Technical Documentation: Generating Unbanked and Underbanked Estimates for Local Geographies Prepared by Haveman Economic Consulting 1 and CFED August 2011 Introduction For years, researchers, policymakers, financial institutions, and community development practitioners have been trying to understand the unbanked and financially underserved populations both the socioeconomic demographics of the populations and the attitudes and behaviors that drive financial decisions. The launch of initiatives to better serve this market has created further demand for data on the unbanked. While the availability of quality data on the unbanked has increased significantly at the national and state levels, there is no current data source that can provide information on the size of the unbanked population at the local level. Haveman Economic Consulting (HEC) has therefore been contracted by CFED, under the U.S. Treasury Department s Bank on USA initiative, to develop a two-stage methodology that provides estimates of the banked status of any geography down to the census tract level, in order to help local stakeholders understand the size of their un- and underbanked populations. The methodology development was carried out with the assistance, and under the guidance, of several expert advisors. 2 The purpose of this memo is to describe the process used to develop the methodology and resulting estimates. Overview of Estimation Procedure Developed by Haveman Economic Consulting In 2009, the FDIC, in conjunction with the U.S. Census Bureau, carried out a survey of household banking habits. The survey was administered as an FDIC-sponsored supplement to the Current Population Survey (CPS) in January 2009. From this survey, the FDIC has published data on the proportion of the population that is unbanked and underbanked for the nation, the 50 U.S. states and the District of Columbia, and the 69 largest metropolitan regions (MSAs). This is now widely accepted as the most representative and reliable data source for banking status at these geographies. 3 As provided in the FDIC report, the definitions of unbanked and underbanked are as follows: 4 Unbanked: Households that would answer no to the following question: Do you or does anyone in your household currently have a checking or savings account? Underbanked: Households that have a checking or savings account but rely on alternative financial services. Specifically, underbanked households have used non-bank money orders, non-bank check-cashing services, payday loans, rent-to-own agreements, or pawn shops at least once or twice a year or refund anticipation loans at least once in the past five years. While data for large geographies are now available from the FDIC, 5 banking status data for smaller geographies, such as the remaining MSAs, counties, cities and census tracts, cannot be derived from the survey. To fill this gap, HEC has been contracted to develop a two-stage methodology that provides estimates of the banked status of any Census-delineated geography with 250 or more households in the country, down to the census tract level. The two-stage process works as follows. During the first stage, a model of how various household characteristics contribute to un- and underbanked status is created based on regression analysis of the FDIC survey data. During the second stage, local estimates of un- and underbanked status are produced using data on household characteristics from the American Community Survey 6 as the inputs to the estimation model developed in the first stage. This technical document was prepared by CFED under contract with the U.S. Department of the Treasury.
In order to provide the best possible estimates, we use separate methods for different geographies, depending on the availability of data. The following methods are used for the specified geographies: 1. Data provided by the FDIC Survey. As described above, data on household banking status for the United States, the 50 states and the District of Columbia, and the largest 69 MSAs are available from the 2009 FDIC National Survey of Unbanked and Underbanked Households. 2. Estimates based on American Community Survey (ACS) Public Use Microdata Sample. Although the ACS does not collect information regarding the banking activities of its respondents, we are able to establish correlates to banking status using regression analysis based on information in the FDIC survey. These regression results are then used to estimate banking status for individual households among survey respondents in the 2005-2009 ACS Public Use Microdata Sample (PUMS). Data from PUMS are released for Public Use Microdata Areas (PUMAs), which are statistical geographic areas with a population of 100,000. 7 From these household level estimates of banking status, we are able to provide regional estimates for counties and MSAs that are larger than 100,000 in population, based on one or more PUMAs. This process produces estimates for the larger MSAs for which the FDIC survey does not publish data, as well as for many counties. When the size of the geography allows for its application, this estimation model is preferred relative to method (3), due to the greater precision of its estimates. 3. Estimates based on the Census Bureau s FactFinder Regional Summary Files of 2005-2009 American Community Survey data. For geographies smaller than 100,000 and for all census places and census tracts, summary estimates at the geographic level, rather than the household level, of correlates to banking status are used. These summary estimates were downloaded from the Census FactFinder website 8 for each census tract 9 and census place. 10 Aggregations of census tract summary data are used to generate estimates for counties and MSAs with a population less than 100,000, or for geographies that do not cleanly map to PUMAs. These estimates are derived from the same set of regressions used in method (2). SUMMARY OF DATA SOURCES AND METHODS FOR DIFFERENT GEOGRAPHIES Data Source/Method FDIC Survey Data Estimates based on ACS Microdata Estimated based on ACS Regional Summary Data Geography United States 50 States and District of Columbia 69 largest MSAs Larger MSAs Larger Counties Smaller MSAs Smaller Counties All Cities / Census Places All Census Tracts Establishing Banked Status Correlates In methods (2) and (3) above, estimates of unbanked and underbanked status are generated as predictions from a set of four regressions. 11 These regressions provide separate estimates of the following: 1. The proportion unbanked in metropolitan areas 2. The proportion unbanked in non-metropolitan areas 3. The proportion underbanked in metropolitan areas 4. The proportion underbanked in non-metropolitan areas The regressors are a set of correlates to banking status, established through the four regressions listed above. Separate regressions are performed for metropolitan and non-metropolitan areas (areas within an MSA and areas not in an MSA) to improve the quality of the estimates. Both intuitively and within the data, the relationship CFED: Technical Documentation: Generating Unbanked and Underbanked Estimates for Local Geographies 2
between banked status and the correlates/regressors used is different for households within an MSA versus those outside of an MSA. The set of regressors was chosen through a three-step process: 1. An extensive search of the literature on determinants of banked status, and HEC s own experience. The literature review found that the household demographics and characteristics most correlated to being unbanked or underbanked are fairly consistent across the different surveys and research. Those most likely to be unbanked are low-income, members of particular racial and ethnic groups, young and less-educated households. In addition, citizenship, nativity, speaking a language other than English at home and household composition have all been found to be related to unbanked status. 2. The regressors were then limited to only those for which comparable data points are available in each of the three data sources (FDIC data, the ACS Microdata and the ACS Regional Summary data available through FactFinder). Citizenship and nativity had to be excluded as regressors for this reason. 3. Standard tests of influence of the regressor on the quality of the estimates. 12 Table 1 provides a comprehensive list of the regressors included in the regressions. Choice of Regression Framework The regression framework is simple Ordinary Least Squares (OLS), or a linear probability model. 13 Comparisons of the performance of various modeling choices were made across state and MSA level estimates. Given that the estimates would be based on both micro-data and geographically based aggregates of the micro-data, both types of estimates from the raw FDIC survey data were produced and compared with the estimates obtained directly from the survey. 14 Although no modeling option provided the best estimates in every case, the linear probability model provided estimates with generally lower error more often than did the other modeling options. 15 Given these results, simplicity was chosen over the use of several different modeling approaches, and the linear probability model is the source of the estimates for geographies that cannot be directly sourced from the FDIC survey data. Tables 2 and 3 compare the final unbanked and underbanked estimates from both models with the FDIC survey data for each state. Procedure This section spells out in more detail the procedure used to generate estimates for each of the three sets of regions indicated above. 1. FDIC survey data (national, 50 states and the District of Columbia, and the 69 largest MSAs) a. These data were sourced directly from the published FDIC survey results. 16 2. American Community Survey (ACS) Microdata (some MSAs, larger counties) a. From the FDIC micro-data, construct the set of correlates indicated in Table 1 and the unbanked and underbanked variables. b. Run the 4 OLS regressions indicated above unbanked and underbanked separately for metro and nonmetro households. c. Obtain the 5-year merged ACS data (2005-2009) for households (within PUMAs) from the Census website. d. Construct covariates/regressors that are analogous to those constructed from the FDIC micro-data. e. Use the coefficients from the regressions to estimate unbanked and underbanked status for each household in the sample. f. Aggregate over households, weighting by the household weight, to obtain estimates for the proportion of unbanked and underbanked for each geography. CFED: Technical Documentation: Generating Unbanked and Underbanked Estimates for Local Geographies 3
3. American Community Survey (ACS) Regional Summary Statistics from FactFinder (smaller MSAs, smaller counties, all places and the individual census tracts 17 ) a. From the FDIC micro-data, construct the set of correlates/regressors indicated in Table 1 and the unbanked and underbanked variables. b. Run the 4 OLS regressions indicated above unbanked and underbanked separately for metro and nonmetro households. c. Obtain the 5-year merged ACS regional summary data (2005-2009) from the Census website. d. Extract summary variables that are analogous to those constructed from the FDIC micro-data for each census tract and census place. e. Use the coefficients from the regressions to estimate unbanked and underbanked status for each census tract and census place. Note: These are the same coefficients as those used to create estimates from the ACS microdata. f. Aggregate over census tracts, weighting by the household weight, to obtain estimates for the proportion of unbanked and underbanked for counties and MSAs. CFED: Technical Documentation: Generating Unbanked and Underbanked Estimates for Local Geographies 4
Table 1: COMPREHENSIVE SET OF REGRESSORS Income Variables Income: LESS THAN $5,000 Income: 5,000 TO 7,499 Income: 7,500 TO 9,999 Income: 10,000 TO 12,499 Income: 12,500 TO 14,999 Income: 15,000 TO 19,999 Income: 20,000 TO 24,999 Income: 25,000 TO 29,999 Income: 30,000 TO 34,999 Income: 35,000 TO 39,999 Income: 40,000 TO 49,999 Income: 50,000 TO 59,999 Income: 60,000 TO 74,999 Income: 75,000 TO 99,999 Income: 100,000 TO 149,999 Income: 150,000 OR MORE Poverty Dummy Age of Household Head Age: 25-34 Age: 35-44 Age: 45-54 Age: 55-64 Age: 65+ Race of Household Head Black Latino Native American Asian Other Educational Attainment of Household Head High School Grad Some College Bachelor s Degree or More Household Characteristics Spanish Only Spoken At Home Female Headed Household Married Couple Children Present Domestic Partnership Single Parent Female Single Parent Household Size 2 People in Household 3 People in Household 4 People in Household 5 People in Household 6 People in Household 7+ People in Household Regional Indicators Census Region: Mid-Atlantic Census Region: East North Central Census Region: West North Central Census Region: South Atlantic Census Region: East South Central Census Region: West South Central Census Region: Mountain Census Region: Pacific State of Louisiana* Variables Omitted from the Regressions Income: nonresponsive 1 person in household Age < 35 Race: White Education: less than high school diploma Census Region: New England * A specific indicator for Louisiana was included because without it, the estimates for Louisiana were quite poor. The inclusion of this indicator significantly improved the state and MSA level estimates for Louisiana relative to the survey data. CFED: Technical Documentation: Generating Unbanked and Underbanked Estimates for Local Geographies 5
Table 2: COMPARISON OF UNBANKED ESTIMATES TO FDIC SURVEY DATA Unbanked Estimates State FDIC Survey Data Estimate (ACS Microdata) Estimate (ACS Regional Summary Data) AK 4.3% 5.3% 5.9% AL 11.6% 11.6% 12.1% AR 10.1% 10.1% 10.2% AZ 7.5% 7.3% 7.5% CA 7.7% 7.5% 7.7% CO 6.9% 5.7% 6.0% CT 5.3% 4.5% 4.5% DC 12.2% 10.6% 11.0% DE 5.6% 6.0% 6.3% FL 7.0% 6.7% 6.9% GA 12.2% 8.7% 9.3% HI 2.9% 3.7% 3.7% IA 4.7% 4.4% 4.5% ID 6.7% 5.4% 5.8% IL 6.2% 6.7% 6.7% IN 7.4% 6.1% 6.2% KS 6.4% 5.4% 5.6% KY 11.9% 10.0% 10.4% LA 8.7% 9.0% 10.3% MA 4.1% 4.1% 4.3% MD 5.6% 5.7% 6.0% ME 2.6% 3.5% 3.7% MI 6.7% 6.2% 6.2% MN 2.6% 4.1% 4.2% MO 8.2% 6.4% 6.5% MS 16.4% 14.7% 15.4% MT 3.8% 5.3% 5.9% NC 8.2% 8.3% 8.8% ND 4.8% 4.9% 5.1% NE 5.4% 5.0% 5.1% NH 2.2% 2.1% 2.2% NJ 7.4% 6.5% 6.6% NM 11.4% 10.5% 10.9% NV 6.9% 7.2% 7.4% NY 9.8% 8.0% 8.1% OH 7.1% 6.2% 6.3% OK 9.8% 8.8% 9.2% OR 5.7% 4.8% 5.1% PA 5.1% 6.0% 6.0% RI 6.2% 5.5% 5.6% SC 10.2% 8.5% 9.3% SD 4.8% 5.2% 5.5% TN 9.9% 10.1% 10.6% TX 11.7% 11.4% 11.7% UT 1.7% 4.2% 4.6% VA 5.1% 5.5% 6.1% VT 4.2% 2.6% 2.8% WA 3.9% 4.2% 4.6% WI 4.3% 4.7% 4.7% WV 6.3% 7.1% 7.6% WY 4.0% 4.6% 5.2% CFED: Technical Documentation: Generating Unbanked and Underbanked Estimates for Local Geographies 6
Table 3: Comparison of Underbanked Estimates to FDIC Survey Data Underbanked Estimates State FDIC Survey Data Estimate (ACS Microdata) Estimate (ACS Regional Summary Data) AK 25.5% 17.6% 18.0% AL 20.2% 22.6% 22.5% AR 22.3% 24.2% 25.1% AZ 16.8% 18.6% 18.6% CA 15.2% 16.6% 16.6% CO 15.3% 17.2% 17.3% CT 13.8% 13.7% 13.6% DC 23.9% 21.0% 21.2% DE 14.7% 18.9% 18.7% FL 16.8% 18.8% 18.7% GA 19.4% 21.2% 21.2% HI 13.8% 15.2% 15.0% IA 16.8% 16.0% 16.3% ID 19.7% 18.3% 18.6% IL 15.7% 18.0% 18.1% IN 16.8% 18.4% 18.7% KS 17.4% 16.3% 16.6% KY 23.7% 20.7% 20.6% LA 22.9% 23.7% 25.4% MA 11.4% 12.8% 12.8% MD 20.0% 18.6% 18.5% ME 18.0% 15.6% 15.7% MI 16.7% 18.5% 18.6% MN 11.1% 15.3% 15.4% MO 19.3% 17.2% 17.4% MS 25.2% 24.4% 24.4% MT 19.7% 18.1% 18.4% NC 20.0% 20.4% 20.4% ND 19.0% 16.2% 16.7% NE 14.9% 16.5% 16.7% NH 12.1% 14.2% 14.2% NJ 12.0% 16.8% 16.7% NM 21.7% 20.7% 20.8% NV 20.5% 19.1% 19.0% NY 19.3% 18.4% 18.5% OH 21.0% 18.4% 18.5% OK 21.9% 24.0% 24.4% OR 14.8% 16.1% 16.2% PA 17.6% 18.2% 18.4% RI 12.2% 14.2% 14.1% SC 24.2% 20.8% 21.0% SD 16.2% 16.7% 17.0% TN 17.5% 21.5% 21.4% TX 24.1% 23.9% 24.1% UT 15.2% 17.7% 17.9% VA 15.5% 18.0% 18.1% VT 12.1% 15.9% 15.9% WA 17.3% 15.5% 15.5% WI 16.0% 17.3% 17.7% WV 20.7% 18.7% 18.7% WY 17.4% 18.3% 18.7% CFED: Technical Documentation: Generating Unbanked and Underbanked Estimates for Local Geographies 7
Endnotes 1 Haveman Economic Consulting (HEC) is a private consultancy based in San Rafael California. The firm is owned by Jon Haveman. Dr. Haveman received his Ph.D. in Economics from the University of Michigan and has been producing data on economic well-being in the United States for over 20 years. Prior to starting HEC, Dr. Haveman was on the faculty at Purdue University, a Senior Economist at the Presidents Council of Economic Advisors, and the director of the Economy Program at the Public Policy Institute of California. 2 HEC and CFED gratefully acknowledge the expert guidance of Carolina Reid from the Federal Reserve Bank of San Francisco, Caroline Ratcliffe from the Urban Institute, Jodie Harris from the CDFI Fund, and Louisa Quittman from the U.S. Treasury Department throughout the development of this methodology. In addition, we thank Anne Stuhldreher, Pamela Chan and Terri Friedline from the New America Foundation, Heidi Goldberg from the National League of Cities, and Leigh Phillips from the City of San Francisco s Office of Financial Empowerment for their thoughtful comments on this technical document. 3 There are five key surveys that have been used to estimate the number of unbanked households nationally over the past 35 years: The Panel Study of Income Dynamics (PSID); The Survey of Income and Program Participation (SIPP); The Survey of Consumer Finances (SCF); The Center for Financial Services Innovation (CFSI) Underbanked Consumer Study; and the 2009 FDIC Survey of Unbanked and Underbanked Households. These surveys vary based on sample size, frequency, scope, population of interest, and additional factors. All five of the surveys define unbanked as households that do not currently have a checking or savings account. The FDIC survey data is the most recent available, has the largest sample size (47,000 households), and is the only one of the five surveys designed to be representative down to the state level in all 50 states. Unlike the PSID and SIPP, its primary focus is on financial services usage, but since it was collected along with the CPS, a wealth of household demographic information can be connected to the financial services usage data. 4 Both of these definitions are taken directly from the FDIC report FDIC National Survey of Unbanked and Underbanked Households, December 2009. Unbanked is defined in footnote 3 on page 10 and underbanked is defined in footnote 6, also, on page 10. 5 Available at http://economicinclusion.gov. 6 The American Community Survey is used as the data source for the regressors because it is the only source for household demographic data that is representative down to the census tract level. 7 Available at http://www.census.gov/acs/www/data_documentation/pums_data/ 8 Available at http://www2.census.gov/acs2009_5yr/summaryfile/ 9 The 2005-2009 ACS 5-year estimates are based on data collected between January 2005 and December 2009 and are based on a sample designed to be representative down to the census block group level. 10 Incorporated places such as cities, towns and villages or their Census-designated statistical equivalent. Because census place boundaries often do not map cleanly to census tracts, regional summary data for census places is used directly for the place estimates, rather than aggregation of census tract estimates. 11 The dependent variable for each of the regressions is based on the definitions of unbanked and underbanked provided above and come from the FDIC survey. 12 Quality was judged on the effect of the average absolute error and mean squared error of the resulting estimates. If a regressor s inclusion reduced the precision of the estimates, the regressor was excluded. The primary effect of this exercise was to affect the form in which information was included in the regression rather than to point to the need to exclude regressors. 13 The choice of OLS was made through an intensive comparison of estimates produced from several alternative modeling options: a) probit, b) logit, and c) ordered logit. In each case, estimates based on a model constructed using a random sample of half of the FDIC data were produced for the geographies available in the FDIC survey (states and some MSAs). Each set of our estimates was compared to the FDIC s estimates based on the survey data in terms of a) absolute deviation from the survey estimates, b) mean squared deviation from the survey estimates, and c) bias (either high or low) relative to the survey estimates. 14 The exercise involved first generating regression coefficients from half of the FDIC survey. Once these coefficients have been produced, they are used to generate estimates of banked and unbanked status using the other half of the household level survey data for each state and MSA. In order to estimate the impact of using aggregated data (such as that from FactFinder), we also aggregated the other half of the household level data and used those aggregates to produce estimates. Each modeling structure was compared with the others based on the following sets of estimates: (1) FDIC reported estimates versus household data for states; (2) FDIC reported estimates versus aggregated household data for states; (3) FDIC reported estimates versus household data for MSAs; (4)FDIC reported estimates versus aggregated household data for MSAs. 15 In particular, the ordered logit model was excluded because it provided estimates of unbanked status with a significant downward bias when using aggregated data. It is also the case that no model predicted dramatically more precise estimates for any particular geography or aggregation of the data. In no case was the difference in mean squared error or average absolute deviation more than 5%. 16 Available at http://economicinclusion.gov. 17 To increase the reliability of the estimates, geographies with fewer than 250 households are excluded. About CFED CFED (Corporation for Enterprise Development) expands economic opportunity by helping Americans start and grow businesses, go to college, own a home, and save for their children s and own economic futures. We identify promising ideas, test and refine them in communities to find out what works, craft policies and products to help good ideas reach scale, and develop partnerships to promote lasting change. We bring together community practice, public policy and private markets in new and effective ways to achieve greater economic impact. www.cfed.org CFED: Technical Documentation: Generating Unbanked and Underbanked Estimates for Local Geographies 8