Common inaccuracies and errors in the application of statistical methods in soil science

• Published in: Бюллетень Почвенного института им. В.В. Докучаева no. 102; pp. 164 - 182
• Main Authors: Samsonova, V. P., Meshalkina, J. L.
• Format: Journal Article
• Language: English
• Published: V.V. Dokuchaev Soil Science Institute 11.07.2020
• Subjects: confidence interval; correlation coefficient; criterion power; hypothesis testing; pseudoreplication; significance level; statistics designations
• ISSN: 0136-1694; 2312-4202
• DOI: 10.19047/0136-1694-2020-102-164-182

The most common inaccuracies and errors in the application of statistical methods found in Russian publications on soil science are considered. When designating random variables and distribution parameters in Greek letters, it is necessary to designate those that refer to general populations, and Latin letters – to sampling ones. A detailed description of the experiment and what the replications relate to allows you to draw correct conclusions from the study. It is necessary to avoid pseudoreplication when results at closely located sampling points are considered as characteristics of soil variability over large distances. Expanding the list of descriptive statistics will allow you to use a specific study in meta-analysis. Calculating the confidence interval for the average using the Student's test at different significance levels expands the scope of possible values of the average, but this approach is justified only if the indicator does not differ too much from the normal distribution. When testing statistical hypotheses, it is necessary to pay attention not only to the level of significance, but also to the power of the criterion. The normality distribution hypothesis can be tested using various criteria. The success of applying the criterion depends not only on the validity of the null hypothesis (a truly normal distribution), but also on other reasons: on the sample size and on the alternatives for which the criterion tests the hypothesis. Any statement about the type of relationship between features based on the correlation coefficient (Pearson or Spearman) is meaningless without specifying the number of replicates, since it is the number of replicates that determines the significance of the difference between the correlation coefficient and zero. It is proposed that authors and reviewers pay closer attention to such errors.