More information about text formats
We read with interest the recent
editorial setting an agenda for future research considering neighbourhood
influences on health. Whilst agreeing with many of the points raised in the editorial we take issue with the proposed use
of multiple membership multilevel models to explore contextual effects at
different points in time.
A recent paper used multiple
membership models to analyse the effect of area of residence observed over 9
years on individuals’ health (measured just at the final time point). An accompanying editorial pointed out the problems with
the use of such models. Since we agree that understanding the longitudinal
influence of area of residence over time will form a key part of the
neighbourhood research agenda we expand on the reasons that make multiple
membership models unsuitable for important research questions pertinent to the
field and propose an alternative model.
If person P1 lived in area A1
at time T1 and in area A2 at time T2 then from a life course
perspective we might expect there to be a contribution of both areas to that
person’s measured or reported health at time T2. Similarly, if person P2 moved in the opposite direction then we would expect
a contribution from both areas. The multiple membership model,
however, makes two simplifying and unrealistic assumptions. Firstly, the effect
of each area is assumed to be the same at both times, having the same effect on
person P1 at time T1 as on person P2 at time T2. This means assuming that
the area effect is not dependent on the period; the regeneration or decay of
areas is ignored. Secondly, the multiple membership model
requires explicit weights to be attached to the contribution of each area to an
individual’s health. Suggesting that this should be done on the basis of the
time spent in each area is to assume that the effect of an area is constant
irrespective of the stage of the life course – in other words, such an analysis
assumes that the risk associated with neighbourhoods accumulates at a steady
rate throughout the life course. So if persons P1 and P2
spent an equal amount of time in each area, the direction of movement (from A1 to A2 or from A2
to A1) is ignored.
Alternative models – such as a critical period model – cannot be investigated
under such a framework. Yet the influence of the social environment along the
life course has been shown to differ for specific diseases.
The authors propose extending the
above model to one where health, as well of area of residence, is measured at
both time points. In addition to the assumptions noted above, the
suggestion that the area of residence at time T2 affects health measured at time T1 clearly departs from the usual model of causation.
A more suitable model for such data
is the cross-classified multilevel model. Figure 1 illustrates the
classification diagram for individuals (at level 1) whose health is measured at
time T2 and who live in a
cross-classification of areas at times T1
and T2. Every individual
lives in an area at each time point, whether these are different areas (P1 and P2) or the same area (P3 and P4).
When health is measured at T1
individuals are nested within areas in a strict hierarchy; when health is
measured at T3 there is a
three-way cross-classification of areas at T1,
T2 and T3. The cross-classified
model does not make the assumption that area effects will be the same at the
two time points; in fact, the assumption that the effects for one area at
different times are uncorrelated is likely to lead to conservative estimates of
the variance components. However, the free estimation of the variances
associated with each time point means that a variety of life course models can
be examined without the need for prior assumptions.
Neighbourhood at time T1
Neighbourhood at time T2
Figure 1 Multilevel structure
of individuals nested within a cross-classification of areas at two time points
(T1 and T2). Diagram shows person P1 moving from area A1
to area A2 and person P2 moving from area A2 to area A1.
Alastair H Leyland
Senior Research Scientist
MRC Social and Public Health
Sciences Unit, Glasgow, Scotland
Epidemiological Division, National
Public Health Institute, Oslo, Norway
MRC Social and Public Health
4 Lilybank Gardens
Glasgow G12 8RZ
1. Kawachi I, Subramanian SV. Neighbourhood influences on
health. Journal of Epidemiology and
Community Health 2007; 61:3-4.
T, Clarke P, Wiggins RD, et al. Who you live with and where you live: setting
the context for health using multiple membership multilevel models. Journal of Epidemiology and Community Health
3. Leyland AH. Assessing the impact of mobility on health: implications for lifecourse epidemiology.Journal of Epidemiology and Community Health 2005; 59:90-1.
D, Ben-Schlomo Y, Lynch J, et al. Life course
epidemiology. Journal of Epidemiology and
Community Health 2003; 57:778-83.
Ø, Strand BH, Davey Smith G. Childhood and adulthood
socioeconomic position across 20 causes of death. A
prospective cohort study of 800 000 Norwegian men and women.Journal of Epidemiology and
Community Health (in press).