We are in full agreement with the recommendation that SIPP become the basis for the nation's official poverty and related statistics. The March CPS does not collect all of the information needed for poverty measurement, has problems with the quality of the information that it collects, and does not have much room for further improvement. In contrast, SIPP collects most of the needed information, has achieved quality improvements, and, because of its focus on income, has ample opportunity for further improvements in both the scope and the quality of income-related data. The best time to put this recommendation into effect would be in 1996, when other changes to the survey are made.
A decision to use SIPP to produce the official poverty data means that all aspects of the survey should be reviewed to determine their suitability for providing the most accurate statistics possible under the proposed measure. A key aspect for review is the proposed redesign of the survey. Although the Census Bureau has accepted many of the recommendations of the CNSTAT Panel to Evaluate SIPP, it has decided against the recommendation for a design that would have two panels of about 27,000 households each in the field each year, with new panels introduced every 2 years. Instead, the Census Bureau has proposed a design that would have one large panel of 50,000 households in the field each year, with new panels introduced every 4 years.
The Census Bureau's design has the advantage of maximum sample size in a single panel for purposes of longitudinal analysis. For cross-sectional analysis, the two designs are equivalent: the two panels in the field each year under the CNSTAT SIPP panel's design can readily be combined to produce the same sample size as the single, larger panel of the Census Bureau's design.
Longitudinal estimates are important, but we believe that the time series of annual poverty rates and other statistics is paramount and that the design must support the production of reliable annual estimates. In this regard, the Census Bureau's proposed design provides no overlap between panels. Hence, every 4 years, it will be hard to determine if changes in the poverty rate are real or due to the introduction of a new panel in place of an old panel that may have uncorrected attrition bias or other problems.25
Since most attrition of sample cases from SIPP occurs by the end of the first year of a panel, there may be problems of attrition bias with the CNSTAT SIPP panel's design as well as the Census Bureau's, as the former does not
Attrition bias can occur when attrition rates differ between groups: for example, higher rates of attrition for low-income people could produce a downward bias in the poverty rates. Adjustments to the survey weights are usually made to compensate for attrition bias, but the adjustments may not be adequate.