For your Likert scored items, the standard procedures is to evaluate each of them with Cronbach's alpha, and if they are internally consistent, then sum the items in each scale. In general, the continuous variables will be better predictors than high/low versions because the latter "throw away data" -- but you could certainly test the effectiveness of the dichotomous versions versus the continuous versions, if only to replicate the previous work.

With regard to whether there are interesting combinations of high-high, low-low etc. The first thing to do is to correlate the two tests. If that correlations is quite high (e.g. .9 and above), then there is little likelihood that your data includes many high-low cases. The reason for choosing such a high value of r is that even .9 accounts for just 81% of the variance, which leaves room some room for unusual cases. The best way to investigate that issue is through outlier analysis, which is available when you run a regression that treats one of the IQ scores as a dependent variable and the other as the predictor.

Finally, to check for gender differences, you need to introduce interaction effects. Two equivalent ways to do this are through regression with dummy variables or Analysis of Covariance (ANCOVA).

Thanks for your answer !

Still some issues are unclear to me. First, the Cronbach's alpha test the internal consistency and  should be taken anyway. But I want to test also the correlations between each of the items in the personality scale with IQ scores, seeing what specific items in this scale are predictors for intelligence. Can I do that using the GLM procedure but in way that would mimic forward regression, just with two or more dependent variables in my case, or that it can only do it with one dependent variable each time using simple linear regression? 

Second, if I have to use the GLM procedure when I test two or more dependent variables at the same time, can it include outlier analysis and would it indicate special cases when the correlations of the intelligence scores with the personality scale are too different from each other? (considering that the correlation between the two intelligence measures is ~0.9).

Last thing, just wanted to mention that each intelligence scale is made of two different sub scales (verbal and quantitative intelligence scores), but guess it doesn't change much the answers, doesn't it? 

I am not familiar with SAS, so I can't comment on how to do your analysis using GLM.

I'm not sure why you want to examine the personality scales on an item-by-item basis. the usual "measurement theory" is that each item in a scale is an indicator of the same general concept, so they should be highly inter-correlated because of that common content. So, if a multi-item scale is correlated with IQ, then you would expect the individual items to be correlated as well. 

In general, you have two options when working with a pair of independent variables. One is to run two separate but parallel analyses, which essentially amounts to an internal replication. The other would be to do some kind of Structural Equation Modeling where you treat the two indicators of intelligence as indicators of the same concept. If you go with SEM, you could include a model that takes into account the sub-scales on IQ..

As far as outlier analysis goes, the idea is to use the one measure as a predictor of the other, and then determine whether there any observations that are not well accounted for (i.e., that are not just small random errors around the regression line). With the high correlation, it doesn't seem likely that there will be enough outliers for a separate analysis, unless you have a very large sample.

Thanks again! 

As for item by item, I should elaborate. The personality scale is made of three sub -scales, I want to measure the correlation between each of them (general score composed of few items actually so it's not really item by item) and between intelligence. Assuming that the sub scales are different in their correlation with intelligence.

I believe that the most straightforward way to do what you are proposing is through a regression that uses all of the personality variables as predictors of IQ. If you use the standardized coefficients from that regression, these will be partial correlations that give each sub-scale's correlation with IQ after controlling for the other sub-scales.

Given the high correlation between the two IQ variables, you could combine them into a single more reliable scale, or if you have some reason to differentiate them, you could run two separate regressions.