statistics problem... math experts needed :)

[biothanasis]

Hi all!!!

I am dealing with a problem on how to treat some data.
I have 6 areas, were pollen is recorded for the same period. About 30 taxa, are found in every area at different concentrations.

How could I analyse all taxa together for each area to check if there are differences in between the areas concerning pollen records for the given period? Since I have done samplings in 2 years, how could I check if there are significant differences in the patterns between the years too?

I do not use anova, manova etc, as there are many zeroes in the dataset and also big st.dev. which lead to biased/unreal results.
I have used Wilcoxon pairs test, but for each taxon separately in between 2 areas of interest. I also used cluster analysis to check if the concentrations recorded for each taxon in the 6 areas differ. But it requires a lot of different graphs to show the results.
Could you suggest an analysis that includes all taxa together for each area, so that I can compare the areas and see if there are differences between them?

I am using Statistica v.7.
Any help would be appreciated! Thanks!

 

[TyroneGenade]

Try http://stat-help.com/ . Sounds like you have a complicated problem. Also, you should think about what hypotheses you want to disprove as this will tell you what to compare and how to compare it.

I like Anovas as they are quite robust when it comes to large standard deviations (i.e. any real biological data).

Sorry I can't be of much help. As an experimental biologist I don't deal very much with the real world and my stats problems tend to be neater.

 

[biothanasis]

Tyrone, thank you very much for the info!
I shall send them a query about my problem.
Thanks again!

 

[naoki]

You can probably used generalized linear mixed model (glmm) with probably Poisson distribution. The idea is similar to ANOVA, but you can assume other distribution (than gaussian) for the error term. So 0 records can be included in the analysis. Similar to ANOVA, you have pollen count as the dependent var., and sampling area, year, and taxa as the explanatory var. (fix vs random effects depends on your question). If you don't consider any effects to be a "random" effect, you can use normal generalized linear model (glm). Significance of interaction terms (e.g. year x species or area x species) could be used to justify your judgement. I've not used Statistica, but I'm sure that it has glm (and probably glmm). If not, R is freely available.

Do you research pollination ecology?

 

[biothanasis]

Hello Naoki!

Thank you very much for this detailed info!!!
I will try it through Statistica!

I study pollen circulation patterns in the atmosphere (volumetric sampler) and pollen production in the area I live in.
Pollen production analysis is very easy, but circulation data can be very tricky to analyse.
Thank you again!!!!!

 

[fbrem]

what patterns are you trying to analyze, species richness in a given sample from a site? would a simple repeated measures anova on some species richness index if you are looking at that or dominance index if you are looking at the amount/proportion of pollen each species contributes to the overall sample work?

 

[biothanasis]

Hi fbrem and thanks for your reply!

Generally I want to check whether concentrations of one taxon (or all of them together) differ in between the stations and years. Should any difference be found, is it statistically significant?

I suppose that from your suggestions, the second one would do for me, but concerning means do not actually "work", is there any analysis that could work with sums?

Thank you again!!!!

 

[ALToronto]

Can you post a chart of the data? Are you familiar with Design of Experiments for multi-plot designs? My concern is that if an ANOVA has not shown anything significant, your experimental error may well be too high to see any relationship. How are you dealing with all the zero values?

As for your question about sums, there are plenty of analysis methods that work well with sums of squares.

 

[biothanasis]

Hi AL-Toronto and thank you for your reply!

Unfortunatelly I am not familiar with all Statistical analysis, hence I am asking for help (do not know about Design of Experiments for multi-plot designs).
The problem with using mean+stdeviation related analysis si that they do not depict the "truth" between the comparing variables.
Zero values are indeed zero values. I have tried several transformations in order to eliminate zeroes and use Anova or related analyses, but I do not find it useful as a method, as I get the same results (expected...)

Could you please indicate those analysis methods you mentioned that work with sums of squares? TIA

I have been suggested to use time-series analysis (since my samplings are made repeatedly, the same day, for 2 consecutive years), but I have no idea how they work.

 

[TyroneGenade]

If the zero values are real zeros (and not sampling error---and there are stats models to tell whether the error is sampling size or experimental error) then you have to use them. You can't cherry-pick your data. To play devil's advocate further, if the methods you are using are not showing a significant difference could it be because there isn't one? If you think there is a difference (or should be) and your current data cannot resolve the issue then, as I would tell me students, you have to go back and get more data.

But, more seriously, I think you do need to show us what your data looks like and what you want to show from it (i.e. what are your hypotheses) as this will determine which statistical model would best suit your data. Simply going along trying different models until you get a significant results is utterly unscientific and the very idea makes the hairs on the back of my neck stand up.

 

[biothanasis]

I see your point Tyrone.

No, fortunatelly, I am not trying to find absolutely significant differences, I just want to use an analysis that suits my data.
I want to avoid misinterpriting my results, in the case I have used an unsuitable analysis or to use an analysis that will give the result in lets say 1 or 5 summarised diagramm(s), instead of 100 individual ones.

The zeroes are real zeroes, as I only use the data from the samplings I've taken. So if nothing is evident as pollen in a specific sample, then it gets zero value. I always used them in the analysis I have done so far.

Unfortunatelly I have not published these data yet, so I cannot give the real ones to anyone yet, but I could make an excel with the nature/image of my data.
How do I upload an excel file? Or I might upload a word file...

 

[biothanasis]

Excel file is 54KB and word file is 266KB.
All data have been changed with random values.
The forum cannot handle them I suppose, so perhaps
you could give me an email through pm?

Please let me know and thank you for the help!

 

[ALToronto]

You can create a public folder on any of the free cloud services - Box, Dropbox, Skydrive and others. Then upload your files to the public folder and post the link to it.

 

[biothanasis]

I'll try that! Thanks!