Results

I present the results of my analyses in three consecutive steps. First, I demonstrate the country-specific impact of parental education on propensity to study. Then, I explore fixed effects for individual-level factors; finally, I introduce the cross-level interaction effects between parental education and the institutional environment of the higher education system.

Turning to the country-specific inequality of higher education access, a multilevel logistic is estimated in which the coefficient of parental education is allowed to vary by country (random slopes). In this model, only the individual-level data taken from the ESS is used. For illustrative purposes, the explanatory variable parental education is treated as continuous in this model (in all other models, it is treated as categorical).

Because at this point I am not interested in the fixed overall effect of parental education, only the coefficients of the mean slope (fixed slope + random slope) are reported in Fig. 8.3. On the x-axis, the steepness of the combined slopes can be observed for each country in the sample. Unsurprisingly, the model shows a strong impact of parental education on propensity to study for all countries in the sample, as all slopes have negative overall coefficients.

However, effect magnitude varies starkly between countries. On the one hand, the effect of parental education is strongest in Eastern European countries and Germany. On the other hand, the most equal countries are Sweden, the Netherlands and—perhaps surprisingly so— the United Kingdom and France. With regard to the higher education ideal types, no clear initial patterns emerge. While mass public models generally display low levels of inequality, the three countries with the biggest impact all belong to different ideal types. However, the volatility of the effect size by country (ranging from —0.18 in Sweden to —0.72 in the

Effect of parental education on propensity to study, by country

Fig. 8.3 Effect of parental education on propensity to study, by country

Slovak Republic) gives credence to the notion that equality of opportunity in access to higher education is realized to very different extents in the sample countries. This raises the question as to whether, despite the initial impression, part of these differences can be explained by the institutional characteristics of the higher education systems.

In order to answer this question, the results of the main analysis are presented in Tables 8.1 and 8.2. In Model 1, only the central independent variable (in its categorical form) is included in the regression. In Model 2, the other individual variables, Female, Foreign Born and Migration Background, are added to the equation. In Model 3, accordingly, the interaction terms with the macro-level variables are added. Coefficients are reported as odds ratios, meaning that point estimates under 1 denote a negative relationship between the variable and the probability to study, and point estimates larger than 1 denote a positive one.

Irrespective of the specification, the models show a strong and highly significant impact of parental education on the likelihood to be a student.

Table 8.1 Multilevel regression results: micro specifications

Model 1

Model 2

Parental Education

Parental Education = 1

0.686*** (0.0422)

0.690*** (0.0418)

Parental Education = 2

0.338*** (0.0183)

0.339*** (0.0181)

Parental Education = 3

0.262*** (0.0174)

0.261*** (0.0171)

Parental Education = 4

0.189*** (0.0134)

0.190*** (0.0132)

Parental Education = 5

0.135*** (0.0154)

0.136*** (0.0151)

Parental Education = 6

0.0985*** (0.00908)

0.0986*** (0.00883)

Female = 1

1.328*** (0.0448)

Foreign Born = 1

0.854+ (0.0792)

Migration Background

One parent born outside

0.944 (0.0664)

country

Both parents born outside

1.039 (0.0886)

country

Number of observations

168,868

16,278

Number of groups

22

22

SD of random intercepts

0.44

0.40

Exponentiated coefficients are reported. Standard errors in parentheses. Yearly dummies and constants are included in the models, but not reported + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001

Compared with the reference category (both parents hold a higher education degree), odds ratios are consistently below 1 and decrease with every additional category denoting a lower level of parental education. In addition, effect sizes barely change as more variables are added to the estimation, indicating little multicollinearity with gender and migration background. In order to illustrate the impact of each category of parental education more intuitively, predicted probabilities for each category were calculated for Model 2 and are graphed in Fig. 8.4.

The model predicts a 62 per cent likelihood of pursuing an academic degree for the five-year age group following secondary school-leaving for survey respondents both of whose parents hold a higher education degree—true across all countries and years. Interestingly, the predicted probability does not decrease linearly, as the drop-off from values 0 to 1 (essentially denoting that one parent holds a higher education degree and the other an upper secondary degree) is just 9 per cent, but the drop-off from values 1 to 2 (which in more than 90 per cent of the cases denotes

e margins of parental education, Model 2

Fig. 8.4 e margins of parental education, Model 2

that both parents hold an upper secondary degree) is much stronger at 17.3 per cent. In other words, experience and success in higher education even by one parent leads to a considerably greater probability of an individual to pursue an academic degree. Once the level of parental education reaches the value of 2, however, the propensity to study decreases rather linearly, with higher categories associated with a decreasing likelihood of close to 5 per cent each.

As for the control variables, the model predicts a highly significant (p < 0.01) effect for gender. According to the corresponding odds ratio, being female increases the likelihood of pursuing an academic degree by a factor of1.3 as compared to males.[1] With regard to migration background, the model detects a marginally significant negative effect for being born in a foreign country. The migration background of the parents, however, does not seem to impact the propensity to be a higher education student.

Turning to the inclusion of macro-level variables in Models 3 and 4 (Table 8.2), the interaction between parental education and enrolment

Table 8.2 Multilevel regression results: micro-macro specifications

Model 3

Model 4

Enrolment Ratio

  • 1.007+
  • (0.00417)

1.005 (0.00360)

Public subsidization per student

  • 0.989*
  • (0.00469)

0.987* (0.00606)

Interaction Parental Education x

Enrolment Ratio

Parental Education = 1

  • 0.994+
  • (0.00308)

Parental Education = 2

  • 0.997
  • (0.00281)

Parental Education = 3

  • 1.000
  • (0.00353)

Parental Education = 4

  • 1.006
  • (0.00399)

Parental Education = 5

  • 0.998
  • (0.00850)

Parental Education = 6

  • 0.996
  • (0.00710)

Interaction Parental Education x

Public Subsidization

Parental Education = 1

0.999 (0.00552)

Parental Education = 2

0.995 (0.00488)

Parental Education = 3

1.004 (0.00602)

Parental Education = 4

1.022** (0.00672)

Parental Education = 5

1.023+ (0.0139)

Parental Education = 6

1.035** (0.0118)

Observations

16,203

16,203

Number of groups

22

22

SD of random intercepts

0.41

0.42

Exponentiated coefficients are reported. Standard errors in parentheses. The micro-specification shown in Model 2 is included in the models, but not reported + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001

ratio is largely insignificant. The sole (marginally) significant coefficient is for Parental Education = 1. The corresponding odds ratio is 0.994, suggesting that with increasing values of enrolment ratio, the gap between individuals from that group and from the reference category widens, albeit at a small margin.[2] Given the small effect size and the fact that the impact

Interaction of parental education and public subsidization. Note

Fig. 8.5 Interaction of parental education and public subsidization. Note: dy/dx for factor levels is the discrete change from the base level

of parental education is not moderated by enrolment for any other group, the overall interaction is judged to possess little explanatory power. This is different, however, for public subsidization. Here, the model yields a moderating impact on the effect of parental education. More precisely, the interaction is significant for values 4 through 6 in parental education.[3] The positive odds ratios for these three groups indicate that the gap between them and the reference group gets smaller as public subsidization increases. Hence, respondents whose parents are not particularly well educated see their likelihood of studying increase at high levels of public subsidization.

To illustrate, marginal effects of Model 4 are plotted in Fig. 8.5. Over the entire range of the variable, the effect of parental education is strongest for low levels of public subsidization, where there are marked differences between all seven groups. As public subsidization increases, the likelihood of individuals with parental education values 1, 2 or 3 to be students is relatively stable, compared to the reference group. However, the remaining groups with poor educational backgrounds catch up considerably. At a public subsidization rate of about 60 per cent of GDP per capita, there are no discernible differences between them and respondents whose parents have an aggregate education score of 2 or 3. At such a rate, these groups are 25 to 30 per cent less likely to pursue an academic degree than individuals from the reference group. While there remains a decided advantage for individuals whose parents have successfully completed a higher education degree, the catch-up effect of low socio-economic status individuals under conditions of high public subsidization is remarkable.

  • [1] When all other variables are held at their means, the coefficient translates into a predicted probability of 39.3 per cent for females and 32.5 per cent for males.
  • [2] Over the entire range of the enrolment variable, the gap increases from 4.8 per cent (EnrolmentRatio = 25) to 14.9 per cent (Enrolment Ratio = 100), with all other variables held at their means.As a reminder, the overall fixed gap between these two groups was 9 per cent (see above).
  • [3] The fact that the interaction is only marginally significant for the value 5 is most likely an artifactof the corresponding group. Only 610 respondents (3.6 per cent of overall sample) belong to thiscategory, naturally leading to larger standard errors.
 
Source
< Prev   CONTENTS   Source   Next >