ISYS3438 Chap.7 Multivariate Views and Calculated Fields
Multivariate Views and Calculated Fields
Multivariate displays reveal profiles and interacting conditions but quickly accumulate clutter. This chapter compares parallel coordinates, multiple lines, combination charts and scatterplot matrices. It explains why axis order and normalisation change a profile view and why exact retrieval may need a companion table or ranking.
Calculated fields receive the same discipline: define the business meaning, formula, aggregation level and zero rule, then recompute a small case before using the field across a dashboard.
What this chapter covers
- 01
parallel coordinates
- 02
combination charts
- 03
scatterplot matrix
- 04
normalisation
- 05
calculated fields
- 06
aggregation level
Worked example · free
Reduce a regional profile view
- +1Reorder axes so related commercial and operational measures are adjacent.
- +1Disclose and test the normalisation used across different units.
- +1Mute all records and highlight only the selected region plus a benchmark.
- +1Add a supporting ranked view when exact comparison matters.
Key terms
- Parallel coordinates
- Parallel coordinates draw each record as a line crossing several variable axes.
- Normalisation
- Normalisation transforms variables to a comparable scale under a disclosed rule.
- Calculated field
- A calculated field derives a business measure using a defined formula and aggregation level.
- Scatterplot matrix
- A scatterplot matrix displays pairwise relationships among several quantitative variables.
Multivariate Views and Calculated Fields FAQ
Why does axis order matter in parallel coordinates?
Relationships are easiest to inspect between adjacent axes, so reordering changes which crossings and profiles are visually prominent. The order should reflect the analytical question rather than software default.
Should I normalise every multivariate view?
Only when comparison across different scales requires it. State the transformation and preserve original values for interpretation. Normalisation can clarify profiles while hiding the business units the audience needs.
Why can averaging ratios be wrong?
Groups with different denominators receive equal weight when ratios are averaged directly. Compare that result with a ratio of aggregated numerators and denominators, then choose the version matching the business definition.
Assessment move
Take one multivariate view and remove one variable at a time. Keep a field only if it contributes a relationship the audience uses. For calculated fields, maintain a four-column audit of meaning, formula, level and edge-case rule. Recompute at least one result outside Tableau and compare it with the displayed mark. Write the analytical purpose of every added variable.
A third or fourth attribute should help locate a subgroup, explain heterogeneity or support a decision; it should not appear merely because colour, size and shape are available. Begin with two coordinated common-scale views and add complexity only when the relationship would otherwise disappear. Keep legends, units and selection state close to the marks so the audience does not have to memorise several encodings at once.
For parallel coordinates, normalise only when relative profiles are the intended question, label the transformation and reorder axes according to relationships worth inspecting. Check whether line crossings become clutter rather than evidence. For scatterplot matrices, identify the few panels that answer the decision and annotate them instead of expecting an executive audience to scan every pair.
Combination charts need compatible meanings and carefully explained axes; a second scale can create a visual relationship through arbitrary range choices. Treat each calculated field as a documented business definition. Expand its name into numerator, denominator, unit, time window and exclusion rules. Determine whether it is evaluated per row, after aggregation or across a table partition.
Test zero denominators, nulls, negative adjustments, duplicate identifiers and groups of unequal size using hand-checkable records. Compare an unweighted average of rates with a rate calculated from combined totals and choose according to the decision, not whichever result looks cleaner. Maintain a calculation register containing formula, grain, owner, validation example and known limitation.
When a filter or table layout changes a result, record the order of operations and partition settings. In the final visual, provide a route back to component measures and sample size. Close the analysis by explaining why each variable earns its channel, what simpler alternative was rejected and how the display constrains interpretation.
Reproducible calculation and restrained encoding are the core of a defensible multivariate argument. After the view is built, perform a variable-removal test. Hide one encoded attribute at a time and ask whether the decision becomes worse. If nothing important is lost, remove that channel or move it into detail on demand.
Repeat the test for calculations by substituting their component measures and checking whether the derived field clarifies or conceals the business meaning. Verify all units and scales after each change. Finally, give the reduced view to a peer and ask them to state the relationship, subgroup and action without assistance. Their explanation should match the registered definitions.
This audit controls clutter while preserving the genuinely multivariate evidence the task requires. Finish by locking a validated calculation example into the project notes. Include input rows, intermediate values and expected result. Recheck it after changing joins, filters or dashboard structure.
This small regression test prevents a visually unchanged workbook from carrying a different business definition after later edits and gives the report a concrete method reference.