Bayesian space-time models for surveillance and policy evaluation using small area data Nicky Best Department of Epidemiology and Biostatistics Imperial College, London Joint work with Guangquan (Philip) Li, Sylvia Richardson, Bob Haining, Anna Hansell, Mireille Toledano, Lea Fortunato Outline Introduction Policy Evaluation: Evaluating Cambridgeshire Constabulary’s ‘no cold calling’ initiative Surveillance: Detecting unusual trends in chronic disease rates Introduction Bayesian space-time modelling of small-area data is now common in many application areas disease mapping small area estimation (official statistics) mapping crime rates modelling population change ..... Key feature is that data are sparse Bayesian hierarchical model allows smoothing over space and time → improved inference Introduction Many different inferential goals description prediction surveillance estimation of change / policy impact ..... Many different ways of formulating the space-time model space + time (separable effects) space + time + interaction space-time mixture models ..... Our set-up Inferential goals: detection of areas with ‘unusual’ time trends Goal 1: Policy evaluation Goal 2: Surveillance a policy or intervention has been implemented in a known subset of areas, and we wish to evaluate whether this has had a measureable impact on the event rate in those areas no a priori subset of areas of interest; we just wish to identify any areas whose event rate differs markedly from the general time trend General modelling framework Assume most areas exhibit a common temporal trend (separable space and time effects) – the ‘common trend’ model For a small subset of areas, assume time trend is unusual (space-time interaction) – the ‘local trend’ model Goal 1: Policy Evaluation Evaluating Cambridgeshire Constabulary’s ‘No Cold Calling’ initiative In collaboration with Guangquan Li*, Robert Haining+, Sylvia Richardson +University *Imperial of Cambridge College, London Definition of a “cold call” A visit or a telephone call to a consumer by a trader, whether or not the trader supplies goods or services, which takes place without the consumer expressly requesting the contact. Not illegal but often associated with forms of burglary and “rogue trading”. To discourage cold calling police have targeted specific neighbourhoods as “no cold calling” (NCC) areas: street and house signage; information packs for residents; informal follow-up meetings. Cambridgeshire Constabulary initiated NCC scheme in parts of Peterborough in 2005 and extended it in 2006. Locations of the NCC areas in Peterborough Summary of NCC-targeted areas Data for evaluation All reported “burglary in a dwelling” events (Home Office classification code 18, sub-codes 0-10, and code 29) used as outcome Surrogate for rouge trading and distraction burglary (very small number or recorded events) Data aggregated to annual counts by Census Output Area (COA) in Peterborough Time period: 2001-2008 Total of 9388 recorded burglaries Median burglaries per area per year = 2 5th and 95th percentiles: 0 – 8 Raw data: individual and aggregated time trends Poisson test RR01-04 = 1.06, p=0.56 RR05-08 = 0.85, p=0.19 Positive impact of policy? Strategy for evaluation Compare burglary rates before and after implementation of NCC scheme Comparison is done after adjustment for systematic changes in burglary rate in other non-NCC areas difference between 2 time periods is indicative of impact of policy use of ‘control’ areas helps to differentiate how much of the change is due to the policy and how much to other external factors Deal with sparsity of the data (i.e. small number of burglary events) by Data aggregation → assessing overall impact Hierarchical modelling of local impacts → assessing both overall and local impacts → Separate signal from noise Constructing the control group Control areas are selected to have similar local characteristics (e.g. burglary rates; deprivation scores) to those in the NCC-targeted group Control areas are chosen to be Lower Super Output Areas (LSOA) to obtain reliable control data (results are similar with COA-level controls) Control Criterion Description No. of LSOAs 1 All LSOAs in Peterborough 88 2 ±10% burglary rate of the NCC group in 2005 9 3 ±20% burglary rate of the NCC group in 2005 20 4 ±30% burglary rate of the NCC group in both 2004 and 2005 7 5 LSOAs containing the NCC-targeted COAs (but excluding the NCC-targeted COAs) 10 6 LSOAs that had “similar” multiple deprivation scores (MDS) as those for the NCC LSOAs in 2004 46 Evaluation procedure Evaluation procedure Evaluation procedure The impact function We consider various functional forms for the impact function (Box and Tiao, 1975) The impact of the policy is quantified through the estimation of the function parameter(s) Model selection via DIC Name No change Step change A linear function of time A generalization function Functional form Full model specification Control areas + NCC areas pre-scheme yit ~ Poisson(ni it ) log(it ) ui t it ~ N(0,1000) (overall intercept) 1:T ~ RW1 ( W, 2 ) (time effect) NCC areas post-scheme (t ≥t0) ykt ~ Poisson(nk kt ) log( kt ) * uk* t* kt* I (t t0 ) f (t , bk ) * t* t ui ~ N(0, u2 ) (area effect) uk* ui k it ~ N(0, 2 ) (overdispersion) kt* N(0, 2 ) f (t , bk ) bk (t t0 1) bk ~ N( b , b2 ) Implementation yit 2 Common trend model u2 2 it t ui Model fitted in WinBUGS Common trend model fitted to control areas (all years) plus NCC areas (years before scheme only) Implementation yit 2 2* Common trend model u2 2 it t * t * uk* b b 2 bk ykt ui for k=i kt* Local trend model , t ≥ t0 Model fitted in WinBUGS Common trend model fitted to control areas (all years) plus NCC areas (years before scheme only) Local trend model (impact function) fitted to NCC areas (years after scheme) Implementation ‘cut’ link * distributional constant (no learning) yit 2 2* Common trend model u2 2 it t * t * uk* b b 2 bk ui for k=i kt* ykt Local trend model, t ≥ t0 Model fitted in WinBUGS Common trend model fitted to control areas (all years) plus NCC areas (years before scheme only) Local trend model (impact function) fitted to NCC areas (years after scheme) ‘Cut’ function used to prevent NCC area (post-scheme) data influencing estimation of common trend model parameters Results: choice of impact function No Change Step Linear Generalization function Dbar 15.27 14.32 9.77 11.75 pD 1.21 2.29 2.25 2.57 DIC 16.49 16.61 12.02 14.33 Linear impact function has smallest DIC No change Posterior probability of “success” i.e. Pr(bk < 0) Heterogeneity of local impacts b = -1.1 95% CI(-2.6, 0.2) f (t, bk) = bk∙(t t0+1); bk = + b xk + dk ; dk ~ N(0, 2) Some of the variability in local NCC impacts may be due to coverage The larger the proportion of properties that were visited in a COA, the greater the impact of the NCC scheme Heterogeneity of local impacts Two possible explanations for coverage effect A “threshold” effect A “dilution” effect NCC scheme does not have a measurable impact (in terms of reducing burglary rates) unless a sufficient number of households in the local area are visited Because the COA is the unit of analysis, the NCC scheme impact could be diluted when the households that are visited are only a small proportion of the total households in the COA Neither of these explanations for the coverage effect undermines our overall assessment of the policy’s success Conclusions: NCC scheme NCC scheme led to overall “success” Overall, NCC-targeted areas experienced a 16% (95% CI: -2% to 34%) reduction in burglary rate per year This suggests a positive impact of the NCC policy which had the effect of stabilizing burglary rate in the targeted areas while overall burglary rates were going up Linear impact function is better at describing the data than the other 3, suggesting a gradual and persistent change There exist different impacts between targeted COAs, perhaps due to local differences in implementing the schemes Assessing NCC impact for whole of Cambridgeshire The NCC scheme was extended to the whole of Cambridgeshire for the period 2005-08 We applied our evaluation model to assess impact of NCC scheme separately for urban and rural areas Overall, schemes in urban areas were more successful than those in rural areas. % change in burglary rates after 1st year of NCC scheme No change Urban Rural No change 12UBFW0011 (0.74) 12UEHH0013 (0.76) 00JAPA0012 (0.81) 12UCGA0013 (0.83) 00JAND0001 (0.85) 12UDGS0014 (0.86) 12UCGD0002 (0.88) 12UDGQ0006 (0.87) 00JANG0024 (0.88) 12UCGA0003 (0.88) 12UDGQ0024 (0.89) 00JANG0009 (0.92) NCC2007 00JAPB0010 (0.76) 00JANG0025 (0.96) NCC2006 00JANY0010 (0.55) 00JANC0016 (0.54) 00JANG0013 (0.85) 00JANE0006 (0.83) 00JANE0010 (0.88) MDI MatchRate 00JANT0027 (0.92) 00JANQ0023 (0.91) NCC2005 Overall Overall(0.96) (0.96) −100 −50 0 50 Percentage change in burglary rate compared to controls Overall (0.38) 100 29 Conclusions: Model Hierarchical model allows borrowing of strength across NCC areas Joint estimation of common trend and local trend models enables full propagation of uncertainty enables evaluation of local impacts even when data are sparse Parameters of common trend model treated as ‘distributional constants’ in local trend model Facilitated using ‘cut’ function in WinBUGS More complex impact functions could be implemented, but need sufficient time points post-policy for reliable estimation Goal 2: Surveillance Detecting unusual trends in chronic disease rates In collaboration with Guangquan Li, Sylvia Richardson, Anna Hansell, Mireille Toledano, Lea Fortunato Imperial College, London Surveillance of small area data For many areas of application, such as small area estimates of income, unemployment, crime rates and rates of chronic diseases, smooth time changes are expected in most areas However, policy makers and researchers are often interested in identifying areas that ‘buck’ the national trend and exhibit unusual temporal patterns These abrupt changes may be due to emergence of localised predictors/risk factors(s) or the impact of a new policy or intervention Detection of areas with “unusual” temporal patterns is therefore important as a screening tool for further investigations Motivating example 1: COPD mortality Chronic Obstructive Pulmonary Disease (COPD) is a common chronic condition characterized by slowly progressive and irreversible decline in lung function responsible for approximately 5% of deaths in the UK Main risk factors include Smoking Occupational exposure to high levels of dusts and fumes Outdoor air pollution “Umbrella” term for broad range of disease phenotypes Time trends may reflect variation in risk factors and also variation in diagnostic practice/definitions Motivating example 1: COPD mortality Objective 1: Retrospective surveillance to highlight areas with a potential need for further investigation and/or intervention (e.g. additional resource allocation) Objective 2: Policy assessment Industrial Injuries Disablement Benefit was made available for miners developing COPD from 1992 onwards in the UK As miners with other respiratory problems with similar symptoms (e.g., asthma) could potentially have benefited from this scheme, there was debate on whether this policy may have differentially increased the likelihood of a COPD diagnosis in mining areas Data Observed and agestandardized expected annual counts of COPD deaths in males aged 45+ years 374 local authority districts in England & Wales 8 years (1990 – 1997) Difficult to assess departures of the local temporal patterns by eye Need methods to quantify the difference between the common trend pattern and the local trend patterns express uncertainty about the detection outcomes Bayesian Space-Time Detection: BaySTDetect BaySTDetect (Li et al 2011) is a novel detection method for short time series of small area data using Bayesian model choice between two competing space-time models Model 1 assumes space-time separablility for all areas → one common temporal pattern across the whole study region Model 2 provides local time trend estimates for each spatial unit individually For each area, a model indicator is introduced to decide whether Model 1 or Model 2 is supported by the data → Quantifying the difference A Bayesian procedure of controlling the false discovery rate is employed → Expressing uncertainty about detected areas BaySTDetect: modelling framework yit ~ Poisson( it Eit ) log( it ) i t model 1 for all i, t i ~ spatial BYM model (common spatial pattern) The temporal trend pattern is the same for all areas t ~ random walk (RW[ 2 ]) model (common temporal trend) log( it ) ui it model 2 for all i, t ui ~ N(0,1000) (area-specific intercept) Temporal trends are independently estimated for each area. it ~ random walk (RW[ i2 ]) (area-specific temporal trend) Model selection A model indicator zi indicates for each area whether Model 1 (zi =1) or Model 2 (zi =0) is supported by the data Implementation Model 2: Local trend Model 1: Common trend t i it ui it[C] it[L] Eit Eit yit yit zi it Selection model it zi it[C ] (1 zi ) it[ L ] Eit yit Prior on model indicator For the model indicator zi, we have zi ~ Bernoulli( ) where 0.95 This prior on zi reflects the surveillance nature of the analysis where we expect to find only a small number of unusual areas a priori ensures that a common trend can be meaningfully defined and estimated Classifiying areas as “unusual” Classification of areas as “unusual” is based on the posterior model probabilities pi = Pr(zi | data) Small values of pi indicate low probability that area i fits the common trend → high probability of being “unusual” Need a rule for calibrating the pi that acknowledges the multiple testing setting How low does pi need to be in order to declare area i as unusual? False Discovery Rate (FDR) is the proportion of detected areas that are false (i.e. not truly unusual) (Benjamini & Hochberg, 1995) Various methods to estimate or control FDR Here we control the posterior expected FDR (Newton et al 2004) Detection rule based on FDR control First rank the areas according to increasing values of pi At a nominal FDR level of , the first k ranked areas are declared as unusual where k is the maximum integer k satisfying p j 1 ( j) k where p(j) is the jth ranked posterior common-trend model probability This procedure ensures that (posterior) expected number of false positives is no more than (k ×) of the k declared unusual areas Simulation study to evaluate operating characteristics of BaySTDetect Simulated data were based on the observed COPD mortality data Three departure patterns were considered When simulating the data, either the original set of expected counts from the COPD data or a reduced set (multiplying the original by 1/5) were used 15 areas (approx. 4%) were chosen to have the unusual trend patterns areas were chosen to cover a wide range expected count values and overall spatial risks Results were compared to those from the popular SaTScan space-time scan statistic Simulation Study: Departure patterns Common trend, exp(t) Departure pattern, exp(t ∙) 2 different departure magnitudes: =1.5 and =2.0 Simulation Study: expected counts Table: Summary of the original set of age-adjusted expected counts used in the simulation Simulation Study: FDR control Empirical FDR vs corresponding pre-defined level: Pattern 2 0.15 0.20 Pre-set FDR level 0.8 0.6 0.4 0.2 0.0 0.0 0.10 Empirical FDR 1.0 1.0 0.8 0.6 95% sampling interval 0.4 Empirical FDR mean 0.2 1.0 0.8 0.6 0.4 0.2 0.0 Empirical FDR 0.05 Reduced expected; =2.0 Original expected; =2.0 Original expected; =1.5 0.05 0.10 0.15 0.20 Pre-set FDR level 0.05 0.10 0.15 0.20 Pre-set FDR level SaTScan: Empirical FDR = 0.19 (0.00 to 0.78) for scenario with original expected counts and =2.0 Sensitivity of detecting the 15 truly unusual areas SaTScan (p=0.05) True departure magnitude: =1.5 E=24 E=33 E=42 E=52 E=80 Expected count quantiles Expected count quantiles Sensitivity 0.0 0.2 0.4 0.6 0.8 1.0 E=24 E=33 E=42 E=52 E=80 0.0 0.2 0.4 0.6 0.8 1.0 Sensitivity Pattern 2 0.0 0.2 0.4 0.6 0.8 1.0 Sensitivity 0.0 0.2 0.4 0.6 0.8 1.0 Sensitivity BaySTDetect (FDR=0.1) True departure magnitude: =2.0 E=24 E=33 E=42 E=52 E=80 E=24 E=33 E=42 E=52 E=80 Expected count quantiles Expected count quantiles Sensitivity of detecting the 15 truly unusual areas: reduced expected counts Pattern 2; True departure magnitude: =2.0 SaTScan (p=0.05) 0.8 0.6 0.2 0.4 Sensitivity 0.6 0.4 0.2 0.0 0.0 Sensitivity 0.8 1.0 1.0 BaySTDetect (FDR=0.1) E=5 E=6 E=8 E=11 E=16 Expected count quantiles E=5 E=6 E=8 E=11 E=16 Expected count quantiles COPD application: Detected areas (FDR=0.05) COPD application: Interpretation Results provide little support for hypothesis regarding the industrial injuries policy only 3 out of 40 ‘mining’ districts detected (Barnsley, Carmarthenshire and Rotherham) unusual trend patterns in these areas are not consistent Two unusual districts (Lewisham and Tower Hamlets) with an increasing trend (against a national decreasing trend) were identified in inner London These areas are very deprived, with high in-migration and ethnic minorities → might expect different trends to rest of country In fact, Tower Hamlets has been commissioning various local enhanced services to tackle high rates of COPD mortality since 2008. This rising trend could potentially have been recognised earlier in the 1990s through using BaySTDetect as a surveillance tool. COPD application: SaTScan Primary cluster: North (46 districts) – excess risk of 1.05 during 1990-92 Secondary cluster: Wales (19 districts) – excess risk of 1.12 during 1995-96 Example 2: Data mining of cancer registries The Thames Cancer Registry (TCR) collects data on newly diagnosed cases of cancer in the population of London and South East England It is one of the largest cancer registries in Europe, covering a population of over 12 million, and holds nearly 3 million cancer registration records. We perform a retrospective surveillance of time trends for several cancer types using BaySTDetect aim to provide screening tool to detect of areas with “unusual” temporal patterns automatically flag-up areas warranting further investigations Cancer data Cancer incidence for population aged 30+ years Breast (female only) Colon (males and females combined) Lung (males and females, separately) South East England, ward level (1899 areas) Period 1981-2008 Data were aggregated by 4-year intervals 7 time periods for the detection analysis Cancer data summary OBS EXP OBS colon EXP Female OBS lung EXP Male lung OBS EXP breast Min 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 Q1 10.0 11.3 5.0 5.7 3.0 4.0 6.0 7.6 Median Mean 16.0 17.6 16.5 17.6 8.0 9.1 8.5 9.1 5.0 6.4 5.9 6.4 10.0 11.8 11.2 11.8 Q3 24.0 23.0 12.0 11.8 9.0 8.3 16.0 15.2 Max 69.0 56.5 42.0 34.6 34.0 24.5 66.0 39.5 Comparable to reduced expected count scenario in simulation study Results: Number of detected areas (out of 1899) Cancer type FDR=0.05 FDR=0.1 FDR=0.15 FDR=0.2 Breast 9 19 35 54 Colon 0 3 5 8 Lung (female) 0 1 2 4 Lung (male) 6 14 24 39 54 Detected areas: breast cancer Summarising the unusual trends With a relatively large number of detected areas (e.g., breast and male lung cancer), examination of the individual trends becomes difficult For the detected areas, the estimated RR trends from the local trend model are fed into a standard hierarchical clustering method (hclust in R) i t log( it ) ui it model 1 model 2 The cluster-specific trends are then compared to the overall RR trend 56 1 2 3 Period 4 5 6 7 0.6 0.6 0.6 0.8 0.8 0.8 1.2 1.4 1.6 1.8 1.6 1.8 1 2 3 clusters 2 3 3 4 Period 4 5 5 6 Overall trend 20 areas 12 areas 10 areas 12 areas 6 7 1.8 1 1.6 d hclust (*, "complete") 1.2 1.4 1.4 0.4 0.4 0.6 0.6 0.8 0.8 29UMGT 22UHHP 26UFGH 22UQGT 00AWGC 26UJFX 00ASHB 26UCHD 26UJGQ 00AGGE 43UDGA 43ULGR 00LCPB 00ADGW 43UMFU 00AUFY 00AJGY 00ALHF00APGK 00ANGA 45UDGQ 00BBGX 29UBHR 29UCGF 26UJGC 00AYGL26UEGJ 00ASGJ 00BKGQ 00AKGP 00BHGR 29UHHE 00BCGU 00BHGK 00BCFZ 00BKGR 00BFGE 43UFGN 00BAGM 00MLNP 00BFGN 45UBFT 00BJGG 00BFGS 21UDFU 00AUGM 00ANGC 29UNHA 21UGGJ 00BJGM 00ATGB 00ALGP 00AFGM 00AFGG 0.0 0.5 1.2 1.0 1.2 Relative Risk 1.0 Relative Risk 1.0 Height 1.5 1.6 1.6 1.4 1.4 2.0 2.5 1.8 1.8 Overall trend 54 areas 1.0 Relative Risk 1.0 Relative Risk 1.2 Relative Risk 1.0 Overall trend 30 areas 12 areas 12 areas 0.4 0.4 0.4 Cluster Dendrogram 1 cluster 2 clusters 7 Overall trend 42 areas 12 areas Breast cancer FDR=0.2 Period 1 1 2 4 clusters 2 3 Period 3 4 Period 4 5 5 6 6 7 5 clusters Overall trend 20 areas 12 areas 10 areas 5 areas 7 areas 7 Black line = common trend Coloured lines = average local trend in each cluster BaySTDetect: Conclusions and Extensions We have proposed a Bayesian space-time model for retrospective detection of unusual time trends Simulation study has shown good performance of the model in detecting various realistic departures with relatively modest sample sizes Possible extensions include: Spatial prior on zi to allow for clusters of areas with unusual trends Time-specific model choice indicator zit, to allow longer time series to be analysed Alternative approaches to calibrating posterior model probabilities, e.g. decision theoretic approach (Wakefield, 2007; Muller et al., 2007) References G. Li, R. Haining, S. Richardson and N. Best. Evaluating Neighbourhood Policing using Bayesian Hierarchical Models: No Cold Calling in Peterborough, England. Submitted G. Li, N. Best, A. Hansell, I. Ahmed, and S. Richardson. BaySTDetect: detecting unusual temporal patterns in small area data via Bayesian model choice. Submitted G. Li, S. Richardson , L. Fortunato, I. Ahmed, A. Hansell and N. Best. Data mining cancer registries: retrospective surveillance of small area time trends in cancer incidence using BaySTDetect. Proceedings of the International Workshop on Spatial and Spatiotemporal Data Mining, 2011. www.bias-project.org.uk Funded by ESRC National Centre for Research Methods