I haven't heard that before. If the sample population meet parametric assumptions, why not? Normality, homogeneity of variance, and independence of observations.

A t-test is not suitable for count data because count data is usually skewed -- many smaller values, few higher values. Non-parametric tests are one approach. Another is to use an appropriate transform, usually the logarithm. Yet another is to use a generalized linear model with a Poisson, Gamma or Negative Binomial distribution.

You can transform data. You can use log transformation (x' = log x+1) or Square root transformation (x'= √(x+0,5)), and then, do the analyses.

To follow up on Jurandy's answer: you do the +1 int he log transform if there are zero counts, because log(0) is undefined. Otherwise, you can just use the log transform. I prefer the log because it has an intuitive interpretation: log(a) + log(b) = log(a*b). Also http://pareonline.net/getvn.asp?v=8&n=6 has some nice points.

Thank you to the contributors so far. I am still waiting for more.

Thank you for that important precision, Nina

If by parametric tests you mean those assuming normally-distributed errors, then the issue is simple: count data only take on discrete (integer) and non-negative values and thus errors are rarely even approximately normally distributed. You can treat the count data as counts per se (non-negative integers). Many processes that generate counts have Poisson distributions, so a parametric test based on Poisson rather than Gaussian (normal) error distribution may be appropriate. These go by various names from Poisson Regression (even if the predictor is categorical) to Generalized Linear (mixed) Models. [Logistic Regression & Generalized Linear Models with binomial error are equivalents for yes/no or presence/absence binomial responses.] You still need to test the distribution of your error variation for meeting the assumptions (overdispersion and zero-inflation are two common cases where Poisson distributions don't fit, and where alternatives have been developed).

Thank you Tom

Counts data are not suitable to be analyzed by parametric test because usually most of them do not follow the major assumption of parametric test, viz: counts must be normally distributed, the variance between groups must be approximately similar.

Counts data usually are over dispersed, that is, the variance increases when the mean increases or decreases when the mean decreases too or rarely when mean is equal to variance. Since counts rarely follow normal distribution, they can be either modeled or analyzed using parametric test that assumes other distribution rather than normal such as: Poisson distribution, Quasi-Poisson distribution, Gamma distribution, Binomial and Negative Binomial distribution (the so-called General Linearized Models); depending on whether you dependent variable is binary, integer or non-negative integer values.

For instance, according to my field experience and has been also found elsewhere, Parasites and Insect counts rarely follow normal distribution (variance used to be greater than the mean). They (counts) used to be successfully modeled using negative binomial distribution.

While I do realize this question has already been answered a long time ago, a Google search directed me here and I'd like to share some interesting information if found here, so others may find it more easily. Specifically there is a lot to say AGAINST (log) transforming count data, nicely summarized in this blog post: http://www.r-bloggers.com/do-not-log-transform-count-data-bitches/ where I specifically like the likelihood ratio testing described to be performed in R. Other relevant information on generalized linear models and poisson errors can be found here: http://www3.imperial.ac.uk/pls/portallive/docs/1/1171927.PDF and in the case where you need to model excess zeros (such as might be possible for next generation sequencing data) I'd like to refer to this package's documentation: http://cran.r-project.org/web/packages/pscl/vignettes/countreg.pdf