For each month of the panel period, we constructed hourly wages for each job and business using detailed employment information in SIPP. SIPP contains direct information on hourly wage rates for the 60 percent of jobholders who could provide wage data in this way. Hourly wage rate information, however, is not available for the remaining 40 percent of jobholders and for all those with businesses. For these workers, we constructed hourly wages by combining information on monthly earnings (which are reported for each month of the panel period) and usual hours worked per week at each job or business during the reference period (topcoded at 84 hours), and assuming that the worker was employed for the entire month.(8) The "earnings-based" hourly wage measure was then constructed for each month by dividing monthly earnings by the number of hours worked in the month.(9)
Our preliminary analysis of the SIPP data showed that hourly wage rates fluctuated considerably over time, and especially for the constructed earnings-based measures. These fluctuations are often due to sudden large changes in wage rates that appear to be due more to reporting errors or SIPP data errors than to real wage changes. Furthermore, they yield more worker transitions into and out of the low-wage labor market than we deem plausible. Consequently, we used several methods to "smooth" the hourly wage rates to identify those who were truly in low-wage jobs:
- We set outliers to missing. Wages below $1 and above $150 were treated as missing, which affected 2.7 percent of workers. Furthermore, the SIPP user notes report a data imputation problem for some jobholders whose earnings information was missing. Earnings are reported as zero for these workers rather than as a positive imputed value. SIPP reports that this problem may have affected around 1.5 percent of the observations in the monthly earnings distributions. However, it is not possible to identify these individuals from those who truly reported zero wages. Thus, we set zero wage values to missing. Finally, SIPP topcoded monthly employment income at $12,500. Due to our focus on low-wage workers, however, this constraint does not materially affect the analysis.
- We smoothed the earnings-based hourly wage rates by averaging positive wage values across the four months within a wave. We smoothed in this way because the earnings-based measure varies by month (because sample members were asked to report their earnings for each month of the reference period), whereas the direct hourly wage measure pertains to the entire wave and not to specific months within the wave. Thus, there is considerably more fluctuation in the earnings-based hourly wage measure than in the direct hourly wage measure, which generates more frequent and shorter spells of low-wage employment using the earnings-based measure.
- We smoothed unusual changes in hourly wage rates in the same job across waves. If wages within a job suddenly increased by 25 percent and then rapidly decreased by 25 percent or vice versa, then we smoothed (imputed) wages at the "spike" points as the average of the surrounding wages on that job. We set a conservative 25 percent threshold value to avoid over-smoothing the data.
Finally, for those with multiple jobs and businesses in a particular month, we selected the hourly wage from the job or business in which the sample member worked the most hours. In March 1996, about 11 percent of workers held multiple jobs and businesses. Thus, we defined whether a worker was a low- medium-, or high-wage worker using the wage on the selected "main" job or business in that month.