Drivers Modelling helps us understand which factors drive and influence an outcome of interest.
Examples of outcomes include:
- Rating of a brand’s reputation – e.g. likelihood to recommend.
- Exhibiting of a certain behaviour (yes or no) – e.g. whether a person is classed as a “problem gambler”.
- Agreement with a statement – e.g. the government needs to do more to tackle climate change.
- The proportion of patients treated with brand X.
This technique is particularly helpful when assessing the unique strength of the relationship between each predictor and the outcome.
There are various types of regression models – linear, logistic, ordinal and multinomial, and the type we use depends on the type of outcome and predictors we are working with.
Here are the three main types of drivers analysis in increasing complexity:
1. Simple KDA
Simple Key Drivers Analysis (KDA) is used for more straightforward projects where:
- We have a large block of similarly scaled potential predictors (our drivers) that all need to be evaluated together and are usually ranges (e.g. 1-5, 1-7 scales) or tick boxes (yes/no or apply/not apply questions);
- There is no need to control for other variables (such as demographics) initially;
- The sample size is small relative to the number of predictors;
- There are many potential predictors, and they are highly correlated with each other;
- Simple output of unique importance of the predictors is preferred over more technical output;
- Need to repeat analysis and compare results for multiple subgroups (e.g. different countries in a multi-country study).
KDA produces an easy-to-understand “percentage importance” output to a Shapley Value Analysis.
We use our correlated component regression (CCR) algorithm, which is designed to produce robust results for most types of survey data, even with many predictors, limited sample and high correlations between predictors.
We offer both a screened and unscreened option for predictors.
The Classic Regression Analysis is used when a more detailed understanding of what drives an outcome is needed.
Used when we are dealing with more complex dependent variable types (e.g. categorical) and where the model outputs need describing in a more detailed, formal way.
Blocks of predictors are entered or screened in a specific order, according to the objectives.
Used when one or more of the following apply:
- Need to control for the effects of certain types of variables (e.g. demographics) on the outcome;
- Have at least several categorical predictors (e.g. demographics, such as working status, age band and region);
- Results are going to be published in full and so we need to use a more formal regression method;
- Interested in exploring interactions between predictors rather than just their main effects.
3. Structured Equation Modelling (SEM)
SEMs can be considered a system of regression models with multiple outcome variables that can be predictors of other outcomes.
It allows us to build and test a system of interlinked predictions that explains outcomes of interest.
It also allows us to create composites to summarise multiple predictors or outcomes as part of the model.
The models can be described via flow diagrams.