RStudio与终端R运行lm()模型结果不一致问题问询

I've hit a puzzling issue where running the exact same no-intercept linear model code on identical data produces drastically different coefficient results in RStudio vs. a terminal R session—and this only happens when using the full dataset. Testing smaller subsets (from 10 rows up to half the data) gives perfectly matching results across both environments. Here's the full breakdown:

Key Context #

• Model Code:

  a = dputdata
  lm.transcprot = lm(formula = I(a$b - 0) ~ 0 + a$a)
  

  (Side note: I(a$b - 0) is redundant here; a$b ~ 0 + a$a works exactly the same)
• Data: dput() output is too large to share; password: r
• Data Summary (identical in both environments):

  > summary(a)
        a                b
  Min.   :  1.002   Min.   :     0.13
  1st Qu.:  5.887   1st Qu.:  5320.27
  Median : 14.551   Median : 11739.61
  Mean   : 25.877   Mean   : 20524.21
  3rd Qu.: 34.424   3rd Qu.: 26430.47
  Max.   :136.997   Max.   :116315.41
  

RStudio Environment & Model Output #

Session Details:

> sessionInfo()
R version 3.6.0 (2019-04-26)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Debian GNU/Linux 9 (stretch)
Matrix products: default
BLAS/LAPACK: /usr/lib/libopenblasp-r0.2.19.so
locale:
[1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C LC_TIME=en_US.UTF-8 
[4] LC_COLLATE=en_US.UTF-8 LC_MONETARY=en_US.UTF-8 LC_MESSAGES=C 
[7] LC_PAPER=en_US.UTF-8 LC_NAME=C LC_ADDRESS=C 
[10] LC_TELEPHONE=C LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
attached base packages:
[1] stats graphics grDevices utils datasets methods base
loaded via a namespace (and not attached):
[1] compiler_3.6.0 rsconnect_0.8.13 tools_3.6.0
> find('lm')
[1] "package:stats"

Model Result:

> lm.transcprot
Call:
lm(formula = I(a$b - 0) ~ 0 + a$a)
Coefficients:
a$a 
462.5 

Terminal R Environment & Model Output #

Session Details:

> sessionInfo()
R version 3.6.1 (2019-07-05)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 18.04.2 LTS
Matrix products: default
BLAS/LAPACK: /opt/intel/compilers_and_libraries_2019.3.199/linux/mkl/lib/intel64_lin/libmkl_rt.so
locale:
[1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C 
[3] LC_TIME=en_US.UTF-8 LC_COLLATE=en_US.UTF-8 
[5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8 
[7] LC_PAPER=en_US.UTF-8 LC_NAME=C 
[9] LC_ADDRESS=C LC_TELEPHONE=C 
[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
attached base packages:
[1] stats graphics grDevices utils datasets methods base
loaded via a namespace (and not attached):
[1] compiler_3.6.1 tools_3.6.1
> find('lm')
[1] "package:stats"

Model Result:

> lm.transcprot
Call:
lm(formula = I(a$b - 0) ~ 0 + a$a)
Coefficients:
a$a 
1889 

Likely Causes & Fixes #

1. Different BLAS/LAPACK Libraries: This is the most probable root cause. RStudio uses openblasp while terminal R uses Intel MKL—these are distinct linear algebra libraries with different numerical algorithms and precision handling. For large datasets, small floating-point errors can accumulate into noticeable differences in coefficient estimates.
2. R Version Gap: R 3.6.0 vs. 3.6.1 might have minor tweaks to lm() or underlying math routines, though this is less impactful than the library mismatch.

To resolve this:

• Align Linear Algebra Libraries: Configure both R environments to use the same BLAS/LAPACK library (either MKL or openblasp). You can adjust this via system environment variables or your .Rprofile file.
• Confirm Exact Data Match: Use all.equal(a_rstudio, a_terminal) to ensure the datasets are byte-for-byte identical in both environments—even tiny floating-point discrepancies can snowball with large data.
• Test Manual Calculation: Compute the no-intercept coefficient manually in both environments with sum(a$a * a$b) / sum(a$a^2)—this will isolate whether the difference comes from the core linear algebra computation.

内容的提问来源于stack exchange，提问作者Hart Radev