An introduction to spatial interaction models: from first principles

What are SIMs?

Spatial Interaction Models (SIMs) are mathematical models for estimating movement between spatial entities developed by Alan Wilson in the late 1960s and early 1970, with considerable uptake and refinement for transport modelling since then Boyce and Williams (2015). There are four main types of traditional SIMs (Wilson 1971):

An early and highly influential type of SIM was the ‘gravity model’, defined by Wilson (1971) as follows (in a paper that explored many iterations on this formulation):

\[ T_{i j}=K \frac{W_{i}^{(1)} W_{j}^{(2)}}{c_{i j}^{n}} \]

“where \(T_{i j}\) is a measure of the interaction between zones \(i\) and \(W_{i}^{(1)}\) is a measure of the ‘mass term’ associated with zone \(z_i\), \(W_{j}^{(2)}\) is a measure of the ‘mass term’ associated with zone \(z_j\), and \(c_{ij}\) is a measure of the distance, or generalised cost of travel, between zone \(i\) and zone \(j\)”. \(K\) is a ‘constant of proportionality’ and \(n\) is a parameter to be estimated.

Redefining the \(W\) terms as \(m\) and \(n\) for origins and destinations respectively (Simini et al. 2012), this classic definition of the ‘gravity model’ can be written as follows:

\[ T_{i j}=K \frac{m_{i} n_{j}}{c_{i j}^{n}} \]

For the purposes of this project, we will focus on production-constrained SIMs. These can be defined as follows (Wilson 1971):

\[ T_{ij} = A_iO_in_jf(c_{ij}) \]

where \(A\) is a balancing factor defined as:

\[ A_{i}=\frac{1}{\sum_{j} m_{j} \mathrm{f}\left(c_{i j}\right)} \]

\(O_i\) is analogous to the travel demand in zone \(i\), which can be roughly approximated by its population.

More recent innovations in SIMs including the ‘radiation model’ Simini et al. (2012). See Lenormand, Bassolas, and Ramasco (2016) for a comparison of alternative approaches.

Implementation in R

Before using the functions in this or other packages, it may be worth implementing SIMs from first principles, to gain an understanding of how they work. The code presented below was written before the functions in the simodels package were developed, building on Dennett (2018). The aim is to demonstrate a common way of running SIMs, in a for loop, rather than using vectorised operations (used in the simodels package) which can be faster.

library(tmap)
library(dplyr)
library(ggplot2)
zones = simodels::si_zones
centroids = simodels::si_centroids
od = simodels::si_od_census
tm_shape(zones) + tm_polygons("all", palette = "viridis")
#> -- tmap v3 code detected --
#> [v3->v4] tm_polygons(): migrate the argument(s) related to the scale of the visual variable 'fill', namely 'palette' (rename to 'values') to 'fill.scale = tm_scale(<HERE>)'
#> [plot mode] fit legend/component: Some legend items or map compoments do not fit well, and are therefore rescaled. Set the tmap option 'component.autoscale' to FALSE to disable rescaling.

od_df = od::points_to_od(centroids)
od_sfc = od::odc_to_sfc(od_df[3:6])
sf::st_crs(od_sfc) = 4326
od_df$length = sf::st_length(od_sfc)
od_df = od_df %>% transmute(
  O, D, length = as.numeric(length) / 1000,
  flow = NA, fc = NA
  )
od_df = sf::st_sf(od_df, geometry = od_sfc, crs = 4326)

An unconstrained spatial interaction model can be written as follows, with a more-or-less arbitrary value for beta which can be optimised later:

beta = 0.3
i = 1
j = 2
for(i in seq(nrow(zones))) {
  for(j in seq(nrow(zones))) {
    O = zones$all[i]
    n = zones$all[j]
    ij = which(od_df$O == zones$geo_code[i] & od_df$D == zones$geo_code[j])
    od_df$fc[ij] = exp(-beta * od_df$length[ij])
    od_df$flow[ij] = O * n * od_df$fc[ij]
  }
}
od_top = od_df %>% 
  filter(O != D) %>% 
  top_n(n = 2000, wt = flow)

tm_shape(zones) +
  tm_borders() +
  tm_shape(od_top) +
  tm_lines("flow")
#> [plot mode] fit legend/component: Some legend items or map compoments do not fit well, and are therefore rescaled. Set the tmap option 'component.autoscale' to FALSE to disable rescaling.

We can plot the ‘distance decay’ curve associated with this SIM is as follows:

summary(od_df$fc)
#>      Min.   1st Qu.    Median      Mean   3rd Qu.      Max. 
#> 0.0002404 0.0256166 0.0801970 0.1495649 0.2035826 1.0000000
od_df %>% 
  ggplot() +
  geom_point(aes(length, fc))

We can make this production constrained as follows:

od_dfj = left_join(
  od_df,
  zones %>% select(O = geo_code, all) %>% sf::st_drop_geometry()
)
#> Joining with `by = join_by(O)`
od_dfj = od_dfj %>% 
  group_by(O) %>% 
  mutate(flow_constrained = flow / sum(flow) * first(all)) %>%
  ungroup()
sum(od_dfj$flow_constrained) == sum(zones$all)
#> [1] TRUE
od_top = od_dfj %>% 
  filter(O != D) %>% 
  top_n(n = 2000, wt = flow_constrained)

tm_shape(zones) +
  tm_borders() +
  tm_shape(od_top) +
  tm_lines("flow_constrained")
#> [plot mode] fit legend/component: Some legend items or map compoments do not fit well, and are therefore rescaled. Set the tmap option 'component.autoscale' to FALSE to disable rescaling.

Validation

od_dfjc = inner_join(od_dfj %>% select(-all), od)
#> Joining with `by = join_by(O, D)`
od_dfjc %>% 
  ggplot() +
  geom_point(aes(all, flow_constrained))

cor(od_dfjc$all, od_dfjc$flow_constrained)^2
#> [1] 0.1735933

References

Boyce, David E., and Huw C. W. L. Williams. 2015. Forecasting Urban Travel: Past, Present and Future. Edward Elgar Publishing.

Dennett, Adam. 2018. “Modelling Population Flows Using Spatial Interaction Models.” Australian Population Studies 2 (2): 33–58. https://doi.org/10.37970/aps.v2i2.38.

Lenormand, Maxime, Aleix Bassolas, and José J. Ramasco. 2016. “Systematic Comparison of Trip Distribution Laws and Models.” Journal of Transport Geography 51 (February): 158–69. https://doi.org/10.1016/j.jtrangeo.2015.12.008.

Simini, Filippo, Marta C González, Amos Maritan, and Albert-László Barabási. 2012. “A Universal Model for Mobility and Migration Patterns.” Nature, February, 812. https://doi.org/10.1038/nature10856.

Wilson, AG. 1971. “A Family of Spatial Interaction Models, and Associated Developments.” Environment and Planning 3 (January): 132.