25
MODELLING REPEATED MEASURES ON FAMILY MEMBERS IN GEOGRAPHICAL AREAS IAN PLEWIS UNIVERSITY OF MANCHESTER PRESENTATION TO RESEARCH METHODS FESTIVAL OXFORD, 2 JULY 2008

MODELLING REPEATED MEASURES ON FAMILY MEMBERS IN GEOGRAPHICAL AREAS

  • Upload
    keiki

  • View
    28

  • Download
    0

Embed Size (px)

DESCRIPTION

MODELLING REPEATED MEASURES ON FAMILY MEMBERS IN GEOGRAPHICAL AREAS. IAN PLEWIS UNIVERSITY OF MANCHESTER PRESENTATION TO RESEARCH METHODS FESTIVAL OXFORD, 2 JULY 2008. There are research questions in the social sciences that require descriptions and explanations of variability in one or - PowerPoint PPT Presentation

Citation preview

Page 1: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

MODELLING REPEATED MEASURES ON FAMILY MEMBERS IN GEOGRAPHICAL AREAS

IAN PLEWIS

UNIVERSITY OF MANCHESTER

PRESENTATION TO RESEARCH METHODS FESTIVALOXFORD, 2 JULY 2008

Page 2: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

There are research questions in the social sciences that

require descriptions and explanations of variability in one or

more outcomes within and between families.

Page 3: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Some of these questions can be addressed by partitioning and modelling:

i. variation within individuals (when we have repeated measures);ii. variation between individuals within families (with the

appropriate design);iii. variation between families within areas/neighbourhoods,

schools, cross-classifications etc. (with a clustered or spatial design).

Page 4: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

The Millennium Cohort Study (MCS), for example, meets these three criteria because, as well as being longitudinal, data arecollected from the main respondent (mothers), their partner (if inthe household), the cohort child (or children if multiple birth) andolder sibs. Also, MCS was originally clustered by ward (althoughresidential mobility reduces the spatial clustering over time).

Page 5: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

So we can model ytijk, assumed continuous, where:

t = 1…Tijk are measurement occasions (level 1);

i = 1…Ijk are individual family members (level 2);

j = 1..Jk are families (level 3);k = 1..K are neighbourhoods (level 4).

This is the usual nested or hierarchical structure.

Page 6: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

For variables that change with age or time, educational

attainment for example, we can use a polynomial growth

curve formulation:

tijkeq

tijka

Q

qqijkb

tijky

0

Page 7: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Do growth rates vary systematically with individual, family and neighbourhood characteristics?

We model variation in the random effects bqijk in terms of variables at:

i. the individual level, e.g. gender,

ii. the family level, e.g. family income,

iii. the neighbourhood level, e.g. Index of Multiple Deprivation.

Page 8: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

We might prefer to model the variation in the time-varying outcome in

terms of one or more time-varying explanatory variables. For example,

we might be interested in how individuals’ smoking behaviour varies as

their income changes.

)1(0

Qusuallytijkeq

tijkx

Q

qqijkc

tijky

This model raises the tricky issue of endogeneity of individual (uijk) and family effects (v0jk) generated by unobserved heterogeneity at these two levels that is correlated with x, especially if we are interested in the causal effect of income on behaviour.

Page 9: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

The growth curve and conditional models are compared in, for example:

Plewis, Multivariate Behavioral Research, 2001.

Page 10: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

These two approaches both ignore the fact that family members havelabels: mother, father, oldest sibling etc.

Often, we would like to know how the behaviour and characteristics ofone family member are related to those of other family members.

a. The influences of parents on children and children on parents: mental health and behaviour.

b. The influence of one parent (or, more generally, partner) on another: quitting (or starting) smoking.

c. The influence of one sibling on others: risky behaviour.d. The association of parents’/partners’ characteristics on each other’s

behaviour: educational qualifications and health behaviours.

Page 11: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

In these cases, a multivariate approach (within a multilevel framework) can be more informative as Raudenbush, Brennan and Barnett, J. Family Psych. (1995) first pointed out.

Page 12: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Their model is based on repeated measures of men and women as members of intact

couples:

1,0,)(0

)()(3

)()(2

)()(1

)(0

)(

1,0,)(0

)()(3

)()(2

)()(1

)(0

)(

)()(1

)(0

)(

)()(1

)(0

)(

kFk

uHizF

kbF

izF

kbM

izF

kbF

kbF

kib

kMk

uHizM

kbF

izM

kbM

izM

kbM

kbM

kib

Ftie

itF

ibF

ibF

tiy

Mtie

itM

ibM

ibM

tiy

Correlations between males and females within individuals - r(eMeF) - and at the family level - r(u0u1) - can be estimated.

The equalities of within and between variances for men and women can be tested.

Time varying variables can be introduced if appropriate.

Page 13: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

This basic model can be extended to:

1. Situations where ‘growth’ isn’t obviously applicable, as might be the case for binary and categorical variables (mover-stayer or mixture models).

2. Couples plus children.

3. Changing household structures (two parents to one parent for example).

Page 14: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Suppose we have repeated measures of whether (and how

much) mothers, fathers and their adolescent children

smoke and that we are interested in influences across

family members over time, and of educational qualifications

(and area of residence) on smoking behaviour.

Page 15: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

M, t-1

M,t

F,t-1

F,t

M,Ed.

A, t

A,t-1

F,Ed.

M,smok

F,smok

Page 16: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

t = 2..Ti; min(Ti = 3); neighbourhood level omitted for ease of exposition.

)()(.

)(04

)(.

)(03

)()(02

)()(01

)(00

)(0

)()(,1

)(1

)(0

)(

)()()(03

)()(02

)()(01

)(00

)(0

)()(,1

)(2

)(,1

)(1

)(0

)(

)()()(03

)()(02

)()(01

)(00

)(0

)()(,1

)(2

)(,1

)(1

)(0

)(

AiuF

iyAaM

iyAaF

izAaM

izAaAaA

ia

AtieA

ityAaA

iaA

tiy

FiuH

izFaM

izFaF

izFaFaF

ia

FtieM

ityFaF

ityFaF

iaF

tiy

MiuH

izMaF

izMaM

izMaMaM

ia

MtieF

ityMaM

ityMaM

iaM

tiy

Page 17: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

The model allows for correlated random effects at the

individual level (ui for M, F and A) and also correlated

residuals at each occasion (eti also for M, F and A),

multivariate Normal in each case.

Page 18: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Estimation and specification issues:

1. Simultaneity – y2 is predicted by y1 but y2 is a predictor of y3 etc. so we need a FIML estimation method. Not a problem with continuous x and y but perhaps more problematic for non-linear models. MCMC needed?

2. Autocorrelation structure for level one residuals? E(etet-

1) < 0?

Page 19: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

3. Endogeneity of family residuals (ui) – could use y1 as a control for this but then min(Ti) =4.

4. Missing data – essentially complete case analysis without imputation.

5. Assumption of Normality for random effects. Clark and Etilé, J. Health Economics, 2006 use a similar model but estimate using a modified EM procedure with non-parametric individual random effects.

Page 20: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Estimates from bivariate probit model: MCS waves 1 and 2,intact couples

MOTHER PARTNER

Educ. Quals., mother -0.089 (0.017) -0.059 (0.017)

Educ. Quals., partner -0.051 (0.017) -0.068 (0.016)

Smoking, mother, S12.6

(0.073)0.41 (0.083)

Smoking, partner, S1 0.45 (0.062)2.6

(0.054)

Smoking interaction, S1 -0.41 (0.10) -0.30 (0.10)

Within h/h correlation 0.21

Page 21: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Effects of gaining, losing, changing partner – previousresearch suggests that there are some for smoking but papers have not linked one family member to another.

Focus here on women because men generally notfollowed up in cohort studies. Studies like BHPS might be more informative.

Page 22: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Changes in household structure, MCS, waves 1 to 2 and 1 to 3.

Waves 1 to 2 (%) Waves 1 to 3 (%)

Same couple 77 74

Same single parent 11 10

Change partner 1.3 1.8

Acquire partner 6.2 9.1

Lose partner 4.5 5.1

Page 23: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

Can accumulate data over time intervals.Can specify more complex models for women who change partners.Expect interactions between δ and other explanatory variables.Single level but can be repeated for women gaining/losing partners more than once.

gainedpartnerforcouplesstablefori

bFixbM

ixbM

iybF

iybbF

iy

lostpartnerforcouplesstableforieaF

ixaM

ixaM

iyaF

iyaaF

iy

1;025

)(4

)(3

)(22

)(110

)(2

(B)

1;01

15)(

4)(

3)(

12)(

110)(

2

(A)

2

Page 24: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

CONCLUDING REMARKS

1. Multilevel modeling is a powerful technique for disentangling sources of variation.

2. Using bivariate (or multivariate) multilevel models for repeated measures data extends the range of hypotheses about within household influences on behaviour that can be tested.

3. It is important to model what actually happens to households over time and not always to rely on analyses of intact couples.

4. What hypotheses might we consider for neighbourhood effects?

Page 25: MODELLING REPEATED MEASURES ON  FAMILY MEMBERS IN GEOGRAPHICAL AREAS

POSSIBLE EXTENSIONS

1. The problem of limited dependent variables, especially for measures of amount smoked which is censored at zero. Multilevel extensions of Tobit, mixture or Heckman selection models might be useful here.

2. Combining growth curve models with models for non-smokers. For example, Carlin et al. (Biostatistics, 2001) suggest a mixture model.