Charting change over time
Time series is the most common thing anyone charts and the easiest to do subtly wrong. The errors are rarely in the data — they are in the axis, the spacing, the stacking, and the second axis nobody labelled.
What a connecting line claims
Drawing a line between two points asserts that intermediate values exist and follow that path. For a quantity measured over time this is usually true: revenue existed between March and April even if you only recorded the month ends, so interpolation is a fair representation of a real underlying continuity.
For separate categories it is false. There is nothing between Product A and Product B, and a slope drawn there implies a trend across an ordering that is arbitrary. Reorder the categories and the trend changes, which is a reliable test — if reordering the X axis changes the story, a line chart is the wrong format.
The same claim applies to spacing. Most charting tools, including these, treat the category column as evenly spaced labels rather than as true dates. When your rows are regular — every month, every day — that is equivalent to a time axis and simpler. When they are irregular, with three readings in one week and the next a quarter later, even spacing flattens the gap and misrepresents the rate of change. The remedy is to include the empty periods as blank rows so the spacing carries the correct information.
Blank cells within a series should render as gaps rather than as zeros. A missing measurement and a measurement of zero are different claims, and a chart that plunges to zero for a month you failed to record is showing a data collection failure as a business event.
The baseline argument
Whether the Y axis must include zero is the most litigated question in charting, and the answer differs by chart type for a reason that is not arbitrary.
Bars encode value as length from a baseline. If the baseline is not zero, the ratio between two bar lengths does not match the ratio between the values, and a bar twice as tall no longer means twice as much. Values of 100, 102, and 104 drawn from a baseline of 98 produce bars in a 1:2:3 ratio for a 4 percent spread. That is a misrepresentation regardless of what the axis label says.
Lines encode value as position, and the reader is being asked about the shape of the trend rather than the ratio of two lengths. A series oscillating between 98 and 102 has real structure that a zero baseline compresses into a flat band, hiding the only thing worth seeing. Truncation here is not a distortion, it is a zoom.
The honest practice is to truncate line charts when the variation is small relative to the level, and to say so. A reader who notices an unlabelled non-zero baseline discounts the entire figure, and reasonably so. Where a chart is going to be read quickly by people who will not check, adding the baseline value to the axis label costs nothing.
What stacking gives and takes
Stacking series into an area chart buys you the total for free: the top edge of the stack is the sum, which is often the number people most want. It also shows each component contribution as a band thickness.
What it costs is the readability of every series except the bottom one. Only the bottom band has a fixed baseline; every other band floats on the one below it, so a component with constant thickness appears to rise and fall with whatever is underneath. Human vision compounds this by reading the perpendicular distance between two curves rather than the vertical distance, which understates thickness wherever the bands are steep.
Two rules follow. Put the largest and most stable series at the bottom, where the distortion is smallest. And keep the band count to five or fewer, grouping the remainder into an Other band, because thin bands between two moving curves are unreadable at any colour.
Stacking is only valid when the components sum to something meaningful. Three independent metrics stacked together produce a top edge that is the sum of things that should not be added, and readers will interpret it as a total. Independent series belong on a line chart.
The 100 percent variant, which normalises each period to full height, makes composition shifts very clear and discards the total entirely. A channel holding a steady 30 percent share looks identical whether the business tripled or halved, so a normalised chart shown without a total alongside it can make a collapse look like stability.
Mixing encodings, and the second axis
A combo chart — bars and lines together — earns its complexity when the two series are different kinds of quantity. Revenue as bars and margin percentage as a line works because one accumulates and the other does not, and the differing encodings tell the reader not to compare them directly. Actual against target works for the same reason.
Two series of the same kind should share an encoding. Splitting two revenue figures across bars and a line implies a distinction that does not exist, and readers will spend effort looking for it.
Mixing a quantity with a rate usually forces a second axis, which is where the format attracts criticism. The relationship between the two scales is arbitrary: whoever sets the ranges decides where the bars and the line appear to cross, which point looks like a turning point, and whether the two series look correlated. The same data with different bounds supports different conclusions, and none of them is more correct.
That does not make dual axes unusable. Label both axes with units, avoid implying that a crossing point is meaningful, and use a single axis wherever the ranges are compatible. Where the chart is going into a document rather than a slide, two stacked charts sharing an X axis remove the ambiguity completely for the cost of some vertical space.
One constraint carries over regardless: bars still need a zero baseline. This is frequently what forces the second axis in the first place, since the line series may need a truncated range to be legible while the bars cannot have one.
How many series before it stops working
Legibility falls off sharply above five or six lines. Beyond that, lines cross often enough that following one requires constant reference to the legend, and colour stops being a reliable distinguisher — roughly 8 percent of men have a red-green colour vision deficiency, and a palette that varies only in hue fails for them entirely.
Two approaches scale better than more colours. Highlighting one or two series in strong colours against a light grey background of the rest keeps the context and makes the comparison immediate. Small multiples — several small charts on a shared scale, one series each — can be scanned far faster than a tangle of overlapping lines, and they preserve every series rather than de-emphasising most of them.
Direct labelling at the end of each line beats a legend whenever the labels fit, because it removes a lookup step from every single reading. A legend is a fallback for when the lines are too dense to label, which is also the point at which you should be considering small multiples.
Frequently asked questions
When should I use a line chart instead of bars?
When the X axis is continuous. If reordering the categories would change the story, the sequence is arbitrary and a line is the wrong encoding.
Does the Y axis have to start at zero?
For bars yes, because length encodes value. For lines no, because position encodes value and the reader is judging shape.
Why does a flat series look volatile in a stacked area chart?
Its baseline is the top of the band below it, so it tracks whatever is underneath. Only the bottom series has a fixed baseline.
Are dual axes acceptable?
With care. The relationship between the scales is arbitrary, so label both and avoid implying that where the series cross is meaningful.
How many lines can one chart carry?
Five or six. Beyond that, highlight one or two against grey, or split into small multiples on a shared scale.