Healthy Lifestyles Synthetic Estimates

Project

Shaun Scholes, Kevin Pickering and Claire Deverill

Outline of presentation

• A ‘small-area estimation problem’ with national surveys such as the Health Survey for England

• Get around this problem by statistical modelling. HSE data is used in combination with other data sources to generate estimates at fine levels of geography

• Importance of quality assurance

• Caveats when using small area estimates

Demand for local area data on health

• Health related behaviour is not uniform across England (‘place matters’)

• NatCen regularly use HSE data to produce health indicators: • England (trend tables)• Larger areas such as: GORs (Annual HSE reports); GORs,

SHAs & Counties (SHA report 2004, Compendium of Clinical Indicators, forthcoming Chief Medical Officer’s report on binge drinking)

Regional indicators using HSE data

U n ita ry A u th o rit ies

W a rds M S O A s

D is tric t s

C o u n ties

G O R s

P C O s

S H A s

E n g la nd

Small area estimation problem

• National surveys such as the HSE are not designed to provide reliable estimates at local level:• Sample sizes typically small or zero within small areas, even

after pooling years of HSE data together• Direct estimates, where we can calculate them, are

unbiased - but have low precision due to wide sampling error

• Local health surveys are an option but various problems:• Costs• Comparability across surveys

• Small area estimation techniques are an alternative

Healthy lifestyles synthetic estimates project

• Information Centre for Health and Social Care and Neighbourhood Statistics commissioned NatCen to use HSE data to produce small area estimates of:• current smoking (adults)• binge drinking (adults)• obesity (adults)• fruit & vegetables consumption (adults & children

separately)

Small area estimation methodology used by NatCen

• Use a statistical model to express the relationship between individual healthy lifestyle behaviour and area-level information

• Outputs from that model used to generate a model-based estimate for all areas

• But must be interpreted differently to direct estimates ~ estimates represent the expected prevalence for an area based on its population characteristics

IllustrationWhether an

individual currently smokes from HSE

2003-2005

Attach area-level information to HSE dataset:

Regional indicator

% 16-74 with no qualifications (Census 2001)

Statistical model run on the subset of areas covered by HSE

2003-2005

Use terms from model to obtain a predicted estimate for all small areas

Area-level characteristics and current smoking

Variable Odds ratio Standarderror

95%confidence

interval

T2 ratio

North East SHA (reference) 1North West SHA 0.92 0.076 0.79-1.09 0.92Yorkshire & The Humber SHA 0.94 0.082 0.79-1.12 0.47East Midlands SHA 0.99 0.089 0.83-1.18 0.02West Midlands SHA 0.95 0.083 0.80-1.13 0.34East of England SHA 0.95 0.085 0.80-1.13 0.35London SHA 0.89 0.086 0.74-1.08 1.35South East Coast SHA 0.97 0.092 0.81-1.17 0.09South Central SHA 0.86 0.086 0.71-1.05 2.24South West SHA 0.89 0.080 0.75-1.06 1.75% of IS claimants ‘carers and others’ 0.99 0.002 0.99-1.00 16.64% White origin 1.01 0.002 1.01-1.02 48.90% unpaid carers 0.94 0.014 0.91-0.97 17.82% highest qualification NVQ Level 1 or noqualifications (aged 16-74)

1.02 0.002 1.02-1.03 132.50

% 16+ residing as couple 0.98 0.003 0.97-0.98 84.47

Implementation

MSOA Region % 16-74with noqualifications

% 16+residingas couple

% currentsmoking

E02000984 North-West 32% 73% 13%E02001007 North-West 52% 64% 24%E02000996 North-West 61% 53% 36%

Quality assurance

• Quality of estimates crucially depend on the quality of the model

• Need for QA measures to provide evidence on plausibility of estimates• Internal checks (examine the residuals, correlation between

direct and model-based estimates)• External checks (no gold standard!)

– correlation with direct estimates from other surveys such as GHS

– correlation with local boost surveys

– correlation with Index of Multiple Deprivation 2004

Residual plot (all areas sampled)

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

0 0.1 0.2 0.3 0.4 0.5 0.6

Model-based estimates for PCS

Area

leve

l res

idua

l

Correlation with IMD 2004 (Bolton)

Rank of smoke estimates compared with rank of IMD

0

5

10

15

20

25

30

35

40

0 5 10 15 20 25 30 35 40

Rank of IMD 2004

Ran

k o

f M

od

elle

d E

stim

ates

Putting estimates in a contextIndicator England E02001007 E02000182

Region - North West London

Local Authority - 00BL 00AG

Survey estimate ofcurrent smoking

24.1% [23,25] - -

Model-based estimate - 23.9% [13,40] 19.5% [10,34]

Rank of MSOA on IMD - 3,965 5,009

% dwellings CTX band Aor B (<£52,000)

44% 74% 5%

% residents 16-74 withno qualifications

46% 52% 21%

% residents 16+ residingas a couple

61% 64% 45%

Estimates have health warnings

• Have to be interpreted differently ~ not estimates of actual prevalence

• Dependent upon quality of the model • Method relies on having powerful predictors of health

variations between areas. A large amount of unexplained between area variation results in wide CIs:• Can compare MSOAs against the national average• Cannot meaningfully compare MSOA X against MSOA Y as

CIs overlap

• But we stress that these estimates are not available elsewhere!