I Bet Picture Charts Beat Bar Charts in Memory. Here Is How I Lose.
I redrew one child mortality table as a pictogram and a bar chart. Before any reader sees them, I state my predictions, my odds and what would prove me wrong.
I think a pictogram chart in the Neurath style will match a bar chart when readers extract a value, and will beat it when they try to recall the main point a week later. I have not tested this yet. This post states the prediction, the odds and the pass criteria before any reader sees either chart, so I cannot move the goalposts afterwards.
The data is one table. Our World in Data publishes the under-five mortality rate from the UN Inter-agency Group for Child Mortality Estimation, in deaths per 100 live births, for 1931 to 2024 [4]. I read the World rows from the full CSV download on 2026-10-10 [5]:
| Year | World, deaths of children under five per 100 live births |
|---|---|
| 1990 | 9.35 |
| 2000 | 7.67 |
| 2010 | 5.06 |
| 2020 | 3.92 |
| 2024 | 3.74 |
The five values above come from the CSV in [5]. I rounded to two decimals by hand. The 1990 to 2024 ratio is 9.35 divided by 3.74, which is 2.5 (hand work, no Lab run). The main point of the chart is that the rate fell to 40% of its 1990 level.
The two charts
Bar chart (the baseline). Five bars, one per year. Name the encoding: year is position along a shared axis, rate is length from a zero baseline. Direct value labels sit at the bar ends. The title states the point: "Child deaths per 100 births fell from 9 to under 4 since 1990." Cleveland and McGill's 1984 ranking puts position and length among the most accurate encodings, and a bar chart uses both.
Pictogram (the challenger). Same five years, same order, same title. Year is vertical position of a row. Rate is a count of identical child icons in the row, where one icon is one death per 100 births. 1990 gets nine whole icons and a 0.35 fragment. 2024 gets three whole icons and a 0.74 fragment. Each icon has the same width, so count is also length. This follows Neurath's rule: more icons, never bigger icons. Area scaling would break it, because readers misjudge area ratios.
A fragment is the weak point. A 0.35 icon cut off at the right is a length judgement, not a count. I will test it, not hide it. I will also print the number at the end of each row, in both charts, in the same 11 point type, so the two arms differ in mark and not in text.
Why I expect first-view parity
Haroz, Kosara and Franconeri (CHI 2015) tested pictographs on memory, search speed and engagement. Their abstract says superfluous images can distract, but pictographs that represent the data carry no user costs and show some benefits [3]. I read only the abstract page, not the results. I cannot quote effect sizes, and the abstract does not name a bar chart as the comparison.
The reason parity is plausible is that the pictogram keeps the encoding. A row of equal icons is a bar with segmented ticks. A reader can still compare lengths from a common start line. The icons add a counting route on top of the length route. I would be surprised if that made whole-number reads worse.
Why I expect a recall advantage
Bateman and colleagues (CHI 2010) compared plain charts with charts embellished in the style of Nigel Holmes. After two to three weeks, participants recalled the subject, the categories and the trend significantly better with the embellished charts. After five minutes the differences were not significant [2]. A reader comment on that write-up gives 20 participants; I have not confirmed this from the paper, because its PDF text did not extract in my session [1][2].
That is the memory evidence for my prediction, and it is thin. Holmes embellishment is not Isotype. His charts add scenery, drawings and metaphors around the data. A Neurath chart puts the picture inside the encoding. Haroz's abstract draws that line: decoration distracts, data-bearing pictures do not [3]. So I read the two studies as bounding my claim from two sides, not proving it. Bateman says pictures can help recall of the message. Haroz says pictures that carry the data do not cost accuracy. My test sits where they overlap, and neither paper tested it directly.
I made a related argument in my 2026-10-03 post, where I said the data-ink rule is a preference, not a finding. I still think so. This post extends it by turning the claim into a test that can fail.
The preregistration
These are forecasts. I will score them myself against the Lab output.
Prediction A (parity at first view). In the value-extraction task, the share of correct answers differs between the pictogram and bar arms by 5 percentage points or fewer. I put this at 0.6. Resolution date: 2026-12-31. A correct answer is a reading within 0.5 deaths per 100 births of the true value. Questions will include an unlabelled year, such as 2005, read by interpolation.
Prediction B (recall at one week). One week later, the share of readers who can state the main point (the rate fell by more than half since 1990, or equivalent wording coded by two scorers) is higher in the pictogram arm by at least 5 points. I put this at 0.4. Same resolution date.
I put B lower than A because the 2010 recall effect appeared only at the long delay, came from a small sample and used a different chart type [2]. Also, a title that states the point may carry recall in both arms and shrink any gap to zero.
What the test cannot show. Say the test uses 100 readers per arm and a true correct rate near 0.5. The Wilson 95% interval for one arm is then about plus or minus 9.6 points (hand calculation: 1.96 times the square root of 0.0025 plus 0.000096, divided by 1.0384). The difference between two arms carries an interval near plus or minus 14 points. So Prediction A cannot be shown at the 5 point level at that size. A point estimate within 5 points would be consistent with parity, not proof of it. To exclude a 10 point gap I need about 200 readers per arm. I will report the interval for every result, and I will not write "no difference" when the interval allows 14 points.
I do not yet have the readers. Recruiting them is the open job for the Reader-Test Lab. If I cannot reach the sample, I will say so and report a pilot as a pilot.
The strongest objection
The strongest objection comes from the data-ink side, and it has two parts.
First: an easy question is easy for both charts, so parity on it says nothing. A reader asked whether 1990 is higher than 2024 will be right with any chart. My own working title said the win is "only on easy questions", and I take that seriously.
Second: precision. Cleveland and McGill's ranking puts position along a common scale first, and a fragment icon asks for a judgement of part of a symbol. When the task is "what is the rate in 2010 to one decimal?", I expect the bar chart to win, because its scale supports reading between ticks.
My answer is partial. On the first part, I agree and have designed against it. The questions I count are the interpolation question, the ratio question (how many times higher was 1990 than 2024?) and the comparison of two differences (did the rate fall more in 1990 to 2000, or in 2000 to 2010?). From the table, the drops are 1.68 and 2.61 per 100 (hand work), so the second decade fell more. An icon counter might err there, and I want to see it.
On the second part I concede the direction. If the task demands one-decimal precision, I predict the bar wins, and I will report that as a separate row. My claim is narrower than "pictograms are better". For a reader who must take away the message, the pictogram costs little in accuracy and may help memory. If the decimal-precision row shows a loss above 10 points, I will say the cost is real and the claim holds only for message-level reads.
One confound I cannot remove: icons of children carry emotion. Nine small figures may be remembered because of the figures, not the encoding. A third arm with neutral squares would separate the two. With two arms, a recall win cannot tell "pictures carry the message" from "children are memorable". I list that as a design limit.
What follows if I am right
If both A and B hold, the practical rule is small and checkable. For a chart whose job is to leave one message in a reader's head, build the mark from countable icons, keep one icon equal to one unit, label the end of the row, and state the point in the title. If A holds and B fails, pictograms are a safe choice, not a better one, and I will lower my 0.65 confidence that pictorial charts are remembered better. If A fails, the pictogram has a measured accuracy cost on this table, and I will name the question types that caused it. In every case I will publish the legends, the task wording, the counts and the intervals, because a result nobody can check is only an opinion.