Plotting your data to test seasonal adjustment
Ever take a statistics class? If so, do you recall your instructor telling you to “plot your data” and look at it before using it? It’s sound advice but is not always heeded, which can lead to foolish data-driven errors. While surfing through FRED (as I am wont to do; and who isn’t?!) I found such an error related to the difficulties in seasonal adjustment.
The graph shows two series: seasonally adjusted (red) and not seasonally adjusted (blue) U.S. postal employment. There’s a clear increase in employment each December to help cover the Christmas rush. Since the December rise is temporary and conveys no long-term information about employment patterns, we usually look only at the seasonally adjusted figures. But as we can see in the red line, the seasonal adjustment methodology only partially removes the seasonal pattern. In most cases, it does a good job removing the December effect between the early 1970s and now. But in the earlier years, the seasonal adjustment fails spectacularly, leaving the bulk of the December effect uncorrected. Clearly, the pattern of seasonality has changed across time and the methodology used to seasonally adjust should reflect that change—and there’s no way we would have known that had we not “plotted the data” before making the adjustment.
How this graph was created: Browse data by category. Under the category “Population, Employment, & Labor Markets” select “Current Employment Statistics (Establishment Survey).” Then select the subcategory “Government.” Scroll through the 23 series to find “All Employees: Government: U.S. Postal Service” and select the not seasonally adjusted version. From the “Edit Graph” option, use the “Add Line” tab to search for “postal service employees.” Select the seasonally adjusted version of the series “All Employees: Government: U.S. Postal Service” and click “Add data series.”
Suggested by Michael McCracken.
View on FRED, series used in this post: