起源故事 · Fisher 让因素"一起上场"
Origin Story · Fisher Brings Factors On Stage Together
单因素 ANOVA 一次只问一个因素。但 Fisher 早就看出真实试验里因素从不孤立 —— 肥料和品种、温度和时间,往往合在一起才见真章。
他把试验设计成因子排列(factorial),让两个因素的所有水平组合都出现,于是不仅能算出各自的主效应,还能算出交互作用这一额外信息。
这正是 DOE 的精髓:用同一批试验,把"谁起作用"和"谁和谁配合起作用"一次性看清。
One-way ANOVA asks about one factor at a time. Fisher saw early on that real experiments never isolate factors — fertilizer × variety, temperature × time, the truth only shows up when they meet.
He redesigned the experiment as a factorial, so every combination of factor levels appears, and from a single dataset you can estimate both main effects and the extra piece — the interaction.
That is the heart of DOE: one batch of runs tells you who matters and who has to be tuned alongside whom.
1 交互作用图:折线平行还是交叉? Interaction Plot: Parallel or Crossed?
平行 · 无交互2 2×2 单元格均值:折线就是从这里连出来的 2×2 Cell Means: The Lines Are Just These Four Dots Connected
四个组合的平均得率。横向看出 B 的效应,纵向看出 A 的效应,对角不一致就是 交互。折线图不过是把这四个格子连起来。 Mean yield for the four combinations. Read across rows for B's effect, down columns for A's effect; a diagonal mismatch is the interaction. The line plot is nothing more than these four cells connected.
3 现实里的双因素 ANOVA Two-Way ANOVA in the Real World
DOE 多因素试验:温度 × 时间对得率,一次试验同时给出两个主效应和交互。
Multi-factor DOE: temperature × time on yield — one designed experiment delivers both main effects plus the interaction.
工艺参数交互:转速和进给量对表面粗糙度,最佳组合常藏在交互里。
Machining parameters: spindle speed × feed rate on surface roughness — the best window almost always lives inside the interaction.
配方优化:原料 A 含量 × 固化温度对强度,交互决定要不要"配套调"。
Formulation tuning: resin loading × cure temperature on tensile strength — the interaction decides whether they must be tuned as a pair.
机台 × 班次:产出按机台和班次双向分组,识别"某机台某班次"特别异常。
Machine × shift: output blocked by machine and shift — flags the "this machine on that shift" cells that are out of family.
一句话In One Line
单因素 ANOVA 像分别问"温度有用吗""时间有用吗",双因素更进一步问"它俩要不要一起调"。
一旦交互显著,就不能再孤立地谈某个因素的"最优值" —— 温度的最佳设定取决于时间设多少,反之亦然。
这就是为什么 DOE 反对"一次只动一个因素":那样永远看不到交互,会错过真正的最优组合。看懂折线交不交叉,你就抓住了双因素分析的全部直觉。
One-way ANOVA asks "does temperature matter?" and "does time matter?" separately. Two-way ANOVA goes one level deeper: "do they have to be tuned together?"
Once the interaction is significant, you can no longer talk about a single factor's "optimum" in isolation — the best temperature depends on what time you've chosen, and vice versa.
This is exactly why DOE rejects the "change one factor at a time" approach: OFAT can never see the interaction and will miss the true sweet spot. Once you can tell whether the lines cross, you own the whole intuition of two-way analysis.
常见误用Common Mistakes
交互显著却只解读主效应。交互显著时,主效应要分水平来看,不能笼统下结论。
Reading only the main effects when the interaction is significant. When A×B is significant, main effects must be interpreted level by level — sweeping statements are misleading.
一次只改一个因素做试验。用因子设计让因素同时变化,才能估出交互。
Running OFAT (one factor at a time). Use a factorial design so factors vary together — only then can the interaction be estimated.
每格只测一次就谈交互。无重复无法分离交互与误差,需要重复或可加性假设。
Claiming an interaction without replication. With one observation per cell you cannot separate interaction from error — add replicates or assume additivity.