NCEA Level 3 Statistics common mistakes
NCEA Level 3 Statistics · Understand the traps and how to avoid them.
When asked to interpret a confidence interval, the candidate writes 'there is a 95% probability that the true mean lies between 12.3 and 15.7' instead of stating that the interval was constructed by a method that captures the true parameter in 95% of repeated samples.
Everyday probability language is applied to the parameter, which is treated as a random variable. The candidate has memorised the phrase '95% confident' but never had the sampling-process idea made explicit, so the parameter feels like the thing that varies.
What to do instead
Before writing, sketch the sampling process: many possible samples, each giving its own interval, 95% of which cover the fixed unknown parameter. Then write the sentence as 'If samples were repeatedly taken and intervals built this way, about 95% of them would contain the true mean.' Practise rewriting three flawed interpretations per exercise and mark exactly which word (probability/confident/chance) is doing the damage.
The candidate states a conclusion about the population after a hypothesis test without linking the p-value to the null model, e.g. 'p = 0.03, so the null hypothesis is true with probability 0.03' or 'p = 0.03, so the drug works 97% of the time'.
The p-value is taught as a decision rule ('p < 0.05, reject') and the conditional structure is never unpacked, so students attach the number to the wrong event. The asymmetry of 'evidence against' versus 'proof of' is also glossed over.
What to do instead
Write the definition as a conditional probability in symbols: p = P(data this extreme or more | H0 true). Then rehearse the standard conclusion template: 'Assuming the null model is true, data this extreme would occur about 3% of the time; this is (strong/moderate/weak) evidence against H0.' Ban the words 'probability that H0 is true' in all written answers and check each practice answer for them.
The candidate checks assumptions only by naming them ('normality', 'independence') and never matches the check to the test used, e.g. quoting a normality condition for a chi-square goodness-of-fit test, or checking a sample mean's distribution rather than the residuals in a regression.
Assumption lists are memorised per topic in isolation, so the candidate cannot map a condition onto the model actually being fitted. The difference between the distribution of the data and the distribution of the estimate or residuals is rarely made explicit.
What to do instead
Build a one-page table with columns: test/model, parameter estimated, what must be independent, what must be approximately normal (data, sample mean, or residuals), and the diagnostic plot or rule of thumb used. For each condition write the consequence of violation (e.g. invalid standard error, biased p-value). Then answer exam-style assumption questions by naming the condition, the check, and the consequence.
The candidate reads a scatterplot or reports a regression slope as if it were the strength of a relationship, or concludes causation from a strong correlation, e.g. 'r = 0.89, so increasing study time causes higher marks'.
Correlation and slope both produce a single impressive number that is easy to over-interpret. Confounding and the direction of the relationship are abstract, and the candidate has no habit of asking what else varies with both variables.
What to do instead
For every scatterplot answer, force three separate sentences: direction, form (linear or not), strength with r or r squared, plus one named possible confounder and one reason causation cannot be concluded from observational data. Distinguish r (standardised strength) from the slope (units per unit) by writing both with units in practice questions.
When combining random variables, the candidate adds standard deviations instead of variances, e.g. for the difference in two independent sample means writes sd = 4 + 3 = 7 rather than the square root of 4 squared plus 3 squared.
The arithmetic of means (which does add and subtract) is transferred to spreads. The variance-addition rule is a formula to memorise rather than a consequence of independence, so its sign and squaring get lost under time pressure.
What to do instead
Derive the rule once from a small simulation or a two-dice example so the squaring is visible. Then drill mixed questions where some variables are independent (variances add) and some are linear transformations (multiply sd by the constant). Always write the variance step before taking the square root, and check that a combined sd is smaller than the sum but larger than each part.
The candidate answers a question about the sampling distribution of the sample mean by describing the distribution of the raw data, e.g. 'the sample means will be skewed because the population is skewed' or confuses the standard error with the population standard deviation.
The three distributions (population, sample, sampling distribution) are visually similar and the notation for sigma versus sigma over root n is easily muddled. The central limit theorem is recited as a slogan without the idea of repeated sampling.
What to do instead
Draw the three-distribution diagram side by side with centre, spread and shape labelled for each, and annotate how each changes as n grows. State the standard error formula with a units check and compute it for n = 4, 25, 100 to see the shrinkage. Then answer 'describe the sampling distribution' questions by naming centre, standard error and shape with the justification (CLT applies because n is large enough).
The candidate reports a probability or a p-value to excessive precision or misuses significance level, e.g. 'p = 0.0437821, therefore highly significant', or treats a p-value just above 0.05 as proof of no effect.
Software output is copied verbatim without judgement about rounding, and the binary 'significant/not significant' framing is learned as a fact about the world rather than a decision threshold. The candidate has not practised interpreting borderline results.
What to do instead
Adopt a reporting convention: p to three decimal places, p < 0.001 for tiny values, and always pair the p-value with the estimate and its confidence interval. For borderline cases, write a sentence that distinguishes 'insufficient evidence of an effect' from 'evidence of no effect' and note the sample size. Practise with two datasets giving p = 0.048 and p = 0.052 to see how little separates them.
The candidate designs or critiques an experiment without identifying the treatment, the response, random allocation, and blocking or pairing, e.g. proposing to compare two teaching methods by letting students choose their class, then concluding the method caused the difference.
Design vocabulary is learned as a checklist rather than as a causal argument. The link between random allocation and the validity of a causal claim is not made, so self-selection and confounding slip through unnoticed.
What to do instead
For every design question, write four labelled lines: units, treatments, response variable, and how allocation and blinding are handled. Then state explicitly what the randomisation buys (groups comparable on average, including unmeasured confounders). For observational studies, name at least two confounders and explain why adjustment cannot fully remove them.
The candidate performs the wrong test for the data structure, e.g. uses a two-sample t-test on paired data, a chi-square test of independence on one categorical variable, or ignores that counts are paired by using an independent-samples method.
Test selection is driven by keywords ('compare two groups') rather than by the number and type of variables and the dependence structure. Paired designs look like two groups in a spreadsheet, so the pairing is invisible unless the candidate asks how the data were collected.
What to do instead
Before any calculation, complete a selection grid: how many variables, categorical or quantitative, independent or paired samples, expected counts or sample sizes. Use it to name the test and the degrees of freedom. Then compute the paired and unpaired analyses on the same dataset once to see how much the pairing changes the standard error and the conclusion.
The candidate answers a contextual question with correct mathematics but no interpretation in context, e.g. reporting 't = 2.31, p = 0.024, reject H0' and stopping, or giving a confidence interval with no sentence about the actual variable and units.
Marking schedules reward the final communication step, but students practise calculations in isolation and never rehearse the closing sentence. Under time pressure the interpretation feels optional compared with the arithmetic.
What to do instead
Use a fixed answer skeleton for every inference question: state the parameter in context with units, give the estimate and interval, state the decision, then write one plain-language conclusion naming the variable and population. Time yourself writing only the interpretation sentences for past questions until they take under a minute, and check each includes the word 'mean' or 'proportion', the units, and the direction of the effect.