You should not be initiating force on a rational agent. Although it's ambiguous if the kid can be considered a rational agent 3, 4 even 5 years old, you should never ever hit kids when they're 8 for example. You should also never steal their property. It seems so blatantly obvious to me, you want your child to respect other people and their property. Hitting them in no way facilitates this, it actually promotes it. It tells them that violence is an acceptable means to an end, more specifically, initiating violence on someone if they do something you don't like.
Interesting how black and white you think it is.
I was caught stealing before, and I was caught using violence. And, I was both punished physically, and had items of equal value reciprocally taken from me. I didn't ever do it again. The reason why? Because it taught me empathy for those my actions had consequence on.
It's kind of funny how the white knights pretend to be experts in this field. Honestly, I don't actually know what quantitative effects hitting children has, but at least I'm not claiming one way or the other with "children are impressionable" as my justification.
Edit:
Having read above, I noticed someone claiming that a sample size of 700 was quite enough, and someone claiming that it wasn't. Let me break it down:
You have a population of 1.9 billion kids or so, according to Ghast. 700 children represents a very, very small portion of that population. Now, certainly, a sample of that size may be used to say something about the population within error, to the point where you can theorise about the population. However, that's only one sample. ONE sample. In statistics, one sample is not hard evidence or proof - it is simply a representation of the population at the particular point in which the sample was taken. You cannot possibly make the sample 100% random, which means in order for the data to mean anything, either a larger sample size proportional to the population size, or more importantly, multiple different samples from all over the place at different times and under different conditions need to be taken.
For example, recently in a laboratory I was testing quality of various thermocouples in aqueous conditions - I took multiple samples both under same and under different conditions and every sample yielded SIGNIFICANTLY different results.
700 people is a perfectly good sample size. 1 sample is not enough samples. Therefore that statistical information is not enough to form any conclusions about the majority of the population. Having said all of this, I didn't actually read the paper mentioned - I'm just going of what the posts were saying - so if I've mistaken something about the data then please feel free to say so.