “A is faster”, which the equal means and medians do not support.

Suppose a new loading procedure adds four minutes to every recorded time. Both means become fourteen, but ranges, IQRs and standard deviations are unchanged. Each observation and its mean move together, so deviations are preserved. This can be reasoned without re-entering all twelve numbers.

Suppose instead the times are converted to seconds. Means become 600 seconds; standard deviations become approximately 77.5 seconds for A and 219 seconds for B. The numerical size increases because the unit is smaller, not because the processes became less reliable. Comparisons of spread require consistent units.

For a decision about deadlines, the mean and standard deviation may still be incomplete. If the deadline were twelve minutes, all six A observations meet it, while five of six B observations do. That empirical deadline proportion answers a different question from the mean. Do not assume a familiar summary is automatically the statistic a practical decision needs.

Finally, these are six observations per process. The worked example supports a description of those observations. It does not prove that A will outperform B on every future day or that the loading procedure causes better reliability. Keeping the conclusion bounded is part of the mathematical task.

Investigation C: a spinner, a score table and a misleading shortcut

An ideal spinner has four equal sectors labelled 0, 1, 2 and 5. Spin twice independently and add the two labels. What is the probability the total is at least six? What is the probability exactly one spin gives five? Why is the probability of total ten not one eleventh?

The sixteen ordered sector pairs are equally likely. Totals of at least six arise from (1,5), (2,5), (5,1), (5,2) and (5,5). Hence the required probability is 5/16. The two pairs (0,5) and (5,0) contain a five but total only five, so they do not qualify.

Exactly one five has six pairs: the first spin is five and the other is 0, 1 or 2, or the reverse. Its probability is 6/16 = 3/8. At least one five instead includes (5,5), giving 7/16. Those event descriptions sound similar, but their sample-space markings differ.

The possible totals are 0, 1, 2, 3, 4, 5, 6, 7 and 10. Totals eight and nine cannot occur. Even the nine possible totals are not equally likely: total two has three contributing pairs, while total ten has one. The probability of total ten is therefore 1/16, not 1/11 and not 1/9.

A full total-frequency table gives counts 1, 2, 3, 2, 1, 2, 2, 2, 1 for those nine totals respectively. The frequencies sum to sixteen. This is a probability sample space represented as a statistical frequency table. The same weighted-total method can calculate the average total across all sixteen equally likely outcomes.

As an optional expected-value extension, that weighted average is four points. You can also see it by noting that one spin has average label (0 + 1 + 2 + 5)/4 = 2, so two spins have average total four. This does not say total four is the most likely result; it occurs in only one pair, (2,2). Mean and mode remain distinct even when the distribution comes from a chance experiment rather than a survey.

Investigation D: how much information does a summary lose?

Consider sets C = 0, 4, 5, 6, 10 and D = 0, 1, 5, 9, 10. Both have mean five, median five and range ten. Are their standard deviations equal? Can mean, median and range reconstruct the original observations?

For C, squared deviations from five are 25, 1, 0, 1 and 25, totalling 52. The population variance is 10.4 and standard deviation approximately 3.22. For D, the squares are 25, 16, 0, 16 and 25, totalling 82. Variance is 16.4 and standard deviation approximately 4.05. The shared centre and endpoints do not force the internal observations to have the same spread.

This answers the reconstruction question: those three summaries do not uniquely determine the dataset. Many lists can share selected summaries while differing elsewhere. A question asking for an exact missing raw value from insufficient summaries may have multiple solutions, and you should look for the additional information that would resolve it.

Now imagine grouping both sets into wide intervals. More information would disappear. A grouped midpoint estimate can be useful, but it cannot recover details that were never retained. This links grouped estimates, box-plot limitations and cautious conclusions into one idea: every representation preserves some information and discards other information.

Investigation E: explain a wrong solution rather than just replace it

A student sees a table of five values with frequencies totalling forty, enters each value once, obtains a mean, then uses that mean in a standard-deviation formula divided by forty. Identify the first invalid step and explain why later arithmetic cannot repair it.

The first invalid step is treating each distinct value as one observation when the frequencies say otherwise. The calculator’s mean describes a five-observation dataset. Using forty later does not retroactively make that mean a weighted mean. The student has combined quantities from two different datasets.

A complete repair is to restore the matching values and frequencies, verify N = 40, calculate the weighted total, and then calculate the appropriate mean and standard deviation from the same data. Merely changing the final denominator can produce another plausible number while preserving the fundamental mismatch.

In probability, the analogous error is to list four colour paths and assign each probability one quarter even though the branch probabilities differ. The paths are exhaustive, but not necessarily equally likely. The correction must change the path weights, not simply add more decimal places to the original answer.

Explaining the first invalid step is valuable because it identifies what to retest. The student does not need to repeat every statistics formula. They need a new table with different frequencies and a check that the same reading mistake no longer occurs. A correction becomes a learning intervention only when it changes a later independent attempt.

Investigation F: one boundary word changes the answer

A fictional twenty-person quiz records scores 0, 1, 2, 3 and 4 with frequencies 2, 3, 4, 5 and 6. Find the proportions scoring more than 3, at least 3, fewer than 3 and no more than 3. Then find the exact median and explain why a straight line drawn between cumulative points can give the wrong answer for this discrete dataset.

Begin with the cumulative frequencies: 2, 5, 9, 14 and 20. At score 3, the cumulative count is fourteen because it includes scores 0, 1, 2 and 3. The count strictly above 3 is therefore 20 − 14 = 6. The required proportion is 6/20 = 0.3.

At least 3 includes both scores 3 and 4, so its count is 5 + 6 = 11 and proportion 11/20 = 0.55. Subtracting the cumulative count at 3 would be wrong for this event because it would remove the five observations equal to 3. Instead, subtract the count below 3: 20 − 9 = 11.

Fewer than 3 means scores 0, 1 and 2, giving nine observations and proportion 0.45. No more than 3 means scores 0 through 3, giving fourteen and proportion 0.7. The four phrases refer to four different events. The arithmetic is easy; the challenge is preserving the equality or strict inequality expressed in the words.

Check the complement pairs. More than 3 and no more than 3 have probabilities 0.3 and 0.7. At least 3 and fewer than 3 have probabilities 0.55 and 0.45. Each complementary pair sums to one. By contrast, more than 3 and fewer than 3 do not cover the whole sample space because both exclude the observations equal to 3.

The exact raw-data median is 3. There are twenty observations, so inspect positions ten and eleven. Cumulative frequency is nine at score 2 and fourteen at score 3; both central observations therefore have score 3. Their average is 3. This conclusion uses the exact frequency information, not an estimate of values inside an interval.

Suppose someone plots the points (2,9) and (3,14), joins them with a straight line and reads cumulative position ten as score 2.2. That interpolation would spread the five observations across scores between 2 and 3. The table says all five have score exactly 3. For this discrete frequency table, the interpolated value does not recover the exact median.

This is why the grouped journey-time example and this quiz example require different reasoning. In the grouped table, individual values inside a continuous interval are unknown, so a graph can provide an estimate. In the quiz table, individual score frequencies are known exactly. Do not discard exact information by imposing an unnecessary continuous approximation.

A median of 3 also does not mean exactly half the students scored strictly below 3. Here nine, or 45%, did. Fourteen, or 70%, scored at most 3. The repeated score at the median creates the difference. Median is a positional summary; ties affect strict counts around its value. If a question requires the number above or below a threshold, use the actual frequencies instead of automatically writing half the sample.

The weighted total is 0 × 2 + 1 × 3 + 2 × 4 + 3 × 5 + 4 × 6 = 50, so the mean is 2.5. The mode is 4. Thus mean, median and mode are 2.5, 3 and 4. The fractional mean is perfectly valid even though each observed score is integral. It summarises the total; it does not assert that anyone scored 2.5.

For a final interpretation task, suppose a student says, “Most people scored above the mean, so the mean is wrong.” Eleven observations, the scores 3 and 4, are indeed above 2.5. But a mean is not defined as a value with exactly half the observations on either side. Its defining property here is total divided by count. The student’s criticism confuses the mean with a rough idea of a median, and even median statements need care with ties.

This investigation provides a compact final check of the whole data-reading sequence. Identify the observations, distinguish exact values from intervals, translate the boundary words, locate the correct positions, then interpret the answer. A calculator cannot make these decisions on the student’s behalf.

25. Revision plans that respond to evidence

A revision plan should answer a learning problem, not merely fill a calendar. The schedules below are suggested ways to organise this guide. They are not tested promises of a particular grade, and they should be adjusted around school assignments, the student’s actual syllabus and available energy. A student who already solves a topic independently does not need the same volume as someone learning it for the first time.

Use three categories for each skill: secure, unstable and not yet learned. Secure means you can solve a different-looking question without help after a delay. Unstable means success depends on recent practice, hints or a familiar layout. Not yet learned means the method or prerequisite is missing. A single correct answer while the example is visible is not enough to mark a skill secure.

A two-week repair route for an otherwise prepared student

First review: attempt a selection from the diagnostic before revising. Include one frequency question, one graph or position question and one probability model. Preserve the first attempt, including crossed-out work. Identify the first wrong step in each unsuccessful response. Choose one main mechanism rather than announcing that the entire strand is weak.

Early in week one: repair that mechanism with a small set. For a frequency error, use tables where the number of rows differs sharply from the total frequency. For a replacement error, compare two versions of the same bag experiment. For cumulative frequency, alternate questions that ask for a value at a position with questions that ask for a count above a value.

Later in week one: revisit the same idea with the topic label removed. Mix one repair question among two unrelated questions. The student must choose the method rather than merely apply the chapter they were told to practise. End with a written sentence explaining why the denominator, position or branch probability is correct.

Beginning of week two: use a delayed independent retest. Change the numbers, the wording and, where appropriate, the representation. If the student fails again, inspect whether the same mechanism returned. Do not simply increase the number of questions. A second failure may mean the first explanation did not address the actual confusion.

Later in week two: complete a short mixed set under a realistic time constraint, then review both errors and time use. Keep only the recurring risks on the personal exam checklist. A successful repair should survive a new question and a more realistic context; it does not require endless repetition of the original example.

A six-week foundation-to-performance route

Week one: reading and averages. Work on identifying observations, units, frequencies and totals. Connect raw lists, stems and frequency tables. Use reverse-mean questions to test whether total = count × mean is understood. A useful checkpoint is explaining why five table rows can represent forty observations.

Week two: spread and calculator control. Compare equal-mean datasets with different dispersions. Calculate one small standard deviation by hand, then verify it with the actual calculator. Add a frequency table and check N before reading the output. The checkpoint is selecting the intended statistic and interpreting its units.

Week three: grouped data and positions. Estimate means with midpoints, reconstruct class frequencies from cumulative counts and locate median and quartile positions. Compare box plots with explicit centre-and-spread language. The checkpoint is distinguishing a frequency position from a measured value and an estimate from an exact observation.

Week four: probability models. Build complete sample spaces and distinguish equally likely outcomes from grouped categories. Compare overlapping events, disjoint alternatives and independent stages. The checkpoint is explaining why a probability expression fits the stated experiment before calculating it.

Week five: sequential probability and mixed problems. Use replacement contrasts, complements and small three-stage experiments. Introduce unknown counts only after the branch model is secure. Revisit earlier statistics questions so they do not disappear while probability receives attention.

Week six: independent performance. Use mixed investigations and timed sections appropriate to the actual course. Analyse the first wrong step, not just the final score. Decide which skills need maintenance and which remain active repair targets. The final checkpoint is a clearer, shorter study plan, not a larger pile of resources.

When the examination is very close

Do not attempt to read the entire guide the night before a paper. Review the small number of errors that still recur, practise a few familiar independent questions, check calculator operation and use the syllabus instructions already established with school. Unlearned extensions should not displace core preparation at the last moment.

A short final review can follow three questions: “What do I tend to misread?”, “Which denominator do I tend to misuse?”, and “Which event words do I tend to confuse?” These are concrete enough to influence behaviour during the paper. “Be more careful” is too broad to tell you what to do.

Busy-week maintenance

On a week crowded by other subjects, use a small maintenance set rather than abandoning the strand or forcing a marathon. One frequency question, one cumulative-frequency reading and one two-stage probability can keep the main distinctions active. The purpose is continuity, not maximum volume.

When a skill is genuinely secure, reduce its practice share. Keep an occasional varied question and move attention to the next constraint. A revision system should become more efficient as problems are solved. Otherwise success paradoxically creates an ever-expanding workload instead of freeing time for other learning.

A compact error record

A reusable error-repair record

Field	Example entry
Question and first wrong step	Used five rows as N in a frequency table.
Reason	Confused distinct values with observations.
Replacement principle	N is the sum of frequencies.
Immediate check	Reconstruct a short raw list from the table.
Delayed retest	Different table, mixed among other topics.
Evidence for reducing practice	Correct independent reading across changed layouts.

Keep the record small enough to use. Copying a complete model solution into a notebook can take time without identifying the mistake. The useful entry is the principle that will change the next attempt and the task that will test it. Archive an entry when evidence shows the issue has stopped recurring; the record should not become a permanent catalogue of failure.

26. How parents and teachers can use the guide without taking over

Parents do not need to remember every statistics formula to support better revision. They can ask the student to identify the observation, denominator, unit and event. These questions reveal whether the problem has been understood while leaving the mathematical work with the learner. The aim is to make independent reasoning more visible, not to turn every homework session into an oral examination.

Agree on a small review point rather than checking every line. The student might bring one successful question, one uncertain question and one corrected error. Ask what changed between the first attempt and the correction. A response such as “I now use total frequency, not the number of rows” provides more useful information than “I did another ten pages”.

Three illustrative learners with different next steps

Imagine Alicia, Tricia and Kai Kai working on the same mixed set. These are fictional teaching situations, not accounts of real students or measured outcomes. Alicia remembers the mean formula but repeatedly counts table rows instead of observations. Tricia reads frequencies accurately but copies the first probability onto every second branch. Kai Kai can solve both topics untimed but spends so long rechecking simple arithmetic that later questions remain unfinished.

Alicia’s next task should make the meaning of frequency concrete. Ask her to expand a small frequency table into its full list, count the observations and compare that count with the number of rows. Then return to a larger table without expanding it. The evidence to look for is correct identification of N in a new layout, not faster recitation of the mean formula.

Tricia needs a state-change task. Place the words “with replacement” and “without replacement” above two versions of the same bag experiment. Ask her to write the bag’s contents after each possible first draw before calculating any path. The evidence to look for is a justified second-stage numerator and denominator in a changed example.

Kai Kai needs a performance review rather than automatic reteaching. Ask which checks actually find errors and which merely repeat trusted work. Use a short mixed set with a planned review period, then inspect where time went. The evidence to look for is maintained accuracy with more completed work, not simply a faster overall attempt.

These learners can make the same final answer wrong for different reasons. Giving all three an identical large worksheet may obscure the distinction. A shared lesson can still use a common topic while assigning different next questions. The quality of the next task matters more than the appearance of uniform activity.

Questions that preserve student ownership

Ask “What does that number count?” before saying the denominator is wrong. Ask “What is left after the first draw?” before supplying the second fraction. Ask “Which axis contains the answer’s unit?” before pointing to the cumulative-frequency reading. A prompt should help the student resume reasoning rather than perform the reasoning on their behalf.

If the student cannot restart after a small cue, provide an explanation and one modelled example. Then close the example and ask for a new independent attempt. There is nothing wrong with teaching; the problem is mistaking supported performance for independent mastery. The later retest tells you whether the explanation became usable knowledge.

What teachers can inspect in written work

Look for the first transition from valid to invalid reasoning. A standard-deviation solution may use the correct formula but the wrong N. A probability solution may multiply correctly but use an unchanged bag composition. A graph solution may find the right cumulative position but report that position rather than the corresponding value. Feedback should name that transition.

Where several students make the same mistake, compare two examples that differ in the critical feature. Equal and unequal frequencies isolate weighting. Replacement and no replacement isolate changing probabilities. “Exactly one” and “at least one” isolate event membership. Keeping other features stable makes the distinction easier to see.

What not to infer from one practice result

A correct answer on one familiar question does not establish complete topic mastery. A wrong answer on one demanding extension does not show that the student cannot manage the core syllabus. A lower score on a harder set is not directly comparable with a higher score on an easier set. Use several pieces of appropriately matched evidence.

Do not turn this guide into a comparison between siblings, schools or tuition groups. Its datasets are invented for teaching. They cannot support claims about any real institution or class. For a student’s actual progress, use their own school work, independent attempts and the demands of their actual course.

Keep the workload bounded

Longform guidance is a library, not a daily assignment. Choose the section that answers the current problem and stop when the task’s purpose is met. A student may need fifteen minutes of frequency reconstruction more than an evening of advanced probability. Another may need one difficult mixed investigation rather than repeated routine means.

For a broader Mathematics map, the eduKate Mathematics Learning System provides a route across topics. For this strand, keep the immediate goal narrower: accurate reading, justified models and independent checking. Wider navigation should help the student locate the next useful lesson, not multiply unrelated tasks.

27. Questions students ask, and a working glossary

Why do I get the wrong median when I know the formula?

The formula often gives a position, not the answer itself. After finding the position, locate the observation occupying it. In a frequency table, use cumulative counts; in a sorted raw list, count entries; on a cumulative graph, move from the frequency position to the measured-value axis. For an even-sized raw list, average the two central values, not merely their position numbers.

Why does my calculator disagree with my teacher’s standard deviation?

Check the dataset, frequency use and population-versus-sample output before assuming either answer is wrong. Verify N and the mean. For grouped data, confirm that midpoints were entered rather than boundaries. If those checks agree, compare rounding and the question’s required convention. A discrepancy is diagnostic evidence; identify its source instead of switching outputs until a desired number appears.

Can a mean be a decimal when every observation is a whole number?

Yes. Total divided by count can be non-integer. A mean of 2.55 completed books describes a group total divided by students; it does not claim any individual reported exactly 2.55 books. Similarly, an even-sized dataset’s median can lie between two observed values. Keep a summary statistic distinct from an individual observation.

Does a smaller standard deviation mean a better group?

It means less dispersion around the mean by that measure. Whether that is desirable depends on the context and the centre. Consistently slow completion times are not automatically better than more variable but faster times. When asked to compare groups, state the mean or median, the spread measure and what those differences mean for the variable being studied.

Why is a grouped mean only an estimate?

The frequencies tell you how many observations lie in each interval but not where they lie inside it. The midpoint substitutes a representative value for unknown observations. Different raw datasets can share the same class frequencies. The estimate is useful, but its calculation does not recover information lost during grouping.

Can a box plot tell me how many students scored exactly the median?

Not generally. The five-number summary does not disclose all repeated values or individual observations. You may know a median position or value without knowing the number of ties at that value. For exact counts, use raw data, a sufficiently detailed frequency table or additional information supplied by the question.

Why do I sometimes add probabilities and sometimes multiply?

Multiplication combines stages along a particular path using the correct next-stage probabilities. Addition combines disjoint successful alternatives. For overlapping events, remove double counting. Do not decide from the word “and” or “or” alone without identifying the outcomes. A tree, table or explicit list often makes the required operation obvious.

Are mutually exclusive events independent?

Not when both have positive probabilities. Mutually exclusive events cannot occur together, so knowing one occurred changes the probability of the other to zero. Independence means that information about one does not change the other’s probability. These are different properties and should be tested separately.

Why does “without replacement” change both the numerator and denominator?

The total falls because one object has been removed. A colour’s numerator falls only if an object of that colour was removed. After red is drawn from four red and three blue, there are three red and three blue among six total. After blue, there are four red and two blue among six. Write the remaining composition before writing the fraction.

Is “at least one” always solved by a complement?

It is often convenient because “none” has few paths, but the complement still needs the correct model. Compare the number of cases. For “exactly one”, the complement includes all other counts, so direct enumeration may be simpler. Choose the shorter justified route rather than treating a technique as mandatory.

Can two different questions have the same numerical answer?

Yes. In the diagnostic club example, the conditional swimming proportion among music participants and the whole-group proportion attending neither happen to both equal 2/5. Their denominators and meanings differ. Equal answers do not make the events or methods interchangeable. Show the model so the reasoning remains visible.

What should I do when there is not enough information?

State what is determined and what is not. A red probability of 1/3 identifies a ratio but not the total number of tokens. A box plot identifies selected positions but not every observation. Do not invent a sample size or raw values merely to produce a number. Explain what extra information would make the answer unique.

How many questions should I do before calling a topic secure?

There is no universal count. Look for independent success after a delay, with changed numbers and wording, and inside a mixed set. Repeating an identical question immediately can test memory of the solution rather than understanding. Reduce practice when the evidence is strong, but keep occasional maintenance so earlier learning stays accessible.

Should I read all the extensions for my examination?

Only after checking your actual course requirements and core readiness. Unequal-width histograms, formal conditional notation and expected-score discussions can deepen understanding, but their presence in an educational guide does not make them examinable. The official syllabus and school coverage should control exam priorities.

A glossary that connects terms to decisions

Observation: one recorded item in the dataset. Before counting observations, identify what one item represents: a student, a journey, a score or another unit.

Variable: the characteristic measured or recorded. Its type and units influence which statistics and graphs are meaningful.

Frequency: the number of observations with a value or in a class. Frequencies supply the denominator for a frequency-table mean.

Cumulative frequency: the running total up to a stated boundary. It connects value thresholds with observation positions.

Mean: total divided by count, using frequencies where needed. It preserves information about the total, not necessarily a typical observed value.

Median: a central value determined from ordered positions. Position and value are separate steps in finding it.

Mode: a most frequent value or category. It is not the frequency itself; a dataset may have more than one mode under the usual definition.

Range: maximum minus minimum. It measures the overall span while ignoring the internal arrangement.

Interquartile range: Q3 minus Q1. It describes the numerical span of the central half under the relevant quartile convention.

Standard deviation: a measure of dispersion around the mean formed from squared deviations. It returns to the original units after taking the square root.

Grouped estimate: a calculation using representative class values, commonly midpoints, because the original individual values are unavailable.

Sample space: the complete set of modelled outcomes. Counting fractions require attention to whether those outcomes are equally likely.

Event: the subset of outcomes satisfying the question’s condition. “Exactly”, “at least” and “not” change that subset.

Complement: all outcomes outside an event but inside the sample space. Event and complement probabilities sum to one.

Mutually exclusive: events sharing no outcome. Their probabilities can be added for the union without an overlap correction.

Independent: events for which knowing one does not alter the probability of the other. Independence justifies retaining the same next-stage probability where the model provides it.

Conditional probability: probability evaluated after specified information restricts the relevant population or state. The denominator must follow that information.

Expected count: a model-based average count over repeated versions of an experiment, not a guarantee for one observed set of trials.

28. Official sources and useful next reading

The cohort and examination-scope notes were checked against the following primary sources on 20 September 2026. Use the document for your actual examination year and subject. The worked datasets and practice questions in this guide were created for teaching; they are not copied examination questions or records of real students.

SEAB: 2026 O-Level school-candidate syllabus listings; Mathematics 4052 syllabus, 2026.

SEAB: 2027 SEC G3 school-candidate listings; Mathematics K310 syllabus, 2027.

SEAB: 2027 SEC G2 school-candidate listings; Mathematics K210 syllabus, 2027. Read the formatting and assessment scheme in the PDF, including the distinction between general and underlined content.

For broader background, OpenStax provides a textbook introduction to descriptive statistics and probability. These are wider textbook chapters, not substitutes for Singapore syllabus boundaries.

Within eduKateSG, use What is G2 Math for Secondary School? for the broader subject route, and The Ultimate Guide to Secondary Math Success in Bukit Timah for a wider secondary Mathematics study map. Return here when the current task is specifically data, distributions or probability.

The next useful action is small: choose one question you recently got wrong, identify whether the failure was reading, modelling, calculation or interpretation, and solve a different example that tests that exact point. Statistics and probability become more reliable when the student knows not only how to obtain a number, but what the number is allowed to say.