sociological classifications and simulation models of social inequality [9.9.2010]
DESCRIPTION
Sociological classifications and simulation models of social inequality [9.9.2010]. Paul Lambert, Mark Birkin and Guy Warner Social Stratification Research Seminar, Utrecht, 8-10 September 2010. Sociological classifications and simulation models of social inequality. NeISS - PowerPoint PPT PresentationTRANSCRIPT
Sociological classifications and simulation models of social inequality
[9.9.2010]
Paul Lambert, Mark Birkin and Guy WarnerSocial Stratification Research Seminar,
Utrecht, 8-10 September 2010
1Lambert, Birkin, Warner: SSRC, Sep 2010
Sociological classifications and simulation models of social inequality1) NeISS 2) Simulation models as longitudinal methods3) ‘Ageing and inequality’ project: Social inequalities
modelled as responses to changing socio-economic / socio-demographic structure
4) BHPS-based transition probabilities5) First evidence on the effects of different sociological
classifications
2Lambert, Birkin, Warner: SSRC, Sep 2010
1) www.neiss.org.uk
A JISC initiative (2009-12) on collaborative research infrastructure in the UK
National e-Infrastructure for Social Simulation
• Expert led simulation demonstrations
• Combining data resources• Workflows for the simulation
analysis Modify and re-specify existing
simulation templates
See Birkin et al. (2010) (includes image source)
3
Birkin et al. (2010: 3808)4Lambert, Birkin, Warner: SSRC, Sep 2010
5
Contributions of the ‘NeISS’ project– Accessing live / newly updated
socio-economic/demographic data– Running/supporting complex simulation models with high
computational requirements– Allowing flexible data management (e.g. in defining
alternative measures of social position, education)– Allowing multiple specifications of related models for
comparisons (e.g. varying a few parameters and re-running)
…this application aims to make new inputs to long-standing questions about the influence of different measures of the stratification structure…
Lambert, Birkin, Warner: SSRC, Sep 2010
2) Social simulation models as longitudinal methods
– Definitions of longitudinal social research often focus on data collected at or about multiple points in time
– Simulation models are often (but not necessarily) based on longitudinal data collections
– The do intrinsically involve: • data simulated (i.e. constructed) forward in time• analysis to summarize pogression through time
E.g. Gilbert and Troitzsch (2005); Gilbert (2008); Zaidi et al. (2009)6Lambert, Birkin, Warner: SSRC, Sep 2010
The contribution of simulation• A simple, and ordinarily daft, simulation is to extrapolate forward in time based
on the perpetuation of current parameters (e.g. using aggregate data)
Plagiarism cases, year 1 essays
020
040
060
080
010
00
1995 2000 2005 2010 2015 2020 2025time
ObservedProjected
Aggregate data projection (for dafties)
7
• These models show different trends if we assume that patterns respond to policy and payoff [ f(O,P) ] and opportunity [ f(t) ]
-20
0-1
00
010
020
030
0
1995 2000 2005 2010 2015 2020 2025time
ObservedProjectedError est.
Banded projected
Simulation model with payoff and policy effects on 2008 projections
- Simulation models find ways to respond to the evolving population structures and constraints, typically in a ‘non-linear’ way- Good for seeing a likely pattern, but weak in terms of realistic margins of error
8
010
020
030
0
1995 2000 2005 2010 2015 2020 2025time
ObservedProjectedError est.
Banded projected
Simulation model using year-on-year time, payoff and policy effects
Simulation approachesThe above are aggregate models bringing in effects from
the contextual average– Graph 1: Projected value = f(time)– Graphs 2 and 3: Projected value = f(time)*f(current proportion)
Modern simulations tend to be either:
• Micro-simulation– {Year-on-year} predicted values for every subject,
carried forward via transition probabilities
• Agent based modelling – {Year-on-year} predicted values for particular
subjects (agents), modelling forces and interactions experienced by the agent 9
The contribution of simulation• The general contribution is to model forwards in
order to see plausible patterns within complex/responsive systems– Needs a good model of influences, projected
influences, contextual effects(serious models take a lot of work – e.g. Euromod; SAGE)
– We ordinarily try various inputs to the system (e.g. what would happen if we did X)
• …data choices (between measurement options) could be very important here… 10
Might social classifications matter in longitudinal simulations?
• Things might be pretty robust regardless of measures
• Different measures tend to correlate with age, gender, and change over time
• Major differences in functional form could be consequential, cf. – Crossing a threshold in a {two}-category model– A continuous model without any thresholds
11Lambert, Birkin, Warner: SSRC, Sep 2010
A selection of possible measures of social inequality using BHPS 1991-2007
1991 Mean 2007 R2 with year2
Intergenerational correlation (CAMSIS) (all adults) 28 25 24 83
Intergenerational correlation (CAMSIS) (men only)
31 28 28 77
Husband-wife homogamy (CAMSIS r) 39 37 33 70
Personal income Gini (all adults) 45 44 43 87
Personal income Gini (men only) 40 40 40 46
Household income Gini (all adults) 38 38 37 55
Occupational Gini (CAMSIS) (all adults) 29 28 27 91
Percent of all adults in EGP I 14 16 18 73
Percent of all adults in EGP I or II 33 37 43 93
Percent of all adults in RGSC I or II 32 37 42 96
Source: BHPS cross-sectional aggregates, weighted using {w}xrwght. 12Lambert, Birkin, Warner: SSRC, Sep 2010
.25
.3.3
5.4
Fitt
ed v
alu
es
0 5 10 15 20 25 30 35wave
Intergenerational correlation
Household inc. gini(projected)(projected)
BHPS - Forward projections using current parameters
13Lambert, Birkin, Warner: SSRC, Sep 2010
Variations in deterministic parameters • Here we’ll include in models the influences of
educational level and family type (and artificially adjust educational qualifications’ prevalence by age cohort)
0 20,000 40,000 60,000 80,000 100000
4. Single parent
3. Couple + children
2. Multi-adult
1. Single adult
..Many more variations of these and other measures are possible, for future consideration…
14
educ2b 209785 2 .2787092 0 1 Low school level or beloweduc2a 209785 2 .1169197 0 1 Degree education or aboveeduc4 209785 4 2.808118 1 4 BHPS 4-fold educational level classificationzqfedhi 209785 12 6.957461 1 12 highest educational qualificationisced 210571 8 20.72271 12 32 Variable Obs Unique Mean Min Max Label
3) Ageing and inequality• Sociological and econometric research agendas studying the
circumstances of social inequality who is advantaged/disadvantaged; why is that?
• We increasingly acknowledge the potential influences of demographic transitions/socio-economic transformations
ageing population; changing family structures;educational expansion; immigrant influxes
• Ample longitudinal survey data resources e.g. BHPS; GHS; LFS; ‘Slow Degrees’ dataset
• Many previous simulation analyses compare the effects of social changes on social inequalities (e.g. Zaida et al. 2009),– to our knowledge, there is little attention paid to, or sensitivity
analysis of, measures of social structure and inequality other than income - such as of occupations, educational levels
16
‘…the interaction between ageing effects and [the] nature and impact of socio-economic inequalities..’
The educ profile represents grade inflation. The income/occ profiles could be one or two things - changing rewards with age; plus or not a general upgrading of rewards across birth cohorts
-1-.
50
.5F
itted
va
lues
20 30 40 50 60 70age at date of interview
Personal income Occupational advantageEducational advantage
BHPS wave 17, unweighted, Males in work aged 16-70. N=3590. R2=0.09 for income, 0.03 for occupation and education.
Lambert, Birkin, Warner: SSRC, Sep 2010
0.1
.2.3
.4F
itted
va
lues
1990 1995 2000 2005 2010year
RGSC % advantaged EGP % advantaged
RGSC gender difference EGP gender difference
Source: BHPS waves 1-18, x-sectional respondent weighting, young adults 16-40 only
Gender trends 16-40yrs, social advantage in 2-category EGP and RGSC
17Lambert, Birkin, Warner: SSRC, Sep 2010
18
‘…the interaction between ageing effects and [the] nature and impact of socio-economic inequalities..’ [ctd]
It proves very difficult to separate the experiences of age cohorts from other time trends (gender; industry; administration)
Unskilled
Skilled manual
Petty-bourg.
Non-manual
Salariat
Source: Females from LFS/GHS, using data from Li and Heath (2008)
percent of year category
Goldthorpe class scheme harmonised over time
Lambert, Birkin, Warner: SSRC, Sep 2010
19
Methodological topics
– Comparison between analyses which use different measures of position within the inequality structure
e.g. occupations; education; income; wealth
– Model of the feedback effects on those positions of trends in national and local distributions
variously measured– Modelling of the feedback effects of trends in
national and local demographics (e.g. family structures; immigration)
variously measured
20
How to proceed? • Specify a simulation of social inequality outcomes using
demographic, economic and geographic indicatorsEstablishes a predicted profile, which is described over timeDeterministic and stochastic components in predicting values (see
O’Donoghue et al, 2009)
• Vary the model according to: – Alternative measures of social inequality– Alternative measures or projections on economic and industrial trends– Alternative measures or projections on demographic trends– Allowing locality variations
21
0.1
.2.3
.4.5
20-40 40-60 60-70
BHPS w17, Males in employment aged 20-70Model 1: Income=0.39Educ + 0.1Age -0.001Age2 + 0.13Age*EducModel 2: Income=0.17Educ + 0.1Age -0.001Age2 + 0.26Age*Educ
Mod1 Mod2 Age group mean
E.g.: This shows projected mean incomes as function of education, with less and more uni. expansion over time
4) BHPS based transition probabilities
In general… – Use a resource like BHPS to calculate year-on-year transition
probabilities from one situation to another – These probability calculations can often be enhanced by
other supplementary micro-data, e.g. on transitions between rarer states (see Zaida et al. 2009b)
– Those probabilities are then applied successively to a baseline dataset, projected forward over time, and that data is then summarised (the simulation)
(running it on an actual dataset reduces the chances of parameters taking the predicted values outside a plausible range)
22Lambert, Birkin, Warner: SSRC, Sep 2010
In the following application…
– BHPS balanced panel • (carry forward all 2007 respondents every year till 2025)
– Predict next year’s outcome from predicted values of a regression with explanatory variables of the current outcome (observed or simulated), plus gender, age, dob, educational level, family type, and age*educ interaction
– Numerous shortcuts: global imputation for family type; ignoring spouse’s changes; …
– Variable parameters summarized below: • 4 different education measures • 4 different treatments (increasing educ for later cohorts only)
Lambert, Birkin, Warner: SSRC, Sep 2010 23
5) First evidence on the effects of different sociological classifications
24Lambert, Birkin, Warner: SSRC, Sep 2010
.25
.3.3
5.4
.45
1990 2000 2010 2020 2030
Observed
Educ4 Educ4++Educ2a Educ2a++Educ2b Educ2b++
ISCED ISCED++
Source: Simulations on the BHPS balanced panel subsample (weighted by xlwght). Data from 1991-2007; simulations from 2008-2025. ++ lines show impact of educational upgradings.
Father-Child CAMSIS r, all adultsIntergenerational correlation simulations
.3.3
5.4
.45
1990 2000 2010 2020 2030
Observed
Educ4 Educ4++Educ2a Educ2a++Educ2b Educ2b++
ISCED ISCED++
Source: Simulations on the BHPS balanced panel subsample (weighted by xlwght). Data from 1991-2007; simulations from 2008-2025. ++ lines show impact of educational upgradings.
Father-Child CAMSIS r, males onlyIntergenerational correlation simulations
Lambert, Birkin, Warner: SSRC, Sep 2010 25
.1.2
.3.4
.5
1990 2000 2010 2020 2030
ObservedEduc4 Educ4++
Educ2a Educ2a++Educ2b Educ2b++ISCED ISCED++
Source: Simulations on the BHPS balanced panel subsample (weighted by xlwght). Data from 1991-2007; simulations from 2008-2025. ++ lines show impact of educational upgradings.
Husband-Wife CAMSIS r, all adultsPartner homogamy
Gini calculations on income and occupations: so far the regression model generating the simulated values isn’t doing a good job of summarising inequality as it tends to reduce inequalities
Lambert, Birkin, Warner: SSRC, Sep 2010 26
.1.1
5.2
.25
.3
1990 2000 2010 2020 2030
ObservedEduc4 Educ4++
Educ2a Educ2a++Educ2b Educ2b++ISCED ISCED++
Source: Simulations on the BHPS balanced panel subsample (weighted by xlwght). Data from 1991-2007; simulations from 2008-2025. ++ lines show impact of educational upgradings.
Gini for current job (CAMSIS), all adultsOccupational inequality
• When greater ‘stochastic’ dependence is used, however, the variable operationalisation impacts diminish
Lambert, Birkin, Warner: SSRC, Sep 2010 27
.36
.38
.4.4
2.4
4
1990 2000 2010 2020 2030
ObservedEduc4 Educ4++
Educ2a Educ2a++Educ2b Educ2b++ISCED ISCED++
Source: Simulations on the BHPS balanced panel subsample (weighted by xlwght). Data from 1991-2007; simulations from 2008-2025. ++ lines show impact of educational upgradings.
Gini for all-sources personal income, all adultsPersonal income inequality
6) Conclusions• Simulations and social classifications
• Simulations offer a tool for evaluating classifications (haven’t previously been used for this?)
• Classification permutations offer new alternatives to the simulation communities
• {NeISS role in infrastructural support}
Preliminary findings suggest: • Measures are important – differences between
measures matter a lot, and they matter more than do differences between treatments!
• Gaps open up: Longer-term longitudinal trends susceptible to differences in measures
28Lambert, Birkin, Warner: SSRC, Sep 2010
References• Birkin, M., Procter, R., Allan, R., Bechhofer, S., Buchan, I., Goble, C., et al.
(2010). The elements of a computational infrastructure for social simulation. Philosophical Transactions of the Royal Society, Series A, 368(1925), 3797-3812.
• Gilbert, G. N. (2008). Agent-Based Models. Thousand Oaks, Ca.: Sage.• Gilbert, G. N., & Troitzsch, K. G. (2005). Simulation for the Social Scientist, 2nd
Edition. Maidenhead, Berkshire: Open University Press.• O’Donoghue, C., Leach, R.H., & Hynes, S. (2009) “Simulating earnings in
dynamic microsimulation models”, in Zaidi, A., Harding, A., & Williamson, P. (Eds.). (2009a). New Frontiers in Microsimulation Modelling. Farnham: Ashgate, pp381-412.
• Zaidi, A., Evandrou, M., Falkingham, J., Johnson, P., & Scott, A. (2009b) “Employment transitions and earnings dynnamics in the SAGE model”, in Zaidi, A., Harding, A., & Williamson, P. (Eds.). (2009a). New Frontiers in Microsimulation Modelling. Farnham: Ashgate, pp351-379.
• Zaidi, A., Harding, A., & Williamson, P. (Eds.). (2009a). New Frontiers in Microsimulation Modelling. Farnham: Ashgate.
29Lambert, Birkin, Warner: SSRC, Sep 2010
• Simulation models can be used to project over time in order to estimate emergent social-structural patterns. The NeISS project (National e-Infrastructure for Social Simulation, www.neiss.org.uk) is a UK initiative in supporting the construction, estimation and interpretation of social simulation models applied to a variety of scenarios. In this paper, I will present results from one of the exemplar projects within NeISS, an analysis of ‘ageing and inequality’, which is designed to model the development of social inequality over time in response to trends in major socio-demographic and socio-economic changes (such as the aging population, changing family formation patterns, changing patterns in educational provision, and changing occupational/industrial opportunity structures). Social inequality indicators used include measures of income inequality, occupational inequality, and social mobility. The data is initially parameterised around annual transition patterns in contemporary Britain, though it should in principle be generalisable to other data scenarios. A unique contribution of the NeISS project is its capacity to support multiple replications of simulations using different underlying measurement instruments of the same concepts – in this paper, we explore the impact of different approaches to measuring occupational circumstances, educational attainment and ethnicity in the context of the simulation model.
30Lambert, Birkin, Warner: SSRC, Sep 2010
31
…from the NeISS application…• “The key substantive question concerns the interaction
between ageing effects and [the] nature and impact of socio-economic inequalities. These issues involve complex, non-linear processes that are suited to simulation approaches. The exemplar will enable study of the impact of alternative socio-economic measures and resources within a micro-level simulation analysis of socio-economic inequalities across age groups, premised upon large scale social survey data (such as British Household Panel Survey, Labour Force Survey, General Household Survey and UK Census based data)” (WP 4.1.4)
Lambert, Birkin, Warner: SSRC, Sep 2010
32
Some specific research questions• How age-qualifications links impact trends in social inequality
– Mass education; admissions policies; cognitive effects • How will (high/low qualified) cohort-specific immigrant
influxes impact upon regional age-occupation-qualification distributions– Simulation: increase or decrease proportions within birth cohorts/ethnic
groups/regions/sectors with certain qualifications • How will fine-grained industrial sector transformations impact
different age cohorts and subsequent stratification positions (e.g. rise of the ‘cultural industries’)– Simulation: Modify national and/or local industrial distributions and project
forward over time• How is long term wealth accumulation influenced by longer
life expectancies (e.g. changing inheritance patterns; longer pension dependence) – Simulation: Model and modify income through work and through inheritance
as it influences relative social position at a national level (e.g. BHPS)
Lambert, Birkin, Warner: SSRC, Sep 2010