How low response among Latino immigrants will lead to differential undercount if the United States’ 2020 census includes a question on sensitive citizenship

1.Introduction

The U.S. Census Bureau has been consistently optimistic about the viability of streamlined census procedures introduced as part of “modernizing” and “re-engineering” Census 2020.

However, serious data quality problems are likely to emerge in 2020 from adopting these methodological approaches which have been driven primarily by cost-containment efforts – especially if the 2020 decennial census includes a sensitive question asking respondents to report the citizenship status of everyone living at the place where they live. Review of the consequences of adopting these “modernized” census implementation methodologies to compensate for low and uneven response rates in any nation where economic, political, and racial/ethnic disparities persist should lead less-developed countries and organizations such as UNESCO’s Institute of Statistics to resist the temptation of following this pathway toward statistical improvement.

In order to assess the impact of adding a citizenship question to Census 2020 on immigrant census response rates in a region of California where there is dense settlement of Mexican-origin immigrants, the San Joaquin Valley Census Research Project (SJVCRP) conducted a survey in September-October, 2018 to examine Hispanic (and other) immigrants’ willingness to self-respond, to respond to enumerator visits, and to requests for proxy interviews seeking information on their neighbors – in a census without the citizenship question or in one that included it.

Using findings from this survey research and Census Bureau descriptions of planned data-gathering and analytic procedures, I have developed a “cascade model” framework to project how differential patterns of census response resulting from a census that includes the sensitive question on citizenship will give rise to differential undercount and distort the reported demographic profile of regions such as the San Joaquin Valley.

Drawing on findings from this large survey of Latino immigrants, I analyze the consequences of adopting these newly “modernized” procedures for Census 2020 enumeration in the large San Joaquin Valley of California, a region where there is a high concentration of immigrants, coupled with stark sociopolitical and economic disparities. The analysis shows that adoption of many of the U.S. Census Bureau’s re-engineered methodological procedures, while promising cost savings, are likely to degrade data quality unacceptably if, as the San Joaquin Valley survey research indicates, the added citizenship question, gives rise to high, but uneven, levels of non-response in a region.

The new Census 2020 procedures for the United States, while probably satisfactory for enumerating easier-to-count populations and communities with a census questionnaire containing only the nine basic questions included in the traditional “simple” census, can be expected to have serious limitations when utilized in immigrant-dense neighborhoods and communities in a census that includes a question on citizenship. Chronic operational problems in census enumeration procedures will inevitably be exacerbated in areas such as the San Joaquin Valley by inclusion of a census question asking respondents about their own citizenship status and the status of everyone living at the same place. It will differentially suppress both willingness to respond and accuracy of responses among sub-populations.

The Cascade Model analysis presented here addresses a central methodological question underlying the ongoing legal controversy about efforts to add the citizenship question that has been considered by the U.S. Supreme Court in April, 2019. That question is whether “modernized” approaches to non-response follow-up (NRFU) and new statistical procedures can successfully compensate for problems of widespread unwillingness to respond to a statistical data collection initiative widely perceived as unwarranted or dangerous by non-citizens, their families, and social networks. In the international context of statistical policy, this relates to underlying issues as to what sorts of central government survey designs are viable and justifiable.

2.Structure of the Cascade Model

The analysis of how differential non-response is transformed into differential census undercount is referred to as a “cascade” model because the level of success and failure at each stage in decennial census operational procedures, determines the parameters for census operations at the next stage.

Level of self-response, for example, determines, extent of reliance on enumerator efforts to secure information from the households that fail to self-respond during NRFU. Enumerator success in this endeavor subsequently determines the extent of reliance on administrative records and the availability of these records determines the extent of reliance on proxy interviews for information on households. Finally, cumulative success in securing data from households in each of these efforts to compensate for lowered levels of self-response, determines the need for reliance on count and whole-person imputation.

The accuracy of the “census count” (as well as the census-based demographic profile of the population) depends on the level of reliance on each data-collection or analytic operation since some of these data-collection and analytic procedures (e.g. proxy interviews, reference to administrative records) are more error-prone than others.

Despite the valiant but compromised efforts by the Census Bureau to generate accurate census tabulations in the face of greatly elevated and uneven non-response, the attrition in data quality at each step in the census process will very likely result in flawed tabulations of both the size and demographic characteristics of the region’s population.

The “cascade” of census stages in decennial census collection/imputation are visualized in the Cascade Model of undercount of the San Joaquin Valley Hispanic 1st and 2nd generation immigrants in Fig. 1 below.

Figure 1.

The Cascade Model of diminishing data quality in successive census operations.

The Cascade Model incorporates empirical data collected in the SJVCRP survey of Hispanic 1st and 2nd generation immigrants to estimate Census Bureau’s success/failure in the following stages of the process: MAF-building, self-response, NRFU “direct” interview response, and proxy interview response. It also relies on the SJVCRP survey data to estimate the systematic undercount resulting from use of hot-deck imputation due to differences in size between the households likely to respond to the census and those likely not respond.

This type of analysis underscores a crucial consideration – that any government survey process involves interactions between often-conflicting systems for visualizing and exchanging information – that of the official institution conducting the survey embedded in the methodology it has chosen to use for data collection and analysis, and the real-world sociocultural system(s) of the respondent population. These systems include language, conceptualization of “household”, literacy, digital literacy, together with the sociological and “structural” context of the communities where a household lives (e.g. size and configuration of immigrant social networks in a community, residence patterns, housing arrangements, patterns of work, postal service, Internet access, interactions with agencies which generate administrative records contemplated for use in imputing household characteristics).

3.SJVCRP study population, survey sample, and why findings are relevant internationally

California’s San Joaquin Valley is a unique region within a highly-developed country, making analysis of factors affecting census accuracy in this particular sub-state region internationally relevant. Although California has large, highly-developed urban centers, it also has extensive agricultural areas. The San Joaquin Valley is the largest of these labor-intensive agricultural areas that have drawn Mexican immigrants northward for almost a century. It has a 2017 population of about 4.2 million persons – more than the population of major urban municipalities such as the City of Los Angeles, Houston, or Chicago [1].

The reason analysis of census operations in this societal context is relevant is that the San Joaquin Valley region has a very high proportion of immigrants (21.7% foreign-born vs. 13.7% for the United States overall), because there is high-level of ethnic/racial diversity, economic inequality, and sociopolitical tension between immigrants and the federal government. These factors play an important role in determining census response rates initially, given the sensitivity of the citizenship question for immigrants, and subsequently in determining the viability of standardized data collection and imputation procedures that have serious limitations in this distinctive sociopolitical context.

The analysis focuses on 1st and 2nd generation Hispanic immigrants (i.e. 1st generation – foreign-born, 2nd generation – sons and daughters of foreign-born parents) – because they make up more than one-third of the region’s entire population. Moreover, a high-proportion of the adult foreign-born Hispanic population, those 18+ years of age who are potential census respondents, are undocumented and, therefore, their census response is likely to be disproportionately affected by inclusion of the sensitive question on citizenship.

Based on expert estimates of numbers of undocumented immigrants by region of origin, prepared by the Center for Migration Studies of New York using ACS data, I estimate that about 8.5% of the region’s entire adult population is made up of undocumented Mexican or Salvadoran immigrants, while about 5.3% are legal residents and 6.2% are naturalized Citizens [2]. Based on Census Bureau analysis, I estimate that about 14.5% of the region’s adult population are sons or daughters of foreign-born Mexican or Central American parents [3].

The cascade model of undercount draws on San Joaquin Valley Census Research Project (SJVCRP) survey data collected in face-to-face interviews with 418 first- and second-generation Latino immigrants during September-October, 2018 [4]. The survey data suggest there may be unprecedented low levels of census response among Hispanic immigrants in Census 2020 stemming from inclusion of the question about citizenship status (which is generally understood as being an inquiry about immigration status). Lowered, and uneven, levels of response within immigrant social networks would initiate a cascade of Census Bureau procedures to compensate for non-response that, unfortunately, have limitations that cannot be easily overcome in these sorts of communities. The Cascade Model then estimates the level of differential census undercount for four groups: undocumented, legally resident, naturalized 1st generation immigrants and the 2nd generation U.S.-born in the event the citizenship question is included.

Because the survey data show differential levels of response within the Hispanic immigrant population as well as in comparison to non-immigrant non-Hispanics, there is reason to be concerned that the resulting bias in census-derived demographic profile of the population will make tabulations of race/ethnic and age characteristics of the entire region unreliable.

Although some coefficients in the cascade model cannot be definitively specified – because the U.S. Census Bureau’s operational procedures may evolve further and the societal context in which the decennial census is embedded may further evolve in the months before Census 2020 takes place (e.g. due to shifts in immigration policy and enforcement), the model does provide a framework for identifying particularly problematic operations and procedures and provides a systematically-derived estimate of resulting differential undercount in Census 2020.

A major uncertainty in projecting patterns of census response in 2020, of course, relates to whether the Supreme Court decision (expected in June, 2019) forces the U.S. Department of Commerce to abandon its plan to include the question on citizenship. However, even if the citizenship question is removed, it is possible that there may be lingering concerns among immigrants about the confidentiality of census responses and suspicion that census response is risky [5].

4.How robust are the planned Census 2020 NRFU procedures in the San Joaquin Valley?

A central concern is that the reliability of the Census Bureau’s planned data collection strategies has not yet been assessed or tested in the distinctive societal context of immigrant communites using the newly-modified survey instrument that now will probably include a highly-sensitive question about citizenship. The 2018 End-to-End Test did not include a questionnaire with the citizenship question.

The U.S. Census Bureau will be conducting research on the impact of the citizenship question on census self-response during the summer of 2019 and reporting preliminary findings about six months before Census 2020 operations begin. But the analysis may well be too late to change Census Bureau staffing level, operational guidelines, and data collection methodology for NRFU [6]. Moreover the Census Bureau’s 2019 study, while well-designed, will only examine self-response and will not generate insights about households’ responsiveness to enumerator contacts during non-response followup or responsiveness to requests for proxy interviews. These are serious limitations – because in a scenario where there is a high level of non-response as is the case in the San Joaquin Valley, the NRFU phase of data collection is likely to account for about half of all data collected.

5.Understanding the dynamics of census undercount in the San Joaquin Valley region

The “cascade model” of undercount described here draws on earlier researchers’ powerful insights that patterns of differential undercount do not stem from certain populations being intrinsically “hard to count” but, rather, from the interactions between the census system of enumeration and the population being enumerated.

Two lines of research have been particularly fruitful in improving researchers’ understanding of the dynamics of undercount – a multi-stranded research initiative in the Los Angeles basin as part of the 1986 Test of Adjustment Related Operations (TARO) and a subsequent national initiative as part of the 1990 Decennial Census – the ethnographic alternative enumeration research conducted by the Census Bureau’s Center for Survey Research.

Based on the Los Angeles research, a sound theoretical framework to understand the causes of census undercount, not simply correlations, emerged. The powerful insight stemming from this research is that undercount results from conflict between the census system processes for enumeration/data collection and perspectives on and patterns of information-sharing at play in the social universe in which potential survey respondents live.

Within this over-arching theoretical framework, the Census Bureau’s ethnographic research program then made important contributions to understanding undercount in a range of specific local conditions. It analyzed how different interactions between the socioeconomic and cultural context of potential census respondents’ lives and census operations gave rise to differential undercount of low-income minority and immigrant families. The program meticulously documented undercount in 31 distinct ethnic groups in different communities across the U.S. [7]. Related ethnographic research conducted by Diego Vigil in connection with the 1986 TARO in the Los Angeles basin and by Phillipe Bourgois in New York documented specifically how distrust in government among immigrants constrained census response even at that point in time – decades before the current national debate on immigration policy began to make daily news headlines [8, 9].

The Cascade Model-based estimates of differential undercount among sub-populations of Hispanic immigrants to the United States are driven by the distinctive conditions observed in the SJVCR survey of the specific sociologically distinct ethnic/racial sub-population of Mexican and Central American immigrants. Although the analysis relates specifically to the San Joaquin Valley context, the over-arching framework provides a useful tool for better understanding the dynamics through which differential non-response is transformed into differential undercount for any low-response population or socially-defined group in any community where relevant empirical data can be secured.

During the non-response follow-up (NRFU) process, the Census Bureau works hard to implement a methodological strategy designed to compensate for household non-response. Nonetheless, census data quality is eroded in communities when there are high levels of non-response among some sub-populations despite the Census Bureau’s best efforts to secure complete enumeration. This is because, when confronted with high levels of household non-response, the Census Bureau is forced to rely on additional operational and statistical procedures – recourse to administrative records, proxy interviews, and, finally, imputation, to generate published tabulations of census data. Each of these efforts, while partially compensating for non-response, also introduces errors into the eventual tabulations of census data that provide the official basis for apportionment and for allocation of federal funding.

The specific ways in which the Census Bureau’s system of data collection and analysis leads to differential undercount stem in part from the characteristics of the population being enumerated but, also, from the “structural” characteristics of a geographic area – housing patterns and living arrangements, as well as local socioeconomic context [10, 11, 12]. For better or worse, the population and housing characteristics of the San Joaquin Valley make it a “natural laboratory” for exploring the extent to which adding a sensitive question such as the citizenship question to the decennial census affects census accuracy in a geographic area with concentrated immigrant settlement.

6.Cascade model coefficients derived from the SJVCRP survey

The SJVCRP survey provides data showing Hispanic immigrants’ willingness to self-respond to Census 2020 if it includes a citizenship question, willingness to respond to an enumerator visit, and willingness to respond to an enumerator’s request for a proxy interview to provide information about their neighbors [4].

Because the SJVCRP survey data include information on respondents’ legal and citizenship status (made possible by reliance on interviewers who are immigrants themselves conducting face-to-face interviews in immigrant-friendly social contexts) it was possible to determine willingness to respond among immigrants by immigration and citizenship status, as well as educational attainment, English-language ability, age, and neighborhood characteristics (represented by the Census Bureau’s LRS indicator and percent non-citizens in a tract).

7.How variations in response rate and “structural” causes of undercount will be transformed into eventual differential undercount in Census 2020

The ways in which differential non-response erodes census accuracy are dynamic and vary in relation to specific population being enumerated, survey design, and community context. But there is no doubt that when any survey – decennial census included – encounters high levels of non-response, accuracy is degraded because the process now relies less on “data” collected, actual empirical data, and more on statistical processes utilized to adjust for lack of data or inaccurate information supplied by reluctant respondents.

Higher levels of non-response in reaction to inclusion of a sensitive survey question result in incremental errors and uneven levels of non-response among different groups. This, in turn, consistently results in differential undercount of “hard to count” sub-populations. However, willingness to respond is not the only cause of undercount. Eventual differential undercount stems from the way in which census operations play out in different neighborhoods, with distinct housing patterns, ethnicities, NRFU workload and staffing (an important reason why extensive questionnaire testing in diverse sociocultural contexts and geographic areas is important).

Factors in the “cascade” of census operations the Cascade Model identifies as determinants of the eventual accuracy of the census count for a neighborhood, community, county, or state include the following:

The Cascade Model indicates that, at each stage, there is erosion in data quality despite Census Bureau efforts to secure information from respondents and from inaccuracies in the surrogate sources of information it utilizes in lieu of household census responses.

8.Structure of the Cascade Model of census undercount in the San Joaquin Valley

The model, first, gives attention to the Census Bureau’s limitations in generating a complete address list that includes low-visibility unconventional or hidden housing units (the “bad MAF” problem). It does not include an estimate of possible erroneous deletions of occupied housing units that are incorrectly believed to be vacant (due either to errors in administrative records, misreporting by proxy sources of information about housing unit vacancy, or enumerator judgement).

The model also gives special attention to the extreme reluctance observed in the San Joaquin Valley Census Research survey to proxy interviews used by the Census Bureau to secure information on non-responding households. As well as being difficult to secure, it is crucial to understand that proxy interviews are error-prone because they are, at best, estimates by neighbors – some accurate, but some inaccurate – of the number and characteristics of neighboring households. The SJVCRP survey shows that less than half believe they know enough about their neighbors to provide accurate responses.

Another important cause of undercount is due to the serious problems with “enumerations” based on securing an apparently matching administrative record that that the Census Bureau envisions using to determine the characteristics of a household that has not self-responded, and has not been contacted. The difficulty is that availability of “high-quality” administrative records is known to vary by household characteristics and consequently, inevitably contributes to differential undercount. Characterizations of household size and demographic profile based on reference to administrative records are, of course, not actually enumerations and are an element of Census 2020 operations that has not been adequately tested.

Even when an apparently matching administrative record is found for an immigrant household, it is suspect as a source of information about household size and composition because it may be out of date and is very unlikely to include peripheral household members who are not part of the primary core family living in the housing unit. Given this consideration, it is worrisome that the U.S. Census Bureau’s Operational Plan is now to seek an apparently matching administrative record even before enumerators go on to seek proxy interviews – although it is known that proxy interviews often result in unreliable data.

When a household has not self-responded and when an apparently high-quality administrative matching cannot be found, then enumerators are asked to seek a proxy interview with a nearby household to determine the size and characteristics of the non-responding household.

Finally, at the end of the cascade of Census Bureau efforts to secure information on non-responding households, there is the last-ditch attempt to use hot-deck imputation to determine the characteristics of households that have not responded, for which there is no information from a proxy interview, and where no matching administrative record can be found.

This final stage in the decennial census process is particularly problematic in the San Joaquin Valley and probably in other communities with dense concentrations of immigrants where low rates of self-response and challenges encountered in implementing NRFU leave a significant proportion of housing units to be “enumerated” via hot-deck imputation.

The ubiquitous problem in use of hot-deck imputation in the San Joaquin Valley is that the non-responding Latino households are systematically larger than those that do respond. Therefore, when a nearby responding household is considered to be the “donor” for imputing the size and characteristics of a non-responding household, each imputation will contribute to systematic differential undercount.

Given survey findings San Joaquin Valley Census Research Project, the Cascade Model generates an estimate of an 12% differential undercount of Mexican and Central American immigrants in the region. This is a conservative projection. The model does not, for example, attempt to estimate the additional negative impact that constrained access to the Internet and low digital literacy, coupled with print literacy, might have on self-response rates although there is little doubt that socioeconomic disparities pose barriers to census response of some who are motivated to participate.11

9.An in-depth look at the cascade of semi-successful efforts in the NRFU process meant to compensate for households’ failure to self-respond

The success of NRFU depends on many factors – some which can be predicted more reliably than others. Key challenges are discussed below.

10.Details on components of the Cascade Model and threats to census data quality

Below are details on key components of the Cascade Model used to estimate how non-response translates into undercount. The model predicts Census Bureau success rate at each step in the census process and estimates the percent of actual population enumerated at each stage. As the analysis moves through each stage in the process, it provides an estimate of cumulative enumeration – after self-response, after response to enumerator NRFU visits, after “enumeration” via proxy interview, and “enumeration” via reference to administrative records.

It is then assumed that the size and characteristics of the remainder of non-responding household that have not been enumerated in any of these operations will need to be imputed.

Historically, the Census Bureau has not had to rely extensively on imputation. But there are many reasons to believe the situation will be different in Census 2020, at least in communities of Hispanic and other immigrants similar to those studied by the SJVCRP due to low levels of self-response among these populations and the challenges encountered in implementing planned NRFU procedures. There are also reasons to believe that some of the standard census procedures such as reminder letters and postcards will be minimally effective in the sociopolitical context of a census that includes the citizenship question – since non-response does not result from “forgetting” but is actually deliberate.

Having determined the proportion of households in each sub-population where size and household characteristics will have to be imputed, the Cascade Model then examines undercount that stems from errors in hot deck imputations of the size and characteristics of the remaining residue of households that have not been enumerated.

11.Putting it all together: Cascade Model coefficients and components

Below, I provide details on the elements in the Cascade Model that provide the basis for the Cascade Model estimate of the specific undercount rates for the overall San Joaquin Valley immigrant population and sub-populations (defined on the basis of legal and citizenship status) within this population.

Table 1 shows coefficients used in constructing the Cascade Model-derived from SJVCRP Latino survey data, from review of the likely efficacy of imputation based on administrative records, and from under-reporting in complex households that do respond to the census.

These coefficients are then used to estimate the errors that arise at each stage in the census process and which give rise to undercount at each stage in the course of decennial census data collection and/or imputation.

Table 2 shows the structure of the model. The estimated data collection success rate is shown for each stage in census enumeration (self-response, enumerator direct interview, proxy interview) and imputation (via reliance on administrative records or hot-deck imputation). Success at each stage then determines the remainder of the universe to be enumerated using the next procedure and, eventually estimated using a fallback option. The table then shows the extent to which each component is estimated to contribute to overall undercount for the designated sub-population: undocumented and legally-resident immigrants, naturalized citizens and U.S.-born 2nd generation immigrants.

12.Practical consequences when differential non-response is transformed into differential undercount

The quest for a fair and accurate national census requires not only national estimates of differential undercount but, robust analyses of undercount in smaller geographic areas, most obviously at the level of the political jurisdictions where compromised census accuracy has the greatest impact on equity – counties, sub-state regions and municipalities where differential undercount is higher than in other areas. The Cascade Model’s projection that the level of Hispanic immigrant undercount in the San Joaquin Valley region might be around 12% if the citizenship question is added to Census 2020 is very serious.

This level of undercount would lead to omission of about 4.1% of the region’s entire population, about 190,000 Hispanic residents of the region. These omissions would, of course, also affect the overall demographic profile of the Hispanic population, skew the racial/ethnic profile of the region, result in a loss of about $280 million per year in census-driven federal funding, and compromise efforts by California’s Citizens Redistricting Commission to draw racially equitable districts.

The cascade model’s projection of the likely magnitude of differential undercount within Latino immigrant communities, even if refinements are required, underscores the need to use ethnographic research, augmented “triple enumeration” survey research similar to that used in the 1986 TARO Causes of Undercount Survey and demographic analysis (DA) in addition to dual-system estimation (DSE) to provide a reliable estimate of differential undercount if Census 2020 includes the citizenship question.

13.Conclusions

The Cascade Model presents a likely scenario regarding differential undercount of sub-populations within the overall Latino 1st and 2nd generation immigrant population of the San Joaquin Valley if Census 2020 includes a citizenship question. The model has practical utility in identifying the relative contribution to undercount associated with different aspects of census operations. As can be seen in Table 2, housing units omitted from the Master Address File, efforts to utilize administrative records for imputing the size and characteristics of non-responding households, and the need to resort to hot-deck imputation at the end of the cascade of enumeration efforts each account for a substantial portion of undercount – although the extent to which each contributes to differential undercount varies for each sub-population within the broader population of Hispanics, given the prevailing concerns about census response, as well as different living, and housing arrangements within each group.

These insights provide guidance for strategic efforts to improve census data quality. Because they take into consideration “structural” and operational causes of undercount they can be used to optimize planning – e.g. underscoring the value of Summer, 2019 in-field address canvassing (to improve the MAF), suggesting the need for expanded U/L (update-leave) operations (to improve mail delivery of census invitations, forms, and NRFU), and to enhance the potential of messaging about census data confidentiality directed specifically to landlords.

There remain uncertainties regarding eventual undercount stemming from unknown factors relating to Census Bureau implementation of census operations in the region such as adequacy of NRFU funding, ability to hire linguistically competent enumerators, density and effectiveness of questionnaire assistance centers to assist in NID processing. There are also uncertainties as to whether the CBAMS-based estimate of the relationship between expressed willingness to respond to a citizenship question and eventual response (either self-response or response to an enumerator contact) will turn out to be correct. Moreover, the SJVCRP analysis of survey respondents’ perspectives on response to a census with a citizenship question suggest that responsiveness will not be affected by Census Bureau messaging as much as by overall federal government stance toward immigrants.

Analytically, there also remain some uncertainties as to the exact proportions of 1st generation immigrants (undocumented, legal residents, and naturalized citizens) in the region – in part because these estimates are based on analyses of American Community Survey data which probably suffers from differential representation of the most socioeconomically marginal households. Other questions about how local labor market, housing arrangements related to housing affordability and local pro-census partnerships affect the Latino immigrants’ willingness to respond could also be explored using ethno-survey research methods. For example, further research would be useful to better understand how social relationships and immigration status within complex households affect partial household undercount.

The current refinement of the initial Cascade Model prototype developed in the SJVCRP also has included examination of two alternative scenarios to rapidly explore the sensitivity of the Cascade Model to key coefficients. One scenario assesses projected differential undercount assuming a level of self-response and response to enumerator interviews that is 15% higher than the SJVCRP survey findings indicate for undocumented respondents, 10% higher for legal non-citizens, unchanged for naturalized citizens (whose response rate is already relatively high) and a 15% higher response rate for the 2nd generation immigrants. The additional scenario includes the additional assumption that willingness to participate in proxy interviews is higher than the SJVCRP survey results suggest. In the first of these scenarios, the differential undercount for all Latino immigrants is 9.1%. In the second scenario, it is 8.8%. The constraints on NRFU efficacy appear to be difficult to overcome using the “modernized” approaches proposed by the U.S. Census Bureau.

It is interesting to note that the SJVCRP survey-based estimate of differential undercount of Latino immigrants in the San Joaquin Valley, as well as the two alternative scenarios, are consistent with analyses of the relationship between mail return rate and differential undercount observed in the decennial censuses conducted in 1990, 2000, and 2010. Analyses of published reports of historical mail return rates and differential undercount (as measured via dual-system estimation) in these censuses show that the ratio of mail return/differential undercount for Hispanic, Asian, and Black populations varied from one census to another but that lowered self-response was consistently translated into differential undercount.

The initial problem identified in the SJVCRP survey findings – serious and variable decrease in self-response, enumerator response, and proxy interview response among sub-populations of Latino immigrants resulting from reluctance to participate in a survey with a highly sensitive question cannot be sidestepped by recourse to procedures such as matching administrative records and hot-deck imputation in order to compensate for non-response. The Cascade Model’s articulation of the dynamics of differential undercount in an ethnically, socioeconomically, and socio-politically diverse region such as the San Joaquin Valley shows there are a number of lessons to be learned for future census planning – in the United States and in other countries. “Modernization” that fails to adequately assess the implications of sociological diversity in 21st century communities and incorporate such insights into strategies for re-engineering statistical procedures for data collection and analysis is dangerous

Furthermore, in the United States, low levels of self-response exacerbated by inclusion of a citizenship question generally viewed as part of a multi-stranded government program targeting immigrants will almost inevitably lead to differential undercount not only between racial/ethnic groups but, also, within a large racial/ethnic group – Hispanics – where responsiveness and opportunities to respond are linked to legal status as well as English-language ability, living arrangements, and neighborhood context.

The SJVCRP survey data also strongly suggest that U.S. Census Bureau reassurances making reference to difficult-to-understand data disclosure avoidance procedures and/or Title 13 protection of confidentiality will have limited effectiveness in decreasing the prevailing perception in the Latino immigrant community that census participation is risky and that responses may have negative consequences. Census questionnaire fields requiring respondents to provide household members’ names, together with information on citizenship, are seen as being an indication that supposedly confidential household information might be used by immigration authorities and a constantly-recurring theme is the idea that if the purpose of a decennial census is to count the population, why are there so many queries about “personal” information – to householders and (in proxy interviews) to their neighbors.

There are tremendous statistical risks inherent in adding a sensitive question related to the legal and citizenship status of foreign-born residents to the decennial census. This makes such a decision and the subsequent assurance that there can be adequate methodological compensation for low and uneven census participation irresponsible.

An important implication for countries throughout the world is that at least some of the challenges encountered in census-taking in culturally and socioeconomically diverse communities with concentrations of immigrant households can and should be foreseen. Migration is ubiquitous globally and, particularly in recent decades, newly-arriving immigrants in many countries lack socioeconomic and/or sociopolitical inequity. In these countries, as is shown with the data from the survey of California’s San Joaquin Valley Mexican and Central American immigrants, real and/or perceived lack of equity and consequences of providing information to the government will suppress cooperation with standard data collection procedures.