如何通过循环正确替换HADS量表的缺失值？

背景与需求 #

我正在撰写医学科研论文，使用HADS量表评估患者的焦虑与抑郁情况。该量表包含14个条目，分为HADS-D（抑郁子量表）和HADS-A（焦虑子量表）两个子量表，各含7个条目，每个条目分值为0-3分。缺失数据需按量表手册规则处理：

• 若单个子量表缺失条目超过1个，则删除该观测；
• 若单个子量表仅缺失1个条目，则用该子量表其余6个条目的均值填补缺失值。

变量对应关系 #

• HADS-D子量表（总分对应hads_anx_score）：hads_tense_rec、hads_glad_rec、hads_omen_rec、hads_laugh_rec、hads_trouble_rec、hads_happy_rec、hads_relax_rec
• HADS-A子量表（总分对应hads_depr_score）：hads_limited_rec、hads_scary_rec、hads_looks_rec、hads_restless_rec、hads_future_rec、hads_panic_rec、hads_enjoy_rec

处理步骤与问题 #

我将处理流程拆解为6步，但第5步的填补循环未能成功替换缺失值，仍存在如hads_limited_rec == .的缺失数据：

1. 初始化子量表得分变量
2. 创建is_missing_前缀变量标记缺失值
3. 统计每个子量表的缺失条目数
4. 删除不符合要求的观测
5. 填补单个缺失条目（问题所在）
6. 计算子量表总分

原代码 #

*STEP 1: Initialize the HADS-A and HADS-D subscales
gen hads_anx_score = .
gen hads_depr_score = .

* STEP 2:Loop over each observation
foreach var in hads_tense_rec hads_glad_rec hads_omen_rec hads_laugh_rec hads_trouble_rec hads_happy_rec hads_relax_rec hads_limited_rec hads_scary_rec hads_looks_rec hads_restless_rec hads_future_rec hads_panic_rec hads_enjoy_rec {
    gen is_missing_`var' = missing(`var')
}

* STEP 3: Calculate the number of missing items per subscale
egen missing_hads_anx = rowtotal(is_missing_hads_tense_rec is_missing_hads_glad_rec is_missing_hads_omen_rec is_missing_hads_laugh_rec is_missing_hads_trouble_rec is_missing_hads_happy_rec is_missing_hads_relax_rec)

egen missing_hads_depr = rowtotal(is_missing_hads_limited_rec is_missing_hads_scary_rec is_missing_hads_looks_rec is_missing_hads_restless_rec is_missing_hads_future_rec is_missing_hads_panic_rec is_missing_hads_enjoy_rec)

* STEP 4. Drop observations with more than one missing item in any subscale
drop if missing_hads_anx > 1 | missing_hads_depr > 1

**STEP 5.** Replace single missing items with the mean of the present six items
foreach var in hads_tense_rec hads_glad_rec hads_omen_rec hads_laugh_rec hads_trouble_rec hads_happy_rec hads_relax_rec {
    qui replace `var' = (hads_tense_rec + hads_glad_rec + hads_omen_rec + hads_laugh_rec + hads_trouble_rec + hads_happy_rec + hads_relax_rec - `var') / 6 if is_missing_`var' == 1 & missing_hads_anx == 1
}

foreach var in hads_limited_rec hads_scary_rec hads_looks_rec hads_restless_rec hads_future_rec hads_panic_rec hads_enjoy_rec {
    qui replace `var' = (hads_limited_rec + hads_scary_rec + hads_looks_rec + hads_restless_rec + hads_future_rec + hads_panic_rec + hads_enjoy_rec - `var') / 6 if is_missing_`var' == 1 & missing_hads_depr == 1
}

问题原因 #

原代码中，当目标变量var为缺失值时，(所有条目总和 - var)的运算结果仍为缺失值（Stata中缺失值参与任何算术运算都会返回缺失），导致无法完成填补。例如hads_limited_rec缺失时，hads_limited_rec + ... - hads_limited_rec等价于. + ... - .，结果还是缺失。

修复后的代码 #

*STEP 1: Initialize the HADS-A and HADS-D subscales
gen hads_anx_score = .
gen hads_depr_score = .

* STEP 2: Create missing value indicators
foreach var in hads_tense_rec hads_glad_rec hads_omen_rec hads_laugh_rec hads_trouble_rec hads_happy_rec hads_relax_rec ///
               hads_limited_rec hads_scary_rec hads_looks_rec hads_restless_rec hads_future_rec hads_panic_rec hads_enjoy_rec {
    gen is_missing_`var' = missing(`var')
}

* STEP 3: Calculate the number of missing items per subscale
egen missing_hads_anx = rowtotal(is_missing_hads_tense_rec is_missing_hads_glad_rec is_missing_hads_omen_rec ///
                                 is_missing_hads_laugh_rec is_missing_hads_trouble_rec is_missing_hads_happy_rec ///
                                 is_missing_hads_relax_rec)

egen missing_hads_depr = rowtotal(is_missing_hads_limited_rec is_missing_hads_scary_rec is_missing_hads_looks_rec ///
                                 is_missing_hads_restless_rec is_missing_hads_future_rec is_missing_hads_panic_rec ///
                                 is_missing_hads_enjoy_rec)

* STEP 4. Drop observations with more than one missing item in any subscale
drop if missing_hads_anx > 1 | missing_hads_depr > 1

* STEP 5. Calculate mean of non-missing items for each subscale (for observations with exactly 1 missing)
egen hads_anx_mean = rowmean(hads_tense_rec hads_glad_rec hads_omen_rec hads_laugh_rec hads_trouble_rec hads_happy_rec hads_relax_rec) if missing_hads_anx == 1
egen hads_depr_mean = rowmean(hads_limited_rec hads_scary_rec hads_looks_rec hads_restless_rec hads_future_rec hads_panic_rec hads_enjoy_rec) if missing_hads_depr == 1

* Replace missing items with the subscale mean
foreach var in hads_tense_rec hads_glad_rec hads_omen_rec hads_laugh_rec hads_trouble_rec hads_happy_rec hads_relax_rec {
    qui replace `var' = hads_anx_mean if is_missing_`var' == 1 & missing_hads_anx == 1
}

foreach var in hads_limited_rec hads_scary_rec hads_looks_rec hads_restless_rec hads_future_rec hads_panic_rec hads_enjoy_rec {
    qui replace `var' = hads_depr_mean if is_missing_`var' == 1 & missing_hads_depr == 1
}

* STEP 6. Calculate subscale totals
egen hads_anx_score = rowtotal(hads_tense_rec hads_glad_rec hads_omen_rec hads_laugh_rec hads_trouble_rec hads_happy_rec hads_relax_rec)
egen hads_depr_score = rowtotal(hads_limited_rec hads_scary_rec hads_looks_rec hads_restless_rec hads_future_rec hads_panic_rec hads_enjoy_rec)

* Optional: Drop temporary variables
drop is_missing_* missing_hads_anx missing_hads_depr hads_anx_mean hads_depr_mean

修复说明 #

1. 先用egen rowmean直接计算子量表中非缺失条目的均值（仅针对有1个缺失的观测），避免缺失值参与运算的问题；
2. 再用该均值替换对应子量表中的缺失条目；
3. 最后用rowtotal计算子量表总分，确保结果准确；
4. 可选择删除临时变量，保持数据整洁。

内容的提问来源于stack exchange，提问作者Gioni_Bletsch