Re: How to compute p-values associated with regression coefficients from multiple linear regression?

From: Dennis Shea <shea_at_nyahnyahspammersnyahnyah>
Date: Mon Jan 13 2014 - 09:52:36 MST

re: "compute the p-value that is associated
      to each regression coefficient"

I was hoping someone else would answer this. I am not a statistician.
**Hopefully**, someone more knowledgeable than I can help here.

The p-values for the partial regression coefficients ('b') can
be calculated via:

        t(m) = b(m)/stderr(b(m))

where 't(m)' is the t-value and 'stderr(b(m))' is the standard
error of the 'b'.

eg: from the R output below for X2: t=0.022573/0.008168=> 2.764

Unfortunately, I am not sure how to calculate stderr(b(m))
Again, perhaps someone could modify the attached NCL script.

================================================================
[1] If you have 'R' available, you could use the following approach.
     If you have used NCL to do the 'heavy lifting' (ie pre-process
     the data), create an ascii (or binary, or netCDF) file which can be
     read by R. Here it is a sample ascii file.

     The sample data set [KENTUCKY.TXT] is the same as is used
      in Example 3 at
http://www.ncl.ucar.edu/Document/Functions/Built-in/reg_multlin.shtml

     Start R, then enter the following

     df = read.table("KENTUCKY.TXT", header=TRUE) # ascii => data frame
     df # print 'df'

     mlr <- lm(Y~., data=df) # linear model
     mlr # output results ... this contains the info you want
         # see 'Coefficients' and significance

+++++++> output from R's 'mlr' function <++++++++++++
Call:
lm(formula = Y ~ ., data = df)

Coefficients:
(Intercept) X1 X2 X3 X4
   X5
   -2.244460 0.005091 0.022573 -0.232804 0.062611
-0.002038
          X6
   -0.116584

> summary(mlr)

Call:
lm(formula = Y ~ ., data = df)

Residuals:
     Min 1Q Median 3Q Max
-5.7159 -2.8806 -0.7836 1.2759 22.1562

Coefficients:
              Estimate Std. Error t value Pr(>|t|)
(Intercept) -2.244460 11.270778 -0.199 0.84309
X1 0.005091 0.010959 0.465 0.64458
X2 0.022573 0.008168 2.764 0.00838 **
X3 -0.232804 0.175949 -1.323 0.19278
X4 0.062611 0.020587 3.041 0.00400 **
X5 -0.002038 0.002186 -0.932 0.35656
X6 -0.116584 0.072610 -1.606 0.11568

---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 5.136 on 43 degrees of freedom
Multiple R-squared:  0.6136,	Adjusted R-squared:  0.5596
F-statistic: 11.38 on 6 and 43 DF,  p-value: 1.394e-07
==========================================================
[2] As noted, I don't know how to compute the standard error of
     the coefficients which is necessary to compute the t-values.
     Hopefully, someone more knowledgeable than I can help!!!!!
     Attached is an NCL script (reg_multlin.ncl_talk.ncl)
     with a local driver function that can be reused.
     The script returns return the (overall) multiple regression
     coef (r) and the F-statistic with dof=(M,N-M-1). You could
     look up the p-value.
     See the output: ncl_talk.reg_multlin (attached)
     The NCL output matches the R output except for the se(b(m))
     related quantities.
Good luck
On 1/8/14, 2:26 PM, Strada, Susanna wrote:
> Hi,
>
> I performed a multiple linear regression analysis using NCL function "reg_multlin" to obtain a 2D (lat,lon) map of standardized regression coefficients (6 independent variables, 30 observations).
>
>       N = dimsizes(y)
>       M = N_vars
>       X = new( (/M+1,N/), "float" )
>       X(0,:) = 1.0
>       X(1,:) = x1
>       X(2,:) = x2
>       X(3,:) = x3
>       X(4,:) = x4
>       X(5,:) = x5
>       X(6,:) = x6
>       beta = reg_multlin(y, X, False)
>
> Now, I would like to compute the p-value that is associated to each regression coefficient.
> I'm thinking about, firstly, computing the t-value by doing the ratio between a regression coefficient and its standard deviation, afterwards I will use the NCL function "betainc" to get the p-value for a Student t-test.
>
> My problem is that I don't know how to obtain the standard deviation of each regression coefficient using NCL. I tried to move the "reg_multlin" function in a loop over the number of observations:
>
>       do nf = 0,nfils-1,1
> ->      b_nf(:,nf) = reg_multlin(y, X(:,nf), False)
>       end do
>
>       fAODbeta(ns,ilat,ilon) = beta(1) + 0.
>       fAODbeta_std(ns,ilat,ilon) = stddev(b_nf(1,:)) + 0.
>
> but I get the following error for the line indicated by an arrow (->)
>
> fatal:Number of dimensions in parameter (1) of (reg_multlin) is (1), (2) dimensions were expected
>
> Could you suggest me how to solve my problem?
> Or do you know a better way to estimate the p-value associated with each regression coefficients?
>
> Many thanks!
>
> Best regards,
> Susanna
> _______________________________________________
> ncl-talk mailing list
> List instructions, subscriber options, unsubscribe:
> http://mailman.ucar.edu/mailman/listinfo/ncl-talk
>





_______________________________________________
ncl-talk mailing list
List instructions, subscriber options, unsubscribe:
http://mailman.ucar.edu/mailman/listinfo/ncl-talk

scatter_kentucky.png
Received on Mon Jan 13 09:52:43 2014

This archive was generated by hypermail 2.1.8 : Sun Jan 19 2014 - 21:56:35 MST