It is the analysis of species differences, α β diversity of differences and diversity of nothing less than the difference based on the difference 16s sequencing. At most, that is the basis of PICRUSt, BUGBASE this type of software complete function prediction, do some functional differences analysis.
Common difference test methods can be divided into two categories, one is the basic difference test method, only the output value p, e.g. t-test, rank sum test. While the other one is a p-value can be output, the output value of R may be, as Anosim, Adonis like. Let's briefly look at these methods.
01. The basic test method
For only the p-value output test methods, and its purpose is very simple, that is whether there are differences similarity test comparing the distance between the groups, which reflect the impact of different environments on the community if the result of the role. Common analytical methods are chi-square test, Student t test, Wilcoxon rank sum test, and so on.
If only two samples compared with chi-square test for, but to tell the truth, nothing reliability test out the results, because at this stage it is not repeated studies 16s "unconvincing" the. They will not speak cheap, do they not repeat the difficulty, that is, from biology, statistical point of view, but also need to do repetitive details, refer to the previous article "shallow talk sequencing sample diversity repeat the question."
If the two samples (at least 3 repeats), you can try Student t, Welch'st and Wilcoxon rank sum test. Student t test requires samples with normal distribution and variance alignment. When the number of different groups of samples, time aligned variances nor, Welch's t test is a good choice.
Wilcoxon rank sum test, also known as Mann-Whitney U test, is a statistical method based on variable ranking, no samples with normal distribution, the sample variance does not need to be aligned, is more extensive testing methods, but also because the test too loose, easy to bring a lot of false positives.
If a plurality of sets of comparison sample, can select one way ANOVA, TURKEY Kruskal-Wallis H test and the like. In fact, one way ANOVA and TURKEY are based on analysis of variance, with the latter only a posteriori, can know the two groups contribution to the overall differences.
Kruskal-Wallis H test nature is also a rank-sum test, the difference between the first two in that it does not require the number of samples and the variance of alignment, more widely used.
These methods Which is better? This is a matter of preference, if in the case of uncertain data model, suggestions can try, after all, is the question of the command line
。
02. Matrix-based test methods
Test methods mentioned above, in fact, can only tell you that if these groups were significantly different (can be simply understood as the presence or absence). At the same time that if you want to know the extent of these differences (can be simply understood as the number of) it, it needs Anosim, Adonis and MRPP other testing methods.
These test methods may be output not only significant result (p value), as well as the extent of the result (R value), R value can be used to determine the contribution of the packet size.
For example, microbial growth by N, P, and other factors influence. Case if there are three sets of samples, namely, control, N process and the P-treated soil, in the N and P treatment relative to the control are significant (p <0.05) in the case, wanted to know in the end process N or P is more important that is, before the R-value depends on the size.
Anosim
Anosim (Analysis of similarities) is a non-parametric test method. It is calculated by the first variable relationship between samples (or similarity), then calculate the ranking relationship, the finally determines whether the inter-group differences permutation test groups were significantly different by rank.
The test has two important values, p is a value to be determined between this group and the group is significant if the comparison; R is a value to be a degree of difference obtained between the two groups and the group.
The actual range is the R-value range (-1,1), but generally is between (0,1), R> 0, specify the presence of differences between groups, typically R> 0.75: Large differences; R> 0.5: moderate differences, R> 0.25: small differences. R is equal to 0 or near 0, indicating no difference between the groups. Occasionally, R <0, the situation is significantly greater than within-group differences between groups, which shows our sampling or packet big problem, can be considered as invalid data (how to do? Redo the experiment ah!).
Anosim principle (similarity left, the right is the relationship between rank)
Adonis
Adonis, the name of this very strange, but it actually is the famous PERMANOVA, and everyone familiar with it. In fact it uses nearly Anosim, it is possible to give a different interpretation of the packet factor (R value) of the sample packets and the difference was significant (P value).
The difference is that different test model applications, ADONIS is essentially based on the F-statistic analysis of variance, so many details are similar to the above analysis of variance.
MRPP
As MRPP, in fact, with almost two above, except that it uses the model test method is based on the delta statistics, the details would not elaborate.
03. application
In many cases, we have done PCoA, after NMDS, although the naked eye can distinguish about whether the grouping can be clearly separated, but in fact, still need rigorous statistical testing, understand that this so-called "clearly separated" in the end significantly insignificant.
Anosim, Adonis These models can be used for multivariate statistical tests is very appropriate. It is worth to note that, Anosim essentially based ranking algorithm, in fact, the best fit with the effect of NMDS. If the analysis is PCoA recommended to apply with Adonis test results.
NMDS with Anosim Display Results [1]
PcoA conjunction Adonis (PERMANOVA) Analysis of [2]
04. Summary
Above a lot of nagging, points can be summarized as:
a. The first dot mentioned T inspection can only tell whether a significant difference, only the second small point mentioned Anosim and other methods in order to have the degree of difference evaluation.
b.Anosim, Adonis (PERMANOVA) and MRPP but different models, basically need complex multivariate analysis methods and the like used with the PCoA. Anosim and more with NMDS, Adonis and more with PCoA.
c. These test methods, not beta diversity of unique, they can be used for analysis in any case, but some are based on the original abundance of data, some of which are based on the distance relationship, some are based on rankings, and so on. Note that the conditions of their analysis, the analysis is not an object.