Top tips for beginners on statistics

2016 03 17 11 40 27 35 2016 03 18 Statistics Hoffman 200

Have you been motivated to start your own research by high quality presentations at the ECR? Have you ever felt completely out of your depth when reading about statistical tests? You are not alone. Now you can get some essential top tips from a statistics expert, Dr. Verena Hoffmann, PhD, who is an expert reviewer for European Radiology. Learn where you can get free online software and interpretation advice, which software packages may suit you, and how essential it is to have the correct statistical test set out from the start of any research project. Read on and don't be afraid!

ECR Today: A famous saying (attributed by some to Winston Churchill) states: "Do not trust any statistics you did not fake yourself." Can you tweak any data just the way you like? When are statistics misused?

Dr. Verena Hoffmann, PhD, is a statistician at the Institute for Medical Information Sciences, Biometry, and Epidemiology at the Ludwig-Maximilians Universität in Munich.Dr. Verena Hoffmann, PhD, is a statistician at the Institute for Medical Information Sciences, Biometry, and Epidemiology at the Ludwig-Maximilians Universität in Munich.

Dr. Verena Hoffmann, PhD: It can be easy to give the reader a wrong impression by using inappropriate statistical methods. Because of that, statistical literacy is important for scientists. Statistical methods that make sense in one context can be completely wrong in another. One popular mistake, or attempt to groom data, is calculating percentages.

If the data of a thousand patients are available it is totally fine to report percentages to the decimal, while it is very misleading to report 66.6% when you should be reporting that two out of three observed patients were affected. My impression is that scientists usually make mistakes because they do not know the correct methods, rather than actually out of fraudulent intent. However, their results and conclusions are equally wrong in both cases.

Which statistical tests are out there and which are actually helpful and meaningful? Are there any "good" or "bad" statistical tests? How do I decide which test to apply?

Almost all available methods are good for some scenario of research. The key is to find the correct method for the respective research question. Regarding statistical tests, the choice of the test depends on the nature of the data. How is the data scaled? Is it normally distributed? How many samples do you have and are the samples dependent or independent of each other? For example, if you want to know if the rates of cured patients after one year of therapy significantly differ between two groups of patients who were treated with different medications, the Chi-squared test would be an option for analysis.

When should one not use statistical tests in research publications?

Statistical tests should not be used for very small sample sizes. Generally, the term "significant" should be reserved for tests used for prospectively planned analyzes of predefined endpoints according to the protocol, to avoid the problems of multiple testing. P-values can also be used descriptively, but other measures (such as the median, mean, or mean differences) are often more meaningful.

Which software could you recommend for statistical analysis?

The most commonly used, commercially available software packages are SAS and SPSS. SPSS is the more beginner-friendly as you can do most analyses by selecting them in menus. Some programming skills are needed to use SAS, but it is more flexible, as is the "R" software, which is freely available and also provides packages including the most recently developed methods. The use of all software requires a solid base of statistical methods so you will be able to interpret the results.

Do editors (or reviewers) check statistics in research manuscripts prior to publication? What do I need to consider when submitting statistical research data? What should I avoid?

A check of the statistical methods is getting more and more common. Also, an increasing number of journals have dedicated statistical reviewers. When submitting a manuscript, make sure you describe what methods were used. Be specific! If the reviewer cannot see what you did, there will be one more round of reviews.

Many famous people have written about the use or abuse of statistics in various settings. What is your favorite statistics-related quote?

Andrew Lang said: "He uses statistics like a drunken man uses a lamp post, more for support than illumination." Unfortunately, this is true for some researchers, who are so very convinced of their hypothesis that they would rather go searching for a statistical method that supports it than stick with an appropriate analysis method that is telling them their hypothesis might not be true.

Could you recommend any books or online resources that are brief and simple enough for beginners and yet helpful? Is there an online forum where I could get specific advice if I am unable to get help from my hospital?

The Institute for Digital Research and Education at UCLA provides a great website covering many statistical methods and tips for software and interpretation and there is a wide range of tutorials available on YouTube, especially on the use of SPSS. While there are online forums and groups (mostly specific to the software you use) where people help each other with complex questions, be sure to do your research first. Just as there are few doctors out there able or willing to answer the question "my knee hurts -- what can I do?" there are also few statisticians who can answer unspecific beginners' questions.

Would you have any other recommendations?

Think about the analysis methods already when planning your research. Define primary and secondary endpoints, and confirmatory and explorative analyses. Make an analysis plan you can stick to. Also make sure to collect all the information you will need for the analysis and to collect the data in a format that can be imported into the analysis software. Do a sample-size estimation so your analysis will not be underpowered. If in doubt, consult a statistician. Most research institutions employ someone who will help you -- and when talking to him or her, don't start the conversation with the Churchill quote.

Interview conducted by Dr. Christiane Nyhsen, consultant radiologist at Sunderland Royal Hospital, U.K.; member of the editorial advisory board of AuntMinnieEurope.com; and former chairperson of the ESR Radiology Trainees Forum.

Originally published in ECR Today on 2 March 2016.

Copyright © 2016 European Society of Radiology

Page 1 of 1243
Next Page