报告数据加载失败，请刷新重试。 Report data could not be loaded; please reload.
消费者问卷 · 经济分析 · 机器学习CONSUMER SURVEY · ECONOMETRICS · MACHINE LEARNING
基于 622 份问卷，分析消费者对智能驾驶安全性、功能价值的评价与支付溢价意愿的关系，并比较有序 Logit 和随机森林的预测表现。An analysis of 622 survey responses examining how ratings of driving safety and feature value relate to willingness to pay a premium. Ordered logit and random forest models are also compared using cross-validation.
研究摘要STUDY SUMMARY
622 份问卷中的技术认知、功能价值与支付溢价意愿。Technology perceptions, feature value and willingness to pay a premium in 622 survey responses.
样本包含 622 份问卷。结果变量为支付溢价意愿，技术认知取 Q15、Q16 的平均值，功能价值取 Q22–Q25 的平均值。The sample contains 622 survey responses. The outcome is willingness to pay a premium. Technology recognition is the mean of Q15 and Q16; feature value is the mean of Q22–Q25.
Q21：愿意为智能驾驶功能支付溢价，保留 1–5 的有序等级。Q21: willingness to pay a premium for intelligent-driving functions, retained as an ordered 1–5 outcome.
主规格为 Q15、Q16 的完整回答均值，代表重要性和安全性认知。The primary proxy is the complete-row mean of Q15 and Q16, representing importance and safety recognition.
主规格为 Q22–Q25 的完整回答均值，代表具体辅助驾驶功能的价值认知。The primary proxy is the complete-row mean of Q22–Q25, representing value recognition for concrete assisted-driving functions.
Y 为 Q21 的支付溢价意愿，保留 1–5 分等级。T 由 Q15“功能重要性”和 Q16“提高安全性”取均值；V 由 Q22–Q25 的自适应巡航、车道保持、自动泊车和拥堵辅助四项取均值。它们分别概括一般技术评价和具体功能评价。M1 为 Q18 的驾驶乐趣评价，M2 为 Q19 的出行效率评价，均为单题指标。Y is the ordered 1–5 response to Q21 on willingness to pay a premium. T averages Q15 (importance) and Q16 (safety). V averages Q22–Q25 on adaptive cruise control, lane keeping, parking assistance and traffic-jam assistance. M1 is Q18 (driving pleasure); M2 is Q19 (travel efficiency), both single-item measures.
从原始题项读取数据，不使用预先生成的处理列。组合指标要求所用题项回答完整；各分析按所需变量的完整记录估计。性别、年龄、学历、收入、驾龄和驾驶频率按类别编码。T 和 V 的 Cronbach α 分别为 0.858、0.871，反映题项内部一致性，但不能单独证明测量效度。V 的原题涉及功能对购买意愿的影响，因此与 Y 的含义有接近之处。The pipeline reads original items rather than pre-generated processed columns. Composites require complete item responses, and each analysis uses complete records for its required variables. Six demographic and driving controls are category-coded. Cronbach’s alpha is 0.858 for T and 0.871 for V; these assess internal consistency, not construct validity. V’s questions already refer to purchase intention, so its meaning overlaps with Y.
主模型使用有序 Logit，并控制性别、年龄、学历、收入、驾龄和驾驶频率。OR 大于 1 表示更高的 T/V 与更高购买溢价意愿的排序相关。The primary ordered-logit model controls for gender, age, education, income, driving experience, and driving frequency. An OR above 1 indicates an association with a higher willingness-to-pay category.
研究问题是：在控制人口特征与驾驶经历后，技术认知和功能价值是否仍与支付溢价意愿有关？Y 有五个有序等级，因此使用有序 Logit，同时纳入 T、V 和六类控制变量。T、V 保持原来的 1–5 分尺度，不做标准化；表中优势比对应指标增加 1 分。The question is whether T and V remain associated with willingness to pay after accounting for demographics and driving history. Ordered logit models the five ordered Y categories using T, V and six categorical controls. T and V remain on their original 1–5 scales; odds ratios refer to a one-point increase.
点估计来自 primary / controlled；虚线为 OR=1。置信区间见下表。Estimates are from primary / controlled; the dashed line is OR=1. Confidence intervals are shown below.
| 变量Variable | OR | 95% CI | p | N |
|---|
优势比表示控制其他变量后的关联，因果方向无法由本次问卷确定。The odds ratios describe adjusted associations. The survey does not establish their causal direction.
在 622 份记录中，T 的 OR 为 2.08（95% CI：1.62–2.68），V 为 3.83（2.84–5.16），两项 p 值均小于 0.001。按模型的比例优势假设，其他变量不变时，T 增加 1 分对应跨越任一 Y 等级阈值的优势乘以约 2.08。OR 不是概率倍数，也不是支付金额的增幅。V 的点估计更大，但其题项与 Y 更接近，不能仅凭 OR 排定真实影响大小。Across 622 records, T has OR 2.08 (95% CI 1.62–2.68) and V has OR 3.83 (2.84–5.16), both p<0.001. Under proportional odds, a one-point increase in T multiplies the odds of exceeding any Y threshold by about 2.08, holding other variables fixed. This is neither a probability ratio nor a change in payment amount. V’s larger estimate must also be read in light of its conceptual overlap with Y.
五条间接路径均使用 5,000 次 Bootstrap 估计 95% 置信区间。结果反映本样本中的间接关联；因果中介关系需要额外的研究设计和假设支持。The five indirect paths use 5,000 bootstrap draws to estimate 95% confidence intervals. The estimates describe indirect associations; causal mediation would require additional design and identification assumptions.
分别估计 T→M1→Y、T→M2→Y、V→M1→Y、V→M2→Y 和 T→V→Y。每条路径单独拟合：先用 X 和控制变量解释中间变量 M，得到 a；再用 X、M 和控制变量解释 Y，得到 b；间接关联取 a×b。两个方程均为含截距的 OLS，把评分作为连续变量近似。按受访者有放回抽样 5,000 次，使用乘积估计分布的 2.5% 和 97.5% 分位数构造区间。Five paths are fitted separately: T→M1→Y, T→M2→Y, V→M1→Y, V→M2→Y and T→V→Y. For each, OLS regresses M on X and controls to estimate a, and Y on X, M and controls to estimate b. Both equations include intercepts and approximate ratings as continuous. The product a×b is bootstrapped by resampling respondents with replacement 5,000 times; percentile intervals use the 2.5th and 97.5th percentiles.
点为间接关联估计，横线为 95% Bootstrap 区间；每条完整路径均以 Y 为终点。Points show indirect estimates; horizontal lines show 95% bootstrap intervals. Every full path ends at Y.
| 路径Path | ab | 95% CI | 有效/总数Valid / total |
|---|
中介方程是探索性的 OLS 近似，不能与有序 Logit 系数直接相乘。Mediator equations are exploratory OLS approximations and are not multiplied by ordered-logit coefficients.
五条路径的区间均高于 0。经驾驶乐趣的估计分别为 0.256 和 0.257，经出行效率为 0.180 和 0.163；T→V→Y 为 0.351。每条路径均得到 5,000 次有效估计。这些数值属于独立的线性路径模型，不与前节的 OR 相加或相乘，也不能加总为“解释比例”。同一时点的问卷无法确认变量发生顺序，因此这里报告探索性间接关联。All five intervals lie above zero. Products through driving pleasure are 0.256 and 0.257, through travel efficiency 0.180 and 0.163, and through T→V→Y 0.351. Each path has 5,000 valid bootstrap estimates. These are separate linear path models; their estimates cannot be combined with the ordered-logit ORs or summed into an explained proportion. The cross-sectional survey cannot establish temporal ordering.
性别、年龄、收入、驾龄和驾驶频率分别进行嵌套有序 Logit LR 检验。五项检验的 p 值使用 Holm 方法进行多重比较校正。Gender, age, income, driving experience, and driving frequency are tested with nested ordered-logit LR comparisons. The five p-values are adjusted for multiple comparisons using the Holm method.
检验技术认知、功能价值与 Y 的关系是否随性别、年龄、收入、驾龄或驾驶频率而变化。每次比较一个含分组主项的有序 Logit 与加入 T×分组、V×分组交互项的模型，使用似然比（LR）检验联合判断斜率差异。五个分组分别检验，再用 Holm 方法校正多重比较。表中 k 为该次分析的组数。For each grouping, a restricted ordered logit containing group main effects is compared with a model adding T-by-group and V-by-group interactions. The likelihood-ratio test jointly assesses slope differences. Five grouping tests are adjusted using Holm’s method. The table’s k is the number of groups.
虚线为 0.05 判断线。本次五项比较校正后均未低于该线。The dashed line marks 0.05. None of the five adjusted comparisons falls below it.
| 分组Grouping | k | LR | p | Holm p |
|---|
五项检验经 Holm 校正后均未达到 0.05 显著性水平。None of the five tests meets the 0.05 significance threshold after Holm adjustment.
性别和年龄的原始 p 值分别为 0.013、0.042，校正后为 0.063、0.170；其余分组的校正值也高于 0.05。因此，本次分析没有得到经多重比较校正后仍显著的人群差异证据。这不等于证明各组完全相同，也不支持依据个别未校正结果给出针对某个人群的结论。Raw p-values for gender and age are 0.013 and 0.042, increasing to 0.063 and 0.170 after adjustment. All other adjusted values also exceed 0.05. The analysis therefore finds no group differences meeting the adjusted significance threshold. This does not establish that groups are identical.
五折折外评估显示，加入 T、V、M1、M2 后模型明显超过多数类基线；在本次运行中，有序 Logit 的 QWK 高于随机森林。Five-fold held-out evaluation shows a clear improvement over the majority baseline after adding T, V, M1, and M2. In this run, ordered logit has a higher QWK than the random forest.
预测目标仍为 Y 的五个等级。比较三组输入：仅控制变量、控制变量＋T/V、控制变量＋T/V/M1/M2；每组均比较多数类基线、有序 Logit 和随机森林。使用随机种子 42 的五折分层交叉验证，每折约四份数据训练、一份评价。随机森林固定为 200 棵树、最大深度 6；表中报告五折指标的平均值。The target remains the five Y categories. Three feature sets—controls, controls plus T/V, and controls plus T/V/M1/M2—are evaluated with a majority baseline, ordered logit and random forest. Five-fold stratified cross-validation uses seed 42, training on four folds and evaluating on the remaining fold. Random forests use 200 trees and maximum depth 6. The table reports fold means.
多数类基线 QWK 为 0；扩展有序 Logit 约为 0.685。The majority baseline has QWK 0; extended ordered logit is about 0.685.
| 特征集Feature set | 模型Model | Acc. | F1 | QWK | MAE |
|---|
图中展示扩展随机森林对 P(Y≥4) 的全部 23 个输入特征，按平均绝对 SHAP 排序。数值表示贡献大小，不表示作用方向或因果关系。The chart shows all 23 inputs used by the extended forest for P(Y≥4), ranked by mean absolute SHAP. Magnitude does not indicate direction or causality.
Accuracy 统计等级完全预测正确的比例；F1 为各等级 F1 的宏平均；QWK 为二次加权 Kappa，考虑等级误差并校正随机一致性；MAE 为平均相差几个等级。仅控制变量的预测接近基线。加入 T/V 后，有序 Logit 的 QWK 为 0.650；再加入 M1/M2 后为 0.685，准确率为 54.5%。同样扩展输入下，随机森林为 0.619 和 53.5%，本次结果没有显示它优于有序 Logit。Accuracy measures exact category matches; F1 is macro-averaged across categories; QWK is quadratic-weighted kappa; MAE is the mean absolute category error. Controls alone perform near the baseline. Ordered-logit QWK reaches 0.650 with T/V and 0.685 with M1/M2 added, with accuracy 54.5%. The extended random forest scores 0.619 and 53.5%, so it does not outperform ordered logit in this run.
对扩展随机森林预测“Y≥4”的概率计算 SHAP，在每折未参与训练的样本上解释，再按平均绝对值汇总排序。功能价值、技术认知、驾驶乐趣和出行效率排在前四。图中也列出分类控制变量编码后的各项，所以共有 23 个输入特征。排序反映模型预测依赖程度，不表示因果效应；交叉验证将每折训练样本与测试样本分开。SHAP explains the extended forest’s predicted probability of Y≥4 on each held-out fold. Features are ranked by mean absolute attribution, with V, T, M1 and M2 first. Category-coded controls bring the total to 23 inputs. The ranking describes predictive attribution, not causal effects. Each fold keeps its test sample outside training.
数据来自 622 份消费者问卷。完整分析代码与双语 Notebook 见项目仓库。The study uses 622 consumer survey responses. Analysis code and bilingual notebooks are available in the repository.
题项、组合变量、控制变量和缺失值处理规则见 src/data_schema.py 与 src/config.py。Item mappings, composites, controls and missing-value rules are defined in src/data_schema.py and src/config.py.
正文统一使用当前主规格：T 由 Q15、Q16 构成，V 由 Q22–Q25 构成，控制变量按类别编码。The report uses the current primary specification: T uses Q15 and Q16, V uses Q22–Q25, and controls are category-coded.
有序 Logit、随机森林和多数类基线使用同一组分层折。Ordered logit, random forest, and the majority baseline share the same stratified folds.
结果反映这份横截面问卷中的关联和预测表现，不能确定因果关系。Results describe associations and predictive performance in this cross-sectional survey. They do not establish causality.