Purpose

Creates a graphical summary of temporal dominance of sensations (TDS) data. 

Data Format

EyeOpenR has a specific temporal data format that should be adhered to

1. TDS EyeQuestion format.xlsx. 
   

Panelist	Product	Session	Replica	Sequence	Intensity	Multiple	Attribute	Start	Stop
J3	P1	1	1		1	Before	B10	0	3
J3	P1	1	1		1	Before	B8	3	5
J3	P1	1	1		1	Before	B6	5	6
J3	P1	1	1		1	Before	B4	6	17
J3	P1	1	1		1	Before	B5	17	28
J3	P1	1	2		1	Before	B6	0	1
J3	P1	1	2		1	Before	B8	1	3
J3	P1	1	2		1	Before	B10	3	5

The table shows the first 8 data rows from the example data set, the first 5 columns, panellist, product, session, replicate, and sequence are used to represent the experimental design used in the study.  The ‘Intensity’ column is not needed for TDS data and can either be empty or should contain a value of 1 in every cell.  If the study design incorporates different occasions or conditions in which products are evaluated, then the ‘Multiple’ column is used to tell EyeOpenR which product evaluations belong to the same condition.  Each condition is analysed separately - for the example data set there are both ‘Before’ and ‘After’ conditions.  The final three columns represent the times at which each attribute is the dominant sensation during each product evaluation.  A value of zero in the Start column indicates the start of a product evaluation, so for example judge J3, product P1 session 1, and replicate 1, the attribute B10 is initially dominant between zero and 3 seconds, and this product evaluation ends with attribute B5 being dominant between 17 and 28 seconds.

The ’attribute’ tab of the Excel lists the attributes that are used in each condition and has the following format (an illustrative sample of 6 data rows are shown):

Note that datatype should be specified as ‘TDS’ and that a column called ‘multi’ is used to represent the different conditions in the experimental design.  The values in the ‘attribute’ and ‘multiple’ columns should be consistent between the two excel tables.

Background

EyeOpenR produces two types of bandplot.  The first type of bandplot is a statistical summary of the data, with one plot for each product that shows the times at which each attribute is significant.  There are several data transformations that need to be carried out to produce the chart, starting with a conversion of the TDS data from the start/stop format to a binary format as follows.  For each 1 second interval, a value of 1 is recorded whenever an attribute is dominant, and a value of 0 is recorded elsewhere.  For each product and each attribute, and at each 1 second interval, the 1s and 0s from all judges and replicates are averaged to give the proportion of product evaluations for which dominance was recorded.  The result is a curve for each attribute and each product showing how the proportion varies with time.  If one of the smoothing options has been selected, the proportions are further processed by applying a weighted moving average algorithm, where the window size is proportional to the degree of smoothing. The window size (n) is as follows:

Smoothing          n(sec)

Extra low            3

Low                     7

Medium              11

High                    15

For each curve the software then decides at which time points, the proportion is significantly higher than is expected by chance.  The chance level is simply 1/p where p is the total number of attributes, since a judge who is randomly moving between p attributes, has a probability of selection of 1/p for each one.  The time points where each attribute is contributing significantly are chosen by performing a 1-sample proportion test (Z-test) to decide where the height of the curve exceeds 1/p with an alpha level set by the “Level filter” option.  A band plot is then used to visualize, for each product, the periods where each attribute is a significant contributor.  Blocks of colour, coded by attribute, show the important periods for each attribute as can be seen from the chart below.

The second type of plot is the individual bandplot which shows the raw data in the start/stop format.  A chart summarises a single product and contains a separate band for each judge and each replicate.  There is a different block of colour for each period during which an attribute is dominant, and to interpret the plot the user should refer to the separate legend to see which colours are associated with which attributes.  An example individual bandplot is shown below.

Options

1. Smoothing – either select ‘no’ for no smoothing or choose one of the other 4 options to apply a weighted moving average algorithm to smooth out the curves of dominance proportion vs time (see background section for more details).

1. Level filter – the alpha level for the 1-sample proportion test that decides at which times an attribute is significant for each product. Default 5%.

1. Scale by judge – whether to standardize each product evaluation in the time domain, so that they all finish at the same time.  If the food samples under test need to be chewed and swallowed, then the time taken to the swallow point can vary considerably between judges – use judge scaling to correct for this.

1. Number of decimals for values – controls the number of decimals printed in all numeric output.

Results and Interpretation

1. The ‘Bandplot’ tab has buttons that allow the user to select a chart for any specific combination of product and condition.  On a chart, each block of colour represents a period of time measured along the horizontal axis, during which an attribute has been selected as dominant more frequently than would be anticipated by chance alone.  A different colour is used to represent each attribute, and the attributes are arranged along the vertical axis in bands.  The chart shows coloured rectangles only where the frequency of selection of the attribute is significant at the percentage level set by the ‘Level Filter’ option.  The 1% level (alpha=0.01) is the most conservative and will show the least amount of colour, while the 30% (alpha=0.3) is the least conservative and will show the most amount of colour.

1. The ‘Individual Bandplot’ tab has buttons that allow the user to select a chart for any specific combination of product and condition.  Each chart is a visual representation of the raw data, with one band for each product evaluation.  A band is multicoloured and continuous, where the change of colours along it represents the transition between dominant attributes during a single product evaluation.  If ‘Scale by Judge’ is set to ‘No’, then each band will be of a different length terminating at the time where the judge stopped recording sensations – often this is the point where the judge swallowed the food sample.  If ‘Scale by Judge’ is set to ‘Yes’, then scaling will be applied to the time scale so that each product evaluation has been stretched or shrunk to an average length.  In this latter instance, the chart will appear as a solid rectangle of colours as all the bands finish at the same time.

1. The ‘Ind. Bandplot Legend’ tab has one button for each condition and gives a legend showing the correspondence between the colours used in the individual band plots and the attribute names.

Technical Information

·         The R functions used for analysis were all developed in house at Qi Statistics.

References

1. Galmarini, M. V., Visalli, M., & Schlich, P. (2017) “Advances in representation and analysis of mono and multi-intake Temporal Dominance of Sensations data” Food Quality and Preference, v56, pp247-255.

1. Monterymard, C., Visalli, M., & Schlich, P. (2010) “The TDS-band plot: A new graphical tool for temporal dominance of sensations data”, Proceedings 2nd Conference of The Society of Sensory Professionals, pp 27–29.