1. Interactions between
the individual and the group:
reflections from multilevel modelling
in educational research
Federica Russo
Center Leo Apostel, VrijeUniversiteitBrussel&
Centre for Reasoning, University of Kent
2. Overview
Philosophy of education and empirical research
Reverse the question:
does empirical research look into philosophy?
Multilevel models in educational research
Definition and examples
The need for an accompanying ‘substantive theory’
A ‘substantive theory’ for multilevel
Recent work by Little and Yilokoski
Main features
2
4. Does PhilEd pay enough attention
to empirical research?
Phillips Hyslop-Margison&Naseem
J Phil Ed 39(4), (2005) Phl Ed Archive (2007)
• PhilEd hasn’t paid enough
and serious attention to • Straw man: wrong selection
empirical research of critics
• It is possible to study • Counterexamples: PhilEd
normative processes does pay attention to
empirically empirical research
• Mutual benefit of PhilEd • Problem of empirical
and empirical research to generalisability
look into real cases
4
5. PhilSci and PhilEdu
Phillips (2005, p.582):
The marked change in doing philosophy of science came
about when it was realised that there was much to gain by taking
scientific research seriously, rather than discussing
an artefact of the philosophers’ imagination. […]
The present essay is making a call for a parallel revolution in
philosophical discussions of educational research, a revolution that
entails taking examples of educational research seriously. […]
Processes that humans engage in, in the real world, whether
normative or cultural or psychological (or all three at once) can be
studied—and probably ought to be studied—empirically, but
they also need to be assessed in terms of the values (and if
relevant the conception of education) that they
embody.
5
6. Turn the question on its head
Does empirical research look sufficiently
into philosophy of education?
(Or, for the matter, into philosophy?)
8. Modelling in the social sciences
Causal relations in social contexts
Marital problems ⇄ migration
Maternal education → child survival
Stress + physical health + …→ self-rated health
Two approaches
Qualitative: smaller and focused samples
Quantitative: statistical analyses of large data
8
10. Why multilevel?
An example of quantitative methods used in
empirical research in education
Going quantitative, the new panacea for evidence
But … is it really panacea?
It models hierarchical structures
Typical of social (and education) contexts
A sounding board for the question:
does empirical research look into philosophy?
10
11. Multilevel models
A special type of statistical model used in causal analysis
to model hierarchical structures:
Individuals / family / local population / national population
Firms / regional market / national market / global market
Pupils / classes / school / school systems
No a priori reason to choose the level of analysis
Actually, good reasons to study the interactions between
the levels
11
12. Traditional approaches
Holism
the system as a whole determines the behaviour of the
parts in a fundamental way; the properties of a given
system cannot be reduced to the mere sum of its
components
Individualism
social phenomena and behaviours have to be explained by
appealing to individual decisions and actions, without
invoking any factor transcending them
12
13. The ‘statistical’ counterparts
Aggregate-level models
explain aggregate-level outcomes through aggregate-level
variables
Individual-level models
explain individual-level outcomes by individual-level
explanatory variables
13
14. Types of variables
Individual: measure individual characteristics, take
values of each of the lower units in the sample.
e.g. income of each individual in the sample
Aggregate: summary of the characteristics of
individuals composing the group
e.g.: mean income of state residents
14
15. Dangers
Atomistic fallacy
wrongly infer a relation between units at a higher level of
analysis from units at a lower level of analysis
Ecological fallacy
draw inferences about relations between individual level
variables based on the group level data
15
16. Robinson:
illiteracy and immigration
1930 census in the US, for each of 48 states + district of
Columbia
Individual correlation: descriptive properties of individuals
Positive correlation: immigrants more illiterate than native citizens
Ecological correlation: descriptive properties of groups
Negative correlation: correlation between being foreign-born and
illiterate magnified and in the reversed direction
Explanation: immigrants tend to settle down in states where
native population is more literate
16
17. Courgeau:
Farmers’ migration in Norway
Data from the Norwegian population registry (since 1964) and
from two national censuses (1970 and 1980)
Aggregate model and individual model show opposite results:
Aggregate: regions with more farmers are those with higher rates
of migrations;
Individual: in a same region migration rates are lower for farmers
than for non-farmers
Reconciliation: multilevel model
aggregate characteristics (e.g. the percentage of farmers)
explain individual behaviour (e.g. migrants’ behaviour)
17
18. Types of models - summary
Individual: explain individual-level outcomes by individual-
level explanatory variables
e.g.: explain the individual probability of migrating through the
individual characteristics of being/not being farmer
Aggregate: explain aggregate-level outcomes through
explanatory aggregate-level variables
e.g.: explain the percentage of migrants in a region through the
percentage of people in the population having a certain occupational
status (e.g. being a farmer)
Multilevel: make claims across the levels, from the aggregate-
level to the individual-level and vice-versa
e.g.: explain the individual probability to migrate for non-farmers
through the percentage of farmers in the same region
18
19. Grouping in multilevel
Units grouped at different levels, a-contextual language
Grouping may be more or less random
Once the grouping is done, differentiation:
group and its member influence and are influenced by the
group membership
19
20. Statistical modelling of hierarchies
Yij 0 j 1 j x ij 2 z j ij
response variable at the
individual level
explanatory variable at the individual level
explanatory variable at the group level
i: index for the individuals
j: index for the group
these vary depending on the group
Errors are independent at each level and between levels 20
21. Goldstein:
Multilevel in educational research
Study school effectiveness, examination results, …
All quantifiable aspects of education
‘Statistical’ advantages of multilevel
Efficient estimates of regression coefficient
Correct standard errors, confidence
intervals, significance tests for the clusters
Enables measuring differences between clusters
http://www.math.helsinki.fi/msm/banocoss/Goldstein_course.pdf
21
22. Hierarchies in educational research
Simple hierarchy:
Pupil / class / school / neighbourhood /
Cross-classified structure
Pupil – ethnicity // school – neighbourhood //
22
23. Goldstein et al: examination results and
school differences
Inner London schools
Response variable: examination results
Explanatory variables: standardised London reading tests, verbal
reasoning category, gender, school gender
(mixed, boys, girls), school religious denomination
(State, Church of England, Roman Catholic, other)
Results:
Small effect of school gender; Roman Catholic slightly better; girls
better than boys; large differences for different verbal reasoning
categories.
Differences between schools in examination results depend
on intake achievement and curriculum subject considered
No single dimension in which schools differ
23
24. Driessen: School composition and
primary school achievement
Dutch primary schools
Response variable: language and math proficiency
Explanatory variables: parental ethnicity and
education, pupils sex and age, school composition, ethnic
diversity
Results:
Quite strong effect of school composition on language, weak
on math; all children, independently of
background, perform worse in schools with high ethnic
diversity
Question about distribution policy and other measures
24
25. […] despite their usefulness, models for
multilevel analysis cannot be a universal
panacea. […]
They are notsubstitutes for well
grounded substantive theories […]
Multilevel models are tools to be used with
care and understanding.
Goldstein, Multilevel statistical models,
http://www.bristol.ac.uk/cmm/team/hg/multbook1995.pdf
25
27. Modelling and explaining
What does a multilevel model model?
Relations between different levels in a hierarchical structure
What does a multilevel model explain?
How group behaviour influences individual behaviour
(but not vice-versa)
Statistically, multilevel achieves both
But the ‘substantive theory’ is still wanting
27
28. What needs the ‘substantive theory’
School religious denomination
What social practices, norms, values are involved?
School composition and ethnicity
How do these influence peer relation among pupils?
…
What is the extra information that we need and that
statistics does not give us?
28
30. Levels, beyond statistics
Dan Little
Levels of the social:
Ontology what social entities?
Explanation reduction?
Causation causal powers?
Inquiry what level?
Description what level requirements?
Generalisation recurrence of types?
Avoid analogies with natural sciences,
don’t reify social phenomena
30
31. Levels, beyond the received views
Methodological individualism Holism
• Social facts must be • Social entities and
reducible to facts about structures have primacy and
individuals are independent
• There is no higher level • Individuals are influenced
without lower level by social facts, but do not
influence them
• E.g.: Austrian school
economics, some political • E.g.: sociologists in the
scientists Durkheim tradition
31
32. Methodological localism
Social structures influence social outcomes, embodied in action of
socially constructed individuals
Individuals are the bearers of social structures and causes, but
individual actors are socially constructed
Emphasis:
Contingency of social processes
Mutability of social structures over space and time
Variability of human social systems (norms, social practices, urban
arrangements, …)
Cast doubt on generalisable theories across many populations,
look for specific causal variation
32
33. Scale-based levels
Petri Ylikoski:
Macro social facts are typically supra-individual
Micro and macro have a part-whole relationship, but not just
mereological constitution
Difference in scale, not categorical
A heuristic, as there is no unique micro-level, context-
relativeness
33
34. Different questions
Constitutive questions Causal questions
• How macro properties are
constituted by smaller-scale • Origin, persistence, and
entities change of macro properties
• How the macro depends on • What the outcome would
the micro have been, had things in the
• How the macro would have causal history been
been different, had the different
micro been different
34
36. Levels and types of question
Give reality to levels
allocation is not random, not a statistical artefact
Articulate the embodied aspects of level interactions
Sociology, anthropology, pedagogy, psychology, …
Look for empirical origins of variations in outcome
Large-scale statistical studies
Small-scale qualitative studies
INTEGRATIONof explanation of socially-constituted behaviours
36
38. In this talk:
PhilEd and empirical research
Disputed question of the relation between PhiEd and empirical research
Turn the question on its head:
does empirical research look into philosophy?
Empirical research
Causal modelling widely used in social research, including education
Sophisticated formalisms are designed to measure, model, explain social
reality, including hierarchical structures
Despite progress and improvements, formal methods are still in need of
‘substantive’ theory
In search of a substantive theory
Recent work by Little and Yilikoski addresses level ontology, it helps find
the main features of the substantive theory
38
39. Trouble shared, trouble halved?
Empirical research does not look outside statistics
sufficiently
A problem shared also by e.g. social epidemiology
Statistical modelling, an alleged gold standard to
generate evidence
A problem shared by e.g. evidence-based medicine
39
40. Remedies?
Qualitative research, philosophical investigations to
feed empirical research
Dismantle evidence hierarchies and gold standards
Build integrated methods
Quantitative and qualitative
Empirical and conceptual
Multiple sources of evidence
40
42. Courgeau D. 1994 Du groupeàl’individue: l’exemple des comportementsmigratoires. Population 1.
Courgeau D. 2007 Multilevel synthesis. From the group to the individual. Springer.
Driessen G. 2002 School composition and achievement in primary education: a large-scale multilevel
approach. Studies in Educational Evaluation 28.
Goldstein H. 1999 Multilevel statistical models. Wiley.
Goldstein et al. 1993 A multilevel analysis of school examination results. Oxford Review of Education
19(4).
Hyslop-Margison EJ and AyazNaseem M. 2007 Philosophy of education and the contested nature of
empirical research: a rejoinder to D.C.Phillips. Philosophy of Education.
Little D. 2006 Levels of the social. In The Philosophy of Anthropology and Sociology, Risjord and
Turner (eds). Elsevier Science.
Phillips D.C. 2005 The Contested Nature of Empirical Educational Research (and Why Philosophy of
Education Offers Little Help). Journal of Philosophy of Education 39(4).
Robinson W.S. 1950 Ecological Correlations and the Behavior of Individuals. American Sociological
Review, 15(3)
Russo F. Causality and causal modelling in the social sciences. Measuring variations. Springer. 2009
Ylikoski P. 2012 Micro, macro, and mechanisms. In Oxford Handbook of Philosophy of Social
Sciences, Kinkaid (ed) OUP 42
Editor's Notes
Few contacts between philed and philsci here is a good occasion personal contact but also proximity of topics and interestsPoint of contact = relation between phil (ed) and empirical researchUse multilevel as a case to illustrate the interactions (or lack thereof)
Coming from PhilSciPartly true, partly not. Still a lot of abstract, toy-example based investigationWhen it is based on empresarch, phil is less precise and sharp, but perhaps more useful. A trade-off that is difficult to accept and that not always isLast para:- yes, so there must be interaction, not just phil looking into emp research. This point will come up again.
Two approaches:Qual more details about internal dynamics of some social system. Very limited generalisabilityQuant big data supposedly allow more generalisabilityAlso, supposedly, quant has more rigorous methodology … so here another pandora’s box. Keep this in mind, I’m coming back in the conclusion
Holism: for instant sociology of DurkheimIndividualism: for imstance meth individualism in economics
Don’t necessarily show this slides, may be too technical.
School composition. Within the PRIMA project, a school composition variable was constructed on the basis of the ethnicity and education of the parents.Ethnic diversity. In addition to the school composition variable. the number of different ethnic groups was also included as a predictor variable at the level of the school. This provides an indicator of the heterogeneity of the school with regard to ethnic origin and thereby a measure of the linguistic and cultural diversity characteristic of the school.
Ontology: what are social entities composed of? Explanation: do social explanations need to “reduce” to facts about the actions of individuals?Causation: do “higher-level” social entities have causal powers? Inquiry: at what level should (a given style of) social inquiry focus its efforts at descriptive and explanatory investigation? Description: are there “level” requirements or constraints on social description?Generalization: are there higher-level “types” of social entities that recur in different historical and social settings?
Explain why I don’t agree with- Just macro properties- counterfactual element in causal question