555 reputation
38
bio website
location
age
visits member for 3 years, 9 months
seen Sep 17 at 14:00

I'm an ABD statistician, and work fulltime in this capacity. I have interest in Computational Statistics, Markov Chain Theory, and Sample Survey.


Sep
6
revised How do I solve $E\left[ E \left(X|Z \right) E\left( Y|Z \right)\right]$?
Added clarification on X, Y, and Z being random variables, removed \times to be more consistent with undergraduate text book notation (Hogg and Craig etc...).
Sep
6
suggested suggested edit on How do I solve $E\left[ E \left(X|Z \right) E\left( Y|Z \right)\right]$?
Sep
4
answered Alternative to z test for massive samples?
Jun
24
awarded  Informed
Jun
24
comment “Better” goodness-of-fit tests than chi squared for histogram modeling?
@AdamO: In model selection usually some form of regularization such as parsimony is employed, this provides uniqueness. If two models really are equivalent, then you can define an equivalence class, but in application there is usually some form of regularization such as distribution preference. I also don't understand you're interpretation of splines. Densities are functions, and splines in particular low rank splines are perfectly good at estimating functions, furthermore with low rank splines you can just choose a fixed number of knots and use the coefficients as a family of densities.
Jun
21
comment Interpreting the mathematical formula of a mixed effect model
Maybe I'm wrong here, but shouldn't $x_{0ij}=1$ or maybe some element of a design matrix? I'm also unsure of the notation $\beta_{0ij}$, as a simple thought experiment how many parameters would you have if this were true, how many observations? You might want to post a link to this notation, maybe something has been taken out of context.
Jun
21
comment “Better” goodness-of-fit tests than chi squared for histogram modeling?
I think this problem isn't entirely well defined. It is entirely possible that your test results could be tied between different distribution families. Furthermore it may be hard to tell a multi-modal distribution from a single modal distribution in general for smaller sample sizes. Is there a particular reason why the distribution is so important? Perhaps fitting a spline or another "non-parametric" approach may be more useful?
Feb
14
comment Spatial autocorrelation of non-homogeneous point data in irregularly shaped sample plots
This question looks fairly interesting; looking at the data within each parcel of land, and your interest only in the presence, it seems that you would be looking at some form of multivariate Poisson process. This unfortunately isn't my area of expertise. Before you jump down the spatial rabbit hole though, you might want to first just ignore the spatial component initially and see how your results hold through a contingency table using the number of occurrences.
Feb
8
comment R is taking ages to cluster documents, what other options do I have?
A reasonable approach to such a problem would to run with a number of randomly selected subsets of the 40,000 documents of known and increasing sizes that run fairly quickly (e.g. 50 documents, 100 documents, 500 documents, 1000 documents, etc...). Given this information you should at least have an approximate estimate for the expected amount of time the problem should take (e.g. fit a set of polynomials to the observed runtimes and extrapolate to 40,000 documents).
Feb
7
comment Normal distribution in R
@Corone don't worry about it; you did the R code which was the actual question. Feel free to add my comments to your answer so we have a comprehensive resource.
Feb
7
comment Normal distribution in R
The intuitive explanation is that the T-distribution accounts for the increased variability due to not knowing the true variance. A slightly more rigorous explanation is that what you are really doing is a hypothesis test on the mean where the variance is a nuisance parameter; the test-statistic $(\bar{x} -\mu) / \sqrt{S^2/n} \sim T(n-1)$ is invariant to the value of $\sigma^2$ and has some other nice properties (Uniformly Most Powerful (Unbiased) ) under certain hypotheses.
Feb
7
comment Normal distribution in R
You might want to try a T-distribution instead since we don't know the variance.
Feb
7
comment Normal distribution in R
Anomalous needs to be defined; since the outcomes (anomalous or not anomalous) is binary this problem should map to a formal T-test. Furthermore, are we making assumptions about independence, and normality with the sample?
Feb
7
awarded  Editor
Feb
7
revised Normal distribution in R
Corrected minor typos, formatting, and clarified incoherent statements.
Feb
7
suggested suggested edit on Normal distribution in R
Feb
1
comment When is quantile regression worse than OLS?
I think most researchers would entertain both OLS and quantile regression; differences between the methods would shine light on what you are trying to model. With respect to OLS, if you toss in normality assumptions you do get a lot of fairly well documented and thorough testing methodology that is available in most statistical packages.
Jan
18
comment Exploratory data analysis
Thanks, you might want to update your question with this information though. So you have 20 covariates or potential covariates, and you also have a set of random effects? something of the form $Y_i = x_i \beta + Z_k \alpha + \epsilon$ where $Z_k$ is your random effect $\sim iid(0,\sigma_{\alpha}^2)$ and $\epsilon_i$ is distributed $\sim iid(0,\sigma^2)$ independent of $Z_k$.
Jan
18
comment Exploratory data analysis
What does your covariate data look like? e.g. continuous, ordinal, categorical (some mix of these). Furthermore you haven't really made any assumptions so it's hard to know what checks are useful, e.g. if you are doing OLS you are really just concerned about the error term having mean zero and constant variance (note that I have not made normality a requirement).
Jan
17
awarded  Critic