Further Statistics: Hypothesis Testing and Chi-Squared Depth - Worksheets, Questions and Revision
20 original exam-style questions - 12 pages of questions with a full mark scheme - free printable PDF.
Further Statistics: Hypothesis Testing and Chi-Squared Depth revision guideWrite-in editionMore A Level Further Maths practiceMock exams
This topic is chapter 6 of A Level Further Maths Further Statistics Topic Workbook and chapter 16 of A Level Further Maths 10-Minute Tests 2.
2.6 Further Statistics: Hypothesis Testing and Chi-Squared Depth
Key facts: chi-squared tests, in more depth
Original content written for Revision Library.
A chi-squared test compares observed frequencies O with the expected frequencies E predicted by a model, using X^2 = sum of (O-E)^2/E over all classes or cells, after any pooling. For a goodness-of-fit test to a fully specified distribution, df = (number of classes) - 1; if a parameter is estimated from the same sample (for example a Poisson mean, a binomial p, or a geometric p), one further degree of freedom is lost for each parameter estimated, because the estimation process artificially improves the apparent fit. For continuous data grouped into classes, expected frequencies are found from probabilities calculated using the standard Normal distribution, via z = (x - mu)/sigma and the function Phi. For an r by c contingency table testing independence between two categorical variables, E = (row total * column total)/grand total (this follows directly from the multiplication rule P(row and column) = P(row) * P(column) applied to the sample), and df = (r-1)(c-1). Classes or cells with an expected frequency below 5 should be pooled with a neighbouring class or column before the test statistic is calculated, and the degrees of freedom adjusted accordingly. Yates' continuity correction, X^2 = sum of (|O-E| - 0.5)^2/E, is applied only when df = 1 (every 2 by 2 contingency table), since this is the only case where the discrete test statistic is poorly approximated by the continuous chi-squared distribution. The signed standardised residual for a cell, e = (O-E)/sqrt(E), shows the direction as well as the size of a departure from the model, and the sum of its squares reproduces X^2. Because X^2 is never negative, only large (upper-tail) values ever provide evidence against H0; the calculated statistic is compared with the critical value from chi-squared tables at the chosen significance level and the stated degrees of freedom, and H0 is rejected if the statistic exceeds this value. As with any hypothesis test, the significance level equals P(Type I error), reducing it increases P(Type II error) for a fixed sample size, and increasing the sample size generally increases the power of the test; a chi-squared statistic also scales with sample size, so a very large sample can make even a small, unimportant deviation from a model 'statistically significant'. Combining ('pooling') data from distinct groups before testing can also mask or even reverse a genuine association found within each group, an effect known as Simpson's paradox.
Attended club: passed 60, did not pass 20, row total 80.
Did not attend: passed 30, did not pass 40, row total 70.
Column totals: passed 90, did not pass 60, grand total 150.
Weight class: under 985g, 985g up to 995g, 995g up to 1005g, 1005g up to 1015g, 1015g or more.
Observed frequency: 40, 38, 50, 45, 27.
GCSE: Monday 38, Tuesday 42, Wednesday 32, Thursday 8, row total 120.
A-level: Monday 22, Tuesday 19, Wednesday 15, Thursday 4, row total 60.
Column totals: Monday 60, Tuesday 61, Wednesday 47, Thursday 12, grand total 180.
Dog owners: insured 42, not insured 18, row total 60.
Cat owners: insured 24, not insured 36, row total 60.
Column totals: insured 66, not insured 54, grand total 120.
For a chi-squared test, the standardised residual for a cell is defined as e = (O-E)/√E. Unlike the term (O-E)2/E used in the test statistic itself, the sign of e shows whether that cell's observed frequency was higher or lower than expected.
Question 1
- (a) H0 for a test of association states precisely that the two categorical variables are independent (oe)B1
- (a) i.e. for every row category and column category, P(that row and that column) = P(that row) * P(that column)B1
- (a) Answer: H0 is exactly the statement that the two variables are independent: P(row and column) = P(row) * P(column) for every cell.
- (b) under independence, P(row 1 and column 1) = (R1/n) * (C1/n)M1
- (b) cso: expected frequency = n * P(row 1 and column 1) = n * (R1/n) * (C1/n) = R1*C1/n, as requiredA1
- (b) Answer: E(row 1, col 1) = n * (R1/n) * (C1/n) = R1*C1/n, matching the standard formula.
- (c) a test statistic of 0 means every observed frequency in the sample exactly equals its expected (independence) value, so this sample gives the strongest possible evidence in favour of H0B1
- (c) however this only describes the particular sample collected; it does not prove that the underlying population variables are truly independent, since a different sample could show some discrepancy. It is evidence for, not proof of, independenceB1
- (c) Answer: Not fully accurate: a value of 0 gives the strongest possible sample evidence for independence, but it does not prove the population variables are independent.
Question 2
- (a) when a parameter (e.g. a Poisson mean or a binomial p) is estimated from the sample, it is deliberately chosen (e.g. by matching the sample mean) so that the fitted distribution reproduces key features of that very sample as closely as possibleB1
- (a) this means the resulting expected frequencies tend to be closer to the observed frequencies, and the test statistic tends to be smaller, than if a genuinely independent, pre-specified parameter value had been usedB1
- (a) reducing the degrees of freedom by 1 for each estimated parameter compensates for this artificial closeness of fit, so the test keeps its stated significance levelB1
- (a) Answer: Estimating a parameter from the data makes the model fit that data unusually well by construction; losing one degree of freedom per estimated parameter corrects for this so the test keeps its stated significance level.
- (b) (i) df = 6 - 1 = 5B1
- (b) (ii) df = 6 - 1 - 1 = 4B1
- (b) (iii) df = 6 - 1 - 2 = 3B1
- (b) Answer: (i) 5 (ii) 4 (iii) 3
Question 3
- (a) H0: attending the Saturday club and passing the mock exam are independent (no association)B1
- (a) H1: attending the Saturday club and passing the mock exam are associatedB1
- (a) Answer: H0: club attendance and passing the mock are independent. H1: they are associated.
- (b) uses E = (row total * column total) / grand totalM1
- (b) all four correct: 48, 32, 42, 28A1
- (b) Answer: 48, 32, 42, 28
- (c) uses X2 = sum of (O-E)2/EM1
- (c) at least 3 terms correct (unsimplified): 144/48, 144/32, 144/42, 144/28A1
- (c) cao, awrt 16.07A1
- (c) Answer: X2 = awrt 16.07 (accept 16.1)
- (d) df = (2-1)(2-1) = 1; 16.07 > 3.841, so reject H0: there is significant evidence of an association between attending the Saturday club and passing the mock examA1
- (d) since 16.07 also exceeds 7.879 (the 0.5% critical value), this is very strong evidence, significant even at the 0.5% levelA1
- (d) Answer: df = 1; 16.07 exceeds even the 0.5% critical value, so there is very strong evidence of an association.
- (e) Yates' correction is usually recommended when the sample size (or one or more expected frequencies) is small, since the discrete chi-squared statistic is then poorly approximated by the continuous chi-squared distributionB1
- (e) the correction can only ever reduce the test statistic slightly (each |O-E| is reduced by 0.5 before squaring); since 16.07 so greatly exceeds even the most stringent critical value quoted (7.879), a small reduction could not change the conclusionB1
- (e) Answer: Recommended when expected frequencies/sample size are small; here the margin over the critical value is so large that a small correction cannot change the outcome.
- (f) the correction compensates for approximating a discrete test statistic by the continuous chi-squared distribution, and this approximation is least accurate specifically when df = 1B1
- (f) for tables with more categories (df > 1) the approximation is already good enough without correction, and there is no single agreed way to extend an analogous correction across several cells, so it is conventionally omittedB1
- (f) Answer: The df = 1 approximation is the one that needs correcting; for larger tables the chi-squared approximation is already adequate and no standard correction is defined.
Question 4
- (a) uses X2 = sum of (O-E)2/E with E = 12.5 in every categoryM1
- (a) cao: X2 = 4.72A1
- (a) 4.72 < 7.815, so do not reject H0: no significant evidence against the categories being equally likelyA1
- (a) Answer: X2 = 4.72; not significant at 5%.
- (b) substitutes O' = 10*O and E' = 10*E into the formula: sum of (10*O - 10*E)2 / (10*E)M1
- (b) simplifies: sum of 100*(O-E)2 / (10*E) = 10 * sum of (O-E)2/EM1
- (b) cso: new test statistic = 10 * 4.72 = 47.2A1
- (b) Answer: New test statistic = 10 * 4.72 = 47.2
- (c) 47.2 > 7.815, so the result is now significant at the 5% level, in contrast to part (a)A1
- (c) the exact same relative (proportional) pattern of deviation from the expected frequencies, which was not significant for n = 50, becomes highly significant purely because the sample size increasedB1
- (c) this shows that chi-squared significance reflects how confidently a deviation can be detected, not the size or practical importance of the deviation itself; a very large sample can make even a small, practically unimportant deviation 'statistically significant'B1
- (c) Answer: 47.2 > 7.815, now significant; the same proportional pattern becomes significant purely because of a larger sample size, showing significance is not the same as practical importance.
Question 5
- (a) 8.24 > 5.991, so reject H0A1
- (a) there is significant evidence of an association between the area a resident lives in and their preferred collection frequencyA1
- (a) Answer: 8.24 > 5.991, reject H0; significant evidence of an association between area and preferred collection frequency.
- (b) rejecting H0 when H0 is actually true (oe)B1
- (b) in context: concluding that there is an association between area and preferred collection frequency when, in reality, no such association existsB1
- (b) Answer: Concluding there is an association between area and collection preference when in reality there is none.
- (c) 0.05B1
- (c) Answer: 0.05
- (d) lower the significance level (e.g. use 1% instead of 5%)B1
- (d) this would increase the probability of a Type II error, since it becomes harder to reject H0, making it more likely that a real association is missed (not detected)B1
- (d) Answer: Use a lower significance level (e.g. 1%); this increases the probability of a Type II error.
- (e) a larger sample gives more precise, reliable estimates of the expected frequencies and increases the power of the testB1
- (e) this makes the test more able to detect a genuine association where one exists, reducing the chance of a false 'no association' conclusion (a Type II error)B1
- (e) Answer: Larger samples increase the power of the test, making a genuine association more likely to be detected, so Type II error probability falls.
- (f) all expected frequencies (under H0, for each area/collection-frequency combination) must be at least 5B1
- (f) Answer: All expected frequencies for the area/collection-frequency cells must be at least 5.
Question 6
- (a) P(X = j) = e-1.504 1.504j/j! for j = 0, 1, 2M1
- (a) P(X ≥ 3) = 1 - P(X ≤ 2)M1
- (a) expected: 26.45, 39.78, 29.91, 22.87A1
- (a) Answer: 26.45, 39.78, 29.91, 22.87
- (b) chi2 = sum of (O - E)2/EM1
- (b) 0.995A1
- (b) degrees of freedom = 4 - 1 - 1 = 2 (the mean was estimated from the data)B1
- (b) do not reject H0: the Poisson model is a suitable fitA1
- (b) Answer: chi2 = 0.995, nu = 2, do not reject H0
Question 7
- (a) H0: p = 0.25B1
- (a) H1: p < 0.25B1
- (a) Answer: H0: p = 0.25, H1: p < 0.25
- (b) P(X ≥ 13) = (1 - 0.25)13 - 1M1
- (b) 0.0317A1
- (b) reject H0A1
- (b) Answer: 0.0317; reject H0
- (c) finds the smallest d with (1 - 0.25)d - 1 ≤ 0.05M1
- (c) critical region X ≥ 12A1
- (c) P(X ≤ 11 | p = 0.1) = 1 - (1 - 0.1)11M1
- (c) 0.686A1
- (c) Answer: X ≥ 12; Type II error 0.686
Question 8
- (a) P(X = j) = e-0.946 0.946j/j! for j = 0, 1, 2M1
- (a) P(X ≥ 3) = 1 - P(X ≤ 2)M1
- (a) expected: 57.86, 54.73, 25.89, 10.53A1
- (a) Answer: 57.86, 54.73, 25.89, 10.53
- (b) chi2 = sum of (O - E)2/EM1
- (b) 0.944A1
- (b) degrees of freedom = 4 - 1 - 1 = 2 (the mean was estimated from the data)B1
- (b) do not reject H0: the Poisson model is a suitable fitA1
- (b) Answer: chi2 = 0.944, nu = 2, do not reject H0
Question 9
- (a) uses B(12, 0.25) cumulative probabilitiesM1
- (a) critical region X ≥ 6A1
- (a) Answer: critical region X ≥ 6
- (b) P(X in critical region | p = 0.25)M1
- (b) 0.0544A1
- (b) Answer: 0.0544
- (c) P(X not in critical region | p = 0.7)M1
- (c) 0.0386A1
- (c) Answer: 0.0386
Question 10
- (a) uses Po(6) probabilities in the appropriate tailM1
- (a) P(X ≤ 1) = 0.0174 ≤ 0.05A1
- (a) critical region X ≤ 1A1
- (a) Answer: critical region X ≤ 1
- (b) the probability of a Type I error = P(X in the critical region | H0 true)M1
- (b) 0.0174A1
- (b) Answer: 0.0174
- (c) P(X not in the critical region | λ = 3)M1
- (c) 0.801A1
- (c) Answer: 0.801
Question 11
- (a) H0: the model (ratio 1:1:1:1) is a suitable fit; H1: it is notB1
- (a) expected = N x (ratio part)/4M1
- (a) expected: 30, 30, 30, 30A1
- (a) Answer: 30, 30, 30, 30
- (b) sum of (O - E)2/EM1
- (b) chi2 = 2.40A1
- (b) degrees of freedom = 4 - 1 = 3B1
- (b) Answer: chi2 = 2.40, nu = 3
- (c) compares 2.40 with 7.815M1
- (c) do not reject H0A1
- (c) there is no evidence against the model: the data are consistent with the ratioB1
- (c) Answer: do not reject H0
Question 12
- (a) H0: the model (ratio 1:1:1:1) is a suitable fit; H1: it is notB1
- (a) expected = N x (ratio part)/4M1
- (a) expected: 16.75, 16.75, 16.75, 16.75A1
- (a) Answer: 16.75, 16.75, 16.75, 16.75
- (b) sum of (O - E)2/EM1
- (b) chi2 = 2.55A1
- (b) degrees of freedom = 4 - 1 = 3B1
- (b) Answer: chi2 = 2.55, nu = 3
- (c) compares 2.55 with 7.815M1
- (c) do not reject H0A1
- (c) there is no evidence against the model: the data are consistent with the ratioB1
- (c) Answer: do not reject H0
Question 13
- (a) H0: the model (ratio 1:1:1:1) is a suitable fit; H1: it is notB1
- (a) expected = N x (ratio part)/4M1
- (a) expected: 40.5, 40.5, 40.5, 40.5A1
- (a) Answer: 40.5, 40.5, 40.5, 40.5
- (b) sum of (O - E)2/EM1
- (b) chi2 = 4.42A1
- (b) degrees of freedom = 4 - 1 = 3B1
- (b) Answer: chi2 = 4.42, nu = 3
- (c) compares 4.42 with 11.345M1
- (c) do not reject H0A1
- (c) there is no evidence against the model: the data are consistent with the ratioB1
- (c) Answer: do not reject H0
Question 14
- (a) H0: p = 0.15B1
- (a) H1: p < 0.15B1
- (a) Answer: H0: p = 0.15, H1: p < 0.15
- (b) P(X ≥ 24) = (1 - 0.15)24 - 1M1
- (b) 0.0238A1
- (b) reject H0A1
- (b) Answer: 0.0238; reject H0
- (c) finds the smallest d with (1 - 0.15)d - 1 ≤ 0.1M1
- (c) critical region X ≥ 16A1
- (c) P(X ≤ 15 | p = 0.08) = 1 - (1 - 0.08)15M1
- (c) 0.714A1
- (c) Answer: X ≥ 16; Type II error 0.714
Question 15
- (a) H0: outcome and group are independent; H1: they are associatedB1
- (a) expected = (row total x column total)/grand totalM1
- (a) E11 = 1280/49 = 26.12, E21 = 680/49 = 13.88A1
- (a) Answer: E11 = 26.12, E21 = 13.88
- (b) sum of (O - E)2/E over all six cellsM1
- (b) chi2 = 0.582A1
- (b) degrees of freedom = (2 - 1)(3 - 1) = 2B1
- (b) Answer: chi2 = 0.582, nu = 2
- (c) compares 0.582 with 9.21M1
- (c) do not reject H0A1
- (c) there is no evidence of an association between group and outcomeB1
- (c) Answer: do not reject H0
Question 16
- (a) H0: outcome and group are independent; H1: they are associatedB1
- (a) expected = (row total x column total)/grand totalM1
- (a) E11 = 1000/29 = 34.48, E21 = 740/29 = 25.52A1
- (a) Answer: E11 = 34.48, E21 = 25.52
- (b) sum of (O - E)2/E over all six cellsM1
- (b) chi2 = 2.30A1
- (b) degrees of freedom = (2 - 1)(3 - 1) = 2B1
- (b) Answer: chi2 = 2.30, nu = 2
- (c) compares 2.30 with 9.21M1
- (c) do not reject H0A1
- (c) there is no evidence of an association between group and outcomeB1
- (c) Answer: do not reject H0
Question 17
- (a) H0: outcome and group are independent; H1: they are associatedB1
- (a) expected = (row total x column total)/grand totalM1
- (a) E11 = 86/3 = 28.67, E21 = 58/3 = 19.33A1
- (a) Answer: E11 = 28.67, E21 = 19.33
- (b) sum of (O - E)2/E over all six cellsM1
- (b) chi2 = 0.0491A1
- (b) degrees of freedom = (2 - 1)(3 - 1) = 2B1
- (b) Answer: chi2 = 0.0491, nu = 2
- (c) compares 0.0491 with 5.991M1
- (c) do not reject H0A1
- (c) there is no evidence of an association between group and outcomeB1
- (c) Answer: do not reject H0
Question 18
- (a) z = (985-1000)/15 = -1 shownM1
- (a) other z-values correct: -1/3 (awrt -0.33), 1/3 (awrt 0.33), 1A1
- (a) Answer: z = -1, -0.333, 0.333, 1
- (b) P(X<985) = Phi(-1) = 1 - Phi(1) = 1 - 0.8413 = 0.1587, shownM1
- (b) uses Phi(-0.3333) = 1 - 0.6306 = 0.3694 and subtraction to find the middle classesM1
- (b) all five probabilities correct: 0.1587, 0.2107, 0.2612, 0.2107, 0.1587A1
- (b) Answer: 0.1587, 0.2107, 0.2612, 0.2107, 0.1587
- (c) multiplies each probability by 200M1
- (c) all correct: 31.74, 42.14, 52.24, 42.14, 31.74A1
- (c) Answer: 31.74, 42.14, 52.24, 42.14, 31.74 (sum = 200.00)
- (d) uses X2 = sum of (O-E)2/E with the correct O and E valuesM1
- (d) at least 3 terms correctA1
- (d) awrt 3.55 (allow awrt 3.56 if unrounded calculator probabilities are used)A1
- (d) Answer: X2 = awrt 3.55 (3.56 using unrounded expected frequencies)
- (e) df = 5 - 1 = 4 (both mean and standard deviation were fixed, claimed values, not estimated); critical value chi-squared4(0.05) = 9.488B1
- (e) ft: 3.55 < 9.488, so do not reject H0; there is no significant evidence against the manufacturer's claimed N(1000, 152) model for the bag weightsA1
- (e) Answer: df = 4, critical value 9.488; do not reject H0, the claimed Normal model is a good fit.
- (f) df would become 5 - 1 - 1 = 3B1
- (f) because one additional parameter (the standard deviation) would then be estimated from the sample data, and each estimated parameter reduces the degrees of freedom by a further 1B1
- (f) Answer: df would decrease to 3, since σ would then be estimated from the data.
Question 19
- (a) uses E = (row total * column total) / grand totalM1
- (a) at least 4 correctA1
- (a) all 8 correct: 40, 40.67, 31.33, 8, 20, 20.33, 15.67, 4A1
- (a) Answer: GCSE: 40, 40.67, 31.33, 8. A-level: 20, 20.33, 15.67, 4.
- (b) the A-level, Thursday cell has expected frequency 4, which is below 5, so the validity condition for the chi-squared approximation is not metB1
- (b) the whole Thursday column should therefore be merged with the neighbouring Wednesday column, giving a combined 'Wednesday or Thursday' category for both rowsB1
- (b) Answer: A-level/Thursday has E = 4 < 5; merge the Thursday column into Wednesday.
- (c) combines the observed frequencies: GCSE 38, 42, 40; A-level 22, 19, 19M1
- (c) combines the expected frequencies correctly: GCSE 40, 40.67, 39.33; A-level 20, 20.33, 19.67A1
- (c) Answer: O: GCSE 38, 42, 40; A-level 22, 19, 19. E: GCSE 40, 40.67, 39.33; A-level 20, 20.33, 19.67.
- (d) df = (2-1)(3-1) = 2B1
- (d) Answer: df = 2
- (e) uses X2 = sum of (O-E)2/E on the merged tableM1
- (e) at least 3 terms correctA1
- (e) cao, awrt 0.465A1
- (e) Answer: X2 = awrt 0.465
- (f) 0.465 < 5.991A1
- (f) ft: do not reject H0; no significant evidence of an association between key stage and preferred session dayA1
- (f) Answer: 0.465 < 5.991, do not reject H0; no evidence of association.
Question 20
- (a) E = 33, 27, 33, 27 (e.g. 60*66/120 = 33)B1
- (a) Answer: 33, 27, 33, 27
- (b) uses e = (O-E)/√E for at least one cellM1
- (b) at least 2 correctA1
- (b) all four correct: 1.57, -1.73, -1.57, 1.73A1
- (b) Answer: Dog/insured: 1.57. Dog/not insured: -1.73. Cat/insured: -1.57. Cat/not insured: 1.73.
- (c) sums the squares of the four residuals: 1.572 + 1.732 + 1.572 + 1.732M1
- (c) awrt 10.9, which exceeds all three critical values, so the result is significant even at the 0.5% levelA1
- (c) Answer: X2 = awrt 10.9, significant at the 5%, 1% and 0.5% levels.
- (d) dog owners are more likely than expected under independence to have insurance (positive residual) and less likely than expected to be uninsured, while cat owners show the opposite pattern; i.e. dog owners are more likely to be insured than cat ownersB1
- (d) Answer: Dog owners are more likely to have pet insurance than cat owners.
Mark your answers
Check answers without an account. Some answers can be checked automatically; use the mark scheme to mark the rest.