8.1 Designing Tables & Statistical Exhibits

Tables Are Part of the Assessed Content

PCPA Domain C is Model Interpretation and Presentation, worth 30% of the exam and 40% of the project rubric. Its first task is to "create and interpret statistical/tabular and graphical/visual representations of data." The assigned readings are Few's Show Me the Numbers, Cairo's How Charts Lie, Knaflic's Storytelling with Data, and Gelman and Unwin on infovis versus statistical graphics.

That means table and chart design are examinable content, not presentation polish. And on the project, with a hard cap of five tables or supporting graphics and 1,250 words including appendices, every exhibit has to earn its place.

Table or Graph?

Few's test is about what the reader must do with the numbers.

A fitted GLM's coefficient table is a table: the reader needs the number. A pattern of loss cost across ten deciles is a graph: the reader needs the shape.

The common failure is the table used as a graph — twenty rows of numbers where the point was simply "it rises then flattens." The reader has to do the visual work the analyst should have done.

Alignment, Precision, Order

Three mechanical choices decide whether a table can be read quickly.

Align numbers right, with a consistent number of decimals. Right alignment puts the ones digit under the ones digit, so magnitude is visible down the column without reading every value. Mixed decimal places destroy that; 1.2, 0.85 and 12.417 in one column force the eye to parse each entry. Align text left. Align column headers with their data.

Show only the precision the decision needs. A territory relativity printed as 1.0837 claims four significant digits of accuracy from data that may contain 300 claims. Round to the precision the decision actually uses — usually three digits for a relativity, whole dollars for a pure premium, one decimal for a percentage. Over-precision is a quiet credibility claim, and a reviewer will notice.

Order rows by something meaningful. Alphabetical order is the default and almost never the most useful. Order by magnitude when the reader is looking for the biggest effect, by a natural sequence (age band, vehicle age, policy year) when the order carries meaning, and keep a consistent order across related tables so the reader can compare them.

Reduce the Non-Data Ink

Few's central design idea is that anything on the page that is not data competes with the data. In practice:

• Drop vertical gridlines entirely; use white space to separate columns.
• Use at most a light horizontal rule under the header and above a total row.
• Avoid fill colour and shading in every other row unless the table is very wide.
• Use one highlight — bold, or a single colour — and use it only on what the reader should notice.
• Repeat units in the header (Pure premium ($)) rather than on every cell.

The goal is not minimalism for its own sake. It is that when almost everything is emphasised, nothing is.

What an Actuarial Exhibit Must Carry

Beyond formatting, an actuarial table has content obligations that a generic business table does not:

1. Volume. Exposure, claim count, or record count for every row. Without it, the reader cannot tell a credible relativity from noise. This single column is the most frequently omitted and the most frequently needed.
2. A basis statement. Policy year or accident year, paid or incurred, gross or net, evaluated as of when.
3. Units. Dollars, car-years, per 100 of payroll, percentage points.
4. A title that states the finding, not the mechanics. "Loss cost rises 3.1x from the best to the worst decile" tells the reader something; "Table 2: Model output by decile" does not.

Two Tables Worth Building for the Project

The coefficient / relativity table. This is the exhibit that satisfies the rubric criterion "candidate generates and provides technical output (e.g. indicate coefficients from the model)".

Everything a reviewer needs is present: the base level is identified, both the coefficient and its exponentiated relativity are shown, exposure lets the reader see that Territory 4's 1.667 rests on thin data, and the standard error confirms it.

The model comparison table. One row per candidate model, columns for the diagnostics you used to choose between them — degrees of freedom, deviance, AIC, BIC, and an out-of-sample statistic. This single table can carry the rubric criteria "describes which model they selected and why" and "explains what they did to iteratively check and build their model" in far fewer words than prose.

[!WARNING] On the project, tables and graphics share one allowance of five. A coefficient table, a model comparison table, and a validation chart already use three. Decide the full exhibit set before you build any of them, and do not spend an allowance on a chart that merely restates a number already in the text.