In the next days I have to give a talk in which I need to explain some of the usual singularities of pairs that one meets when dealing with the minimal model program: KLT, DLT and LC pairs. In particular I would like to give an intuition (also to non specialists) of the reason why an LC pair is much more difficult to treat than a DLT pair, for example. In the same context I would also like to give an intuition of what the LC centers of a pair are and why they are "special" subvarieties for a pair. Note that I'm always considering normal (possibly nonsmooth) projective varieties. Do you have any ideas?

For references, probably you already know things like KollarMori and Kollar's "Singularities of pairs", for LCcenters and subadjunction, Kollar has some notes on that as well. There's some sections on this in the recent book by HaconKovacs too that I've looked at recently. With regards to your specific questions I have some comments but probably VA will have better comments: 1. an LC pair is much more difficult to treat than a DLT pair The archetypical LC pair is probably $(\mathbb{A^2}, Div(x) + Div(y))$, the two coordinate axes (EDIT: this example is also DLT, and doesn't really represent the most poorlybehaved properties of LCpairs). This is a simple normal crossings (SNC) pair and we understand how they behave pretty well with respect to numerous operations. On the other hand, if you have a pair $(X, D)$ where the pair is KLT, then even if the singularities of $D$ are bad, the fact that $(X, D)$ is klt means that you can perturb the coefficients of $D$ in many ways. This gives one a great amount of flexibility in numerous situations. DLT, by Szabo's criterion, is a combination of KLT and SNC with coefficients $\leq 1$. Basically, a DLT pair is LC and where it is not KLT, it is SNC (simply normal crossings), which we understand. Dually, where it's not SNC, it's KLT, and we can perturb things and do those other tricks. 2. what the LC centers of a pair are and why they are "special" subvarieties for a pair. The most basic example of a LCcenter of a pair $(X, D)$ is a prime divisor $D_i$ which is a component of $D$ and such that the $D_i$coefficient of $D$ is $1$. For example, if $X$ is smooth, and $D$ is a prime divisor, then $D$ is a LCcenter of $(X, D)$. Why is this nice? This lets us relate the canonical divisor $K_X + D$ of $(X, D)$ and the canonical divisor of $D$. By definition, LCcenters are exactly the images of divisors $D_i'$ where $D_i'$ is a component of $D'$ with coefficient $1$, and $(X', D')$ is a pair obtained by taking a log resolution $\pi$ of $(X, D)$ and setting $D' = \pi^*(K_X + D)  K_{X'}$. In particular, if $W$ is an LCcenter of a pair $(X, D)$, then one has $$(K_X + D)_W = K_W + (\text{some correction terms}).$$ Again, many properties of the pair $(X, D)$ can be translated to properties of $(W, (\text{correction terms}) )$, this isn't trivial to show. One should note that even if you have a 1dimensional LCcenter, then there can still be a correction factor (just not in the smooth example I described above). 


I would add an important distinction of LC singularities to Karl's excellent answer above: DLT (and in particular KLT) singularities are rational, while LC are not. This already appears when the boundary is empty. LC singularities are not even necessarily CohenMacaulay, such as for example a cone over an abelian variety of dimension at least 2. I think that this is the main reason for LC pairs being so much harder. For LC centers, it's also a good idea to look at Ambro's work on quasilog varieties. That seems to be a very effective way of handling LC centers. This is evidenced by this recent paper in JAMS and Kollár's local KawamataViehweg vanishing result (on the arXiv). Another comment on LC centers: this is probably obvious for most people, but the point of an LCcenter is that that's why a pair is not KLT. They are also called "nonklt" centers. I believe Christopher Hacon prefers this terminology. As far as I understand, LC centers are similar in spirit to associated primes. The simplest thing about a module is its support and in particular its irreducible components. However, considering associated primes gives a better understanding as there are components that kind of should be considered part of the support, but they are embedded, sort of "shadowed" by other components. The union of LC centers is exactly the nonklt locus, but just like with associated primes, there may be some that are embedded, so knowing the LC centers gives more refined information about the failure of the pair to be klt. 


I know very little of this subject, but I can attest to the value of it, since many of us studied the singularities of theta divisors for years without noticing or even conjecturing some basic results that followed from Kolla'r's observation that (A,D) is a log canonical pair if D is a theta divisor on a principally polarized abelian variety A. As Kolla'r put it with extreme simplicity and clarity: "we know a highly singular divisor moves, but a theta divisor doesn't move, so it should not be very singular." He deduced that the points of D of multiplicity ≥ k have codimension ≥ k in A. I particular there are no points on D of multiplicity > dim(A). The generalization of this result by Ein and Lazarsfeld that the extreme case occurs only for a product p.p.a.v., established the normality of an irreducible theta divisor. To me, this is a beautiful example of the concept of l.c. pairs which allowed the characterization of indecomposable p.p.a.v's in terms of the geometry of their theta divisors. This is one of the first cases of a technique that allowed the analysis of theta divisors that could not be understood in terms of curves, e.g. as Jacobians or Pryms. 

