I'm using regsubsets from the leaps library to perform the best subset selection. I need to compare the coefficients it generates to the "true" coefficients I specified when simulating the data (by comparison, meaning, the difference between them squared, and the square root taken of the sum), for each number of predictors.
Since there are 16 different models that regsubsets generated, I use a loop to do this automatically. It would work except that when I extract the coefficients from the best model fit with x predictors, it only gives me the non-zero coefficients of the polynomial fit. This messes up the size of the coefi vector causing it to be smaller in size than the truecoef true coefficients vector.
If I could somehow force all coefficients to be spat out from the model, I wouldn't have an issue. But after looking extensively, I don't know how to do that.
Alternative ways of solving this problem would also be appreciated.
library(leaps)
regfit.train=regsubsets(y ~ poly(x,25, raw = TRUE), data=mydata[train,], nvmax=25)
truecoef = c(3,0,-7,4,-2,8,0,-5,0,2,0,4,5,6,3,2,2,0,3,1,1)
coef.errors = rep(NA, 16)
for (i in 1:16) {
coefi = coef(regfit.train, id=i)
coef.errors[i] = mean((truecoef-coefi)^2)
}
The equation I'm trying to estimate, where j is the coefficient and r refers to the best model containing "r" coefficients:
Thanks!
This is how I ended up solving it (with some help):
The loop indexes which coefficients are available and performs the subtraction, for those unavailable, it assumes they are zero.
truecoef = c(3,0,-7,4,-2,8,0,-5,0,2,0,4,5,6,3,2,2,0,3,1,1)
val.errors = rep(NA, 16)
x_cols = colnames(x, do.NULL = FALSE, prefix = "x.")
for (i in 1:16) {
coefis = coef(regfit.train, id = i)
val.errors[i] = sqrt(sum((truecoef[x_cols %in% names(coefis)] -
coefis[names(coefis) %in% x_cols])^2) + sum(truecoef[!(x_cols %in% names(coefis))])^2)
}
Related
New to stackoverflow. I'm working on a project with NHIS data, but I cannot get the svyglm function to work even for a simple, unadjusted logistic regression with a binary predictor and binary outcome variable (ultimately I'd like to use multiple categorical predictors, but one step at a time).
El_under_glm<-svyglm(ElUnder~SO2, design=SAMPdesign, subset=NULL, family=binomial(link="logit"), rescale=FALSE, correlation=TRUE)
Error in eval(extras, data, env) :
object '.survey.prob.weights' not found
I changed the variables to 0 and 1 instead:
Under_narm$SO2REG<-ifelse(Under_narm$SO2=="Heterosexual", 0, 1)
Under_narm$ElUnderREG<-ifelse(Under_narm$ElUnder=="No", 0, 1)
But then get a different issue:
El_under_glm<-svyglm(ElUnderREG~SO2REG, design=SAMPdesign, subset=NULL, family=binomial(link="logit"), rescale=FALSE, correlation=TRUE)
Error in svyglm.survey.design(ElUnderREG ~ SO2REG, design = SAMPdesign, :
all variables must be in design= argument
This is the design I'm using to account for the weights -- I'm pretty sure it's correct:
SAMPdesign=svydesign(data=Under_narm, id= ~NHISPID, weight= ~SAMPWEIGHT)
Any and all assistance appreciated! I've got a good grasp of stats but am a slow coder. Let me know if I can provide any other information.
Using some make-believe sample data I was able to get your model to run by setting rescale = TRUE. The documentation states
Rescaling of weights, to improve numerical stability. The default
rescales weights to sum to the sample size. Use FALSE to not rescale
weights.
So, one solution maybe is just to set rescale = TRUE.
library(survey)
# sample data
Under_narm <- data.frame(SO2 = factor(rep(1:2, 1000)),
ElUnder = sample(0:1, 1000, replace = TRUE),
NHISPID = paste0("id", 1:1000),
SAMPWEIGHT = sample(c(0.5, 2), 1000, replace = TRUE))
# with 'rescale' = TRUE
SAMPdesign=svydesign(ids = ~NHISPID,
data=Under_narm,
weights = ~SAMPWEIGHT)
El_under_glm<-svyglm(formula = ElUnder~SO2,
design=SAMPdesign,
family=quasibinomial(), # this family avoids warnings
rescale=TRUE) # Weights rescaled to the sum of the sample size.
summary(El_under_glm, correlation = TRUE) # use correlation with summary()
Otherwise, looking code for this function's method with 'survey:::svyglm.survey.design', it seems like there may be a bug. I could be wrong, but by my read when 'rescale' is FALSE, .survey.prob.weights does not appear to get assigned a value.
if (is.null(g$weights))
g$weights <- quote(.survey.prob.weights)
else g$weights <- bquote(.survey.prob.weights * .(g$weights)) # bug?
g$data <- quote(data)
g[[1]] <- quote(glm)
if (rescale)
data$.survey.prob.weights <- (1/design$prob)/mean(1/design$prob)
There may be a work around if you assign a vector of numeric values to .survey.prob.weights in the global environment. No idea what these values should be, but your error goes away if you do something like the following. (.survey.prob.weights needs to be double the length of the data.)
SAMPdesign=svydesign(ids = ~NHISPID,
data=Under_narm,
weights = ~SAMPWEIGHT)
.survey.prob.weights <- rep(1, 2000)
El_under_glm<-svyglm(formula = ElUnder~SO2,
design=SAMPdesign,
family=quasibinomial(),
rescale=FALSE)
summary(El_under_glm, correlation = TRUE)
Bert-toolkit is a very nice package to call R functions from Excel. See: https://bert-toolkit.com/
I have used bert-toolkit to call a fitted neuralnet (avNNnet fitted with Caret) within a wrapper function in R from Excel VBA. This runs perfect. This is the code to load the model within the wrapper function in bert-toolkit:
load("D:/my_model_avNNet.rda")
neuraln <- function(x1,x2,x3){
xx <- data.frame(x1,x2,x3)
z <- predict(my_model_avNNET, xx)
z
}
Currently I tried to do this with a fitted GAM (fitted with package mgcv). Although I do not succeed. If I call the fitted GAM from Excel VBA it gives error 2015. If I call the fitted GAM from a cell it gives #VALUE! At the same time the correct outcome of the calculation is shown in the bert-console!
This is the code to load the model in the wrapperfunction in bert-toolkit:
library(mgcv)
load("D:/gam_y_model.rda")
testfunction <- function(k1,k2){
z <- predict(gam_y, data.frame(x = k1, x2 = k2))
print (z)
}
The difference between the avNNnet-model (Caret) and the GAM-model (mgcv) is that the avNNnet-model does NOT need the Caret library to be loaded to generate a prediction, while the GAM-model DOES need the mgcv library to be loaded.
It seems to be not sufficient to load the mgvc-library in the script with the GAM-model which loads the GAM-model in a wrapper function in bert-toolkit, as I did in the code above. Although the correct outcome of the model is shown in the bert-console. It does not generate the correct outcome in Excel.
I wonder how this is possible and can be solved. It seems to me that maybe there are two instances of R running in bert-toolkit.
How can I load the the mgcv-library in such a way that it can be used by the GAM-model within the function called from Excel?
This is some example code to fit the GAM with mgcv and save to model (after running this code the model can uploaded in bert-toolkit with the code above) :
library(mgcv)
# construct some sample data:
x <- seq(0, pi * 2, 0.1)
x2 <- seq(0, pi * 20, 1)
sin_x <- sin(x)
tan_x2 <- tan(x2)
y <- sin_x + rnorm(n = length(x), mean = 0, sd = sd(sin_x / 2))
Sample_data <- data.frame(y,x,x2)
# fit gam:
gam_y <- gam(y ~ s(x) + s(x2), method = "REML")
# Make predictions with the fitted model:
x_new <- seq(0, max(x), length.out = 100)
x2_new <- seq(0, max(x2), length.out = 100)
y_pred <- predict(gam_y, data.frame(x = x_new, x2 = x2_new))
# save model, to load it later in bert-toolkit:
setwd("D:/")
save(gam_y, file = "gam_y_model.rda")
One of R's signatures is method dispatching where users call the same named method such as predict but internally a different variant is run such as predict.lm, predict.glm, or predict.gam depending on the model object passed into it. Therefore, calling predict on an avNNet model is not the same predict on a gam model. Similarly, just as the function changes due to the input, so does the output change.
According to MSDN documents regarding the Excel #Value! error exposed as Error 2015:
#VALUE is Excel's way of saying, "There's something wrong with the way your formula is typed. Or, there's something wrong with the cells you are referencing."
Fundamentally, without seeing actual results, Excel may not be able to interpret or translate into Excel range or VBA type the result R returns from gam model especially as you describe R raises no error.
For example, per docs, the return value of the standard predict.lm is:
predict.lm produces a vector of predictions or a matrix of predictions...
However, per docs, the return value of predict.gam is a bit more nuanced:
If type=="lpmatrix" then a matrix is returned which will give a vector of linear predictor values (minus any offest) at the supplied covariate values, when applied to the model coefficient vector. Otherwise, if se.fit is TRUE then a 2 item list is returned with items (both arrays) fit and se.fit containing predictions and associated standard error estimates, otherwise an array of predictions is returned. The dimensions of the returned arrays depends on whether type is "terms" or not: if it is then the array is 2 dimensional with each term in the linear predictor separate, otherwise the array is 1 dimensional and contains the linear predictor/predicted values (or corresponding s.e.s). The linear predictor returned termwise will not include the offset or the intercept.
Altogether, consider adjusting parameters of your predict call to render a numeric vector for easy Excel interpretation and not a matrix/array or some other higher dimension R type that Excel cannot render:
testfunction <- function(k1,k2){
z <- mgcv::predict.gam(gam_y, data.frame(x = k1, x2 = k2), type=="response")
return(z)
}
testfunction <- function(k1,k2){
z <- mgcv::predict.gam(gam_y, data.frame(x = k1, x2 = k2), type=="lpmatrix")
return(z)
}
testfunction <- function(k1,k2){
z <- mgcv::predict.gam(gam_y, data.frame(x = k1, x2 = k2), type=="linked")
return(z$fit) # NOTICE fit ELEMENT USED
}
...
Further diagnostics:
Check returned object of predict.glm with str(obj) and class(obj)/ typeof(obj) to see dimensions and underlying elements and compare with predict in caret;
Check if high precision of decimal numbers is the case such as Excel's limits of 15 decimal points;
Check amount of data returned (exceeds Excel's sheet row limit of 220 or cell limit of 32,767 characters?).
I am new to neural networks and the mxnet package in R. I want to do a logistic regression on my predictors since my observations are probabilities varying between 0 and 1. I'd like to weight my observations by a vector obsWeights I have, but I'm not sure where to implement the weights. There seems to be a weight= option in mx.symbol.FullyConnected but if I try weight=obsWeights I get the following error message
Error in mx.varg.symbol.FullyConnected(list(...)) :
Cannot find argument 'weight', Possible Arguments:
----------------
num_hidden : int, required
Number of hidden nodes of the output.
no_bias : boolean, optional, default=False
Whether to disable bias parameter.
How should I proceed to weight my observations? Here is my code at the moment.
# Prepare data
train.mm = model.matrix(obs ~ . , data = train_data)
train_label = train_data$obs
# Normalize
train.mm = apply(train.mm, 2, function(x) (x-min(x))/(max(x)-min(x)))
# Create MXDataIter compatible iterator
batch_size = 128
train.iter = mx.io.arrayiter(data=t(train.mm), label=train_label,
batch.size=batch_size, shuffle=T)
# Symbolic model definition
data = mx.symbol.Variable('data')
fc1 = mx.symbol.FullyConnected(data=data, num.hidden=128, name='fc1')
act1 = mx.symbol.Activation(data=fc1, act.type='relu', name='act1')
final = mx.symbol.FullyConnected(data=act1, num.hidden=1, name='final')
logistic = mx.symbol.LogisticRegressionOutput(data=final, name='logistic')
# Run model
mxnet_train = mx.model.FeedForward.create(
symbol = logistic,
X = train.iter,
initializer = mx.init.Xavier(rnd_type = 'gaussian', factor_type = 'avg', magnitude = 2),
num.round = 25)
Assigning the fully connected weight argument is not what you want to do at any rate. That weight is a reference to parameters of the layer; i.e., what you multiply in the inputs by to get output values These are the parameter values you're trying to learn.
If you want to make some samples matter more than others, then you'll need to adjust the loss function. For example, multiply the usual loss function by your weights so that they do not contribute as much to the overall average loss.
I do not believe the standard Mxnet loss functions have a spot for assigning weights (that is LogisticRegressionOutput won't cover this). However, you can make your own cost function that does. This would involve passing your final layer through a sigmoid activation function to first generate the usual logistic regression output value. Then pass that into the loss function you define. You could do squared error, but for logistic regression you'll probably want to use the cross entropy function:
l * log(y) + (1 - l) * log(1 - y),
where l is the label and y is the predicted value.
Ideally, you'd write a symbol with an efficient definition of the gradient (Mxnet has a cross entropy function, but its for softmax input, not a binary output. You could translate your output to two outputs with softmax as an alternative, but that seems less easy to work with in this case), but the easiest path would be to let Mxnet do its autodiff on it. Then you multiply that cross entropy loss by the weights.
I haven't tested this code, but you'd ultimately have something like this (this is what you'd do in python, should be similar in R):
label = mx.sym.Variable('label')
out = mx.sym.Activation(data=final, act_type='sigmoid')
ce = label * mx.sym.log(out) + (1 - label) * mx.sym.log(1 - out)
weights = mx.sym.Variable('weights')
loss = mx.sym.MakeLoss(weigths * ce, normalization='batch')
Then you want to input your weight vector into the weights Variable along with your normal input data and labels.
As an added tip, the output of an mxnet network with a custom loss via MakeLoss outputs the loss, not the prediction. You'll probably want both in practice, in which case its useful to group the loss with a gradient-blocked version of the prediction so that you can get both. You'd do that like this:
pred_loss = mx.sym.Group([mx.sym.BlockGrad(out), loss])
I have fitted a GARCH process to a time series and analyzed the ACF for squared and absolute residuals to check the model goodness of fit. But I also want to do a formal test and after searching the internet, The Weighted Portmanteau Test (originally by Li and Mak) seems to be the one.
It's from the WeightedPortTest package and is one of the few (perhaps the only one?) that properly tests the GARCH residuals.
While going through the instructions in various documents I can't wrap my head around what the "h.t" argument wants. It says in the info in R that I need to assign "a numeric vector of the conditional variances". This may be simple to an experienced user, though I'm struggling to understand. What is it that I need to do and preferably how would I code it in R?
Thankful for any kind of help
Taken directly from the documentation:
h.t: a numeric vector of the conditional variances
A little toy example using the fGarch package follows:
library(fGarch)
library(WeightedPortTest)
spec <- garchSpec(model = list(alpha = 0.6, beta = 0))
simGarch11 <- garchSim(spec, n = 300)
fit <- garchFit(formula = ~ garch(1, 0), data = simGarch11)
Weighted.LM.test(fit#residuals, fit#h.t, lag = 10)
And using garch() from the tseries package:
library(tseries)
fit2 <- garch(as.numeric(simGarch11), order = c(0, 1))
summary(fit2)
# comparison of fitted values:
tail(fit2$fitted.values[,1]^2)
tail(fit#h.t)
# comparison of residuals after unstandardizing:
unstd <- fit2$residuals*fit2$fitted.values[,1]
tail(unstd)
tail(fit#residuals)
Weighted.LM.test(unstd, fit2$fitted.values[,1]^2, lag = 10)
I am running a regression with 37 variables, and I am using stepAIC to perform model selection. I do NOT want a predictive model. I just want to find out what varibles have the best explanatory power.
My current code looks like:
fitObject <- lm(mydata)
DEP.select <- stepAIC(fitObject, direction = 'both', scope= list(lower = ~AUC), trace = F, k = log(obs))
# DEP is my dependent variable, and AUC is an independent variable I was want to have in my model.
The problem is that a lot of my variables have high correlation, and the result stepAIC gives me contains several of those highly correlated variables. Notice that I have forced AUC in the model, multicollinearity is a problem especially when those variables highly correlated with AUC are chosen in the model.
Is there a way to specify in the function some thresholds for correlation or p-value of the coefficients?
Or any comments on other approaches that can solve my problem are welcome.
Thank you!
Perhaps Variance Inflation Factor will work better for you. This article explains some of the logic. http://en.wikipedia.org/wiki/Variance_inflation_factor
Example use:
v=ezvif(df,yvar ='columnNameOfWhichYouAreTryingToPredict')
Here is the function I wrote that combines VIF::vif with cross validation.
require(VIF)
require(cvTools);
#returns selected variables using VIF and kfolds cross validation
ezvif=function(df,yvar,folds=5,trace=F){
f=cvFolds(nrow(df),K=folds);
findings=list();
for(v in names(df)){
if(v==yvar)next;
findings[[v]]=0;
}
for(i in 1:folds){
rows=f$subsets[f$which!=i]
y=df[rows,yvar];
xdf=df[rows,names(df) != yvar]; #remove output var
vifResult=vif(y,xdf,trace=trace,subsize=min(200,floor(nrow(xdf))))
for(v in names(xdf)[vifResult$select]){
findings[[v]]=findings[[v]]+1; #vote
}
}
findings=(sort(unlist(findings),decreasing = T))
if(trace) print(findings[findings>0]);
return( c(yvar,names(findings[findings==findings[1]])) )
}
I would recommend to remove the variables with high correlations. The libraries caret and corrplot can help:
library(corrplot)
library(caret)
dm = data.matrix(mydata[,names(mydata) != 'DEP'] #without your outcome var
Visualize your correlations clustering highly correlated together
corrplot(cor(dm), order = 'hclust')
And find the indices of variables that you could remove due to high (>0.75) correlations
findCorrelations(cor(dm), 0.75)
Removing these variables can improve your model. After removing the variables, continue doing the stepAIC as you described in your question.
To assess multicollinearity between predictors when running the dredge function (MuMIn package), include the following max.r function as the "extra" argument:
max.r <- function(x){
corm <- cov2cor(vcov(x))
corm <- as.matrix(corm)
if (length(corm)==1){
corm <- 0
max(abs(corm))
} else if (length(corm)==4){
cormf <- corm[2:nrow(corm),2:ncol(corm)]
cormf <- 0
max(abs(cormf))
} else {
cormf <- corm[2:nrow(corm),2:ncol(corm)]
diag(cormf) <- 0
max(abs(cormf))
}
}
then simply run dredge specifying the number of predictor variables and including the max.r function:
options(na.action = na.fail)
Allmodels <- dredge(Fullmodel, rank = "AIC", m.lim=c(0, 3), extra= max.r)
Allmodels[Allmodels$max.r<=0.6, ] ##Subset models with max.r <=0.6 (not collinear)
NCM <- get.models(Allmodels, subset = max.r<=0.6) ##Retrieve models with max.r <=0.6 (not collinear)
model.sel(NCM) ##Final model selection table
This works for lme4 models. For nlme models see: https://github.com/rojaff/dredge_mc