methods for investigating zoning effects mark tranmer ccsr
TRANSCRIPT
Methods for investigating zoning effects
Mark Tranmer
CCSR
Allowing for ‘area effects’
• Suppose we have some area level information
• Such as: aggregate information for a particular set of ‘areal units’ e.g. wards; EDs; Output Areas; Districts
• Or individual level data with area indicators.
Allowing for these effects in our analyses
• Then we might say ‘great! I’ll fit a multilevel model’ – especially if we have individual level data with area indicators.
• Or we might calculate correlations etc at the area level from aggregate area level data.
But …
• If we calculate area level correlations because we want to make inferences about individuals that live in those areas but we only have area level data… problem: ecological fallacy
• So let’s suppose we can actually do an analysis using individual level data with area indicators … e.g. a multilevel model. Hence simultaneously allowing for individual and area level effects.
• Does that solve the problem?
No, because …
• What do we mean by an ‘area’?
• Modifiable Areal Unit Problem (MAUP)
• Analyses that involve areas are affected by
The average population size of those areas: ‘scale effects’
No, because …
• Once we choose a particular scale, they are also affected by the way in which those areas are defined. I.e. the choice of boundaries: ‘Zoning effects’.
• Also: Scope effects? What is the overall region of study? This will have implications for the extent of variation.
Zoning effects example
• Suppose we have a region that contains a 9 areal units of equal population, and we want to make a ward from three of these contiguous units.
Zoning effects example
• Ward A1
Zoning effects example
• Ward B1
Zoning effects example
• Overlay wards A1 and B1
Zoning effects example
• We can also do the same thing for the other wards: e.g.
I’m interested in developing a statistical framework to investigate these effects
• I think a cross-classified multilevel model might be the way to tackle the problem
• What I hope to do is to find a way to assess the nature and extent of zoning effects at a particular scale.
Two level model(s)
ijAjAijA euy 2)( uAjAuVar
2)( eAijAeVar
ijBjBijB euy 2)( uBjBuVar
2)( eBijBeVar
Cross-classified model
),(),(),( jBjAijBjAjBjAjBjAi euuuy
2)( uAjAuVar 2)( uBjBuVar 2
),( )( uABjBjAuVar
2)( )( eABjAjBieVar
How to test this idea
• Simulated data: I set up a simulation study• I generated some simulated data for a normally
distributed variable. Each of the 9 cells in the grid has a different (but known) mean and within each of the 9 cells I set the variance to be equal (25).
• So I aimed to simulate complex between-cell variation (whilst knowing the procedure I had applied to induce that variation).
I assumed these zonings
ResultsTwo level models * Variance component estimates
Ward Indiv
A, person 427 366
B, person 176 616
Cell,person 642 150
Cross-classified models
Estimated parameter:
Var(A) Var(B) Var(A*B) Var(Indiv)
A,B,cell,person 333 4 309 150
Conclusion
• I think we have a framework for investigating the causes of zoning effects
• It seems to work for simulated data, though I have yet to fully work out what these results mean
• Can anyone suggest to me some real data that investigate using this methodology.