Programming the statistical procedures from SAS

Re: Overlay historgram and distributions

Reply
Contributor
Posts: 40

Re: Overlay historgram and distributions

Thanks for the "1 minus."

I just kicked myself for not catching that.

 

I attached out SAS output if needed.

In summary, none of the 3 are fitting the data well.

Not sure what to do.

I wish the Univ. System that I work for would employ a consulting statatician, but I digress.

 

Travis

SAS Super FREQ
Posts: 3,317

Re: Overlay historgram and distributions

Sometimes (actually, often) data do not fit a "named" distribution. Not every data set is a random sample from a simple theoretical model.  When you say "I don't know what to do," it sounds like you can't proceed without a parametric model, but parametric models are just one kind of statistical analysis. If you tell us what you are trying to do with these data, that might help.

 

I am going to move this thread from the Graphics Community to the Statistical Community to give you greater access to statistical experts.  After I do that, please explain that nature of the data and what you are trying to accomplish.

 

Contributor
Posts: 40

Overlay historgram and distributions; no Continuous Distribution fits

Appologies in advance for multiple things scrunched into 1 post.

 

Need help detemerming what to do if no distribution (continuous ) fits the data (attached).

Most of the variables are not normal. Thus, running PROC Univariate to see what structures best fit my continuous data before I run GLIMMIX. Lots of variables, but for now, I'll focus on suppDMIkg.

 

Study design (lamb feeding trial): Effects of using 2 different feed ingredients (juniper and urea) in supplements fed to ewe lambs on the following dependent variables: intake (supplement, hay, and total), growth (BW gain, efficiency, etc.), and blood serum (e.g., glucose).

  • Animal = exp. unit; each lamb has own unique ID
  • All lambs fed hay. Each lamb also fed respective treatment: 1 of 8 different feeds in a 4×2 factorial: 4 juniper levels (15, 30, 45, or 60%) and 2 urea levels (1 or 3%).Feed Intake and Growth evaluated on d 0, 5, 12, 19, 26, 33, and 40.
  • Focusing right now on supplement intake (SAS name = suppDMIkg)

 

Objectives:

  1. Does increasing level of juniper in the supplement result in linear, quadratic, or cubic trends in intake, growth, or blood serum componenets?
  2. Does the response (increasing juniper) change over days on trial?
  3. Does the response (increasing juniper) change due to level of urea?

 

I struggled with the contrast statements for suppDMIkg, but I think I got it figured out. When I used "e" it gave me coef. that didn't work (didn't equal 0), thus I calculated them myself.

 

QUESTION: Is the following method to construct the coeff. correct?

QUESTION: How is a 3-way interaction handled? e.g., JUNIPER x UREA x day

I ran PROC IML for juniper and then for day. I then cross multiplied and got my coefficients (final result = 0) for linear, quad, and cubic.

CONTRAST 'LINEAR JUNIPER'                JUN       -3  -1   1  3/e;

CONTRAST 'QUADRADIC JUNIPER'             JUN        1  -1  -1  1/e;

CONTRAST 'CUBIC JUNIPER'                 JUN       -1   3  -3  1/e;

CONTRAST 'LINEAR JUN*day' JUN*day

1.792842      1.075707      0.358569      -0.3585687    -1.0757058    -1.7928429

0.597614      0.358569      0.119523      -0.1195229    -0.3585686    -0.5976143

-0.597614     -0.358569     -0.119523     0.1195229     0.3585686     0.5976143

-1.792842     -1.075707     -0.358569     0.3585687     1.0757058     1.7928429/e;

 

CONTRAST 'QUADRADIC JUN*day' JUN*day 

0.5455447     -0.109109     -0.436436     -0.436436     -0.109109     0.5455447

-0.5455447    0.109109      0.436436      0.436436      0.109109      -0.5455447

-0.5455447    0.109109      0.436436      0.436436      0.109109      -0.5455447

0.5455447     -0.109109     -0.436436     -0.436436     -0.109109     0.5455447/e;

 

CONTRAST 'CUBIC JUN*day' JUN*day

0.372678      -0.5217492    -0.2981424    0.298142      0.521749      -0.372678

-1.118034     1.5652476     0.8944272     -0.894426     -1.565247     1.118034

1.118034      -1.5652476    -0.8944272    0.894426      1.565247      -1.118034

-0.372678     0.5217492     0.2981424     -0.298142     -0.521749     0.372678/e;

 

 

  DAY            
linear 5 12 19 26 33 40  
juniper -0.597614 -0.358569 -0.119523 0.1195229 0.3585686 0.5976143  
-3 1.792842 1.075707 0.358569 -0.3585687 -1.0757058 -1.7928429 -1.7292429
-1 0.597614 0.358569 0.119523 -0.1195229 -0.3585686 -0.5976143 -0.5764143
1 -0.597614 -0.358569 -0.119523 0.1195229 0.3585686 0.5976143 0.5764143
3 -1.792842 -1.075707 -0.358569 0.3585687 1.0757058 1.7928429 1.7292429
  0 0 0 0 0 0 0.00000000

 

 

I'm ultimately trying to run PROC GLIMMIX, using the correct distribution.

QUESTION: If an interaction is not significant in the model (e.g., JUNxDAY), but contrast statement shows a linear JUNxDAY interaction, can one still discuss that linear interaction or is it protected by the model P-value (must be < 0.05)?

 

DATA LAMB; SET grow;
oneMinusSuppDMIkg = 1-suppDMIkg;
PROC SORT; BY DAY ID JUN UREA;  RUN;

PROC GLIMMIX;
CLASS ID JUN UREA DAY;
MODEL oneMinusSuppDMIkg = DAY JUN UREA DAY*JUN DAY*UREA JUN*UREA/dist=LOGNORMAL ddfm=kr solution;

*/original model, the 3-way interaction was not sign., thus I dropped it/*
Random day /residual subject = ID type =UN;  */what is the difference between this and "Random _residual_/subject ID(DAY) ..."/*
CONTRAST 'LINEAR JUNIPER'            JUN       -3  -1   1  3/e;
CONTRAST 'QUADRADIC JUNIPER'     JUN        1  -1  -1  1/e;
CONTRAST 'CUBIC JUNIPER'              JUN       -1   3  -3  1/e;
CONTRAST 'LINEAR JUN*day' JUN*day
1.792842    1.075707    0.358569    -0.3585687    -1.0757058    -1.7928429
0.597614    0.358569    0.119523    -0.1195229    -0.3585686    -0.5976143
-0.597614    -0.358569    -0.119523    0.1195229    0.3585686    0.5976143
-1.792842    -1.075707    -0.358569    0.3585687    1.0757058    1.7928429/e;

CONTRAST 'QUADRADIC JUN*day' JUN*day  
0.5455447    -0.109109    -0.436436    -0.436436    -0.109109    0.5455447
-0.5455447    0.109109    0.436436    0.436436    0.109109    -0.5455447
-0.5455447    0.109109    0.436436    0.436436    0.109109    -0.5455447
0.5455447    -0.109109    -0.436436    -0.436436    -0.109109    0.5455447/e;

CONTRAST 'CUBIC JUN*day' JUN*day
0.372678    -0.5217492    -0.2981424    0.298142    0.521749    -0.372678
-1.118034    1.5652476    0.8944272    -0.894426    -1.565247    1.118034
1.118034    -1.5652476    -0.8944272    0.894426    1.565247    -1.118034
-0.372678    0.5217492    0.2981424    -0.298142    -0.521749    0.372678/e;

LSMEANS /DIFF ADJUST=simulate (REPORT SEED = 121211) CL LINES ILINK;
LSmeans JUN UREA DAY/DIFF adjust=SIMULATE  (REPORT SEED = 121211) CL LINES adjdfe=row ILINK;
RUN;  Quit;

Contributor
Posts: 40

Re: Overlay historgram and distributions; no Continuous Distribution fits

Also, when I use "oneMinusSuppDMIkg = 1-suppDMIkg" to make data skew differently, then use iLink in GLIMMIX LSmeans statement, the values are negative.

I tried iLink+1, but didn't work.

Any help would be apprecited.

Ask a Question
Discussion stats
  • 3 replies
  • 230 views
  • 0 likes
  • 2 in conversation