1
$\begingroup$

I'm trying to regress some simple pooled data. My data has 60 observations and three columns: Weight, Height, and Sex (female=1, male=0).

If I regress thus, Weight ~ Height + Sex, my model is fairly satisfactory, but the residuals are not homoscedastic (green errors are male, blue female):

plot

I tried regressing on the log of Weight and/or Height, but that didn't do much. What should I do to make the residuals homescedastic and/or make my model more accurate? Any help would be appreciated.

Edit

Doing a generalized regression model gives the following.

Generalized least squares fit by REML
  Model: Weight ~ h + s 
  Data: P149 
       AIC      BIC    logLik
  514.2221 524.4374 -252.1111

Variance function:
 Structure: Different standard deviations per stratum
 Formula: ~1 | Sex 
 Parameter estimates:
        0         1 
1.0000000 0.6685307 

Coefficients:
                 Value Std.Error   t-value p-value
(Intercept)  27.197499  51.88129  0.524226  0.6022
h             1.852382   0.75634  2.449128  0.0174
s           -25.284478   5.53300 -4.569755  0.0000

 Correlation: 
  (Intr) h     
h -0.997       
s -0.524  0.466

Standardized residuals:
       Min         Q1        Med         Q3        Max 
-1.6655243 -0.6879858 -0.1839396  0.5628971  3.9857544 

Residual standard error: 22.13369 
Degrees of freedom: 60 total; 57 residual

With this s. residual plot:

Could someone please explain how precisely this model is different from a standard multiple regression model? Thanks.

$\endgroup$
1
  • $\begingroup$ You should really consider adding an interaction term to the model. I would expect the relationship between height and weight to be different between the sexes. So, you should at least test that. $\endgroup$ – Roland Feb 19 '14 at 7:59
1
$\begingroup$

You should use a model framework that allows you to model the different variation in males and females. Package nlme provides a generalized least squares implementation (gls) that can be used in conjunction with the varIdent function to specify the appropriate variance structure. E.g., you could try something like

library(nlme)
gls(Weight ~ Height + Sex, weights=varIdent(form = ~ 1 | Sex), data=dat)
$\endgroup$
2
  • $\begingroup$ Thanks. This is the R output. Is gls correcting the variance or just telling me the structure? Generalized least squares... Model: Weight ~ Height + Sex AIC BIC logLik 514.2221 524.4374 -252.1111 Variance function: Structure: Different standard deviations per stratum Formula: ~1 | Sex Parameter estimates: 0 1 1.0000000 0.6685307 Coefficients: Value Std.Error t-value p-value (Intercept) 27.197499 51.88129 0.524226 0.6022 Height 1.852382 0.75634 2.449128 0.0174 Sex -25.284478 5.53300 -4.569755 0.0000 $\endgroup$ – N4v Feb 18 '14 at 14:26
  • $\begingroup$ It is including the different variance in the model. The output (which is difficult to read in a comment) tells you that the residual variance of sex 1 is 0.669 times the residual variance of sex 0. The value of 0.669 is fitted. $\endgroup$ – Roland Feb 18 '14 at 17:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.