How do you compare statistics across revisions?

Comparing statistics across revisions in ISIC Rev. 5, UN Statistics Division, requires more than placing two tables side by side, because the categories themselves may have changed between the periods compared. The honest method is to use a correspondence table, to note where splits and mergers break comparability, and to compare at the broadest level that remains comparable where detail cannot be reconciled. A comparison that hides a break in series misleads even when its arithmetic is correct.

Why can two correct tables still refuse to line up?

Each table can be internally correct and still fail to support a direct comparison, because the category boundaries moved between the two periods. A category that kept its name may have gained activities from a neighbour or lost activities to a new category, and a category that vanished may have been divided among several successors. The numbers therefore describe different baskets of activity even when the label on the row looks unchanged. This is why a comparison must begin with the structure, not with the figures: read the correspondence table first, identify which rows map one to one, which split, and which merged, and only then decide which comparisons the data can honestly support. Skipping that step and comparing row by row produces a trend that may reflect reclassification rather than real change, and no amount of careful arithmetic can repair a comparison whose rows were never equivalent.

Which level should a cross revision comparison use?

Use the broadest level at which the categories still cover the same scope, and be explicit about the step up you took. If a class split in Rev. 5, the class figures before and after the change cannot be placed in one series, but the group or division that contains all the pieces of the split may still be comparable, provided the correspondence table shows that the broader total was unaffected. That trade is worth making: a broader comparison that is sound tells a reader more than a detailed comparison that is false. The cost is real, since detail is lost, so say so in a footnote rather than letting the reader assume the detailed series survived. Where even the broader level was affected, for example where a merger crossed a division boundary, the honest presentation is a break in series, with the two periods shown as separate segments and the break labelled, not joined by a line that implies continuity.

What must a footnote about a break in series actually say?

A vague note that a revision occurred is not enough. A useful footnote names the versions on each side of the break, the level at which the comparison is being made, and the nature of the change: which categories split, which merged, and whether any part of the old total has no counterpart in the new structure. It should also state the reference periods clearly, because national adoption dates can differ, and a country that adopted Rev. 5 later will show the break in a different year from a country that adopted it earlier. Readers comparing countries therefore need the adoption timing as well as the structural change. A footnote written this way does more than warn: it lets a reader judge how much weight the comparison can bear, and it lets a later analyst reproduce the comparison rather than guessing at the choices behind it. Keep the version label, ISIC Rev. 5, UN Statistics Division, attached to the Rev. 5 side of every such note, so the label is never left to be inferred from context.

A series that spans a revision, handled in stages

Consider an analyst who wants a ten year trend for a sector that was restructured in Rev. 5. She starts with the correspondence table and marks the rows that map one to one: those can be carried straight through. Next she marks the split rows and the merged rows, and tests whether a broader aggregate covers each split completely; where it does, she builds the trend at that broader level and footnotes the loss of detail. Where it does not, she leaves a visible break in the series and presents the two segments side by side without joining them. The published chart is less smooth than the one she first imagined, with a broader line and one visible break, but every segment on it means what its label says. That is the whole craft of cross revision comparison in one workflow: map first, aggregate where the map allows it, and break openly where it does not.

Anyone reading a long series should therefore ask three questions before trusting a trend: which revision produced each segment, at what level the comparison was made, and where the breaks are marked. If those answers are not visible, the trend is not yet ready to be quoted.

This guide is general information about statistical classification, not a ruling on any particular record. It does not assign a code, a principal activity, or a registration to any reader, and the code that belongs on a specific record is a matter to confirm with the national statistical institute or business register that holds the record.