📚 Year 13 Edexcel Statistics: Experimental & Practical Assessment Essentials | 高三 Edexcel 统计：实验与实践考核要点
In Year 13 Edexcel Statistics, the ability to design robust experiments and evaluate practical investigations is a critical assessment focus. Whether you are tackling questions on clinical trials, agricultural studies or quality control, examiners expect you to demonstrate a thorough understanding of experimental principles, data collection methods and the mitigation of bias. This article distils the essential concepts, common pitfalls and effective exam strategies, and extends them with the formulas, structures and evaluation language that turn a competent answer into a top-band one.
在高三 Edexcel 统计学中，设计严谨的实验并评估实践调查是关键考核点。无论是解答临床试验、农业研究还是质量控制问题，考官都期望你展示对实验原则、数据收集方法以及减少偏差的透彻理解。本文提炼了核心概念、常见陷阱和高效的考试策略，并进一步补充公式、答题结构与评估用语，帮助你把"合格答案"提升为"高分答案"。
1. Core Principles of Experimental Design | 实验设计核心原则
Sound experimental design rests on three pillars: randomisation, replication and control. Randomisation ensures that treatments are assigned without bias, balancing out unknown confounding variables. Replication involves repeating the experiment on multiple independent units to estimate variability and increase precision. Control, often achieved through control groups, provides a baseline against which treatment effects are measured. In Edexcel exams, you should be able to explain why each principle is essential and recognise flaws when they are absent.
稳健的实验设计建立在三大支柱之上：随机化、重复和对照。随机化确保处理分配无偏，平衡未知的混杂变量。重复指在多个独立单元上重复实验，以估计变异并提高精确度。对照（通常通过对照组实现）为衡量处理效应提供了基线。在 Edexcel 考试中，你应能解释每个原则的重要性，并在它们缺失时识别出缺陷。
Before you can evaluate any design, you must be able to name its structural components precisely: the experimental units (the individual plots, patients or components being studied), the treatments (the conditions deliberately imposed), the response variable (what is measured) and the extraneous variables (everything else that could influence the response). A well-written answer always identifies these four elements before commenting on quality.
在评估任何设计之前，你必须能准确说出它的结构成分：实验单元（被研究的个体地块、病人或部件）、处理（人为施加的条件）、响应变量（被测量的量）以及外部变量（其他可能影响响应的一切因素）。高分答案总会在评价设计优劣之前，先明确这四个要素。
You should also be able to distinguish an experiment from an observational study. In an experiment the researcher controls the allocation of treatments, so a statistically significant difference can be attributed to a causal effect, provided randomisation was genuine. In an observational study the researcher merely records what happens, so association can be established but causation cannot, because lurking variables may explain the pattern.
你还应能区分实验研究与观察性研究。在实验中，研究者控制处理的分配，因此只要随机化真实可靠，显著差异就能被归因于因果效应。在观察性研究中，研究者只是记录发生的情况，因此只能确立相关性，无法确立因果关系，因为潜藏的第三变量可能解释这一模式。
2. Control Groups and Placebos | 对照组与安慰剂
A control group is a standard of comparison that receives no treatment or a standard treatment. In medical experiments, a placebo is a harmless, inert treatment that looks identical to the real one. Because patients often improve simply through expectation, the placebo effect, the control group allows the genuine treatment effect to be isolated. If the treatment group improves by more than the placebo group, the excess is evidence of a real pharmacological effect.
对照组是接受无处理或标准处理的比较基准。在医学实验中，安慰剂是与真实处理外观完全相同、但无活性成分的无害处理。由于病人常因心理预期而好转（即安慰剂效应），对照组才能把真实的处理效应对离出来。若处理组的改善超过安慰剂组，超出的部分就是真实药效的证据。
Two refinements are worth memorising. First, the nocebo effect shows that negative expectations can produce adverse symptoms, which is another reason why a control arm is essential. Second, the double-dummy technique is used when two treatments have different forms: each group receives one real treatment and one matching placebo, so no participant can tell which arm they are in.
有两个细化点值得牢记。第一，反安慰剂效应表明负面预期也会产生不良症状，这同样是必须设置对照组的理由。第二，当两种处理剂型不同时，会采用双模拟技术：每组各接受一种真实处理和一种与之匹配的安慰剂，使受试者无法判断自己属于哪一组。
Ethically, a control group may raise concerns. Where a proven treatment already exists, a placebo control is often unacceptable, and the correct design compares the new treatment against the current standard rather than against nothing. Mentioning this trade-off between statistical purity and ethical constraint is a reliable way to earn evaluation marks.
在伦理层面，对照组可能引发争议。当已有确切有效的治疗方案时，设置安慰剂对照通常不可接受，正确做法是将新疗法与现行标准疗法比较，而不是与"什么都不做"比较。在答案中点出统计严谨性与伦理约束之间的这一权衡，是稳拿评估分的有效方法。
3. Randomisation Techniques and Allocation | 随机化技术与处理分配
Randomisation is not the same as sampling. Random sampling concerns how units are selected from a population; random allocation concerns how selected units are assigned to treatments. Exam questions frequently conflate the two, so always state which one you are describing and why it matters. Random allocation is what licenses causal inference; random sampling is what licenses generalisation to the population.
随机化不等同于抽样。随机抽样关注如何从总体中选取单元；随机分配关注如何把已选单元分配到各处理组。考题常把两者混为一谈，因此务必说明你在描述哪一者以及为何重要。随机分配是因果推断的依据；随机抽样则是推广到总体的依据。
The practical mechanics are simple and worth quoting. Number the experimental units from 1 to n, generate random digits from a random number table or a calculator function, and allocate the first n₁ units encountered to treatment A and the remainder to treatment B. If a designed balance is required, draw without replacement until each group reaches its required size.
操作机制很简单，值得在答案中写出：把实验单元编号为 1 至 n，用随机数表或计算器生成随机数字，按出现顺序把前 n₁ 个单元分到处理 A，其余分到处理 B。若要求各组数量精确平衡，则不放回抽取，直到每组达到规定数量为止。
| Method | 方法 | Procedure | 做法 | Strength | 优点 | Weakness | 缺点 |
| Simple random 简单随机 | Every unit equally likely to be chosen 每个单元被选中的机会均等 | Unbiased, easy to analyse 无偏，易分析 | Needs a full list; may miss subgroups 需完整名册；可能漏掉子群 |
| Stratified 分层 | Split into strata, sample within each 先分层，再在各层内抽样 | Guarantees representation, reduces variance 保证代表性，降低方差 | Requires known stratum sizes 需已知各层规模 |
| Systematic 系统 | Every k-th unit from a random start 随机起点后每隔 k 个取一个 | Quick, spreads across the list 快捷，覆盖整份名单 | Dangerous if the list has periodicity 名单若含周期性则风险大 |
| Cluster 整群 | Randomly select whole groups 随机选取整个群体 | Cheap when travel is costly 差旅成本高时省钱 | Higher sampling error 抽样误差更大 |
4. Replication, Sample Size and Precision | 重复、样本量与精确度
Replication serves two purposes: it provides an estimate of the natural variability of the response, and it increases the precision of any estimate of the treatment effect. Precision here means the reciprocal of the variance of the estimator, so precision increases as the standard error falls. The standard error of a sample mean is the key quantity.
重复有两个作用：提供响应自然变异的估计，并提高处理效应估计的精确度。这里的精确度指估计量方差的倒数，因此标准误越小，精确度越高。样本均值的标准误是核心量。
SE(x̄) = σ ÷ √n ≈ s ÷ √n
Because the standard error is inversely proportional to the square root of n, quadrupling the sample size only halves the standard error. This is the single most quotable fact in practical-design questions, and it explains why huge samples give diminishing returns and why the cost of an investigation must be weighed against the extra precision obtained.
由于标准误与 n 的平方根成反比，样本量变为四倍只能把标准误减半。这是实践设计题中最值得引用的一条结论，它解释了为什么超大样本的收益递减，也解释了为什么必须把调查成本与所获得的额外精确度放在一起权衡。
a proportion, n = z²p(1 − p) ÷ E², and if p is unknown the conservative choice p = 0.5 maximises the product p(1 − p) and therefore gives the largest, safest sample size.
当题目给定目标误差界限 E 时，所需样本量可直接计算。估计均值时，n = (zσ ÷ E)²。估计比例时，n = z²p(1 − p) ÷ E²；若 p 未知，保守做法是取 p = 0.5，因为此时 p(1 − p) 最大，得到的样本量最大、最稳妥。
Finally, remember that replication must be independent. Measuring the same plant five times is not replication; it is repeated measurement and it inflates the apparent sample size without adding any new information. Examiners award marks specifically for spotting this pseudo-replication.
最后，切记重复必须相互独立。对同一株植物测量五次不是重复，而是重复测量，它虚增了表观样本量却未增加任何新信息。考官会专门为识别这种"伪重复"给分。
5. Blocking and Matched Pairs | 区组设计与配对设计
Blocking is a deliberate design device that groups experimental units which are similar in some nuisance variable, such as soil type, age or machine operator. Randomisation then takes place within each block. Because the nuisance variation is removed from the residual, the standard error falls and the test becomes more sensitive to a real treatment effect. This is sometimes called local control or the reduction of variance.
区组设计是一种有意的设计手段，把在某些干扰变量（如土壤类型、年龄或机器操作员）上相似的实验单元归为一组，然后在每个区组内部进行随机化。由于干扰变异被从残差中剔除，标准误会下降，检验对真实处理效应也更敏感。这有时被称为局部控制或方差缩减。
Matched pairs is the extreme case where each block contains exactly two units, ideally twins, before-and-after measurements on the same subject, or two halves of the same sample. The paired t-test then analyses the differences rather than the raw values, which removes between-subject variability entirely.
配对设计是区组设计的极端情形，每个区组恰好包含两个单元——理想情况下是同卵双胞胎、同一受试者的前后测量，或同一样品的两半。此时配对 t 检验分析的是差值而非原始数值，从而彻底消除个体间变异。
t = d̄ ÷ (sd ÷ √n), df = n − 1
nces and uses sₚ² = ((n₁ − 1)s₁² + (n₂ − 1)s₂²) ÷ (n₁ + n₂ − 2) with df = n₁ + n₂ − 2. State the assumption explicitly: if the variances look very different, or the data are strongly skewed with small samples, prefer the unpooled form or a non-parametric alternative such as the Mann-Whitney test.
若无法配对、两组相互独立，则使用双样本 t 检验。合并方差形式假定两总体方差相等，使用 sₚ² = ((n₁ − 1)s₁² + (n₂ − 1)s₂²) ÷ (n₁ + n₂ − 2)，自由度 df = n₁ + n₂ − 2。务必明确写出该假设：若两方差差异明显，或数据严重偏斜且样本量小，应改用非合并形式，或使用 Mann-Whitney 检验等非参数替代方法。
6. Blinding: Single, Double and Triple | 盲法：单盲、双盲与三盲
Blinding conceals treatment identity from the people involved, and it attacks bias at its source rather than trying to correct it afterwards. In a single-blind trial the participants do not know which treatment they receive, which controls the placebo effect and reduces reporting bias. In a double-blind trial the assessors or clinicians are also unaware, which removes observer bias and the unconscious tendency to record favourable outcomes for the new treatment.
盲法向相关人员隐藏处理身份，它从源头上打击偏差，而不是事后补救。在单盲试验中，受试者不知道自己接受的是哪种处理，这控制了安慰剂效应并减少报告偏差。在双盲试验中，评估者或临床医生同样不知情，从而消除观察者偏差，以及无意识地为新疗法记录有利结果倾向。
| Design | 设计 | Who is blinded | 谁被设盲 | Bias controlled | 控制的偏差 |
| Open label 开放标签 | Nobody 无人设盲 | None; only suitable when blinding is impossible 无；仅在无法设盲时使用 |
| Single blind 单盲 | Participants only 仅受试者 | Placebo effect, reporting bias 安慰剂效应、报告偏差 |
| Double blind 双盲 | Participants and assessors 受试者与评估者 | Placebo effect and observer bias 安慰剂效应与观察者偏差 |
| Triple blind 三盲 | Also the data analysts 还包括数据分析者 | Also analysis-stage bias 还包括分析阶段偏差 |
A strong answer always adds one sentence of limitation: blinding may be impossible or unethical when the treatment is a surgical procedure, a diet change or a physiotherapy programme, and in such cases the design must compensate by using objective outcome measures, automated measurement devices and blinded assessment of recorded data.
高分答案总会补一句局限说明：当处理是外科手术、饮食改变或物理治疗时，设盲可能根本无法实现或不合伦理，此时设计必须通过客观结局指标、自动化测量设备和盲法评估记录数据来补偿。
7. Confounding Variables and Sources of Bias | 混杂变量与偏差来源
A confounding variable is associated with both the treatment and the response, so its effect cannot be separated from the treatment effect. If fitter volunteers are more likely to be assigned to the exercise programme, fitness confounds the comparison and the apparent benefit is exaggerated. Randomisation addresses confounding on average, blocking addresses it by design, and matching addresses it by construction.
混杂变量同时与处理和响应相关，因此其效应无法与处理效应分离开来。如果体质更好的志愿者更可能被分配到运动组，体质就混杂了比较，表观收益被夸大。随机化在平均意义上解决混杂，区组设计通过设计解决混杂，配对则通过构造方式解决混杂。
| Bias | 偏差 | Cause | 成因 | Mitigation | 缓解措施 |
| Selection bias 选择偏差 | Non-random allocation or self-selection 非随机分配或自愿加入 | Random allocation; concealed allocation 随机分配；分配隐藏 |
| Measurement bias 测量偏差 | Faulty or uncalibrated instruments 仪器失准或未校准 | Calibration; standard protocols 校准；标准化流程 |
| Observer bias 观察者偏差 | Expectations influence recording 预期影响记录 | Blinding; automated recording 盲法；自动记录 |
| Non-response bias 无应答偏差 | Certain groups fail to reply 某些群体不予回应 | Follow-up; report response rate 追踪；报告应答率 |
| Recall bias 回忆偏差 | Memory is selective or hazy 记忆具选择性或不清晰 | Contemporaneous records; diaries 同期记录；日志 |
When asked to identify a weakness in a given study, use the four-step structure that examiners reward: name the bias, explain the mechanism that creates it, state the direction of its likely effect on the conclusion, and propose a specific practical remedy. Vague answers such as "the sample was too small" earn little; precise answers such as "attrition was 30% in the treatment arm versus 8% in the control arm, which probably exaggerates the effect because dropouts tend to be those who did not improve" earn full marks.
当被要求指出给定研究的缺陷时，请使用考官认可的四步结构：指出偏差名称、解释其产生机制、说明它对结论可能影响的方向，并提出具体的可行补救措施。像"样本太小"这样的笼统答案得分很低；而"处理组失访率为 30%，对照组仅为 8%，这很可能夸大效应，因为退出者往往是未见改善的人"这样的精确表述则能拿满分。
8. Data Collection Methods and Measurement Quality | 数据收集方法与测量质量
Primary data are collected first-hand by the investigator for the specific question at hand; secondary data already exist and were gathered for another purpose. Primary data give control over definitions and timing but cost time and money; secondary data are cheap and often extensive but may use different definitions, be out of date, or carry unknown errors.
一手数据由研究者针对当前问题亲自收集；二手数据已经存在，原本为其他目的而采集。一手数据让研究者掌控定义与时间，但耗费时间与金钱；二手数据廉价且往往规模庞大，但可能定义不一致、时效过期，或含有未知误差。
When designing a questionnaire or interview schedule, avoid leading questions, double-barrelled questions and overlapping response categories, and pilot the instrument on a small group. When designing an observation or sensor-based study, define the measurement protocol in advance: who measures, with what instrument, at what time of day, and to how many decimal places.
设计问卷或访谈提纲时，要避免诱导性问题、双重问题以及相互重叠的选项类别，并先在小范围做预试。设计观察法或传感器研究时，需事先定义测量规程：由谁测量、用何仪器、在一天中的哪个时段、记录到小数点后几位。
Distinguish accuracy from precision, and random from systematic error. Accuracy is closeness to the true value; precision is closeness of repeated readings to one another. Random error scatters readings symmetrically about the true value and can be reduced by taking the mean of many readings. Systematic error shifts all readings in the same direction, so averaging does not remove it; it must be eliminated by calibration, by correcting for a zero error, or by using a different instrument.
请区分准确度与精确度，以及随机误差与系统误差。准确度是接近真值的程度；精确度是重复读数彼此接近的程度。随机误差使读数围绕真值对称散布，可通过多次读数取均值来减小。系统误差使所有读数朝同一方向偏移，取平均无法消除，必须通过校准、修正零点误差或更换仪器来消除。
