1 of 15

Geography and Intersectionality

Andrew Bell (with Dan Holman, Sarah Salway and Mark Green)

Sheffield Methods Institute, University of Sheffield

Tweet @andrewjdbell

Email Andrew.j.d.bell@Sheffield.ac.uk

​

2 of 15

So far:

  • Social processes can be intersectional – we are a combination of our characteristics, and more than the sum of our parts
  • These social processes affect health outcomes
  • We’ve thought about various attributes of identity… but what about place?

3 of 15

Aims

  • To include place in our intersectional identities
    • Where we live is an important part of who we are
    • Neighbourhood affect different sorts of people differently
    • Some sorts of people more intergrated into (and more affected by?) neighbourhoods
    • Eg: high pollution levels = worse health… but more so for particular groups? (eg those more at risk of respiratory problems)
    • Living in a poor area might have a different effect to being poor.

4 of 15

Challenges in this

  • What do we mean by neighbourhoods
    • Partly defined by what data we have – LSOAs?
    • 35,000 LSOAs in England and Wales – even with a big dataset, we’ll struggle to find intersections
    • Use bigger Geography? – but then it’s no longer really about neighbourhoods (and still intersections are too small)

5 of 15

What did we do

  • Use LSOAs to define neighbourhoods (but not intersections
  • But also use Index of Multiple Deprivation values to define intersections

Intersections

People

6 of 15

What did we do

  • Use LSOAs to define neighbourhoods (but not intersections
  • But also use Index of Multiple Deprivation values to define intersections

Intersections

People

Neighbourhood

7 of 15

What did we do

  • Data: UK Biobank
    • Upside: It’s big (500k), has biomarker measures, links to LSOA (and so IMD)
    • Downside: it’s unrepresentative

​

  • Use a range of biomarkers (HbA1C, C-Reactive Protein, Systolic blood pressure, Waist circumference, Grip strength)

​

8 of 15

What did we do

  • Range of Multilevel models, both including and not including Geography in our models
    • Compare levels of variance when including and not including Geography – to see, to what extent Geography matters generally
    • See differences between specific geo-intersections
    • Reduced risk of multiple testing (see Bell et al 2019)

​

    • Models fitted with MCMC (5k burnin, 50k chain) using MLwiN, non-informative priors, convergence checked on all parameters etc.

​

9 of 15

Results

  • Geography (defined by IMD) does matter a bit for all outcomes

 

HbA1c (mmol/mol)

CRP (mg/L)

SBP (mm Hg)

Waist (cm)

Grip strength (kg)

Age 50-59

Ref

Ref

Ref

Ref

Ref

Age 60-69

2.98 (2.63 - 3.32)

.070 (.052 - .087)

7.92 (7.28 - 8.56)

2.33 (1.75 - 2.94)

-2.64 (-2.93 - -2.34)

Age 70-79

4.32 (3.94 - 4.71)

.102 (.082 - .121)

14.86 (14.45 - 15.55)

3.65 (3.02 - 4.30)

-5.57 (-5.89 - -5.25)

Women

Ref

Ref

Ref

Ref

Ref

Men

.980 (.687 - 1.27)

-.089 (-.104 - -.075)

4.47 (3.89 - 5.01)

8.38 (7.85 - 8.91)

15.44 (15.18 – 15.71)

White British

Ref

Ref

Ref

Ref

Ref

Chinese

2.13 (1.59 - 2.66)

-.253 (-.277 - -.228)

-2.07 (-3.28 - -.854)

-8.38 (-9.33 - -7.45)

-3.10 (-3.61 - -2.60)

Indian

4.46 (4.05 - 4.87)

.099 (.077 - .121)

-.314 (-1.08 - .464)

-.216 (-.932 - .486)

-5.55 (-5.91 - -5.20)

Pakistani

6.37 (5.85 - 6.90)

.179 (.141 - .217)

-1.47 (-2.62 - -.312

2.92 (2.00 - 3.84)

-5.07 (-5.56 - -4.58)

Caribbean

4.20 (3.76 - 4.66)

-.027 (-.049 - -.004)

2.29 (1.43 - 3.15)

1.02 (.239 - -1.78)

2.21 (1.81 - 2.59)

African

3.76 (3.25 - 4.27)

.036 (.009 - .064)

5.05 (4.08 - 6.01)

3.33 (2.47 - 4.17

-.874 (-1.31 - -.445)

Low education

Ref

Ref

Ref

Ref

Ref

Med. education

-.990 (-1.36 - -.629)

-.068 (-.084 - -.051)

-1.28 (-1.98 - -.575)

-1.64 (-2.29 - -.951)

1.42 (1.10 - 1.77)

High education

-1.49 (-1.87 - -1.10)

-.132 (-.148 - -.115)

-3.05 (-3.77 - -2.33)

1.63 (.774 - 2.47)

1.77 (1.43 - 2.12)

V. low neigh. dep.

Ref

Ref

Ref

Ref

Ref

Low neigh.

.035 (-.459 - .518)

.018 (-.006 - .042)

.408 (-.554 - 1.39)

.933 (.098 - 1.78)

-.218 (-.640 - .206)

Med. neigh. dep.

.549 (.073 - 1.02)

.042 (.018 - .066)

.482 (-.396 - 1.38)

1.63 (.774 - 2.47)

-.475 (-.908 - -.047)

High neigh. dep.

1.06 (.593 - 1.55)

.066 (.041 - .091)

.902 (.016 - 1.79)

2.44 (1.61 - 3.27)

-1.09 (-1.51 - -.670)

V. high neigh. dep.

1.47 (.994 - 1.95)

.122 (.096 - .148)

.222 (-.667 - 1.13)

3.16 (2.34 - 4.00)

-1.91 (-2.33 - -1.49)

n

362495

348710

375312

385250

383536

10 of 15

Results

  • How much Geography matters – not a huge amount, but a bit (once other variables accounted for – still big Geographical differences)

 

Neighbourhood variance

Intersectional variance

Individual variance

Hba1c, model 1

.738 (.674-.803)

-

42.43 (42.23-42.63)

Hba1c, model 2

.395 (.345-.448)

-

42.44 (42.25-42.64)

Hba1c, model 3

.326 (.277-.377)

10.27 (7.74-13.57)

40.17 (39.98-40.36)

Hba1c, model 4

.324 (.273-.372)

.838 (.540-1.24)

40.17 (39.98-40.36)

Hba1c, model 5

.146 (.109-.185)

12.21 (10.51-14.17)

40.10 (39.91-40.29)

Hba1c, model 6

.150 (.107-.195)

1.44 (1.06-1.89)

40.11 (39.93-40.30)

Model 1 = Two-level null neighbourhood model.�Model 2 = Model 1 + IMD quintiles as main effects.�Model 3 = Cross-classified null model, IMD not used to define intersections.�Model 4 = Cross-classified main effects model (gender, ethnicity, education, age), IMD not used to define intersections.�Model 5 = Cross-classified null model, IMD used to define intersections.�Model 6 = Cross-classified main effects model (gender, ethnicity, education, age, IMD), IMD used to define intersections.

11 of 15

Results

12 of 15

Results

Shallower deprivation gradients (and lowest risk) for white British

13 of 15

Results

Deprivation gradient steepest for Pakistani women

14 of 15

Results

Very high average HbA1cfor high ed, high deprivation Pakistani men

15 of 15

Why?

  • Are these results real, or the result of small samples/chance?
  • Our models correct somewhat for multiple testing, but not perfectly
  • But it’s clear that simplistic narratives are too simplistic – careful targeting if particular groups and particular places might be worthwhile
  • Need for more (qual) research on particular groups highlighted here

​