The correlation coefficient is not a measure of correlation, but of co-linearity, and coefficients < 0.5 are typically bunk, and the entire system of computing correlation coefficients is bunk if the data is nonlinear. Almost all interesting phenomena outside are non-linear, esp. biology and higher, where nonlinearities arise from the mutual interaction of parts of the system. Here, the correlation coefficients are in the ~0.2 range. You might as well be staring at monkey entrials.
Methodologically, you cannot construct explanatory models of weak-effect non-linear phenomena from observational data. Basically, you can fit any explanatory model you like, since you can make any parts of the system interact with any strength, and since they are non-linear, this will reproduce any distribution you so wish.
You can entirely reproduce any heritability distribution, genetic covariance, "shared" parent-child, "unshared" child-child, etc. you want by changing this model. Observational data here is basically useless at discriminating. ie., i can make a model where genes are 100% irrelevant, or 100% determinative, entirely consistent with the observational data.
The only method which can distinguish here is interventional, ie., you have to actually control the causes of the system. However, since we cannot breed different groups with different genes; nor take the same person and run their life with differnet parents, friends, etc. you're basically out of luck.
I'd prefer we closed this whole field down, and any person mentioning "heritability" outside of a wheat breeding lab, shuffled off to some discipline less catastrophically detrimental to social policy.