计算McFadden's R-squared遇ValueError：长度需1却给定118649的解决咨询

问题背景 #

尝试手动计算逻辑回归的McFadden's R-squared时触发维度不匹配错误，相关代码及错误信息如下：

原代码 #

# Calculate McFadden's R-squared
log_likelihood_model = -np.sum(y * np.log(y_pred_prob) + (1 - y) * np.log(1 - y_pred_prob))
y_pred_null = np.mean(y)  # Predicting the mean of y as a null model
log_likelihood_null = -np.sum(y * np.log(y_pred_null) + (1 - y) * np.log(1 - y_pred_null))
mcfaddens_r2 = 1 - (log_likelihood_model / log_likelihood_null)

print("McFadden's R-squared:", mcfaddens_r2)

错误输出 #

/usr/local/lib/python3.10/dist-packages/pandas/core/ops/__init__.py in to_series(right)
    237         else:
    238             if len(left.columns) != len(right):
--> 239                 raise ValueError(
    240                     msg.format(req_len=len(left.columns), given_len=len(right))
    241                 )

ValueError: Unable to coerce to Series, length must be 1: given 118649

错误原因 #

错误核心是y为pandas Series/DataFrame类型，而y_pred_null是单个标量（均值）。pandas执行元素级运算时会严格校验维度：若左侧是多列DataFrame，要求右侧长度等于列数，但此处y_pred_null长度为1，与y的118649条样本量冲突，触发广播逻辑错误。

修复方案 #

有两种简洁的修复方式：

方案1：将y转换为numpy数组 #

绕过pandas的维度校验逻辑，直接用numpy数组计算：

import numpy as np

# 若y是pandas对象，先转为numpy数组
y = np.array(y)

# Calculate McFadden's R-squared
log_likelihood_model = -np.sum(y * np.log(y_pred_prob) + (1 - y) * np.log(1 - y_pred_prob))
y_pred_null = np.mean(y)
# numpy会自动将标量广播为与y匹配的形状
log_likelihood_null = -np.sum(y * np.log(y_pred_null) + (1 - y) * np.log(1 - y_pred_null))
mcfaddens_r2 = 1 - (log_likelihood_model / log_likelihood_null)

print("McFadden's R-squared:", mcfaddens_r2)

方案2：生成与y同形状的空模型预测值 #

让y_pred_null的维度与y完全一致，避免广播冲突：

import numpy as np

# Calculate McFadden's R-squared
log_likelihood_model = -np.sum(y * np.log(y_pred_prob) + (1 - y) * np.log(1 - y_pred_prob))
# 生成和y形状完全相同的数组，每个元素都是y的均值
y_pred_null = np.full_like(y, np.mean(y))
log_likelihood_null = -np.sum(y * np.log(y_pred_null) + (1 - y) * np.log(1 - y_pred_null))
mcfaddens_r2 = 1 - (log_likelihood_model / log_likelihood_null)

print("McFadden's R-squared:", mcfaddens_r2)

额外优化提示 #

若使用scikit-learn逻辑回归模型，可直接调用log_loss函数简化计算：

from sklearn.metrics import log_loss

log_likelihood_model = -log_loss(y, y_pred_prob, normalize=False)
log_likelihood_null = -log_loss(y, [np.mean(y)]*len(y), normalize=False)
mcfaddens_r2 = 1 - (log_likelihood_model / log_likelihood_null)

内容的提问来源于stack exchange，提问作者Kezz