Criteria for selecting a suitable method for producing post-2000 county population estimates: A case study of population estimates in Illinois

Abstract

We compared 2000 county population estimates for Illinois against 2000 census counts. Administrative records (ADREC) and ratio correlation (Ratio-CORR) methods were used to produce two sets of controlled county estimates for 2000; a third set represented an average of the estimates reached using these methods. Another set using the ADREC method was not controlled to any estimate. Also, the 2000 estimates were adjusted for undercount in the 1990 census. We compared performance of these methods with the performance of two naive models: (i) do nothing and (ii) constant growth rate. ADREC estimates were more accurate than estimates from the Ratio-CORR or Average method in terms of Mean Absolute Percent (MAPE) or weighted MAPE. Undercount adjustment in general improved the accuracy of the estimates for all three methods. A top-down or bottom-up approach worked equally well. As a single method, ADREC performed best.

Indirect Estimates of Total Fertility Rate Using Child Woman/Ratio: A Comparison with the Bogue-Palmore Method

“…In this study, comparisons of absolute and signed numeric and percentage differences, and estimates of moments of these distributions such as averages or the root mean-squared error [27]–[28] were employed to evaluate the accuracy and bias associated with estimates of TFR in these estimates. Signed, algebraic errors (mean algebraic error–MAlgE–and mean algebraic percentage error–MALPE) proposed to capture bias [21], [25] while absolute numeric (mean absolute error–MAE) and percentage (MAPE) were utilized to capture accuracy. The Root Mean Squared Error (RMSE) is reported in both numeric and percentage terms as a measure of the robustness of the method across the diverse nations for which it was tested [27]–[28]; a robust estimate should have a relatively low RMSE indicating the presence of few unusually large errors.…”

“…The Root Mean Squared Error (RMSE) is reported in both numeric and percentage terms as a measure of the robustness of the method across the diverse nations for which it was tested [27]–[28]; a robust estimate should have a relatively low RMSE indicating the presence of few unusually large errors. Here, we summarize results in percentage terms to avoid size-related bias in numeric comparisons [21], [25], [36]. In this analysis, comparisons of the population characteristics –suggested to be associated with fertility variation in the Proximate Determinants model of Bongaarts and Potter [19]– associated with countries where each method (the algebraic percentage error scores in particular) performed best were made using contingency table analysis based on the chi-squared statistic [37]–[38] and the Wilcoxon ranked sums test [39].…”

“…The performance of the method is characterized and then compared (where data permit) to that of the Bogue-Palmore method. Statistical measures of accuracy and bias such as the mean absolute and algebraic percentage errors (MAPE & MALPE) [21], [25]–[26] and the root mean-squared error (RMSE) [27]–[28] are computed, comparing the estimated values to those reported in the IDB. The relative performance of each method is considered in light of the characteristics of countries where each method is more accurate, by considering data from the Demographic and Health Surveys (DHS–www.measuredhs.com) on population characteristics.…”

The impact of incomplete geocoding on small area population estimates

“…Differences between estimated and observed 2000 household populations are used to measure the accuracy of estimates in terms of average overall numeric and percentage errors across all block groups. Because error distributions are asymmetric and tend to be right-skewed (Shahidullah and Flotow 2005;Tayman and Swanson 1999), both mean and median errors were considered in evaluating the accuracy of estimates. Differences in these error distributions between unremediated and remediated estimates were used to evaluate improvements associated with the application of Horvitz-Thompson adjustments.…”

“…This also avoids the additional challenge of estimating the group quarters population (National Research Council 2010), for which estimation errors related to surveillance and geocoding are best considered separately. Accuracy is evaluated as the discrepancy between estimates and observed Vintage 2000 household populations, measured as the mean numeric and percentage errors across all block groups (Shahidullah and Flotow 2005;Smith and Mandell 1984;Smith et al 1999;Smith and Shahidullah 1995). The spatial patterning of these errors is explored using heat maps, in which increasingly dark coloration of block groups represents increasing magnitude of error (Berke 2005;Fotheringham et al 2002).…”