In an experiment where subjects are divided in several groups, we have a random factor A (subjects) nested within fixed factor B (groups). So random is nested within fixes. This I understand. But in the linked example a fixed factor is supposed to be nested within a random factor. How should I think about it? Something fixed nested within subjects? Is it even possible?
I expect this to be a real dumb question that somebody can quickly answer here...
@wolframalpha "formal" is almost an antonym of "rule of thumb"
@ssdecontrol a typical way to handle irregularly spaced time series is with continuous-time methods. (If it's a regularly spaced series with a lot of missing values you can easily deal with them using state-space methods; consequently an irregularly-spaced series with all inter-observation gaps a multiple of some fixed interval don't necessarily require anything more than standard ways of dealing with missingness in time series)
In part it depends on what models you want to estimate.
As long as you have the data to estimate the model you want to fit, a likelihood's a likelihood and a posterior is a posterior; there's EM algorithms and MCMC and so on that can sometimes be adapted.
I haven't heard of Eckner ... what sort of models does he use?
...
e.g. see Peter Brockwell "On the Use of Continuous-time ARMA Models in Time Series Analysis" Chapter in *Athens Conference on Applied Probability and Time Series Analysis*, pp 88-101 ... http://link.springer.com/chapter/10.1007%2F978-1-4612-2412-9_7
The reference list at that link gives lots of relevant books and papers.
@mark999 I spent several hours trying to understand this issue, so I guess I should post a proper question, even if it's dumb. Here it goes: stats.stackexchange.com/questions/232109.
@amoeba Yes the formula is correct. I don't have access to it at the moment, but a while ago I verified that using that aov formula gives results that are the same as those given in an example in "Applied Linear Statistical Models" by Kutner et al., and the same as what you get with an equivalent model with lme if the data is balanced.
Strange, here is a post stats.stackexchange.com/questions/60108 that contains a link to another post but it does not appear in the Linked list on the right. Is it a bug?
he hasn't published anything for quite some time. i was using "Algorithms for Unevenly Spaced Time Series: Moving Averages and Other Rolling Operators" to implement moving average smoothing
@Kodiologist indeed. i was leery of doing something like running a regression against a bunch of month/year indicators; would like to actually try and make reasonably accurate forecasts. maybe i'll throw some mock data together and post a question about it
@ssdecontrol The usual approach with lots of (usually imbalanced) timepoints is not to treat months or years as dummy variables but to treat time as a continuous variable and use transformations such as trig functions, polynomials, and splines, or model the relationship between time and outcomes nonparametrically.
Hi, I am using GradientBoosting and RandomForest Regressors , but I want a general method for tuning parameters( for ensemble methods) , @Glen_b I used exhaustive grid search and even randomized grid search using pre-defined sets of parameters which I guess is a formal way ? But it's sheer brute force :( , Is there something which continously evaluates the score to reach out the best possible model ?
@Kodiologist thanks. i've seen that done in papers but it's good to know that there are real humans doing it
@wolframalpha nothing wrong with brute force! as long as you're using a principled technique like cross-validation and an evaluation metric that reflects what you're interested in
if you're concerned about the stability of your parameter estimates, you can try taking high-scoring hyperparameter values and perturbing them slightly to see how much predictive performance degrades. also you can run repeated or nested cross-validation to estimate the variance of your predictive performance
in order to continuously evaluate the score at all possible parameter values in some domain, you'd need a closed-form solution for that score. we can only do that in very specific cases, like lasso/ridge regression
BTW if you're using gradient boosting and random forests, you might want to check out XGBoost. it's nothing conceptually new (gradient boosted trees) but it performs incredibly well in my experience. it's a C++ lib, with Python and R wrappers
Well, I don't know. It's not really cheating, you did solve the problem. But you didnt need all the fancy neural network technology to do so. You solved it by sitting around and thinking about it, and then writing down a solution.
It turned up on the Late Answer queue, I think I put it down as "No Action Needed" in the sense that it did seem to be a valid answer to the question, though I am somewhat sceptical of the degree to which it is a useful one..