Reading a table in a document is a four-pass routine: read the title and column headers, check the units and what each row represents, trace one row and one column to confirm how the data is indexed, then read any footnotes before you quote a number. Do it in that order and you'll catch most misreadings before they reach a report. The routine works on screen and on paper, in PDFs and in word processors, and the only real hazard is skipping a step because the table looks simple.

Start With the Table Title and Its Sourcing Line

The title tells you what the rows and columns measure. A title like "Average household expenditure by category, 2023" locks three things in place: the statistic is an average, the unit is a household, and the vintage is 2023. Read it before anything else and write down the three facts.

Below the title sits the source line, and it decides how much weight the table can carry. A table sourced to a national statistics office and a table labelled "internal estimates" can carry identical numbers and demand different levels of trust. If no source line exists, treat the vintages and definitions as unknown.

Treat the title as a contract you will check the body against. If the title says "average" and the cells hold counts, you have found an error in the document, not in your reading. Say so in your notes rather than picking whichever number suits you.

Digital readers lose this page furniture first. In a forwarded spreadsheet or a screenshot pasted into a slide, the title and source often get cropped, leaving a grid of numbers with no anchor. Before you interpret such a grid, go back to the original file β€” the caption usually survives in the PDF or the source workbook.

Confirm Column Headers and the Unit in Every Header

Every column header carries two pieces of information: the measure and, usually, its unit. "Revenue (Β£m)" tells you the cells are millions of pounds, so a cell reading "12.4" means Β£12.4 million, not Β£12.40. "Share (%)" tells you the cells are percentages and should total 100 across a complete set of rows.

Headers repeat, abbreviate, or shift across a multi-page table, and those shifts cause quiet errors. A column labelled "2023" on page one can appear as "23" on page two, and a currency marker stated once in the header may not repeat. If a value looks ten or a thousand times too large, check the unit before you check your arithmetic.

Watch for stacked headers, where a top row groups columns ("North", "South") and a second row subdivides them ("2022", "2023"). Read stacked headers top-down, left to right, and don't treat the grouping row as a data row. When a header cell is blank, assume it inherits the cell to its left until the document says otherwise β€” and note the assumption.

Right-aligned text in a column is a useful signal. Numbers are conventionally right-aligned and labels left-aligned, so a right-aligned column under a label header usually means the header is mislabelled or the cells are codes. That mismatch is worth a sentence in your notes, because whoever built the table may need to fix it.

Work Out What Each Row Represents and How Rows Are Ordered

The row labels tell you what is being counted, but the ordering tells you how to read the table. Rows are commonly sorted by value (largest first), alphabetically, or by a fixed category order that does not change between editions. Sorting by value is the trap: it makes rows look rankable when they are just ordered for readability, and the ranking can flip in next year's edition.

Two row-label conventions catch readers. First, "Total" or "All" rows sit at the bottom or top and duplicate the sum of the rows above β€” they are not extra data, and including a total row in an average of rows double-counts everything. Second, indented sub-rows belong to the row above and should be read as a breakdown, not as peers.

Row labels also carry definitions. "Small businesses" might mean firms with fewer than 50 employees in one table and fewer than 250 in another, and the label rarely spells this out. The footnote or the surrounding text usually holds the cut-off; find it before you compare the row to anything outside the table.

Row order matters for comparison across tables. If Table 4 lists categories alphabetically and Table 7 lists them by size, a row-for-row comparison between them is meaningless. Line up the labels first, then compare the values.

Check Units, Scale, and the Value of a Missing Cell

Units live in three places: the column header, a scale note under the header (such as "Γ—1,000"), and the cell formatting. A column headed "Population (thousands)" with a cell of "48" means 48,000 people. Strip the scale and you have an error of three orders of magnitude.

Formatting carries meaning too. A cell showing "1,234" uses a thousands separator; a cell showing "1.234" in a European document may mean one thousand two hundred thirty-four, not one point two three four. Currency symbols, percent signs, and plus or minus signs can be applied by cell format rather than typed into the cell, so the underlying value may differ from what you see β€” check by clicking the cell in a spreadsheet.

Missing values are their own vocabulary. Common conventions are an en dash for "not applicable", a blank for "not available", and "0" for a genuine zero. Some documents use ".." for "not collected". Read the legend that defines these symbols; if there isn't one, treat every dash and blank as ambiguous and say so.

Symbol or format Usual meaning What to do
En dash (–) Not applicable to this row Don't treat as zero when summing
Blank cell Not available or not collected Check the legend; query the source
0 Genuine zero Include in sums and averages
".." Not collected Exclude from calculations
"c." before a figure Approximate or suppressed Don't quote to more precision than given

Negative values deserve the same care. A minus sign rendered as a hyphen is easy to miss in a dense grid, and parentheses around a number ("(412)") conventionally mean negative in financial tables. Both conventions can appear in the same document, so confirm which one the table uses.

Read One Row Across and One Column Down Before Trusting the Rest

The fastest way to confirm you understand a table's structure is to trace a single row and a single column. Take the first data row, read it left to right, and say out loud what each cell means. Then take a column, read it top to bottom, and do the same. If either pass produces a sentence you can't finish, the structure isn't clear yet.

Trace a total column next. A "Total" column should equal the sum of the year columns to its left, and a "Total" row should equal the sum of the rows above, within rounding. If the totals don't reconcile, you've found either a rounding convention, a suppressed value, or a genuine error in the document β€” and each demands a different response.

Reconciliation also exposes hidden exclusions. A total that falls short of the visible rows usually means some rows were omitted for confidentiality or because they fell below a reporting threshold. Note the gap; don't distribute it silently across the visible rows.

Do this check on a working copy of the numbers, not in your head. Copy the row into a scratch spreadsheet, sum it, and compare. The arithmetic takes a minute and catches the errors that survive a visual scan.

Use the Footnotes Before You Interpret Anything

Footnotes carry the definitions the table can't fit. The usual suspects are population bases, rounding rules, break points in a time series, and revised figures. A note reading "figures from 2020 onward include a methodology change" means the series is not continuous, and a year-on-year comparison across that boundary is invalid.

Rounding notes change how you should read the cells. If the table states "figures rounded to the nearest thousand", a cell of "12" covers anything from 11,500 to 12,499. Two adjacent cells reading "12" and "13" may therefore differ by almost nothing or by nearly two thousand, and you cannot tell which without the unrounded source.

Revision notes matter for anything you quote publicly. Statistical agencies routinely revise earlier periods, and a table published in one edition can show different figures for the same year in the next. Record the edition and the revision date alongside the number, not just the year it refers to.

Locate the footnote markers before you read the data. Markers are usually superscript letters or numbers attached to a specific header or cell, and they can hide under a column header where the eye skips. Scan the headers first, collect every marker, then read the notes at the bottom.

Compare Two Rows or Two Editions Without Being Fooled

Comparisons fail for four predictable reasons: different units, different bases, different definitions, and different vintages. A percentage change and a percentage-point change are different quantities β€” going from 4% to 6% is a two-percentage-point rise and a 50% increase, and both figures are correct for their own question.

Index numbers need their base year. A series set to 100 in 2015 and a series set to 100 in 2020 cannot be compared directly; you must re-base one to match the other. The base year is normally stated in the title or a footnote, and it changes between editions.

Check whether the comparison is like-for-like before you compute anything. If one row covers full-time employees and another covers all employees, the difference in averages partly measures the base, not the phenomenon. Write down the base for each row, then compare.

For two editions of the same table, align the rows by label rather than by position. A category added in the new edition shifts every row below it, so a positional comparison pairs the wrong labels and produces a fake change. Match labels, then match values.

Pull Numbers Into a Calculation Without Losing Precision

Quoting a table cell means deciding how much precision to carry forward. Carry the unrounded figure through intermediate calculations and round only the final result, or your total can drift from the table's own total by a visible amount. A competent reader notes the rounding rule and applies it once, at the end.

Percentages computed from rounded cells inherit the rounding. Two cells of "3" and "7" (rounded to the nearest whole number) could represent 2.6 and 7.4, making the true share 26% rather than the 30% you get from the rounded cells. When a document publishes both counts and percentages, use the published percentage or the unpublished counts, not your own division of the rounded counts.

Average-of-averages is the classic error. Averaging the averages of unequal groups gives a number that belongs to no group. If you need an overall average, go back to the underlying counts and weights where the table provides them, and state the weights you used.

Keep a short audit trail alongside any extracted figure: table title, edition, row label, column header, and unit. That record answers the question a reviewer will ask, and it lets you find the number again when the document is revised.

Adapt the Method to PDFs, Word Tables, and Spreadsheets

Screen reading adds failure modes that paper does not have. In a PDF, a visually clean table can be an image, a set of text boxes, or an untagged structure that screen readers and copy-paste routines cannot follow in order. Copying a row from such a PDF can paste cells in reading order rather than visual order, so check any pasted row against the visible grid before you use it.

In a Word or Google Docs table, headers may not repeat across page breaks, and a row can split across two pages so that the top half of a row appears under one page's headers and the bottom half under the next. Turn on the header-row repeat setting if you are editing, and if you are reading, treat a split row as a single row.

Spreadsheets introduce hidden columns, filtered rows, and formatted cells whose displayed value differs from the stored value. Before you read a spreadsheet table, check for filter arrows, unhide all columns, and click a sample cell to see the formula bar value. A filtered view can make a total look small because the visible rows exclude the rest.

Print and photocopy add their own damage. Faint gridlines disappear, a minus sign scans as a speck, and a footnote marker becomes a smudge. If you are reading from a scan, verify any ambiguous cell against a second copy or an original digital file, and flag what you could not verify.

FAQ

What is the first thing to read in a table?

Read the title and its source line first, because they fix the statistic, the unit, and the vintage before any number can mislead you. A title containing "average", "per household", and a year does more interpretive work than the column headers do. Skipping it means you can read every cell correctly and still answer the wrong question. Write those three facts down before moving to the headers.

Are blank cells the same as zero in a table?

No β€” blank cells conventionally mean "not available" or "not collected", while a printed 0 means a genuine zero, and mixing them changes sums and averages. Most statistical documents define their symbols in a legend, often beneath the table or at the front of the publication. If no legend exists, treat blanks as unknown and exclude them from calculations rather than assuming zero. State the treatment in your notes so the omission is visible to whoever reads your work.

How do I read a table spread across two pages in a PDF?

Find the repeated header row, then match columns by label rather than by position, because a two-page table often shifts a column sideways. If headers don't repeat, reconstruct them from the first page and verify against a cell whose value you can check independently. Note any row that splits across the break and read it as one row. Where the structure cannot be reconstructed, request the original spreadsheet from the author instead of guessing.

What is the difference between a percentage change and a percentage point change?

A percentage-point change is the simple arithmetic difference between two percentages, while a percentage change is that difference divided by the starting value. Moving from 4% to 6% is 2 percentage points and a 50% increase, and both descriptions can appear in the same report. Pick the one that matches the question being asked and name which you used. Mixing them overstates small changes and understates large ones.

Why don't the rows in a table add up to the total shown?

Rows usually fail to add to the total because of rounding, suppression, or an "other" category that isn't shown as a row. Rounding alone can create a gap of a few units across a long column, and confidentiality rules can remove a row without leaving a visible marker. Check the footnotes for both before treating it as an error. If neither explains the gap, report the discrepancy rather than silently adjusting the figures.

Can I compare two tables published in different years?

Only after checking definitions, units, and base years, because statistical tables are revised and redefined between editions. A category added in a later edition shifts row positions, so match rows by label, not by where they sit. Where the methodology changed, the footnote usually dates the break, and comparisons across it need a caveat. If the definitions can't be reconciled from the published notes, contact the publishing body rather than assuming continuity.

What does an en dash mean in a table cell?

An en dash conventionally means "not applicable" β€” the category doesn't exist for that row β€” rather than zero or a missing observation. Some documents use a blank for "not available" and a dash for "not applicable", which is a distinction worth preserving in any figure you extract. Check the legend for the document's own convention, since usage varies between publishers. Treating a dash as zero is the most common extraction error in financial and statistical tables.

How much precision should I carry from a table into a calculation?

Carry the full published value through intermediate steps and round only at the end, or you can drift from the table's own totals. Where the table shows only rounded figures, remember that a cell of "12" may stand for anything from 11.5 to 12.49, so derived percentages inherit that uncertainty. Publish your result to the same precision as the source, not more. Quoting four decimal places from a table rounded to whole numbers implies accuracy the table never claimed.