Missing Data Analysis and Multiple Imputation with Chained Equations in Social Work Research

Analyses of multivariate data are often hampered by missing values in social work research. Previous common practices that address missing data include available data approaches (e.g., listwise deletion, missing data indicators), deterministic imputation methods (e.g., mean substitution, single regression imputation, and hot deck imputation), and maximum likelihood methods (e.g., expectation maximization approach and direct ML methods). However, those practices all have limitations. For example, available data approaches tend to lower statistical power and increase biases in parameter estimations, leading to problematic conclusions about the impact of interventions and policies. Deterministic imputation methods and maximum likelihood methods may address some of the biases in parameter estimations yet underestimate the standard errors of coefficients. Multiple imputation has been recognized as the current "best practice" in handling data that are Missing at Random (MAR) or Missing Completely at Random (MCAR) yet remains underutilized among social work researchers. The purposes of this workshop are to review the common approaches of addressing missing values and to introduce multiple imputation with chained equations in social work research. This workshop will start with a brief introduction of missing data mechanisms, followed by a critical review of common practices in addressing missing values and their inherent limitations. We will then use our own research examples to demonstrate multiple imputation approach, particularly multiple imputation with chained equations, to address missing values. We will conclude with a discussion of multiple imputation use in advanced analyses in survey research and clinical trials.

Topics Review of missing data mechanisms (MCAR, MAR, and NMAR) Summary of common practices and their problems (including available data methods, deterministic imputation methods, and maximum likelihood methods) Overview and demonstration of multiple imputation in Stata (including imputation, analysis, and pooling) Additional issues for multiple imputation in survey research and clinical trials (including interaction terms, nonlinear relationships, data transformation, multilevel data, experimental designs, and propensity score analysis)

Career Level and Prerequisites Participants are expected to have basic knowledge of regression analysis using generalized linear models. Experience with Stata is recommended, but not required.

Methods and Approach We will deliver this hands-on workshop in an interactive manner. A PowerPoint presentation and data analysis demonstration will be used in conjunction with interpretations and discussions. Attendants are encouraged to ask questions during the session.