Flash Profiling is a rapid type of descriptive method adapted from free choice profiling. Normally used with untrained judges, each individual selects their own attributes to describe at set of products. They then use these attributes to rank the products (for Flash Profiling) according to their intensity on an ordinal scale, or for Free Choice Profiling (FCP) they instead rate the products.
- demo data set.xlsx
Due to the way the data is collected (i.e. assessors use their own words – and may use different numbers of attributes), the data format for Flash/FCP is particular to only that method in EyeOpenR®.
The columns
of data are as follows:
Judge,
Attribute, Session, Replica, Sequence, then the Flash/FCP values - one column
for each product.
Ensure you
select the right data type when you bring the data in – by selecting Flash/FCP
in the analysis method.
Background
This analysis module can be used for both Flash and Free Choice Profiling (FCP) data as the underlying method is the same, regardless of the data. The analysis uses a method called Multiple Factor Analysis (MFA) to overlay data matrices with the same products in each matrix BUT variables in each matrix may be different. MFA allows the relationship between the blocks of data to be established through seeking common structures. It provides useful information regarding how close each block of data is to each other.
These matrices may arise in a range of different approaches to “sensory” data collection. The data is presented in multiple blocks of data, and so the data analysis should take into account this structure – so we use MFA as it is one of the multiple-block data analysis methods. (Previously, and in other software, Generalised Procrustes Analysis (GPA) has also been used for analysing this type of data.)
For both Flash and FCP, scores are considered as intensity (although in Flash, scores are ranks) – and their analysis is the same. Contrary to classical QDA® data, the information cannot be aggregated across assessors, and the individual tables should be considered separately.
The analysis performs PCA on the individual data tables and then combines the tables using different weights (this is the basis of the MFA method). The weighting of the tables makes it possible to ensure the tables with more variables, or a stronger correlation structure, do not weight too heavily in the analysis. Within a table, the variables must be of the same type (quantitative or qualitative), but the tables can be of different types.
Options
The module is easy to use as there are not many options needed.
The user can select to analyse using Multiple Factor Analysis (MFA) with or without the standardization of the data for the first part – the PCAs on the individual data tables. When choosing, recall that:
• Standardized = rescaled = correlation based
• Unstandardized = unscaled = covariance based
This is a sensory question – is it right with the objectives/design of the study that all the variables are assessed on the same scale? If so, then choose standardized to run correlation based PCAs.
If instead, you want the attributes to be assessed on their individual scales, then choose unstandardized and run covariance based PCAs.
If you are still unsure then try both and see what the differences are!
Results and Interpretation
The outputs do not show all the individual PCA outputs as this is an internal step of the analysis, but starts with the outputs from the final weighted PCA across all the tables of data.
- Eigenvalues: The first output is the table of eigenvalues for the final PCA, which also shows the percentage of variation explained by each dimension. (Sum of Eigenvalues = total variance of samples across variables/attributes). The percentages (given individually per dimension and also as cumulative percentages so you can see their sum down the last column) are based on the eigenvalues.
- Products: The Product coordinates on each dimension are then provided, listed by product and these are used to create the product map, which is also on this tab of the output under the heading ‘graph’, for the first 2 dimensions. The coordinates are also listed underneath the product chart. The product map helps visualise the relative positioning of the products, and in common with other multivariate tools, products that lie close together are more similar than those that are far apart. The contributions and squared cosines are also provided in tables in case the user wishes to look at these. If the squared cosines associated with the axes used on a chart are low, the position of the product in question should not be interpreted.
- Attributes: The coordinates used to create the attribute map are provided on the Attributes tab, listed for each attribute and data table. Again the contributions and squared cosines are also provided in tables in case the user wishes to look at these, before the attribute chart (on the graph tab). If the squared cosines associated with the axes used on a chart are low, the position of the attribute in question should not be interpreted. The squared correlations are the squared correlations between the factors (dimensions) and the variables/attributes. To understand the product positions the attribute map can be examined. Here the attributes are identified according to the judges using them so it is important to look for common trends in the descriptions for areas of the sensory map as well as the unique attributes that some judges use.
- Group: Next are the Lg and RV Coefficients, followed by the Group (Judge/Panellist) map (with the coordinates listed beneath). The RV coefficient is a multi-dimensional correlation coefficient ranging from 0 (perfect orthogonality) to 1(homothetic). Where the RV coefficient between two judges is high this means their product maps were similar. Likewise low RV coefficients indicate lack of agreement in the product map structure for two judges. Look at comparing the judges to each other as well as to the consensus solution. We do not always expect the RV to be high for this if judges are using different attributes. The LG coefficient, is interpreted similarly - they measure how much the judges/tables are related in a pairwise fashion (i,j). The more attributes of judge i that are related to the attributes of judge j, the higher the Lg coefficient. The group map shows the position of the judges, where judges close to each other had similar sensory mapping of the products. If a judge lies out on their own it means they saw the product set differently.
- Partial Axes: The coordinates and contributions for the partial axes are provided, together with the chart of the partial axes representation. The partial axes representation, correspond to the projection as illustrative of the dimensions related to each group separately (i.e. results of the separate analyses on each group) into the overall solution of the MFA. It is useful to look at which of the individual judges dimensions correspond to the first and second dimensions of the overall solution for a deeper understanding of the data. It may be that some judges are showing more importance on one dimension compared to the other, even though they are in a similar space to the consensus.
- R packages: SensoMineR, MFA (FactoMineR), Averagetable (SensoMineR)
- R function settings that are not otherwise visible to the user:
- Groups
should be defined as such in the metadata:
- “standardized”
- “centered”
- “nominal”
- “frequency”