There exist estimates for the size of the largest root. The most general go back to the idea that $z$ is not a root of
$$
p(z)=a_nz^n+a_{n-1}+...+a_1z+a_0
$$
if $|z|>R>0$ with an outer root radius $R$ that satisfies the intequality
$$
|a_n|R^n\ge |a_{n-1}|R^{n-1}+...|a_1|R+|a_0|
$$
This polynomial inequality for $R$ is easier to solve numerically than zeroing in on any specific root of the original polynomial. Especially as for the further numerical purposes only a low relative accuracy is needed. The smallest $R$ is obtained as the only positive root of a polynomial with only one sign change in the coefficient sequence, meaning there is exactly one positive root. This situation allows for the secure use of simple scalar root-finding methods like the Newton method.
But one can also obtain simple (over-)estimates like
$$
R=\max\left(1,\frac{|a_{n-1}|+...+|a_0|}{|a_n|}\right)
$$
or
$$
R=1+\frac{\max_{k<n} |a_k|}{|a_n|}
$$
These estimates support the general idea, if the coefficients are small relative to the leading coefficient, then the roots will also be small.
The last bounds only give $R\ge 1$. To get beyond that restriction, "guess" a smaller scale $\rho$ and compute the root bound $R_\rho$ from $p_\rho(z)=p(\rho z)$. Then $R=\rho R_\rho$ gives a better bound. $\rho$ can be estimated as power of 2 from the exponent of the coefficients as floating-point numbers. The aim is that the coefficient sequence of $p_\rho$ is about balanced with the leading coefficient dominating and at least one other coefficient of similar magnitude.
I'd recommend studying the techreport to the Jenkins-Traub RPOLY method. I have some of that also reproduced in the corresponding Wikipedia article.