[R] difference in results from R vs SPSS

[(Ted Harding)]

On 24-Jun-06 Fiona Sammut wrote:
> Thanks for the reply.  The mistake in 1 resulted to be due to using a 
> different number of decimal places in the data.

If using a different number of decimal places results in such
a large difference, you need to seriously consider whether a
regression can give meaningful results at all with your data.

What do the SEs of the coefficients look like? In R, do

  summary(lm(y ~ x))

For example, I tried the following (in R):

  x<-0.1*(0:10)
  y<-x + rnorm(11)
  y3<-round(y,3)
  y1<-round(y,1)

so

 y3:
 0.058  0.165  1.329 -0.964 -0.896  0.888
       -1.124  0.630 -0.544  0.702  0.944

 y1:
 0.1    0.2    1.3   -1.0   -0.9    0.9
       -1.1    0.6   -0.5    0.7    0.9

and then:

summary(lm(y3~x))
  Coefficients:
              Estimate Std. Error t value Pr(>|t|)
  (Intercept) -0.07014    0.51113  -0.137    0.894
  x            0.35627    0.86396   0.412    0.690

summary(lm(y1~x))
  Coefficients:
              Estimate Std. Error t value Pr(>|t|)
  (Intercept) -0.05455    0.50562  -0.108    0.916
  x            0.32727    0.85465   0.383    0.711

so here too there are differences between y3~x and y1~x
not unlike yours (23% in intercept compared with your 29%,
8% in x-coefficient compared with your 9%) purely as a
result of rounding. But the SEs of intercept and x-coeff
are large, so there is very little information about what
their values should be: in round figures, a 95%CI for the
intercept would be -0.06 +/- 2.26*0.5 = (-1.19, 1.07), and
for the x-coeff 0.34 +/- 2.26*0.86 = (-1.60, 2.28). Compared
with this, the distinction between the regressions y3~x and
y1~x is at the limit of subtlety -- otherwise put, it wouldn't
matter a lot what the data were!

Best wishes,
Ted.

> As regards 2, will check it out again.
> 
> Thanks again.
> 
>> Fiona Sammut wrote:
>>> Hi all,
>>>
>>> 1.  I am doing some data analysis using both R and SPSS, to get used 
>>> to both software packages.  I had performed linear regression in R 
>>> and SPSS for a number of times before this last one and the resulting
>>> coefficient values always matched.  However, this last data set I was
>>> analyzing using simple linear regression and using the command 
>>> lm(y~x), gave me different readings from R and SPSS:
>>>
>>> R:                y= 33.803 + 6.897x
>>>
>>> SPSS:             y= 47.589 + 6.263x
>>>
>>> I have no doubts regarding the procedure employed in SPSS and I am 
>>> sure I am using the same dataset for both software packages.
>>> [...]


--------------------------------------------------------------------
E-Mail: (Ted Harding) <Ted.Harding at nessie.mcc.ac.uk>
Fax-to-email: +44 (0)870 094 0861
Date: 25-Jun-06                                       Time: 11:34:30
------------------------------ XFMail ------------------------------