Paulo,

I see the benefit of having readear that are "comfortable with" and "used to" barcharts.

But I don't subscribe to your claim "because they are very easy to understand". I'd rather say that they are easy to misunderstand. They are often harder to decipher, and the relevant information is obscured rather than highlighted.

There are surely some not-so-easy-to-interpret plots, especially, as you noted, for "non-IT users or non-data-visualization-experts". But usually the story behind the graph is not-so-easy, too. One must then find a way to visualize the relevant information is the most simple way possible (as always). Chosing an inappropriate chart type just because this is well-known (for totally different purpuses/data/stories) is usually not a good idea. The reader may then think he understands it, but he likely misses the essential information that actually should (but is not, or not recognizeably) be visualized.

And I am not talking about chart-junk. Charts that are unneccesarily difficult or use gimmicks that do not help to clarify the relevant information are bad charts anyway. But I think we agree here :)

Dear Veera Manikandan,

As I am just beginner in data analysis, I do not know much about using bargraphs for continuous data. However, the question you asked is quite useful for me and I hope good answers will follow soon.

Sailesh

Dear Veera, the paper is very pertinent; it deals with presenting continuous data in small sample size studies. This means bar graphs being not meaningful or even misleading whether the number of continuous data points is small - such as 2-6 - as mentioned somewhere in the paper. Also they are not reflecting the differences between groups of individuals whether data is paired. Again, it's about small data sets in the first place, secondly it is about the type of the data being analysed. 

I am not an expert in doing the data analysis, but what I recommend is you should know very well which kind of data you are going to analyse - discrete, continuous, mixed, qualitative, quantitative, paired, etc. Then set clearly your goal versus your hypothesis (why you have got those data, which hypothesis you started from - e.g., those values vary with time, age, experimental condition,...). I assume you finally would like to see if your findings have a statistical value or not, i.e. if your data set could be extrapolated from your sample group to a larger population. Once you know it precisely you may use a simple approach by yourself  starting from a tutorial such as http://statistics.about.com/od/HelpandTutorials/a/7-Common-Graphs-In-Statistics.htm, or you ask a bio-statistician for help. 

If you think of qPCR data, there will be a different story.  

The barchart is the most powerful technique to hide relevant information from the reader. The scatterplots show all the information and allow to judge the quality of the data and the adequacy of the conclusions, to spot possible problems or other interesting features or patterns. It is the very first and one of the most important steps of any data analysis to visualize the data, and these scatterplots are a convenient tool for that.

Barcharts are often phsychologically inappropriate because they focus on an "amount", where there exists no such amount and/or where the amount is not not of primary interest (example: any arbitrarily normalized measures; from my field of research I could name "relative protein or gene expression", ratio measures in general like fold-changes, fluorescence measures (often even presented in arbitrary units!), etc etc).

Even when the amount of data or the number of groups will make a scatterplot too busy or when the visualization of larger scale patterns is the aim of the plot, barcharts are generally the worst alternative. Firstly because of their inapproriateness, secondly because there are far better plots available to highlight the relevant information while using less ink (e.g. boxplots, dotplots and alikes).

My conclusion: barcharts are almost always the worst solution, although there might be some cases where they could be ok.

Btw: The linked paper presents a plot showing "paired data" that I find particularily bad. There are at least three much better alternatives: a) use a 2D scatterplot showing the pairs as coordinate-points, b) use a Bland-altman plot (also a 2D scatterplot showing the differences versus the averages within the paires), and c) a 1D scatterplot showing the pairwise differences directly.

I like combining both scatter and bar charts. However, PrismGraph way of doing this displeases me as they do not show the SD, SE or CI95% bellow the line representing the mean. This prevents from seen if means tend to overlap or not (I always use CI95%).

I therefore stopped using PrismGraph and coded my own graphs on STATA. Here is an example of such a code.

------------------------------------

Example from STATA (4 conditions)

------------------------------------

* Generate means, standard deviations, and count

egen mean = mean(outcome), by(group) 

egen n = count(outcome), by(group) 

egen sd = sd(outcome), by(group) 

* Compute confidence intervals

gen upper = mean + invttail(n-1,0.025) * sd / sqrt(n) 

gen lower = mean - invttail(n-1,0.025) * sd / sqrt(n)

* Generate scatter 

dotplot outcome, over(group) ny(50) nx(0) center ///

xlabel( 0 "Condition 1" 1 "Condition 2" 2 "Condition 3" 3 "Condition 4") ///

xtitle("Experimental conditions") scheme(lean2) ///

ytitle("Measure [unit]") ///

ylabel( 0 "0" 0.25 "25" 0.5 "50" 0.75 "75",nogrid) ///

legend(off)

* Overlay bar graph

graph addplot bar mean group, barwidth(0.80)  fcolor(none) lpattern(dash) ///

xlabel( 0 "Condition 1" 1 "Condition 2" 2 "Condition 3" 3 "Condition 4") ///

xtitle("Experimental conditions") scheme(lean2) ///

ytitle("Measure [unit]") ///

ylabel( 0 "0" 0.25 "25" 0.5 "50" 0.75 "75",nogrid) ///

legend(off)

* Overlay CI95% bars

graph addplot rcap upper lower group, ///

xlabel( 0 "Condition 1" 1 "Condition 2" 2 "Condition 3" 3 "Condition 4") ///

xtitle("Experimental conditions") scheme(lean2) ///

ytitle("Measure [unit]") ///

ylabel( 0 "0" 0.25 "25" 0.5 "50" 0.75 "75",nogrid) ///

legend(off)

* Overlay regression line (only if conditions are dose-dependant)

graph addplot lfit outcome group, lcolor(gs8) lpattern(dash) lwidth(thin)  ///

xlabel( 0 "Condition 1" 1 "Condition 2" 2 "Condition 3" 3 "Condition 4") ///

xtitle("Experimental conditions") scheme(lean2) ///

ytitle("Measure [unit]") ///

ylabel( 0 "0" 0.25 "25" 0.5 "50" 0.75 "75",nogrid) ///

legend(off)

* Save graph

graph export Figure1.tif, replace

drop mean n sd upper lower

graph drop _all

Hello,

I do not totally agree with this point of view.

Bar chart, for me, and other "simple" charts solutions like for example pie charts and so on, may be basic in terms of information showed. However, this does not mean they cannot be useful. In my opinion, these kind of traditional charts are very helpful simply because they are very easy to understand. Depending on the scenario they can be the best solution to quickly have an idea of the meaning of a certain dataset while, in other scenarios they are not the best solution.

Its use does not only depends on the data we are analyzing. It also depends on the user profile who will analyse and try to understand the meaning of the data. If we are talking about non-IT users or non-data-visualization expert users, than a simple chart (like bar chart) can be the best solution just because the user will not have the knowledge to understand the meaning of a more complex chart (e.g. scatter plot, weighted map, tree map, 3D charts and so one).

So, in conclusion, in my opinion, the best chart to use in order to analyze a dataset, depends on the data itself but also on the user which will analyze this data.

Paulo,

I see the benefit of having readear that are "comfortable with" and "used to" barcharts.

But I don't subscribe to your claim "because they are very easy to understand". I'd rather say that they are easy to misunderstand. They are often harder to decipher, and the relevant information is obscured rather than highlighted.

There are surely some not-so-easy-to-interpret plots, especially, as you noted, for "non-IT users or non-data-visualization-experts". But usually the story behind the graph is not-so-easy, too. One must then find a way to visualize the relevant information is the most simple way possible (as always). Chosing an inappropriate chart type just because this is well-known (for totally different purpuses/data/stories) is usually not a good idea. The reader may then think he understands it, but he likely misses the essential information that actually should (but is not, or not recognizeably) be visualized.

And I am not talking about chart-junk. Charts that are unneccesarily difficult or use gimmicks that do not help to clarify the relevant information are bad charts anyway. But I think we agree here :)

 Dear Mr. Wilhelm,

I totally agree with what you wrote.

Best regards

I was going to say more or less what Paulo said.  What works best depends on what information you are trying to impart and on your audience.  More is not always better.

regards to all,

Well i truly think what paulo said and all that subscribed to his idea makes greater sense, just to remind ourselves of the basics in life and particulalry in mathematics, at some point we were told that in mathematics there is also an answer "it can't" e.g 5 minus 8 , later that answer made no sense to those that proceeded with mathematics.

Applying the same thinking here about graphical presentations that seem basic or not making sense all depends with the user or beneficiary ,so we can agree that some will never get to understand whenthese basic graphical presentations will stop making sense while others will appreciate when that finally happens.

kind regards

It doesn't make sense to use bar graph in visualizing continuous data, because bar graphs are suitable for count data such as discrete data. Histogram or box plot is more appropriate..