September 27, 2005

Rise, Peak and Decline: Trends in U.S. Immigration 1992 - 2004

VIII. Immigration Measures from the Census, ACS and CPS

Our goal of measuring trends and patterns in annual immigration—specifically in the number of foreign-born persons coming to the United States in a given year with the intention of living in the country—and its characteristics thus requires expansion of the data and methods brought to bear on the problem. In assessing immigration levels and trends, we use three sources of data:

(1) The March Supplements to the Current Population Survey, the annual socioeconomic supplement to the U.S. government’s monthly labor force survey, for 1994–2005, with two different weighting schemes, one covering 1994–2001 and the other overlapping to cover 2000–2005;

(2) Sample data from Census 2000; and

(3) The American Community Survey (ACS) national sample for 2000–2004.8

Two different types of data from these sources enter into our measures. The March CPSs and the ACSs include a question asked of all respondents at least 1 year old about residence one year earlier. Thus, the number of foreign-born who lived abroad one year before the survey data provides an assessment of the inflow of immigrants in the year before the survey. The estimates coming from the analyses of this question, particularly from the CPS,9 give figures slightly lower than other measures, but, as we will see, the pattern of change across the 1990s and 2000s is consistent with other measures.

The CPS, Census 2000 and the ACS all ask immigrants when they “came to live in the United States.” While there is some controversy about the interpretation of these data in relation to new arrivals of immigrants,10 their use to assess migration flows is well established and the results track well with other measures. Our approach involves averaging arrival cohorts across several survey years for both the ACS and CPS. It is also necessary to annualize the CPS estimates because the data are only available for arrival in groups of years, not for single years.

Further complications are introduced by the use of CPS data spanning the 1994– 2005 period. CPS population figures arise out of a complicated weighting process that uses postcensal population estimates, generated by the Census Bureau and benchmarked to the previous census. In most decades, this switch in weighting from the previous census to a new one introduces only minor discontinuities. However, the Census Bureau underestimated the population count from Census 2000 by almost 5 million persons (Passel 2001) so that CPS data developed with weights based on the 1990 Census differ considerably from similar data developed using weights based on Census 2000. The two groups most affected by the reweighting that occurred officially for March 2002—Asians and Hispanics—are immigrant-dominated populations. Thus, the reweighting process had a marked effect on the size of the foreign-born population as measured by the CPS. Because of the magnitude of the changes introduced by the weighting change, the Census Bureau has released two transitional, overlapping CPS datasets. For March 2001, there is an entire alternative dataset available for producing measures consistent with data from 2002 and later.11 For March 2000, the Census Bureau supplied an alternative set of weights consistent with the regular March supplement and Census 2000 (Passel 2001). Thus, we have two CPS measures for 2000 and 2001 to help assess the impact of the new census results on measures of immigrant inflows.

Year/Period of Arrival

The CPS, the ACS and Census 2000 all include a series of questions asking place of birth, citizenship and the year that respondents born outside the United States “came to live” in the United States. These questions define the immigrant or foreign-born population as persons born outside the United States who were not US citizens at birth. The estimates in Detailed Tables 2a–2d3 labeled “arrival” are derived from the Census, ACS and CPS questions on the year that respondents came to live in the United States.

Census 2000. The estimates from Census 2000 use the 5-percent Public-Use Microdata Sample (PUMS) as augmented with assignments of legal status (Passel, Van Hook and Bean 2005). The estimates are straightforward tabulations of the foreign-born population for single years of arrival by race/Hispanic origin,12 legal status, country of birth or state of residence. The only estimation required is for the year 2000. Census data for arrivals in 2000 are incomplete because the census enumerates persons at their residence as of April 1, 2000. Persons entering the U.S. after April 1 should not be included in the census count; however, the census enumeration process actually spreads over a period of several months after April 1 as enumerators collect information from persons who failed to respond initially to the Census. Accordingly, we treat the data for persons arriving in 2000 as representing one third of the entrants during the calendar year rather than the one fourth share implied by the Census date. Thus, the estimate from the 2000 PUMS of 589,000 immigrants arriving in 2000 is inflated to the figure of 1,765,000 for the full year shown in Detailed Table 1a.

American Community Surveys, 2000–2004. The detailed data collected in the American Community Survey are basically the same as collected in the sample phase of the Census. The ACS design, however, differs in that the survey consists of non-overlapping samples collected on a monthly basis over the course of the calendar year rather than the traditional census or survey method of collecting all data for a specific reference date. For our ACS-based estimates, we again use tabulations from the PUMS. For 2000–2004, the ACS was not in full production; the sample size will be 250,000 households per month, or 3 million per year, but during testing the ACS had a nominal sample size of 700,000 households over the course of the calendar year. Because only about one third of nonresponses are followed up in the ACS, the actual sample size for each year was about 480,000.13 The ACS, as implemented for these years, covers only the household population, not the whole population, of the United States; accordingly, the ACS figures for immigrants are slightly less than the Census values in 2000 and are a bit lower in other years than the entire U.S. total would be, were it available.

As with the Census-based estimates, we use straightforward tabulations of the foreign-born population by single years of arrival from the ACS as our initial estimates for race/Hispanic groups,14 countries of birth and states of residence. There are several data-driven limitations to the ACS estimates, though. Because of the smaller sample size of the ACS, not all of the countries of birth coded in Census 2000 were available in the ACS; the impact of this limitation is small, however, since all large sending countries are separately identified in the data. We did not generate estimates of inflows by legal status because the status assignments have not been done with ACS data.

The ACS estimates for each year of arrival presented in Detailed Tables 2a–2d3 are an average of the five individual ACS estimates (i.e., from the 2000–2004 ACSs). The principal technical issue in developing these averages is the handling of estimates of the number of immigrants arriving in the year of the survey, because the data are collected throughout the year rather than at the end of the year. Thus, the tabulated ACS data from 2003, for example, do not generate an estimate of the full number of immigrants arriving in 2003. This deficiency affects the estimates for 2000–2004 only, not those for years before 2000. Immigrant arrivals in each year through 1999 are averages of five individual estimates. For the 2000–2003 estimates, the survey conducted in the year of the estimate is dropped from the average; that is, the average estimate for 2000, as shown in Detailed Tables 2a–2d3, is based on four separate estimates from the 2001–2004 ACSs; for 2001, from the three ACSs of 2002–2004; for 2002, from the two ACSs of 2003–2004, and for 2003, from the 2004 ACS alone. The estimate shown for 2004 is obtained by inflating the partial-year data from 2004 to a full-year estimate. The inflation factor for 2004 is the average of factors for 2000–2003 obtained by computing the ratio of the partial-year estimate from the year’s survey to the full-year estimate from the next year’s ACS. For example, the inflation factor for the 2000 ACS is the ratio of immigrant entries in 2000 from the 2000 ACS (a partial year) to immigrant entries in 2000 from the 2001 ACS (a full year).

Current Population Surveys, 1994–2005. The Current Population Survey is the U.S. government’s labor force survey that provides the monthly data used for the official measure of the unemployment rate. Beginning in 1994, the information needed to identify immigrants and to measure immigration (country of birth, citizenship and year of entry) was added to the data collected in the basic CPS and its supplements as part of the redesign process done every decade. The regular monthly CPS is a complex state-based sample of about 50,000 households; each month’s sample overlaps with the previous year and adjacent months to minimize the variance of estimates of month-to-month change in employment. In March of every year, the sample size is increased by augmenting it with CPS households from other months; the data collected are greatly expanded also. The additional data include information on income (by source of income), health insurance, participation in social programs, detailed labor force history and a variety of other items. The CPS universe is the civilian, noninstitutional population; this universe is larger than the ACSs—the household population—but smaller than the Census’s universe—the total population.

Through 2001, the March CPS expansion increased the sample size by about 10% over the regular monthly sample to 55,000 households by doubling the sample of Hispanic households. The weights for the CPS are based on the Census Bureau’s population estimates obtained by updating the previous census. So the population estimates for the March 1994–2001 CPSs are based on the 1990 Census, as corrected for undercount, using detailed national estimates by age, sex and race; national estimates by age, sex and Hispanic origin; and state estimates of the population aged 16 and over. In general, figures for the foreign-born population from these 1990-based CPSs are less than those from 2000-based CPSs.

In the post-2000 redesign of the CPS, the March supplement sample size was increased substantially to about 80,000 households by oversampling households with children and minority households other than Hispanics (which were already oversampled).15 The weighting scheme was also changed. The weights for the post-2000 redesign are based on Census 2000, without a correction for undercount, using the same detailed national categories for age, sex and race; national estimates for age, sex and Hispanic origin; and state estimates but for all ages and with detail by broad age groups, sex and simplified race groups. The redesigned questionnaire, expanded sample and new weighting scheme were introduced for the March 2002 supplement. Beginning in 2003, the race question was changed to allow for multiple responses as in Census 2000 and the ACS; to incorporate these new race data, the weighting categories were adjusted slightly by placing all persons with multiple race responses into a separate weighting group. For March 2001, the Census Bureau tested the entire redesign including the greatly expanded sample (but not the new race question) and released a separate set of CPS data for that year that are consistent with the data for March 2002 and later. In addition, for the March 2000 supplement, the Census Bureau produced a special set of CPS weights consistent with the population as enumerated in Census 2000 rather than the official CPS data for March 2000 which are weighted to a population estimate based on the 1990 Census carried forward to 2000. As a result of the redesign and testing, we have a series of March CPS supplements for 2000– 2005 that are consistent with Census 2000 and that overlap for two years with the 1990-based series of 1994–2001.16 We use the CPS supplements to produce two separate series of immigration estimates by race, by country of birth,17 by legal status18 and by state of residence— a 1990-based series and a 2000-based series.

Producing annual estimates of immigrants arriving from the CPS question on year of arrival in the United States is more complicated than the analogous process for Census 2000 and for the ACSs of 2000–2004. Although the CPS collects information on year of arrival for every year, the data released to the public group the responses into intervals. Thus, for the period of interest (i.e., 1992 and later), the CPS provides information only for two-year periods of arrival, beginning with even-numbered years; that is, the data responses are coded as arrived in 1992–93, 1994–95, 1996–97, etc. To produce annual estimates of arrivals, we assign each year of the two-year period the same value equal to one half of the estimate for the two-year period. A further complication is introduced because the most recent period of arrival in every CPS is expanded to include the current year of the survey. Also, since all of the coded intervals must begin with an even numbered year and must include at least two full years (according to the Census Bureau’s data disclosure standards), the most recent period of arrival is coded to cover three full years plus the partial year of the survey in odd-numbered years. Since the March CPS interviews are done during the week including March 19, the most recent arrival period shown in even-numbered years covers two full years and 2.5 months (i.e., 1992–94 for the March 1994 CPS; 1994–96 for March 1996, 1996–98 in March 1998, etc.). In odd-numbered years, the most recent period of arrival encompasses three years and 2.5 months (i.e., 1992–95 for the March 1995 CPS; 1994–97 for March 1997, 1996–99 for March 1999, etc.). To estimate annual arrivals by calendar year for these most recent periods, we divide the number of immigrants arriving in the interval by the number of years covered (either 2.208 or 3.208) and assign the average value to each year covered by the interval, including the year of collection. This averaging across periods tends to smooth out the time series of estimates and produce estimates that change in “steps” when compared with the series from the ACS and Census. (Note, for example, in Detailed Tables 2a–2d3, that the 2000-based CPS values for immigrants arriving during 1990–1999 are identical in adjacent years whereas the ACS and Census estimates vary from year to year.)

The CPS-based estimates from the period of arrival data are also averaged across multiple CPSs. Because of sampling variability and attrition of some arrival cohorts due to the combined effects of small sample sizes, emigration and mortality, not all of the CPSs could be used for every estimate year. However, all periods of arrival, with a very few exceptions, draw on data from at least three CPS supplements; the following table details the supplements used to estimate arrivals for each calendar year:

In averaging the annual estimates across multiple CPS observations, each CPS is given an equal weight except that the annualized estimates for the year of the CPS supplement (e.g., the estimate for 2002 based on the March 2002 CPS) are only weighted at 0.208 or a weight equivalent to an observation covering 2.5 out of 12 months.

Because of the multiple averaging involved in developing estimates for calendar year arrivals from CPS data, the CPS-based time series of arrivals tend to be smoother than the other estimates (Detailed Tables 2a–2d3). Because they are averaged across time and CPSs, the increases to peak values tend to start a year earlier than other estimates, the peak values sometimes last a year longer than other estimates, and the declines from the peaks tend to start a year later than in some series. There are substantial differences in the estimates developed from 1990-based CPS weights (the series for 1994–2001 CPS data) and those developed from the 2000-based CPS weights (the series for 2000–2005 CPS data). The 1990-based population estimates were substantially different from the population figures that emerged from Census 2000. Specifically, the 1990-based population estimates understated the Hispanic and Asian populations by a bit more than 10%, with some of the age-sex groups (e.g., Hispanic males ages 20–29) falling as much as 20 to 30% below the Census 2000 counts. Further, the state-level estimates for states that were new destinations for immigrants, especially for Hispanic immigrants, were lower than Census 2000 counts, especially for the minority populations in these states. The introduction of Census 2000-based population data into the weighting process thus led to significantly higher figures for immigrants. The overall foreign-born totals, in Detailed Table 3, show that the 1990-based totals for 2000 and 2001 are about 6% less than the 2000-based figures. The impact of new weights is, of course, much larger for Hispanic and Asian populations and for more recent arrival cohorts. Thus, comparison of the averaged estimates for arrival cohorts in Detailed Table 2a shows that the Hispanic figures for 1990–2001 arrivals run about 20% lower for the 1990-based estimates; for Asians and blacks, the 1990-based estimates also are lower but with more variability than for Hispanics; for white immigrants, the 1990-based estimates are actually slightly larger than the 2000-based series. Whereas some of these differences are substantial, they should not be taken as indicative of flaws in the overall estimates; rather, they demonstrate the degree of error that had built up in the population estimates and data series for the 1990s due to inadequate information about immigrant flows to the United States (Passel 2001).

Residence 1 Year Ago

The ACS asks respondents “Where did this person live 1 year ago?” while the CPS asks “Where did [respondent] live on March 1, [insert previous year]?” This question is asked of respondents more than 1 year old who report living in a different house or apartment on the reference date. We use these data to identify immigrants during a year—defined as foreign-born individuals who report living outside the United States in response to these questions. Census 2000 has a similar question but the reference date is five years before the census (i.e., April 1, 1995) so we do not use these data in this study.

American Community Survey, 2000–2004. The development of estimates from the ACS information on residence one year earlier is very straightforward compared with the year of immigration variables. We first tabulate foreign-born individuals who report living outside the United States a year ago by race, by country of birth and by state of residence using the procedures described above for these variables. The only issue in combining the estimates is the date of reference for the migration period. Since the ACS data are collected through the calendar year, we assume that, on average, the data were collected at the midpoint of the survey year; for example, we assume that the 2003 ACS data were collected as of July 1, 2003. Thus, the reference period during which migration is assumed to have occurred is July 1 of the year before the survey through June 30 of the year of the survey. To convert these estimates to calendar years, we average estimates from adjacent ACS years with weights of 0.5; for 1999 estimates, we use one half of the value from the 2000 ACS and assume that it applies to the entire calendar year. Likewise for 2004, we assume that the ACS value for migration applies to the whole year. The results of this averaging process are shown as estimates of immigration (labeled as “yr ago”) for 1999–2004 in Detailed Tables 2a–2d3.

Current Population Surveys, 1994–2005. Again, the CPS estimates are somewhat more complex than the ACS estimates, but the overall process is very similar. Again, we first tabulate the number of foreign-born individuals who lived outside the United States one year before the survey from each March CPS Supplement by race, by legal status, by country of birth, and by state. We again have two separate series of estimates—based on the 1994–2001 CPSs for 1990-based weights and on the 2000–2005 CPS for the 2000-based weights. As before, each series is treated separately. Because of the data deficiencies noted above, we do not use country of birth data from the 1994 CPS, and the values for other years have been edited to reduce the number of unknowns; the race data for 1994–95 have been edited to correct Census Bureau processing errors and the data for 2003–05 edited to assign multiracial individuals to single races. We do not have legal status information for the 1994, 1997 and 2005 CPSs, so the annualized estimates for years affected by these surveys draw on less information.

Finally, the CPS data required one further modification. In processing the data, it became apparent that there were serious problems with the residence one year earlier data from the March 1995 CPS. Specifically, the estimated annual immigration based on this survey was only about 55% of the estimated values from the 1994 and 1996 CPSs. However, there was no indication from any other data source (including the year of arrival data) to suggest that the numbers were so low during the period. It appears that an additional question in the March 1995 CPS asking about residence five years earlier may have affected responses to the residence one year ago question. Because of these data problems, we dropped the 1995 data from this analysis.19

The conversion of basic CPS tabulation of immigrants in the year before the survey followed the annualization process described above for averaging ACS data. The reference period of the migration is specified more clearly in the CPS than in the ACS, so the process for developing calendar year estimates is also better specified. Specifically, the annual estimate of immigration for the year before the CPS using the residence one year ago data weights the current CPS for 10 months and the previous CPS for 2 months. For example, the number of immigrants arriving in calendar year 1998 is computed as 10 times the foreign-born population living outside the U.S. one year ago identified in the March 1999 CPS plus twice the number from the March 1998 CPS with the total divided by 12 (i.e., 10+2). If either year is missing (as with 1995 data as just noted or, for example, 1997 data on legal status), an adjustment is made in the denominator so that, in essence, the migration rate observed in the available CPS is assumed to apply to the entire calendar year.

Detailed Table 2a–2d3 shows the two separate series of annualized estimates based on CPS data for residence one year ago. Again, the weighting has an impact on the estimates as the 2000-based weights give measures of immigration about 10% higher than the 1990-based weights. There is, however, little overlap in the series as only the 2000 and 2001 CPSs carry two sets of weights. In contrast to the period of arrival estimates, the ACS tends to yield somewhat higher estimates of immigration from the residence one year ago question than the CPS estimates. Because the time reference in the CPS is more specific than in the ACS (i.e., “March 1” versus “1 year ago”), the ACS may pick up more migrants and movers whose actual date of moving occurred slightly more than one year before the interview.20 Alternatively, the ACS-CPS difference may be due to sampling variability or differences in the population estimates used in weighting the surveys.

Combined Estimates

To arrive at a single set of immigration values for calendar years 1992–2004, we take a simple average of the estimates available for each year, but restrict the averaging to estimates based on weights consistent with Census 2000. Specifically, the estimates incorporated into our overall average are: (a) Census 2000 estimates based on year of arrival for 1992–2000; (b) average ACS estimates based on year of arrival for 1992–2004; (c) average CPS 2000-based estimates based on year of arrival for 1992–2004; (d) average ACS estimates based on residence abroad 1 year ago for 1999–2004; (c) average CPS 2000-based estimates based on residence abroad 1 year ago for 1999–2004. In averaging these estimates (most of which are averages themselves), we are trying to address questions surrounding the trend in immigration over this period, not the level of immigration. The different measurement techniques provide somewhat different information about levels of immigration but all show the same trends over time. Thus, the average across methods should not be treated as the single best measure of the number of immigrants entering the country, but rather as an indicator of the relative contributions of different groups (e.g., race, legal status, geography) to immigration and of how these contributions have changed over time.

Figure 10 shows the estimated annual level of immigration for all of the series of annual estimates based on different methods; Figure 11 shows the same level of detail, but for immigration from Mexico. The variability across different estimate methods is clear in both figures as the lines show a range of several hundred thousand from highest to lowest for a number of years. The largest high-low differences tend to be in peak immigration years, however. What is also apparent from the figures is that all of the separate estimates series show similar patterns of a relatively low plateau at the beginning, an increase to peak values, followed by a decline from the peak.

  1. For 2000 and 2001, the ACS was designated as the Census Supplementary Survey. For clarity and conciseness, we refer to both as ACS samples, for 2000 and 2001 respectively.
  2. The overlapping sample design of the CPS may affect measures of residence one year earlier because half of the sample housing units stay in the CPS from the previous year. Also, edit and imputation schemes may affect the final data.
  3. Redstone and Massey 2003; Myers 2004.
  4. This CPS, designated by the Census Bureau as the SCHIP (State Children’s Health Insurance Program) sample, is completely separate from the regular March 2001 Supplement. The SCHIP sample incorporates the sample expansion introduced in March 2002 and uses the new weighting methods introduced in 2002 as well as population controls based on Census 2000. See below.
  5. The race data in Census 2000 allow for individuals to be assigned to more than one race group. For consistency with historical data and for simplicity of presentation, we assign multiple-race individuals to single races using a hierarchical assignment process. The hierarchy for assignment is: Black, Asian or Pacific Islander, white or “some other race” (assigned as white), and American Indian or Alaska Native. (See Suro et al., 2005.)
  6. For 2001–2004, data from all ACS households were released in the PUMS. For the Census 2000 Supplementary Survey (C2SS) which represented the ACS for 2000, only about one quarter of the households in the full ACS sample were released in the PUMS.
  7. All years of the ACS allow multiple-race responses. Accordingly individuals in the PUMS were classified by race using the same hierarchical methodology as applied to Census 2000.
  8. For details of CPS design, weighting and methods, see U.S. Census Bureau 2000 and 2002.
  9. For 1994 and 1995, the official CPS data from the Census Bureau and the Bureau of Labor Statistics (BLS) have weights and race data that are incorrect and inconsistent with subsequent data because of errors in the collection and editing of the race item. We use a set of weights produce by the author for the Urban Institute that correct the race data and CPS weights (Passel 1996). The error mainly affects data for Asians and American Indians; the correction adds substantially more than 2 million persons to the foreign-born population.
  10. We augment the official data on country of birth by editing the data to assign country of birth to about three quarters of the more than 1 million persons reported in the CPSs of 1999 and later as “country of birth unknown” or “some other country.” The Census and ACS data do not include equivalent categories because they incorporate more extensive edits to eliminate such responses (or, more properly, such nonresponses). The data on country of birth collected and released for March 1994 cover only about 30 specific countries, rather than the full complement of countries and regions of birth covered in other CPSs, Census 2000, and the ACS. Accordingly, we do not use data by country of birth from the 1994 CPS.
  11. The legal status data use the methodology first created by Passel and Clark (1998) for the Urban Institute and later adapted to Census 2000 (Passel et al. 2004, 2005). The legal status assignments have not been implemented with CPS data for 1994 and 1997 and the assignments for March 2005 have not been completed, so the estimates by legal status developed here exclude those CPSs.
  12. The March 2005 CPS Supplement also included this question about residence 5 years ago in addition to the question on residence 1 year ago. There are some indications that the residence 1 year ago data from 2005 may be somewhat “out of line” with other data, but the 2005 figures are higher than expected, not substantially lower.
  13. The CPS changed the wording of its question in March 2005 to correspond to the ACS wording. The March 2005 CPS shows much higher levels of immigration in response to this question than does the March 2004 CPS. This difference may represent a real increase, but part of it may be attributable to the response pattern noted in the text.