This question is somewhat related to [this one][1].

As Alan has already said, following the actual path of the light ray through each layer leads to more physically accurate results. I will base my answer on a paper by Andrea Weidlich and Alexander Wilkie (["Arbitrarily Layered Micro-Facet Surfaces"][2]) that I have read and implemented. In their paper they assume that the distance between two layers is smaller than the radius of a differential area element. This simplifies the implementation because we do not have to calculate intersection points separately for each layer, actually we assume that the intersection points are the same for all layers. 

According to the paper, two problems must be solved in order to render multilayered material. The first one is to properly sample the layers and the second is to find the resulting BSDF generated by the combination of the multiple BSDFs that are found along the sampling path.

**Sampling**

In this first stage we will determine the actual light path through the layers. When a light ray is moving from a less dense medium, e.g. air, to a more dense medium, e.g. glass, part of its energy is reflected and the remaining part is transmitted. You can find the amount of energy that is reflected through the [Fresnel reflectance equations][3]. So, for instance, if the Fresnel reflectance of a given dielectric is 0.3, we know that 30% of the energy is reflected and 70% will be transmitted:

[![enter image description here][4]][4]

When the light ray is moving from a more dense to a less dense medium, the same principle described by the Fresnel reflectance applies. However, in this specific case, [total internal reflection][5] (a.k.a TIR) might also happen if the angle of the incident ray is above the critical angle. In the case of TIR, 100% of the energy is reflected back into the material:

[![enter image description here][6]][6]

When light hits a conductor or a diffuse surface, it will always be reflected (being the direction of reflection related to the type of the BRDF). In a multilayer material, the resulting light path will be the agregate result of all those possibilities. Thus, in the case of a 3-layer material, assuming that the first and secong layers are dielectrics and the third layer is diffuse, we might end up with the following light path (a tree actually):

[![enter image description here][7]][7]

We can simulate this type of interaction using recursion and weighting each light path according to the actual reflectance/transmitance at the corresponding incident point. A problem regarding the use of recursion is that the number of rays increases with the deepness of the recursion, concentrating computational effort on rays that individually might contribute almost nothing to the final result. On the other hand, the aggregate result of those individual rays at deep recursion levels can be significant and should not be discarded. In this case, we can use Russian Roulette (RR) in order to avoid branching and to probabilistic end light paths without losing energy, but at the cost of a higher variance (noisier result). In this case, the result of the Fresnel reflectance, or the TIR, will be used to randomly select which path to follow. For instance:

[![enter image description here][8]][8]

As can be seen, TIR or Fresnel reflectance might keep some rays bouncing indefinitely among layers. As far as I know, [Mitsuba][9] implements plastic as a two layer material, and it uses a closed form solution for this specific case that accounts for an infinity number of light bounces among layers. However, Mitsuba also allows for the creation of multilayer materials with an arbitrary number of layers, in which case it imposes a maximum number of internal bounces since no closed form solution seems to exist for the general case. As a side effect, some energy can be lost in the rendering process, making the material look darker than it should be. 

In my current multilayer material implementation I allow for an arbitrary number of internal bounces at the cost of longer rendering times (well... actually, I've implemented only two layers.. one dielectric and one diffuse :).

An additional option is to mix branching and RR. For instance, the initial rays (lower deep levels) might present substantial contribution to the final image. Thus, one might choose to branch only at the first one or two intersections, using only RR afterwards. This is the case with [smallpt][10].

An interesting point regarding multilayered materials is that individual reflected/transmitted rays can be importance sampled according to the corresponding BRDFs/BTDFs of the current layer. 

**Evaluating the Final BSDF**

Considering the following light path computed using RR:

[![enter image description here][11]][11]

We can evaluate the total amount of radiance $L_r$ reflected by a multilayer BSDF considering each layer as a individual object and applying the same approach used in ordinary path tracing (i.e. the radiance leaving a layer will be the incident radiance for the next layer). The final estimator can thus be represented by the product of each individual Monte Carlo estimator:

$$ L_r =  \left( \frac{fr_1 \cos \theta_1}{pdf_1}  \left( \frac{fr_2 \cos \theta_2}{pdf_2} \left( \frac{fr_3 \cos \theta_3}{pdf_3} \left( \frac{fr_2 \cos \theta_4}{pdf_2} \left( \frac{L_i fr_1 \cos \theta_5}{pdf_1} \right)\right)\right)\right)\right)$$

The [paper][2] by Andrea Weidlich and Alexander Wilkie also takes absorption into consideration, i.e. each light ray might be attenuated according to the absorption factor of each transmissive layer and to the distance traveled by the ray within the layer. I've not included absorption into my renderer yet, but it is just a real coefficient computed according to the [Beer's Law][12].

**Alternate approaches**

The Mitsuba renderer uses an alternate representation for multilayered material based on the "tabulation of reflectance functions in a Fourier basis". I have not yet dig into it, but might be of interest: "[A Comprehensive Framework for Rendering Layered Materials][13]" by Wenzel Jacob *et al*. There is also an expanded version of this paper.


  [1]: https://computergraphics.stackexchange.com/q/4611/5681
  [2]: https://dl.acm.org/citation.cfm?id=1321292
  [3]: https://en.wikipedia.org/wiki/Fresnel_equations
  [4]: https://i.sstatic.net/S4zHY.png
  [5]: https://en.wikipedia.org/wiki/Total_internal_reflection
  [6]: https://i.sstatic.net/Li0lu.png
  [7]: https://i.sstatic.net/WBz0l.png
  [8]: https://i.sstatic.net/t5ewd.png
  [9]: https://www.mitsuba-renderer.org
  [10]: http://www.kevinbeason.com/smallpt
  [11]: https://i.sstatic.net/GWv6b.png
  [12]: https://en.wikipedia.org/wiki/Beer%E2%80%93Lambert_law
  [13]: https://dl.acm.org/citation.cfm?id=2601139