本文根据《Journal of the American College of Cardiology》上曾发表的一篇文章《Making Sense of Statistics in Clinical Trial Reports》，来全面而具体地说明临床试验论文中，各种类型数据与结果使用图表的正确展示方法。
This article will focus on the presentation of baseline data, trial information, and dichotomous frequency and measurement data in outcome data.
1. Baseline data display
Baseline data is the information that must be provided and presented for all types of trials, and plays an important role in describing the characteristics of the samples included in the study, and indicating the variables of interest in the study.
The presentation of baseline data is mostly in the form of a table, with the names of various variables as the headings of each row of the table, and different group names are the headings of each column of the table, and the variables that need to be covered include demographic information, variables that may affect outcome events, and related medical records, each column provides a group of data, and it is not necessary to provide aggregate data or compare differences between groups in RCTs.
A table of the study is provided below as a template, and the data presentation method, possible scenarios, and relevant considerations are indicated in the corresponding positions (Table 1).
(Click on the image to enlarge it)
二、试验信息（Trial profile）展示
The trial profile is also a part that needs to be displayed, which is used to clearly and intuitively describe the operation at each time point in the trial process, as well as the changes and reasons for the number of people in each group after each operation, so that readers can have an intuitive understanding of the study.
The trial profile is mostly displayed in the form of a flow chart, arranged in chronological and operational order, clearly listing the number of people in each part, and the important operation nodes that need to be displayed should at least include random grouping, completion of follow-up and data analysis.
The following figure shows a trial profile provided in the study, which can be used as an example for reference, and the possible display methods are not limited to examples, but the main information should be clearly provided (Figure 1).
(Click on the image to enlarge it)
3. The results of the two-classification frequency data are displayed
Dichotomous frequency data, such as gender, occurrence or absence of various events, are often included in pilot studies, and are an important part of the presentation of results, and both the primary outcome data table and the side effect table contain a large amount of dichotomous data.
In the primary outcome data table, similar to the baseline data format, the names of various variables or events are used as the headings of each row of the table, while the different groups are named the headings of each column of the table, except that two additional columns need to be added to report the size and P-value of the effect measure, respectively.
It is important to note that in the presentation of events, the primary endpoint of many studies is a composite endpoint, in which case the number and percentage of each event occurring separately should be provided in the following order at the same time.
Efficacy measures need to include point estimates and interval estimates. Here, we will emphasize the importance of interval estimation and explain the relationship between point estimation and interval estimation: point estimation can obtain a specific value that represents the magnitude of the result, and there is a large degree of uncertainty in this result, and interval estimation can express this uncertainty, usually with a 95% confidence interval.
The reason why 95% was chosen was mainly to ensure the consistency of the studies and facilitate comparison with other studies, because most of the studies chose the value of 95%; At the same time, it is consistent with the usual significance level of 0.05, which is convenient for correlation with P<0.05. The larger the sample size, the smaller this uncertainty, the narrower the confidence interval, and the more accurate the point estimate estimates. Therefore, it is not rigorous enough to report the point estimates separately and the results of the interval estimates need to be reported at the same time.
The figure below also provides a table for the study as a model, with notes noted next to it (Table 2).
(Click on the image to enlarge it)
A variety of different effect indicators can be used for the effect analysis of the study, and Table 3 summarizes the concepts and calculation methods of each effect measure. Usually, when there are two groups, the four-grid table of the data is listed first, the incidence rate is calculated, and then the corresponding effect indicators are calculated.
(Click on the image to enlarge it)
效应指标包括绝对数（如危险度差（Risk difference, RD/Difference in percentages）、需治疗人数（Number needed to treat, NNT）等）和相对数（如相对危险度（Relative risk, RR）、相对危险降低（Relative risk reduction, RRR）和比值比（Odds ratio, Relative odds, OR）等）两类。 相对数指标具有统计学优点，且对于不同类型的人群的一致性较好，结果易于推广，而绝对数指标则更有实际价值和意义。
(See: Summary: Indicators that can be used to evaluate the effectiveness of interventions)
In addition, in addition to the detailed records and primary outcome measures, safety information is also reported, and the safety tables are mostly dichotomous.
Similar to the primary outcome event table, the names of various variables or events are used as the headings of each row of the table, while the different groups are the headings of each column of the table, except that only the test for differences between groups is required to report the size of the P-value.
The following figure shows the safety table shown in the study as a reference for the content and style (Table 4).
(Click on the image to enlarge it)
Fourth, the results of the measurement data display
1. Comparison of means between groups at a single time point
When analysing quantitative outcomes, a common analysis strategy is to directly compare the differences between different intervention groups for outcomes. However, considering that in most cases these outcomes will be measured at baseline, a more reasonable approach would be to compare the mean change from baseline in the outcomes.
But there is also a bug here – this change is often influenced by the baseline level, as we often call "regression to the mean" – and participants with higher baseline levels of outcomes may experience greater reductions under the same intervention conditions. To address this problem, an alternative statistical analysis method, analysis of covariance (ANCOVA), is needed, in which the baseline level of the outcome measure is adjusted when the change in the value is compared.
As an example, the SYMPLICITY HTN-3 study [6] is a randomized, double-blind, sham-controlled trial that enrolled a total of 535 patients with severe refractory hypertension who were randomized 2:1 to perform either nephrosympathetic or sham treatment. The primary endpoint of the study was a decrease in systolic blood pressure (SBP) at 6 months of treatment.
(Click on the image to enlarge it)
As shown in Table 1, a between-group comparison of SBP at 6 months in the nephrosympathetic and sham treatment groups was performed first, and secondly, the baseline level of SBP was adjusted when comparing the 6-month change in SBP between the two groups. It is evident that the 95% CI was "wider" than the latter two in the first group, and the 95% CI was the narrowest in the third case after adjusting for baseline SBP.
Unfortunately, the above 95% CI still crossed "0", i.e., there was no statistically significant difference between groups. These results suggest that nephrosympathetic debridement did not reduce systolic blood pressure levels at 6 months compared with sham surgery in patients with refractory hypertension.
Is it really necessary to adjust the baseline level of SBP? The two regression lines in the nephrosympathetic and sham treatment groups in the figure below show that subjects with higher baseline SBP levels in both groups also had a greater SBP decline after 6 months. Without adjustment for baseline SBP levels, the actual effect size may have been misestimated (4.07 vs. 4.11 adjusted mmHg).
In addition, it is also clear from the chart below that the results of the study vary widely from study to study subjects, which is why clinical trials often need to include enough study subjects (too small a sample size can lead to inconsistent results).
(Click on the image to enlarge it)
Another question that a careful partner may ask is whether to choose the difference in outcome change or the percentage change from baseline in the actual analysis. Statistically speaking, this is when it comes to which case is more appropriate to use covariance analysis. (See:
2. Methods for analyzing and displaying results of multi-point measurement data
We discussed how baseline data for outcome measures can be used to make estimates of the clinical efficacy of interventions more reasonable. In fact, many times a clinical study is designed to collect data at just two time points, the beginning (baseline) and the end (follow-up endpoint). When data are presented at multiple time points for an outcome measure, different approaches are required, depending on the purpose of the study.
(1) The trend of the mean values of the two groups over time
The measurement data at multiple time points can be displayed in a line chart with time abscissa and indicator value as the ordinate, and the following figure is a graph of the results of a study, and the mean and standard error are depicted at each time point (figure below).
(Click on the image to enlarge it)
and (2) comparison between groups of different decline (or increase rates).
In many studies, the percentage of change from baseline (rate of decline or increase) at subsequent time points was clinically significant, such as the rate of decline in forced vital capacity in studies of respiratory impairment, in which case the rate of decline or increase in the outcome at different time points could be calculated for comparison between groups and subsequent analysis.
and (3) the unique value at a particular point in time during follow-up
For example, when glycosylated hemoglobin is measured at 18 months to evaluate the efficacy of a diabetes drug, the time-specific data should be analyzed.
Because there is often a correlation between measurements at multiple time points, unlike the usual statistical analysis methods that require data to be independent of each other, repeated measures analysis should be chosen. (See:)
3. Analysis method with skewed distribution of measurement data
In some cases, the data of the continuous data are skewed, and the traditional analysis method of means comparison between groups may be distorted by some extreme values, in which case the following treatment methods can be selected:
(1) Appropriate data conversion is taken
For example, when the raw values are taken as natural logarithms, the data are normally distributed, and the geometric means are used to compare the differences between groups.
(2) Use nonparametric tests
In nonparametric tests, the median is more commonly used to describe the efficacy of the two groups, and the difference in efficacy between groups is analyzed by nonparametric methods (such as the commonly used rank-sum test) to avoid the influence of extreme values.
(3) Set a specific threshold value to convert the original continuous variable into a dichotomous variable
For example, the percentage of people who exceed the upper limit of liver function is calculated, and the chi-square test is used to compare the percentages between groups.
bibliography
1. J Am Coll Cardiol 66(22):2536–2549
2. N Engl J Med 2014; 371:993–1004.
3. N Engl J Med 2013; 369:1317–26.
4. Lancet 2014; 384:1849–58.
5. N Engl J Med 2013; 368:1303–13.
6. N Engl J Med. 2014; 370:1393–401.