What are the important conditions in ANOVA fixed effects?

Question

I am working with an ANOVA model. I want to run a fixed effects ANOVA in which I have a ratio dependent variable and three independent variables with two and three levels. Obviously, before analyzing the results, I want to check the assumptions of factorial ANOVA. Reviewing some handbooks, I found some divergences in their explanations about assumptions of ANOVA. Moreover, I have doubts about the underlying assumptions of factorial ANOVA. The major ones are:

1. All of the handbooks that I check point out that the dependent variable in ANOVA models should be at least an interval variable. I work with a count variable, in fact converted to a ratio variable (i.e. a percentage). So, is ANOVA appropriate in this case?

2. All of the handbooks stress the importance of checking the assumptions of the ANOVA model for inference: mainly a) independence, b) normality, and c) homogeneity of variances. However, they examine this aspects in different ways. Some of them check the data, i.e. the independence of cases, normality of each group and homogeneity of variances between groups. But others examine only the residuals (error) derived from the analysis (i.e. the independence, normality and homoscedasticity of the residuals).

So, I am confused about the appropriateness of my approach, but also about the assumptions that I should review. What does the ANOVA model require? Parametric assumptions for variables, only for residuals or both? References are welcome.

Answer 1

It should be clear that a counted fraction $X$ from $n$ can't follow a normal distribution however it's expressed—it has a discrete probability mass function whereas the normal has a continuous density. Nevertheless, the distribution of the proportion $P=\frac{X}{n}$ will approximate that of the normal more closely as the sample size $n$ increases. A bigger problem for the general linear model, of which ANOVAR is an instance, is heteroskedasticity: the bounds on proportions imply that the variance of $P$ varies with its mean. The main motivation for the angular transformation you mention is to stabilize the variance. If the function $f(P)$ is approximated by a Taylor series around $\pi$, the mean of $P$,

$$\newcommand{\d}{\mathrm{d}}\newcommand{\var}{\operatorname{Var}}f(P) \approx f(\pi) + (P-\pi)\frac{\d f(\pi)}{\d \pi}$$

then its variance, assuming a binomial distribution for $X$, is given by

$$\var f(P) \approx \var\left[ f(\pi) + (P-\pi)\frac{\d f(\pi)}{\d \pi}\right] \\ \approx \left(\frac{\d f(\pi)}{\d \pi}\right)^2 \var(P) \\ \approx \left(\frac{\d f(\pi)}{\d \pi}\right)^2 \frac{\pi(1-\pi)}{n}$$

For this function to achieve approximately constant variance you require $$ \left(\frac{\d f(\pi)}{\d \pi}\right)^2\propto \frac{1}{\pi(1-\pi)} $$ & thus $$ f(\pi) \propto \int{ \sqrt{\frac{n}{\pi (1-\pi)}}\, \d \pi} \\ \propto\operatorname{asin}\sqrt{\pi} $$

When the machines turn against us this will again be invaluable knowledge; until that time comes follow @Henrik's advice & model discrete data as what they are.