The dichotomous variable I want to use for a point-biserial correlation has very unequal n's ($n_1=140$, $n_2=6$). Would this be breaking any rules? What would this mean for my results? Would it just limit the statistical power?
I would first look at a scatterplot of the variables to see if they are linear before running an analysis. In addition, see Kraemer's 1980 paper,Robustness of the Distribution Theory of the Product Moment Correlation Coefficient, in which it is noted,
Robustness of normal test theory for correlation coefficients is at least asymptotically ensured for bivariate distributions satisfying a linearity and a homoscedasticity condition for the null theory and a further kurtosis condition for the nonnull theory. If any one of these conditions fail, it may be demonstrated that robustness may fail as well. (Abstract)