Brainstorm of the methodological approach

[martinfleis]

Brainstorm of the methodological approach

• The paper deals with the geographical differentiation of the relation between people and places
• We are interested in the ways different socioeconomic groups of people occupy different types of urban form and whether and how this changes across different regions in Germany

Input data

• Census information on 100m grid covering built-up areas.
  • Traditional census data but also rent price
• Morphometric assessment of Germany
  • Individual morphometric characters
  • Morphotopes as the smallest urban localities of distinctive character
  • Taxonomy combining morphotopes into a hierarchical classification system

Approach

Options

• Reduce complexity of census data to an index of deprivation or a subset of indices.

1. Modelling of a set level of morphological taxonomy based on census data

   • This would require a reasonable amount of raw data from Census, so a single index is not viable for the modelling step.
   • Could be either a set of indices or a selection of raw measurements + potentially their lag
   1. Global baseline model
   2. Global model with spatial fixed effects (based on Landkreise (400))
   3. Localised models following the GWR logic but based on a Landkreise geometry (so a submodel would be trained on a Lankreise and its neighbours following some neighborhood definition)

   • SHAP analysis on each model and a comparison of localised effects

   • LISA of local SHAP values

   • Spatial cross-validation based on morphotope delineation.

   • We can say the degree in which each census variable affects the morphology in each place but not much about the differences of who lives where. The interpretation value of the result might be a bit low.

   • We could do modelling per type and compare those across the country but that could result in a lot of model. But also possibly into a lot of interesting and interpretable data.

2. Modelling of deprivation levels based on morphology

   • Census is squeezed to an index (or a set of them but then we'd do the same analysis for all)
   • We'd use morphometric characters aggregated onto the grid
   • It is turned into an analysis of features of urban form that explain deprivation - we get a regional shape of deprivation
   1. Global baseline model
   2. Global model with spatial fixed effects (based on Landkreise (400))
   3. Localised models following the GWR logic but based on a Landkreise geometry (so a submodel would be trained on a Lankreise and its neighbours following some neighborhood definition)

   • SHAP analysis on each model and a comparison of localised effects

   • LISA of local SHAP values

   • Spatial cross-validation based on a grid.

   • We can say how deprivation in each area looks like from the perspective of morphometrics, which may be tricky to interpret reasonably as morphometrics is easier to grasp when looking at typology.

3. No modelling, just a relationship analysis

   • Regional summary of average population per a morphological type

   • Spatial analysis of the averages (probably LISA)

   • This is simple and straightforward so we should probably do it either way.

   • It is directly interpretable

   • We would have to work on an aggregated level - on a set level of the taxonomy. The issue is that this aggregation affects the composition and we lose the variability within the type we still work with within the modelling.

---

Comment as of 20/8/2024

I am currently leaning towards a combination of 3 and 1, where 1 is based on a set of domain indices (employment index, education index, rent index etc). It may be more work but if we go all the way to the localised models done per type of form, it could be really interesting. The key in the paper will be to ensure that the relatively complex methods are well explained. The story behind is still clear (at least in my head) and singular, so it could be a nice paper.