[4Q2-GS-10] AI application:

[4Q2-GS-10-02] Relative Evaluation of Items Using Respondent-Specific Score Distributions in a Regional Well-Being Survey

〇Aoi Hagita¹, Yuki Takeishi¹, Ryoichiro Yamazaki¹, Yuki Yamagishi^1,2, Kazuhito Sasamoto³, Shigeki Aoki⁴, Masahiro Hashimoto^4,5,6 (1. Shizuoka Institute of Science and Technology, 2. RIKEN, 3. Shizuoka City, 4. MaOI, 5. Hosei University, 6. Institute of Science Tokyo)

Keywords:Likert scale, Information content, Questionnaire Survey

In typical Likert scale evaluations used in surveys on well-being and life satisfaction, there are few respondents who utilize all levels of the scale, and even fewer who use multiple levels uniformly; instead, many tend to repeatedly choose either the same level or the extreme levels. Therefore, by using each respondent’s final probability distribution of ratings and calculating the information content of the cumulative relative frequency for each scale level, we defined this information content as an index of ``rarity.'' This information content is additive, allowing for comparisons of totals, and the cumulative relative frequencies from the upper and lower ends of the scale can be separated as positive and negative information, respectively. One interpretation of the ranking of items based on these positive/negative indicators is as an estimated reordering from ``good to average'' (for positive) and from ``average to bad'' (for negative). In evaluation experiments using actual well-being and life satisfaction survey data, comparisons between the ranking of items by average score and the ranking by the proposed indicator were conducted, thereby revealing the relative positive and negative impressions associated with each item.