Defn: A field extension $K/F$ is obtained by adjoining $n$th roots if there is a tower of fields $F=K_1\subset\cdots\subset K_n=K$ such that for each $i$, $K_{i+1}=K_i(\alpha_i)$ and there exists some $n_i>1$ such that $\alpha_i^{n_i}\in K_i$.
I'm interested in whether subextensions $K/F$ of an extension $E/F$ obtained by adjoining $n$th roots must also be obtained by adjoining roots.
When $K/F$ is Galois, we can make use of the fact that Galois extensions are obtained by adjoining roots iff their Galois groups are solvable. Extend $E/F$ to its Galois closure, which is also obtained by adjoining roots, so $\mathrm{Gal}(E^{\mathrm{Gal}}/F)$ is solvable. Since $\mathrm{Gal}(K/F)$ is a quotient of $\mathrm{Gal}(E^{\mathrm{Gal}}/F)$ it is solvable as well, thus $K/F$ is obtained by adjoining roots.
This reduces the problem to the case $E=K^{\mathrm{Gal}}$. However, I have no idea how to usefully characterize non-Galois extensions which are obtained by adjoining roots, so I don't know how to approach this case.
I'm mostly interested in the case where $K$ is separable, but I'd also be interested in a counterexample where this is not the case.
Let $f(x)=x^3-x-1$, let $\alpha$ be a zero of $f$ in an extension of the rationals, let $E={\bf Q}(\alpha)$, let $K$ be a splitting field for $f$ over $E$. Then $K$ is obtained from the rationals by adjoining $n$th roots, but $E$ isn't.