As we explained in chapter 2, to process our data we will use a cumulative odds ordinal logistic regression with proportional odds. For this method to be valid, some assumptions must be met. Those assumptions are the following 4:
1. There is one dependent variable that is measured at the ordinal level.
2. There are one or more independent variables (continuous, ordinal or categorical). 3. There should be no multicollinearity.
4. We have proportional odds.
Assumptions 1 & 2
The first two assumptions relate to our choice of variables and it is easy to see if they are met. In our study they are. Our dependent variable, diffuse support is indeed our only dependent variable and it is measured in the ordinal level due to the Likert type question used to capture the data. Moreover, we have a number of independent variables which are continuous, ordinal or categorical. When it comes to the last two assumptions, they relate to how well our data fits the ordinal logistic regression model. The better the fit the more we can be sure that our results are valid. Unfortunately, testing for assumptions 3 and 4 is not as obvious as for 1 and 2 and requires statistical tests. The next paragraphs will do just that.
Assumption 3
Firstly, we will test our model for multicollinearity. Multicollinearity occurs when two or more independent variables are highly correlated with each other, making it difficult to know which independent variable is responsible for the variance in the dependent variable. Moreover, it causes technical problems in running an ordinal logistic regression. To test for multicollinearity we will create dichotomous dummy variables for each of our categorical independent variables. We will then run a linear regression in order to obtain the “Coefficients” table. By examining the VIF value for every variable we are able to determine if we violate the assumption of no
multicollinearity and in the cases that we do, we can remove the variables that cause the problem. For us not to have multicollinearity in our model the VIF values should be lower than 10. As the coefficients table shows in the Appendix, our model is free of multicollinearity, thus meeting the third assumption.
Assumption 4
Having met the third assumption, it is time to test for the 4th and final assumption; that of proportional odds. What this assumption entails is that each independent variable has an identical effect on the dependent variable for each cumulative split of the dependent variable. We can test this assumption using a full likelihood ratio test which compares the fit between the proportional odds model and a model with varying location parameters:
Test of Parallel Linesa
Model
-2 Log
Likelihood Chi-Square df Sig. Null Hypothesis 32581.125
General 30794.773b 1786.353c 312 .000
The null hypothesis states that the location parameters (slope coefficients) are the same across response categories.
a. Link function: Logit.
b. Maximum number of iterations were exceeded, and the log-likelihood value and/or the parameter estimates cannot converge.
c. The Chi-Square statistic is computed based on the log-likelihood value of the last iteration of the general model. Validity of the test is uncertain.
Table 4
The difference in fit between the two models is large and statistically significant which means that the assumption of proportional odds is violated according to this test. However, the full likelihood ratio test which we performed may register violations that do not exist, especially when the sample size is very large, as is the case for our dataset. We will therefore continue with testing the goodness of fit before we evaluate the overall suitability of our method.
Goodness of Fit
To test the goodness of fit we will use the Pearson and Deviance statistics as well as the Likelihood ratio test. The Pearson and Deviance statistics measure how poorly the model fits the data. This is done by comparing our model with the full model (a model that fits the data perfectly). As a result, we want this difference to not be statistically significant.
Goodness-of-Fit
Chi-Square df Sig. Pearson 58886.006 62944 1.000
Deviance 32581.125 62944 1.000
Link function: Logit.
Table 5
As shown in the above table, both tests give a clear result which can be interpreted as our model being a good fit to the observed data. Despite those results we cannot be sure if the model is indeed a good fit because when we analyze the data we get the following warning message:
Warnings
There are 63052 (80.0%) cells (i.e., dependent variable levels by observed combinations of predictor variable values) with zero frequencies.
This message warns us of having too many cells with zero frequencies and neither the Pearson nor the Deviance test give reliable results if there are many cells with zero frequencies. Consequently we will make use of the Likelihood-ratio test to see if we get similar results. The Likelihood-ratio test looks at the change in model fit when comparing the final model to the intercept-only model:
Model Fitting Information
Model
-2 Log
Likelihood Chi-Square df Sig. Intercept Only 41502.532
Final 32581.125 8921.407 104 .000
Link function: Logit.
Table 6
The results of this test show a clear statistically significant result. This means that the independent variables add statistically significantly to the model. The three tests give us the same result so we can be confident that the model is a good fit for the data.
Taking all into account, we can conclude that the cumulative odds ordinal logistic regression with proportional odds is a valid method for our research. It may not be ideal since the assumption of proportional odds may be violated but when one is dealing with real world data, especially in the social sciences, this kind of problems are unavoidable. In other words, this method is the best feasible option for the kind of data that we have.
Hypotheses Testing
Hypothesis 1: Civic knowledge in the EP is causally related to diffuse support for the EP. This is our main hypothesis. We expect to see a strong causal relationship between understanding how the EP works and supporting it. Since it is almost impossible to measure understanding per se, we use knowledge of how the regime works as a proxy. Holding correct information about something doesn’t always lead to understanding it, but in our case, it probably does. The previous statement would mostly be true about certain highly abstract concepts that we usually find in science and that require high levels of intelligence to process. Understanding the workings of an institution however, requires far less processing power than theoretical physics for example. For something as material and concrete as the EP, holding the right information should be enough for the vast majority of the population to lead to
understanding. It is therefore safe to assume that knowledge leads to understanding when it comes to political objects such as the EP.
Our main independent variable, civic knowledge is measured with the use of Likert type questions. The number of questions answered correctly is used to create a score of 0 to 6, with 0 meaning no civic knowledge whatsoever and 6 meaning the highest possible level of civic knowledge that we can measure. The variable that is created can be understood both as a scale and as an ordinal variable. Scale, because the distance between 1 and 2 is exactly the same as the distance between 5 and 6; namely, one correct answer. Nominal, because the underlying concept is impossible to have been captured perfectly by the questions and therefore the amount of civic knowledge between 1 and 2 may differ from the amount of civic knowledge between 5 and 6. For this reason and to be completely sure of the results of our first and most important hypothesis we treat the variable as both scale and ordinal. We run two ordinal regressions and get two Parameter Estimates tables which will tell us if there is a statistically significant relationship between the dependent variable, diffuse support, and our main independent variable, civic knowledge.
Parameter Estimates
Estimate Std.
Error Wald df Sig.
95% Confidence Interval Lower Bound Upper Bound Threshol d Diffuse Support = -2 -2.209 .034 4100.814 1 .000 -2.277 -2.141 Diffuse Support = -1 -.461 .027 287.320 1 .000 -.514 -.408 Diffuse Support = 0 1.759 .029 3585.806 1 .000 1.702 1.817 Diffuse Support = 1 4.262 .045 9041.507 1 .000 4.174 4.350
Location Civic Knowledge .187 .007 782.052 1 .000 .174 .201 Link function: Logit.
Table 7 Parameter Estimates Parameter B Std. Error 95% Wald Confidence
Interval Hypothesis Test Exp(B)
95% Wald Confidence
Interval for Exp(B)
Lower Upper
Wald Chi-
Square df Sig. Lower Upper Threshold Diffuse Support = -2 -3.306 .0388 -3.382 -3.230 7271.101 1 .000 .037 .034 .040
Diffuse Support = -1 -1.556 .0313 -1.617 -1.494 2474.464 1 .000 .211 .198 .224 Diffuse Support = 0 .668 .0299 .610 .727 500.395 1 .000 1.951 1.840 2.069 Diffuse Support = 1 3.172 .0442 3.085 3.258 5148.476 1 .000 23.847 21.868 26.006 Civic Knowledge = 0 -.894 .0564 -1.005 -.784 251.585 1 .000 .409 .366 .457 Civic Knowledge = 1 -1.021 .0513 -1.121 -.920 396.274 1 .000 .360 .326 .398 Civic Knowledge = 2 -.860 .0441 -.947 -.774 380.885 1 .000 .423 .388 .461 Civic Knowledge = 3 -.525 .0405 -.604 -.445 167.982 1 .000 .592 .547 .641 Civic Knowledge = 4 -.255 .0376 -.328 -.181 45.843 1 .000 .775 .720 .835 Civic Knowledge = 5 -.166 .0376 -.239 -.092 19.342 1 .000 .847 .787 .912 Civic Knowledge = 6 0a . . . . 1 . . (Scale) 1b
Dependent Variable: Diffuse Support for the EP
Model: (Threshold), Total Civic Knowledge about the EP a. Set to zero because this parameter is redundant. b. Fixed at the displayed value.
Table 8
The results of both tables show us that civic knowledge has a statistically significant effect on diffuse support no matter how we treat it. Despite that, these regressions don’t take into account any possible confounders of this relationship. In other words, we cannot be sure that the effect can really be attributed to civic knowledge because we have not controlled for the effect of any other variables. We therefore run more regressions testing for possible confounders. First we will treat civic knowledge as a scale variable and we will get the following tables:
Tests of Model Effects
Source
Type III Wald Chi-
Square df Sig. Total Civic Knowledge about the EP 19.304 1 .000
Political Interest Index 3.861 3 .277
Evaluation of Societal Utility from membership 418.039 1 .000
Optimism about the national economy 61.676 3 .000
Optimism about personal utility 85.116 4 .000
Perception of personal [my voice counts] inclusion in the democratic process of the EU
394.535 3 .000
Perception of the country's [my country's voice counts] inclusion in the democratic process of the EU
96.883 3 .000
Political Ideology – Left / Right placement 13.091 9 .159
Political Outcomes evaluation - in our country .960 2 .619
Political Outcomes evaluation - in the EU 455.015 2 .000
Mobile and Landline access 6.109 2 .047
Country 522.549 29 .000
Evaluation of the level of cooperation – facing the economic crisis 15.717 2 .000
Gender 27.005 1 .000
Occupation of respondent 15.743 16 .471
Level in society - self placement 38.252 9 .000
Respondent cooperation 8.628 3 .035
Information Access 3.547 1 .060
Intensity of internet use 11.540 1 .001
Education .364 1 .547
Age .882 1 .348
Duration of Interview 1.326 1 .250
Dependent Variable: Dif_Sup_EP
Model: (Threshold), Total Civic Knowledge about the EP, POLITICAL INTEREST INDEX (D71 SUMMARIZED), Evaluation of Societal Utility from membership until now, EUROPE - BRINGS CITIZENS TOGETHER, RETURN TO GROWTH - PERSPECTIVE, FINANCIAL SITUATION HH - NEXT TWO YEARS, Perception of personal [my voice counts] inclusion in the democratic process in the EU, Perception of the country's [my country's voice counts] inclusion in the democratic process in the EU, LEFT-RIGHT PLACEMENT, DIRECTION THINGS ARE GOING - IN (OUR COUNTRY), DIRECTION THINGS ARE GOING - IN THE EU, Mobile and Landline access, COUNTRY/SAMPLE ID (SERIES STANDARD), FACING THE CRISIS - EU MEMBER STATES ACTION (TOT), GENDER, OCCUPATION OF RESPONDENT, LEVEL IN SOCIETY - SELF PLACEMENT, RESPONDENT COOPERATION, requires hardware (PC / laptop or smartphone / tablet) and connection to the internet (at home / at work or other place), intensity of internet use, AGE EDUCATION, AGE EXACT, DURATION OF INTERVIEW
Table 9
In this regression, civic knowledge’s effect appears to be statistically significant with a Wald x2 value of 19.304. To be completely sure about the validity of this conclusion we will run another regression, identical to the previous one only this time our main independent variable, civic knowledge will be treated as an ordinal variable.
Tests of Model Effects
Source
Type III Wald Chi-
Square df Sig. Total Civic Knowledge about the EP 35.583 6 .000
Political Interest Index 3.768 3 .288
Evaluation of Societal Utility from membership 417.441 1 .000
European Identity 256.335 3 .000
Optimism about the national economy 61.451 3 .000
Optimism about personal utility 83.801 4 .000
Perception of personal [my voice counts] inclusion in the democratic process of the EU
392.890 3 .000
Perception of the country's [my country's voice counts] inclusion in the democratic process of the EU
97.367 3 .000
Political Ideology – Left / Right placement 13.136 9 .157
Political Outcomes evaluation - in our country 1.017 2 .602
Political Outcomes evaluation - in the EU 455.146 2 .000
Mobile and Landline access 6.117 2 .047
Country 521.503 29 .000
Evaluation of the level of cooperation – facing the economic crisis 15.741 2 .000
Gender 26.883 1 .000
Occupation of respondent 15.502 16 .488
Level in society - self placement 38.390 9 .000
Respondent cooperation 8.500 3 .037
Information Access 3.446 1 .063
Intensity of internet use 11.345 1 .001
Education .419 1 .517
Age .757 1 .384
Dependent Variable: Diffuse Support for the EP
Model: (Threshold), Total Civic Knowledge about the EP, POLITICAL INTEREST INDEX (D71 SUMMARIZED), Evaluation of Societal Utility from membership until now, EUROPE - BRINGS CITIZENS TOGETHER, RETURN TO GROWTH - PERSPECTIVE, FINANCIAL SITUATION HH - NEXT TWO YEARS, Perception of personal [my voice counts] inclusion in the democratic process in the EU, Perception of the country's [my country's voice counts] inclusion in the democratic process in the EU, LEFT-RIGHT PLACEMENT, DIRECTION THINGS ARE GOING - IN (OUR COUNTRY), DIRECTION THINGS ARE GOING - IN THE EU, Mobile and Landline access,
COUNTRY/SAMPLE ID (SERIES STANDARD), FACING THE CRISIS - EU MEMBER STATES ACTION (TOT), GENDER, OCCUPATION OF RESPONDENT, LEVEL IN SOCIETY - SELF PLACEMENT, RESPONDENT
COOPERATION, requires hardware (PC / laptop or smartphone / tablet) and connection to the internet (at home / at work or other place), intensity of internet use, AGE EDUCATION, AGE EXACT, DURATION OF INTERVIEW
Table 10
The result is a clear statistical significance with a Wald x2 value of 35.583 (almost double from the previous one). The key takeaway from those regressions is that civic knowledge is linked with a real causal relationship to diffuse support, no matter how we decide to treat it and taking into account possible confounders. Our first hypothesis is confirmed.
Before we move on to the next hypothesis, it is interesting to describe briefly what the rest of our results mean and how they compare to the rest of the bibliography. As we said on an earlier chapter, for some scholars demographic factors are important in determining support for a regime. From our data however we do not get as clear a picture. We can see that some demographic factors are important while others are not. For example, the effect of age is not statistically significant but the effect of gender is. These findings offer little support to psychological explanations that have been proposed. Other factors that can be connected to psychological explanations and appear frequently in the bibliography are identity and trust. Our results show a strong causal relationship between feeling European and supporting the EP. Similarly, trust is causally related to support with a very strong relationship, just like Easton had suggested (Easton, 1975, p. 447).
Moving away from psychological explanations, a market liberalization explanation of support also seems inadequate. Education for instance seems to have a small and not statistically significant effect on support. Optimism about the national economy and household
finances on the other hand are good predictors of support. Occupation is another factor that can explain to a small degree the variation in support.
Instead of focusing on market liberalization as an explanation that stands alone, we can consider market explanations in general to be a narrower subset of explanations focusing on utility or performance. Then our research offers further evidence in favor of such explanations. Variables that measure evaluations of societal and personal utility for the past and for the future have a strong and statistically significant effect on support. Another interesting observation related to evaluations of utility or performance is the fact that EU citizens appear to be aware of the distinction between European and National affairs, with the former having a strong and significant effect on support for the EP and the later having a weak and not significant effect.
In this paragraph we would like to make mention to the effect that the various forms of inclusion to the democratic process can have on support for the EP. In our opinion inclusion could be part of both psychological and utilitarian explanations. A psychological explanation could be that, when someone is included he or she feels part of the polity and therefore supports it. A utilitarian explanation could be that people only extend their support to a regime that serves their interests well, or in other words, a regime that they can control. Our data point towards the conclusion that inclusion however broadly or narrowly defined is an important predictor of support.
Finally, we should acknowledge that the results presented above are just a first picture and should be taken with a fair amount of healthy skepticism. In order to avoid multicollinearity we excluded some interesting factors, demographic or otherwise, such as the degree of urbanization, attachment to one’s state or city and having children or being married. The potential relationships of support are practically infinite and it is impossible to measure them all. We are mainly interested in one of them and in order to have a better image of how civic knowledge affects support we need to compromise on our ability to evaluate which theory offers the best overall explanation of support.
Before we move on to the next hypothesis, let us point out one interesting anomaly. In table 8, we see that people with 0 correct answers have a smaller coefficient than those with 1 correct answer (highlighted boxes). One possible explanation for those results is that many people may have given wrong answers on purpose, thus being placed in a lower category of knowledge than their actual one but on their actual category of support (supposing that they didn’t answer falsely the question that measured support as well). In the author’s opinion however, it is more probable that we do not understand the full mechanism of how knowledge impacts support. This is why we see people with no knowledge at all being more likely to support the EP than those with very low knowledge. It is possible that people with very low knowledge are not aware of how much they do not know and aside from being less inclined to support the EP, they are also driven by their false belief that they are knowledgeable and don’t second guess themselves, resulting in a more extreme expression of their opposition. In contrast, people with no knowledge may be relatively more aware of their ignorance. This ignorance leads them to low levels of support for the EP but the very realization of the fact that they know nothing about the EP might lead them to take a less extreme stance towards it. Perhaps if we want to understand how the basic principles of the relationship between knowledge and support work, it is better to study it around the center and not in the extremes.
Hypothesis 2: The effect of civic knowledge on support for the EP does not differ much from country to country.
Given that the mechanism that we study only makes sense on the individual level, and given that human nature doesn’t change, we expect the effect of knowledge to be the same or similar across states, regardless of the overall levels of support or knowledge.
We know from the previous hypothesis that the variable “country” has a strong and statistically significant effect on support. This means that the country one lives in has a causal