I know that you may think that this paper is a bit off from the primary aim of this blog. However, statistics is everywhere and it is very important since our interpretation and conclusions of our data is dependent on the statistic applied.
Therefore, I have decided to share this very easy to follow paper that reviews the major errors in statistic observed in the papers published in the high standard quality Infection and Immunity journal. Since we can learn from errors, here is a brief summary of the most common errors:
1) Fail in the adjustments of the P values when doing multiple analysis. Since the majority of the analysis performed among different groups are done independently and then taken together there is an increased probability of false significance. Therefore, an appropriated statistical test has to applied in order to compare all the groups and additionally the statistic test applied should correct for p-values errors.
2) Several conclusions are based only in the interpretation of the dispersion of the data (such as SD or SEM). Standard deviation bars contain no information regarding the precision of the mean. In a matter of fact, it can occur considerable overlap of confidence intervals even when exist statistically significant differences between groups.
3) Application of parametric analysis in data with clear skewed distribution. Data such as cell or CFU counts, titers and percentages are normally skewed distributed, therefore, standard parametric tests shouldn’t be applied. However, generally, if transformation (log transformation for instance) of the data is performed a parametric data can be applied.
4) Wrong way to represent the data. These errors are associated with the representation of the measures of variability such as standard deviation, standard error of the mean and % of intervals of confidence. Data analyzed by non-parametric analysis should be reported in tables or depicted in graphs with median along with an appropriated range (minimum and maximum values, upper and lower quartiles, etc). Means and standard deviation are not appropriated to represent data that do not follow a Gaussian distribution.
5) The determination of paired and unpaired samples is often done wrong and therefore, the statistical analysis applied is not the appropriated.
Finally, one of the biggest critics was that a great number of papers do not present a detailed description of the statistical analysis performed and therefore, is difficult to analyze if the statistical analysis applied is adequate or not for their particular case.
Is a very short ans objective but still I found it very helpful.
Reference of the paper summarized above:
RM: AF 1372