Damon and Rogers Power Point

1
Damon and Rogers Power Point

• UCLA
• 2007

2
Modelling Strategy

• Modelling GCSE results as a Binary Outcome vs.
  Modelling GCSE Scores
• Fitting standard logistic and linear regression
  models without school effects
• Fitting two-level logistic and linear regression
  models with school effects
• Parental Education vs. Parental Class which
  provides more explanatory power?

3
Initial Logistic Regression including Parental
Education
4
Final Logistic Regression including Parental
Education
5
Interpretation of Final Logistic Regression
including Parental Education

• Gender significant females outperform males
• Age significant younger respondents report
  better results than older ones
• Parental education significant those
  respondents with at least one parent with a
  degree have significantly better results
• Model explains 11 of variation in outcome

6
Final Logistic Regression including Parental Class
7
Interpretation of Final Logistic Regression
including Parental Class

• Parental Class included instead of Parental
  Education
• A similar range of significant explanatory
  variables emerge, including age and gender
• Parental Class significant those from business,
  professional or managerial families score better
  at GCSE than those from other class categories
• This model explains 12 of variation

8
Two-Level Logistic Regression including Parental
Education
9
Interpretation of Two-Level Logistic Regression
including Parental Education

• Include school-level explanatory variables in
  two-level logistic regression model incorporating
  a school-specific random effect
• Final model reveals importance of gender, age and
  parental education
• No school effects significant
• Model explains significant amount of variation in
  outcome, whilst leaving unexplained a significant
  amount of between-school variation

10
Two Level Regression Model including Parental
Class
11
Interpretation of Two-Level Regression including
Parental Class

• A similar pattern emerges gender, age and
  parental class significant

12
Conclusions from fitting one-level and two-level
logistic regressions

• Age, gender and family background are significant
  factors in determination of GCSE results when
  measured as a binary outcome
• Parental class provides greater explanatory power
  than parental education
• Difference in resolution a six-category
  variable should, pari passu, provide greater
  predictive power than a simple dichotomy
• No school-level factors significant

13
A Note on Ethnicity

• Ethnicity appears significant in final one-level
  models Indians perform significantly worse than
  other ethnic groups
• This ethnic effect ceases to be significant in
  two-level models
• This suggests that initial (spurious) Indian
  effect is a result of school-level factors

14
Two Level Random Intercept Model including
Parental Education
15
Interpretation of Two-Level Random Intercept
Model including Parental Education

• Gender, age and parental education again prove
  significant in a similar pattern to earlier
  logistic regression models
• Some school-level effects are now significant

16
Significant School Effects

• Respondents attending positively selective
  schools score an average of about 19 points more
  than others
• Those at Christian schools score almost 8
  additional points
• Those at 11-18 schools and mixed-sex schools also
  perform significantly better

17
A Further Note on Ethnicity

• Ethnic effect now appears rather different
• Pakistanis obtain significantly better results
  than other ethnic groups
• This suggests that the apparent Indian effect
  in earlier logistic regression models was
  spurious and was an artefact of ignoring
  school-level effects

18
Two Level Random Intercept Model including
Parental Class
19
Interpretation of Two-Level Random Intercept
Model including Parental Class

• A more complex, fine-grained picture emerges
• Gender, age and parental class again significant
• Example children from professional family
  backgrounds obtain almost 8 additional points
  (equivalent to an A)

20
More Interpretation of Two-Level Random Intercept
Model including Parental Class

• Ethnicity also significant
• Pakistanis and Bangladeshis score significantly
  more on average than Whites
• School effects also significant
• Example respondents at positively selective
  schools score an average of almost 20 additional
  points

21
Conclusions

• The continuous measure provides increased
  resolution.
• Some school effects are significant others are
  not
• Some general factors are significant others are
  not

22
Advantages of a Statistical Modelling Framework

• We can examine the relative significance of a
  wide range of explanatory factors simultaneously.
  i.e. we can control/adjust for a range of
  potential explanations
• We can test the explanatory power of a wide range
  of different sociological theories of the
  determinants of educational attainment
• We can also investigate interactions
• Does the gender difference vary significantly
  between ethnic groups?

23
EFFNATIS SAMPLE
Table A The Main Sample
24
EFFNATIS SAMPLE
Table A The Main Sample
25
Ethnicity and Religion
Table B Ethnic Grouping and Religious
Affiliation
26
Language Spoken at Home with Parents
Table C Language Used at Home with Parent(s)