Regularized cross-sectional network modeling with missing data: A comparison of methods

Many applications of network modeling involve cross-sectional data of psychologicalvariables (e.g., symptoms for psychological disorders), and analyses are often conductedusing a regularized Gaussian graphical model (GGM) employing a lasso, alsoknown as the graphical lasso or glasso. Appropriate methodology for handling missingdata is underdeveloped while using glasso, precluding the use of planned missingdata designs to reduce participant fatigue. In this research, we compare three approachesto handling missing data with glasso. The first resembles a two-stage estimationapproach—–borrowed from the covariance structure modeling literature—–wherebya saturated covariance matrix among the items is estimated prior to using glasso. Thesecond and third approaches use glasso and the expectation-maximization (EM) algorithmin a single stage and either use EBIC or cross-validation for tuning parameterselection. We compared these approaches in a simulation study with a variety of samplesizes, proportions of missing data, and network saturation. An example with datafrom the Patient Reported Outcomes Measurement Information System is also provided.The EM algorithm with cross-validation performed best, but all methods appeared to beviable strategies under larger samples and with less missing data.