Journal:
Open Quantitative Sociology & Political Science

Authors:
Noah Carl

Title:
Net Opposition to Immigrants of Different Nationalities Correlates Strongly with Their Arrest Rates

Abstract:
Public beliefs about immigrants and immigration are widely regarded as erroneous. For example, members of the public typically overestimate the immigrant fraction of the population by ~10–15 percentage points. On the other hand, popular stereotypes about the respective characteristics of different groups (e.g., sexes, races, nationalities) are generally found to be quite accurate. The present study shows that, in the UK, net opposition to immigrants of different nationalities correlates strongly with the log of immigrant arrests rates (r = .77; p = 0.00002; 95% CI = [.52, .90]) and with the log of their arrest rates for violent crime (r = .77; p = 0.00001; 95% CI = [.52, .90]). This is particularly noteworthy given that Britons apparently think that an immigrant’s criminal history should be one of the most important characteristics when considering whether he or she should be allowed into the country. Moreover, the correlations are not accounted for by a general opposition to non-Whites, non-Westerners, or foreigners who do not speak English. While circumstantial in nature, the study’s findings suggest that public beliefs about immigrants are more accurate than is often assumed.

Key words:
Immigrants; attitudes; stereotypes; crime rates

Length:
~2,400 words; 7 pages

Files:
https://osf.io/mpq5n/

I read the paper. I have no particular objection to the methods or interpretation.

However, it's a shame that this data would only be used for this small analysis. This paper is, to my knowledge, the first to present and analyze immigrant performance data for broken down by country of origin/nationality for the UK, and so it would be very relevant to see if the usual predictors -- national IQ and Islam prevalence -- perform well in the UK sample as well. It would also be very relevant to see if there is an S factor, as there was for Denmark and Norway.

Thus, I propose that you extend this study to include more diverse outcome data and analyze whether these can be predicted by national IQ/Islam prevalence as previously reported. For this reason, I spent some time searching for data for you. I used search terms like "united kingdom", "country of birth", "country of origin", "nationality", "statistics" and then added a word for the particular type of outcome I was looking for, e.g. crime, income, employment. This produced the following results:

You have crime rates, but you can get prisoners per capita by nationality:

https://www.gov.uk/government/collections/offender-management-statistics-quarterly

These numbers cover a lot more countries. See table 1.11 (in the yearly version spreadsheet). There's prisoner rate data for about 200 countries. You just need to pair it with matching population count data for the same periods. The data covers 2002 to 2016, so one can average out a lot of the randomness for the smaller groups.

This source has welfare use by country of origin for a small number of countries.

https://www.migrationwatchuk.org/briefing-paper/367

It's probably possible to find some more data. I particularly would like to see:
- (un)employment rate
- average income
- educational attainment -- most relevant when they attended school in the UK
- cognitive/scholastic scores, GCSE for instance
- health indicators, such as smoking

It is possible to impute data for groups that have e.g. 2/3 datapoints.

Let me know what you think. I realize this would take a lot more work, but I think it is worth it. Recall how much your previous terrorism paper improved by using multiple data sources.

However, it's a shame that this data would only be used for this small analysis. This paper is, to my knowledge, the first to present and analyze immigrant performance data for broken down by country of origin/nationality for the UK, and so it would be very relevant to see if the usual predictors -- national IQ and Islam prevalence -- perform well in the UK sample as well. It would also be very relevant to see if there is an S factor, as there was for Denmark and Norway.

Thus, I propose that you extend this study to include more diverse outcome data and analyze whether these can be predicted by national IQ/Islam prevalence as previously reported.

While I agree an analysis of how well country-of-origin variables predict immigrant performance in the UK would be very interesting, the purpose of my paper was somewhat different, namely to explore whether Britons' immigration attitudes are based on well-informed beliefs about the respective characteristics of different immigrant groups––which I believe is also an interesting question. I would therefore prefer to keep the focus of the analysis roughly as it is now.

For this reason, I spent some time searching for data for you. I used search terms like "united kingdom", "country of birth", "country of origin", "nationality", "statistics" and then added a word for the particular type of outcome I was looking for, e.g. crime, income, employment. This produced the following results:

You have crime rates, but you can get prisoners per capita by nationality:

https://www.gov.uk/government/collections/offender-management-statistics-quarterly

These numbers cover a lot more countries. See table 1.11 (in the yearly version spreadsheet). There's prisoner rate data for about 200 countries. You just need to pair it with matching population count data for the same periods. The data covers 2002 to 2016, so one can average out a lot of the randomness for the smaller groups.

I considered using prison population data in my analysis, but the sample sizes were indeed very small for some countries. For example, there are only 3 Argentinian prisoners and 4 Japanese prisoners in 2012. Because prison population is a stock measure, averaging over multiple years doesn't make that much difference––many of the prisoners from a given country who are incarcerated in one year will also be incarcerated in the next year. And obviously, with n = only 23 cases in the analysis, measurement error on just a few cases can make a lot of difference to the estimated correlations and regression coefficients. The arrest data, which I chose to utilise instead, have the virtue of much larger sample sizes: no country in the sample had fewer than 100 total arrests over the time period in question.

This source has welfare use by country of origin for a small number of countries.

https://www.migrationwatchuk.org/briefing-paper/367

It's probably possible to find some more data.

These data are probably of reasonably high quality. However, they are only available for about half of the 23 countries in my analysis. And, obviously, n = 23 already represents a small sample size.

I particularly would like to see:
- (un)employment rate
- average income
- educational attainment -- most relevant when they attended school in the UK
- cognitive/scholastic scores, GCSE for instance
- health indicators, such as smoking

It is possible to impute data for groups that have e.g. 2/3 datapoints.

Let me know what you think. I realize this would take a lot more work, but I think it is worth it. Recall how much your previous terrorism paper improved by using multiple data sources.

I agree an S-factor analysis of UK data would be highly interesting and informative. But I would claim that investigating how opposition to immigrants from different nationalities relates to their involvement in crime––the purpose of my study––is also interesting and informative. My hunch is that many social scientists, especially those ignorant of the stereotype accuracy literature, would be quite surprised to learn that net opposition to different immigrant groups is reasonably well-calibrated to their rates of criminality.

Obtaining reliable data at the level of the immigrant nationality appears to be quite difficult. The UK census only reports figures at the level of the broad region (e.g., South Asia), which isn't particularly helpful. And most social surveys don't have enough cases to get a reliable average for some of the smaller immigrant groups.

I would reiterate that, if reliable data on a large number of immigrant groups within the UK do become available, it would certainly be worth conducting an S-factor analysis. But in the meantime, I believe my analysis of public attitudes merits publication in its own right.

Fair enough. Let me replicate your analysis in R and then you have my approval.

Some comments.

Raw support data
Raw data for net support/oppose are not given. Please give the raw support numbers, so that others may calculate other metrics for net support. Your operationalization, while reasonable, is not the only reasonable one.

More exact sources
The spreadsheet gives sources in Sheet 2, but there are no links. There should be. If the source files are xlsx files or similar (and not e.g. a web browser interface), please include them as well.

White
The White majority variable is not explained in the paper. Was this filled out by the author's judgment or how was it done? Note that "White" is a kind of sociological category, not necessarily a homogeneous genetic category. For instance, persons who self-identify as non-Hispanic White in the United States are about 99% European genetically. However, persons who self-identify as Blanca (White) in Latin America are often only about 80% European genetically. You could use the World Migration Matrix to fill in the data. It's the Putterman and Weil dataset.

http://www.brown.edu/Departments/Economics/Faculty/Louis_Putterman/world%20migration%20matrix.htm

English language
No source is given for the categorical English language variable. This variable is also not explained in the paper. Since it's coded as categorical, presumably it reflects official language status. Official language status, however, is a poor correlate of actual speakers. Consider using proportion of English speakers by country, such as:

https://en.wikipedia.org/wiki/List_of_countries_by_English-speaking_population

It's doubtful that most people know that English is an official language in e.g. Pakistan, as you have coded it.

https://en.wikipedia.org/wiki/List_of_territorial_entities_where_English_is_an_official_language

Reversely, English is not an official language in any Nordic country (+Netherlands), but these countries have the best English speakers of non-native countries. Of course, English ability is to some degree confounded with cognitive ability.

https://en.wikipedia.org/wiki/EF_English_Proficiency_Index

Analytic replication
No inconsistencies.

Raw data for net support/oppose are not given. Please give the raw support numbers, so that others may calculate other metrics for net support. Your operationalization, while reasonable, is not the only reasonable one.

These data have now been included in the supplementary data file.

The spreadsheet gives sources in Sheet 2, but there are no links. There should be. If the source files are xlsx files or similar (and not e.g. a web browser interface), please include them as well.

Links to data sources have now been included in the supplementary data file.

The White majority variable is not explained in the paper. Was this filled out by the author's judgment or how was it done? Note that "White" is a kind of sociological category, not necessarily a homogeneous genetic category. For instance, persons who self-identify as non-Hispanic White in the United States are about 99% European genetically. However, persons who self-identify as Blanca (White) in Latin America are often only about 80% European genetically. You could use the World Migration Matrix to fill in the data. It's the Putterman and Weil dataset.

Percentage of the population that is white has now been used in place of white majority. These data were taken from Wikipedia and the CIA World Factbook.

No source is given for the categorical English language variable. This variable is also not explained in the paper. Since it's coded as categorical, presumably it reflects official language status. Official language status, however, is a poor correlate of actual speakers. Consider using proportion of English speakers by country

Percentage of the population that speaks English has now been used in place of English as an official language. These data were taken from the Wikipedia page recommended by Emil. Note that the coding for Western country has now been altered slightly (Israel = non-Western), in accordance with Huntingdon's (1996) classification.

All changes seem good to me.

I had a look a the datafile too. Seems to be in order.

I saw that the numbers were updated in the paper for the regression models. I re-run my code and my numbers fit again. The only discrepancy was that you changed the variable name from "Western" to "West", which broke my code.

I approval of publication.

One final suggestion. Add some identifier for the UK in the title. I imagine that I will run a replication study about this for Denmark some time next year, so it would be nice if they could be easily grouped with similar, but distinctive titles.

One final suggestion. Add some identifier for the UK in the title. I imagine that I will run a replication study about this for Denmark some time next year, so it would be nice if they could be easily grouped with similar, but distinctive titles.

I have changed the title to: 'British Opposition to Immigrants of Different Nationalities Correlates Strongly with Their Arrest Rates'.

Sounds good.

Many thanks for the review.

1. I think that the manuscript should disclose that, for each of the 23 dataset countries, the YouGov immigration item had "Don't Know" responses for 22% to 29% of observations.

This has been noted in Section 2.

It might be beneficial to the reader to note that the manuscript's main inferences do not change with particular alternate specifications, whether the aforementioned specifications or other specifications.

This has been noted in Section 2.

3. It might be valuable to note in the manuscript the countries coded as Western. The countries furthest below the regression line in Figure 1 are those most plausibly considered Western, but other plausibly Western countries are near or above the line, so it might be valuable to know which countries are coded Western.

This information has been included in a footnote in Section 2.

4. Ideally, in Figure 1, the country labels would not be on top of one another, and the figures would include the correlation coefficients.

The correlation coefficients have been included in the text, as well as in the Abstract, so I do not believe there is a strong rationale for also including them in the Figures. I would also claim that the country labels on the graphs are reasonably legible. I would therefore prefer not to change them.

5. The abstract and conclusion provide two implications of the manuscript: accurate stereotypes, and opposition to immigration from certain groups being informed by rational beliefs about that group. Presumably, according to some perceptions, irrational reasons for opposition to immigration from a particular country are the race or ethnicity of the immigrants, so there might value in reporting standardized effects for the Western country variable, to see the extent to which immigrants from Western countries are favored, or, for that matter, to include regional controls to assess the extent to which particular regions are favored, although the small sample size and lack of relevant controls might not be conducive enough to this sort of analysis.

Standardised coefficients for control variables are now displayed in the tables. I agree with the reviewer's intuition that, given the small sample size, adding regional controls would offer very little conceptual purchase.

6. The final sentence of the manuscript notes the lack of data for characteristics of immigrant groups in the UK, but the percentage white and percentage English speakers controls used in Table 1 were based on the home countries and not on immigrants in the UK, so it's not clear why controls could not be included with data on characteristics from the home countries for, say, mean country IQ or international test scores.

The regression models now also control for national IQ.

Many thanks for the review.

There are no page numbers. Add page numbers.

Page numbers have been added.
 

I do not understand this sentence: ”Moreover, the correlations are not accounted for by a”
You mean are not modified by considering/controlling for? Are stable?

Yes, that is what I mean. I would prefer not to change the phrasing here, as I believe it conveys the results of the regression analyses in a way that is easier for the non-specialist to understand. In my opinion, relatively straightforward language is desirable in the Abstract.

Introduction: ”Indeed” should be replaced by something like ”This is backed by”.

Again, I would prefer not to change the phasing here, as I believe "Indeed" is more appropriate than "This is backed by". The reason being that the Sides and Citrin's (2008) result is not technically an example of stereotype accuracy.
 

In Figure 1 add for both figures in the heading the sources of data and the correlations, e.g.:
Scatterplots of the relationship of net opposition (YouGov poll) with log of immigrant arrest rates (r=.77) and log of immigrant arrest rates for violent crime (r=.77) (both Metropolitan Police).

The title for Figure 1 has been changed accordingly.

Both r=.77?

Yes, it appear so. I guess this isn't that surprising given that the correlation between the two measures of arrest rates is r = .95.

Add a correlation table for all used variables.

This has been added in Appendix A.

Maybe a further good control: percentage of Muslims in each country of origin.

Percentage Muslim has now been controlled for in the regression models.

Published.

https://openpsych.net/paper/48

Following correspondence with Richard Lynn, I have added a post-publication supplement to the paper's OSF page.

Please find a point-by-point rebuttal below.

1. Ecological study design problem

The research design is ecological.  This means that the data are not individuals, but aggregates (groups) of individuals.  This is probably the weakest study design in the social sciences; it is not only observational, but does not actually measure anything about people, but rather, just aggregations of people.  One consequence of this is that these research designs tend to over-estimate model fit.  That is, any effects estimated tend to fit more poorly in the real world than they do in the model.  This is because these study designs under-estimate variability.  In this example, had the author used individual data on the perception of immigrants rather than averages to fit his models, he probably would have seen a weaker relationship than he observed.

Researchers interested in stereotype accuracy generally distinguish between consensual stereotypes on the one hand and personal stereotypes on the other (Jussim et al., 2015). Consensual stereotypes correspond to the average beliefs about some group of people, whereas personal stereotypes correspond to a particular individual’s beliefs about some group of people. (Note that this distinction was already mentioned in footnote 1 of my paper.) As Jussim et al. (2015) note, consensual steretypes are “usually assessed by sample means” [emphasis added]. 

Although my study was not concerned with the accuracy of consensual stereotypes per se, it was concerned with a related concept, namely average preferences for or against certain immigrant groups. And just as it is interesting to examine average beliefs about certain groups, it is also interesting to examine average preferences for or against those groups. Note that Ford (2011; published in a ‘mainstream’ journal) also examined average preferences for or against certain immigrant groups. 

2. Small sample problem

In addition to being a weak study design, the author relies on 23 observations to draw his conclusions.  Statistics can make up for small samples when study designs are strong and variables are measured without systematic error, but the small sample size used in this study is particularly troubling when combined with all the other problems with the study.  Small samples are a multiplier of all other problems.

Of course, it is always better to have more data, but my analysis was limited by what was available at the time. In particular, the 2016 YouGov poll that my study drew upon only asked bout 23 different immigrant nationalities. Hence, n = 23. 

However, note that the raw correlations were significant at the 0.1% level, and many of the associations in the multivariate models were significant at the 5% level or lower. Moreover, most of these associations were in the ‘expected’ direction (e.g., positive effects of percentage white, Western country and percentage English speakers; negative effect of percentage Muslim). These observations indicate that, as a matter of fact, there was no “small sample problem”. 

If there had been a “small sample problem”, the associations would have been only borderline significant, and the ones in the multivariate models would have been as likely to go in the ‘expected’ direction as in the opposite direction.

3. Bad sample

The researcher did not look at all immigrant data in the UK, but a small non-random sample of 23 countries.  There are a large number of Italian and Portuguese immigrants to the UK, but these data are not included in the study.  If they were included, the results may have looked different.  When the data we use are not exhaustive (complete) and not selected randomly, there is always the possibility that the selection of data used will affect out findings in a systematic way.  This is particularly problematic when the sample is small; a small non-random sample is the holy grail of statistical badness.

Once again, I looked at all data that were available at the time (n = 23 immigrant groups). It is true that the results might have been different if the number of immigrant groups had been larger. But note that in our unpublished study on immigration policy preferences in Denmark (which had a larger n = 32), we again observed a very strong association between level of opposition to immigrant groups from different origin countries and those groups net fiscal contributions in Denmark. 

Furthermore, although the aforementioned 2016 YouGov poll only asked about 23 different immigrant nationalities, it did encompass many of the nationalities with large immigrant populations living in the UK (e.g., Poland, Nigeria, France, Romania, Pakistan; see ONS, 2012). In addition, it included nationalities from nearly all the major world regions: Europe, the Middle East, South Asia, East Asia, Africa, North America, South America, Oceania.  

4. Missing variable problem

The author uses multiple regression model to control for the ‘confounding’ effect of things like whiteness, English speaking, being from a Western country, and religion on his observation that crime rates influence perception of immigrants.  He did not control for other potential confounding effects, however–like the economic wealth / productivity of the country of origin, historical tensions or media portrayals.  I added per capital GDP to his data set an observed that the log of per capita GDP correlates more strongly with perception of immigrants than the the log of crime rate.  It’s hard to know what variables to include in an analysis like this, but it matters, as the inclusion and exclusion of variables can change how data are interpreted.

It is not possible to anticipate every alternative model specifications that someone may prefer. The data were published online precisely in order to allow other researchers to run their own analyses. Moreover, there is disagreement about whether to control for a variable like GDP per capita because it is arguably endogenous (i.e., more of an ‘outcome’ measure than an ‘input’ measure). 

In addition, the multivariate models are arguably less interesting than the raw associations between crime rates and net opposition. As noted in the paper, YouGov asked the British public to say how important each of 14 characteristics should be when considering whether or not an economic migrant should be allowed into the UK. The two most important were ‘criminal record (major/violent)’ and ‘criminal record (minor/non-violent)’. Thus, even if respondents were using GDP per capita as a proxy, the stereotypes underling their immigration policy preferences can still be seen as ‘rational’. 

5. Non sequitur

Carl draws the conclusion that ‘public beliefs about immigrants are more accurate than often assumed’, but the bulk of his analysis does not meaningfully address this claim.  Carl has not defined what ‘accurate’ is, but no reasonable definition can be boiled down only to crime rate–that is, the negative contributions of immigrants.  If public opinions were really ‘accurate’, their perceptions would also correlate with the positive contributions of immigrants, and in fact there would be strong correlation between net utility of immigrants and the perception of immigrants.  But Carl focuses only on one possibly useful measure, and ignores the rest.  As such, even if the technical difficulties above were overlooked, his conclusion is a misdirection since he’s not really measuring accuracy of public opinion.

As mentioned above, accuracy was defined in exactly the same way as it is defined in the literature on stereotype accuracy, namely in terms of the correlation (or ‘correspondence’) between average beliefs/preferences and average criterion values (see Jussim et al., 2015). 

Given that Britons say an immigrant’s criminal history should be one of the most important characteristics when decided whether he should be admitted to the country, and there is a strong correlation between crime rates and net opposition, it seems reasonable to claim that their immigration policy preferences are informed––at least to some extent––by rational beliefs. 

Note that the paper acknowledged (in the Abstract, Introduction and Conclusion) that the British public systematically overestimates the percentage of immigrants in the population. It simply concluded that “public beliefs about immigrants are more accurate than is often assumed” [emphasis added]. Nowhere did it state that the public are entirely accurate about all aspects of immigration. 

Update: the statement above was later modified to "public beliefs about the relative positions of different immigrant groups may be reasonably accurate" 

References

Ford, R. (2011). Acceptable and unacceptable immigrants: How opposition to immigration in Britain is affected by migrants' region of origin. Journal of Ethnic and Migration Studies, 37, 1017–1037.

Jussim, L., Crawford, J.T., Rubinstein, R.S. (2015). Stereotype (in)accuracy in perceptions of groups and individuals. Current Directions in Psychological Science, 24, 490–497.

ONS. (2012). Population by country of birth and nationality. Office for National Statistics, published online.