First it was 500 dog owners out of ~1000 respondents ~= +/— 4.4%. So, Double that to 2000 people and you can get 3% accuracy on overall per ownership, assuming no systemic bias. But, slice that into 30-39 year olds and you have ~1/7 the data and thus much higher error bars.

Now, if they had sampled ~20,000 people then sure you can slice and dice like that. But slicing data always increases the risk of sampling bias as would anything that targets 30-39 year olds specifically.