Flaherty Brian P, Shono Yusuke
University of Washington.
Huntsman Cancer Institute and University of Utah.
J Surv Stat Methodol. 2021 Apr;9(2):231-256. doi: 10.1093/jssam/smaa047. Epub 2021 Mar 1.
Scientists use latent class (LC) models to identify subgroups in heterogeneous data. LC models reduce an item set to a latent variable and estimate measurement error. Researchers typically use unrestricted LC models, which have many measurement estimates, yet scientific interest primarily concerns the classes. We present highly restricted LC measurement models as an alternate method of operationalization. MACREM (Many Classes, Restricted Measurement) models have a larger number of latent classes than a typical unrestricted model, but many fewer measurement estimates. Goals of this approach include producing more interpretable classes and better measurement error estimates. Parameter constraints accomplish this structuring. We present unrestricted and MACREM model results using data on activities of daily living (ADL) from a national survey ( = 3485). We compare a four class unrestricted model with a 14 class MACREM model. The four class unrestricted model approximates a dimension of functional limitation. The 14 class model includes unordered classes at lower levels of limitation, but ordered classes at higher levels of limitation. In contrast to the four class model, all measurement error rates are reasonably small in the 14 class model. The four class model fits the data better, but the 14 class model is more parsimonious (43 vs. 25 parameters). Three covariates reveal specific associations with MACREM classes. In multinomial logistic regression models with a no limitation class as the reference class, past 12 month diabetes only distinguishes low limitation classes that include cutting one's own toenails as a limitation. It does not distinguish low limitation classes characterized by other common limitations. Past 12 month asthma and current disability status perform similarly, but for heavy housework and walking limitation classes, respectively. These limitation specific covariate associations are not apparent in the unrestricted model analyses. Identifying such connections could provide useful information to advance theory and intervention efforts.
科学家使用潜在类别（LC）模型来识别异质数据中的亚组。LC模型将项目集简化为一个潜在变量并估计测量误差。研究人员通常使用无限制的LC模型，这种模型有许多测量估计值，但科学兴趣主要集中在类别上。我们提出高度受限的LC测量模型作为一种替代的操作方法。MACREM（多类别、受限测量）模型的潜在类别数量比典型的无限制模型更多，但测量估计值要少得多。这种方法的目标包括产生更具可解释性的类别和更好的测量误差估计。参数约束实现了这种结构。我们使用一项全国性调查（n = 3485）中关于日常生活活动（ADL）的数据展示了无限制和MACREM模型的结果。我们将一个四类无限制模型与一个14类MACREM模型进行了比较。四类无限制模型近似于功能受限的一个维度。14类模型在较低受限水平包含无序类别，但在较高受限水平包含有序类别。与四类模型相比，14类模型中的所有测量误差率都相当小。四类模型对数据的拟合更好，但14类模型更简约（43个参数对25个参数）。三个协变量揭示了与MACREM类别的特定关联。在以无限制类别作为参考类别的多项逻辑回归模型中，过去12个月患糖尿病仅能区分包括自己剪脚趾甲作为一项限制的低受限类别。它无法区分以其他常见限制为特征的低受限类别。过去12个月患哮喘和当前残疾状况分别对繁重家务和行走受限类别有类似作用。这些特定于受限情况的协变量关联在无限制模型分析中并不明显。识别此类联系可为推进理论和干预工作提供有用信息。
