I believe your last sentence is the most important one. You want consistency in publications.

+1 for your reasons. Also, the y-axis being lower than zero is not just a cosmetic problem, but also shows some statistical assumption is wrong with the error bars. Most likely, non-parametric or a non-normal distribution is needed for the error bars.

@ Richard Erickson, would you mind elaborating a bit more? These error bars are just standard errors (accounting uncertainty in flux estimation (curve fitting)), in this sense I think error bar extending into the negative is definitely possible.

@ychung Nitrogen flux cannot be negative, correct? If your Nitrogen flux can be negative, then you're good to go. If N flux should only be positive, the Normal distribution is wrong for your data.

The standard errors making sense as error bars in this sort of plot assumes that the errors around the curve fit are normally distributed. Now it maybe the case that the errors in your curve fit tend to normality via the central limit theorm in the case of a high enough N, but clearly that estimate isn't perfect here - as these error bars are suggesting that, for example, on the 13th of September, the flux could have conceivably been -100.

@ychung see this post on another SE site for more details: mathoverflow.net/questions/4840/…

Isn't it amazing that in 2020, Microsoft manages to produce one graph for which one can list several paragraphs of reasons why it's bad :-(

@WolfgangBangerth For all the unique problems with Excel, an ability to produce graphs deserving paragraphs of criticism is certainly not one of those unique ones. :)

@ychung I'd consult with a colleague with some statistical expertise in your field, but naively I'd suggest you consider a log transformation of your data. Logarithms are often great for measurements that can't feasibly be zero or negative and that you would feel comfortable describing in terms of ratio changes. If "half" is basically the same magnitude change as "doubling" - you should probably be on a log scale. I'm not familiar with your research area, though, so these descriptions might not apply to your data. Good luck!

As for what to use instead - R/ggplot (which is free) and matlab are all well and good if you are already a coder - and if you are not it might be a good excuse to learn. However, you might find that intimidating or too much effort for the reward. In which case I recommend having a look at GraphPad Prism. Its expensive, but if you are already at a university, you might find they have a site licence already.

@RichardErickson I assume these are N2O fluxes from soil (to atmosphere). That happens to be my area of research and they can be negative (but possible magnitude of negative fluxes is limited). I'd present these error bars as shown (but would never use Excel for this).

@WolfgangBangerth Excel is not optimized for academia! There is nothing wrong with this kind of chart in a business environment (except that many people won't understand error bars)

@Roland Thank you for providing the domain specific answer to my question.

@Borgh: Surely there is no distinction between academia and business in this regard -- good figures are ways to communicate meaning without confusing the viewer unnecessarily. I can't see why the environment in which a figure is used has any implication on this as a universal truth.

@WolfgangBangerth in boardrooms it is far more important too make your report look pretty and to have it scale nicely into powerpoint. Graphs like these exist solely to emphasise a point usually made in text or speech, they never need to stand alone.

@borgh: That means that there is absolutely no difference to any other use of graphics -- to support a point made in text or speech. There is zero excuse for using poor graphics in any context.

Heh .. instantly reminded me of my studies and using those graphs .. those unfitting horizontal grid lines, why fit them at 50, 150, 250 and so on, not the far more reasonable 0, 100, 200 - Yes, I know that they weren't put there by you, but by Excel instead, which is why I hate the Excel automatism - the different distance from 0 to the first horizontals , compared to the distance between the other horizontals make it appear primitive and uneven.. and as already mentioned - doesn't show much effort

@IanSudbery The Excel default error bars are worse that that! They assume that all variation in your data is due to error. Try plotting 1,2,3,4,5 vs 1,2,3,4,5 (i.e. an errorless straight line) then turn on the error bars and you'll find the result is stddev(1,2,3,4,5). It's hard to think of an example when they'd actually be right.