Dynamite plot

Dynamite plot is a bar plot with error bars. And it is rarely the best choice.

It hides

Rafael Irizarry wrote an open letter to journal editors titled “Dynamite Plots Must Die”. His main argument: dynamite plots hide details—individual data points and the underlying distribution. By employing scatter-like plots, density estimates, and box plots (or some combinations of these), we can reveal a lot more details about the data.

This is an excellent point and whenever there’s an opportunity to show more details about underlying data distribution or individual data points, you should try to show them!

Dot (point-range) plot is usually better

Even when we are ignoring the issue of hiding, say we have plenty of underlying data points that are following perfect normal distributions beautifully, dynamite plot is still not ideal in many situations due to its physical semantic. The information that a dynamite plot conveys is essentially a point estimate and error. For this, a simple dot plot (or point-range plot) suffices; it is often cleaner and free of semantic issues when dealing with non-zero baselines or nonlinear scales.

For instance, it sometimes doesn’t make sense to start the scale at zero. Bars mislead in such cases because bars carry a visual metaphor of physical accumulation—like liquid filling a container from the bottom. This works when zero is meaningful (counts, amounts, durations), but when our axis doesn’t start from zero¹, it is visually misleading. Points are free of such visual metaphor.

Another issue arises with nonlinear axes. When plotting on a log scale, bars are again misleading due to its physical semantics. On a log scale,

• There’s no natural “ground”—log(0) is undefined
• What matters is the ratio between values, not distance from some baseline
• A bar starting from 1 (or 10, or wherever) implies that baseline is special, when it isn’t.

In the titer example, the vaccine group is 8× higher than placebo. A point-range plot shows this directly; bars obscure it by drawing attention to the arbitrary baseline.

Almost always, replacing bars with points makes the plot simpler, neater, and less misleading. Additionally, because point-range is minimal, it can be more easily stacked or combined with other plots.