estimate the difference in the biases of the two designs by conducting surveys from the 1990 and 2003 frames simultaneously, exploring the causes of the biases in either design would be quite difficult. Information that is useful for the reduction of nonsampling errors is best obtained from special evaluation studies such as cognitive laboratory investigations, reinterview surveys, record check studies, and reliability analysis. Therefore, the committee questions the usefulness of continuing the old panel indefinitely as suggested in the hybrid option.
As an alternative to continuing the old panel indefinitely, the committee suggests carrying forward a small sample of the old panel on a one-time basis if the SRS staff considers it important to do so for the following purposes: (1) using comparisons of the new and old panel estimates to generate hypotheses regarding the sources and causes of the nonsampling error that could then be explored in more detail using the special evaluation study methods noted above; and (2) using simple ratios of the estimates derived from the old and new samples in 2003 to adjust estimates from the old data series so that they can be compared with estimates from the new design. In this way, the biases in contrasts of estimates from the old and new data series due to survey design differences could be substantially reduced.
Recommendation 2: If SRS staff confirm that a targeted sample could be useful for the purpose of adjustment, SRS should consider surveying in 2003 a small and carefully designed subsample of the current panel to study biases in the current sample, possibly to use for the purpose of adjustment.
This sample could help direct efforts to evaluate nonsampling error in the new panel and would also provide a means to adjust the old data series for comparisons with the new data series.
Recommendation 3: A cost-benefit analysis should be conducted to optimize the relative allocation of resources between the National Survey of College Graduates and the National Survey of Recent College Graduates. Also, additional oversampling should be applied to capture adequate numbers for small domains for which increased interest has become apparent since the last design.