Date: Thu, 28 May 2009 05:43:53 +0530
Reply-To: Bhupinder Singh <bhupi80singh@YAHOO.CO.IN>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Bhupinder Singh <bhupi80singh@YAHOO.CO.IN>
Subject: Split-Split-plot mixed model
Content-Type: text/plain; charset=iso-8859-1
I have ran the Split-Split-plot mixed model experiment and I am not sure if it is correct or not so need the help of experienced person who can just make a look and tell me if it is right or wrong. I am little confused over the error terms generated by the model.
InputYear location env Plot_no Block Tillplc$ Prate$ Krate$ Yield;
procmixeddata= yield method= Type3;
classenv Block Tillplc Prate Krate ;
modelyield = Tillplc|Prate|Krate/ddfm=kr;
Help required on repeated measures:-
I have data on soil properties from four consective depths at two different times of the year. Do I need to check UN, CS, AR(1), AR(1)+RE, and TOEP structures and see which one give lower values of AIC, AICC and BIC. As this is a complex experiment so I don't think it would be feasible to visualize patterns of correlation between observations at different times.
Similarly, I have data of tissue analysis at various times of the year and I want to analyze it for repeated measures.
I have never done the repeated measures but have studied in the course so little confused to put my knowledge in to practical experience.
Thanks in advance
From: Susan Durham <sdurham@BIOLOGY.USU.EDU>
Sent: Wednesday, May 13, 2009 3:31:25 PM
Subject: Re: Split-Split-plot mixed model (problem to write random terms)
On Wed, 13 May 2009 11:47:00 -0500, Bhupinder Farmaha
>From: Susan Durham [mailto:sdurham@BIOLOGY.USU.EDU]
>Sent: Wednesday, May 13, 2009 11:08 AM
>To: SAS-L@LISTSERV.UGA.EDU; Bhupinder Farmaha
>Cc: Susan Durham
>Subject: Re: Split-Split-plot mixed model (problem to write random terms)
>On Mon, 11 May 2009 01:39:59 -0500, Bhupinder Farmaha
>>Hi Susan and SAS group
>>I am hereby writing the full model I have written for my study. In this
>>model Y=yield E=env, B=block, T=tillage, P=phosphorus, K=potassium, and
>>Err=error term. Env and blocks are random terms while all other three
>>factors are fixed.
>>Ylmijk = ï¿½ + El + B(l)m + Ti + ETli + Err1[BT(l)mi] +/ Pj + EPlj + TPij +
>>ETPlij + Err2[BP(l)mj + BTP(l)mij]
>>288 1 1 4 2 2 8 3
>>3 6 6 36
>>+/ Kk + EKlk + TKik + ETKljk + PKjk + EPKljk + TPKijk + ETPKlijk +
>>Err3[BK(l)mk + BTK(l)mik + BPK(l)mjk + BPTK(l)mijk
>> 3 3 6 6 9 9 18
>>I think this model should use following terms to test the effects and their
>>El, Ti & ETli ï¿½ will be tested by Err1
>>Pj , EPlj , TPij & ETPlij will be tested by Err2
>>Kk , EKlk , TKik , ETKljk , PKjk , EPKljk , TPKijk & ETPKlijk will be tested
>>SAS coding for this is:-
>>Options ps=50 ls=74 pageno=1;
>>Infile "F:\Sp_project.csv" delimiter="," firstobs=2;
>>Input Year location env Plot_no Block Tillplc$ Prate$ Krate$ Yield;
>>proc mixed data = soybean_yield method = Type3;
>>class env Block Tillplc Prate Krate ;
>>model yield = Tillplc|Prate|Krate/ddfm=kr;
>>random env block(env)
>>lsmeans tillplc/adjust =tukey;
>>lsmeans prate/adjust =tukey;
>>lsmeans krate/adjust =tukey;
>>Could you guys please help me to find errors in this model and validate
>>through SAS that correct errors terms are used to test the main effects and
>>From: Susan Durham [mailto:sdurham@BIOLOGY.USU.EDU]
>>Sent: Sunday, May 10, 2009 8:41 PM
>>To: SAS-L@LISTSERV.UGA.EDU; Bhupinder Farmaha
>>Cc: Susan Durham
>>Subject: Re: Split-Split-plot mixed model (problem to write random terms)
>>On Fri, 8 May 2009 22:26:34 -0500, Bhupinder Farmaha
>>>I am having hard time to find error in my SAS mixed model. Here is my
>>>problem. I am
>>>looking the effect of three variables on soybean yield.
>>>I have RCBD with split-split-plot arrangement. I have two different
>>>environments (random) and blocks (random) are nested within the
>>>Main plots are tillage (3 levels), first split is Phosphorus (4 levels) and
>>>second split is Potassium (4 levels). Tillage, Phosphorus and potassium are
>>>I have written the following model but unable to find the error.
>>>Options ps=50 ls=74 pageno=1;
>>>Infile "F:\Sp_project.csv" delimiter="," firstobs=2;
>>>Input Year location env Plot_no Block Tillplc$ Prate$ Krate$ Yield;
>>>proc mixed data = soybean_yield method = type3;
>>>class env Block env Tillplc Prate Krate Yield;
>>>model yield = Tillplc|Prate|Krate/ddfm=kr;
>>>random env block(env)
>>>env*tillplc tillplc*block(env) env*Prate env*tillplc*Prate
>>>env*Krate env*tillplc*Krate env*Prate*Krate env*tillplc*Prate*Krate ;
>>>Any help will be much appreciated.
>>The most obvious problem is that your response variable YIELD is listed in
>>the CLASS statement. The response variable cannot be a classification
>>factor in the MIXED procedure.
>>As a second issue, the choice of terms in the RANDOM statement is puzzling.
>> If a study is a "true" split-plot where treatments are randomly assigned to
>>experimental units, random terms within a hierarchical level (i.e., whole
>>plot, subplot, sub-subplot) of the design are often (but not always) pooled.
>> You have an odd combination of terms; you include some that I might pool
>>into other terms and omit some that I probably would include.
>>If you'd like feedback from the SAS-L group about the RANDOM statement,
>>you'll want to provide more detailed information about your study, including
>>why you're specifying your RANDOM statement this way.
>>Hope this helps,
>>Utah State University
>In my experience, having both ENV and BLOCK(ENV) as random-effects factors
>in a study like yours generates statistical tests that can be, at first
>look, non-intuitive. Hopefully, we can resolve any apparent confusion by
>thinking about the scope of inference for the study.
>I'm assuming that this is an agricultural field study, done in two different
>locations (i.e., ENVs). Let me know if this is not the case. Regardless,
>the comments below are pertinent.
>If you were to analyze the data for each ENV level separately, BLOCK would
>serve as the replicating factor, and error terms (and denominator degrees of
>freedom) would be defined by interactions of the fixed-effects factors and
>BLOCK. If you wanted more independent information about the experimental
>treatments, you would add another BLOCK (i.e., you would add another
>replicate). In this scenario, BLOCKs are assumed to be a random sample, and
>as such, BLOCKs delineate the scope of inference--the statistical population
>to which inference is made--for the analysis. For your study, the scope of
>inference would be limited to the particular environment in which the blocks
>Now say that you expand the scope of inference and that you consider the
>levels of ENV to be a random sample from a larger statistical population.
>Notice that you have changed the scale and nature of your scope of
>inference. ENV is now the replicating factor: if you wanted more
>independent information about this larger statistical population, you would
>add another ENV. You would not get more information by adding another
>BLOCK, because additional BLOCKs only give more information about the same
>ENV; in other words, BLOCKs are subsamples, not replicates. In this
>scenario, error terms (and denominator degrees of freedom) would be defined
>by interactions of the fixed-effects factors and ENV; with only two levels
>of ENV (i.e., two replicates), the denominator df for fixed-effects tests
>will be small.
>Your MIXED procedure code is pooling terms in accordance with your
>mathematical model. But I think you are finding that tests of fixed effects
>are not using the error terms that you had intended. Rather than reflecting
>BLOCKs as replicates, the error terms are using ENVs as replicates. You
>could verify this by looking at the Expected Mean Squares generated by the
>METHOD=TYPE3 option that you've specified, if you haven't already. So your
>code is doing "a" right thing, but it might not be "the" right thing, and
>it's probably not what you had intended it to do.
>What is "the" right thing? Without knowing more about the details of your
>study design (especially, what ENV represents, spatial layout, etc.), I
>can't offer a definite opinion. But here's a short list of possible
>1. Use ENV as a random-effects replicating factor, with BLOCK as a
>random-effects subsample factor. With only two levels of ENV, this approach
>could limited, both by the extent to which two environments represent the
>statistical population of environments and by the imprecision of variance
>2. Use ENV as a fixed-effects treatment factor, with BLOCK as a
>random-effects replicating factor. This approach restricts statistical
>inference to just the two environments used in the study.
>3. Do separate analyses for each level of ENV with BLOCK as a random-effects
>replicating factor. This approach does not include statistical comparisons
>The following two references dive a bit deeper into multi-location analysis:
>Littell, RC; Stroup, WW; and Freund, RJ. 2002. SAS for Linear Models, 4th
>ed. SAS Press. See Section 11.8.
>Littell, RC; Milliken, GA; Stroup, WW; Wolfinger, RD; and Schabenberger, O.
> 2006. SAS for Mixed Models, 2nd ed. See p8 and Sections 6.6 and 6.7.
>Hope this helps!
>Thanks for spending time and made through look at my experiment. Here, I am
providing more information that may help you to guide me better. This is an
agricultural experiment and response and rate variables are same as I have
specified here in the heading. There are two adjacent fields in the study.
First year, experiment was conducted on field I and second year conducted on
field II. The reason being for different fields is that corn and soybean
were rotated. This study is of soybean only. I am dong with two years of
experiment. In the third year, soybean will be planted back on field I. The
treatments will be applied in the same fashion as was designed originally no
matter what is the crop and year. I have a made env as random variable and
it is product of year*location. Blocks, year and env are all random
variables. I understand that env * main effects will be my error terms for
testing. But looking at the results of this simulation, it made me feel that
there might be something wrong that I am unable to catch. I was thinking
that as we do pooling of error terms, the same way I might have to pool the
random terms. These are few confusions that are bothering me. I have already
taken the help of both references you have mentioned but still I have the
above doubts in my mind.
What specifically do you see in the results that causes you to be concerned?