Psychometricians working on International Large Scale Assessments (ILSAs) typically specify latent ability factors with distinct and correlated constructs for test domains, such as reading, mathematics and science. A construct for general ability is not specified. However, several country-specific studies conclude that ILSAs largely reflect general ability. We extend such studies and examine the dimensionality of the 2018 PISA assessment in 33 OECD countries examining three models: three-dimensional IRT model, the bifactor IRT model and the bifactor (S-1) IRT model. A four-tiered approach was adopted. First, models were compared using an information criterion (AIC). Second, the correlations from the multidimensional model were estimated to assess in which countries the three dimensions are sufficient discriminant validity. Third, a variety of bifactor indices were utilized to establish the explanatory power and reliabilities of the latent dimensions generated by the three models. Finally, the statistical relationships between the latent factors derived from the three models and educationally relevant covariates were estimated. The bifactor model fits the data better than standard multidimensional model or S-1 model in every country investigated. The correlations in the correlated factor model are above 0.8 in all 33 countries. The symmetrical bifactor general ability model shows that 80%, or more, of the common variance in student responses to the PISA instruments is accounted for by a general ability factor. On average, 27% of variance in the mathematics items is independent of the general factor and can be attributed to a specific mathematics ability factor. The respective estimates for reading are 12% and science is 17%. Relationships for selected covariates with the PISA domains follow the same pattern as general ability in the bifactor model.

从事国际大规模评估 (ILSA) 工作的心理测量学家通常会为阅读、数学和科学等测试领域指定具有不同且相关结构的潜在能力因素。未指定一般能力的构造。然而，一些针对特定国家的研究得出结论，ILSA 在很大程度上反映了一般能力。我们扩展了此类研究并检查了 33 个经合组织国家 2018 年 PISA 评估的维度，检查了三个模型：三维 IRT 模型、双因子 IRT 模型和双因子 (S-1) IRT 模型。采用了四级方法。首先，使用信息标准 (AIC) 比较模型。其次，估计来自多维模型的相关性，以评估在哪些国家这三个维度具有足够的区分效度。第三，各种双因素指数被用来确定三个模型产生的潜在维度的解释力和可靠性。最后，估计了来自三个模型的潜在因素与教育相关协变量之间的统计关系。在所调查的每个国家，双因子模型比标准多维模型或 S-1 模型更适合数据。所有 33 个国家的相关因子模型中的相关性均高于 0.8。对称双因素综合能力模型表明，学生对 PISA 工具的反应中 80% 或更多的共同方差是由综合能力因素解释的。平均而言，数学项目中 27% 的方差与一般因素无关，可归因于特定的数学能力因素。阅读分别估计为 12%，科学为 17%。所选协变量与 PISA 域的关系遵循与双因子模型中的一般能力相同的模式。