I appear to finish up ceaselessly explaining to college students and colleagues that it’s a good suggestion to spend a great deal of time to make your scientific figures essentially the most informative and engaging doable.
Nevertheless it’s a tremendous stability between overly flashy and downright boring. For sure, empirical accuracy is paramount.
However why do you have to care, so long as the mandatory info is transferred to the reader? Crucial reply to that query is that you’re attempting to be a focus for editors, reviewers, and readers alike in a extremely aggressive sea of data. Certain, if the work is nice and the paper well-written, you’ll nonetheless garner a readership; nevertheless, for those who give your readers a little bit of visible pleasure within the course of, they’re more likely to (a) keep in mind and (b) cite your paper.
I attempt to ask myself the next when making a determine — with out pointless bells and whistles, would I current this determine in a presentation to a bunch of colleagues? Would I current it to an viewers of non-experts? Would I need this determine to seem in a information article about my work? After all, all of those venues require differing levels of accuracy, complexity, and aesthetics, however a great determine ought to ideally serve to coach throughout very totally different audiences concurrently.
A sub-question price asking right here is whether or not you suppose a colleague would use your determine in one in every of their displays. Consider the final time you made a presentation and located that excellent determine that brilliantly portrays the purpose you are attempting to get throughout. That’s the sort of determine it’s best to attempt to make in your personal analysis papers.
I subsequently are inclined to spend fairly a little bit of time crafting my figures, and after years of creating errors and getting just a few issues proper, and retrospectively discovering which figures seem to garner extra consideration than others, I can supply some primary recommendation in regards to the DOs and DON’Ts of determine making. All through the next part I present some examples from my very own papers that I feel exhibit among the ideas.
tables vs. graphs — The very first query it’s best to ask your self is whether or not you may flip that boring and ugly desk right into a graph of some type. Do you really want that desk? Are you able to not simply translate the cell entries right into a bar/column/xy plot? Should you can, it’s best to. When a desk can’t simply be translated right into a determine, more often than not it in all probability belongs within the Supplementary Info anyway.
white area — White area is a type of features that you don’t essentially realise is the rationale you don’t just like the look of a selected graph. In case your axis scales are such that a lot of the information seem at one excessive, in case your panels have enormous gaps between them (see subsequent entry), or there’s only a massive gap someplace within the determine, you want to rethink the configuration of the knowledge. You are able to do numerous issues to take away white area, together with shifting elements nearer collectively, or including icons (see beneath), altering axis scales (see additionally beneath). A pleasant, tight (however not too cluttered) determine is way more visually interesting than one the place massive white holes distract your consideration.
panels — In case your figures look cluttered at one excessive, or a bit bare on the different, it’s time to think about multi-panel plots. Such plots assist you to put a variety of info in a single determine, offered you don’t attempt to swamp your reader with all the pieces and the kitchen sink in a single go. Some suggestions for good multi-panel figures embody: avoiding panel titles (see extra beneath; panel letters or numbers of adequate measurement often, however not at all times suffice), standardising panel measurement, avoiding repetition of axis labels and titles amongst panels (see extra beneath), and standardised axis scales (the place doable).
titles — Determine or panel titles are often pointless and distracting, however you’ll need to embody a straightforward strategy to determine what totally different symbols/traces/colors point out by way of a legend, and naturally, an in depth follow-up clarification within the caption. Easy letters, numbers, or symbols for sub-components usually do the trick and keep away from cluttering the determine with an excessive amount of annotation.
captions — Talking of captions, the age-old advice {that a} determine needs to be stand-alone actually comes into play when crafting a determine. Can informal observer skimming via your paper perceive the which means based mostly on the determine and caption collectively, or are they required to learn your entire textual content to get it? If the latter, your determine will not be stand-alone and needs to be fleshed out somewhat extra.
abbreviations — aside from panel indicators, I have a tendency to not use abbreviations/acronyms/initialisms in my graphs for the easy motive that it’s not speedy obvious what they imply. I detest these kinds in just about all scientific work anyway, so I additionally advise holding them out of your figures (my Australian state abbreviations proven beneath however 😉 ).
keep away from repeating labels — As talked about above, keep away from repeating labels and titles amongst axes which might be the identical in (often) multi-panel plots. If the axis scale is identical throughout, say, the rows of panels, then all you want is the title and labels on the primary panel on the left, with all subsequent panels merely repeating the axis ticks. The identical applies within the x axis for columns of panels. Not solely does this simplify the design, it additionally saves an enormous quantity of white area.
to log or not log — Usually, a pleasant logarithmic (or different) transformation of an axis can tighten up the show and render a wonky distribution extra visually interesting. It might additionally do away with pointless white area. Nevertheless, bear in mind that any transformation adjustments the graph’s interpretation, in order that you need to be very clear what the pattern signifies.
axis segments — In reference to transformations, in case you are involved about deceptive interpretation, or a metamorphosis fails to resolve the white-space downside, a segmented axis can produce a way more interesting determine. Say 90% of your information fall between 1 and 10, however you may have just a few information within the 100s or 1000s. Breaking the axis up so that the majority of it refers back to the 1:10 vary, with somewhat devoted to the acute values, can actually assist interpretation.
uncertainty — Do your pattern traces have any related uncertainty (e.g., normal deviations/errors)? Do your bars have measurement error? In case you have ANY related information errors, don’t simply present the central tendency. Add all uncertainty within the type of error bars, shaded uncertainty areas, and so on.
information distributions — Many journals lately require you to show all the info uncertainty in a plot, such that bar graphs with little T error bars are now not acceptable. Nice methods to show the info distribution is thru issues like boxplots, however even higher are violin plots now rising in reputation. If I’ve a distribution, I now often embody a all of the jittered information on high of the violin plot itself.
to 3D or not 3D — You’ve seen it on the telly hundreds of instances earlier than: a bar graph with a mysterious third dimension displaying ‘columns’ as a substitute of bars. Don’t do that. Until you may have a 3rd dimension in your information, don’t make one up. Three-dimensional graphs may look interesting, however they’re often empirically deceptive.
color — Within the not-too-distant previous, color was usually frowned upon for scientific papers, primarily on account of the price of reproducing color pictures in print. As of late that limitation is much less and fewer relevant, as a result of most publication is now on-line, and color prices not more than greyscale/black-and-white figures. That mentioned, don’t go loopy with colors. Many people are relatively color blind, and luckily, many colourblind-friendly color schemes at the moment are out there on most graphing purposes. The opposite motive too many colors might be distracting is that they don’t conform to any empirical symbolisation. In different phrases, do your totally different colors point out some component of the info (categorisation, origin, and so on.)? If not, hold them to a minimal. Simply in case somebody must print nonetheless lately, additionally take into consideration whether or not all the knowledge shall be retained in your color determine ought to somebody want to provide it in greyscale. If that proves difficult, rethink your color scheme.
borders — Usually I attempt to hold borders so simple as doable. There is no such thing as a want for a whole field in a bivariate plot, however a map usually has ‘boundary’ results (e.g., the sudden disappearance of a shoreline), which might be solved elegantly with a easy line border. Too many borders makes a determine look cumbersome and blocky. Too few can result in misinterpretation of parts aren’t simply separated upon first look.
font — Usually journals require any quantity/phrase fonts within the graph to be in step with the font of the principle textual content. If that’s the case, it’s best to observe their conference. If not, then a easy, interesting, but non-flashy font needs to be used for all determine parts (axis titles, legends, axis labels, and so on., and so on.). Don’t combine and match fonts on the identical determine.
are the info steady? — I usually see graphs the place single values (e.g., frequencies, discrete temporal values, and so on.) are joined by some kind of line, implying that you’ve got information between the discrete values. Should you don’t, don’t attempt to indicate a steady distribution between the adjoining classes. Select a format that shows the info most precisely. Alongside these identical traces, nice, bloody excessive bars from zero to the worth at hand are inclined to condense all the knowledge into one excessive of the graph. Right here, a degree is way more appropriate.
pointless capitalisation — I see this lots. Axis labels, axis titles, panel titles, and so on. with capitalised first phrases. It doesn’t assist that the majority purposes mechanically capitalise the primary phrase in a textual content field. Ask your self whether or not it’s a correct noun; if not, don’t capitalise. Most labels will not be the primary phrase of sentences, so standardise and hold your capitalisation just for the phrases requiring it.
icons/pictures — I discussed above that icons can typically that gaping white-space challenge. A cleverly positioned icon or simplified picture of the topic at hand can usually accompany a formidable graph and make it pleasing to peruse and reproduce. Once more, use with moderation, and check out to ensure your icons are high-resolution (in any other case, they have an inclination to look novice).
shading — Do you may have icons, arrows, and so on. that appear just a bit too boring? Usually a really delicate shadow can present somewhat perspective. However just like the 3D challenge, keep away from inferring an empirical dimension. One other highly effective use of shading (drop shadows, glows, and so on.) is to assist differentiate textual content from background element.
backgrounds — It’s typically tempting to incorporate a background color and even a picture behind your graph. This is usually a powerfully aesthetic element if carried out subtly, however actually distracting if carried out with out care.
use a number of purposes — I’ve but to search out the ‘excellent’ graphing software, so I have a tendency to make use of many on the identical time to provide the best-quality figures. The generic R plotting services are crap, though ggplot makes figures much more aesthetically pleasing (however requires much more coding know-how). Excel is loathesome for figures. I usually use R to provide the abstract info, then import the info right into a devoted graphing software (e.g., GraphPad Prism, and so on.), which I can then import right into a GIS software if I want to mix issues with maps. Or, I can produce subplots in a single software and mixture them in Powerpoint, or some such. The important thing right here is to be versatile, and ensure the ultimate output might be exported at excessive decision (vector or a minimum of 600 dpi).