Now I'm going to go on to talk about tables and figures in peer reviewed journals. Tables are usually a listing of numbers that reflect the summary of the data from the clinical trial. One thing you don't want in tables is errors. One tip I have for you is, as much as possible, try to use the automated output from statistical packages to create the tables. The results are really inserting output from the statistical program and not having someone copy it which is very prone to errors. Programs like SAS have ODS, which is their output delivery system, a little complicated but it does work and R they even have a program that's called Table1 that is designed to put together your Table 1. Use a consistent layout generally and I'll point this out in some tables that I show you, the intervention should be the columns and the characteristics that you're comparing between the interventions are the rows. Use footnotes generously, you want people to understand what's in the table so define any acronyms or abbreviations. You can put notes about where model estimates came from and what covariates were adjusted for. Table should always include some measure of precision of your summary measures so you don't want to mean without a confidence interval. Always include the units and the number of participants whose data is in this table. Back to Table1, I mentioned before that we usually don't recommend significance testing when we're comparing the baseline characteristics between groups. Although sometimes journals ask you to include them. Like the abstract, tables should be standalone documents that you can look at this table and understand what is presented in it without reading the text. Here's an example of a baseline table from one of the studies I was involved with. This is table 1, characteristics of the patients, we have three treatment groups across the columns. Then the characteristics we want to compare them on are in the rows along with some measure of variability and their unit. Also, unfortunately, we have P-values here to see if there's any difference in, let's say, the age between the three different treatment groups. The reason we don't like to include P-values in table 1, is that because people were randomized to their treatment, any differences between the characteristics of the group is something that happened by chance, so this P-value is literally a measure of chance. It's really not a measure of treatment effect, which it is when we're looking at it in terms of the results. If I was able to resubmit this table, I would take out the P-values, I would report confidence intervals instead of standard deviations and I would think about the level of precision that should be reported. I'm not sure that 29.3 is that much different than being 29 years old. Here is an example of a results table. These are much more complicated to put together, there is not a straight forward formula or program like the one in R that spits out table 1. You have to figure out the best way to present your results but you should still follow model that the columns should be the different interventions. Here we have placebo is one group and omeprazole is the other group. We have the Ns in each group and then the things that we're comparing these two treatment groups are across the rows. But in the results table, we also have to communicate the treatment effect. In this case, it was a incidence rate ratio, comparing the rate of events in the omeprazole group to the placebo, so we have that with a confidence interval. We also have a P-value for that effect indicating that it was not a significant effect. Because we were very interested in whether the subgroup of people with gastroesophageal reflux had a different outcome than people without reflux, we had a formal test of the interaction that was specified in the paper. I also want to point out in this table that the primary outcome here is asthma episodes. Episodes of poor asthma control. But that's a composite outcome that consists of several different types of events. You could have had a drop in your peak flow or you could have had an urgent care visit for asthma, and you can see many more people had drops in their peak flow and the rate of those events was much higher than having to receive urgent care for your asthma. It's important if you're presenting an overall composite outcome that you actually present the components as well so the reader can judge what were the most important events in that composite outcome. An improvement I would've made to this table, is I would've added a confidence interval to these event rates since this really is a summary statistic. I want to go over one more results table, because this is results on a continuous outcome. The last table we looked at was an event-based outcome. This one is a continuous outcome of visual acuity. I also think it helps us to understand about the different ways to express treatment effects. Let's just go through the table. The primary outcome is visual acuity, down the rows, we have the different times it was measured, in the first column, we have the number of observations at each measurement point, and then here we have in the columns the two treatment groups and their estimated mean. Here just means at baseline, the implant groups visual acuity was about 61 letters and the systemic group was about 64, that's how they came into the study. Then we see what their mean visual acuity for the groups were at the different time points. It's a little hard to look at those numbers and figure out is there a difference? May be easier if we looked at the estimated mean change from enrollment. If we knew at six months that the implant group had a six-letter improvement in visual acuity whereas the group that received systemic treatment only had about a two-letter improvement, that's a little easier to compare than 66.9 and 66.4. These are the change from baseline at the different time periods in each group. But really the overall treatment effect is best expressed in this last column which is actually the difference in the change. It's the difference and the differences which we can see at the first time point is about 3.91 and you can roughly guess that by thinking that's going to be 5.88 minus 1.97, but this is really the treatment effect. The implants effect was not to improve vision from baseline by 3.91 letters, we can see here it actually improved vision from baseline by 5.88 letters. The treatment effect is the difference in the improvement between the implant group versus the systemic group. This is the money column with the important results of the trial. I want to say a few words about figures for reports of randomized clinical trials. Obviously, a figure is meant to convey a message. In the context of clinical trials, it's usually about the design of the trial, or the treatment effects on the primary outcomes. The figure should have clear informative captions and legends so that you can again understand the figure by just looking at the figure without referring to the text. Also, when you're summarizing numbers, you should use appropriate visualizations depending on the particular metric you're using and I will give you some examples of that. Here are two examples of designed slides. On the left-hand side of the slide, we have a depiction of a three-group parallel design. Indeed, you can see that participants entered, they were on a running period where everyone received the same drug, they were randomized to one of three drugs and followed for up to 16 weeks. You get a lot of information from this particular figure. On the right side of the slide, I have a designed figured for more complicated trial. It was a modified factorial trial with five groups. If you recall in a factorial trial, a participant is randomized, in this case, to either active drug or a placebo. They are also randomized to receive an enhanced presentation versus of regular presentation. In each of the groups, there are a 120 individuals. But because this was a modified factorial, we had another treatment group of usual care. I think for the complicated design that it was just trying to depict the treatment groups was a good idea for this particular figure as opposed to doing what was done in the other figure that also depicted a lot of information about the interventions and the follow-up period. I also mentioned that it was important to pick the correct mode of visualizing your data for the type of data at this. On this slide, I have three different ways to graphically depict the same data and it is baseline data from a clinical trial on a continuous outcome. Often people like these bar graphs, but they really are inappropriate for continuous data and I think the reason should be obvious. The bar graph here it goes up to the mean and has a 95 percent confidence interval. But actually, if we move to the middle, we can see a much improved depiction of these data because now we can see where the mean is in relationship to most of the data, we can see where the median is, the straight line and the whole extent of the data. This depiction is much more informative. Bar graphs should never be used for depicting continuous events. The second graph is what we usually call boxplots to depict the distribution of the variable in each treatment group. A more modern way that is coming into vogue for depicting the distribution are the spinal plots on the right-hand side that's a violin plot. It's more complicated in that it shows you the density of the values. Even though both of these graphs depict a median of about 15, we can get a better feel for how the actual data are distributed within that density. You are likely to be seeing many more violin plots, sometimes called mandolin plots. My second point was that you need to emphasize the treatment effect in figures. I also have three different graphs to show you how they may depict the treatment effect. In the first graph, if we're looking at change from baseline in the outcome, as I said these boxplots are good for showing the distribution of the data, we can see the means and that these groups look quite different. However, the information on the outcome is pretty limited to comparing the two groups overall. In the second graph, we have change from baseline that includes subgroups. Now the change is along the x-axis, there's a reference line at zero, meaning there was no change in the outcome. Then we are able to look at the overall effect in this first part of the graph which shows you basically is the same depiction here except for now it has a 95 percent confidence interval. It would tell you that the control group, which is depicted in blue, looks different than the experimental group. Many times we are also interested in seeing if the treatment effects that we see overall hold true in subgroups of the population. Do males have the same treatment effect as females? We have the comparison here among males, the control group versus the experimental group. Then finally, the comparison in females with the upper line being the control group and the lower being the experimental group. As indicated by this P-value for interaction, there really isn't a difference in how men and women reacted to the treatment. You can see the overall treatment effects and the differences in the men and the women seemed quite similar. However, that graph emphasizes change from baseline. The final graph I'm going to show you focuses on the actual treatment effect, which is the differences in the difference from baseline. Both these groups change from baseline but we want to see the difference in the change and that is really the treatment effect. To get oriented again, the y-axis is again the difference of the differences in outcome. It's the difference in the difference of outcomes. How much different is the experimental group minus the control group? We have a reference line at one, we start with the overall difference and see that here, the point estimate for the difference is about minus one, so that the experimental group had less progression of disease than the control group and indeed the confidence interval excludes zero. It does appear to be a significant effect associated with treatment. We can then look at the effect in the different subgroups and see very easily that the effect in males is quite similar to the effect in females. If we want to depict event-based outcomes in graphs, we might do something on the left side of the slide and have cumulative percentages, but that's not a very informative graph. Usually when we're looking at events, we'd like to include some type of time dimension. That is why we're giving it the real estate of a graph so we can depict two-dimensions. Here we have a Kaplan-Meier curve. The y-axis is the cumulative percentage of patients that have failed and the x-axis is time. We can see the behavior of the three different treatment groups. We also have below the figure a table that tells us the actual numbers involved at each time period. This includes much more information than a bar graph of events. I also want to mention that figures can be used to mislead and I personally can't think of any pictures I used to mislead, so I had to find some in the literature. But basically, what they tend to do is to somehow change the scale to emphasize differences, we have two depictions of different size circles. The depiction on the top is showing that this circle with a value of 30 is twice as large as this circle with a value of 15 and this circle has a value of 10 so it is 1/3 the size of the largest circle. You could also depict that by drawing the circle not based on the actual area of the circle but on the radius of the circle, which is a common algorithm in some statistical programs, but it ends up distorting the differences. Because when we use the radius to draw the circle, the differences between the four circles look much greater than the actual differences based on their values. Another common way to mislead people is to fool around with the axis. Here you can see we've got a full axis from 0-100 and most of the bars are about 80, they all are pretty even, no big differences. On the bottom set of bars, you can see where the graph was constrained just to show values above 50 percent so it doesn't start at zero. The axis is unnaturally expanded and it makes the differences between the groups look larger. Just be careful about interpreting graphs to make sure they aren't deceptive and also don't create deceptive graphs. Some final notes on figures is use colors judiciously, they should be adding to the meaning of the graph, not to its aesthetic sensibilities. They can be very distracting. You should always try to use high contrast, you should avoid inserting unuseful information in your graphs, commonly known as chart junk. I really advise you not to use Excel. Like tables, you should start from a statistical package and have that produce the graphic from the raw data to ensure that you really have an accurate depiction. Finally, I wanted to give you an example of a bad graph and one that was a little bit improved. This first graph in the middle is the bad graph and it's got all these bars depicting how much people trusted different sources of information in different time periods which you can see the time periods in this legend, if you could possibly read it and decipher the colors. A better way, but not perfect of depicting similar types of information, was to connect the information over time in a line. Here in the graph on the very far right, instead of having the source of the information on the axis, we have the time periods on the axis. The percentage of people who trusted information from different sources is on the y-axis as it was in the graph next to it. But instead of using bars, we used a line graph to show how trust in different sources changed over time during the pandemic. You can see that the trust in social media was never high and decreased over time. Still a little bit too busy but certainly better than that middle graph.
