Explaining immigration preferences: Disentangling skill and prevalence

Abstract

One of the most important and consistent findings to emerge from the study of immigration politics over the past decade is the seemingly uniform preference among mass citizenries for high-skilled immigrants. One potential conceptual flaw in this mounting body of literature is that skill is confounded with prevalence: people may prefer high-skilled immigrants not because they are skilled but because there are not very many of them. To address this possibility, we conducted an original experiment within a nationally representative survey of over 12,000 respondents. We conducted three main empirical tests and found that the skill premium is not confounded by prevalence. However, low-skilled Mexican immigrants specifically are disadvantaged when people are told that they are prevalent, a finding that comports with extant research on the construction of Latino immigration as a unique threat to American society.

Over the past decade, scholarship on immigration politics has rendered an important and highly consistent finding: the seemingly uniform preference for high-skilled immigrants among the mass public in the United States (US) (e.g. Hainmueller and Hopkins, 2015; Wright et al., 2016) and many other industrialized nations (e.g. Hainmueller and Hangartner, 2013; Iyengar et al., 2013). Utilizing sophisticated experimental designs, these works demonstrate that, across a host of immigrant-receiving nations, citizens’ judgments about whether or not an immigrant should be admitted are significantly impacted by an immigrant’s skill level independent of their nationality, skin tone, or religious practices—and irrespective of citizens’ own economic positions, ethnic prejudice, or partisanship.

The purpose of this article is to address a potential conceptual flaw in this mounting body of literature, and to assess whether this flaw is empirically consequential. The “skill premium” repeatedly observed in empirical studies could mean that people prefer high-skilled immigrants to low-skilled immigrants because of skill per se. However, a potential problem with this theoretical interpretation is that skill is confounded with prevalence: people may prefer high-skilled immigrants because there are not very many of them.

This possibility is plausible given that in the US, as well as in many other immigrant-receiving nations, low-skilled immigrants comprise the majority of foreign-born persons. Data from the Current Population Survey indicate that roughly 28% of immigrants, compared to only 7% of native-born citizens, lack a high school education (Camarota, 2012). Additionally, immigrants are more than twice as likely to work in low-skill occupations such as construction and extraction than native-born citizens (11.4% vs. 4.1%) and half as likely to work in high-skilled occupations such as management (8.1% vs. 16.7%). ¹ Individuals from Mexico are by far the most-prevalent group, comprising roughly 28% of the foreign-born population, with the second and third most-prevalent groups being individuals from China (5.9%) and India (5.1%). ² Focusing on Mexican immigration, the potential confound of skill with prevalence is clear: Mexican immigrants are disproportionately represented in low-skill occupations and roughly 60% lack a high school education (Passel et al., 2012). Immigrants from China and India combined (~4.5 m) constitute less than half the number from Mexico (~11.6 m). Further, these immigrants are much more likely to be highly skilled: roughly 47% of Chinese and 76% of Indian immigrants hold a bachelor’s degree or higher. ³

It is plausible that, when considering the admission of immigrants in recent experimental studies, American citizens compare presented immigrant profiles to the stored mental image of the modal immigrant residing in the US—a low-skilled Latino. As a result, an observational equivalence problem may exist, where we are unable to disentangle (a) preference for skill from (b) preference for immigrants unlike the most-prevalent immigrant group, as both lead to the same observed outcome—the selection of high-skilled applicants. If this potential alternative was supported by empirical evidence, then the skill premium, rather than indicating an impartial preference for skill, could be interpreted as a convoluted manifestation of citizens’ aversion toward the immigrants most prevalent in their nation. This would represent yet another piece of evidence demonstrating that prejudice—in one form or another—largely underlies public opinion on immigration (Hainmueller and Hopkins, 2014). In this article, we conduct a series of empirical tests to disentangle these competing explanations.

Confounding is a common problem in survey experiments (Dafoe et al., 2015). Even though experimental treatments are randomly assigned, it is often difficult to identify the underlying theoretical mechanism because a treatment could bring to mind correlated factors. When respondents read information about a high-skilled immigrant—and no information about prevalence—they may think that there are few of these immigrants (especially relative to low-skilled Latinos) and therefore be reacting to the low prevalence and not the high-skill level. Although the difference between the treatment and control groups provides an unbiased estimate of the word(s) representing the concept of “high skill” (versus “low skill”), it does not provide the effect of skill per se. This issue is similar to the problem of omitted-variable bias in observational studies. Scholars have attempted to address this issue by providing additional information in the form of controlled comparisons (Dafoe et al., 2015) or conjoint designs (Hainmueller et al. 2014). For example, previous studies have included controls for characteristics related to skill such as potential dependence on the welfare state, cultural similarity, and assimilation potential. However, none of the studies on skill preference include information on prevalence. This study assesses whether accounting for information about prevalence—a potential omitted-variable—affects inference.

Existing research documents a link between the prevalence of an immigrant group and public antipathy toward the group, which establishes a basis for the expectation that people—particularly American citizens—may be motivated to embrace the admission of non-prevalent immigrants. As a general point, scholarship finds that the perceived amount of immigration is positively associated with holding negative views toward immigrants among citizens in Europe and the US (Herda, 2010; Sides and Citrin, 2007). Focusing specifically on the US and Latino immigration, there is evidence that prevalence is associated with stigmatization and public antipathy. Latinos are mentioned more than any other group in media coverage of immigration (Valentino et al., 2013) and are constructed by academic, media, and political elites as a threat to the nation (Chavez, 2013; Haynes et al., 2016; Huntington, 2004). Other prominent though less-prevalent immigrant groups, such as Asians, are constructed as “model minorities” (Lee, 2015; Taylor and Stern, 1997). Perhaps as a result of such media and elite discourse, there is mounting evidence of the uniquely threatened response of American citizens to Latino immigrants. Scholarship finds that residing near large Latino populations is associated with greater prejudice toward immigrants (Ha, 2010) and that negative affect toward Latinos is a prepotent predictor of opposition to immigration (Valentino et al., 2013). Complementing this, experimental research finds that citizens are more incensed by information about the costs of immigration when the group in question is Latino versus white (Brader et al., 2008), and express greater disapproval of transgressive behaviors when engaged in by a Latino versus white immigrant (Hartman et al., 2014). Finally, research demonstrates that Americans hold negative implicit (i.e. unconscious) attitudes toward Latino immigrants, and that judgments about immigration policy are tightly linked to implicit anti-Latino bias (Pérez, 2010). ⁴

An additional limitation of some previous studies is that the outcome variable is whether a particular immigrant should be admitted (e.g. Hainmueller and Hopkins, 2015; Iyengar et al., 2013; Wright et al., 2016). Subjects are presented with a decision task that is susceptible to “person positivity bias” (Iyengar et al., 2013) and that fundamentally differs from what typically underlies common immigration policies, which is the question of the admission of a large number of immigrants rather than a single immigrant. For example, in the controversy over the Obama Administration’s request to state governments to shelter unaccompanied minors crossing the border to flee violence in Central America, local lawmakers and citizens were confronted with the choice of accepting thousands of persons (Etter and Oldham, 2016). The same is true when setting per-country admission ceilings, as well as refugee agreements and asylum policies, as each concerns the entry of large numbers of persons (American Immigration Council, 2016). People may be willing to admit a particular person, but may not want to apply those criteria to a larger number of similar immigrants. We address this concern by including additional outcome variables measuring support for the admission of numerous applicants of a particular type.

Data and study design

We conducted an original survey experiment including 12,052 respondents, administered by Morning Consult, a major survey research firm and polling organization. Details about the survey methodology, sampling procedures, and questionnaire can be found in Online Appendix 1. Descriptive statistics of the sample are in Online Appendix 2. The sample size was purposely large so that we would have adequate statistical power to detect subtle differences in effect sizes.

The design of our study employed Dafoe et al.’s (2015) controlled comparisons approach, where we fixed a set of attributes of the immigrant profile and varied the ones that were relevant for our analysis. ⁵ The experimental design is summarized in Online Appendix 3. We employed a 3 x 3 x 2 factorial design in which we manipulated three factors. First, the country of origin of the immigrant: Canada, a predominantly white nation that is culturally similar to the US and that mainly produces skilled immigrants; India, a predominantly non-white nation that mainly produces skilled immigrants; and Mexico, a predominantly non-white nation that mainly produces unskilled immigrants. Our assignment of skill/prevalence to countries of origin is based on data on the skill composition of the immigrant groups. ⁶ Second, we manipulated two levels of skill: high and low. The high-skilled immigrant presented to respondents had completed a graduate degree and was a doctor. The low-skilled immigrant had no formal education and was a farm laborer. ⁷ We therefore operationalize skill as education and occupation jointly. As shown Online Appendix 3, we fixed the other features of the immigrant. ⁸

The third manipulated factor was whether the respondent received information about the prevalence of the immigrant described in the profile. Respondents in the treatment group saw the following text: “Of the 42.3 million foreign-born individuals currently residing in the US there are [very many/only a few] [low-skill/high-skill] persons from [COUNTRY].” For India and Canada, respondents were told that high-skilled immigrants were prevalent when presented with a high-skilled profile, and that low-skilled immigrants were not prevalent when presented with a low-skilled profile. For Mexico, respondents were told that low-skilled immigrants were prevalent and high-skilled immigrants were not prevalent. In other words, there was not a full manipulation of prevalence information. Respondents were provided accurate information with no deception, ⁹ and those in the control group did not see any information about prevalence. There were about 1000 respondents per cell (cell counts are presented in Online Appendix 4). There was balance between conditions across demographic variables (see Online Appendix 5).

We measured four main outcome variables. In addition to the binary admission question and ordinal admission scale used in previous work (e.g. Hainmueller and Hopkins, 2015), we measured two further outcome variables: (1) how many more immigrants like the one described in the profile should be admitted; and (2) support for a policy to admit 10,000 additional, similar immigrants. Full question wordings are presented in Online Appendix 1. The latter two dependent variables represent the contribution of this paper—do respondents go beyond accepting a given immigrant if they believe that the type of immigrant presented is going to arrive in large numbers? We also created an additive scale averaging these four variables. All variables were recoded to lie between 0 and 1.

Test #1: Do people only prefer skilled immigrants because they are not prevalent?

We first examined the control group that was provided with no prevalence information to assess whether non-prevalent immigrants are preferred to prevalent immigrants regardless of skill level. As shown in the top half of Table 1, pooling across ethnic groups, respondents exhibit a strong preference toward accepting high-skilled immigrants over low-skilled immigrants (11.1%, p < .001). ¹⁰ If prevalence were confounding skill, we should expect to see this skill effect concentrated among Mexicans and less pronounced for Indians and Canadians. This is because high-skilled Mexicans are less prevalent and high-skilled Indians and Canadians are more prevalent. However, as shown in the top half of the table, the skill premiums are substantively similar across ethnic groups. Further, there are no statistically significant differences in the skill premiums between ethnic groups (see Online Appendix 7b). These findings not only replicate those from previous work, but also comport with the prevailing interpretation of the skill premium as an impartial preference for skill regardless of nationality. ¹¹

Table 1.

Immigrant admission rates by experimental condition.

	Full Sample	Mexico	India	Canada
No Information
High Skill	83.1%	83.2%	80.4%	85.9%
Low Skill	72.0	70.5	69.8	75.6
Difference (Skill Premium)	11.1	12.7	10.6	10.3
	p < .01	p < .01	p < .01	p < .01
Prevalence Information
High Skill	81.5%	81.1%	78.6%	85.2%
Low Skill	68.4	62.7	66.8	75.6
Difference (Skill Premium)	13.2	18.4	11.8	9.5
	p < .01	p < .01	p < .01	p < .01
Difference-in-Difference	2.0%	5.7%	1.3%	−0.7%
	p = .18	p = .04	p = .65	p = .77

Note: N = 12,052. Two-tailed p-values of differences based on regression results reported in Online Appendix 7.

Test #2: Do people prefer admitting skilled immigrants as a matter of general policy?

As mentioned above, in prior research on the skill premium respondents are asked about whether a particular immigrant should be admitted, not about whether many immigrants like the one presented to the respondent should be admitted. It is possible that people want to admit a particular skilled immigrant but would not support admitting this group as a matter of policy if this group were to be made prevalent in the country. Consequently, in addition to asking about admission about the immigrant presented in the profile, we also asked whether more immigrants like the one should be admitted and whether respondents supported a policy admitting 10,000 such immigrants. As shown in Online Appendix 7a, the results are consistent across all outcome variables. ¹² Respondents exhibit a strong skill-premium (ranging from 7–11% across all outcome variables), and there is no substantively or statistically significant heterogeneity across ethnic groups (see Online Appendix 7b). In fact, for each of the 12 tests we ran (three ethnicities x four outcome variables), the skill premium was highly significant and never smaller than 6%. Thus, results from previous studies are not an artifact of the particular outcome variables employed.

Test #3: Does providing people with prevalence information affect the skill premium?

In our final test, we assessed whether providing information about the prevalence of the immigrant in question affects people’s willingness to admit him/her as well as views toward broader immigration policies. As shown in the first column of Table 1, pooling all ethnicities together, the skill premium increases by 2.0% (p = .18). However, this is not the correct estimand because the prevalence information indicates that high-skilled Mexicans are rare and high-skilled Indians/Canadians are common. Consequently, if prevalence were confounding the skill premium, then we would expect the difference-in-difference to be positive for Mexicans and negative for Indians/Canadians. Accordingly, we estimated a model where we recoded the prevalence treatment to take on the value of -1 for respondents in the treatment group who were presented an Indian/Canadian profile, +1 for respondents in the treatment group who were presented a Mexican profile, and 0 for respondents in the control group who received no prevalence information. This recoding effectively signs the differences-in-differences of the three individual ethnicities in a consistent manner. This recoding yielded a significant pooled difference-in-difference of 3.4% (p = .002). As shown in Online Appendix 7c, results are similar for the other outcome variables except for one.

However, as shown in Table 1, this effect is completely driven by the Mexican profiles (see Online Appendices 7d and 7e). The difference-in-differences are statistically insignificant and close to zero for Indian and Canadian profiles, which suggests that for these predominantly high-skilled groups, prevalence is not confounding the skill premium. In other words, admission support for high-skilled Indian and Canadian applicants does not significantly decrease in response to being reminded that high-skilled immigrants from these nations are more common than low-skilled immigrants. However, Table 1 makes clear that providing information about the prevalence of different skill categories increases the skill premium for Mexicans by a substantively large 5.7% (p = .036). As shown in the table, this is mainly due to the prevalence information decreasing support for low-skilled Mexican immigrants from 70.5% to 62.7% (p < .01); support for high-skilled immigrants is basically unaffected by the prevalence information (p = .20). This is consistent across all the outcome variables (see Online Appendix 7d). Further, the difference-in-difference is statistically significantly greater for Mexicans than Indians/Canadians (p = .09). ¹³ We now turn to the implications of this finding.

Conclusion

Overall, our study suggests that the skill premium is not confounded by prevalence. People prefer high-skilled immigrants not only for origin countries that mainly produce low-skilled immigrants (Mexico) but also white (Canada) and non-white (India) countries for which high-skilled immigrants are prevalent. Further, people support the admission of not only particular high-skilled immigrants, but also their admission as a general and broad policy.

What should we make of the fact that low-skilled Mexican immigrants specifically are disadvantaged when people are told that they are prevalent? One possibility is that sociotropic evaluations of immigration are tinged with prejudice. Adding to this, it may be possible that the effects of prevalence are asymmetrical as a function of skill level, and thus only influence public opinion when relating to low-skilled workers. Future research is therefore warranted on the skill premium. Although it is unlikely that existing results in the literature are due to a preference for prevalence (as opposed to high-skilled) immigrants, there does seem to be some differences in how people respond to immigrants from more stigmatized groups (Mexicans) than to ones representing either culturally similar whites (Canadians) or predominantly high-skilled model minorities (Indians).