Hi, I have relative abundances of OTUs under different conditions and I´d like to see how some of the OTUs differ between treatments. Is it legit if I just choose one of the OTUs of interest and test its abundance undr different treatments with ANOVA? If not, why and what other test I can use? Thanks!
Hi Lenka, It sounds like you are using OTU (operational taxonomic unit) as a proxy for "species" for microbial communities. When you say relative abundance I imagine auger plates with various percentages of OTU's (microbes) for a given treatment. If this is the case then yes a one-way ANOVA would be appropriate for comparing differences among categorical independent variables (example: low, med, high treatments). If you are only looking at two treatments (example: present Vs. absent) a T-test would be more appropriate. However, you must make sure that your data is distributed normally (verify with histogram/bell shaped distribution) and has no difference in variance among treatments (Levene's Test) to meet parametric assumptions of ANOVA and T-test. If you don't meet these assumptions a non-parametric Kruskal-Wallis Test (ANOVA equivalent) or Mann Whitney-U Test (T-Test equivalent) will be appropriate provided the variance isn't too terribly different among treatments. If your independent variable is quantitative you can also run a simple linear regression (also parametric) to look for correlations (example: OTU abundance as a function of temperature, nutrients, etc.). Couple of things to think about, these types of tests are geared more towards an experimental design where you are only looking at how one particular OTU responds to empirical manipulation. My primary concern involves potential interactions among multiple OTU's within a specific treatment.
Hope this Helps!
Best
Drew
Hi Lenka, It sounds like you are using OTU (operational taxonomic unit) as a proxy for "species" for microbial communities. When you say relative abundance I imagine auger plates with various percentages of OTU's (microbes) for a given treatment. If this is the case then yes a one-way ANOVA would be appropriate for comparing differences among categorical independent variables (example: low, med, high treatments). If you are only looking at two treatments (example: present Vs. absent) a T-test would be more appropriate. However, you must make sure that your data is distributed normally (verify with histogram/bell shaped distribution) and has no difference in variance among treatments (Levene's Test) to meet parametric assumptions of ANOVA and T-test. If you don't meet these assumptions a non-parametric Kruskal-Wallis Test (ANOVA equivalent) or Mann Whitney-U Test (T-Test equivalent) will be appropriate provided the variance isn't too terribly different among treatments. If your independent variable is quantitative you can also run a simple linear regression (also parametric) to look for correlations (example: OTU abundance as a function of temperature, nutrients, etc.). Couple of things to think about, these types of tests are geared more towards an experimental design where you are only looking at how one particular OTU responds to empirical manipulation. My primary concern involves potential interactions among multiple OTU's within a specific treatment.
Hope this Helps!
Best
Drew
With one-way ANOVA or one of the non-paramtric tests cited you will see if there are almost one treatment different from the remaining ones. However, to prove which specifi treatments are different in average, you should apply a multiple comparison test, such as Tukey, t-test, Bonferroni, Skott-Knott, and others.
It is legitimate if you just choose one. The responses by Robson and Andrew should work. The problem is that people have trouble stopping. So you have 25 OTUs, or maybe 50. The more OTUs that you test the more likely it is that you will find a significant result. The problem is that multiple comparison tests like Tukey can control the experimentwise error rate for one test (with multiple treatments), but they do not control the manuscriptwise error rate. Your manuscript has 25 Tukey tests, and each test uses the magical 0.05 significance then for your manuscript you have an alpha of 1-(1-0.05)^25 = 0.72. It is therefore fairly likely that at least one of your 25 Tukey tests will have a significant difference even if no such difference exists. The problem gets much worse if you use a t-test, have 25 OTUs, and have 5 treatments for each OTU. The t-test does not control the experimentwise error, so you have ten tests per OTU times 25 OTUs and the resulting mess is uninterpretable.
If Andrew is correct in his guess .... try using treatment as a dependent variable and all the OTUs as independent variables in a multivariate analysis. Depending on your data you could also try things like cluster analysis or discriminate analysis. Be careful about possible interactions between your OTUs. If OTU1 is present does that reduce the probability (or colony size) of OTU7? You are looking to answer questions like:
1) Do the OTUs form functional groups?
2) Are some of the OTUs more responsive to the treatments?
3) Are some of the OTUs more important than others?
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis