2
$\begingroup$

Given a linear model and the following hatvalues and influence.measures, how can I say which measurements are suspicious? I mean which ones should be revisited/normalized/filtered out?

> hatvalues(...)
0.53 0.34 0.27 0.31 0.31 0.70 0.53 0.34 0.27 0.31 0.31 0.70

and

> influence.measures(...)
 dfb.1_ dfb.x   dfb.I..2 dfb.Sd.L dfb.a.S.   dffit  cov.r   cook.d   hat inf
 0.2636 -0.3194   0.4629   0.3333  -0.6360  1.2330 3.4266 0.321323 0.609   *
-0.2368  0.3619  -0.3439   0.0838   0.0613  0.4313 1.8680 0.061618 0.311    
-0.3098  0.1384  -0.0849   0.3061  -0.2986 -0.4969 1.9330 0.064004 0.336    
-0.0646  0.0814  -0.0491   0.0303  -0.1101  0.2312 2.8601 0.013303 0.318    
 0.2636 -0.3028   0.3940   0.0932   0.0693 -0.6024 1.6638 0.063608 0.311    
-0.1263  0.3613  -0.3031   0.3368  -0.1429  0.6283 1.0418 0.100366 0.266    
-0.4291  0.8386  -1.1344   0.6684  -1.2916 -2.3642 1.3039 1.001394 0.609   *
-0.0433  0.0662  -0.0380  -0.0236   0.0893  0.1949 2.9614 0.008832 0.318    
-0.6140  0.2936  -0.1603  -0.8081   0.4926 -1.1868 0.4303 0.226898 0.336    
 0.0686 -0.2240   0.2414   0.2103  -0.0942 -0.3433 1.9311 0.033001 0.266    
 0.0336 -0.0321   0.0262   0.0380  -0.0301  0.0602 3.6663 0.000834 0.439   *
 2.1140 -1.4469   1.2603  -1.8339   1.3619  2.9196 0.0869 0.896901 0.439   *

As far as I'm concern, Cook's distance greater than 1 is exactly what I'm looking for. However, I've read somewhere that 3p/n is often employed. In addition, R provides some * besides the output lines, hence I assume those measurements are influent even though the distance is lower than 1.

How can I interpret the output? Are other numbers such as DIFFT important?

Note: I used tag R since I can't figure out any suitable one and because of my low reputation I can't create new ones. If you know some better, please let me know.

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ Though the help page for influence.measures is rather scanty, you can paw through the code to find the criteria for being awarded a star. $\endgroup$
    – Scortchi ♦
    Commented May 13, 2015 at 13:20
  • $\begingroup$ @Scortchi Really? Could you post it (possibly as an answer)? All I found is The optional infl, res and sd arguments are there to encourage the use of these direct access functions, in situations where, e.g., the underlying basic influence measures (from lm.influence or the generic influence) are already available. which is (from my point of view) unclear. Thanks $\endgroup$
    – petrbel
    Commented May 13, 2015 at 16:26
  • $\begingroup$ I haven't pawed through it myself: type in influence.measures & you'll see the code. $\endgroup$
    – Scortchi ♦
    Commented May 13, 2015 at 16:29

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.