Small area modelled estimates

Modelled estimates, also called synthetic estimates, involve two key steps. First, a robust national survey that gives data on the behaviour(s) in question and on a number of demographic, socio-economic, etc. factors that, in statistical terms, 'predict' much of the behaviour in question. For tobacco smoking, for example, these might include gender, age, social status and area deprivation. These independent 'predictive' variables are combined into a statistical model that best predicts the behaviour (the dependent variable). Note that prediction in this context is not about causality but about strength of statistical association.

The second step is to apply the national model to local areas using local data for the independent 'predictive' variables. There is no point in constructing a national model that uses data that is not available for local areas, so strong predictive variables may not be included in the model because they are not available locally. Past modelling has relied heavily on the census to provide the predictive variables, though the improved availability of local administrative data through Scottish Neighbourhood Statistics may now give other options.

The model gives a prediction of what prevalence of a behaviour would be expected locally if the national relationships are also true locally. Like all statistical models it does not predict 100% of a behaviour. Also, like all statistical models, it has a margin of error and interpretation should have due regard to appropriately calculated confidence intervals. The national survey underpinning the model may have a low response rate (for example, the latest Scottish Health Survey has an individual response rate of only 56%), introducing a further element of uncertainty through response bias. There are mixed views regarding whether modelled estimates can provide a good level of accuracy.

Modelled estimates can give an indication of the likely level of a behaviour in an area based on the characteristics of the local population. They cannot be used to look at change over time in local areas, because the characteristics of an area (used to predict the behaviour) do not change quickly. Even when used to estimate current prevalence, by the very nature of the method there is scope to question the validity of the results for any particular area, and debate is particularly likely if resource allocation is an issue. 

It is important to note that synthetic estimates cannot be used to monitor the effectiveness of interventions.