Canal
Adjusted Seasonal Moving Average Model for the Current Population Survey (303775)
*James Lawrence, US Census Bureau
Keywords: response rates, paradata, time series, real time forecasting
For many years, the Current Population Survey (CPS) has had the option of extending its data collection period in times of sluggish response due to inclement weather or other unforeseen obstacles. Historically, the decision to extend interviewing was made by survey managers' review of key metrics. The Census Bureau developed a model that calculates a projected response rate for each CPS interview period a few days before closeout. This projection is intended to help reach a data-driven decision on whether to extend interviewing, based on how interviewing has been progressing on the ground. The model that we use is time series-driven and aims to account for the dynamically evolving interviewing environment. Each month’s projection starts with a moving average of the prior year’s response rates. A seasonal adjustment is then applied to account for seasonal trends. Finally, an adjustment is made for the amount of interviewing progress has been made in the current month compared to the amount typically made by then. We present this model and its predecessor, discussing results showing that this model produces more accurate projections and provides them at an earlier date.