I can't actually run your code without editing it. FWIW, when providing example code it is best to provide a reproducible example (see e.g. http://stackoverflow.com/questions/5963269/how-to-make-a-great-r-reproducible-example). For example, you use computer specific commands like `setwd()` which are guaranteed to fail if I try and run it.

In any case, I expect that what is happening is that the `kde2d()` function does some smoothing to produce kernel density estimates on the regular grid. If you look at the help page (`?kde2d`), or better yet have a look at the reference (Venables and Ripley), you'll see that one of the arguments describes the "bandwidth" to use for the x and y directions. Likely this bandwidth describes the amount of smoothing (akin to a search radius, I expect) desired for the kernel density estimate.

A contour of what? I would imagine that you mean that you are slicing the likelihood surface, and making contours on it, but you say you do not want to assume normality. Do you have a posterior likelihood?  Can you compute the value of your surface at given points?

I do have a posterior likelihood using the Bayesian Multivariate Logit Normal model and yes I can evaluate the posterior at any given points

You don't say if you can evaluate the posterior at any set of inputs. If so, and if you only have two parameters, then one approximation that could be useful is to find the mode of the posterior, then  multiply the modal likelihood by 0.05. Any point with posterior greater than that number is in a 95% credible region for the pair of parameters.

@James, Thanks for your thought. What would be the theoretical explaination behind it?