
#1




Fall 2019 MASII Thread
Looks like we're in need of this thread.
I'm starting with the syllabus source texts. Anyone have recommendations on other materials? I'd prefer to have a large pool of high quality questions. My current goal is to finish reading all of the source texts and any problem sets they include by July 31, leaving August, September, and October for review and practice exams. 
#2




Welp, active thread.
I ended up purchasing the Infinite Actuary course. I think I'm one of two people so far subscribed for Fall, because as I finish lessons they switch to saying "50% have finished," so the other person may not have started yet for Fall. No wonder this forum is empty. I've only watched the prereq vids so far which was a nice refresher, most of it's relatively still in my head from MASI. Tomorrow I'm finishing up Buhlmann credibility in the Tse textbook. I'm looking to look to get a minimum of 4 hours of study total, so if there's extra time I'll go back and do additional credibility problems from the old SOA C problem set, and watch TIA credibility videos and work examples up through where I'm at in the text. I'm just working at a layered approach, read textbooks first and do all problems there according to my schedule, then where I'm ahead of schedule watch TIA for what I've covered earlier and try to embed the concepts through follow up exposure. 
#3




Hi, I'll help this thread be more active.
Going to start looking through the source material this week. Not sure if I will buy any materials yet. TIA seems useful but is also extremely expensive. This might be the only exam where I go 100% source material/past exams. How are you finding the TIA material? 
#4




Hey, glad to have additional participants.
The TIA seminar has 202 sample questions, and I think 104 inlecture examples that can be added for a total of 306 practice problems, and one practice exam. Dave Revelle is the instructor and I have a high opinion of him from previous exposure. I've only made it through the prerequisite videos but they've been helpful. There are flashcards, but unless I'm mistaken they only cover credibility and Linear Mixed Models, which leaves the biggest section of the exam on Bayesian Models uncovered. It is expensive but discounted to about $700 since it's not as full of a course as their other more established ones. I mainly am concerned that textbook problems and 2 official exams alone will leave me without enough context to know what types of problems to expect and whether I'm fast enough, so I'm hoping I can use the TIA problems and practice exam to beef that up, and videos certainly make review speedier, but I'm sticking to the source materials for a first learning experience on each section, just to make sure I'm seeing exactly what the exam committee is seeing. I think ACTEX is the only other manual option out there and it's tempting to buy that as well because I'm concerned I'll go through every problem I have available and then still have 2 months left where I'm only rehashing the same problems. This might be an exam where I'll have to create my own problems, and make my own formula sheets, which would be a new experience, but maybe would really force me to learn the material vigorously. Last edited by grape; 05202019 at 12:05 PM.. 
#5




So I found (and pretty sure others felt a similar way) that there was not a sufficient pool of MASII practice question just due to the nature of the material and the fact the exam is relatively new. I found it MUCH more helpful to go through the source material at least twice, and time permitting go through sections you still don't understand more.
I really only found the questions in Tse's book (source material for objective A.) to be around what is expected on the exam (it actually has old exam questions so you know it is somewhat accurate). I felt as though the questions from other sources, while informational and would be helpful to take a look at/work through them, focused too much on setting up R/software code which is not on the exam. These can be helpful if you work through them an try to understand the output, but again you can't really be sure if you are interpreting it right if you don't have the correct setup or if it is a subject response. Since R is the format for any output they would expect you to interpret on the exam, I would highly suggest working through any R code that is in the text itself. It is helpful to actually see the results firsthand as opposed to just looking at a table in the book, and you get a lot more output if you do it on your own so you may find something that makes the idea click or lead to more questions. I do not recommend, the Actex study manual at all. It really only provided a summary of the source materials and nothing beyond that. I thought the questions were impeccably easy and again not really accurate to types of questions on the exam for the most part. I did not use TIA so I cannot comment on that. In my opinion, I would definitely spend a lot more time focusing on the source materials, really reading and making sure you grasp the information in those as opposed to what types of questions they will ask. Objective A was really the only objective I think that can have a variety in the type of questions asked, but again as long as you understand the fundamentals, you should be able to complete any question asked. Not sure how everyone else studies, but I also found it helpful to highlight things as I read and then go back and create my own summary sheet for each chapter/section, to help it stick in my brain. I sat for the Spring examg, so still waiting for results and depending on that, my opinions could be completely irrelevant if I failed miserably. Either way good luck to all of you on this exam!! 
#6




Quote:
My first readthrough end date target is August 1, but my guess is that I will exhaust the available question sources in the first couple weeks in August if I'm exclusively drilling problems. After that I might have to just reread material and summarize it in different ways, maybe make flash cards to drill key concepts and even important sentences/paragraphs that hint at possible questions. I did happen upon what could be a good resource for learning, Richard McElreath (our textbook author for Bayesian models) has a course up on youtube covering his textbook with 20 lectures available. If I watch those once or twice in addition to doing my own careful reading of the text that could help. Statistical Rethinking Winter 2019 I'm also considering installing all the recommended R packages and following along with all of the examples. I know R isn't on the exam, but maybe actually doing all of the examples will force some familiarization with the topics. I'm just looking for a strategy to cover all of the info thoroughly so many times in different ways that it gets ingrained by exam day. The credibility section is comforting in that I have no concerns about it. But I'll be starting the Linear Mixed Model text this weekend, with no idea how best to approach it, other than reading, taking notes, and answering questions in the text. 
#7




My recommendation would be to actually start with the Statistical Rethinking text rather than Linear Mixed Models. It's a better written and more initially accessible textbook, and the principles covered in later chapters help explain the usefulness of linear mixed models more than the LMM text itself.
McElreath's lectures are a fantastic resource, agreed. You definitely have to read the text to get the full picture, but the later chapters made more sense with the accompanying lectures in mind. I'd do the coding exercises as well, if just to get some muscle memory behind the examples he provides. Keep in mind with the LMM text that it's front loaded. The second chapter covers 95% of the technical material, and it's a total slog to get through. The remaining chapters are applied exercises that touch on some useful material, but seem to mostly be there for reinforcing the concepts from chapter 2. I found it more effective to skim the chapter, read everything else, and come back to it once I knew how its concepts would be used. 
#9




Really appreciate the insight but I did not follow your advice, no offense. I worked out my schedule already and felt like rearranging everything wouldn't be worth it, and I don't mind hitting the hard material sooner to give me more chances to digest it over longer periods of time.
Anyway, got through chapter 2 of the LMM text this weekend and found it pretty engaging and interesting. It does feel that the bulk of syllabus items in the LMM section are referring to Chapter 2 concepts, and I went through the Fall 2018 exam questions on LMM and felt they were nearly all answerable with what I already knew after that chapter. Still, this week and into next I'll be working through all of the case studies and cementing everything. I've been thinking about whether an LMM collapses to a traditional general least squares linear model after you test the random effects and determine that you want to remove them from the model. I suppose if you still have the residual variance matrix R then it's still an LMM model even though you only have fixed effect parameters, right? In that case though, we already have residuals in the normal linear model that are normally distributed, so wouldn't that be considered a mixed model? What if your R matrix is diagonals only? Is the marginal model considered a standard linear model? Another question that I had that hasn't been addressed in the material yet, can we use all of the different response distributions in LMM that we learned in the MASI material for the general linear model? If we use a distribution other than g(u)=u, Do all of the model specifications and assumptions needed for LMM still hold up? I suppose once I dig into the software I may see whether these things are available. One thing that frustrates me about this book a little is that most of the diagnostic tools and hypothesis tests and parameter estimators they do give you, they don't prove, and in some case they define one term that I would expect to be on the exam with a term that's not defined in the text that sounds like it has a technical meaning. Things like Hessian Matrix, PositiveDefiniteness, Rank of a Matrix, etc. This list is not exhaustive. I'm unclear how far I should dig into these things outside of the text since it doesn't seem reasonable to me they could ask questions on them, but I feel like my knowledge is incomplete because they aren't defined. Anyway, good to know I'm over the worst hump (chapter 2) and excited to get through to the Bayesian textbook given the reputation it has for being more readable. 
#10




I think I got an answer to one of my questions above in the TIA materials for LMM. Dave goes through an example with the Rat Pups to explain how you could use general least squares to predict birth weight without the random effects, and you get one model, call it version a, with fixed beta parameters with values {a}. Then you fit the model with liter specific random effects and everything else the same, say model b, the beta parameters come out a bit different {b}, and you have a marginal model for that model that works basically like a general least squares model for predictions, but because you fit the model with random effects, the fixed beta values are all slightly different.
So the marginal model is not equivalent to a least squares model. I think maybe I had it backwards too, you might not start with a LMM and then test the random effects, remove them, and then have a GLS model. You would more likely start with the GLS model and then look at the residuals to see if random effects are needed. I'm really confused about the difference between correlated u terms and correlated residuals, but hopefully that will make sense after some more digging in the examples. I guess with the rat pup birth weight, I can't imagine why one litter's random effect would be correlated with another if we're already controlling for litter size. I'll have to see some examples where different variancecovariance matrices setups are tested. 
Thread Tools  Search this Thread 
Display Modes  

