A-Level Further Maths Statistics: Poisson, Chi-Squared and Hypothesis Tests — 进阶数学统计备考指南

一、9665 统计模块考什么：考试大纲与题型结构 | What the 9665 Statistics Option Covers: Syllabus and Question Patterns

9665 是牛津AQA国际A-Level进阶数学（OxfordAQA International A-Level Further Mathematics）的课程代码。在完成纯数（Pure Mathematics）与力学（Mechanics）等必修内容之后，统计模块是进阶数学中最常用的应用分支之一，也是许多大学数学、经济、工程与数据科学专业明确看重的部分。统计选项的考查范围高度集中：离散随机变量、泊松分布、卡方检验、相关与回归、假设检验，五大板块反复出现在历年试卷中。

9665 is the specification code for the OxfordAQA International A-Level Further Mathematics qualification. After completing the compulsory Pure Mathematics and Mechanics content, the Statistics option is one of the most popular applied branches of Further Maths, and it is explicitly valued by many university courses in mathematics, economics, engineering and data science. The statistics option has a tightly focused syllabus: discrete random variables, the Poisson distribution, chi-squared tests, correlation and regression, and hypothesis testing. These five blocks recur on past papers year after year.

考试题型通常分为两类：一类是纯计算题，直接考查公式运用与查表能力；另一类是情境应用题，把统计方法嵌入现实场景，例如工厂产品的缺陷数、医院急诊的到达人数、网站点击量等。后者更看重学生能否正确选择模型、写出假设并解释结论，这正是很多中国学生容易失分的地方，因为结论解释需要用规范的统计语言而非大白话。

Exam questions come in two broad types. The first is pure calculation, which tests formula manipulation and table-reading skills directly. The second is applied or contextual, embedding statistical methods in real situations such as the number of defective items from a factory line, arrivals at a hospital emergency department, or website clicks. The applied type rewards students who can choose the correct model, write down hypotheses and interpret conclusions properly. This is where many students lose marks, because the conclusion must be expressed in precise statistical language rather than everyday prose.

从分值占比看，统计模块在整套进阶数学中约占四分之一到三分之一，具体比例随考试局版本略有差异。对志在拿到 A* 的学生来说，统计部分几乎是必须满分的目标区，因为它的套路固定、计算量适中，远没有纯数的证明题那样难以预测。建议备考时按板块逐个击破，先掌握分布与检验的适用条件，再刷历年 topic test 与真题。

In terms of marks, the statistics module accounts for roughly a quarter to a third of the whole Further Maths qualification, with minor variation between exam board versions. For students aiming at an A*, the statistics section is almost a mandatory full-marks target: the patterns are fixed, the computation load is moderate, and it is far more predictable than the proof questions in Pure Mathematics. The recommended approach is to master each block in turn, first learning the conditions under which each distribution or test applies, then working through past topic tests and real papers.

二、离散随机变量回顾：E(X) 与 Var(X) 的计算与线性变换 | Discrete Random Variables Refresher: Computing E(X) and Var(X), and Linear Transformations

统计模块的一切都建立在离散随机变量的期望与方差之上。设随机变量 X 的取值为 x₁, x₂, …, xₙ，对应概率为 p₁, p₂, …, pₙ，则期望 E(X) = Σ xᵢpᵢ，它衡量分布的中心位置；方差 Var(X) = E(X²) − [E(X)]²，它衡量分布的离散程度。这里最容易出错的是：必须先算 E(X²) = Σ xᵢ²pᵢ，再减去期望的平方，绝不能把 E(X²) 误写成 [E(X)]²。

Everything in the statistics option rests on the expectation and variance of discrete random variables. If X takes values x₁, x₂, …, xₙ with probabilities p₁, p₂, …, pₙ, then the expectation E(X) = Σ xᵢpᵢ describes the centre of the distribution, while the variance Var(X) = E(X²) − [E(X)]² measures its spread. The most common error is forgetting that you must first compute E(X²) = Σ xᵢ²pᵢ and then subtract the square of the expectation; E(X²) is not the same as [E(X)]².

线性变换规则在考试中几乎必考：若 Y = aX + b，则 E(Y) = aE(X) + b，Var(Y) = a²Var(X)。注意方差对平移 b 不敏感，却对伸缩 a 取平方。例如把温度从摄氏度换算成华氏度 F = 1.8C + 32，期望按同样公式换算，但方差要乘以 1.8² = 3.24。理解这条规则后，很多看似复杂的题目可以直接化简。

Linear transformations are almost guaranteed to appear: if Y = aX + b, then E(Y) = aE(X) + b and Var(Y) = a²Var(X). Notice that the variance is unaffected by the shift b but is multiplied by a² under scaling. For example, converting temperatures from Celsius to Fahrenheit via F = 1.8C + 32 transforms the expectation with the same formula, but the variance is multiplied by 1.8² = 3.24. Once this rule is understood, many apparently complicated questions simplify immediately.

典型例题：掷一枚均匀六面骰子，令 X 为点数。则 E(X) = 3.5，E(X²) = (1+4+9+16+25+36)/6 = 15.1667，故 Var(X) = 15.1667 − 12.25 = 2.9167。若每次掷骰奖励 2X + 1 元，则期望奖励为 2×3.5 + 1 = 8 元，方差为 4×2.9167 = 11.6667。这类小计算看似简单，却是后面泊松分布与正态近似的运算基础，务必做到又快又准。

Worked example: roll a fair six-sided die and let X be the score. Then E(X) = 3.5, E(X²) = (1+4+9+16+25+36)/6 = 15.1667, so Var(X) = 15.1667 − 12.25 = 2.9167. If the prize money for one roll is 2X + 1 yuan, the expected prize is 2×3.5 + 1 = 8 yuan and the variance is 4×2.9167 = 11.6667. These small calculations look trivial, but they are the computational foundation for the Poisson distribution and the normal approximation, so they must be done quickly and accurately.

三、泊松分布的核心条件：稀有事件、独立性与均值等于方差 | The Poisson Distribution: Rare Events, Independence and the Mean Equals Variance Property

泊松分布是描述稀有事件在固定时间或空间内发生次数的经典模型。随机变量 X 服从参数为 λ 的泊松分布，记作 X ~ Po(λ)，其概率质量函数为 P(X = r) = e^(−λ) λʳ / r!，其中 r = 0, 1, 2, …。λ 是单位时间（或单位空间）内事件的平均发生次数，也是该分布唯一的参数。

The Poisson distribution is the classic model for the number of times a rare event occurs in a fixed interval of time or space. A random variable X follows a Poisson distribution with parameter λ, written X ~ Po(λ), with probability mass function P(X = r) = e^(−λ) λʳ / r! for r = 0, 1, 2, … . The parameter λ is the average number of occurrences per unit time (or unit space) and is the only parameter of the distribution.

使用泊松分布前必须验证三个条件：第一，事件在不相交的区间内独立发生，即一个区间内的发生数不影响另一个区间；第二，事件不能同时发生，也就是在极短的时间内至多发生一次；第三，事件以恒定的平均速率发生，λ 不随时间或空间位置变化。考试中常见的设题场景包括：每分钟到达服务台的顾客数、每页印刷错误数、放射性物质单位时间的衰变数、道路单位长度的坑洞数。

Before using the Poisson distribution you must check three conditions: first, events occur independently in disjoint intervals, so the count in one interval does not affect the count in another; second, events cannot occur simultaneously, meaning at most one event in any very short interval; third, events occur at a constant average rate, so λ does not change over time or position. Typical exam scenarios include the number of customers arriving at a desk per minute, the number of printing errors per page, the number of radioactive decays per unit time, and the number of potholes per unit length of road.

泊松分布最著名的性质是均值等于方差：E(X) = Var(X) = λ。这一性质有两个用途：一是检查数据是否可能来自泊松分布（若样本均值与方差相差悬殊，则模型不适用）；二是在选择题或短答题中快速验证答案是否合理。当题目给出样本均值和方差并要求判断分布类型时，均值与方差接近相等就是选择泊松分布的重要依据。

The most famous property of the Poisson distribution is that the mean equals the variance: E(X) = Var(X) = λ. This has two uses: it helps you check whether data could plausibly come from a Poisson distribution (if the sample mean and variance differ wildly, the model is inappropriate), and it gives a quick sanity check for answers in multiple-choice or short questions. When a question gives a sample mean and variance and asks you to identify the distribution, near-equality of mean and variance is a strong signal for the Poisson model.

四、泊松分布查表与计算器技巧：累计概率 P(X ≤ k) 与补事件 | Poisson Tables and Calculator Skills: Cumulative Probabilities P(X ≤ k) and Complementary Events

考试提供泊松分布累计概率表，给出 P(X ≤ k) 在不同 λ 下的数值。读表的关键是搞清楚题目要的是哪种概率：P(X = k) 要用 P(X ≤ k) − P(X ≤ k−1)；P(X ≥ k) 要用 1 − P(X ≤ k−1)；P(X > k) 要用 1 − P(X ≤ k)；P(X < k) 要用 P(X ≤ k−1)。把边界条件写清楚，是这类题不丢分的前提。

Exams provide cumulative Poisson probability tables giving P(X ≤ k) for various values of λ. The key to reading the table is knowing exactly which probability the question wants: P(X = k) is P(X ≤ k) − P(X ≤ k−1); P(X ≥ k) is 1 − P(X ≤ k−1); P(X > k) is 1 − P(X ≤ k); and P(X < k) is P(X ≤ k−1). Writing the boundary conditions down clearly is the prerequisite for scoring full marks on these questions.

当 λ 不在表列出的整数中时，可以取相邻的两个 λ 做线性插值，但考试通常会把 λ 设计成表内数值。若 λ 超过表的范围（例如 λ = 20），则应改用正态近似（见下一节）。另外，很多现代图形计算器内置 PoissonCDF 功能，可以直接输出 P(X ≤ k)，建议平时练习就熟悉自己计算器的菜单路径，考试时先用计算器算一遍，再与查表结果互相印证，避免低级误差。

When λ is not one of the tabulated integers, linear interpolation between the two neighbouring values is acceptable, but exams usually set λ to a tabulated value. If λ exceeds the range of the tables (for example λ = 20), you should switch to the normal approximation described in the next section. Moreover, many modern graphical calculators include a PoissonCDF function that outputs P(X ≤ k) directly; it is wise to learn the menu path of your own calculator during practice, then cross-check the calculator result against the tables in the exam to avoid careless errors.

示例：设 X ~ Po(3)，求 P(X = 2) 与 P(X ≥ 3)。查表得 P(X ≤ 2) = 0.4232，P(X ≤ 1) = 0.1991，所以 P(X = 2) = 0.4232 − 0.1991 = 0.2241；而 P(X ≥ 3) = 1 − P(X ≤ 2) = 1 − 0.4232 = 0.5768。注意 P(X ≥ 3) 包含 X = 3、4、5、… 所有值，所以用的是 P(X ≤ 2) 而非 P(X ≤ 3)，这一字之差正是最常见的陷阱。

Example: let X ~ Po(3) and find P(X = 2) and P(X ≥ 3). From the tables, P(X ≤ 2) = 0.4232 and P(X ≤ 1) = 0.1991, so P(X = 2) = 0.4232 − 0.1991 = 0.2241, while P(X ≥ 3) = 1 − P(X ≤ 2) = 1 − 0.4232 = 0.5768. Note that P(X ≥ 3) includes X = 3, 4, 5, … and therefore uses P(X ≤ 2) rather than P(X ≤ 3); that one-word difference is the most common trap in this type of question.

五、正态近似泊松：近似条件、连续性修正与标准化 | Approximating Poisson by the Normal Distribution: Conditions, the Continuity Correction and Standardisation

当 λ 足够大时（多数考试局以 λ > 10 为界），泊松分布的形状趋于对称，可以用正态分布近似：X ~ Po(λ) 近似为 Y ~ N(λ, λ)。此时所有计算都转成标准正态分布 Z = (Y − λ)/√λ，配合 Z 值表完成概率求解。近似的好处是摆脱了泊松表的 λ 上限限制。

When λ is large enough (most boards use λ > 10 as the rule of thumb), the Poisson distribution becomes roughly symmetric and can be approximated by a normal distribution: X ~ Po(λ) is approximated by Y ~ N(λ, λ). All calculations then convert to the standard normal Z = (Y − λ)/√λ using Z-tables. The advantage of the approximation is that it removes the upper limit on λ imposed by the Poisson tables.

由于泊松是离散分布而正态是连续分布，近似时必须做连续性修正：把离散值 k 视作连续区间 (k − 0.5, k + 0.5)。具体规则为：P(X ≤ k) ≈ P(Y ≤ k + 0.5)；P(X < k) ≈ P(Y ≤ k − 0.5)；P(X ≥ k) ≈ P(Y ≥ k − 0.5)；P(X > k) ≈ P(Y ≥ k + 0.5)。漏掉这 0.5 的修正会导致答案偏差，在只差 0.01 的临界题上足以改变结论。

Because Poisson is discrete and the normal distribution is continuous, the approximation requires a continuity correction: the discrete value k is treated as the continuous interval (k − 0.5, k + 0.5). The rules are: P(X ≤ k) ≈ P(Y ≤ k + 0.5); P(X < k) ≈ P(Y ≤ k − 0.5); P(X ≥ k) ≈ P(Y ≥ k − 0.5); and P(X > k) ≈ P(Y ≥ k + 0.5). Omitting the half-unit correction biases the answer, and on a borderline question separated by 0.01 it can change the conclusion.

完整示例：设 X ~ Po(15)，求 P(X ≤ 12)。用 Y ~ N(15, 15) 近似，先修正边界：P(X ≤ 12) ≈ P(Y ≤ 12.5)。标准化得 Z = (12.5 − 15)/√15 = −2.5/3.873 = −0.645。查标准正态表，P(Z ≤ −0.645) = 1 − Φ(0.645) ≈ 1 − 0.7405 = 0.2595。若忘记连续性修正而直接用 12，则 Z = −0.775，概率为 0.2192，两者相差 0.04，足以让答案失分。

Full example: let X ~ Po(15) and find P(X ≤ 12). Using Y ~ N(15, 15) as the approximation, first correct the boundary: P(X ≤ 12) ≈ P(Y ≤ 12.5). Standardising gives Z = (12.5 − 15)/√15 = −2.5/3.873 = −0.645. From the normal tables, P(Z ≤ −0.645) = 1 − Φ(0.645) ≈ 1 − 0.7405 = 0.2595. If you forget the continuity correction and use 12 directly, Z = −0.775 giving a probability of 0.2192; the 0.04 difference is enough to cost marks.

六、卡方拟合优度检验：检验观测数据是否符合理论分布 | The Chi-Squared Goodness-of-Fit Test: Does the Observed Data Fit the Theoretical Model?

拟合优度检验回答的问题是：一组观测频数是否与某个理论分布（均匀、泊松、正态、二项等）一致。检验统计量为 X² = Σ (Oᵢ − Eᵢ)² / Eᵢ，其中 Oᵢ 是第 i 类的观测频数，Eᵢ 是理论频数。X² 越小说明拟合越好；X² 超过临界值则拒绝原假设，认为数据不符合该分布。

The goodness-of-fit test answers the question: do a set of observed frequencies agree with a theoretical distribution (uniform, Poisson, normal, binomial and so on)? The test statistic is X² = Σ (Oᵢ − Eᵢ)² / Eᵢ, where Oᵢ is the observed frequency in class i and Eᵢ is the expected frequency. A small X² means a good fit; if X² exceeds the critical value, the null hypothesis is rejected and the data is deemed inconsistent with the distribution.

自由度（degrees of freedom）的计算是本题型的核心考点：df = 类别数 − 1 − 被估计参数的个数。若理论分布的参数（如泊松的 λ、正态的均值和标准差）是从数据中估计出来的，每估计一个参数就多减去 1。例如用样本均值估计 λ 后检验泊松拟合，df = k − 2；若 λ 是事先给定的理论值，则 df = k − 1。自由度的细微差别会改变临界值，进而改变结论。

Degrees of freedom are the core of this question type: df = number of classes − 1 − number of parameters estimated from the data. If parameters of the theoretical distribution (such as λ for Poisson, or the mean and standard deviation for normal) are estimated from the data, subtract 1 for each estimated parameter. For example, testing a Poisson fit after estimating λ from the sample mean gives df = k − 2, whereas a pre-specified theoretical λ gives df = k − 1. The subtle difference in degrees of freedom changes the critical value and therefore the conclusion.

使用条件必须写清楚：每个期望频数 Eᵢ 应不小于 5，否则要把相邻类别合并，使合并后的期望频数达到要求。检验步骤为：先设 H₀（数据服从某分布）与 H₁，再计算各组期望频数、检验统计量 X²，查表得临界值，最后比较并下结论。结论必须用情境语言表述，例如“在 5% 显著性水平下，没有证据表明骰子不公平”。

The conditions of use must be stated clearly: every expected frequency Eᵢ should be at least 5; otherwise adjacent classes must be merged until the merged expected frequencies satisfy the requirement. The procedure is: state H₀ (the data follows the distribution) and H₁, calculate the expected frequencies and the test statistic X², look up the critical value, then compare and conclude. The conclusion must be expressed in context, for example “at the 5% significance level there is no evidence that the die is biased”.

七、卡方列联表检验：检验两个分类变量是否独立 | Chi-Squared Contingency Tables: Testing Whether Two Categorical Variables Are Independent

列联表检验用于判断两个分类变量是否独立。设有 r 行 c 列的表格，行变量与列变量的独立性是原假设，备择假设是两者相关。每个单元格的期望频数为 E = (行合计 × 列合计) / 总样本量，自由度 df = (r − 1)(c − 1)。检验统计量同样是 X² = Σ (O − E)² / E。

The contingency table test judges whether two categorical variables are independent. For a table with r rows and c columns, the null hypothesis is independence of the row and column variables, and the alternative is that they are associated. The expected frequency of each cell is E = (row total × column total) / grand total, with degrees of freedom df = (r − 1)(c − 1). The test statistic is again X² = Σ (O − E)² / E.

与拟合优度检验不同，列联表的期望频数没有参数估计的扣除，直接套公式即可。但注意两个细节：第一，如果超过 20% 的单元格期望频数小于 5，或任一单元格期望频数小于 1，检验结果不可靠，应合并行或列；第二，2×2 表格有时会要求使用耶茨连续性修正，具体以考试局规范为准，AQA 体系通常不强制，但题目会明确提示。

Unlike the goodness-of-fit test, the contingency table has no deduction for estimated parameters; you simply apply the formula. Two details matter though: first, if more than 20% of cells have expected frequencies below 5, or any cell below 1, the test is unreliable and rows or columns should be merged; second, 2×2 tables sometimes require Yates’s correction for continuity, depending on the board specification. AQA-based specifications usually do not force it, and the question will state clearly if it is needed.

示例：调查 200 名学生，考察性别与是否选修进阶数学是否独立。设男性中选修 70 人、未选修 30 人，女性中选修 50 人、未选修 50 人。行合计分别为 100 与 100，列合计分别为 120 与 80。则男性选修格的期望频数 E = 100×120/200 = 60，观测值 70 与期望 60 的偏差贡献为 (70−60)²/60 = 1.667。逐格计算后求和得 X²，与 df = 1 的临界值 3.841（5% 水平）比较，即可判断性别与选课是否相关。

Example: a survey of 200 students asks whether gender and choosing Further Maths are independent. Among males, 70 chose it and 30 did not; among females, 50 chose it and 50 did not. Row totals are 100 and 100; column totals are 120 and 80. The expected frequency for the male-chose cell is E = 100×120/200 = 60, and its contribution is (70−60)²/60 = 1.667. Summing the contributions of every cell gives X², which is compared with the critical value 3.841 (5% level) at df = 1 to decide whether gender and subject choice are related.

八、相关与回归：PMCC、Spearman 秩相关与最小二乘回归线 | Correlation and Regression: PMCC, Spearman’s Rank and the Least-Squares Line

相关分析衡量两个变量的线性关联强度。皮尔逊积矩相关系数（PMCC）r 的计算公式为 r = Sxy / √(Sxx·Syy)，其中 Sxx = Σ(x − x̄)² = Σx² − (Σx)²/n，Syy 同理，Sxy = Σxy − (Σx)(Σy)/n。r 的取值范围是 [−1, 1]，r = 1 为完全正线性相关，r = −1 为完全负线性相关，r 接近 0 表示线性关系很弱。

Correlation analysis measures the strength of the linear association between two variables. Pearson’s product-moment correlation coefficient (PMCC) is r = Sxy / √(Sxx·Syy), where Sxx = Σ(x − x̄)² = Σx² − (Σx)²/n, Syy is analogous, and Sxy = Σxy − (Σx)(Σy)/n. The value of r lies in [−1, 1]: r = 1 is perfect positive linear correlation, r = −1 perfect negative linear correlation, and r near 0 means a weak linear relationship.

当数据包含异常值或并非线性关系时，PMCC 可能产生误导，此时应使用 Spearman 秩相关系数 ρ。做法是先把两组数据分别按大小排序并赋予秩次（并列取平均秩），再对秩次计算 PMCC 公式。Spearman 系数对异常值不敏感，且能捕捉单调（不一定线性）的关系，是稳健性的首选。

When the data contains outliers or the relationship is not linear, the PMCC can mislead, and Spearman’s rank correlation coefficient ρ is the better tool. You first rank each data set separately (ties take the average rank), then apply the PMCC formula to the ranks. Spearman’s coefficient is insensitive to outliers and captures monotonic rather than purely linear relationships, making it the robust first choice.

回归部分要求掌握最小二乘回归线 y = a + bx，其中斜率 b = Sxy/Sxx，截距 a = ȳ − b·x̄。回归线用于在给定 x 时预测 y；反方向预测（给定 y 求 x）不能用反解，必须另算 x on y 的回归线 x = a′ + b′y。这是高频考点：用错了回归方向，预测值就是错的。另外，回归模型只在观测数据范围内可靠，外推预测要谨慎表述。

For regression you must master the least-squares line y = a + bx, with slope b = Sxy/Sxx and intercept a = ȳ − b·x̄. The line predicts y for a given x; predicting x from y requires the separate regression line x = a′ + b′y rather than solving the first line for x. This is a frequent exam point: using the wrong regression direction gives the wrong prediction. Also, the regression model is reliable only within the range of the observed data, so extrapolation should be described cautiously.

九、假设检验五步法：原假设、检验统计量、临界值与结论 | The Five-Step Hypothesis Test: Null Hypothesis, Test Statistic, Critical Values and Conclusion

假设检验是统计模块的灵魂，几乎所有板块都以它收尾。标准五步法为：第一步，写出原假设 H₀ 与备择假设 H₁，例如对泊松均值检验 H₀: λ = 3, H₁: λ > 3；第二步，确定显著性水平（通常为 5% 或 1%）并明确单尾或双尾；第三步，计算检验统计量（如 X = 观测的事件数）；第四步，求临界区域，即拒绝 H₀ 的取值集合；第五步，比较并写出情境化结论。

Hypothesis testing is the soul of the statistics option, and nearly every block ends with it. The standard five-step procedure is: first, state the null hypothesis H₀ and the alternative H₁, for example H₀: λ = 3 against H₁: λ > 3 for a Poisson mean; second, fix the significance level (usually 5% or 1%) and state whether the test is one- or two-tailed; third, compute the test statistic, such as the observed count X; fourth, find the critical region, the set of values that leads to rejecting H₀; fifth, compare and write the conclusion in context.

以泊松均值检验为例：某服务站平均每分钟接待 3 位顾客，怀疑改造后客流增加。设 X ~ Po(λ)，H₀: λ = 3, H₁: λ > 3，取 5% 显著性水平。查表找最小的 k 使 P(X ≥ k) ≤ 0.05。由累计表 P(X ≤ 6) = 0.9665，故 P(X ≥ 7) = 0.0335 ≤ 0.05，临界区域为 X ≥ 7。若改造后某分钟观测到 8 位顾客，则 8 落入临界区域，拒绝 H₀，结论为“有证据表明客流显著增加”。

Take a Poisson mean test as the example: a service desk averages 3 customers per minute, and after a renovation we suspect the flow has increased. Let X ~ Po(λ) with H₀: λ = 3 and H₁: λ > 3 at the 5% significance level. From the tables, find the smallest k with P(X ≥ k) ≤ 0.05. Since P(X ≤ 6) = 0.9665, we have P(X ≥ 7) = 0.0335 ≤ 0.05, so the critical region is X ≥ 7. If 8 customers are observed in one minute after the renovation, 8 lies in the critical region, H₀ is rejected, and the conclusion is “there is evidence that the customer flow has increased significantly”.

写结论时要注意两点：一是必须回到问题情境，不能只写“拒绝原假设”；二是措辞要区分“证据”与“证明”，统计检验只能提供证据，不能证明事实。双尾检验的临界区域在分布两端各占 α/2 的概率，找临界值时上下尾都要查，切勿只查一侧。

Two points matter when writing conclusions: first, always return to the context of the question rather than merely writing “reject the null hypothesis”; second, distinguish “evidence” from “proof”, since a statistical test provides evidence, never proof. In a two-tailed test the critical region is split between the two tails with probability α/2 each, so both tails must be examined when finding critical values; never check only one side.

十、相关系数的显著性检验：从样本 r 判断总体相关 | Testing the Significance of a Correlation Coefficient: From the Sample r to a Population Conclusion

样本相关系数 r 不为 0 并不能直接说明总体相关，因为抽样波动也会产生非零的 r。显著性检验的做法是：设 H₀: ρ = 0（总体相关系数为 0，即无线性相关），H₁: ρ ≠ 0（双尾）或 ρ > 0 / ρ < 0（单尾），然后查相关系数临界值表，表中给出不同样本量 n 与显著性水平下的临界 r 值。

A sample correlation coefficient r different from 0 does not by itself establish population correlation, because sampling variation also produces non-zero r values. The significance test proceeds as follows: set H₀: ρ = 0 (no linear correlation in the population) against H₁: ρ ≠ 0 (two-tailed) or ρ > 0 / ρ < 0 (one-tailed), then read the critical value table for correlation coefficients, which lists critical r values for different sample sizes n and significance levels.

判定规则很直接：若 |r| 大于临界值，则拒绝 H₀，认为存在统计上显著的线性相关；否则没有足够证据。例如 n = 10 时，5% 双尾检验的临界 r 约为 0.632。若算得 r = 0.71，则 0.71 > 0.632，拒绝 H₀，结论为“有证据表明两变量存在正的线性相关”。若 r = 0.5，则不能拒绝 H₀，只能说样本证据不足以支持相关结论。

The decision rule is straightforward: if |r| exceeds the critical value, reject H₀ and conclude that there is statistically significant linear correlation; otherwise there is insufficient evidence. For example, with n = 10 the critical r for a two-tailed 5% test is about 0.632. If you compute r = 0.71, then 0.71 > 0.632, so H₀ is rejected and the conclusion is “there is evidence of a positive linear correlation between the two variables”. If r = 0.5, H₀ cannot be rejected; the sample evidence is insufficient to support a correlation conclusion.

这一检验在数据科学中同样重要：筛选特征时，先对每个候选变量与目标变量做相关显著性检验，能快速排除“纯属巧合”的相关系数。A-Level 阶段只需会查表比较，但理解其思想有助于衔接大学统计课程。注意：显著性相关不代表因果，题目若追问解释，要回答“可能存在共同原因或第三变量影响”。

This test matters in data science too: when screening features, running a significance test between each candidate variable and the target quickly eliminates correlation coefficients that are pure coincidence. At A-Level you only need to read the table and compare, but understanding the idea bridges smoothly into university statistics. Remember that significant correlation does not imply causation; if a question asks for interpretation, answer that a common cause or a third variable may be at work.

十一、真题常见陷阱：自由度、合并类别与尾概率方向 | Common Exam Traps: Degrees of Freedom, Merging Classes and Tail Directions

统计模块失分往往不是不会算，而是踩了固定的坑。第一个高频陷阱是自由度。拟合优度检验中忘记减去估计参数的个数，或列联表中把 (r−1)(c−1) 错算成 rc − 1，都会导致查错临界值。建议把两类检验的自由度公式单独抄在笔记首页，考前默写一遍。

Lost marks in the statistics module usually come from falling into fixed traps rather than not knowing how to calculate. The first frequent trap is degrees of freedom. Forgetting to subtract the number of estimated parameters in a goodness-of-fit test, or computing rc − 1 instead of (r−1)(c−1) for a contingency table, sends you to the wrong critical value. Write the two degree-of-freedom formulas separately on the first page of your notes and recite them before every exam.

第二个陷阱是合并类别。期望频数小于 5 的类别必须合并，合并后要重新计算合并类的期望频数，并相应减少类别数 k（自由度随之变化）。有些学生只合并观测频数小的类而忘记调整期望值，导致 X² 计算错误。第三个陷阱是尾概率方向：单尾与双尾的临界值不同，题目写“是否与……不同”是双尾，“是否大于……”是单尾，读题时先圈出方向词。

The second trap is merging classes. Classes with expected frequencies below 5 must be merged; after merging you must recompute the expected frequency of the combined class and reduce the number of classes k accordingly (so the degrees of freedom change). Some students merge only the classes with small observed frequencies and forget to adjust the expectations, corrupting the X² calculation. The third trap is tail direction: one-tailed and two-tailed tests have different critical values. A question asking “is it different from …” is two-tailed, while “is it greater than …” is one-tailed; circle the direction word when reading the question.

第四个陷阱是连续性修正缺失。凡是用正态分布近似离散分布（二项或泊松），必须带 0.5 修正；批卷时这一点几乎必扣。第五个陷阱是回归方向：用 y on x 的回归线反推 x，必须换线。第六个陷阱是结论措辞：忘记情境、把“无证据”写成“证明无关”、把“显著”理解成“重要”，都属于失分点。考前把这份陷阱清单过一遍，比多刷一套题更有效。

The fourth trap is a missing continuity correction. Whenever a normal distribution approximates a discrete distribution (binomial or Poisson), the 0.5 correction is mandatory; examiners almost always deduct for omitting it. The fifth trap is regression direction: to predict x from y you must switch to the x-on-y line rather than inverting the y-on-x line. The sixth trap is conclusion wording: forgetting the context, writing “proved unrelated” instead of “no evidence”, or confusing “significant” with “important”. Reviewing this trap list before the exam is more effective than one more past paper.

Summary | 总结

这篇指南覆盖了牛津AQA国际A-Level进阶数学 9665 统计模块的五大核心板块：离散随机变量的期望与方差、泊松分布及其近似、卡方拟合优度与列联表检验、相关与回归、假设检验五步法。每个板块的适用条件、公式与常见陷阱都已逐条展开，并配有完整的计算示例。

This guide has covered the five core blocks of the OxfordAQA International A-Level Further Mathematics 9665 statistics option: expectation and variance of discrete random variables, the Poisson distribution and its approximations, chi-squared goodness-of-fit and contingency tests, correlation and regression, and the five-step hypothesis test. For every block the conditions of use, formulas and common traps have been laid out one by one, with complete worked examples.

备考建议：先把每一类检验的步骤写成固定模板，再带着模板刷 topic test 和真题；做完后对照评分标准，重点检查自由度、连续性修正、尾方向与结论措辞。统计模块是进阶数学中最容易拿满分的部分，只要条件判断准确、步骤完整规范，A* 的统计分数就能稳稳收入囊中。

Study advice: write each type of test as a fixed template first, then work through topic tests and past papers with the template at hand; afterwards compare with the mark scheme, checking degrees of freedom, the continuity correction, tail direction and conclusion wording in particular. The statistics option is the easiest block in Further Maths to score full marks on. With accurate condition checks and complete, standardised steps, the statistics marks needed for an A* are safely within reach.

更多咨询请联系16621398022（同微信）