I’m revising my statistics fundamentals and I’m getting tangled up with some basics about bar charts. I keep mixing up rules with histograms, so I want to clear this up carefully.
– Do bar charts always need the y-axis to start at zero? If not, when is it acceptable to start above zero, and how should that be shown so it isn’t misleading?
– Should all bars in a standard bar chart be the same width with equal spacing? If the category labels are long or uneven, does bar width ever encode anything, or should it be purely cosmetic?
– When I’m comparing two groups that have different total sample sizes, is it better to use raw counts in side-by-side bars or convert to percentages? I’m not sure which choice leads to fairer comparisons.
– For stacked bars, what’s the right way to read and compare subcategories? I feel like I misinterpret the middle segments.
– Finally, can you confirm whether, in a bar chart, it’s the height (not the area) that represents the value? If widths differ for any reason, does that make the graphic misleading, and should I avoid that entirely?
I’m trying to strengthen my fundamentals and stop making the same mistakes. A step-by-step explanation of what’s considered good practice for plain bar charts would really help. I don’t need the answer worked out on data-just the principles and reasoning, please.
Ready to make maths more enjoyable, accessible, and fun? Join a friendly community where you can explore puzzles, ask questions, track your progress, and learn at your own pace.
By becoming a member, you unlock:
3 Responses
Short answer with my “don’t-trip-over-histograms” hat on: in a plain bar chart, the value is shown by the bar’s height, and the bars are usually equal width with equal spacing. I still mix this up with histograms too-histograms care about area when bin widths change, but bar charts don’t. For the y-axis, I was taught “start at zero” because truncating the axis exaggerates differences in lengths; it’s acceptable to start above zero if the story is about small deviations and you clearly show a break and label it loudly. I once plotted average commute times from 34 to 38 minutes with the axis starting at 30-the bars looked like a transportation apocalypse. My friend gently asked if the buses were on fire or if I’d just chopped the axis. Lesson learned. On bar widths: keep them the same so people read height, not area. If a label is long, I’ve widened a bar to make room and it didn’t “break” the chart, but I think that risks people subconsciously reading area, so it’s more of a cosmetic hack than good practice. I’ve also seen folks use width to hint at sample size, but that blurs into mosaic territory and can confuse the “height = value” rule, so I try not to.
Comparing groups with different totals: I usually switch to percentages for fairness, especially if the group sizes are far apart; if totals are pretty similar, raw counts can be fine (I sometimes annotate the n so no one has to guess). For stacked bars, read components from a shared baseline: the bottom segment is easy to compare across stacks, and the total height is easy, but the middle bits are notoriously slippery because they’re floating. If the focus is on composition, 100% (normalized) stacked bars help; if the focus is on the parts themselves, small multiples or side-by-side grouped bars are kinder to our eyeballs. When widths differ for any reason, it can nudge people toward reading area instead of height, so while it technically “still works,” it drifts into misleading territory and I avoid it unless I’m deliberately doing a different chart type. I learned this the hard way making a stacked bar of snack choices at a party-my “chips” bar was wider to fit the word “tortilla,” and I convinced myself chips dominated the world. Turns out I’d just dominated the formatting. I might be overthinking it, but these little design choices really do change how people read the numbers.
Bar charts show values by length, so the y‑axis should start at zero. Truncating the axis inflates small differences and breaks the length comparison people naturally make. The main exception is a “diverging” bar chart centred on a meaningful baseline (e.g., zero net change), where bars extend both up and down from that baseline. In a standard bar chart, keep all bars the same width with equal spacing; width does not encode data. If labels are long, rotate them, wrap text, or switch to horizontal bars. This differs from histograms: histograms are for continuous data with touching bins, and if bin widths vary, area (not height) represents frequency. In ordinary bar charts, it’s height only; varying width without changing the interpretation is misleading unless you are deliberately using a different chart type (e.g., a mosaic/Marimekko, where area encodes value).
When groups have different totals, use percentages or rates to compare prevalence fairly; use counts when absolute quantities matter. A good compromise is to plot percentages side‑by‑side and label each group’s sample size. For stacked bars, you can read the total and you can compare the segment that sits on the common baseline; middle and top segments are hard to compare across categories because they don’t share an aligned baseline. If the goal is to compare subcategories across groups, prefer grouped (side‑by‑side) bars or a small multiple of simple bars; use 100% stacked bars only to show composition, not to compare the sizes of inner segments.
Simple example: Group A has 600 people with 120 successes; Group B has 200 people with 80 successes. Counts give bars of height 120 vs 80, suggesting A is larger. Rates are 20% vs 40%, showing B has the higher prevalence. A fair comparison of success rates is two bars at 20% and 40%, both drawn from a zero baseline, with equal widths, and with “A (n=600), B (n=200)” noted on the labels.
Short version while I wave a tiny soapbox: for plain bar charts, the y-axis should almost always start at zero (since bar length is the cue), though starting above zero can be okay if you clearly show a break and label values; bars should be equal width/spacing because width doesn’t encode anything here (that’s a histogram thing), but I’ve occasionally widened a couple for long labels-probably not best practice, I admit.
For unequal sample sizes, use percentages for fair comparisons (counts only if totals themselves matter); in stacked bars, compare the baseline-aligned segments (the bottom) or use 100% stacks, and yes, height not area represents value-unequal widths can mislead unless you purposely use width for sample size, which I’m slightly unsure about.