Estimating Gross Flows Consistent with Stocks in the CPS: The Basic Gross-Flow Table Formerly Used by the Bureau of Labor Statistics in Examining Labor Market Flows Was Expanded and the Resulting Tables Were Raked Iteratively in Order to Produce Labor Market Flow Statistics Compatible with CPS Stocks

Article excerpt

The Current Population Survey (CPS) is primarily a cross-sectional survey signed to estimate the distribution of labor force states--employed (E), unemployed (U), or not in the labor force (N)--among the population (1) for a given month. However, the CPS also can be used to examine the number of persons who change their labor force state between months.

Gross-flow estimates describe the month-to-month transitions from one labor force state to another. The following 3 x 3 matrix gives an example in which EU represents the number of persons who were employed in the previous month (May) and are unemployed in the current month (June), and similarly for the other entries:

               Current month (June)

               E     U    N

Previous  E    EE    EU   EN
month     U    LTE   UU   UN
(May)     N    NE    NU   NN

Gross-flow estimation is possible in the CPS because households are interviewed for 4 consecutive months, are then rotated out of the survey for 8 months, and are then interviewed for another 4 consecutive months. About three-fourths of the sample households are in common across 2 consecutive months. Household records can be linked, and month-to-month labor force transitions determined, for most persons in those households.

Gross-flow statistics from the CPS were published from 1948 until 1952. Publication was stopped because there were clear discrepancies between labor force changes derived from the flows and labor force changes derived from the monthly stock estimates. (The sources of these differences are explained later.) Over the years, many analysts have called for the Bureau of Labor Statistics (BLS) to resume publishing gross flows. This article describes a new method of obtaining flow statistics that are compatible with the monthly stock numbers. Seasonal adjustment of gross-flow series also is discussed.

Existing gross-flow data problems

The Census Bureau generates unpublished gross-flow estimates as part of its monthly production of CPS data. The current procedure used by the Census Bureau to generate the tabulations each month starts by matching respondents in the current month to respondents in the previous month; about 72 percent are matched. Next, the sampling weights of the matched respondents are adjusted so that weighted sample totals, by sex, match known population totals. The adjusted weights are then used to compute weighted estimates of labor force transition flows. The analysis that follows focuses on two types of error inherent in this procedure: classification error and margin error.

Classification error. For a variety of reasons, some CPS respondents may be classified into the wrong labor force state. Errors in classifying the respondent can have large effects on gross-flow calculations. In stock data, classification errors tend to offset each other, whereas in flow data, errors tend to be additive. For example, if equal numbers of respondents are erroneously classified as employed when they are unemployed and as unemployed when they are employed, stock data will be unaffected, but both EU and UE flows will be increased.

Although research indicates that classification error may have large effects on gross flows, the Bureau of Labor Statistics has no current plans to publish classification-error-corrected flows. While measurement error probabilities could be derived from reinterview data, it is not entirely clear how such data should be used. In their attempts to correct for classification error, John Abowd and Arnold Zellner, (2) and James Poterba and Lawrence Summers, (3) used "reconciled" reinterview data, whereby the interviewer attempted to establish a true labor force state in the case of contradiction between the original survey and the reinterview. Because of data quality problems, however, reconciled reinterview data are no longer being produced. Tin Chiu Chua and Wayne Fuller (4) used unreconciled reinterview data, but doing so requires additional statistical assumptions. …