can we estimate coefficients, r2, residuals for rolling window regression in single function?-CodePudding

I ran this code to generate the coefficients, r2, adj.r2, and residuals for rolling window regression.

window <- (13)
ler <- function(j) {
  model <- dynlm(y ~ L(y,1)   x1   x2   x3, data= dd[(1:window) j,])
  c <- model$coefs
  r2<- model$r.squared
  a.r2<- model$adj.r.squared
  e <- model$residuals
  list(c,r2,a.r2,e)
})
out <- pbapply::pblapply(0:(nrow(dd)-window), ler)

while I am able to get residuals, the other (coefs, r2, adj.r2) were coming as NULL. how can I get all these for the rolling window model?

CodePudding user response：

The input data is missing from the question so we will use the built in 11x8 anscombe data frame with a window size of 5 and just use the y1 and x1 columns. (All columns in anscombe are numeric so if the actual input you are using has any non-numeric columns remove them first.)

We use the dyn package which also loads zoo. This gives an 11 row matrix with one row for each input row. Note that dplyr clobbers R's lag so in case you have dplyr loaded we have restored it.

library(dyn)
lag <- stats:::lag # this line is only needed if dplyr is loaded

out <- rollapplyr(anscombe, 5, function(x) {
  fm <- dyn$lm(y1 ~ lag(y1, -1)   x1, zoo(x))
  s <- summary(fm)
  c(coef(fm), r2 = s$r.squared, a.r2 = s$adj.r.squared, e = resid(fm))
}, by.column = FALSE, fill = NA)
out

giving:

      (Intercept) lag(y1, -1)         x1         r2       a.r2         e.2        e.3         e.4         e.5
 [1,]          NA          NA         NA         NA         NA          NA         NA          NA          NA
 [2,]          NA          NA         NA         NA         NA          NA         NA          NA          NA
 [3,]          NA          NA         NA         NA         NA          NA         NA          NA          NA
 [4,]          NA          NA         NA         NA         NA          NA         NA          NA          NA
 [5,]    5.737196   0.2051911 0.05566628 0.04018973 -1.8794308 -0.88226314 -0.3069362  1.01645850  0.17274084
 [6,]    2.674691   0.6139670 0.09652978 0.29958165 -1.1012551 -0.61664914  0.6126708 -0.81556822  0.81954660
 [7,]    9.614862  -0.3998369 0.24367243 0.92448518  0.7734555  0.03285018 -0.4426953  0.26436572  0.14547936
 [8,]   -2.807541   0.7134064 0.47205105 0.98990153  0.9697046 -0.34013131  0.2161506  0.10970662  0.01427410
 [9,]    3.886087  -0.1126389 0.55864348 0.86605967  0.5981790 -0.80881355  1.1239357 -1.04515518  0.73003301
[10,]    3.504280  -0.2036625 0.68004056 0.82863726  0.4859118  1.68395442 -0.4899263  0.04283509 -1.23686318
[11,]    4.151978  -0.3805723 0.69030113 0.99924805  0.9977442  0.10216091  0.0256465 -0.03868229 -0.08912512

CodePudding user response：

The answer is simply that a dynlm object does not contain members called coefs, r.squareds or adj.r.squareds. Look:

library(dynlm)
#> Loading required package: zoo
#> 
#> Attaching package: 'zoo'
#> The following objects are masked from 'package:base':
#> 
#>     as.Date, as.Date.numeric
data("UKDriverDeaths", package = "datasets")
uk <- log10(UKDriverDeaths)
model <- dynlm(uk ~ L(uk, 1)   L(uk, 12))

names(model)
#>  [1] "coefficients"  "residuals"     "effects"       "rank"         
#>  [5] "fitted.values" "assign"        "qr"            "df.residual"  
#>  [9] "xlevels"       "call"          "terms"         "model"        
#> [13] "index"         "frequency"     "twostage"

However, its summary object does:

summary_mod <- summary(model)

names(summary_mod)
#>  [1] "call"          "terms"         "residuals"     "coefficients" 
#>  [5] "aliased"       "sigma"         "df"            "r.squared"    
#>  [9] "adj.r.squared" "fstatistic"    "cov.unscaled"  "frequency"

So you probably need:

ler <- function(j) {
  model <- dynlm(y ~ L(y,1)   x1   x2   x3, data= dd[(1:window) j,])
  model <- summary(model)
  c <- model$coefficients
  r2<- model$r.squared
  a.r2<- model$adj.r.squared
  e <- model$residuals
  list(c,r2,a.r2,e)
}

^{Created on 2022-03-05 by the reprex package (v2.0.1)}