Addressing the Challenges of Assessing Forecasting and Time Series Methods Using COVID-19 Pandemic-Era Data

• Activity Number: 517 - New Approaches for Correlated and Clustered Data
• Type: Contributed
• Date/Time: Thursday, August 11, 2022 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Niloofar Ramezani* and Karen Traxler
• Companies: George Mason University and U.S. Department of Justice
• Keywords: COVID-19; Response Rate; Forecasting; Time Series; Smoothing; Accurate Prediction

Abstract

Researchers using archival data to predict count variables face unique challenges due to lower response rates during the COVID-19 pandemic. Combining pre and during pandemic data to build forecasting models may introduce significant error to the forecasting due to stability assumption violation. Currently, not enough evidence-based guidance about properly accounting for the pandemic effect on future count data exists. To provide applied researchers reliable recommendations to estimate future counts when these data are included, authors used publicly available data from 2018 through 2021 collected by the Department of Justice’s Executive Office of Immigration Review to predict the future number of respondents in immigration removal cases by fiscal year. Accurate immigration respondent predictions inform budgeting efforts as well as public policy. Using these immigration data, authors examined measures of model fit for three data forecasting methods: Exponential smoothing, INGARCH time series, and Nonlinear denoising via wavelet shrinkage techniques. Methodological recommendations are provided regarding evidence-based best practices for forecasting precise respondent counts.