A problem often noted with a "sandbox" is that the players flounder, not being sure of what they could do, much less what they want to do, because they don't have the knowledge of the sandbox to form proactive goals and assess how to implement them.
The biggest problem with this discussion, which someone above also noted, is that we often proceed on the basis that "Railroad always bad, Sandbox always good".
Yeah. We really need better words to describe what we are talking about.
Reading a script is not a fun game, but neither is wandering about aimlessly with no clue what to do next.
It's not merely a matter of lack of proactivity. Most people playing an RPG unconsciously expect that the game will develop some sort of story through play and that this story will in some way be like the stories generated by other literary media. But if you run a pure sandbox without scripting and without effectively chasing down the players and making things happen to them, the big problem is that you tend to produce a long series of disconnected events that don't collectively form a story. What I found, both as a player and a DM, that this works well only for inexperienced players for which the experience of an RPG is fairly new. When you are 12, you don't really care that there isn't a story, you are too engrossed in fighting goblins, looting dungeons, and so forth. All of the game is both mechanically and thematically fresh. After a few years of doing that hardcore 8 hours a week though, it sort of runs into the problem that you've already done all this before. You've fought the monsters. You've seen all manner of traps. You've got high level characters with great gear. Now what? All the proactivity in the world doesn't necessarily create a story, any more than ones own real life easily novelizes into a compelling narrative with rising and falling action, resolution and epiphany.
I have tried to run sandboxes before. And the problem you run into is player buy in when they look around and don't see a story to pursue and nothing with real lasting meaning seems to happen. Even if you've got these cool plots hidden in various parts of the sand, if the players don't engage with them then they might as well not exist. One of the problems I had was that I had all these secrets that were legitimately secret, but no means or reason why the PC's would uncover the secrets or having uncovered them recognize them as having meaning. Real sandboxes can just overwhelm players with detail. There aren't any red herrings because there is no plot! That's in practice the same as saying "Everything is a red herring." How are the players to know which of these grains of sand are special and important and more importantly interesting enough that they'll allow the players to participate in a heroic story like what they would read in a fantasy story?
Some tables and some systems try to get around this problem by moving tangentially to the linear/non-linear axis. In both the sandbox and the adventure path you have tons of 'myth'. That is to say, on the slice of possibilities we are discussing the notion that that there is a concrete setting preexisting the players is largely assumed. The idea is do away with the setting, which in theory allows you to produce stories with lots of play buy in, but without fixed goals or events because the setting itself isn't fixed and is generated in response to player proactivity. It sounds good in theory, but in practice it has a lot of huge traps. One of them is that when you do away with a concrete setting, you remove the one security that the players have that they characters actually have agency. Theoretically, if the players are being allowed to create the myth based on their metagame desires, then the players have agency. But if the GM is not being constrained by his prior agreed upon myth, the problem is the GM is not being constrained at all. The myth that a GM creates for himself and his commitment to stay true to that myth is the one real limit on GM power. How can you say you overturning the GM's 'will' for the game, if the GM is completely free to create any contingency that they want?
In fact, what 'no myth' games seem to do is get players to accept buying a ticket on a railroad, and then letting the GM railroad however he likes. They even openly promote railroading techniques as ways to improve the story, with the GM empowered to create whatever he wants at any moment to have the story move in the direction the GM at that moment thinks will be best. If the GM wants the bad guy to escape because it's good for the story - a classic sign of railroading - in these games that's what he's supposed to do. Now, I'm not saying that's necessarily wrong. Railroading techniques can be used to improve the story. But fundamentally, all the nar techniques in the world are little different than saying, "Find a GM that creates fun railroads." A GM that can't create a really fun railroad for you to buy into, also won't be able to sit in the director's chair of a nar game and make it fun for the players. Why? Because the tool set that both GM's are using will be identical. In both games, if the GM decides that the ground opening up and swallowing you makes for a better story, that's what happens and both GMs will justify it in the same way.
A few systems even note this potential trap and so try to come up with ways to route around it. The usual way they do this is have players be able to decide what is best for the story, so that they aren't just at the mercy of the GMs designs. But the ways to route around it have additional traps, the most obvious of which is that it's just no fun to be the person who both introduces the conflict or problem, and the person who resolves it. If you wanted to just tell a story to yourself, you'd write a novel. Stray too far into PC's being able to control the myth, and pretty soon there is no real exploration, no real secrets to discover, and no real surprises. Even worse, too much of the play described by those systems reminds me not of the experience of being in a fantasy story, but rather of the experience of sitting around a table collaboratively working on a movie script. Even though the end transcription of play might look like a good story, the way you arrived at that story and the experience of arriving at the story is very different.
To short version of this post is I don't think we've described the space of GMing well enough to outline the shape of 'good GMing practice' with confidence. This single axis of 'Railroad/Sandbox' is not only not well labeled, because it's not objectively true that a sandbox is better than a railroad, but it's only one dimension that describes a very complicated space of GMing technique. There are things out there that we don't describe as railroads, even though they rely heavily on techniques that in another context, we'd call railroading, because they also have these additional techniques we don't normally use in railroads - like giving player's metagame agency. But we don't have the language to talk about that with shared understanding. I can say to someone, "All no-myth games are railroads", and I know what I mean. But they don't know what I mean, and so hear something like, "All no-myth games are bad.", when what I mean is that projected on to a single axis no-myth games almost invariably use the same GMing techniques that railroads use. That isn't to say that they are indistinguishable, because there is at least one and probably many ways they can differ. But if you are measuring just one dimension of complex things with just one ruler, you end up treating herrings and laptops as the same thing because 'length'.