Centre for Market and Public Organisation

1
Centre for Market and Public Organisation
Measuring socio-economic position in ALSPAC Liz
Washbrook, CMPO Liz.Washbrook_at_bristol.ac.uk ESRC/
ALSPAC Large Grant Meeting 5th November 2008
2
But first! US cohort studies

• Early Child Longitudinal Study Birth Cohort
  (ECLS-B)
• 10,000 children born 2001, nationally
  representative when weighted
• Over-samples of low birth weight babies, twins,
  some ethnic groups (e.g. Native Americans,
  Chinese)
• Samples from birth certificates, follow-ups at 9
  months, 2 years, Fall prior to kindergarten
  (4y), Fall of kindergarten year (5y). But no
  more!
• Data from parent CAPI, direct child assessments,
  child care providers and teachers. Some resident
  and non-resident father questionnaires.
• Early Child Longitudinal Study Kindergarten
  Cohort (ECLS-K)
• 20,000 children starting kindergarten in 1998 (b.
  1992/3)
• Children sampled from 1277 schools in 100
  counties. Target 24 children per school.
  Nationally representative when weighted.
• Follow ups at Fall Spring kindergarten year
  (5-6y), Fall Spring 1st grade (6-7y), Spring
  3rd grade (9y), 5th grade (11y), 8th grade
  (14y)
• Data from direct child assessments, parental
  phone interviews, teacher and school
  administrator questionnaires.
• Data is publicly available (on CD). See
  http//nces.ed.gov/ECLS/index.asp

3
US cohort studies

• Fragile Families
• 5000 children born 1998-2000 in large US cities
• Designed to follow children born to unmarried
  parents but includes control sample of married
  parent families (25).
• Focus on deprived families 44 mothers at
  baseline black, 35 Hispanic, 27 teenagers, 79
  high school or less
• Detailed information on fathers roles and
  involvement
• Parent interviews in hospital at birth, follow
  ups at 1, 3, 5 and 9. Includes direct in-home
  child assessments.
• Data publicly available www.fragilefamilies.princ
  eton.edu/index.asp

4
Aims

• Aim to stimulate discussion about the
  construction of an index of parental
  socio-economic position (SEP) from the ALSPAC
  data
• Talk will cover
• The range of indicators available and their
  features
• Sample selection/missingness issues (multiple
  imputation)
• Combining the indictors into a single index
  (principal components analysis)
• Illustrated using a case study Measures of
  social inequality in Key Stage 2 exam results
  (age 11)
• Would a standard SEP variable available to all
  ALSPAC researchers be useful?
• If so, how should it be constructed?
• Input, feedback, discussion would be appreciated!

5
What is SEP?

• Extensive literature on theories of social
  stratification (Galobardes, Lynch and Davey
  Smith, 2007 Bradley and Corwyn, 2002).
• Socially derived economic factors that influence
  what positions individuals or groups hold within
  the multiple-stratified structure of society
  (Galobardes et al)
• In practice researchers have used a multitude of
  individual indicators to measure SEP, each of
  which captures a different aspect of
  stratification
• Composite SEP is a relative measure, whereas some
  indicators (income, education) measure absolute
  levels of resources. This may have implications
  when thinking about policy.

6
Why measure parental SEP?

• SEP as a summary measure of family background
  that defines sub-groups of the population.
• Social mobility/life chances
• Nature vs. nurture
• Example Joint CMPO project on the role of
  attitudes and aspirations in explaining the
  educational deficits of children in poverty
• SEP as a way of capturing long-term access to
  resources over the life course, e.g. permanent
  income in economics
• To classify deprived or vulnerable groups in a
  way that captures the idea of multiple risks
• As a control for confounding influences (e.g.
  studying the effects of smoking)? Disaggregated
  sets of control variables may be more appropriate

7
SEP indicators in ALSPAC

• How is the indicator constructed from multiple
  pieces of information? (High frequency of
  measurement in ALSPAC)
• How is the indicator distributed? (E.g.
  discrete/continuous)
• For whom is it available? (Differential
  missingness)
• How well does it distinguish between high- and
  low-performing children? (KS2 is an example
  relationships will differ with different
  outcomes)

8
The sample

• 11 071 children with
• A valid Key Stage 2 score
• Minimum of 2 (out of 10) non-missing SEP
  indicators (30 complete cases)
• Sample is 69 of the eligible birth cohort (15
  994 in NPD)
• Key Stage 2 score derived from exam marks in
  English, maths and science in Year 6 (age 11).
  National tests compulsory in all state schools.
• Test scores are averaged and normalised to mean
  zero, standard deviation 1 on the full eligible
  population of 15 994
• The working sample is not randomly selected
• Mean KS2 (SD)
• Working sample (N11071) 0.11 (0.95)
• lt2 SEP indicators (N4923) -0.26 (1.05)

9
Household income
Measures Take home weekly family income at 33,
47, 85, 97 months 11 years
Proportion of valid responses in bands
Failure to update the bands means that the
usefulness of the 85 and 97 month income measures
is limited.
10
Household income
The age 11 income measure is better
11
Household income

• The SEP index uses
• Log average real equivalised weekly take home
  income at 33 47 mths
• Median income for band imputed using FES data for
  households containing a child of the cohort
  members age, in the relevant year and income
  interval
• Adjustment made for housing benefit income if
  respondent reports zero housing expenditures and
  lives in rented accommodation (predicted value
  from FES for HB recipients in the Southwest,
  varying with year, lone parent status and number
  under 16s in household)
• Expressed in 1995 prices using All Items RPI
• Equivalised using modified OECD scale
• Averaged and logged
• Nominal banded income at 85 months
• Nominal continuous income at 11 years, using band
  midpoints

12
Average KS2, by preschool income quintiles
13
Average KS2, by nominal income at age 7
14
Average KS2, by nominal income quintiles at age 11
15
Parental education

• Measures Mother and partner reports for both
  spouses qualifications antenatal, 61 and 97
  months.
• The SEP index uses maternal reports of own and
  partners highest qualification at 32 weeks
  gestation.
• Issues
• Non-response to the question is coded as no
  qualifications (dont know, no quals and no
  partner were all possible responses)
• Possible discrepancies between own and partner
  report
• Possible changes in the identity of the partner
  over time
• Possible changes in qualifications over time

16
Average KS2, by mothers highest qualification
17
Average KS2, by partners highest qualification
18
Parental social class

• Measures Mother reports of own and partners
  occupation antenatal, 8 and 97 months. Partner
  reports more frequent but not coded.
• The SEP index uses maternal reports of own and
  partners social class at 32 weeks gestation.
• Question related to occupation in current or last
  job
• Occupations coded according to 1991 SOC
  classification
• Used to derive Registrar Generals Social Class
  this is what is available in the datafiles.
  Hierarchical measure.
• No other data on occupation is currently coded

19
Average KS2, by mothers social class
20
Average KS2, by partners social class
21
Housing tenure

• Measures Mother reports of tenure 8, 21, 33 and
  61 months.
• The SEP index uses a derived variable
• Always owner-occupier mortgaged/owned
  outright/buying from council at all 4 dates
• Ever in social housing council rented/Housing
  Association rented at any of 4 dates
• Other not otherwise classified and at least
  one valid response (other responses private
  rented furnished/unfurnished, other). Includes
  all people with a missing value who were never
  observed in social housing, as well as renters.

22
Average KS2, by housing tenure 8-61 months
23
Local deprivation/affluence

• Measures Ward-level Index of Multiple
  Deprivation (IMD) currently matched at birth, age
  5 and age 8, but postcodes available on an annual
  basis
• The SEP index uses the (continuous) rank of the
  IMD for ward at birth
• IMD provided by government statistics. Derived
  from data in 6 domains income, education,
  employment, housing, health, access to services
• Wards in England (approx. 5500 individuals)
  ranked on basis of deprivation from 1 to 8414.
  This allows definition of national quantiles.
• Can be matched to ALSPAC via postcode data

24
Average KS2, by national quintiles of IMD
25
Subjective financial hardship
Measures Mother-completed financial difficulties
questionnaires at 8, 21, 33, 61 and 85 months
Format How difficult at the moment do you find
to afford these items food clothing heating
rent/mortgage things for child? Very (3)
Fairly (2) Slightly (1) Not difficult
(0) Responses to the 5 items at each date summed
to give to score between 0 and 15

• The SEP index uses the mean score across the 5
  dates
• The 61 and 85 month measures include questions on
  educational courses, medical care, child care and
  other things
• Do not pay for this/DSS pays options for rent
  and heating coded as 0
• The distribution of the resulting variable in
  highly skewed

26
Average KS2, by quintiles of financial
difficulties score
27
Multiple Imputation by Chained Regression
SEP indicators missing (out of 10)

• Iterative multivariable regression technique
  switching regression
• Statas ice command
• Specify a prediction equation for each variable
• Randomly allocate values to missing cases
• Predict values for missing cases
• Update RHS variables and repeat cycle (10 times)
• Options allow choice of estimation method,
  passive imputation and substitution of RHS
  dummies, constrained intervals for predicted
  values

Current method Imputation carried out using 10
SEP variables only does not use other
information Only a single imputed dataset created
28
The ice command
29
Prediction equations
30
Principal components analysis

• PCA provides a way of combining (weighting) the
  individual components into a single index
• PCA conducted on the 10x10 polychoric correlation
  matrix
• Standard PCA techniques assume continuous,
  normally distributed variables.
• Polychoric correlation can be used when there are
  binary and categorical components (e.g.
  education).
• It assumes that ordinal variables obtained by
  categorizing an normally distributed underlying
  variable.
• PCA extracts a single component that maximises
  the explained proportion of the variation in the
  (standardised) components
• Each component is assigned a scoring coefficient
  that is used as a weight in the construction of
  the SEP index

31
Principal components analysis
Scoring coefficients
SEP index explains 46 of total variation in
components
32
Average KS2, by quintiles of SEP index
33
Summary

• ALSPAC contains numerous indicators that can be
  used to construct an SEP index
• Indicators vary in
• The type of resources they measure
• The sections of the population they distinguish
  (e.g. tenure appears good at picking out the very
  disadvantaged, but does not discriminate at the
  top of the distribution)
• The likelihood of non-response by different
  groups
• Issues that need to be considered when
  constructing an index
• Which components should be included? (Should
  education be separate?)
• How should observations at multiple dates/by
  multiple respondents be treated?
• How should missing values be dealt with?
• How should the components be combined?