{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 课堂练习：帕尔默群岛三种企鹅的形态数据（参考答案）\n",
    "\n",
    "**仅供教师使用，课前不要发给学生。** 学生版是同一目录下的 `exercise.py` / `exercise.ipynb`，\n",
    "题目文字与本文件逐字相同（`../tools/check_course.py` 会核对）。\n",
    "\n",
    "数据：`../data/penguins.csv`，即 palmerpenguins v0.1.0 中的 penguins 表（CC0），\n",
    "原始研究为 Gorman、Williams 和 Fraser（2014，PLoS ONE 9: e90081）。\n",
    "练习的结论以本课程所依据的论文初稿为准：每一问的数值都应与论文正文、表 1 或图 1 图注一致。\n",
    "\n",
    "练习分三部分，分别对应单元第 1–3 周的上机课，每部分约 80 分钟：\n",
    "\n",
    "| 部分 | 内容 | 对应周次 |\n",
    "|---|---|---|\n",
    "| A | 读数据、缺测值、样本构成、描述统计、分组作图 | 第 1 周 |\n",
    "| B | 两性体重差异：正态性检查、Welch t 检验、置信区间、Holm 校正、Mann–Whitney U | 第 2 周 |\n",
    "| C | 回归：物种内与合并拟合、辛普森悖论、把物种放进模型 | 第 3 周 |\n",
    "\n",
    "运行方式（任选其一）：\n",
    "\n",
    "- `uv run solution.py`：按文件头里的版本自动装好依赖后运行；\n",
    "- `pip install pandas matplotlib scipy`，再 `python solution.py`；\n",
    "- 在 Jupyter 中打开 `solution.ipynb`，或在 VS Code / PyCharm / Spyder 中按 `# %%` 分块逐块运行。\n",
    "\n",
    "作为脚本运行时，图保存在 `solution_output/`，所有答案数值写入 `solution_output/answers.csv`。\n",
    "\n",
    "每题代码后面的“自查”格与学生版逐字相同，由 `../tools/build_exercise.py` 生成：正确答案只以哈希值写在格子里。\n",
    "参考答案运行时会执行全部自查，有一项不通过就报错退出，所以参考答案能跑完就说明自查格与答案一致。"
   ],
   "id": "cell-00"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "import sys\n",
    "from pathlib import Path\n",
    "\n",
    "import matplotlib\n",
    "\n",
    "if \"ipykernel\" not in sys.modules:  # 作为脚本运行时不弹窗口，只存图\n",
    "    matplotlib.use(\"Agg\")\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "from scipy import stats\n",
    "\n",
    "try:\n",
    "    HERE = Path(__file__).resolve().parent\n",
    "except NameError:  # Jupyter 中没有 __file__\n",
    "    HERE = Path.cwd()\n",
    "\n",
    "\n",
    "def find_data(name=\"penguins.csv\"):\n",
    "    \"\"\"依次在 ../data/、当前目录、./data/ 中找数据文件。\"\"\"\n",
    "    for p in (HERE.parent / \"data\" / name, HERE / name, HERE / \"data\" / name):\n",
    "        if p.exists():\n",
    "            return p\n",
    "    raise FileNotFoundError(f\"找不到 {name}，请把它放在本文件旁边或 ../data/ 中\")\n",
    "\n",
    "\n",
    "OUT = HERE / \"solution_output\"\n",
    "OUT.mkdir(exist_ok=True)\n",
    "\n",
    "SPECIES = [\"Adelie\", \"Chinstrap\", \"Gentoo\"]\n",
    "MORPH = [\"bill_length_mm\", \"bill_depth_mm\", \"flipper_length_mm\", \"body_mass_g\"]\n",
    "COLORS = {\"Adelie\": \"#D55E00\", \"Chinstrap\": \"#0072B2\", \"Gentoo\": \"#009E73\"}  # 与论文图 1 相同\n",
    "LABEL = {\"Adelie\": \"Adélie\", \"Chinstrap\": \"Chinstrap\", \"Gentoo\": \"Gentoo\"}\n",
    "\n",
    "plt.rcParams.update({\n",
    "    \"figure.dpi\": 110, \"savefig.dpi\": 200, \"font.size\": 10,\n",
    "    \"axes.spines.top\": False, \"axes.spines.right\": False,\n",
    "})\n",
    "\n",
    "ANSWERS = {}  # 键 -> 数值（或数值列表）；最后写入 answers.csv，供 ../tools/check_course.py 与论文核对\n",
    "\n",
    "\n",
    "def record(key, *values):\n",
    "    ANSWERS[key] = [float(v) for v in values]\n",
    "\n",
    "\n",
    "def finish(fig, name):\n",
    "    \"\"\"保存图；作为脚本运行时随即关闭，在 Jupyter 中留给 inline 后端显示。\"\"\"\n",
    "    fig.savefig(OUT / f\"{name}.png\", bbox_inches=\"tight\")\n",
    "    if \"ipykernel\" not in sys.modules:\n",
    "        plt.close(fig)"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-01"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查工具：运行这一格即可，不用读懂或修改。\n",
    "# 每题代码后面的“自查”格调用这里的 check()：数值算对了打印 ✓ 通过，不对就打印一句提示。\n",
    "# 正确答案只以哈希值保存在自查格里，看不到具体数字。\n",
    "import hashlib\n",
    "\n",
    "SELFCHECK_FAILED = []\n",
    "\n",
    "\n",
    "def _flat(v):\n",
    "    if hasattr(v, \"to_numpy\"):\n",
    "        v = v.to_numpy()\n",
    "    if isinstance(v, np.ndarray):\n",
    "        v = v.ravel().tolist()\n",
    "    if isinstance(v, (list, tuple)):\n",
    "        return [x for item in v for x in _flat(item)]\n",
    "    return [v]\n",
    "\n",
    "\n",
    "def _key(v, digits):\n",
    "    parts = []\n",
    "    for x in _flat(v):\n",
    "        if x is Ellipsis:\n",
    "            raise ValueError(\"ellipsis\")\n",
    "        s = f\"{float(x):.1e}\" if digits == \"e\" else f\"{float(x):.{digits}f}\"\n",
    "        parts.append(s[1:] if s.startswith(\"-\") and float(s) == 0 else s)\n",
    "    return hashlib.sha256(\"|\".join(parts).encode()).hexdigest()[:10]\n",
    "\n",
    "\n",
    "def check(qid, items):\n",
    "    print(f\"—— 自查 {qid} ——\")\n",
    "    for label, get, digits, good, hint, *known in items:\n",
    "        try:\n",
    "            k = _key(get(), digits)\n",
    "        except NameError as e:\n",
    "            msg = f\"还没有定义 {getattr(e, 'name', '')}：先运行上面的代码格\"\n",
    "        except Exception as e:\n",
    "            if \"ellipsis\" in str(e).lower():\n",
    "                msg = \"还没完成：上面的代码里还有 ... 没有填\"\n",
    "            else:\n",
    "                msg = f\"算不出来（{type(e).__name__}）：确认上面的代码已运行，变量名和结构与框架一致\"\n",
    "        else:\n",
    "            if k == good:\n",
    "                print(f\"✓ 通过　{label}\")\n",
    "                continue\n",
    "            msg = (known[0] if known else {}).get(k, hint)\n",
    "        SELFCHECK_FAILED.append(f\"{qid} {label}\")\n",
    "        print(f\"✗ {label}：{msg}\")\n",
    "\n",
    "\n",
    "def reflect(qid, prompts):\n",
    "    print(f\"—— 自查 {qid} ——（讨论题不自动判分，交之前对照检查）\")\n",
    "    for p in prompts:\n",
    "        print(\"□\", p)\n",
    "\n",
    "\n",
    "def _mf(sp):\n",
    "    \"\"\"B3、B4 自查用：sexed 中某个物种雄性、雌性的体重。\"\"\"\n",
    "    d = sexed[sexed[\"species\"] == sp]\n",
    "    return d.loc[d[\"sex\"] == \"male\", \"body_mass_g\"], d.loc[d[\"sex\"] == \"female\", \"body_mass_g\"]\n",
    "\n",
    "\n",
    "def _line_slopes(ax):\n",
    "    \"\"\"C3 自查用：图中每条直线的斜率，从小到大。\"\"\"\n",
    "    out = []\n",
    "    for ln in ax.get_lines():\n",
    "        x, y = np.asarray(ln.get_xdata(), float), np.asarray(ln.get_ydata(), float)\n",
    "        if len(x) >= 2 and x[-1] != x[0]:\n",
    "            out.append((y[-1] - y[0]) / (x[-1] - x[0]))\n",
    "    return sorted(out)"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-02"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## A 部分　认识数据（第 1 周）\n",
    "\n",
    "### A1　读入数据\n",
    "\n",
    "用 `pd.read_csv` 读入 `penguins.csv`（缺测值在文件里写作 `NA`），回答：\n",
    "\n",
    "1. 数据有几行、几列？每一行代表什么？\n",
    "2. 哪些变量是分类变量，哪些是数值变量？`year` 在文件里是整数，分析时应当把它当作哪一类？"
   ],
   "id": "cell-03"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "df = pd.read_csv(find_data(), na_values=[\"NA\"])\n",
    "print(df.shape)\n",
    "print(df.dtypes)\n",
    "print(df.head())\n",
    "record(\"n_total\", len(df))"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-04"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（A1）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('A1', [\n",
    "    ('行数和列数', lambda: df.shape, 0, '73e90472b8',\n",
    "     '对照 penguins.csv 检查：是不是读错了文件，或者把表头当成了数据？'),\n",
    "    ('四项形态测量是数值型', lambda: [df[c].dtype.kind == 'f' for c in MORPH], 0, 'bd85fe66f5',\n",
    "     '形态测量列应当是浮点数；如果是文本（object），检查 na_values 有没有写对'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-05"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（A1）**\n",
    "\n",
    "- 344 行 × 8 列，每一行是一只被测量的繁殖成体企鹅（观测单位是“个体”，不是“巢”或“岛”）。\n",
    "- 分类变量：`species`、`island`、`sex`（名义变量）。数值变量：`bill_length_mm`、`bill_depth_mm`、\n",
    "  `flipper_length_mm`、`body_mass_g`，均为连续变量，只是按测量精度记录（喙 0.1 mm，鳍 1 mm，体重按弹簧秤读数）。\n",
    "- `year` 只有 2007、2008、2009 三个取值，代表三个繁殖季，应当作分类（或有序分类）变量，而不是做算术的连续变量。"
   ],
   "id": "cell-06"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### A2　缺测值\n",
    "\n",
    "1. 每一列各缺多少个值？\n",
    "2. 有几条记录四项形态测量（喙长、喙深、鳍长、体重）全部缺失？分别是哪个物种？\n",
    "3. 有几条记录四项形态测量齐全、只缺性别？分别是哪个物种？有没有只缺一部分形态测量的记录？\n",
    "4. 如果一开始就写 `df = df.dropna()`，还剩几条记录？对“体重对鳍长的回归”来说，这样做有什么问题？"
   ],
   "id": "cell-07"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "print(df.isna().sum())\n",
    "\n",
    "no_morph = df[MORPH].isna().all(axis=1)\n",
    "sex_only = df[\"sex\"].isna() & df[MORPH].notna().all(axis=1)\n",
    "partial = df[MORPH].isna().any(axis=1) & ~no_morph\n",
    "\n",
    "print(\"四项全缺：\", no_morph.sum(), df.loc[no_morph, \"species\"].value_counts().to_dict())\n",
    "print(\"只缺性别：\", sex_only.sum(), df.loc[sex_only, \"species\"].value_counts().to_dict())\n",
    "print(\"只缺部分形态：\", partial.sum())\n",
    "print(\"dropna() 后：\", len(df.dropna()))\n",
    "\n",
    "morph = df.dropna(subset=MORPH)            # 形态比较、回归用（n = 342）\n",
    "sexed = df.dropna(subset=MORPH + [\"sex\"])  # 两性比较用（n = 333）\n",
    "print(\"形态齐全：\", len(morph), \" 形态与性别齐全：\", len(sexed))\n",
    "\n",
    "record(\"miss_nomorph\", no_morph.sum())\n",
    "record(\"miss_sex\", sex_only.sum())\n",
    "for sp in SPECIES:\n",
    "    record(f\"miss_nomorph_{sp}\", (no_morph & (df[\"species\"] == sp)).sum())\n",
    "    record(f\"miss_sex_{sp}\", (sex_only & (df[\"species\"] == sp)).sum())\n",
    "    record(f\"x_n_morph_{sp}\", (morph[\"species\"] == sp).sum())\n",
    "    record(f\"x_n_sexed_{sp}\", (sexed[\"species\"] == sp).sum())\n",
    "record(\"x_partial\", partial.sum())\n",
    "record(\"n_morph\", len(morph))\n",
    "record(\"n_sexed\", len(sexed))"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-08"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（A2）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('A2', [\n",
    "    ('四项全缺的记录数', lambda: no_morph.sum(), 0, 'd4735e3a26',\n",
    "     '四项要“全部”缺失才算：看看你对 df[MORPH].isna() 用的是 all 还是 any，axis 对不对',\n",
    "     {'4fc82b26ae': '你数的是“任何一列有缺测”的记录；这里只看四项形态测量是否全部缺失'}),\n",
    "    ('只缺性别的记录数', lambda: sex_only.sum(), 0, '19581e27de',\n",
    "     '两个条件要同时满足：sex 缺失，并且四项形态测量都不缺',\n",
    "     {'4fc82b26ae': '这是 sex 缺失的总数；其中有的记录连形态测量也没有，要排除'}),\n",
    "    ('morph 的记录数', lambda: len(morph), 0, '023849c389',\n",
    "     'morph 只要求四项形态测量齐全：看 dropna 的 subset',\n",
    "     {'556d7dc3a1': '你把缺性别的记录也删了：morph 不要求有性别'}),\n",
    "    ('sexed 的记录数', lambda: len(sexed), 0, '556d7dc3a1',\n",
    "     'sexed 要求四项形态测量和性别都齐全：看 dropna 的 subset',\n",
    "     {'023849c389': 'sexed 还要去掉没有性别的记录'}),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-09"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（A2）**\n",
    "\n",
    "1. 四项形态测量各缺 2 个，`sex` 缺 11 个，`species`、`island`、`year` 不缺。\n",
    "2. 2 条记录四项全缺：阿德利企鹅 1 条、巴布亚企鹅 1 条（这两条也没有性别）。\n",
    "3. 9 条记录只缺性别：阿德利企鹅 5 条、巴布亚企鹅 4 条。没有只缺一部分形态测量的记录。\n",
    "4. `dropna()` 后剩 333 条。体重对鳍长的回归用不到性别，本可以用 342 条记录；\n",
    "   一刀切地删掉缺任何变量的记录，会无故丢掉 9 只企鹅。正确做法是每项分析只要求它用到的变量齐全，\n",
    "   并像论文图 2 那样报告每一步去掉了哪些记录。"
   ],
   "id": "cell-10"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### A3　样本构成\n",
    "\n",
    "1. 做出物种 × 岛屿的交叉表（带行列合计），以及各年份的记录数。\n",
    "2. 能否用这批数据比较“岛屿”对体重的影响？如果能，只能在哪个物种内比较？"
   ],
   "id": "cell-11"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "tab = pd.crosstab(df[\"species\"], df[\"island\"], margins=True, margins_name=\"Total\")\n",
    "print(tab)\n",
    "years = df[\"year\"].value_counts().sort_index()\n",
    "print(years)\n",
    "\n",
    "for sp in SPECIES:\n",
    "    record(f\"n_{sp}\", (df[\"species\"] == sp).sum())\n",
    "    for isl in [\"Biscoe\", \"Dream\", \"Torgersen\"]:\n",
    "        k = ((df[\"species\"] == sp) & (df[\"island\"] == isl)).sum()\n",
    "        if k:\n",
    "            record(f\"isl_{sp}_{isl}\", k)\n",
    "for isl in [\"Biscoe\", \"Dream\", \"Torgersen\"]:\n",
    "    record(f\"isl_total_{isl}\", (df[\"island\"] == isl).sum())\n",
    "for yr in [2007, 2008, 2009]:\n",
    "    record(f\"n_year_{yr}\", (df[\"year\"] == yr).sum())"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-12"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（A3）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('A3', [\n",
    "    ('物种 × 岛屿交叉表', lambda: tab.loc[SPECIES, ['Biscoe', 'Dream', 'Torgersen']], 0, '642ae2c4cb',\n",
    "     '行是物种、列是岛屿，用全部记录 df',\n",
    "     {'4e1ed1c889': '这是用 morph 做的；样本构成要用全部记录 df'}),\n",
    "    ('各年份记录数', lambda: years.loc[[2007, 2008, 2009]], 0, '7609773af2',\n",
    "     '对 df 的 year 列计数',\n",
    "     {'45d1bba5a2': '这是用 morph 数的；这里要用全部记录 df'}),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-13"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（A3）**\n",
    "\n",
    "- 阿德利企鹅 152 只（Biscoe 44、Dream 56、Torgersen 52），帽带企鹅 68 只（全部来自 Dream），\n",
    "  巴布亚企鹅 124 只（全部来自 Biscoe）；岛屿合计 Biscoe 168、Dream 124、Torgersen 52。\n",
    "  2007、2008、2009 年分别为 110、114、120 只。\n",
    "- 除阿德利企鹅外，物种和岛屿完全重合：帽带企鹅与巴布亚企鹅之间的任何差异，都分不清是物种效应还是岛屿效应（混杂）。\n",
    "  岛屿之间的比较只能在阿德利企鹅内部进行。"
   ],
   "id": "cell-14"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### A4　复现论文表 1\n",
    "\n",
    "按物种和性别分组，计算四项形态测量的均值和标准差（保留与表 1 相同的小数位），以及每组的只数 n。\n",
    "每个物种再加一行“全部”（含未记录性别的个体）。\n",
    "\n",
    "提示：`groupby(...).agg([\"mean\", \"std\", \"count\"])`。思考：`pandas` 的 `.std()` 和 `numpy` 的 `np.std()`\n",
    "默认结果为什么不一样？表 1 用的是哪一种？"
   ],
   "id": "cell-15"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "by_sex = morph.dropna(subset=[\"sex\"]).groupby([\"species\", \"sex\"])[MORPH].agg([\"mean\", \"std\", \"count\"])\n",
    "by_species = morph.groupby(\"species\")[MORPH].agg([\"mean\", \"std\", \"count\"])\n",
    "print(by_sex.round(1))\n",
    "print(by_species.round(1))\n",
    "\n",
    "SHORT = {\"bill_length_mm\": \"bl\", \"bill_depth_mm\": \"bd\", \"flipper_length_mm\": \"fl\", \"body_mass_g\": \"bm\"}\n",
    "for sp in SPECIES:\n",
    "    record(f\"t1_{sp}_all_n\", by_species.loc[sp, (MORPH[0], \"count\")])\n",
    "    for v in MORPH:\n",
    "        record(f\"t1_{sp}_all_{SHORT[v]}\", by_species.loc[sp, (v, \"mean\")], by_species.loc[sp, (v, \"std\")])\n",
    "    for sx in [\"female\", \"male\"]:\n",
    "        record(f\"t1_{sp}_{sx}_n\", by_sex.loc[(sp, sx), (MORPH[0], \"count\")])\n",
    "        for v in MORPH:\n",
    "            record(f\"t1_{sp}_{sx}_{SHORT[v]}\", by_sex.loc[(sp, sx), (v, \"mean\")], by_sex.loc[(sp, sx), (v, \"std\")])\n",
    "\n",
    "x = morph.loc[morph[\"species\"] == \"Adelie\", \"body_mass_g\"]\n",
    "print(\"pandas std:\", round(x.std(), 1), \" numpy std:\", round(np.std(x.to_numpy()), 1),\n",
    "      \" numpy std ddof=1:\", round(np.std(x.to_numpy(), ddof=1), 1))\n",
    "record(\"x_adelie_bm_sd_ddof0\", np.std(x.to_numpy()))"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-16"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（A4）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('A4', [\n",
    "    ('雌雄各组的只数', lambda: [by_sex.loc[(sp, sx), ('body_mass_g', 'count')] for sp in SPECIES for sx in ('female', 'male')], 0, '17240ab62d',\n",
    "     '按 (物种, 性别) 分组，数据用形态测量和性别都齐全的记录'),\n",
    "    ('雌雄各组的均值', lambda: [by_sex.loc[(sp, sx), (v, 'mean')] for sp in SPECIES for sx in ('female', 'male') for v in MORPH], 1, '58b39d1444',\n",
    "     '检查分组（物种, 性别）和数据子集；agg 里要有 \"mean\"'),\n",
    "    ('雌雄各组的标准差', lambda: [by_sex.loc[(sp, sx), (v, 'std')] for sp in SPECIES for sx in ('female', 'male') for v in MORPH], 1, '779ea83633',\n",
    "     '检查分组和数据子集；agg 里要有 \"std\"',\n",
    "     {'937075dc48': '你算的是 ddof=0 的标准差（除以 n）；表 1 用的是样本标准差'}),\n",
    "    ('“全部”一行的只数', lambda: [by_species.loc[sp, ('body_mass_g', 'count')] for sp in SPECIES], 0, '38ffdbc141',\n",
    "     '“全部”按物种分组，用形态测量齐全的记录',\n",
    "     {'5cf662e89b': '你用的是有性别的记录；“全部”应当包含未记录性别的个体'}),\n",
    "    ('“全部”一行的均值和标准差', lambda: [by_species.loc[sp, (v, s)] for sp in SPECIES for v in MORPH for s in ('mean', 'std')], 1, 'd06d4e3cad',\n",
    "     '“全部”按物种分组，用形态测量齐全的记录，agg 里要有 \"mean\" 和 \"std\"',\n",
    "     {'1293a62d38': '你用的是有性别的记录；“全部”应当包含未记录性别的个体',\n",
    "      '52c26a0dab': '标准差用的是 ddof=0（除以 n）；表 1 用的是样本标准差'}),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-17"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（A4）**\n",
    "\n",
    "- 结果应与论文表 1 逐格一致，例如阿德利企鹅全部 151 只：喙长 38.8 ± 2.7 mm，喙深 18.3 ± 1.2 mm，\n",
    "  鳍长 190.0 ± 6.5 mm，体重 3701 ± 459 g；巴布亚企鹅雄性 61 只，体重 5485 ± 313 g。\n",
    "- `pandas` 的 `.std()` 默认 `ddof=1`（除以 n − 1，样本标准差），`np.std()` 默认 `ddof=0`（除以 n）。\n",
    "  表 1 用的是 `ddof=1`。n 较大时两者差别很小（阿德利企鹅体重 459 对 457），n 小时差别明显。\n",
    "- 注意“全部”一行的 n（阿德利企鹅 151、巴布亚企鹅 123）大于雌雄之和（146、119），因为它包含未记录性别的个体。"
   ],
   "id": "cell-18"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### A5　体重的中位数与四分位距\n",
    "\n",
    "1. 计算每个物种体重的中位数和四分位距（第 1、第 3 四分位数）。\n",
    "2. 帽带企鹅的第 1 四分位数为什么不是一个实际测到的体重值？换一种分位数算法会不会得到不同的数？\n",
    "3. 巴布亚企鹅的体重中位数比阿德利企鹅高多少克？相当于阿德利企鹅中位数的百分之几？"
   ],
   "id": "cell-19"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "q = morph.groupby(\"species\")[\"body_mass_g\"].quantile([0.25, 0.5, 0.75]).unstack()\n",
    "print(q)\n",
    "for sp in SPECIES:\n",
    "    record(f\"b_med_{sp}\", q.loc[sp, 0.5])\n",
    "    record(f\"b_iqr_{sp}\", q.loc[sp, 0.25], q.loc[sp, 0.75])\n",
    "gap = q.loc[\"Gentoo\", 0.5] - q.loc[\"Adelie\", 0.5]\n",
    "gap_pct = gap / q.loc[\"Adelie\", 0.5] * 100\n",
    "record(\"b_gap\", gap)\n",
    "record(\"b_gap_pct\", gap_pct)\n",
    "print(\"差：\", gap, \"g，\", round(gap_pct, 1), \"%\")\n",
    "\n",
    "chin = morph.loc[morph[\"species\"] == \"Chinstrap\", \"body_mass_g\"].to_numpy()\n",
    "for method in [\"linear\", \"lower\", \"midpoint\", \"weibull\"]:\n",
    "    print(method, np.percentile(chin, 25, method=method))\n",
    "    record(f\"x_q1_chin_{method}\", np.percentile(chin, 25, method=method))"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-20"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（A5）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('A5', [\n",
    "    ('各物种的三个四分位数', lambda: q.loc[SPECIES, [0.25, 0.5, 0.75]], 1, '6d6ba64981',\n",
    "     '用 morph 按物种分组，对 body_mass_g 取 0.25、0.5、0.75 分位数（保持默认算法），再 unstack'),\n",
    "    ('中位数之差', lambda: gap, 0, 'a037a7dfaf',\n",
    "     '巴布亚企鹅的中位数减阿德利企鹅的中位数，都从 q 的 0.5 那一列取',\n",
    "     {'d818f221cb': '你用的是均值；题目问的是中位数'}),\n",
    "    ('差值占阿德利企鹅中位数的百分比', lambda: gap_pct, 1, '6cc32a9a06',\n",
    "     '差值除以阿德利企鹅的中位数，再乘 100',\n",
    "     {'80a753d4a4': '分母应当是阿德利企鹅的中位数'}),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-21"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（A5）**\n",
    "\n",
    "1. 中位数：阿德利企鹅 3700 g（四分位距 3350–4000 g），帽带企鹅 3700 g（3487.5–3950 g），巴布亚企鹅 5000 g（4700–5500 g）。\n",
    "2. `pandas`/`numpy` 默认用线性插值（`method=\"linear\"`），第 1 四分位数落在两个观测值之间时取插值。\n",
    "   换成其他算法（如 `lower`、`weibull`）会得到不同的数；报告分位数时应说明算法，或至少说明所用软件。\n",
    "3. 高 1300 g，相当于阿德利企鹅中位数的 35.1%。"
   ],
   "id": "cell-22"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### A6　分组作图\n",
    "\n",
    "1. 画出不分物种的体重直方图。\n",
    "2. 再按物种分开画（叠加直方图、分面直方图或箱线图加散点均可）。\n",
    "3. 对比两张图：不分物种时的分布形状说明了什么？用“全部 342 只的均值 ± 标准差”来概括体重合适吗？"
   ],
   "id": "cell-23"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "fig, axes = plt.subplots(4, 1, figsize=(6, 6.6), sharex=True, sharey=True)\n",
    "bins = np.arange(2600, 6500, 100)\n",
    "axes[0].hist(morph[\"body_mass_g\"], bins=bins, color=\"#777777\", edgecolor=\"white\", linewidth=0.5)\n",
    "axes[0].set_title(f\"All species pooled (n = {len(morph)})\", loc=\"left\", fontsize=10)\n",
    "for ax, sp in zip(axes[1:], SPECIES):\n",
    "    d = morph.loc[morph[\"species\"] == sp, \"body_mass_g\"]\n",
    "    ax.hist(d, bins=bins, color=COLORS[sp], edgecolor=\"white\", linewidth=0.5)\n",
    "    ax.set_title(f\"{LABEL[sp]} (n = {len(d)})\", loc=\"left\", fontsize=10)\n",
    "axes[-1].set_xlabel(\"Body mass (g)\")\n",
    "fig.supylabel(\"Number of penguins\")\n",
    "fig.tight_layout()\n",
    "finish(fig, \"A6_body_mass_hist\")\n",
    "print(\"合并均值 ± SD：\", round(morph[\"body_mass_g\"].mean()), \"±\", round(morph[\"body_mass_g\"].std()))\n",
    "record(\"x_pooled_bm\", morph[\"body_mass_g\"].mean(), morph[\"body_mass_g\"].std())"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-24"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（A6）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('A6', [\n",
    "    ('四张直方图各画了多少只', lambda: [sum(p.get_height() for p in a.patches) for a in fig.axes], 0, '3a7a05b339',\n",
    "     '自上而下应分别是：形态齐全的全部企鹅、阿德利、帽带、巴布亚企鹅（顺序同 SPECIES）；数据用 morph 的 body_mass_g，bins 用框架里给的'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-25"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（A6）**\n",
    "\n",
    "- 不分物种的直方图右偏，右侧拖着一个“肩”：它是三个物种分布的混合，巴布亚企鹅的中位数比另两种高 1300 g，堆在右侧。\n",
    "- “全部 342 只的均值 ± 标准差”把三个物种混在一起，标准差主要反映物种间差异，不能代表任何一个物种。\n",
    "  应按物种（必要时再按性别）分别概括，即论文表 1 的做法。\n",
    "- 第 2 张图用分面（每个物种一行、共用横轴）比叠加半透明直方图清楚，叠加时重叠部分的颜色会混在一起。\n",
    "- 评分时看图是否有坐标轴标题和单位、分组是否可辨（图例或分面），不要求颜色与参考答案相同。"
   ],
   "id": "cell-26"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## B 部分　两性体重差异（第 2 周）\n",
    "\n",
    "本部分只用形态测量和性别都齐全的 333 只企鹅（A2 中的 `sexed`）。\n",
    "\n",
    "### B1　描述两性差异\n",
    "\n",
    "对每个物种，计算雌、雄体重的均值、标准差和只数，雄雌均值差（雄 − 雌），\n",
    "以及雄性均值比雌性均值高出的百分比：（雄性均值 / 雌性均值 − 1）× 100%。\n",
    "三个物种中，哪个物种体重的两性差异最大？哪个最小？"
   ],
   "id": "cell-27"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "g = sexed.groupby([\"species\", \"sex\"])[\"body_mass_g\"].agg([\"mean\", \"std\", \"count\"])\n",
    "print(g.round(0))\n",
    "diff, pct = {}, {}\n",
    "for sp in SPECIES:\n",
    "    f, m = g.loc[(sp, \"female\"), \"mean\"], g.loc[(sp, \"male\"), \"mean\"]\n",
    "    diff[sp] = m - f\n",
    "    pct[sp] = (m / f - 1) * 100\n",
    "    record(f\"e_diff_{sp}\", diff[sp])\n",
    "    record(f\"e_pct_{sp}\", pct[sp])\n",
    "    print(f\"{sp:10s} 差 {diff[sp]:6.0f} g   高出 {pct[sp]:5.1f} %\")"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-28"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（B1）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('B1', [\n",
    "    ('雌雄体重的均值', lambda: [g.loc[(sp, sx), 'mean'] for sp in SPECIES for sx in ('female', 'male')], 0, 'e128bc7934',\n",
    "     '用 sexed 按 (物种, 性别) 分组，对 body_mass_g 求均值'),\n",
    "    ('雄雌均值差', lambda: [diff[sp] for sp in SPECIES], 0, '5a118a04d4',\n",
    "     '每个物种的雄性均值减雌性均值',\n",
    "     {'945f2ff772': '方向反了：题目要的是雄 − 雌'}),\n",
    "    ('雄性高出的百分比', lambda: [pct[sp] for sp in SPECIES], 1, 'cb20d77874',\n",
    "     '对照题目里的公式：雄性均值除以雌性均值，减 1，再乘 100',\n",
    "     {'be6c918128': '分母应当是雌性均值',\n",
    "      'e638c0f53f': '别忘了减 1：要的是“高出”的百分比'}),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-29"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（B1）**\n",
    "\n",
    "雄雌均值差：阿德利企鹅 675 g、帽带企鹅 412 g、巴布亚企鹅 805 g；\n",
    "雄性比雌性重 20.0%、11.7%、17.2%。按百分比，阿德利企鹅最大、帽带企鹅最小。\n",
    "注意按绝对差（克）排序时巴布亚企鹅最大，两种尺度的排序不同，所以要说明用的是哪种尺度；\n",
    "论文用百分比，与原始研究的两性体型差异指数（SDI）排序一致。"
   ],
   "id": "cell-30"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### B2　检查正态性\n",
    "\n",
    "对 6 个物种 × 性别组分别做 Shapiro–Wilk 检验，并画出正态 Q–Q 图（`scipy.stats.probplot`）。\n",
    "Shapiro–Wilk 检验 P > 0.05 能不能说明“数据服从正态分布”？"
   ],
   "id": "cell-31"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "fig, axes = plt.subplots(2, 3, figsize=(9, 5.4))\n",
    "sw = []\n",
    "for j, sp in enumerate(SPECIES):\n",
    "    for i, sx in enumerate([\"female\", \"male\"]):\n",
    "        y = sexed.loc[(sexed[\"species\"] == sp) & (sexed[\"sex\"] == sx), \"body_mass_g\"].to_numpy()\n",
    "        p = stats.shapiro(y).pvalue\n",
    "        sw.append(p)\n",
    "        stats.probplot(y, dist=\"norm\", plot=axes[i, j])\n",
    "        axes[i, j].get_lines()[0].set(color=COLORS[sp], markersize=3)\n",
    "        axes[i, j].get_lines()[1].set(color=\"#555555\")\n",
    "        axes[i, j].set(title=f\"{LABEL[sp]} {sx} (Shapiro–Wilk P = {p:.2f})\", xlabel=\"\", ylabel=\"\")\n",
    "        axes[i, j].title.set_fontsize(9)\n",
    "fig.supxlabel(\"Theoretical normal quantiles\")\n",
    "fig.supylabel(\"Body mass (g)\")\n",
    "fig.tight_layout()\n",
    "finish(fig, \"B2_qq\")\n",
    "record(\"e_sw_range\", min(sw), max(sw))\n",
    "print(\"Shapiro–Wilk P 范围：\", round(min(sw), 2), \"–\", round(max(sw), 2))"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-32"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（B2）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('B2', [\n",
    "    ('6 组 Shapiro–Wilk 检验的 P 值', lambda: sorted(sw), 2, '79978914f0',\n",
    "     '对 6 个物种 × 性别组的体重（来自 sexed）分别做 stats.shapiro，把 .pvalue 依次放进 sw'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-33"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（B2）**\n",
    "\n",
    "6 组的 Shapiro–Wilk P 在 0.20–0.99 之间，Q–Q 图上的点大体沿直线排列。\n",
    "P > 0.05 只说明“没有检测到明显偏离正态”，不能证明服从正态；小样本时检验的功效低，大样本时又会对很小的偏离敏感。\n",
    "所以论文把 Shapiro–Wilk 检验当作辅助检查，并用不依赖正态假设的 Mann–Whitney U 检验复核（B5）。"
   ],
   "id": "cell-34"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### B3　Welch t 检验与均值差的置信区间\n",
    "\n",
    "1. 对每个物种，用 `scipy.stats.ttest_ind(雄, 雌, equal_var=False)` 做双侧 Welch t 检验，记下 t、自由度和 P。\n",
    "2. 不调用现成函数，按公式自己算出 t、Welch–Satterthwaite 自由度和雄雌均值差的 95% 置信区间：\n",
    "   SE = √(s₁²/n₁ + s₂²/n₂)，t = (x̄₁ − x̄₂)/SE，\n",
    "   df = (s₁²/n₁ + s₂²/n₂)² / [(s₁²/n₁)²/(n₁−1) + (s₂²/n₂)²/(n₂−1)]，\n",
    "   95% CI = (x̄₁ − x̄₂) ± t₀.₉₇₅,df × SE。和第 1 步的结果对一下。\n",
    "3. 再用 `equal_var=True`（Student t 检验）算一遍。哪些物种的 t 值完全没变？为什么？"
   ],
   "id": "cell-35"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "def welch(a, b, level=0.95):\n",
    "    \"\"\"a − b 的 Welch t 检验：返回 t、自由度、双侧 P、均值差及其置信区间的下限和上限。\"\"\"\n",
    "    a, b = np.asarray(a, float), np.asarray(b, float)\n",
    "    n1, n2 = len(a), len(b)\n",
    "    va = a.var(ddof=1) / n1  # s₁²/n₁\n",
    "    vb = b.var(ddof=1) / n2  # s₂²/n₂\n",
    "    diff = a.mean() - b.mean()\n",
    "    se = np.sqrt(va + vb)                                          # 填空 1\n",
    "    df_w = (va + vb) ** 2 / (va**2 / (n1 - 1) + vb**2 / (n2 - 1))  # 填空 2\n",
    "    t = diff / se\n",
    "    p = 2 * stats.t.sf(abs(t), df_w)\n",
    "    half = stats.t.ppf(0.5 + level / 2, df_w) * se                 # 填空 3\n",
    "    return t, df_w, p, diff, diff - half, diff + half\n",
    "\n",
    "\n",
    "rows = []\n",
    "for sp in SPECIES:\n",
    "    d = sexed[sexed[\"species\"] == sp]\n",
    "    m = d.loc[d[\"sex\"] == \"male\", \"body_mass_g\"]\n",
    "    f = d.loc[d[\"sex\"] == \"female\", \"body_mass_g\"]\n",
    "    t, df_w, p, diff, lo, hi = welch(m, f)\n",
    "    sp_res = stats.ttest_ind(m, f, equal_var=False)\n",
    "    st_res = stats.ttest_ind(m, f, equal_var=True)\n",
    "    assert np.isclose(t, sp_res.statistic) and np.isclose(p, sp_res.pvalue)\n",
    "    rows.append(dict(species=sp, n_f=len(f), n_m=len(m), t=t, df=df_w, p=p, diff=diff, lo=lo, hi=hi,\n",
    "                     t_student=st_res.statistic, df_student=len(m) + len(f) - 2, p_student=st_res.pvalue))\n",
    "    record(f\"e_t_{sp}\", t)\n",
    "    record(f\"e_df_{sp}\", df_w)\n",
    "    record(f\"e_ci_{sp}\", lo, hi)\n",
    "    record(f\"x_e_p_{sp}\", p)\n",
    "    record(f\"x_e_tstudent_{sp}\", st_res.statistic)\n",
    "tests = pd.DataFrame(rows).set_index(\"species\")\n",
    "print(tests.round(3).to_string())"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-36"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（B3）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('B3', [\n",
    "    ('t 统计量', lambda: [welch(*_mf(sp))[0] for sp in SPECIES], 1, 'f84e3a44cf',\n",
    "     't = diff / se，t 不对说明填空 1 的标准误不对：对照题目里 SE 的公式',\n",
    "     {'f8477d58c6': '填空 1：标准误要开平方根',\n",
    "      '8e2663caec': '填空 1：先把两项 s²/n 相加，再开平方根'}),\n",
    "    ('自由度', lambda: [welch(*_mf(sp))[1] for sp in SPECIES], 1, '058e6b7330',\n",
    "     '检查填空 2：对照题目里 df 的公式，分子、分母各是什么',\n",
    "     {'d94c9f90e3': '这是 Student t 检验的自由度；Welch 检验要用题目里的 Welch–Satterthwaite 公式',\n",
    "      '567085cd6f': '填空 2：分母两项各自除以 n − 1，不是 n'}),\n",
    "    ('双侧 P 值', lambda: [welch(*_mf(sp))[2] for sp in SPECIES], 'e', '084cc6eb08',\n",
    "     'P 由 t 和自由度算出（这一行框架已写好），先让上面两项通过'),\n",
    "    ('均值差的 95% 置信区间', lambda: [welch(*_mf(sp))[4:6] for sp in SPECIES], 0, '8720a79fcc',\n",
    "     '检查填空 3：用的是哪个分布、哪个分位数、哪个自由度',\n",
    "     {'5c3f84de3a': '填空 3 用了正态分布的 1.96；这里要用 t 分布的分位数',\n",
    "      'e2f8a893ac': '95% 双侧区间用的是 0.975 分位数，不是 0.95'}),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-37"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（B3）**\n",
    "\n",
    "| 物种 | t | 自由度 | 雄雌均值差（95% CI） |\n",
    "|---|---|---|---|\n",
    "| 阿德利企鹅 | 13.1 | 135.7 | 675 g（573–776） |\n",
    "| 帽带企鹅 | 5.2 | 62.6 | 412 g（254–570） |\n",
    "| 巴布亚企鹅 | 14.8 | 116.6 | 805 g（697–913） |\n",
    "\n",
    "- 手算结果与 `scipy` 一致（代码里用 `assert` 做了核对）。\n",
    "- 阿德利企鹅（73/73）和帽带企鹅（34/34）的 Student t 值与 Welch t 值完全相同：两组 n 相等时，\n",
    "  两种检验的 t 统计量在代数上相等，只有自由度不同（Student 为 n₁ + n₂ − 2）。巴布亚企鹅 58/61，t 值略有差别。\n",
    "  n 不等且方差不等时两者差别才大；Welch 检验不要求方差相等，代价很小，所以论文默认用它。"
   ],
   "id": "cell-38"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### B4　多重比较：Holm 校正\n",
    "\n",
    "三个物种各做了一次检验。自己写一个函数实现 Holm 校正：把 m 个 P 值从小到大排序，\n",
    "第 k 小的乘以 (m − k + 1)，再从小到大取累计最大值，超过 1 的记为 1。\n",
    "校正后的 P 值是多少？和不校正、Bonferroni 校正（都乘以 m）相比有什么不同？"
   ],
   "id": "cell-39"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "def holm(pvals):\n",
    "    p = np.asarray(pvals, dtype=float)\n",
    "    m = len(p)\n",
    "    adj = np.empty(m)\n",
    "    running = 0.0\n",
    "    for k, idx in enumerate(np.argsort(p)):\n",
    "        running = max(running, min(1.0, (m - k) * p[idx]))\n",
    "        adj[idx] = running\n",
    "    return adj\n",
    "\n",
    "\n",
    "tests[\"p_holm\"] = holm(tests[\"p\"])\n",
    "tests[\"p_bonf\"] = np.minimum(1, tests[\"p\"] * len(tests))\n",
    "print(tests[[\"p\", \"p_holm\", \"p_bonf\"]].to_string(float_format=lambda v: f\"{v:.2g}\"))\n",
    "for sp in SPECIES:\n",
    "    record(f\"e_padj_{sp}\", tests.loc[sp, \"p_holm\"])\n",
    "    record(f\"x_e_pbonf_{sp}\", tests.loc[sp, \"p_bonf\"])\n",
    "record(\"e_padj_max\", tests[\"p_holm\"].max())"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-40"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（B4）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('B4', [\n",
    "    ('holm() 处理示例 [0.04, 0.01, 0.03]', lambda: holm([0.04, 0.01, 0.03]), 2, '96119003e7',\n",
    "     '对照题目里的步骤：排序、乘 (m − k)、取累计最大值、放回原来的位置',\n",
    "     {'b1a004b3ab': '漏了“累计最大值”：排在后面的校正 P 值不能比前面的小',\n",
    "      'ad999c02c7': '这是 Bonferroni（每个都乘 m）；Holm 的乘数随排名递减',\n",
    "      'dcd39a3549': '结果要放回每个 P 值原来的位置，不是按大小排好输出'}),\n",
    "    ('holm() 处理示例 [0.6, 0.7, 0.02]', lambda: holm([0.6, 0.7, 0.02]), 2, '6cb9a4c474',\n",
    "     '对照题目里的步骤逐一检查：乘数、累计最大值、上限 1、放回原来的位置',\n",
    "     {'567a478a2e': '校正后的 P 值不能超过 1'}),\n",
    "    ('三个物种的 Holm 校正 P 值', lambda: holm([stats.ttest_ind(*_mf(sp), equal_var=False).pvalue for sp in SPECIES]), 'e', 'a8e50abee3',\n",
    "     '上面两个示例通过后这一项应当也通过；不通过时检查 holm() 是否改动了传入的数组'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-41"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（B4）**\n",
    "\n",
    "Holm 校正后 P：阿德利企鹅 1.3 × 10⁻²⁵，帽带企鹅 2.3 × 10⁻⁶，巴布亚企鹅 5.6 × 10⁻²⁸，全部 < 0.001。\n",
    "最小的 P（巴布亚企鹅）乘 3，第二小的（阿德利企鹅）乘 2，最大的（帽带企鹅）乘 1，所以帽带企鹅的 P 校正前后不变。\n",
    "Bonferroni 把三个都乘 3，结果不比 Holm 小；Holm 同样控制族错误率，但更不保守。\n",
    "本例效应很大，校正与否都不改变结论；但养成“做了几次检验就要说明怎么校正”的习惯很重要。"
   ],
   "id": "cell-42"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### B5　非参数复核：Mann–Whitney U 检验\n",
    "\n",
    "对每个物种做双侧 Mann–Whitney U 检验，同样做 Holm 校正。结论和 Welch t 检验一样吗？"
   ],
   "id": "cell-43"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "pu = [stats.mannwhitneyu(sexed.loc[(sexed[\"species\"] == sp) & (sexed[\"sex\"] == \"male\"), \"body_mass_g\"],\n",
    "                         sexed.loc[(sexed[\"species\"] == sp) & (sexed[\"sex\"] == \"female\"), \"body_mass_g\"],\n",
    "                         alternative=\"two-sided\").pvalue for sp in SPECIES]\n",
    "pu_holm = holm(pu)\n",
    "print(dict(zip(SPECIES, [f\"{v:.2g}\" for v in pu_holm])))\n",
    "record(\"e_pu_max\", pu_holm.max())\n",
    "for sp, v in zip(SPECIES, pu_holm):\n",
    "    record(f\"x_e_pu_holm_{sp}\", v)"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-44"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（B5）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('B5', [\n",
    "    ('Mann–Whitney U 检验的 P 值（未校正）', lambda: pu, 'e', 'c3cbceebb4',\n",
    "     '每个物种做 stats.mannwhitneyu(雄, 雌, alternative=\"two-sided\")，按 SPECIES 的顺序放进 pu'),\n",
    "    ('Holm 校正后的 P 值', lambda: pu_holm, 'e', 'a52be04055',\n",
    "     '把 pu 交给 B4 的 holm()',\n",
    "     {'c3cbceebb4': '还没做 Holm 校正'}),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-45"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（B5）**\n",
    "\n",
    "三个物种 Holm 校正后的 P 都 ≤ 5.3 × 10⁻⁶，结论与 Welch t 检验相同：三个物种的雄性都重于雌性。"
   ],
   "id": "cell-46"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### B6　讨论题（写在答题区，每题 2–4 句）\n",
    "\n",
    "1. 原始研究在每个巢中同时测量了两只亲鸟（通常一雌一雄）。这对 B3 的检验意味着什么？用这份数据能不能补救？\n",
    "2. 阿德利企鹅两性体重差异的 P 值小到 10⁻²⁵ 量级。“P 值很小”和“差异很大”是一回事吗？你会怎样报告这个结果？"
   ],
   "id": "cell-47"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（B6）：讨论题不自动判分，运行后对照检查你的回答。\n",
    "reflect('B6', [\n",
    "    '第 1 题：说清楚了哪一条检验假设受影响吗？说明了这份数据能不能补救、为什么吗？',\n",
    "    '第 2 题：改写后的结论里，除了 P 值，有没有效应大小和它的置信区间？数字带单位了吗？',\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-48"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（B6）**\n",
    "\n",
    "1. 同一巢的雌雄个体可能共享食物条件和繁殖时间，雌雄样本并不完全独立，而 Welch t 检验和 Mann–Whitney U 检验都假定个体独立。\n",
    "   这份简化数据表没有巢编号，无法还原配对，所以补救不了；若能从原始数据包中还原巢的配对关系，可改用配对分析或混合模型。\n",
    "   由于效应量很大，配对结构不太可能改变差异的方向，但会影响 P 值的准确程度（论文 4.5 节）。\n",
    "2. 不是。P 值同时取决于效应大小和样本量，只说明“差异不太可能是零”。应报告效应量及其置信区间，\n",
    "   例如“雄性平均重 675 g（95% CI 573–776 g），即重 20.0%；Welch t = 13.1，自由度 135.7，Holm 校正后 P = 1.3 × 10⁻²⁵”。"
   ],
   "id": "cell-49"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## C 部分　回归与辛普森悖论（第 3 周）\n",
    "\n",
    "本部分用四项形态测量齐全的 342 只企鹅（A2 中的 `morph`）。\n",
    "\n",
    "### C1　喙深对喙长：不分物种\n",
    "\n",
    "用 `scipy.stats.linregress` 拟合喙深（y）对喙长（x）的普通最小二乘回归，报告斜率、R² 和斜率的 P 值。\n",
    "只看这个结果，你会怎样描述喙长和喙深的关系？"
   ],
   "id": "cell-50"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "pool = stats.linregress(morph[\"bill_length_mm\"], morph[\"bill_depth_mm\"])\n",
    "print(f\"斜率 {pool.slope:.3f}  R² {pool.rvalue**2:.3f}  P {pool.pvalue:.2g}\")\n",
    "record(\"a_pool_slope\", pool.slope)\n",
    "record(\"a_pool_r2\", pool.rvalue**2)\n",
    "record(\"a_pool_p\", pool.pvalue)"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-51"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（C1）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('C1', [\n",
    "    ('斜率和 R²', lambda: [pool.slope, pool.rvalue ** 2], 3, '02defb4e38',\n",
    "     'x 是喙长、y 是喙深，数据用 morph',\n",
    "     {'f61cf7e808': 'x 和 y 放反了：y 是喙深',\n",
    "      '254cffc076': '结果是 nan：数据里还有缺测值，要用 morph'}),\n",
    "    ('斜率的 P 值', lambda: pool.pvalue, 'e', 'ddc7f3a16f',\n",
    "     'P 值在 linregress 结果的 pvalue 里'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-52"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（C1）**\n",
    "\n",
    "斜率 −0.085 mm/mm，R² = 0.055，P = 1.1 × 10⁻⁵。只看这一步，会得出“喙越长，喙越浅”的结论，\n",
    "而且这个负相关“统计显著”。C2 会说明这个结论是错的。"
   ],
   "id": "cell-53"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### C2　喙深对喙长：分物种\n",
    "\n",
    "在每个物种内分别拟合，报告斜率、斜率的 95% 置信区间（t 分布，自由度 n − 2）、R² 和 P。\n",
    "提示：`linregress` 的结果里有斜率的标准误 `stderr`。"
   ],
   "id": "cell-54"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "def fit_table(data, x, y, by=\"species\"):\n",
    "    keys, out = [], []\n",
    "    for key, d in data.groupby(by):\n",
    "        r = stats.linregress(d[x], d[y])\n",
    "        half = stats.t.ppf(0.975, len(d) - 2) * r.stderr\n",
    "        keys.append(key)\n",
    "        out.append(dict(n=len(d), slope=r.slope, intercept=r.intercept, lo=r.slope - half,\n",
    "                        hi=r.slope + half, r2=r.rvalue**2, p=r.pvalue))\n",
    "    return pd.DataFrame(out, index=pd.Index(keys))\n",
    "\n",
    "\n",
    "bill = fit_table(morph, \"bill_length_mm\", \"bill_depth_mm\")\n",
    "print(bill.round(3))\n",
    "for sp in SPECIES:\n",
    "    record(f\"a_slope_{sp}\", bill.loc[sp, \"slope\"])\n",
    "    record(f\"a_ci_{sp}\", bill.loc[sp, \"lo\"], bill.loc[sp, \"hi\"])\n",
    "    record(f\"a_r2_{sp}\", bill.loc[sp, \"r2\"])\n",
    "record(\"a_slope_range\", bill[\"slope\"].min(), bill[\"slope\"].max())"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-55"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（C2）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('C2', [\n",
    "    ('三个物种的斜率', lambda: bill.loc[SPECIES, 'slope'], 2, 'fc536f32f6',\n",
    "     'fit_table 的 x 是喙长、y 是喙深，数据用 morph'),\n",
    "    ('斜率的 95% 置信区间', lambda: bill.loc[SPECIES, ['lo', 'hi']], 2, 'db6a3688ca',\n",
    "     '检查 fit_table 里的 half：用的是哪个分布、哪个分位数、哪个自由度',\n",
    "     {'8d00972235': '用了正态分布的 1.96；斜率的置信区间用 t 分布，自由度 n − 2',\n",
    "      '52c73f9031': '95% 双侧区间用的是 0.975 分位数，不是 0.95'}),\n",
    "    ('R²', lambda: bill.loc[SPECIES, 'r2'], 2, '90ee546298',\n",
    "     'R² 是 r.rvalue 的平方'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-56"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（C2）**\n",
    "\n",
    "| 物种 | n | 斜率（95% CI），mm/mm | R² |\n",
    "|---|---|---|---|\n",
    "| 阿德利企鹅 | 151 | 0.18（0.11–0.25） | 0.15 |\n",
    "| 帽带企鹅 | 68 | 0.22（0.16–0.29） | 0.43 |\n",
    "| 巴布亚企鹅 | 123 | 0.20（0.16–0.25） | 0.41 |\n",
    "\n",
    "三个斜率都为正（P 均 < 0.001），与 C1 的合并斜率方向相反。这就是辛普森悖论：\n",
    "每个物种内喙越长喙越深，但巴布亚企鹅“喙长而浅”、阿德利企鹅“喙短而深”，物种间的差异把合并后的趋势拉成了负的。"
   ],
   "id": "cell-57"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### C3　把悖论画出来\n",
    "\n",
    "画喙深对喙长的散点图，点按物种着色；叠加三条物种内回归线（实线）和一条合并回归线（灰色虚线）。\n",
    "图中要有坐标轴标题（带单位）和图例。"
   ],
   "id": "cell-58"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "fig, ax = plt.subplots(figsize=(5.6, 4.2))\n",
    "for sp in SPECIES:\n",
    "    d = morph[morph[\"species\"] == sp]\n",
    "    ax.scatter(d[\"bill_length_mm\"], d[\"bill_depth_mm\"], s=12, color=COLORS[sp], alpha=0.7,\n",
    "               edgecolors=\"white\", linewidths=0.3, label=LABEL[sp])\n",
    "    xs = np.linspace(d[\"bill_length_mm\"].min(), d[\"bill_length_mm\"].max(), 2)\n",
    "    ax.plot(xs, bill.loc[sp, \"intercept\"] + bill.loc[sp, \"slope\"] * xs, color=COLORS[sp], lw=2)\n",
    "xs = np.linspace(morph[\"bill_length_mm\"].min(), morph[\"bill_length_mm\"].max(), 2)\n",
    "ax.plot(xs, pool.intercept + pool.slope * xs, color=\"#555555\", lw=2, ls=\"--\", label=\"All species pooled\")\n",
    "ax.set(xlabel=\"Bill length (mm)\", ylabel=\"Bill depth (mm)\")\n",
    "ax.legend(frameon=False, fontsize=9)\n",
    "fig.tight_layout()\n",
    "finish(fig, \"C3_simpson\")"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-59"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（C3）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('C3', [\n",
    "    ('图中直线的斜率', lambda: _line_slopes(ax), 2, '09be624873',\n",
    "     '先让 C1、C2 的自查通过；图中应有 4 条直线：三个物种各一条（截距、斜率取自 bill），一条合并回归线（取自 C1 的 pool）'),\n",
    "    ('散点的组数', lambda: len(ax.collections), 0, '4e07408562',\n",
    "     '每个物种画一组散点，共 3 组'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-60"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（C3）**\n",
    "\n",
    "见 `solution_output/C3_simpson.png`。三条彩色实线都向右上，灰色虚线向右下。\n",
    "评分看三点：点按物种区分、物种内与合并两类直线都在且可区分、坐标轴有单位。"
   ],
   "id": "cell-61"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### C4　体重对鳍长\n",
    "\n",
    "1. 在每个物种内拟合体重对鳍长的回归，报告斜率（95% CI）和 R²。\n",
    "2. 三个物种合并拟合，R² 是多少？合并后 R² 变高了，能不能说“合并分析时鳍长更能解释体重”？"
   ],
   "id": "cell-62"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "mass = fit_table(morph, \"flipper_length_mm\", \"body_mass_g\")\n",
    "print(mass.round(3))\n",
    "pool_m = stats.linregress(morph[\"flipper_length_mm\"], morph[\"body_mass_g\"])\n",
    "print(f\"合并：斜率 {pool_m.slope:.1f}  R² {pool_m.rvalue**2:.2f}\")\n",
    "for sp in SPECIES:\n",
    "    record(f\"c_slope_{sp}\", mass.loc[sp, \"slope\"])\n",
    "    record(f\"c_ci_{sp}\", mass.loc[sp, \"lo\"], mass.loc[sp, \"hi\"])\n",
    "    record(f\"c_r2_{sp}\", mass.loc[sp, \"r2\"])\n",
    "record(\"c_r2_range\", mass[\"r2\"].min(), mass[\"r2\"].max())\n",
    "record(\"c_pool_r2\", pool_m.rvalue**2)\n",
    "record(\"c_pool_slope\", pool_m.slope)"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-63"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（C4）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('C4', [\n",
    "    ('物种内斜率及其 95% 置信区间', lambda: mass.loc[SPECIES, ['slope', 'lo', 'hi']], 1, 'efacc57b86',\n",
    "     'fit_table 的 x 是鳍长、y 是体重，数据用 morph'),\n",
    "    ('物种内 R²', lambda: mass.loc[SPECIES, 'r2'], 2, '2cc3a45b7d',\n",
    "     'R² 来自 fit_table 的 r2 列'),\n",
    "    ('合并拟合的斜率', lambda: pool_m.slope, 1, 'fab41ffb8b',\n",
    "     '三个物种合并：对 morph 全体做 linregress，x 是鳍长、y 是体重'),\n",
    "    ('合并拟合的 R²', lambda: pool_m.rvalue ** 2, 2, '342a441fd8',\n",
    "     'R² 是 rvalue 的平方'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-64"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（C4）**\n",
    "\n",
    "1. 斜率：阿德利企鹅 32.8 g/mm（22.8–42.9），帽带企鹅 34.6 g/mm（24.4–44.7），巴布亚企鹅 54.6 g/mm（44.7–64.6）；\n",
    "   R² 分别为 0.22、0.41、0.49。\n",
    "2. 合并后 R² = 0.76（斜率 49.7 g/mm）。不能那样说：巴布亚企鹅鳍又长、体又重，物种之间的均值差异\n",
    "   贡献了大部分“被解释的变异”。高出的 R² 描述的是物种之间的差异，而不是同一物种内鳍长对体重的解释力。\n",
    "   （巴布亚企鹅的斜率点估计最大，但论文没有正式检验斜率之间的差异，这里也不要下“斜率不同”的结论。）"
   ],
   "id": "cell-65"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### C5　把物种放进模型\n",
    "\n",
    "论文建议“至少应把物种作为分组变量或模型中的固定效应”。拟合多元回归\n",
    "喙深 = b₀ + b₁·喙长 + b₂·[帽带企鹅] + b₃·[巴布亚企鹅]（方括号为 0/1 指示变量，阿德利企鹅为参照），\n",
    "可以用 `np.linalg.lstsq`。b₁ 是多少？与 C1、C2 的斜率比较。"
   ],
   "id": "cell-66"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "X = np.column_stack([\n",
    "    np.ones(len(morph)),\n",
    "    morph[\"bill_length_mm\"],\n",
    "    (morph[\"species\"] == \"Chinstrap\").astype(float),\n",
    "    (morph[\"species\"] == \"Gentoo\").astype(float),\n",
    "])\n",
    "beta, *_ = np.linalg.lstsq(X, morph[\"bill_depth_mm\"].to_numpy(), rcond=None)\n",
    "resid = morph[\"bill_depth_mm\"].to_numpy() - X @ beta\n",
    "s2 = resid @ resid / (len(morph) - X.shape[1])\n",
    "se_b1 = np.sqrt(s2 * np.linalg.inv(X.T @ X)[1, 1])\n",
    "half = stats.t.ppf(0.975, len(morph) - X.shape[1]) * se_b1\n",
    "print(f\"b1 = {beta[1]:.3f}  95% CI {beta[1] - half:.3f}–{beta[1] + half:.3f}\")\n",
    "print(f\"b2（帽带 − 阿德利）= {beta[2]:.2f}  b3（巴布亚 − 阿德利）= {beta[3]:.2f}\")\n",
    "record(\"x_c5_b1\", beta[1])\n",
    "record(\"x_c5_ci\", beta[1] - half, beta[1] + half)"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-67"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（C5）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('C5', [\n",
    "    ('b₁（喙长的系数）', lambda: beta[1], 2, '0e1a3f9422',\n",
    "     '设计矩阵应有 4 列：截距、喙长、两个物种指示变量；y 是喙深'),\n",
    "    ('b₂、b₃（两个指示变量的系数）', lambda: beta[2:4], 2, '87d97d75cf',\n",
    "     '第 3 列是帽带企鹅的指示变量，第 4 列是巴布亚企鹅；阿德利企鹅是参照组，不单独设列',\n",
    "     {'8d10187c0e': '两个指示变量的顺序反了：先帽带企鹅，再巴布亚企鹅'}),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-68"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（C5）**\n",
    "\n",
    "b₁ = 0.20 mm/mm（95% CI 0.17–0.23），为正，落在 C2 三个物种内斜率（0.18–0.22）之间，\n",
    "与合并斜率（−0.085）方向相反。加入物种指示变量后，每个物种有自己的截距，\n",
    "b₁ 估计的是“同一物种内”喙长每增加 1 mm 喙深的平均变化（假定三个物种斜率相同）。\n",
    "这正是把物种作为固定效应的含义。"
   ],
   "id": "cell-69"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### C6　挑战题：再把性别分开\n",
    "\n",
    "雄性在四个性状上都大于雌性，所以物种内的回归也混入了两性差异。在 6 个物种 × 性别组合内\n",
    "分别拟合体重对鳍长、喙深对喙长，回答：\n",
    "\n",
    "1. 体重对鳍长的斜率和 R² 落在什么范围？有几个组合的斜率和 R² 都低于该物种合并两性时的值？\n",
    "2. 喙深对喙长的关系在几个组合中 P < 0.05？阿德利企鹅雄性的斜率是多少？"
   ],
   "id": "cell-70"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "sx_mass = fit_table(sexed, \"flipper_length_mm\", \"body_mass_g\", by=[\"species\", \"sex\"])\n",
    "sx_bill = fit_table(sexed, \"bill_length_mm\", \"bill_depth_mm\", by=[\"species\", \"sex\"])\n",
    "print(sx_mass.round(3))\n",
    "print(sx_bill.round(3))\n",
    "lower = sum((sx_mass.loc[(sp, sx), \"slope\"] < mass.loc[sp, \"slope\"]) and (sx_mass.loc[(sp, sx), \"r2\"] < mass.loc[sp, \"r2\"])\n",
    "            for sp, sx in sx_mass.index)\n",
    "record(\"sx_c_slope_range\", sx_mass[\"slope\"].min(), sx_mass[\"slope\"].max())\n",
    "record(\"sx_c_r2_range\", sx_mass[\"r2\"].min(), sx_mass[\"r2\"].max())\n",
    "record(\"sx_c_lower\", lower)\n",
    "record(\"sx_a_sig\", (sx_bill[\"p\"] < 0.05).sum())\n",
    "record(\"sx_a_adm_slope\", sx_bill.loc[(\"Adelie\", \"male\"), \"slope\"])\n",
    "record(\"sx_a_adm_p\", sx_bill.loc[(\"Adelie\", \"male\"), \"p\"])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-71"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 自查（C6）：运行后 ✓ 表示通过，✗ 后面是提示。\n",
    "check('C6', [\n",
    "    ('体重对鳍长：6 组斜率的最小值和最大值', lambda: [sx_mass['slope'].min(), sx_mass['slope'].max()], 1, '10e5686261',\n",
    "     '用 sexed，by=[\"species\", \"sex\"]，x 是鳍长、y 是体重'),\n",
    "    ('体重对鳍长：6 组 R² 的最小值和最大值', lambda: [sx_mass['r2'].min(), sx_mass['r2'].max()], 2, '751fb6f8df',\n",
    "     'R² 来自 sx_mass 的 r2 列'),\n",
    "    ('喙深对喙长：P < 0.05 的组合数', lambda: (sx_bill['p'] < 0.05).sum(), 0, '4e07408562',\n",
    "     'sx_bill 是喙深对喙长（x 是喙长）；数 p 列中小于 0.05 的行'),\n",
    "    ('阿德利企鹅雄性的喙部斜率', lambda: sx_bill.loc[('Adelie', 'male'), 'slope'], 3, '95619a4916',\n",
    "     '在 sx_bill 中找 (Adelie, male) 这一行'),\n",
    "])"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-72"
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**参考答案（C6）**\n",
    "\n",
    "1. 斜率 12.0–40.3 g/mm，R² 0.06–0.44；6 个组合中有 5 个的斜率和 R² 都低于该物种合并两性时的值。\n",
    "2. 只有 3 个组合 P < 0.05；阿德利企鹅雄性的斜率为 −0.017 mm/mm（P = 0.75），接近于零。\n",
    "   所以第 3 周课上的物种内斜率不宜直接解释为个体水平的形态协变，模型中还应纳入性别（论文 4.3 节）。"
   ],
   "id": "cell-73"
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "if SELFCHECK_FAILED:\n",
    "    raise SystemExit(\"参考答案没有通过这些自查：\" + \"；\".join(SELFCHECK_FAILED))\n",
    "pd.DataFrame(\n",
    "    [(k, \";\".join(repr(v) for v in vals)) for k, vals in ANSWERS.items()], columns=[\"key\", \"values\"]\n",
    ").to_csv(OUT / \"answers.csv\", index=False)\n",
    "print(f\"已写入 {len(ANSWERS)} 个答案数值和图到 {OUT}\")"
   ],
   "execution_count": null,
   "outputs": [],
   "id": "cell-74"
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}
