and Estimating Benchmarking the Trend in the Poverty Rate from the Panel Study of Income Dynamics
Lloyd D. Griegera Sheldon Danzigerb Robert F. Schoenic
Population Studies Center Research Report 08-633
March 2008
a
Population Studies Center, Gerald R. Ford School of Public Policy, Department of Sociology, University of Michigan National Poverty Center, Gerald R. Ford School of Public Policy, Population Studies Center, University of Michigan
b
c
Institute for Social Research, Gerald R. Ford School of Public Policy, Department of Economics, University of Michigan
Direct correspondence to Lloyd Grieger, Population Studies Center, University of Michigan, 426 Thompson Street, Ann Arbor, MI 48106-1248. Tele: [phone]. Fax: [phone]. E-mail: [e-mail].
Estimating and Benchmarking Poverty Rate Trend from PSID
2
ABSTRACT
We describe how to accurately estimate poverty rates using the Panel Study of Income Dynamics (PSID) because changes in the PSID over its 40-year history have created confusion for researchers. We discuss a number of issues related to measuring poverty and benchmark a new PSID poverty estimate with published rates from the Current Population Surveys (CPS). We demonstrate that to produce a consistent time series over the entire time period covered by the PSID that is similar to the Census Bureaus official approach, researchers should use the poverty thresholds presented in this paper. For example, the correlation between the PSID and CPS rates using one of the two currently available PSID thresholds is only 0.43 over the 1967-2002 period, whereas it is 0.82 using our new PSID threshold. The second of the two thresholds in the PSID archive is only available from 1989 onwards. For this period, poverty rates based on this second threshold have a correlation of 0.97 with CPS rates, which is similar to the new rates we propose here. Furthermore, as previously documented, the PSID yields poverty rates that are lower than the CPS rates, which is also the case with our new threshold.
Estimating and Benchmarking Poverty Rate Trend from PSID
3
1. INTRODUCTION
Since 1968, the Panel Study of Income Dynamics (PSID) has collected economic, demographic, and social data on a national sample of the U.S. population. The PSID was established to examine the dynamics of economic life, particularly dynamics in poverty status. As a result, the PSID has been widely used by poverty researchers, with 410 publications on poverty listed in the PSID bibliography as of February 2008. This paper has two goals: to guide researchers through the process of calculating the poverty rate from the PSID on a consistent basis for each year from 1968 to the present, and to compare the level and trend in PSID poverty rates to that of the official Census Bureau rates that are based on the March Current Population Survey (CPS). Two different poverty thresholds are available on the PSID data files in some years, and the thresholds available throughout the entire study period are not the ones used by the Census Bureau in calculation of official poverty rates. For this reason and others, there is some confusion among users about how best to use the PSID to estimate poverty. Our first goal is to eliminate this confusion. Poverty rates in the first few years of the PSID were benchmarked by Lane and Morgan [2]. However, we do not believe that anyone has updated these analyses despite the many changes in the PSID over the past nearly 40 years. The PSID was collected with an in-person interview until 1972 and a telephone interview in all subsequent years. Paper and pencil interview format was used until computer assisted interviewing was adopted in 1993. Until an immigrant refresher sample was added in 1997/1999, the PSID was not representative of individuals arriving in the US after 1968. In 1997, due to budgetary constraints, the PSID
Estimating and Benchmarking Poverty Rate Trend from PSID
4
eliminated roughly two-thirds of the low-income Survey of Economic Opportunity sample, and after 1997 respondents were interviewed biennially instead of annually. Moreover, cumulative selective attrition over the nearly 40 years of data collection, particularly selection not captured by the PSID weights, may have biased PSID estimates of poverty. Thus, an updated comparison of the trend in the PSID and official Census Bureau poverty rates is warranted. The next section describes the PSID data, emphasizing the two key elements that are required to determine poverty status: the poverty thresholds and the measure of total family income. Section III presents our results, with final observations provided in Section IV.
2. DATA
A person is counted as poor if his/her total family income is less than or equal to the familys poverty threshold. Below we discuss each of these concepts the poverty threshold and family income as reported in the PSID. The official CPS-based poverty rates can be found at the Census Bureaus web site: http://www.census.gov/hhes/www/poverty/histpov/hstpov2.html .
2.1. Poverty Threshold
Over the last four decades, the Census poverty thresholds have operated under two different regimes. Before 1980, a familys poverty threshold was determined by the number of people related by blood, marriage or adoption who resided in the same housing unit, the number of children in the family, the gender of the family head, the age of the family head, and whether the family lived on a farm. After 1980, distinctions between families headed by males and females and farm/non-farm families were dropped. For each year prior to 1980, Census Bureau thresholds are presented in four different matrices per year (each matrix includes separate thresholds as family size and the number of children present in each family size varies), one each per male/female headed by farm/non-farm family. Thresholds also differ between families
Estimating and Benchmarking Poverty Rate Trend from PSID
5
headed by a person less than 65 years of age and families headed by an elderly person. After the gender and farm distinctions were dropped, there was one matrix per year. The poverty threshold matrices for each year since 1980 are available from the Census Bureaus website at http://www.census.gov/hhes/www/poverty/threshld.html. The more complicated matrices for the years preceding 1980 can be obtained by examining the published annual poverty reports (P-60 series). Note, however, that only the reports from 1972 onward contain detailed matrices. These reports are publicly available at http://www.census.gov/hhes/www/poverty/publications.html. For every year dating back to 1959, the weighted average poverty threshold for a non-farm family of a given size is available at http://www.census.gov/hhes/www/poverty/histpov/hstpov1.html. When the PSID began in 1968, it was decided that the poverty thresholds put in the PSID data archive would use the low-cost food budget of the U.S. Department of Agriculture rather than the economy food budget which forms the basis of the official Census Bureau thresholds. The original 1968 documentation states that this decision was made because: It was the opinion of Faith Clark of the Department of Agriculture that the latter standard (i.e., economy level) was too spartan [3]. The economy budget is 80 percent of the low cost budget. These decisions are discussed on page 82 of the 1990 documentation (http://psidonline.isr.umich.edu/Data/Documentation/pdf_doc/psid90w23.pdf) and page 39 of the 1974 documentation, (http://psidonline.isr.umich.edu/Data/Documentation/pdf_doc/psid74w7.pdf). In our analyses below, we refer to the low-cost thresholds on the PSID data file as PSID-1. Note, however, that the PSID poverty threshold and the associated income-to-needs variable used to calculate PSID-1 are stated in each year in 1967 dollars, leaving it to users to
Estimating and Benchmarking Poverty Rate Trend from PSID
6
choose the appropriate inflation adjustment. This is an important fact that might not have been recognized by some data users. In our calculations, we use the CPI-U (all items using current methods, series CUUR0000SA0, available at: http://data.bls.gov/cgi-bin/surveymost?cu) to inflate this threshold to the relevant year. The attributes of PSID-1 and the three other thresholds discussed below are reported in Table 2. The PSID names for each variable used to compute the poverty rates are listed in Table 3. Note that threshold PSID-3 is available only for interview years 1990 onwards. One difference between Census Bureau procedures and the way the PSID develops the threshold is that partial year co-residence is accounted for in the PSID, but not in the CPS. That is, family membership in the PSID is determined for each month during the calendar year prior to the interview, creating a separate threshold for each of the 12 months. For each family, these 12 thresholds are averaged to determine the threshold for that family for the year as a whole. Similarly, the PSID measure of family income reflects the income of all persons living in the family unit during calendar year t, regardless of whether that person was living in the family at the time of interview in year t+1. Income for each family member includes only the amount accrued during the months that the person resided with the other family members. In contrast, the Census Bureau measures family composition at the time of the March CPS interview to determine who is in the family. Annual family income for the previous calendar year is measured as the total for all persons residing in the family unit at the time of the March interview, regardless of where they lived during the pervious calendar year. Consider a couple with a small child and assume that total family income in calendar year t was comprised solely of the husbands earnings. Assume the man earned $5000 per month but died on November 30. The PSID would consider this three-person family not to have been poor
Estimating and Benchmarking Poverty Rate Trend from PSID
7
in that year since the husband is counted in the threshold for 11 months and the total income over those months greatly exceeds the poverty threshold. However, in March t+1, the CPS would interview the widow who had no income at all in the previous year and count her and her child as a poor two-person family. Another important difference between the PSID and CPS is how each defines family, as both the poverty threshold assigned to each person and her/his family income depend on the number of people in the family. In the CPS, the family includes people who are related and reside together. Individuals living alone and unrelated individuals are not included in the calculation of CPS poverty rates, but are included as one person families in the official poverty rate for all persons. The PSID defines family more broadly and includes unrelated people who live together and share resources (like cohabiting partners) and single-person families. The CPS and the PSID also differ in their treatment of related subfamilies. For example, consider a grown child in a PSID family who moves away from her parents and marries. There are now two PSID families: the parents family and the grown childs own family. After some time, the now married child and her family return to live in the original household with her parents. The PSID would count this household as having two families, each with its own poverty threshold, whereas the CPS would count only one with the threshold based on the total number of household residents, as related subfamilies are considered part of the primary family. These situations are not very common, but they are more common among lower-income than higherincome families [1]. The PSID documentation provides guidance on how to make PSID-1 more comparable to the official Census Bureau threshold. Our second threshold, PSID-2, follows this guidance by multiplying the PSID-1 threshold by 0.8. This reduction transforms the PSID-1 from a threshold
Estimating and Benchmarking Poverty Rate Trend from PSID
8
based on the low-income food budget to one based on the economy food budget similar to the CPS thresholds. Thus, PSID-2 is identical to PSID-1 except that it uses a lower cost food budget. By definition, the poverty rate must be lower under PSID-2 than PSID-1.1 Beginning in survey year 1990, the PSID data file contains the Census Bureau threshold, which we refer to as PSID-3. This threshold is discussed on page 82 of the 1990 documentation (http://psidonline.isr.umich.edu/Data/Documentation/pdf_doc/psid90w23.pdf). Like PSID-1 and PSID-2, this threshold uses PSID methods to account for part-year co-residence. Also note that survey year 1990 corresponds to income year 1989 because respondents in any given survey year report income received during the prior calendar year. Given that PSID-3 is not currently available before 1990 and we want a Census threshold for the entire PSID data period, we developed PSID-4, a simplified version of the official Census poverty threshold. PSID-4 incorporates the weighted average poverty threshold for all families of size n, rather than assigning the specific poverty threshold for a family of size n, with x family members over age 65 and y children under 18. 2 Specifically, PSID-4 is the weighted average poverty threshold for a non-farm family (this differs from Census Bureau procedures only prior to 1980) of a given size in each year, where the family size is measured at the time of the interview following Census Bureau procedures. PSID-4 thresholds do not differ between elderly and nonelderly unrelated individuals, by number of related children under 18, or by family size
1
The income-to-needs ratio on the PSID data file is the family income divided by PSID-1 except for the case of farming families where the income-to-needs is equal to 1.25 times the family income divided by PSID-1. According to the PSID documentation, this adjustment accounts for the lower food cost for farmers and, therefore, yields lower poverty rates. Because of the low (and decreasing) prevalence of farming families, the difference in the poverty rates using the income-to-needs variable vs. PSID-1 is negligible and we do not show results using this measure. Like the variable used to construct PSID-1, the needs part of the income-to-needs variable on the PSID file is given in 1967 dollars and must be adjusted to account for inflation. This income-to-needs variable is available in most years up until the mid-1990s. 2 This simplification is particularly important in years prior to 1980, when the poverty threshold incorporated differences by farm/non-farm status and gender of the household head in addition to the number of family members above age 65 and below age 18.
Estimating and Benchmarking Poverty Rate Trend from PSID
9
greater than 9, as is the case for the official Census thresholds and PSID-1, PSID-2, and PSID-3 thresholds. While it is possible to match individuals in the PSID to their specific poverty threshold from the detailed Census matrix for each year after 1972, only the weighted averages for non-farm families are available on the Census Bureau web site before this time. The differences between the thresholds of families of the same size but different composition are relatively small and hence have a small effect on the calculated poverty rate for all persons.
2.2. Income and Weights
Family income in the PSID is defined as the sum of all labor, asset, and transfer income (AFDC/TANF, Social Security, etc.) for the head, spouse, and all others living in the family unit at any point during the calendar year. While the PSID collects data on food stamp and other non-cash government benefits, these benefits are not included in total family income because the Census Bureaus official poverty is rate based on money income, which excludes all non-cash income. PSID income is adjusted for partial year co-residence of family members as described above. Note that differences in the definition of family between the PSID and the CPS affect the number of people over whom income is summed and the relevant poverty thresholds. The names for the PSID income variables are provided in Table 2. We compute the poverty rate for all persons using the PSID individual weights. Several individual weights are available that account for sample design and selective attrition. The core PSID sample includes both the original 1968 Survey Research Center sample, which is a nationally representative cross-section, and the Survey of Economic Opportunity over-sample of low income households3. In 1990, a Latino sub-sample was introduced. Because this sample was discontinued after 1995, we do not include these respondents. In 1997/1999 an immigrant sub3
For more information on the sub-samples that make up the PSID or about the sampling frame, see http://psidonline.isr.umich.edu/Guide/ug/stdydsgn.html.
Estimating and Benchmarking Poverty Rate Trend from PSID
10
sample was added and has been included in every subsequent wave. We include these respondents because they have been interviewed just like permanent members of the core sample and PSID plans to continue to interview them indefinitely. The poverty rates that are reported in Table 3 for the years 1968 to 1996 include only core sample members and use the core sample individual weights. From 1997 onwards, we use the combined immigrant and core samples and associated weights. The variable names for the weights are listed in Table 2. The PSID staff recently developed two additional weights: the new longitudinal weight and a cross-sectional weight. The new longitudinal weight addresses some limitations of the old longitudinal weights for years after 1992. The cross-sectional weights assign a positive value to all individuals regardless of whether they have the PSID gene (i.e., whether they are PSID sample members). This approach, described in the PSID weight documentation, maximizes sample size. The cross-sectional weights and the old and new longitudinal weights are identical prior to 1993. These two weights are being beta tested and will be available on the PSID web site in 2008.
3. RESULTS
Table 3 reports the PSID poverty rates using each of the four thresholds discussed above. For comparison, the CPS-based official poverty rates are reported in the next-to-last column of the table. All rates are plotted in Figure 1, and the correlations between the series are reported in Table 4. According to PSID-1, the poverty rate was 17.98 percent in 1967; the rate then fell to 11.39 percent in 1973. Following an increase in 1974 and 1975, PSID-1 fell to 10.53 percent in 1980, rose to 13.79 percent in 1984, and then fell to 11.88 percent in 1989. The recession of the
Estimating and Benchmarking Poverty Rate Trend from PSID
11
early 1990s pushed poverty rates up; the economic expansion of the mid- to late 1990s pushed the rates back down to 10.12 in 2000, the lowest rate in the series. PSID-2, which uses the economy instead of the low-cost food budget, implies a lower poverty rate: in 1967 the rate is 11.81 instead of 17.98. PSID-1 is 40-70% higher than PSID-2, depending on the year. However, the time patterns of PSID-1 and PSID-2 are very similar, with a simple correlation of 0.87 over the period 1967 to 2002. Prior to 1973 the trends in PSID-1 and PSID-2 differ significantly from the official CPS series, with PSID-1 and PSID-2 showing greater declines in poverty during this period. The correlation between PSID-1 and the CPS-based estimate over the entire period is only 0.43; the correlation rises to 0.81 for the period from 1973 to 2002. The correlations for PSID-2 and the CPS are higher: 0.72 for the entire period and 0.91 after 1972. PSID-3, currently available only since survey year 1990, yields a time series of poverty rates that is highly correlated with the official series because it incorporates the Census thresholds. The correlation of PSID-3 and the CPS-based estimates for the 1989-2002 income years is 0.97. Table 4 shows that all four measures have a correlation of at least 0.90 during this period. Poverty rates using PSID-4 are highly correlated with the CPS series over the 1967-2002 perioda simple correlation of 0.82. PSID-4-based rates are correlated at 0.96 with PSID-3based rates for the 1989 to 2002 income years. In sum, the PSID can be used to estimate a time series of poverty rates that is similar to the published series. However, the PSID-1 series has the lowest correlation with the official rates -- only 0.43 over the entire period and 0.81 since 1973. If the goal is to achieve comparability with the official Census Bureau poverty series, then one should use PSID-3. However, the
Estimating and Benchmarking Poverty Rate Trend from PSID
12
PSID-3 thresholds are only available for income years after 1989. Thus, to create a series comparable to the Census over the years since 1968, PSID-4 should be used. In addition to concerns about comparisons between PSID- and CPS-based trends in poverty, the estimated levels of poverty are different in any given year. This fact was identified over 30 years ago by Lane and Morgan: The Panel Study finds somewhat fewer people poor. Whether the Panel Study or the Census is more accurate is uncertain. Unearned and irregular income which is important to low-income people tends to be underreported in surveys. It is possible that reporting improved through repeated interviews. This would suggest that the Panel Study data are more accurate. On the other hand, the Census samples are much larger. Furthermore, very poor people may be among those most likely to drop off a panel study, and this loss may not be completely compensated for by adjustments which have been made for nonresponse. These considerations would suggest that Census is more accurate. [2] Furthermore, if Lane and Morgans hypothesis that income reporting improves as respondents are repeatedly interviewed is correct, then this could explain why the PSID-based estimates show greater decline than CPS-based estimates prior to 1973. After having been in the PSID for a few years, PSID respondents may have become more accurate income reporters. A final table compares estimated poverty rates using the three different weights: the longitudinal weight that has been on the PSID file through the years, the new longitudinal weight, and the cross-sectional weight. For this exercise we examine threshold PSID-3, the method most comparable to the official Census Bureau approach. Recall that the weights are identical prior to income year 1992, so we only report rates from 1992 onwards (Table 5). We find that the rates are nearly identical regardless of the weight that is used. In addition, the correlation between the CPS-based estimate and PSID-based estimate with the historical
Estimating and Benchmarking Poverty Rate Trend from PSID
13
longitudinal weight, the new longitudinal weight, and the cross-sectional weight are quite similar at 0.972, 0.959, and 0.934, respectively.
4. CONCLUSION AND RECOMMENDATION TO USERS
There has been some confusion among users about the appropriate way to calculate poverty rates using the PSID. This report guides users through this process, identifying some pitfalls that may arise and considering alternative ways to calculate poverty rates. We hope this report will reduce confusion about how to identify which PSID individuals are in poverty. An important lesson is that if users want estimates using an approach similar to the Census Bureaus official approach, they should use PSID-3 for the period 1990 onwards; if they want to examine earlier years, then PSID-4 is the best currently available option; PSID-4 is available from the authors and will be posted to the PSID web site in 2008. In the future, the PSID staff will create the PSID-3 version of the thresholds for years prior to survey year 1990. The annual poverty rates derived from the PSID are lower than the rates in the CPS, an observation that was made in the 1970s. Most importantly, the gap between the PSID and the CPS that existed in the mid-1970s has remained steady through the most recent period, and, as a result, the PSID shows trends quite similar to trends observed with the Census Bureaus official rates.
Acknowledgements Rebecca Blank, Johanne Boisjoly, Greg Duncan, Peter Gottschalk, Ann Huff-Stevens, and Kate McGonagle provided helpful comments on previous versions of this paper. This project was supported in part by funds provided by the U.S. Department of Health and Human Services, Office of the Assistant Secretary for Planning and Evaluation, #5 U01 PE000001-05 and #1 U01 AE000002-01. Any opinions and conclusions expressed are solely those of the authors and should not be construed as representing the opinions or policy of any agency of the Federal government.
Estimating and Benchmarking Poverty Rate Trend from PSID
14
References
[1] Gouskova, Elena, and Robert F. Schoeni, Comparing Estimates of Family Income in the Panel Study of Income Dynamics and the March Current Population Survey, 1968-2005 (2007). Available at PSID website: http://psidonline.isr.umich.edu/Publications/Papers/Report_on_income_quality_v3.pdf [2] Lane, Jonathan, and James Morgan, Patterns of Change in Economic Status and Family Structure,...