PRIMER-E - Science topic

Hello! I hope you've found a solution already, but if not, here are a few tips:

PRIMER can be quite demanding with large datasets, so try running it in priority mode with as much memory and CPU as possible. Close other programs, disable energy saving mode, and consider running it overnight.. (with Windows updates disabled).

Check your data for missing values or redundant rows, as these can slow down processing significantly.

If it’s an option, you could subsample or average your data to reduce size but keep in mind this may reduce variability, so consider if that fits your needs.

Finally, you might try the vegan package in R to run PERMANOVA. With one factor, it’s straightforward and you can adapt permutations and residuals options to be similar to PRIMER.

Best from Ulm, Katja

Yes. Usually this is done in conjunction with a cluster analysis to define groups (with or without SIMPROF). It's all under the Graph>Special>Overlays menu

ANOSIM is good for skewed data. But you must look at all samples and factors at the same time because of interactions between factors and to avoid bias. If there are only slight differences among sample sizes between treatment combinations, you could impute missing values using the mean or median of the treatment levels. However, this approach should be used with caution and should only be used when sample sizes are nearly equal to begin with. If the sample sizes are not equal and the assumption of equal variances is violated, you could instead perform a non-parametric equivalent to an ANOVA such as the Kruskal-Wallis test :)

Bit late to the party, but effectively using the right resemblance measure (for biological samples generally one that ignores joint absences) is how we deal with the inflated zero problem. That is why we use a measure like Bray-Curtis to define resemblance, rather than a measure like Euclidean distance. The zero-adjusted measure is for specific circumstances where there are few or no occurrences in samples, which is not the same thing as many zeros overall.

Hi Rowan, I am in a similar situation. What I did I used an average of the response variables. But I do not know if it the optimal solution. Did you solve this riddle at the end?

To me there appear to be two questions. One concerns differences in the detected community in space and time, which can simply be addressed using Permanova (or Anosim). The other concerns which of the measured variables 'best explain' the observed differences in the fish community. I wouldn't try to address this using Permanova at all, but some of the other tools in the box such as BEST, which can look for the subset of measured 'environmental variables' (I wouldn't call them covariates) best explainingg the community pattern, Linktree which looks for thresholds in individual variables explaining divisions in the community data, or even just bubble plots in MDS as a first mode of exploration. I'm never comfortable with trying to set up a single test that ticks all boxes, and prefer a piecemeal approach where one builds an explanation using a variety of tests and approaches.

The analysis is fine. The position of the vector diagram relative to the ordination is arbitrary - it could just as easily be in a separate key. The diagram indicates the direction across the ordination plane in which values of the selected variables increase. The length of the lines indicates the amount of total variation in each variable is explained in the chosen ordination plane. If all of the variation is explained, the line reaches the circle.

LDM

Yes, the null hypothesis in RELATE is that the correlation is zero. As you have a lot of samples you have a lot of statistical power to detect departure from the null. The significance tells you how likely it is that you would observe a value as high (or higher) than the value you have observed if the true correlation was actually zero.

There are many other tools in Primer you could also try (RELATE for 2-way designs, BIO-ENV, etc.), depending on your survey design. It might be that you get more meaningful results if you combine samples, maybe pooling nearby samples for example. It is likely, given that you have so many samples, that the signal (from the environment) is being lost in the noise (many samples each dominated by one or two species of fish).

There are several ways to select variables: significance criterion (based on p-values), information criterion, penalized likelihood, change-in-estimate criterion, and background knowledge. As Dr. Booth suggested, variable selection should not be based only on p-values. Check this article for details:

The short answer is yes, it could. The routine finds a subset of variables that, in combination, most closely match a predetermined pattern (represented as a resemblance matrix). So you could make resemblance matrices based on the condition indices and biomarkers (singly or in combination) and then ask which set of gut content items maximises the match. How you treat your data, and which resemblance measures would be appropriate, will depend on the data you have and the specific questions you want to address.

SIMPER is used after the differences between factors are confirmed using an ANOSIM or PERMANOVA test, the factors sometimes or frequently are defined a priori (before to start the research or analysis) according to some geographical, physics or chemistry characteristics that show an effect on the community as a driver of its structure. If there is not a clear effect for any factor selected a priori SIMPROF test allows to observe groups with significant differences, without significant differences SIMPER analysis has no sense.

It really depends what the nature of the data is, and what hypothesis you want to test. For example, a t-test may appropriate, or Wilcoxon-Mann-Whitney, or Mood's median test, and so on.

I've just been doing this for a paper. Primer and Permanova+ are not designed to produce final publication quality graphics and the figures often need a little work, although generally they are pretty good. The easiest way I find is to copy the figure (ctrl+C) and paste it into Powerpoint. You can then use drawing tools>Group>ungroup (you need to do this twice, the first time it tells you that it's not a drawing, asking if you want to convert it, so say yes, then do it again to ungroup). You can then tidy the figure, move labels around, change fonts, add information, and so on. The final step (for me) is to use Adobe Illustrator to produce a final tiff, but first get the figures the way you want them.

Enrique, i am very grateful for your answer, we finally used PERMANOVA and included Anderson et al. 2008 as follows: " Possible correlation structure among samples, derived from the stratified-random sampling methodology, would be ignored by PERMANOVA under permutation (Anderson et al. 2008).". Yet I found that PERMANOVA is unaffected by differences in correlation structure for balanced designs in Anderson & Walsh (2013) - http://coralreefdiagnostics.com/s/Anderson_et_al-2013-ANOSIM-vs-PERMANOVA.pdf

Thus i interpreted that differences found by PERMANOVA reflect changes in community structure and carried out SIMPER to detect the variables that contributed the most.

Best regards

Thank you Amy i'll have a look at it ASAP.

I agree with you on the SE thing which, in my mind also, has to be calculated from the expression values obtained for each of the biological replicates. just a classical SE calculation, relevant of the biological diversity, and not a SE based on the qPCR technique itself.

I've been in touch with other researchers since my last message. If I get anything new I'll post here some explanations of what I've been doing and what solutions I found.

Thanks again

A couple of additional notes: R is a scaled measure of the separation between groups of samples, so if the Global test is significant (there are differences between groups) you can interpret the relative values (close to 0, not different, close to 1, very different) even if you don't have sufficient power (samples) to run the pairwise tests.  Simprof tests a different hypothesis, so be careful that you understand what each test is actually testing.

The package is called Primer, not Primer-e (which is the company that sells it).

Jayachandran,

You may use PC scores to reduce variables (i.e. the correlated ones). firstly, you should do a PCA on your parameters using predetermined number of PCs (lets say N= 10 PCs). Then check to see how many PCs are needed to describe 100% cumulative %variation (for example the first 5 out of 10 PCs). In each PC (1st to 5th) choose the variable with the highest score (irrespective of its positive or negative sign) as the most important variable. Since PCs are orthogonal in the PCA, selected variables will be completely independent (non-correlated).

Hope this helps you

Amir

To obtain your "indicator" data matrix:

1º Edit -> Indicators.. etc, the same way than factors

2º and then go to tools -> sum (or maybe average) -> and in variable square click on "sums for indicator" and select your indicator -> OK.

You will obtaain your new raw data for indicator levels.

I am not familiar with the specifics of ANOSIM, but I think a GLMM might also be able to provide your answer, by using the colony as a random effect to compensate for  differences in initial colony size.

The nMDS can appear like showing there is 2 separate groups, but the way you read a nMDS, is that 2 samples close to each other show high similarity, and 2 samples far from each other show low similarity. If you look carefully to your nMDS, you will see that L7 and G5 are really close (highly similar), while L6 and L5 are apart showing low similarity. Overall, most of your samples show similarity between group and high variability within, so it is not surprising to have a non significant ANOSIM even if the green look to be on he left and the blue on the right.

Why do you have more green site than blue? Do the numbers correspond to specific location, time..?

As 2D stress of 0.24 only mean that your nMDS is not good to represent your data in 2D. 3D might improve this. It is not surprising that ANOSIM might reveal things you can't see with nMDS, because nMDS loose information by working in 2D, while ANOSIM will work will all the dimension and complexity of your data. You could try other methods than nMDS, such cluster, CAP, PCoA.

Merci Phil, je pense avoir la solution! Merci !

BIOENV and BVSTEP have the same aim: finding correlations between 2 matrices but BIOENV carries out a full search of all possible combinations, while BIOSTEP carries out a stepwise search where the best variables is first selected, then the second one is added and so on for up to 5 variables usually, which the forward selection. There is also the backward selection, removing one by one the variables that affect the least the rho value when omitted. Usually the limit of variables to use BVSTEP instead of BIOENV is 16 variables, although this is not an absolute cut off and is more a practical cut off. The main reason behind using one instead of the other is because of the number of combinations.

When you actually have lot of variables, you should first have a look if some variables are inter-correlated, which is likely to be the case with environmental variables. So you can remove one of the inter-correlated variables to reduce the total number of variables. BIOSTEP may not be valuable when you have lot of inter-correlated variables as you are likely to have many equally good solutions. But in other situation, when you have not lot of variables which are not highly inter-correlated, BIOSTEP would be a good option.

to determine if you need to transform specific variables, you need to have a look on 2 things (Clarke and Gorley, 2006):

1) “are the samples roughly symmetrically distributed across the range of each variables” (i.e. are your samples skewed to one side)?

2) “if there are strong relationship between some pairs of variables, are these roughly linear rather than strongly curvilinear?”

Based on your draftsman plot, your samples are not really skewed on one side. So you do not necessarily need to transform your data, although you do have some outliers stuck to one side. So you could try a mild transformation (e.g. square root), and see if it improve or not the draftsman plot.

Sometimes if you have a doubt about transformation of your environmental variables, perform a mild (square root) or strong (log) transformation and run again the draftsman plots and see if you improve or not the plots. If not, do not bother with transformation. Sometimes just trying will help you to make your choice. Otherwise, as I mentioned, you mainly need to check if your samples are not skewed to one side or if you do not have outliers that skewed your data.

Hope that help,

Aimeric

Actually, the variation explained of factors can be obtained directly by dividing the sum of squares of the factor by the total sum of squares. In adonis function of vegan package in R, the R2 is part of the output. Hope I could help...

Hi Nayan,

As César said, ANOSIM give you the P value (i.e. significance levels) and a R value (i.e. the strength of the factors on the samples). R value is supposed to vary between 0 and 1 (not between -1 and +1) but you can obtained negative values but they are always close to 0. R value close to 1 indicates high separation between levels of your factor (e.g. control vs treatment samples), while R value close to 0 indicate no separation between levels of your factor. Be careful with the significance level as it is express as %, so if you want the actual P value you have to divide by 100. ANOSIM (one-way) gives you 2 windows: one with the detail results and one with a graph (or 2 graphs if you perform two-way ANOSIM). The result window show you “TESTS FOR DIFFERENCES BETWEEN xx GROUP” which give you the overall P value and R value for your factor. Then you get the “Pairwise Tests” which give you the P values and R values for each level of your factors. For example if you have 1 control and 2 treatments in your factor, the pairwise tests will tell you if the samples from control and treatment 1 are significantly different (P value) and how strongly they are different from each other (R value), and will give you the same for control vs treatment 2, and treatment 1 vs treatment 2. The number of permutations you perform (Actual permutation) and the number you could perform (Possible permutations) are also given and are important. It help you to know if the ANOSIM is relevant in your case, by indicating if you could perform more permutations (you obtain more accurate P value) and if the number of possible permutations is low (e.g. <100) you should not trust the results of the pairwise ANOSIM. The graph window shows you the R value for your factor/matrix (doted black line) and the R value obtained by permutation of the similarity matrix. If the black line (i.e. R value of your similarity matrix) is not overlaying the R values obtained by permutation of the similarity matrix (blue bars) you have significant difference.

The SIMPER analysis gives you the percentage of similarity and dissimilarity or your factors, between levels of your factors and for specific levels of your factors. Then it gives you which variables in your data explain the similarities or dissimilarity: the percentage of contribution of your variables (Contrib%) that explain this similarity. The variables are classified from the highest to the lowest contribution. It also show the cumulative contribution (Cum.%) so that you know how many variables explain for example 90% of the similarity... Do not forget that the % similarity = 100 - % dissimilarity and inversely.

For your results and discussion:

- for ANOSIM you need to first say if there is significance difference and if yes, you need to give the R value. You could have a significance difference, but if the R value is low (e.g. 0.2) this mean your factor have small effect on your variables and is not really important. So with ANOSIM the R value is more important to some extent than the P value.

- You can use SIMPER to support your ANOSIM as you can say how similar are the different levels of your factor and more importantly, which variables explain the difference or similarities.

Hope that help,

Aimeric

PS: César A. Cárdenas thanks for putting the pdf of PRIMER. Do you have also the pdf of the manual/tutorial? I would be grateful if you have it to send it.