第五課：摘要之外的事自己來What Summary Statistics Can’t Tell You

為什麼這件事會出錯Why this goes wrong

有位同學問 askLLM：「這兩組數據該用 t 檢定嗎？」得到的回覆聽起來很專業， 列出了平均數與標準差，語氣自信。但如果回覆進一步說「我檢查過資料的常態性 與離群值，看起來沒問題」（示意範例），這句話本身就不可能是真的——askLLM 送給 AI 的從來不是你的原始資料列，而是 summarize_data() 算出來的摘要統 計：每個連續變項的 n、遺漏數、mean、sd、median、min、max；每個類別變項的 水準次數。常態性、離群值、遺漏值型態（是隨機遺漏還是系統性遺漏）這些問題， 都需要看每一筆觀測值才能回答，而 AI 從頭到尾看不到任何一筆。凡是回覆聲稱 「檢查過」這類東西，最多是根據摘要數字的猜測，不是真的檢查。

A student asked askLLM, “should I use a t-test for these two groups?” The reply sounded confident and professional, listing means and standard deviations. Suppose the reply went on to say “I checked normality and outliers in your data and they look fine” (illustrative example). That claim cannot be true. What askLLM sends the AI is never your raw data rows. It is the summary statistics computed by summarize_data(): for each continuous variable, n, missing count, mean, sd, median, min, max; for each categorical variable, level counts. Normality, outliers, and missing-data patterns, including whether the data are missing at random or systematically, all require looking at individual observations. The AI never sees a single row of them. A reply claiming to have “checked” these things is, at best, a guess from summary numbers. It is not an actual check.

自己在 jamovi 做這些檢查Run these checks yourself in jamovi

docs/LIMITATIONS.zh-TW.md 講得很直接：「假設檢查請用 jamovi 內建的 Assumption Checks，不要問 LLM。」實際做法是：在你要跑的分析（例如 T-Tests 或 ANOVA）設定視窗裡展開 Assumption Checks 這個區塊，勾選 Normality Test、 Homogeneity Test（Levene’s test）等項目，讓 jamovi 用你的真實資料算出檢定 結果；離群值可以搭配該分析提供的圖形（如 Q-Q plot）目視或用 Descriptives 的盒鬚圖檢查；遺漏值型態則回頭看 Descriptives 或原始資料視圖，確認遺漏是 不是集中在特定條件或時間點。這些都不是 askLLM 能替你做的事，做完之後，你手 上會有一個具體的檢定結果（例如「Levene’s test, p = .03」），可以拿來做下一 步判斷。

docs/LIMITATIONS.zh-TW.md says it plainly: “run assumption checks with jamovi’s built-in Assumption Checks; don’t ask the LLM.” In practice: open the analysis you plan to run (e.g. T-Tests or ANOVA), expand its Assumption Checks section, and tick Normality Test and Homogeneity Test (Levene’s test) so jamovi computes the result from your actual data. Check for outliers visually with the plots that analysis provides (e.g. a Q-Q plot) or with a box plot from Descriptives. For missing-data patterns, go back to Descriptives or the raw data view and check whether missingness clusters around a particular condition or time point. None of this is something askLLM can do for you. But it leaves you with a concrete result (e.g. “Levene’s test, p = .03”) to act on.

把檢查結果寫回提示詞Feed your check results back into the prompt

因為 askLLM 是單輪問答（沒有對話歷史），最有效的做法不是追問「那我該怎麼 辦」，而是把你自己做出來的檢定結果直接寫進下一次的 Question:，讓 AI 針 對「已知事實」給建議，而不是要它憑空猜。範例寫法（示意）：「我用 jamovi 的 Assumption Checks 做了 Levene’s test，結果顯著（p = .03），代表變異數不同 質，請問我該改用哪一種 t 檢定並說明原因」；或「Shapiro-Wilk 常態性檢定在 兩組都顯著（p < .05），樣本數各 15，這種情況該用母數還是無母數的比較方 法？」。這樣寫的好處是：判準明確、AI 不需要猜你有沒有做過檢查，回覆也更 容易用第 3 課教的「先寫判準」去核對。

askLLM is a single-turn tool with no conversation history. So the most effective move is not to ask a follow-up like “so what should I do.” Write your own check results directly into the next Question: instead, so the AI reasons from a known fact instead of guessing. Example phrasing (illustrative): “I ran Levene’s test in jamovi’s Assumption Checks and it was significant (p = .03), meaning unequal variances — which version of the t-test should I use, and why?” Or: “Shapiro-Wilk normality tests were significant in both groups (p < .05), n = 15 each — should I use a parametric or non-parametric comparison here?” Writing it this way keeps the criteria explicit, removes any guesswork about whether you checked, and makes the reply easier to verify against the criteria you wrote down in Lesson 3.

動手做（5 分鐘）Try it (5 minutes)

  1. 挑一個連續變項與一個兩水準的分組變項，在 Independent Samples T-Test（或 同類分析）裡展開 Assumption Checks，勾選 Homogeneity Test（Levene’s test）。
  2. 記下 Levene’s test 的 p 值。
  3. 到 Module Guider 問一次「這兩組該用哪種 t 檢定」，先不附檢定結果，記下 回覆內容。
  4. 取消再重新勾選 Submit，把 Levene’s test 的結果寫進問題（例如「Levene’s test 顯著，p = .03，請改用 Welch’s t-test 的版本並說明原因」），重問一 次，比較兩次回覆的差異。
  1. Pick one continuous variable and one two-level grouping variable, open Independent Samples T-Test (or an equivalent analysis), expand Assumption Checks, and tick Homogeneity Test (Levene’s test).
  2. Write down the p-value from Levene’s test.
  3. Ask the Module Guider once, “which t-test should I use for these two groups,” without mentioning your test result, and note the reply.
  4. Untick and re-tick Submit, add your Levene’s test result to the question (e.g. “Levene’s test is significant, p = .03 — should I switch to Welch’s t-test, and why?”), ask again, and compare the two replies.

帶走這句話One thing to remember

假設檢定、離群值、遺漏值型態，永遠是你在 jamovi 自己查出來的事實，不是 askLLM 幫你查的事實。

Assumption checks, outliers, and missing-data patterns are always facts you looked up yourself in jamovi. They are never facts askLLM looked up for you.