Open Dataset for Meta-Analysis of AI-Assisted Programming Learning and Students’ Computational Thinking

This repository contains the data and supporting materials for the meta-analysis titled The Effect of Artificial Intelligence-Assisted Programming Learning on Students’ Computational Thinking. The main research hypothesis was that artificial intelligence-assisted programming learning has a positive effect on students’ computational thinking (CT), and that the magnitude of this effect may vary across study and instructional characteristics. The dataset includes study-level information extracted from 19 empirical studies and 19 independent effect sizes. The coded variables include title, publication year, authors, educational level, teaching strategy, programming environment, sample size, intervention duration, CT measurement tool, instructional function of AI, type of AI, methodological quality scores, and the effect-size data used for synthesis. According to the study protocol, two researchers coded all included studies independently, and inter-coder reliability was high (Cohen’s kappa = 0.902). Methodological quality was assessed using the Kmet et al. checklist, and all included studies were rated as high quality, with scores ranging from 0.818 to 0.917; inter-rater reliability for quality assessment was 0.925. The repository also contains supporting documentation to improve transparency and reusability, including the coding sheet for included studies, the list of full-text articles excluded after eligibility assessment, the PRISMA 2020 checklist, and the PRISMA flow diagram. These files document how studies were identified, screened, assessed for eligibility, coded, and included in the final synthesis. The meta-analysis was conducted in Comprehensive Meta-Analysis 3.0 using a random-effects model because substantial variation across studies was expected in terms of participants, interventions, and educational contexts. Hedges’ g was used as the effect-size index. Publication bias was assessed through funnel plot inspection, Begg’s test, Egger’s test, fail-safe N, and trim-and-fill analysis. The results suggested that publication bias was unlikely to materially affect the findings. The pooled overall effect was large and positive (Hedges’ g = 0.949, 95% CI [0.650, 1.247], p < .001), indicating that AI-assisted programming learning significantly improved students’ CT. However, heterogeneity was substantial (Q = 129.398, I² = 86.1%), so moderator analyses were conducted. Significant subgroup differences were found for CT measurement tool and instructional function of AI, whereas educational level, teaching strategy, programming environment, sample size, intervention duration, and AI type did not show significant between-group differences. A leave-one-out sensitivity analysis further showed that the overall result was robust. This dataset can be used to verify the reported meta-analytic results, inspect coding decisions, reproduce subgroup analyses, and understand the study selection process.本仓库包含题为《人工智能辅助编程学习对学生计算思维的影响》的元分析所用数据与补充材料。本研究的核心假设为，人工智能辅助编程学习对学生计算思维（computational thinking，CT）具有正向影响，且该影响的效应量大小可能因研究特征与教学特征而异。本数据集涵盖从19项实证研究中提取的研究层面信息，共包含19个独立效应量。经编码的变量包括：文献标题、发表年份、作者、教育阶段、教学策略、编程环境、样本量、干预时长、计算思维测量工具、人工智能的教学功能、人工智能类型、方法学质量评分，以及用于效应量合成的效应量数据。根据研究方案，两名研究者独立对所有纳入研究进行编码，编码者间信度（inter-coder reliability）良好（科恩kappa系数（Cohen’s kappa）=0.902）。本研究采用Kmet等人编制的检查表对文献方法学质量进行评估，所有纳入研究均被评为高质量，评分区间为0.818至0.917；质量评估的评分者间信度为0.925。本仓库同时包含提升研究透明度与可复用性的辅助文档，包括纳入研究的编码表、经资格审查后排除的全文文献清单、PRISMA 2020检查表，以及PRISMA流程图。这些文件完整记录了研究的筛选、资格评估、编码及纳入最终合成分析的全过程。本次元分析采用随机效应模型（random-effects model），在Comprehensive Meta-Analysis 3.0软件中完成，因预期不同研究在参与者、干预措施与教育场景间存在显著异质性（heterogeneity）。本研究选用赫奇斯g值（Hedges’ g）作为效应量指标。通过漏斗图检查（funnel plot inspection）、Begg检验（Begg’s test）、Egger检验（Egger’s test）、失效安全N值（fail-safe N）以及修剪填充法（trim-and-fill analysis）对发表偏倚（publication bias）进行评估，结果显示发表偏倚未对研究结论产生实质性影响。合并后的总体效应量为正向且效应显著（赫奇斯g值=0.949，95%置信区间（95% CI）[0.650, 1.247]，p<.001），表明人工智能辅助编程学习可显著提升学生的计算思维。但研究间异质性较高（Q=129.398，I²=86.1%），因此开展了调节效应分析（moderator analyses）。分析发现仅计算思维测量工具与人工智能的教学功能存在显著的亚组差异，而教育阶段、教学策略、编程环境、样本量、干预时长及人工智能类型均未表现出显著的组间差异。留一法敏感性分析（leave-one-out sensitivity analysis）进一步证实，本研究的总体结果具有稳健性。本数据集可用于验证已报道的元分析结果、核查编码决策、复现亚组分析，以及理解研究筛选的完整流程。