前后干预Likert量表调查分析：样本量不匹配时的统计方法选择

First, let's anchor on your core constraint: your pre-intervention (n=14) and post-intervention (n=18) samples are unmatched (independent)—there's no paired link between individual respondents across the two groups. This immediately rules out tests designed for paired data:

• McNemar's Test: Exclusively for paired binary data (same subjects measured before/after). Your independent samples make this test irrelevant here.
• Paired t-test: Requires continuous, paired measurements from identical subjects. Not applicable to your unmatched groups.
• Wilcoxon Signed-Rank Test: Another paired test, for non-normal ordinal data. Doesn't work for independent samples.

Now, let's cover the appropriate methods for your goal: comparing the proportion of "agree/strongly agree" responses (or full Likert scale distributions) between two independent groups.

Recommended Methods #

1. Chi-Square Test of Independence (or Fisher's Exact Test)

   • Use this for each individual Likert item after converting responses to binary: agree/strongly agree = 1, neutral/disagree/strongly disagree = 0. It compares the binary outcome distribution across your pre and post groups.
   • Critical note for small samples: If any expected cell count in your 2x2 contingency table is <5, switch to Fisher's Exact Test—it’s more accurate for small, sparse datasets like yours.

2. Mann-Whitney U Test (Wilcoxon Rank-Sum Test)

   • If you want to retain information from the full Likert scale (instead of collapsing to binary), this non-parametric test compares the ordinal response distributions between your two independent groups. It’s robust to small sample sizes and doesn’t require normality assumptions.

Key Additional Considerations #

• Multiple Comparisons: Running 13 separate tests (one per item) increases your risk of Type I errors (false positives). Adjust your significance level using methods like Bonferroni correction or FDR (False Discovery Rate) to account for this.
• Low Statistical Power: Small sample sizes mean your tests will have limited power to detect true effects. Interpret non-significant results cautiously—they may reflect low power rather than no actual intervention effect.

Example Workflow (R Syntax) #

For a single binary-collapsed item:

# Create 2x2 contingency table
response_table <- matrix(c(pre_agree, post_agree, pre_not_agree, post_not_agree), nrow=2)

# Check expected counts; use Fisher's Exact if any <5
if (all(chisq.test(response_table)$expected >=5)) {
  chisq.test(response_table)
} else {
  fisher.test(response_table)
}

For full Likert scale analysis:

# pre_scores and post_scores are vectors of ordinal Likert values (1-5)
wilcox.test(pre_scores, post_scores, paired = FALSE)

内容的提问来源于stack exchange，提问作者user194003