Відео

@@kalechips965 Excellent question! It depends on your project. I would alter weights to address imbalance. If you need model score to equal the true probability, I would rescale model scores so that the mean training score equals the fraction of positive training. I would not about this if you do not need to interpret model scores. Find the cutoff that gives the best sensitivity, specificity, precision, recall etc.

@statisticsninja I appreciate the tips! One more question that’s a bit more complex… As mentioned above, my sample has a 5:1 class imbalance. I've done a stratified split according to this imbalance, creating (1) a training set for model development (within which cross-validation subsets are themselves split using stratified k-folds) and (2) a test/holdout set purely for final model evaluation. For reproducibility, I have set a fixed seed variable and passed it to any method that has a "random_state" argument. There are two issues. First, my cross-validated training and test metrics are similar, but these same metrics can be more than 10% lower for my final test set. Second, multiple runs of my script using different random seeds causes the overall results to vary appreciably. I think these issues are related. I believe the relatively matched validation training/test scores reflects good capability of my models to learn from each validation subset. However, the fact that the validation test scores tend to be higher than the (unseen) holdout test scores indicates to me that these models do not generalise well. I think this points to a data issue - I would guess that the overall sample is unbalanced or biased in ways that are not captured by the class-stratified sampling. For example, some important features may be underrepresented in certain subgroups of a class, and their distribution would vary significantly across splits, affecting classifier performance. Does this interpretation make sense to you?

@@wanqingtai1490 excellent!!!

@@minghuachang8126 You can use the citation() function in R to get the citation information for a package. Typically it references the journal article that announced the package.

@@nasheedjafri3564 If a variable has a steep slope positive or negative then you want to include it in your model.

@@fathymohamed4312 What type of spatial data do you have? The simplest approach would be to treat your spatial data as predictor variables. R has a lot more spatial tools than Python because R is more common in science.

@@statisticsninja thank you Sir

@@nadimalfana That is a good question. It is a subjective decision. Try to balance the goals and constraints of your project, and what you see in the data. Make the best call you can. There is usually a range of reasonable values.

@@statisticsninja Woah, that's tough decision haha... Thanks!

The easiest way is to get an sf object with the spatial features and a data.frame, then merge your data.frame with merge.sf(). If you have the spatial features without a data.frame(). Then you need to match your data.frame() to the geometries.

For that situation the data are all on the same measurement scale, so I would not normalize. If I had data on different scales, height in inches and weight in lbs, then I normalize.

I could not find a function to do this for you. You could manually compute regression diagnostics using the equation in BayesFM::befa Model specification. Use your befa to fill in everything except errors. Be sure to switch columns and signs of your befa output first.

How would I do this? The only output of BEFA are the factor loadings, variances of error terms, factor correlations and the indicator values. Is it possible to manually generate with only these values?

@@ThankfulAlways I found a better way. Use parameters::model_parameters, and parameters::efa_to_cfa to get the parameters from your befa model. Then fit a confirmatory factor analysis model using your preferred package. This will give your the full power of your favorite factor analysis package.

@@statisticsninja oh really? I will do some research and try it out. Never tried doing confirmatory factor analysis before. Not sure how it's different from exploratory factor analysis. Will read on that. Thanks a lot!

After you fit your model, enter your model into the str() function. It will print the slots of your model. You can use the slot names to extract what you need.

@@statisticsninja Thank you for reply, I hope you would share a video on how to obtain item and test information functions in multidimensional confirmatory IRT. 😀

@@MrArdahazal Which function are you using to fit your model?

@@statisticsninja Hi, I constructed a model as follows as far as I understand your presentation. Besides, I want to learn the test information function and item function parameters, but, I havent understood them. I tried to code them at below of the syntax but I am not sure. Model -------------------------------------- library(mirt) library(latticeExtra) cfa <- mirt::mirt.model( input=' pl = 1-9 sb = 10-15 db= 16-21 dk= 22-28 oy= 29-34 COV = pl*sb, pl*db, pl*dk, pl*oy, sb*db, sb*dk, sb*oy, db*dk, db*oy, dk*oy') ACS<- mirt(data=thesis, model=cfa, method= "MHRM", itemtype = 'graded', SE=FALSE, SE.type="MHRM", TOL=1e-2) ACST <- coef(ACS, IRTpars = TRUE, simplify = TRUE) options(max.print = 1000000) print(ACST, digits = 2) For the whole scale: test information matrix --------------------------------------- Theta <- matrix(seq(-4,4, by = .01)) thetas <- fscores(ACS, method = "EAP", rotate = "oblimin", QMC=TRUE) tinfo <- testinfo(ACS, thetas, degrees = c(0, 0, 0, 0, 0)) plot(thetas, tinfo, type="1") For Item 1: item information matrix ---------------------------------------- Theta <- as.matrix(expand.grid(-4:4, -4:4, -4:4, -4:4, -4:4)) iteminfo1 <- extract.item(ACS, 1) iteminfo<- iteminfo(iteminfo1, thetas, degrees = c(0, 0, 0, 0, 0), total.info= TRUE, multidim_matrix = TRUE) options(max.print = 1000000)

@@MrArdahazalI hope this helps. I posted the RMarkdown file on my website's shared files section. ua-cam.com/video/k_oNhQ9Fy6w/v-deo.html

The predictor matrix is all of your predictor predictors in a matrix or data frame. For multiple imputation, I prefer missRanger. A way to compare imputation methods would be to copy your data, ramdomly replace values with missing values, try several imputation methods, then compare the imputed values to the original values.

Thank you!

I have used R for analyzing satellite data. The project was a success. I would start by analyzing your data independentl of spatial coordinates, then analyze coordinate distribution independentl of data, perform full spatial data analysis.

t@@statisticsninja Thank you so much for your kind response and i hope to find your help in terms of guide me what course i should take to master R software analysing remote sensing data in short time as i started long time ago but unfortunitly i found i have to learn many things to be able to use R to analysis rmote sensing data .

Make sure your data is in a data.frame and not a tibble.

Make sure your data frame has only numeric or integer columns and pass your data frame to to.matrix() or matrix(). Make sure you are not using a tibble

@@statisticsninja okay let me check, thanks

​@@statisticsninjaWe are using a spatial polygon & point to intersect which is not a matrix. coordinates(o_yb) <- ~easting+northing #convert the locations into a SpatialPoint proj4string(o_yb) <- CRS("+init=epsg:27700") #for each order, get the ycodes which intersect the building boundaries o_yb <- do.call(rbind,lapply(o$order_id[o$in_building == 1], function(x) { #x <- 1 #testing t_bld <- bld[bld$fid %in% o_bld$fid[o_bld$order_id == x],] #get the building boundaries for the order do.call(rbind,lapply(o_yb@data$key[o_yb@data$order_id == x], function(x1) { # x1 <- "YDLHP" #testing t_o_yb <- o_yb[(o_yb@data$order_id == x & o_yb@data$key == x1),] t1_bld <- r_intersect(t_bld,t_o_yb) #check if the ycode is in the building if (length(t1_bld) == 0) return(NULL) t_o_yb@data[,c("order_id","key")] #if length is greater than 0, make a data.frame, else return null })) }))

You can check to see if there is a statistical difference between imputed and nonimputed data; you can run anomaly detection and see if imputed records are proportionally a lot of anomalies, you can also train another model on non-imputed data to predict the imputed column and check the residuals when you predict imputed values

@@statisticsninja Thank You for your help. Do you know of the what test I could you use ? I am a passionate amateur so I have some gap that I'm trying to fill.

@@abdulbouraa4529 I would compare the preimputed column to postimputed column by comparing the histograms, and running a Kolmogorov-Smirnov test, ks.test()

@@statisticsninja Thank you very much !

Python

For a mirt model from the mirt package, the itemfit() and M2() functions extract test statistics. You could also use the @Fit slot from the mirt object directly.

@@statisticsninja But is there a way to evaluate the Test Information Curves? Or some analogous function?

@@andresimi Which package and function are you using to fit your model?

@@statisticsninja I am using mirt package with the plot(type = "info") function. Right now, I split a multidmensional model into unidimensional ones so the curves are interpretable. I was wondering if this is ok, or if I have another way of doing this.

You can look at 2x2 subtables and run hypothesis tests to identify which conditional distributions differ from the Fisher exact test null hypothesis. Consider using a p-value correction such as Bonferroni. If your variables have an independent-dependent variables relationship, you can run chi-squared test on pairs of conditional distributions. The standardized residuals will show which cells are most different from the Fisher null hypothesis.

I posted the .dat files in the shared files section of my website. You can load them the same way you would load a .txt file

If you set save.lvs = TRUE when you fit your model, blavaan::blavPredict() with type = "lvmeans" will give the factor score of your fitted data. I could not get blavaan::blavPredict() nor predict() to work with newdata. For new newdata, extract coefficient estimates and multiply the observed values by their coefficients

The data set is on the book's website asdar-book.org The homepage has links to each chapter's data

@@statisticsninja Wow, thank you so much! I really appreciate the quick reply and your fabulous videos!

I posted my R markdown file to the shared files section of my website.