Introduction

Demographic forecasts suggest an increasing need to better understand ethnically diverse societies. The globalized nature of today’s economic value chains, travel and more unfortunate drivers of human mobility send international migration to record levels each year. As a result, European societies are now more ethnically diverse than ever, and by 2044, for example, non-white residents are set to constitute the majority in the United States (Fibbi, Midtbøen, and Simon 2021; Jardina 2019). It remains unclear though what such trends imply for social and political dynamics in receiving societies. The literature offers two contending hypotheses, the Contact Hypothesis and the Group Threat Hypothesis, which will be further explored in this statistical study of immigration attitudes across the United States (US). The subsequent analysis will thus be guided by the following research question: What factors influence attitudes on the effects of immigration? Specifically, I will investigate the role of the percent change in the immigrant population of each US State and individual perceptions about the national economy’s performance.

Current scholarship on the effects of immigration within host societies has turned to indicators of integration as a way of speaking to the drivers and inhibitors of social cohesion. However, much of the extant literature is focused on outcomes such as language acquisition, naturalization, intermarriage and employment to assess integration, but this overlooks the role of the local population in defining and upholding the social barriers to assimilation. I propose that broadening the scope of analysis to include this majority population allows for greater theorization about intergroup interactions that ultimately translate to the political pressures policymakers must contend with. In order to do so, we must turn to insights from social psychology that pre-date this more recent attention to immigrant integration. This is where the Contact and Group Threat hypotheses offer potential utility.

First, the Contact Hypothesis developed in Gordon Allport’s seminal work on prejudice argues that intergroup exposure (in this case locals and immigrants) reduces prejudice, especially when in pursuit of a common goal (Allport 1979; Fibbi, Midtbøen, and Simon 2021; Gabrielli, Gsir, and Zapata-Barrero 2017; Graefe et al. 2008). His theory predicts that tensions ease through repeated interaction with members of the opposite group, as the degree of uncertainty diminishes. This suggests that increased immigration should translate to more intergroup interaction and therefore decrease prejudice overall. In contrast, increasing diversity may actually trigger the opposite reaction, per the Group Threat Hypothesis, whereby the majority population perceives possible threats to material security, group position or cultural erosion. Such a reaction serves to bolster ingroup ties and adherence to dividing lines (Blalock 1967; Blumer 1958; Schlueter, Meuleman, and Davidov 2013; kesller and bloemraad).

Specifically, I will test the following hypotheses to understand which best explains the US context:

H1: The larger the percent change of the immigrant population in each state, the more attitudes on the effects of immigration will be negative

H2: the more negative views of the economy’s performance, the more individuals will hold negative views about the effects of immigration

The second hypothesis above stems from the functional aspects of identity formation and stereotyping underpinning these ingroup and outgroup perceptions. As alluded to in Allport’s Contact Hypothesis and subsequently validated in a number of studies, group boundaries and their salience wax and wane depending on how situational characteristics rely on or ignite their activation (Hale 2004; Rabbie and Horwitz 1969). If two groups benefit from working together towards common goals, Allport argues that prejudice will be reduced. Conversely, when group interests are in conflict, competition and negative attitudes towards outgroups arise from a perceived sense of threat. Material stakes that hinge upon group categorization also tend to increase the importance individuals attach to these divides (Fibbi, Midtbøen, and Simon 2021). Therefore, by accounting for one’s view of the health of the economy, we gain a sense of the extent to which material security is perceived to be at risk.

The findings of this study reveal that percent change in the immigrant population has a much more limited effect on attitudes than views on the health of the economy. However, the relationship between attitudes on the economy and immigration’s effects cannot simply be reduced to feelings of material insecurity. I conclude that it is much more fruitful to think about the notion of security generally in relation the majority group’s status when considering how a the receiving population feels affected by growing ethnic diversity.

Data

In order to evaluate these hypotheses, I draw upon data from the American National Election Studies (ANES) survey and the Migration Policy Institute (MPI) reports to test the effects of change in state populations due to immigration (MPI) upon views of immigration (ANES). The MPI assembled data from previous US Census results to list the percent change in the immigrant population for each US state. Thus, the data included reflects population change per 10-year period because the Census is taken each decade. ANES surveys though are conducted every four years around US presidential elections. Respondents are surveyed two-months prior to election day and two months afterwards, but the responses are compiled in a single dataset. Questionnaires are presented both in person and online.

The questions posed each time cover the same broad issue areas, but the specific aspects probed may differ. I will make use of the ANES 2012 dataset, with the potential for further time-series studies in mind. I selected 2012 given its temporal proximity to the most recent US Census data available spanning the 2000-2010 time frame, but I made individual variable selections based off the consistency of questioning across 2008 and 2016 ANES surveys as I believe it will be important to understand the effect of time on responses in further research. Thus, considering that ANES questions are inconsistent across years, I selected relevant variables that were common across all three datasets to allow for expansion of the present study. Only one immigration-related question met this criterion: “how likely is it that recent immigration levels will take jobs away from people already here”. I thus chose the corresponding individual-level responses as the indicator for my dependent variable, attitudes on the effects of immigration. Respondents could choose between “extremely”, “very”, “somewhat”, “not at all”, “don’t know” or “refuse”. The distribution of responses is shown below:

Given the nearly empty categories and thin conceptual distinction between levels such as “very” and “extremely”, I chose to simplify the response variable by collapsing the six response options into three, “somewhat”, “not at all, and “extremely”. I do present a model that distinguishes between “extremely” and “very” for comparison below, but the difference in AIC supports my choice to simplify these classifications. This will be elaborted upon in subsequent sections. The three-level response variable changes the distribution to the following:

All ANES surveys also ask respondents about their views of the economy as well. Thus, I followed the same selection method of identifying a common question probing the state of the economy to speak to H2. My main independent variables for this study are (1) percent change in immigrant population per state and (2) perceived performance of the national economy. Perceived performance of the national economy corresponds to the ANES 2012 question: “What do you think about the state of the economy these days in the United States?” Response options included “very good”, “good”, “neither good nor bad”, “bad” “very bad”, “don’t know” and “refused”. I collapsed non-responses into the “neither” option, in addition to combining “good” and “very good” as well as “bad and very bad” given the thin conceptual distinction here again. Percent change was the only continuous variable included, while all others were treated as categorical.

Additionally, I also introduced a number of control variables in line with the literature. First, I add a respondent’s highest level of education, as this is argued to have a bearing on exposure and tolerance to outgroups, as well as the degree of understanding for the implications of immigration in an economy (Mansfield and Mutz 2009). In ANES 2012, participants could answer “less than high school”, “high school diploma”, “some post high school, no bachelor’s”, “bachelor’s degree”, “graduate degree”, or “refused”.

The second control variable is informed by two recent studies in particular that examine the role of a receiving society’s integration policy landscape, defined as “the institutional practices adopted by [government] agencies to deal with the settlement of immigrants in host societies” (Schlueter, Meuleman, and Davidov 2013, 670). Schlueter et al. are also interested in comparing the explanatory power of the Contact and Group Threat hypotheses, but their study of immigration attitudes across Europe uses policy landscape as the explanatory variable rather than immigration flows. They find that countries with more inclusive or welcoming policies towards immigrants are associated with more positive perceptions of immigration, while the opposite holds for more restrictive landscapes. The authors argued that this relationship could be attributed to the norm-shaping role of policies regarding the role of immigrants in a society.

Prior to 2019 when Filindra and Manatschal’s paper was published, it would have been difficult and time-consuming to come by this type of policy classification for the subnational level in the United States (Filindra and Manatschal 2019). This is because the majority of work on integration suffers from methodological nationalism, and scholars are only beginning to consider how national positions on immigrant integration get translated to local levels. In the US, integration is regarded as a private matter that the government should not have a role in, so the only federal directives are those codified in the 1964 Civil Rights Act, which outlawed discrimination on the basis of race, color, religion, sex or national origin (Fibbi, Midtbøen, and Simon 2021). The result is a patchwork of integration frameworks across each state. Filindra and Manatschal looked at every integration-related law passed in the 50 US states from 2004 to 2012 and categorized each one as either restrictive or inclusive of immigrant rights and access to services. Using their data, I then constructed a single variable to indicate whether a state had majority inclusive, restrictive or neutral integration laws. Neutral states are those where there are either no integration-specific policies, or the number of restrictive and inclusive was equal.

In addition to disaggregating by these state-level characteristics, I also included a variable to indicate which geographic region an individual respondent lived in. I did so according to the US Census list of countries within each of the four regions, North, South, West and Midwest (US Census Bureau n.d.). This variable “Region” offers a means of accounting for varying histories with ethnic diversity across the country that may hold influence today. The West, for instance, witnessed an influx of immigrant labor during the California Gold Rush and railroad construction of the 1800s, while around the same time, the South was desperately fighting to uphold slavery. Scholars have found evidence of a stronger white identity in the modern South relative to the rest of the country due to this distinct position during the Civil War (Jardina 2019).

Lastly, I also included a control variable for political party, which is recorded in the ANES 2012 dataset on a scale of “Republican”, “Democrat”, “Independent/other” and by “Strong Republican”, “not very strong Republican”, “Independent Republican”, “Independent”, “Independent Democrat”, “not very strong Democrat” and “Strong Democrat” in 2012. A number of studies demonstrate that partisan divide in the US is associated with positions on immigration (see Dougherty, Nienhusser, and Vega 2010; Monforti, Orey, and Conroy 2009; Sikes and Valenzuela 2016 for more). For instance, Hawley (2011) concluded that Republicans are more likely to support immigration restrictions compared to Democrats and Independents, and further, these views diverged even more significantly in contexts marked by greater immigration.

Each of these variables was incorporated into models of the ANES 2012 survey responses. The final sample size included 283 individual-level observations for 2012. Unfortunately, a number of observations had to be removed in cases where the geographic location was not indicated. This was an insurmountable obstacle given the primary objective of this analysis to study the effect of state-level change in immigration population size on attitudes.

It should also be noted that the use of categorical variables absorbs a degree of nuance that may exist between each measure. For instance, there is an important difference in integration policy landscapes that have only inclusive laws compared to those that have both restrictive and inclusive, but according to my measurement methods, as long the number of inclusive policies outnumbered restrictive ones, the state would be considered overall inclusive. The potential for measurement error is further increased by my choices to combine several of the response options into one such as was done for the dependent variable and education level control. However, I believe doing so benefits interpretation, and as a robustness check, I display model results using the disaggregated, more nuanced response variable for comparison as well.

Methods

Given that my dependent variable, attitudes on the effects of immigration, is categorical with more than two response options, I chose multinomial modelling to fit this data. I considered ordinal logit and a possible nested design as alternatives given the structure of the data, but these fell outside the scope of the class. Nevertheless, I aim to account for the fact that each individual survey respondent is housed inside a larger unit through state and regional-level control variables. The results of the multinomial models presented below point to valuable avenues for further study using more these advanced techniques.

Further, despite the fact that an ordinal logit would have arguably been more fitting given the Likert scale response variable, the two key assumptions for the multinomial logit are nevertheless satisfied in this case. First, the response options are both exhaustive and mutually exclusive, as the broadest range of opinions from “extremely” to “not at all” are included, and selecting both would be contradictory. Of course, more nuance is possible, such as if distinguishing between the effects of high-wage and low-wage immigrants, but the nature of the question investigates overall views, thereby upholding the assumption. Secondly, the conditions for the independence of irrelevant alternatives (IIA) are met as well. While there are ways to statistically verify this assertion, such as the Hausman-McFadden or Small-Hsaio tests, McFadden is clear that outcome categories need only be “plausibly assumed to be distinct and weighed independently in the eyes of each decision maker” (Cheng and Long 2007, 598). Thus, we can say through logical deduction that the introduction of the category “somewhat” should not motivate an individual to change their view from “not at all” to “extremely” and vice versa.

Upon establishing that the primary assumptions are not violated by this dataset and selection of variables, I proceeded to run a range of models to test my hypotheses regarding the relationship between immigration flows, perceived health of the economy and attitudes on the effects of immigration. The first model conveys the effects of including all controls, while Model Two includes only the two main explanatory variables. Model three introduces only individual-level controls (education and party affiliation). Model four considers only controls pertaining to attributes at the state and regional level (region and integration policy landscape). Of the four models constructed, the best fit will be assessed based on AIC values. A final model, Model Five, is presented separately with disaggregated data for the response variable as an additional robustness check.

The reference category for each variable is set to reflect the option closest to neutral, where applicable, so as to reflect the odds according to moves towards greater polarity. The dependent variable, for instance, is set to “somewhat” as a base so that results can be interpreted as the effects of the independent variables on having more extreme negative or positive views about the effects of immigration on job availability. For view of the economy, the reference is “neither”, and thus, the reported odds reflect choosing towards either more positive or negative perceptions. Similarly, integration policy landscape is set to “neutral”. For party identity, the reference is “Independent”. For variables that do not have an obvious middle value on the corresponding Likert scale, I proceeded with the reference categories automatically selected by R. This was “no high school” for education and “Midwest” for region.

Results

An overview of model results is listed in the table below.

In terms of the two main independent variables for this study, perceived performance of the economy and percent change in the immigrant population by state, views of the economy seem to be a stronger predictor of attitudes on the effects of immigration than changes in the immigrant population. Having a good view of the economy leads to a statistically significant (at a 99% confidence level) increase in the odds an individual responds “very” over “somewhat” to the question about the likelihood immigrants will take away jobs across each of the four models. For instance, Model Three, which has the lowest AIC, tells us that having a positive opinion of the economy increases the odds of responding “very” as opposed to “somewhat” by 16.95%. None of the reported effects for having a good view of the economy’s performance on answering “not at all” over “somewhat” are statistically significant.

Reporting a bad view of the economy, however, does have a statistically significant effect with p-values less than 0.01 for both “very” and “not at all” over choosing “somewhat” in each of the four models. The odds of responding “very” over “somewhat” though are higher in each case. In Model Three for example, considering economic performance bad increases the odds of responding “very” over “somewhat” by 19.32% and only 10.95% for answering “not at all” over “somewhat”. Considering the reference category for the economic views variable is “neither (good nor bad)”, we can see that having a strong opinion on the economy’s performance, at least in the case of negative opinions, increases the likelihood that an individual has a stronger view on the effects of immigration on jobs than “somewhat”.

The variable “percent change”, on the other hand, has no significant effects in any of the models above conventional levels. There is an association in Model Three to suggest a one unit increase in percent change of immigration in a state increases the odds of responding “very” over “somewhat” by 12.24%, but this is only reported at the 90% confidence level.

The only classifications of highest education to have statistically significant effects are having high school only and having a graduate degree. Having only a high school education increases the odds of responding “very” over “somewhat” by almost 16% rather than having no high school at all in both models controlling for individual characteristics (Model One and Model Three). There was also a statistically significant effect in having only high school upon responding “not at all” over “somewhat”, but the level of increased odds is less in both cases and is only significant at the 95% confidence level whereas the effects on answering “very” were significant at a 99% level.

Having a graduate degree increases the odds of responding “not at all” over “somewhat” by 18.54% (Model One) and 15.68% (Model Three). Both are significant results at the 95% confidence level, whereas the odds were much smaller and not significant for responding “very” over “somewhat”.

The second individual-level control variable, political party identification, also has significant effects in Models One and Three. At 99% confidence levels, being a Democrat increases the odds for responding both “very” and “not at all” over “somewhat”, as opposed to Independents. However, the increase in odds was greater in both models for “not at all”. Identifying as Republican though appears to have a much more straightforward relationship with immigration attitudes, as being Republican over Independent increases the odds of responding “very” over “somewhat” by 17.84% in Model Three for example, while the odds of answering “not at all” over “somewhat” increase by only 9.13% for Republicans. The effect of answering “very” is statistically significant at the 99% confidence level, while it is not statistically significant for “not at all” in Model Three and only at the 90% level in Model One.

Regarding state-level control variables, region appears to have a clear influence on attitudes about the effects of immigration, but integration policy landscape, though significant, appears more mixed. Living in the North or West is associated with a statistically significant (at 99% confidence levels) increase in the odds of responding “not at all” over “somewhat”, as opposed to living in the Midwest. This was true for both models controlling for state-specific attributes, Models One and Four. The increase in odds is near 10% more than the increase in odds for “very” over “somewhat” in both models for both regions. Being in the South, on the contrary, increases the odds of reporting “very” over “somewhat” more than it increases the odds of “not at all” over “somewhat”. Results for the former are statistically significant at conventional levels (95%), while the latter is not.

Model results for the effects of integration policy landscape are similar to those of negative views of the economy’s performance whereby living in a policy landscape that is either more inclusive or restrictive increases the odds of answering either more positively or more negatively about the effects of immigration, as opposed to the more neutral position of “somewhat”. This is perhaps a more applicable observation for restrictive policy landscapes, as living in this state-level environment is linked to a similar increase in odds for both responding “not at all” over “somewhat” and “very” over “somewhat” - at a 99% confidence level in both models. For instance, in Model One the odds increased 10.46% for “very” and 14.01% for “not at all” over “somewhat”. The gap is larger, however, for inclusive integration policy environments. The odds increase for responding “very” as well as “not at all” over “somewhat” in both models, but the difference is a 16.81% increase in odds for choosing “not at all” over “somewhat” compared to choosing “very” in Model One and 17.02% difference in Model Four. These results are significant at the 95% levels for “very” responses and 99% for “not at all” over “somewhat”.

Given the multitude of interesting insights indicated by the four models above, model selection criteria will be very useful in deciphering the most suitable one for the ANES 2012 data. Model Three, which controls for only individual-level characteristics, has the lowest AIC and thus the most parsimonious fit. I further confirmed this by running a confusion matrix to predict the rate of true positives or accuracy for each model. Model Three again stood out as being the best fit for the data, with a 53.9% accuracy rate. Although I previously deduced that the main criteria for the multinomial model was not violated given the nature of the response variable, I do want to ensure that the assumption regarding multicollinearity of the predictors has not been violated either. The GVIF values below indicate low to moderate correlation, which is not surprising given the natural correlation we might expect for socio-demographic variables such as education, income and party identification for instance. Since none of the GVIF values exceed 5, I conclude that there is no substantial cause for concern.

##                      GVIF Df GVIF^(1/(2*Df))
## econ_ecnow2      5.326165  2        1.519161
## Percent_Change   6.840215  1        2.615380
## dem_edugroup_x2 16.557010  4        1.420276
## pid_self2        5.331261  2        1.519524

The marginal effects of Model Three with respect to the primary explanatory variables are displayed in the plot below. Visual inspection reveals that having this more neutral stance of “neither” regarding the performance of the economy is most closely associated with responding more moderately to the immigration question as well via “somewhat”. Responding either “good” or “bad” about the economy has a very similar effect on the odds of thinking immigrants are very likely to take jobs away. The fact that having a positive view of the economy is also associated with this negative view of the effects of immigration falls contrary to the hypthesized relationship between these two variables outlined in H2. This suggests that although I would accept H2 based on the relationship between reporting “bad” on the economy and “very” about immigration’s effects, a more nuanced interpretation is required since “good” has a very similar effect as well. This speaks to the broader view of insecurity posited by the Group Threat Hypothesis as opposed to only material security which is most directly related to economic health. This leads me to believe that an overall threat to the majority’s group position offers a better explanation for opinions on the effects of immigration rather than material threat alone. Additionally, the marginal effects show that the probabilities of responding “very” clearly increases with increase in percent change of the immigration population, while there is a negative relationship for responding “not at all”.

To check the robustness of these results, I also tested model three using the disaggregated response variable, which includes six response categories. Model Five below presents the robustness results for the disaggregated response variable in the 2012 dataset.

We find a similar relationship for all variables, except percent change in the immigrant population by state does have a significant effect at conventional levels, whereas it did not using the collapsed measure. The relationship is significant for responding “very” rather than “somewhat”, which indicates the most extreme views in the negative direction are influenced by size of the immigrant population though not for others. In fact the increase in odds for responding “very” is only 7.36%, while it is more than double (18.45%) for “extremely”. However, the AIC for Model Five is much higher than Model Three, so although the results are intriguing, Model Three offers a much better fit to the data. I nevertheless suggest that further research, namely using nested models, will be needed to respond more conclusively to H1. In the present study, I lean to reject the hypothesis, but the robustness check should drive further exploration.

Conclusion

In the present study, I used multinomial modeling to investigate whether the Contact Hypothesis or Group Threat Hypothesis better explains the reactions of Americans to increasing immigrant populations by state. I tested two hypotheses pertaining to my primary explanatory variables of, percent change in immigrant population per state (H1) and perceived performance of the national economy (H2). I included a number control variables as well to reflect alternative explanations in the literature. These controls were introduced based on the level of analysis they are most applicable (either individual or state-level), given the nested structure of the data. Results indicate that for H1, percent change is not associated with a significant effect on attitudes about the effects of immigration on jobs, while for H2, we do see that more negative views of the economy’s performance translate to increased odds of believing that immigrants take away jobs. Yet, having positive perceptions of the economy’s performance has a very similar effect as well. I draw two important conclusions from these outcomes: (1) percent change may not be the most efficient predictor for assessing between the Contact and Group Threat Hypothesis, and (2) views of national economic health serve as more of an indicator of the importance of the majority population’s general status rather than the threat to material security specifically. Thus, when the economy is considered to be doing well, the locals want to protect this, while if it the economy is perceived as poor, immigrants are not believed to improve the situation for locals. The insight on H2 in particular leads me to deduce that the Group Threat Hypothesis is more applicable to the US environment, but further research is certainly still needed.

The abundance of information available through ANES surveys, which have been conducted every four years since 1948, is ideal for time-series analysis. Future work should take advantage of this and employ a panel structure to investigate H1 and H2 across US states over time. This should account for the state-specific effects that could not be as effectively addressed in the present study without incorporating a nested design. Models One and Four though indicate interesting and significant associations of region and state policy landscapes that a hierarchical method might better address. Future studies would also benefit from a larger sample size. Nevertheless, the analysis above provides productive insights on the value Americans place on their economic status, as well as how individual characteristics such as education and party affiliation mitigate such an effect on the views of immigration.

References

Allport, Gordon W. 1979. The Nature of Prejudice. Unabridged, 25th anniversary ed. Reading, Mass: Addison-Wesley Pub. Co.

Cheng, Simon, and J. Scott Long. 2007. “Testing for IIA in the Multinomial Logit Model.” Sociological Methods & Research 35(4): 583–600.

Dougherty, Kevin, Kenny Nienhusser, and Blanca Vega. 2010. “Undocumented Immigrants and State Higher Education Policy: The Politics of In-State Tuition Eligibility in Texas and Arizona.” The Review of Higher Education 34(1): 123–73.

Fibbi, Rosita, Arnfinn H Midtbøen, and Patrick Simon. 2021. Migration and Discrimination: IMISCOE Short Reader. https://doi.org/10.1007/978-3-030-67281-2 (May 13, 2021).

Filindra, Alexandra, and Anita Manatschal. 2019. “Coping with a Changing Integration Policy Context: American State Policies and Their Effects on Immigrant Political Engagement.” Regional Studies: 1–12.

Gabrielli, Lorenzo, Sonia Gsir, and Ricard Zapata-Barrero. 2017. Political and Civic Participation of Immigrants in Host Countries. An Interpretative Framework from the Perspective of the Origin Countries and Societies. 1st ed. 2017. Cham: Springer International Publishing : Imprint: Springer.

Graefe, Deborah Roempke et al. 2008. “Immigrants’ TANF Eligibility, 1996–2003: What Explains the New Across-State Inequalities? ^{.” International Migration Review 42(1): 89–133.}

Hawley, George. 2011. “Political Threat and Immigration: Party Identification, Demographic Context, and Immigration Policy Preference*: Political Threat and Immigration.” Social Science Quarterly 92(2): 404–22.

Jardina, Ashley. 2019. White Identity Politics. https://doi.org/10.1017/9781108645157 (May 13, 2021). Mansfield, Edward D., and Diana C. Mutz. 2009. “Support for Free Trade: Self-Interest, Sociotropic Politics, and Out-Group Anxiety.” International Organization 63(3): 425–57.

Monforti, Jessica Lavariega, Bryon D’Andra Orey, and Andrew Conroy. 2009. “The Politics of Race, Gender, Ethnicity and Representation in the Texas Legislature.” The Journal of Race and Policy 5(1): 35–53.

Schlueter, Elmar, Bart Meuleman, and Eldad Davidov. 2013. “Immigrant Integration Policies and Perceived Group Threat: A Multilevel Study of 27 Western and Eastern European Countries.” Social Science Research 42(3): 670–82.

Sikes, Chloe, and Angela Valenzuela. 2016. “Texas Dream Act.” Texas State Historical Association. https://tshaonline.org/handbook/online/articles/mlt03.

US Census Bureau. “Census Regions and Divisions of the United States.” https://www2.census.gov/geo/pdfs/maps-data/maps/reference/us_regdiv.pdf.

R Packages Usedt

Tidyverse Haven Readxl Stargazer sjPlot table1 ggplot2 scales gridExtra car nnet survey