Dear Yugenraj,

We do a lot of biofilm characterisation studies using a combined biofilm assay in which we sequentially measure the strength of air-liquid (A-L) interface biofilms, attachment levels to the microcosm walls at the meniscus, and total growth or biomass in the microcosm, and use an ANOVA approach with post hoc Tukey-Kramer HSD or Dunn's method with a control to investigate differences between means (strains).

This approach, which sounds very similar to yours, generally works, but there are two significant assumptions we make. The first is simply a statistical requirement that the data (or more often, the residuals generated by the ANOVA) are Normally distributed, and we check this using a Shapiro-Wilk W test. If the data or residuals are not Normally distributed, it is possible sometimes to undertake an outlier analysis to identify which data are outliers and remove them one at a time until the new data set is Normal (or until you have removed so much data that you can no longer investigate differences between strains. If the latter happens, you could try to transform your data to make it more suitable (repeating the whole data-residual checking process), or abandon the parametric approach and choose an appropriate non-parametric test instead.

The second assumption is biological and is suggests that your experimental approach is biased towards expecting that there should be significant differences between your strains, or that you assume that you can grow cultures and undertake the assay so well that there is going to be no significant variation in your replicate measurements. We often find that our pseudomonad all produce pretty much the same strength biofilms which are often so weak that we are unsure if the maximum deformation mass assay (dropping small glass balls on to the biofilm surface until it breaks) is sensitive enough. If the data is Normal, then we can check this using a T-test and asking whether the mean is significantly different to zero. More often we see some variation in the attachment level data, but this can vary a lot better replicates as it is hard too was the microcosm vials out properly. Again, we can ask if a strain is producing a significant level of attachment by including blank microcosms and measuring the level of attachment for these (this is a Crystal violet assay). We can then undertake an ANOVA (assuming Normality) with a post hoc test of means using Dunn's method with a control, and this will indicate all strains which have attachment levels above the control (we have never seen a strain that has a level lower, and this makes no experimental sense). Finally, we can also look at growth levels, but here there is no 'lower limit' we might expect as all of our strains grow under the conditions we use. If we were concerned about that, I would measure growth levels by OD at the beginning of the assay (using replicate microcosms) and again at the end of the assay. These data could be compared a number of ways, for example by using a paired T-test, but I'd probably simply look at the difference in OD values and as whether the means differed significantly from zero using T-tests.

I suspect that you problem is that you are assuming that you can differentiate your strains in to high and low categories of biofilm-production, but this would require small variations in replicate measurements and high variation between strain means. If your data-residuals are Normal, you could look at this directly using a three-way ANOVA or GLM with replicate measurement, strain, biofilm category as effects. In order for this to work, you need sufficient replicates or there will be no statistical power available to differentiate between the effects. It may simply tell you that your strains are pretty much all the same.

You could then redo your catagorisation and try and split the strains into two groups, low and high biofilm producers. Use the mean of all data as your cut-off figure unless you know that the data-residuals in not Normal, in which case using the median would be a better approach. If you cannot differentiate your strains like this, it looks as if they are all too similar.

Regards, Andrew

Dear Yugenraj,

We do a lot of biofilm characterisation studies using a combined biofilm assay in which we sequentially measure the strength of air-liquid (A-L) interface biofilms, attachment levels to the microcosm walls at the meniscus, and total growth or biomass in the microcosm, and use an ANOVA approach with post hoc Tukey-Kramer HSD or Dunn's method with a control to investigate differences between means (strains).

This approach, which sounds very similar to yours, generally works, but there are two significant assumptions we make. The first is simply a statistical requirement that the data (or more often, the residuals generated by the ANOVA) are Normally distributed, and we check this using a Shapiro-Wilk W test. If the data or residuals are not Normally distributed, it is possible sometimes to undertake an outlier analysis to identify which data are outliers and remove them one at a time until the new data set is Normal (or until you have removed so much data that you can no longer investigate differences between strains. If the latter happens, you could try to transform your data to make it more suitable (repeating the whole data-residual checking process), or abandon the parametric approach and choose an appropriate non-parametric test instead.

The second assumption is biological and is suggests that your experimental approach is biased towards expecting that there should be significant differences between your strains, or that you assume that you can grow cultures and undertake the assay so well that there is going to be no significant variation in your replicate measurements. We often find that our pseudomonad all produce pretty much the same strength biofilms which are often so weak that we are unsure if the maximum deformation mass assay (dropping small glass balls on to the biofilm surface until it breaks) is sensitive enough. If the data is Normal, then we can check this using a T-test and asking whether the mean is significantly different to zero. More often we see some variation in the attachment level data, but this can vary a lot better replicates as it is hard too was the microcosm vials out properly. Again, we can ask if a strain is producing a significant level of attachment by including blank microcosms and measuring the level of attachment for these (this is a Crystal violet assay). We can then undertake an ANOVA (assuming Normality) with a post hoc test of means using Dunn's method with a control, and this will indicate all strains which have attachment levels above the control (we have never seen a strain that has a level lower, and this makes no experimental sense). Finally, we can also look at growth levels, but here there is no 'lower limit' we might expect as all of our strains grow under the conditions we use. If we were concerned about that, I would measure growth levels by OD at the beginning of the assay (using replicate microcosms) and again at the end of the assay. These data could be compared a number of ways, for example by using a paired T-test, but I'd probably simply look at the difference in OD values and as whether the means differed significantly from zero using T-tests.

I suspect that you problem is that you are assuming that you can differentiate your strains in to high and low categories of biofilm-production, but this would require small variations in replicate measurements and high variation between strain means. If your data-residuals are Normal, you could look at this directly using a three-way ANOVA or GLM with replicate measurement, strain, biofilm category as effects. In order for this to work, you need sufficient replicates or there will be no statistical power available to differentiate between the effects. It may simply tell you that your strains are pretty much all the same.

You could then redo your catagorisation and try and split the strains into two groups, low and high biofilm producers. Use the mean of all data as your cut-off figure unless you know that the data-residuals in not Normal, in which case using the median would be a better approach. If you cannot differentiate your strains like this, it looks as if they are all too similar.

Regards, Andrew

Added file.