The analysis of phenotypic data should always begin by checking the quality of the data.
First, one should check whether the data are within the range that is expected.
For example, a pH of 30 should ring many bells; the value 30 is probably obtained by misplacing the decimal point.
Measurement scale
Changing the measurement scale may give a totally different look at the data.
Our perception of an outlier may change if the data are transformed.
Distributions
Most standard analyses of phenotypic data require that the observations follow more or less a normal distribution with constant variance,
i.e. the variance does not depend on the mean.
In many situations we have to deal with observations of which the distribution of the data is far from normal.
For example, for counts the Poisson distribution may be more appropriate.
For the Poisson distribution the variance is equal to mean. So, if the mean increases (e.g. due to a change in environment),
also the variance increases. In order to obtain a constant variance we would need to take the square roots of the observations.
However, thinking of genotypic and environmental effects in terms of square roots makes no sense.
The current approach to analyzing count data would be to assume that the variance is proportional to the mean and that differences
between genotypes and environments are considered on a logarithmic scale (which means that on the original scale we must not look at
differences but at ratios). Such models are usually referred to as generalized linear models.
Model building and model checking
After a model has been fitted to the data we have to check whether the assumptions we have made in the model are supported by the data.
A simple plot of standardized residuals (i.e. standardized differences between observations and fitted values) versus fitted values may show
whether the model assumptions about the variance are appropriate. It may also reveal outliers.
Influential data
We should also check whether one or a few observations have a large influence on the results.