Table 1 A comparison of multilevel discrete-time and Bayesian spatial survival models used in this case study
| Â | Multilevel | Bayesian spatial |
|---|---|---|
Software | MLwiN 2.261 | WinBUGS version 1.4 |
Cost | Once-off purchase | Free |
Available interfaces | Stata | Stata, SAS, R |
Initial data structure | Â | Â |
Retains multilevel structure (Patients nested in higher-level units) | Yes | No |
Data expansion required | Yes | No |
Geographical Structure | None | Preserves adjacent areas |
Explanatory variables | Unit Record Individual and higher-level | Aggregated at individual-level and higher-level |
Modelled Outcome | Individual deaths | Aggregated deaths |
Random Effects | Yes | Yes |
Prior distributions | Gamma, Uniform | Any including Gamma, Uniform, CAR |
Default Priors | Yes | requires user specification of priors; greater flexibility |
Estimation Method: MCMC | Yes | Yes |
Number of MCMC chains | Single only | Single (multiple allowed also) |
Level of random effects | Individual and higher-level | Higher-level |
Within-area correlation | Yes | No |
Between-area correlation | No | Yes |
Adjacency matrix | No | Yes |
Computational efficiency ( 5Â year data)2 | 5-7 days | 5-7 days |
Ease of Implementation | R equires prior data expansion | Requires specification of model including prior distributions |
Diagnostic Tests/ convergence plots | Yes | Yes |
Questions answered: | Â | Â |
Do area- and individual-level factors impact survival for individual patients? | Yes | No |
Extent to which between-individual variability is explained by covariates at both levels | Yes | No |
Estimates unexplained area-level spatial variation after adjusting for parameters | No | Yes |
Map spatial variation by small-areas | No | Yes |
Cross level interactions | Yes | No |
Allow unit record individual-level inferences | Yes | No |
Parameter estimates | Odds ratio (OR) | Relative risk (RR) |