eduKate Singapore

VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Next useful routeCanonical hubHELP

Mathematics in Bukit Timah | Hypothesis Testing, P-Values, Errors and Statistical Power

—

by

eduKate Asia
Original conceptual editorial illustration comparing dispersed independent survey observations with observations clustered into twenty small groups.

Did you know that seeing 15 heads in 20 tosses is unusual for a fair coin, but not impossible? A mathematical hypothesis test can calculate how unusual it is under the fairness assumption. When the coin tosses are independent and each has probability one-half of heads, the exact two-sided probability of 15 or more heads or five or fewer heads is about 4.14%. That probability is called a p-value for the specified test—not the chance that the fairness claim is true or false.

This Mathematics in Bukit Timah guide uses fictional coins, numerical scores and invented experiments to explain null and alternative hypotheses, exact binomial tests, significance levels, one-sided versus two-sided tests, p-values, Type I and Type II errors, statistical power, confidence intervals, effect size, multiple comparisons and study design. None of the data are real examination outcomes, tutor claims or findings about Bukit Timah pupils. The purpose is statistical reasoning, not declaring that any real teaching practice works.

Quick answer: In a pre-specified two-sided test of H₀: p=0.5 against H₁: p≠0.5, independent 20-toss results with 15 heads give p-value=2Σ from k=15 through 20 of C(20,k)(0.5)²⁰≈0.0413895. At a preselected 5% significance threshold, this rejects the null model. It does not prove a coin is biased, give a 95.86% probability the alternative is true, or guarantee the result will recur in another experiment.

Article contents

Original conceptual graphic comparing clustered and independent sampling, accompanying fictional hypothesis testing and statistical power exercises.

Why hypothesis testing begins with a precise question

An experiment produces observations. A hypothesis specifies a mathematical model against which those observations are evaluated.

The null hypothesis, often H₀, describes the reference claim being tested. The alternative hypothesis, H₁, describes the departure the test is designed to detect.

Neither hypothesis becomes meaningful merely by naming it. We must define the variable, the sampling process, the probability model and the exact direction of departure.

Bukit Timah is context, not an invented research study

A learner near Sixth Avenue, King Albert Park or Beauty World may encounter statistics in school or in news articles.

But no pupils were measured for this article. The coins and index scores are hypothetical, with inputs deliberately chosen to illustrate methods.

A hypothesis-testing calculation is conditional on its assumptions. It cannot provide evidence about real school performance unless relevant real measurements and defensible study designs are supplied.

Begin with the simplest fair-coin model

Suppose the invented experiment tosses one coin 20 independent times.

Under the null hypothesis H₀, every toss has probability one-half of heads and one-half of tails.

Let X be the number of heads. Under H₀, X follows Binomial(n=20,p=0.5). This specifies the distribution of all possible head counts from zero through twenty.

Worked example 1: a central outcome

Under the fair-coin model, expected heads are n×p=20×0.5=10.

Observing ten heads is entirely compatible with fairness, although it is not the only possible result.

The standard deviation of the head count is √[np(1−p)]=√[20×0.5×0.5]=√5≈2.236. A head count several units away from ten can occur by random variation.

A hypothesis test needs a statistic

A test statistic is a function of the observed data used to assess compatibility with H₀.

In our coin example the statistic is simply X, the number of heads. Under fairness, its exact probabilities are supplied by the binomial distribution.

For other questions, a standardised sample mean, difference in means or regression coefficient may serve as the statistic.

The binomial probability formula

For X~Binomial(n,p), the probability of exactly k successes is

P(X=k)=C(n,k)pᵏ(1−p)ⁿ⁻ᵏ.

Here C(n,k) counts how many arrangements of k successes occur in n positions.

With p=0.5 and n=20, every particular 20-toss sequence has probability 1/2²⁰, while many sequences may share the same number of heads.

Worked example 2: the probability of exactly fifteen heads

For 20 fair independent tosses,

P(X=15)=C(20,15)/2²⁰.

C(20,15)=15,504, and 2²⁰=1,048,576. The probability is approximately 0.014786.

That is the chance of exactly fifteen heads. A two-sided p-value also includes other outcomes at least as extreme under the specified test.

The significance level is chosen in advance

A significance level α is the preselected upper bound on the test’s probability of rejecting H₀ when H₀ is true, under the assumed model.

Common examples are α=0.05 or α=0.01. They are conventions for error control, not universal boundaries between real and unreal effects.

For exact discrete tests, the actual false-rejection probability can be less than the nominated α because attainable tail probabilities come in discrete jumps.

Why the alternative decides which tail matters

If H₁ says p>0.5, unusually high head counts provide evidence against H₀ in the intended direction.

If H₁ says p<0.5, unusually low counts matter.

If H₁ says p≠0.5, both unusually high and unusually low counts matter. One-sided and two-sided p-values answer different questions.

Worked example 3: set up the two-sided claim

Let H₀:p=0.5 and H₁:p≠0.5.

We observe X=15 of 20 independent tosses. The null distribution is symmetric around ten.

Outcomes at least as far from ten as fifteen are X≥15 or X≤5. These two sets are mutually exclusive and symmetric under the fairness model.

Calculate the exact upper tail

For X~Binomial(20,0.5),

P(X≥15)= [C(20,15)+C(20,16)+C(20,17)+C(20,18)+C(20,19)+C(20,20)]/2²⁰.

The numerator is 15,504+4,845+1,140+190+20+1=21,700.

Therefore P(X≥15)=21,700/1,048,576≈0.0206947.

Worked example 4: combine two tails

Under the symmetric fair-coin null, P(X≤5)=P(X≥15).

The two-sided p-value is 2×21,700/1,048,576=43,400/1,048,576≈0.0413895.

That is about 4.14%. It is a probability of selected possible data outcomes under the null, not a probability assigned to the null being true.

Compare the p-value with the preselected level

Suppose α=0.05 was chosen before observing the coin results.

Since p≈0.04139<0.05, the test rejects H₀ at the 5% level.

If α=0.01 had been preselected, the same result would not reject H₀ because 0.04139>0.01. The data have not changed; the decision threshold has.

“Reject” is not the same as “prove false”

A false-rejection decision remains possible even when the null model is true, because rare outcomes do occur under true models.

Therefore rejection is a statement about evidence relative to the test procedure, not logical proof that the null cannot be true.

The result calls for clear wording: “The stipulated data are inconsistent enough with H₀ to reject at the chosen level,” not “the coin is definitely unfair.”

“Fail to reject” is not the same as “accept as true”

If a test does not cross its significance threshold, the data have not given sufficient evidence for rejection by that test.

This does not prove H₀ is correct. The test may have limited power, the sample may be small or an actual effect may be modest.

The careful conclusion is “fail to reject H₀”, not “H₀ has been proved.”

Worked example 5: an alternative head count

Suppose we again toss a fictional fair-or-unfair coin 20 independent times but see only 12 heads.

A two-sided exact fairness test would treat that as much closer to the null expectation ten than fifteen heads.

The p-value would be much larger than the p-value for fifteen heads. The same conclusion about fairness cannot be drawn from both outcomes without recalculating the evidence.

A p-value needs a complete test specification

A reader cannot determine a unique p-value from “a coin had fifteen heads” alone.

They must know the number of tosses, the null probability, the alternative’s direction, assumptions about dependence and the definition of “at least as extreme.”

In discrete settings, different valid two-sided test conventions may also differ for asymmetric null distributions. Here symmetry makes the two-tail doubling method straightforward.

Why p-values are probabilities about data

The null hypothesis gives a sampling distribution for possible test-statistic results.

The p-value is calculated by adding probabilities of results at least as incompatible with the null in the direction or directions specified by the test.

The hypotheses themselves are not random variables in a conventional frequentist p-value calculation. That is why the p-value is not P(H₀|data).

Worked example 6: diagnose an incorrect sentence

A pupil writes: “p=0.041 means the probability the coin is fair is 4.1%.”

That is not what the exact binomial test calculated.

A correct statement is: “If the fair independent-toss model were true, the chance of a head-count outcome at least as extreme as this under the specified two-sided test is approximately 4.1%.”

A Type I error is a false rejection

A Type I error occurs when H₀ is actually true but the test rejects it.

A properly calibrated 5% significance test controls the probability of that event at or below 5% under the specified null model and testing rule.

That does not mean five percent of all published positive results are false. Such a fraction depends additionally on prevalence of true effects, study selection and other factors.

A Type II error is a missed departure

A Type II error occurs when a specified alternative is true but the test fails to reject H₀.

Its probability is commonly denoted β for a particular alternative parameter value.

Unlike a simple fixed null model, there can be many possible alternatives with different effect sizes. The Type II error probability therefore depends on which alternative is being considered.

Statistical power is one minus beta

For a specified alternative, power=1−β is the probability the test correctly rejects H₀ when that alternative is true.

Higher power is desirable for detecting a meaningful departure, but it is not the probability a rejected finding is true.

Power depends on sample size, variation, significance level, effect size and the planned test.

Worked example 7: distinguish the two errors

Imagine a fictional quality-control test with H₀ meaning a machine meets a reference specification.

Type I error: the reference specification is actually met, but the procedure rejects H₀.

Type II error: the specified alternative is actually true, but the procedure fails to reject. Both are possible even with a mathematically valid test because the data vary randomly.

Why sample size changes sensitivity

With more independent observations under suitable conditions, an estimator’s standard error often becomes smaller.

That can allow a fixed effect to produce a larger standardised test statistic and more power.

However, increasing n does not automatically remove selection bias or model misspecification. A large sample can still give a precise answer to the wrong question.

A familiar z-test for a mean

Suppose observations are independent and normally distributed with known population SD σ=10.

Test H₀:μ=50 against H₁:μ≠50, using sample mean x̄ and sample size n.

Under H₀, the statistic z=(x̄−50)/(10/√n) follows the standard normal distribution. This supplies a separate continuous-model example from the discrete coin test.

Worked example 8: a mean from one hundred observations

Let n=100, x̄=52 and σ=10.

Standard error is 10/√100=1. Therefore z=(52−50)/1=2.

For a two-sided standard normal test, p≈0.0455. At preselected α=0.05, this rejects H₀ under the model assumptions.

The confidence interval gives a matching check

For the same known-σ setting, a two-sided 95% confidence interval for μ is

x̄±1.96σ/√n=52±1.96, giving [50.04,53.96].

The null value μ=50 lies just outside this interval. That agrees with the two-sided 5% z-test rejection.

The equivalence depends on matching the test, interval construction and assumptions. It is not an automatic identity between every possible test and interval method.

Worked example 9: the same mean with fewer observations

Keep x̄=52 and σ=10, but set n=25.

Now SE=10/5=2, so z=(52−50)/2=1.

The two-sided p-value is about 0.3173, much larger than 0.05. The data do not produce a 5% rejection under this different sample-size model.

Significance can change without changing effect size

Both mean examples have estimated difference x̄−μ₀=2 index units.

The difference in p-values arises because the standard error changes with the independent sample size.

Thus a p-value alone does not reveal the magnitude or practical importance of the observed effect. Report the estimate, its units and uncertainty.

Effect size answers a different question

An effect size expresses how large a difference or relationship is on a specified scale.

For a mean test, the raw effect can be x̄−μ₀. A standardised effect may divide by an appropriate SD, with the denominator and interpretation clearly stated.

Statistical significance depends on both effect and uncertainty. An enormous sample can make a tiny effect statistically detectable even when it has little practical relevance.

Worked example 10: two-unit mean difference

With x̄=52 and μ₀=50, the raw estimated mean difference is two fictional index units.

If the known population SD is ten, the observed standardised difference is 2/10=0.2 SD units.

A z-test’s z-statistic for n=100 is two because it divides by the SE one, not by the individual SD ten. These numbers answer different questions.

A one-sided test has a smaller same-level critical value

Under a standard normal test with α=0.05 and an upper-tailed alternative, the critical threshold is about z=1.645.

For a two-sided α=0.05 test, the critical magnitudes are approximately ±1.96.

One-sided testing concentrates the allowed false-positive probability in a specified direction. It must not be selected only after observing which way the result moved.

Worked example 11: compare one-sided and two-sided decisions

Suppose the observed z-statistic is 1.8 under a standard normal null.

The upper-tail p-value is about 0.0359. The two-sided p-value is about 0.0719.

At α=0.05, a pre-specified upper-tailed test rejects; a two-sided test does not. The difference reflects the planned alternative, not permission to choose whichever result is more attractive.

Changing the hypothesis after seeing data is a problem

If a researcher originally planned to test either direction but later reports a one-sided test only because the observed statistic is positive, the stated error control may not apply as intended.

Good practice defines the primary hypothesis, outcome, analysis method and decision level before looking at results when possible.

The lesson is mathematical as well as procedural: a probability calculation needs the correct sample-selection and decision context.

Power can be calculated in a simple normal example

Consider an upper-tailed 5% z-test of H₀:μ=50 with known σ=10 and n=100 independent normal observations.

The rejection threshold for x̄ is 50+1.645×1=51.645.

Suppose the true alternative mean is μ=53. Then x̄~N(53,1²). Power is P(x̄>51.645 | μ=53), which is approximately 0.912.

Worked example 12: translate the threshold under the alternative

Under μ=53 and SE=1, standardise the rejection threshold:

z=(51.645−53)/1=−1.355.

Therefore power is P(Z>−1.355)=Φ(1.355)≈91.2%.

Type II error probability at this particular alternative is about 8.8%. A different true mean would give different power.

A test’s power is not one fixed number for all alternatives

If the true mean were 50.1 rather than 53, the upper-tailed test would be much less likely to reject.

If the true mean were 60, rejection would be extremely likely under the same assumptions.

A complete power statement specifies the alternative effect, variability, sample size and threshold. “The test has 90% power” without a target effect is incomplete.

Type I and Type II errors trade off

Holding sample size and test statistic fixed, making the rejection region more permissive can increase power but also increase the Type I error probability.

Making the threshold more stringent, such as α=0.01 instead of 0.05, reduces false-rejection probability under H₀ but often reduces power for a given alternative.

Better study design or larger independent samples can improve that trade-off. No arbitrary threshold abolishes both error types.

A p-value is not the chance a result will replicate

A significant result may fail to replicate because the original estimate was noisy, the effect was smaller than assumed, the study was biased or conditions differed.

A p-value computed from one dataset is not a direct forecast of the probability the next experiment will have the same sign or cross the same threshold.

Replication questions require additional modelling and evidence.

Sampling assumptions matter

For an exact binomial coin test, the trials were assumed independent with a shared heads probability under H₀.

If the observations are correlated or the probability changes between trials, the simple binomial null distribution may be invalid.

Using a perfectly evaluated formula with the wrong sampling distribution can produce misleading p-values.

Worked example 13: copied records are not new evidence

Imagine one observation copied into a spreadsheet one hundred times.

The file contains one hundred rows but only one independent outcome. Treating the copies as one hundred independent samples falsely narrows the standard error.

The issue arises in the sampling model, not in the arithmetic of a z-test. Diagnose how the rows were generated.

A confidence interval reports more than a threshold

The interval [50.04,53.96] states an estimated range for the unknown mean under the specified 95% procedure.

It shows how close the tested null value is to the boundary and how uncertain the effect remains.

A simple “significant” label loses that information. Statistical interpretation should preserve the estimate and interval rather than only a yes-or-no decision.

Multiple testing increases the chance of false findings

Suppose twenty truly null tests are independent and each rejects falsely with probability exactly 0.05.

The chance that none falsely rejects is 0.95²⁰≈0.3585.

Thus the chance at least one falsely rejects is 1−0.95²⁰≈0.6415, about 64.15%. The independence assumption is used for this particular calculation.

Worked example 14: one test versus twenty

For one true null test calibrated at five percent, false rejection occurs with probability 0.05.

For twenty independent true null tests, the chance of at least one false rejection is about 0.6415.

This does not mean every batch of twenty tests produces a false result. It shows how a family of tests can create a much higher chance of at least one error.

Familywise error control is a separate objective

If a project tests many hypotheses and wants to limit the probability of any false rejection across the family, its analysis plan may use an adjustment.

The simple Bonferroni method tests each of m hypotheses at level α/m to control familywise error at most α under broad dependence conditions.

That can be conservative, but its logic is transparent: the union bound limits the chance that any individual false-rejection event occurs.

Worked example 15: Bonferroni for ten tests

Suppose the desired familywise significance level is 0.05 and ten tests are planned.

Bonferroni sets the per-test threshold to 0.05/10=0.005.

An individual p-value 0.004 crosses this adjusted threshold; p=0.02 does not. The adjustment changes the decision rule, not the original p-value calculation.

False discovery rate is another family-level concept

Some procedures control the expected proportion of false discoveries among findings declared significant, in a specified statistical sense.

This is different from controlling the probability of any false discovery.

The appropriate method depends on whether the project prioritises stringent familywise control, discovery across many signals or another objective.

P-hacking changes the probability story

Repeatedly trying different outcomes, samples, cutoffs or analyses until one produces a small p-value can undermine the nominal interpretation of an isolated reported test.

The final p-value may be mathematically correct for the final test conditional on a fixed analysis, while the overall selection procedure is not represented in that calculation.

This is why transparent protocols and comprehensive reporting matter.

An example of selective reporting

Imagine examining twenty unrelated fictional measures but publishing only the one with p<0.05.

Even if every null is true, the chance of such a selected result can be substantial, as the twenty-test calculation shows.

The problem is not that one arithmetic p-value was incorrectly evaluated. It is that selecting the most favourable result changes how the reported evidence should be interpreted.

Bayesian posterior probability is a different quantity

Bayesian inference can calculate probabilities assigned to hypotheses under a specified prior and likelihood model.

A frequentist p-value is instead a tail probability of data under H₀.

The two answer different questions, and neither can be obtained by simply subtracting the other from one. The Bukit Timah Bayes’ theorem guide explains the prior–likelihood distinction.

A small p-value does not measure practical benefit

Imagine a fictional policy produces an estimated mean improvement of 0.01 index units with a tiny p-value because the sample is immense.

The observed effect might be statistically detectable yet smaller than a meaningful real-world threshold.

The decision requires the effect size, uncertainty, costs and context—not merely the probability of data under a null reference model.

A large p-value does not prove equality

Suppose two imaginary methods have estimated difference 2 units but a very wide interval from −8 to 12.

The data are compatible with a range of negative and positive differences, including zero.

Failing to reject a zero-difference null does not prove the methods are equivalent. An equivalence question requires a preselected meaningful equivalence margin and an appropriate procedure.

Equivalence testing is a separate design

To demonstrate equivalence, one may test whether the difference lies within a chosen acceptable range such as [−δ,+δ].

This is not accomplished simply by obtaining p>0.05 in an ordinary difference test.

A student should understand the distinction between insufficient evidence of difference and evidence that differences are sufficiently small.

Test statistics and p-values require calibrated null distributions

A test statistic value of two may be unusual under a standard normal null but ordinary under another distribution.

The p-value comes from the statistic’s null distribution, including its degrees of freedom, variance and other assumptions.

Comparing every number two with 1.96 without knowing the test model is invalid.

Worked example 16: known versus estimated variation

With known σ=10 and n=100, our mean-test z-statistic was two.

If σ were unknown and estimated by the sample SD, a suitable normal-population mean test would generally use a t statistic and Student’s t null distribution with n−1 degrees of freedom.

The p-value can differ. Choosing z versus t is not a stylistic preference; it follows from the assumptions and variance information.

An exact binomial test is discrete

Our 20-toss fairness test has only 21 possible head counts.

Its p-values do not vary continuously through every number between zero and one. They reflect sums of the probabilities of count outcomes.

A normal approximation may be useful under suitable conditions, but near a fixed decision threshold, an exact discrete test can give a different result from an approximation.

Worked example 17: a boundary case with sixty heads

Suppose X=60 heads in 100 independent tosses under H₀:p=0.5.

The exact two-sided binomial p-value is approximately 0.05689, so an exact 5% test does not reject.

An uncorrected normal approximation z=(60−50)/5=2 gives a two-sided p≈0.0455, which crosses 5%. The disagreement illustrates why the chosen test and approximation matter near a threshold.

The continuity correction explains part of the difference

The binomial count is discrete, while the normal approximation is continuous.

When approximating P(X≥60) under n=100,p=0.5, a continuity correction uses threshold 59.5 rather than 60, giving z=(59.5−50)/5=1.9.

Its two-sided approximate p≈0.0574, close to the exact 0.05689. The correction is an approximation technique, not a substitute for exact counting when the exact distribution is available.

A result should state how the p-value was obtained

“p=0.046” and “p=0.057” may arise from different approximations for a discrete test near the threshold.

A responsible report names the exact binomial test or the approximation, including any continuity correction.

The message is not that one number should be chosen for a desired conclusion, but that the procedure must be defensible and reported consistently.

Hypothesis tests do not fix poor measurement

If a measuring process systematically records values three units too high, a large sample can make the biased mean estimate very precise.

A hypothesis test may then confidently reject a reference value for the wrong reason.

Statistical formulas quantify variation under a model. Correct data definitions and measurements are essential before interpreting results.

Observational comparisons need causal caution

A small p-value for an association between two observed variables does not prove changing one causes the other to change.

Confounding, reverse direction and selection bias can produce strong patterns without the proposed causal relationship.

That caution is especially important in educational reporting. The fictional numerical exercises here are not evidence that tutoring, study time or any named method causes an outcome.

The original null must be stated, not hidden

A reported test result without the null value and alternative direction is incomplete.

For the mean example, H₀:μ=50, H₁:μ≠50, known σ=10 and n=100 were essential pieces.

Without them, the p-value has no identifiable reference. Readers cannot know what was actually tested.

Pre-registration is a useful research practice

A study plan can state primary outcomes, sampling rules, hypotheses, thresholds and analysis choices in advance.

That helps readers distinguish preplanned tests from exploratory patterns discovered after looking at the data.

Exploration can be valuable, but it should be labelled. Reusing the same data to both discover and confirm a hypothesis without appropriate adjustment can overstate evidence.

Replication provides new information

An independent repeat of a study can test whether a finding is robust under a new sample or context.

The repeat need not reproduce exactly the same p-value, because random data vary.

A stronger interpretation combines effect estimates, uncertainty, study quality and consistency across relevant results rather than treating significance as a permanent badge.

Common error 1: interpret p as probability the null is true

The coin p-value 0.04139 is a probability of selected head-count outcomes assuming the null is true.

It is not the chance the null is true given the result.

The repair is to read every conditional probability in full and distinguish P(Data|H₀) from P(H₀|Data).

Common error 2: ignore the alternative direction

A result of fifteen heads is evidence against p=0.5 in an upper-tailed or two-sided test, not in a lower-tailed test favouring fewer heads.

The mathematical tail must match the hypothesis defined before the data were inspected.

A student’s first correction should be to write H₀ and H₁ explicitly before evaluating a test statistic.

Common error 3: accept H₀ whenever p is large

A large p-value means the test has not provided evidence sufficient to reject H₀ at the chosen threshold.

It does not establish the null’s truth, especially if the sample was too small to detect meaningful alternatives.

The correct conclusion is “fail to reject H₀,” along with effect estimates and uncertainty when available.

Common error 4: confuse Type I error with false discovery percentage

The preselected α level controls false-rejection probability under a true null, in a properly calibrated test.

It does not directly tell us the proportion of significant published findings that are false.

That latter proportion depends on how many nulls are true, power, selection and study design.

Common error 5: confuse power with confidence

Power describes how often the test rejects under a specified alternative.

A confidence level describes long-run parameter coverage of an interval procedure.

They are related through test design but are not numerically interchangeable. One cannot infer 90% power merely from a 90% confidence interval.

Common error 6: ignore multiple testing

Reporting the smallest p-value after checking twenty hypotheses does not preserve the same overall false-positive interpretation as one pre-specified test.

The chance of at least one false rejection increases with the number of opportunities.

State the size of the analysis family and use appropriate multiplicity methods when the research question requires them.

Common error 7: treat statistical significance as educational impact

A very small difference in a large sample may be statistically significant while being unimportant in practice.

Conversely a wide interval may fail to reject despite allowing meaningful positive or negative effects.

A statistical test does not replace domain knowledge, well-defined outcome measures or honest interpretation.

Common error 8: claim the test repairs sampling bias

A correct binomial or z formula cannot compensate for correlated, selected or mismeasured observations unless those issues are properly modelled.

The simplest place to diagnose failure is the data-collection process itself.

A smaller p-value from more biased observations does not make the underlying claim more trustworthy.

Primary preparation: chance and unusual outcomes

Younger pupils can toss imaginary coins, count outcomes and notice that some sequences are less common than others.

They can compare expected and observed counts without formal hypothesis labels.

The early transferable skill is recognising that an unusual event can still happen by chance, even under a true probability model.

Lower Secondary: binomial intuition and fair comparisons

Students can calculate simple independent-event probabilities and distinguish an observation from a model expectation.

They can use a small table to count possible head totals, then understand why both tails matter for a two-sided fairness question.

Technical p-values and formal power calculations belong to later learning when probability distributions and sampling ideas are secure.

Advanced statistics and syllabus boundaries

Hypothesis testing, exact binomial inference, power and multiplicity form part of advanced statistics learning.

Use the SEAB 2027 G3 Mathematics syllabus directory for school prerequisites and the 2027 A-Level syllabus directory for higher-level pathways.

This article does not assert that power analysis or multiple-testing adjustments are mandatory in every SEC G3 Mathematics course.

A twenty-minute parent diagnostic

Minutes 1–5: State H₀:p=0.5 and two-sided H₁:p≠0.5 for a fictional 20-toss independent experiment.

Minutes 6–10: Define the extreme outcomes X≥15 or X≤5, and calculate the exact two-sided p-value about 0.04139.

Minutes 11–15: Compare the value with preselected α=0.05 and α=0.01, giving the correct decisions without claiming proof of a hypothesis.

Minutes 16–20: Explain Type I versus Type II errors and why a small p-value is not the probability the null is false.

A seven-day review route

Day one defines null and alternative hypotheses. Day two reviews binomial counts.

Day three calculates tail probabilities. Day four compares one-sided and two-sided tests.

Day five studies Type I, Type II and power. Day six uses a mean z-test and matching confidence interval. Day seven diagnoses a multiple-testing or nonindependent-sample example, explaining which interpretation has become invalid.

Practice questions and checked answers

A. Expected heads in 20 fair independent tosses? Answer: 10.

B. Exact two-sided p-value for 15 heads in 20 with p₀=0.5? Answer: about 0.04139.

C. At preselected α=0.05, reject the null? Answer: yes for that specified two-sided exact test.

D. At preselected α=0.01, reject? Answer: no.

E. A true null rejected by the test is what error? Answer: Type I.

F. A specified false null not rejected is what error? Answer: Type II.

G. Define power. Answer: 1−β for a specified alternative.

H. Known σ=10,n=100,x̄=52,H₀:μ=50: z? Answer: 2.

Frequently asked questions

What is a null hypothesis?

The reference statistical claim against which a specified test evaluates sample evidence.

What is a p-value?

The probability, under the null model, of a test-statistic outcome at least as extreme as the observed one under the specified test.

Does p=0.04 mean a four-percent chance the null is true?

No. That reverses the conditioning and is not a frequentist p-value interpretation.

Does failing to reject prove the null is correct?

No. The test may have limited power or the sample may be uninformative.

What is the difference between one-sided and two-sided testing?

A one-sided test targets departures in a specified direction; a two-sided test targets departures in either direction.

What is a Type I error?

Rejecting a true null hypothesis under the defined test.

What is a Type II error?

Failing to reject a false null under a specified alternative.

Why does power depend on sample size?

Under appropriate independent sampling, larger samples often reduce standard error, making a fixed alternative easier to detect.

Can a small p-value prove cause and effect?

No. Causal conclusions need appropriate design and evidence beyond statistical association.

Are these real Bukit Timah educational results?

No. The coin tosses, scores, effect sizes and test decisions are all original fictional mathematical exercises.

Official context and connected Mathematics in Bukit Timah

Consult the SEAB 2027 G3 Mathematics syllabus directory and 2027 A-Level syllabus directory for the relevant Singapore course pathways. Our data and p-values are teaching constructions, not official examination statistics.

Read Confidence Intervals, Standard Errors and Coverage, Normal Distribution and Z-Scores, Bayes’ Theorem and False-Positive Base Rates and Statistical Correlation and Causation for connected but different questions.

The immutable Clementi Secondary 1 Mathematics small-group tutorial reference anchors concept-first explanations, verified assumptions, careful error diagnosis and independent practice.

The lasting habit is to define hypotheses before testing, choose the correct null distribution and tail, interpret p-values conditionally, report effect size and uncertainty, and never use significance as a substitute for evidence about real learners.

Companion Mathematics in Bukit Timah articles from this four-part batch: Hamming Distance, Parity Checks and Error-Correcting Codes · Simpson’s Paradox, Weighted Averages and Hidden Groups · Fourier Series, Harmonics and Signal Reconstruction. Together these original guides explore reliable digital coding, statistical evidence, subgroup-weighted interpretation and Fourier signal representation, with worked checks and clearly fictional examples.