fixed effect with ancova

[ro]

Hi all,

I need calculate fixed effect with ancova analysis (spss or genstat).

My experiment was on fishes weight with differents diets (8): 15 fishes in each tank (each diet in triplicated, 3 tanks for each diet, total tanks:24). I need calculate p value for fixed effect: tanks and diets wit ancova analysis. the correct or desired model is

weight= Intercept + Diet_Effect + tank_effect+ epsilon

when i do this analysis with just one effect (tank or diet separatedly) i found significants differences (this is the expected result), but when i do the analysis with both effect in the same ancova analysis (the correct method) i dont found significant diferences in diet...

I did read that the correct methods for this kind of problem coul be with blocks or nested method or just interaction tanks*diet, i hope capture variance produced by tanks

my spredsheet looks like

diet     tank     weight
1        1
1        2
1        3
2       4
2       5
2       6
3       7
3       8
3       9
...
Any recomendation is valuable.

[Maguin, Eugene]

Ro,

I don't think you have an ancova design. Why do you think you do?
I disagree with your model statement. Your design statement implies that you
have the weights of the individual fish; so, 15*8*3=360 data points. You
don't. You have the weights of a tank of fish; so, 8*3=24 data points. I say
you have a one factor anaova in which the factor is diet and you have three
data points (tank weights) per level of diet.

Gene Maguin


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of ro
Sent: Thursday, August 18, 2011 6:57 PM
To: [hidden email]
Subject: fixed effect with ancova

Hi all,

I need calculate fixed effect with ancova analysis (spss or genstat).

My experiment was on fishes weight with differents diets (8): 15 fishes in
each tank (each diet in triplicated, 3 tanks for each diet, total tanks:24).
I need calculate p value for fixed effect: tanks and diets wit ancova
analysis. the correct or desired model is

weight= Intercept + Diet_Effect + tank_effect+ epsilon

when i do this analysis with just one effect (tank or diet separatedly) i
found significants differences (this is the expected result), but when i do
the analysis with both effect in the same ancova analysis (the correct
method) i dont found significant diferences in diet...

I did read that the correct methods for this kind of problem coul be with
blocks or nested method or just interaction tanks*diet, i hope capture
variance produced by tanks

my spredsheet looks like

diet     tank     weight
1        1
1        2
1        3
2       4
2       5
2       6
3       7
3       8
3       9
...
Any recomendation is valuable.

--
View this message in context:
http://spssx-discussion.1045642.n5.nabble.com/fixed-effect-with-ancova-tp471
3723p4713723.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[ro]

Hi,
Yep, i have the weight of each fish

[ro]

Sorry if for dont give the inforamtion exactly.... my real datasheet looks like:

animal     diet   tank weight
1             1       1
2             1       1
...15        1       1
1             1       2
2             1       2
...15        1       2
1             1       3
2             1       3
...15        1       3
1             2       1
2             2       1
...15        2       1
1             2       2
2             2       2
...15        2       2
1             2       3
2             2       3
...15        2       3
...until diet 8 ...

[Maguin, Eugene]

Ro,

Why ancova? I think this is your minimum syntax.

Glm weight by diet tank.

You get diet and tank main effects and diet-tank interaction.

OR, is this a pre-post design? How many weight measurements do you have for
each fish?

Gene Maguin

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of ro
Sent: Friday, August 19, 2011 12:22 PM
To: [hidden email]
Subject: Re: fixed effect with ancova

Hi,
Yep, i have the weight of each fish

--
View this message in context:
http://spssx-discussion.1045642.n5.nabble.com/fixed-effect-with-ancova-tp471
3723p4716197.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[ro]

Hi Gene,

I have to adjust a lineal model for genetic variance, i have to adjust one model for this purpose, maybe you are thinking in the covariate (exist other variable in my experiment that i did measure: lenght, before this analysis i did run one correlation analysis and have a very good correlation with weight...)
then for my genetic lineal model i have to found fixed effect and covariate to input to the model, for this reason ancova. i have measured for all fishes (360 animals) each 3 months and the end of experiments...

i know that i get diet and tank main effects and diet-tank interaction. but i have problems when i run this model, when i run ancova whit only diet or tank effects separatedly i have p levels <0.001 for both fixed effects (expected results)

but when i run
weight= u+ diet_effect+ tank_effect+error
i get:
tank_effect: p.value: <0.01
diet_effect: p.value: 0.89
diet_effect*tank_effect: p.value: <0.01
covariate: p.value: <0.01

so why  diet_effect disapear in ancova analysis?, i will hope get both fixed effect...

[ro]

i could make a nested model tanks grouped by dieta, or this is just interaction diet*tank?? or maybe block???

[Maguin, Eugene]

Ro,

Please post the command used to run the ancova model.

Gene Maguin

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of ro
Sent: Friday, August 19, 2011 3:46 PM
To: [hidden email]
Subject: Re: fixed effect with ancova

i could make a nested model tanks grouped by dieta, or this is just
interaction diet*tank?? or maybe block???

--
View this message in context:
http://spssx-discussion.1045642.n5.nabble.com/fixed-effect-with-ancova-tp471
3723p4716827.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Rich Ulrich]

Maybe I'm missing something, but the results do make sense to me.

There is a large variation between Tanks; there is an interaction
between Tanks and Diet.  I'm not sure what you do have as error term
for D, but  T x D  looks reasonable.  That would undermine the difference
that was seen directly when looking at Diet alone. 
 
--
Rich Ulrich

> Date: Fri, 19 Aug 2011 12:42:04 -0700

> From: [hidden email]
> Subject: Re: fixed effect with ancova
> To: [hidden email]
>
> Hi Gene,
>
> I have to adjust a lineal model for genetic variance, i have to adjust one
> model for this purpose, maybe you are thinking in the covariate (exist other
> variable in my experiment that i did measure: lenght, before this analysis i
> did run one correlation analysis and have a very good correlation with
> weight...)
> then for my genetic lineal model i have to found fixed effect and covariate
> to input to the model, for this reason ancova. i have measured for all
> fishes (360 animals) each 3 months and the end of experiments...
>
> i know that i get diet and tank main effects and diet-tank interaction. but
> i have problems when i run this model, when i run ancova whit only diet or
> tank effects separatedly i have p levels <0.001 for both fixed effects
> (expected results)
>
> but when i run
> weight= u+ diet_effect+ tank_effect+error
> i get:
> tank_effect: p.value: <0.01
> diet_effect: p.value: 0.89
> diet_effect*tank_effect: p.value: <0.01
> covariate: p.value: <0.01
>
> so why diet_effect disapear in ancova analysis?, i will hope get both fixed
> effect...
>

[ro]

Gene,
here is command
"General Analysis of Variance."
BLOCK "No Blocking"
TREATMENTS tank+diet+tank.diet
COVARIATE longcms
ANOVA [PRINT=aovtable,information,effects,%cv; FACT=32; CONTRASTS=7; PCONTRASTS=7;\
 FPROB=yes] logpeso

[ro]

sorry thi was the genstat command, this is the spss command:

UNIANOVA logpeso BY dieta tanque WITH longcms
  /METHOD=SSTYPE(3)
  /INTERCEPT=INCLUDE
  /PRINT=HOMOGENEITY DESCRIPTIVE
  /CRITERIA=ALPHA(.05)
  /DESIGN=dieta tanque longcms.

[ro]

Hi Rich,
so its normal that disappear diet_effect when i run ancova with both effects diet and tanks effects?..  i should choice just diet as fixed effect?, what happen with variability from tank source??  this is reflect from Tankx diet interaction??
Thanks

[parisec]

Happy Friday,

I'm trying to grasp how one would interpret the inclusion of subject ID into a linear mixed model. It needs to be included because I need to take into account the correlation that would occurr because the same subjects repeating a task multiple times.

A papar I am reading states that  ID was included in the mixed model for this reason, but there was no estimate reported in the results table.

For a variable like year, I believe the resulting covariance parameter estimate would be interpreted as how much average task time varies within the years included in the model.

Would subject be intrepreted the same way?  I'm wordering if because it is "the given" that subjects times would vary within different subjects that it is included in the model but not interpreted.

Thanks much.
Carol

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Art Kendall]

You have a finding that the relation between diet and weight differs by tank.
IF the finding does not make sense then
do a graph with weight on the vertical, tanks on the horizontal with dots for the diet by tank means.
connect the three dots for each diet.  Your results are saying that those "curves" are not parallel.
Does it look that way visually?

It may be that one cell has a data error, maybe a slipped decimal point.  Did you proofread your data?

However, from your present description, it sound like you missed putting a repeated measure factor in your model (time with 4 or 5 levels).
Between group factors: diet (8) and tank (3). (24 cells with 15 fish in each). Within fish specimen repeated weights (4?, 5?).  You also have some continuous variables that describe the 360 fish.  Are they constant over the repeats or do they vary over the repeats? I.e., do you have 360 or 1440 measures of the covariates?


Art Kendall
Social Research Consultants

On 8/19/2011 3:42 PM, ro wrote:

Hi Gene,

I have to adjust a lineal model for genetic variance, i have to adjust one
model for this purpose, maybe you are thinking in the covariate (exist other
variable in my experiment that i did measure: lenght, before this analysis i
did run one correlation analysis and have a very good correlation with
weight...)
then for my genetic lineal model i have to found fixed effect and covariate
to input to the model, for this reason ancova. i have measured for all
fishes (360 animals) each 3 months and the end of experiments...

i know that i get diet and tank main effects and diet-tank interaction. but
i have problems when i run this model, when i run ancova whit only diet or
tank effects separatedly i have p levels <0.001 for both fixed effects
(expected results)

but when i run
weight= u+ diet_effect+ tank_effect+error
i get:
tank_effect: p.value: <0.01
diet_effect: p.value: 0.89
diet_effect*tank_effect: p.value: <0.01
covariate: p.value: <0.01

so why  diet_effect disapear in ancova analysis?, i will hope get both fixed
effect...

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/fixed-effect-with-ancova-tp4713723p4716821.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD

[Maguin, Eugene]

Ro,

Thank you for posting your syntax. Ok, now I understand why you keep
referring to a covariate. What is longcms?

I'd like to see something else. Please run this syntax and post the
resulting table to the list.

Means logpeso BY dieta by tanque/cells=count mean stddev.

I thing there is something extremely odd going on. I assume that fish were
randomly assigned to tank and diet (or, you had a big tank of fish and you
caught 360 and put them in the 24 tanks). I'd expect that the tank effect
would be not significant and near 0.0 in magnitude. Likewise, the
interaction. As I understand it, you aren't getting that. So, it's past time
to look at the summary data (cell statistics).

Gene Maguin


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of ro
Sent: Friday, August 19, 2011 5:04 PM
To: [hidden email]
Subject: Re: fixed effect with ancova

sorry thi was the genstat command, this is the spss command:

UNIANOVA logpeso BY dieta tanque WITH longcms
  /METHOD=SSTYPE(3)
  /INTERCEPT=INCLUDE
  /PRINT=HOMOGENEITY DESCRIPTIVE
  /CRITERIA=ALPHA(.05)
  /DESIGN=dieta tanque longcms.


--
View this message in context:
http://spssx-discussion.1045642.n5.nabble.com/fixed-effect-with-ancova-tp471
3723p4717017.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Susan Maree Cotton]

I am currently away from the office recovering from a knee reconstruction. I will be only checking emails sporadically.

[ro]

Hi all,

Gene, Longcms is the lenght, this was the variate that have good correlation with weight, in a GLM analysis i can add a covariate...for me this is an ancova analysis, i am right?...

my data was checked twice for 2 different people, but iam gonna review again...

Art, i plot data and lines are not parallel, interaction exist!, but i am not agree with a point (mean), i will check again.

anyway, can i add interaction to a lineal model (remember that i am try to find the best lineal model that adjust to my data), i only thought that fixed effects and covariates are main effects to my model...

Thanks Gene, Rich and Art for you support and you spend time, i will check all data (previously was checked but i am gonna do again with original data...).
any news i ll stay in touch...

[ro]

Gene,
yes, I found others good correlations but with no sense eg: fish_color (melanin content and hue (color parameters)), but they dont have sense with weight of fishes...then dont add to the model...

[E. Bernardo]

I tried to run the syntax below using version 18 to generate Forest Plot.  Unfortunately, it does not work.  A warning appeared on the screen as follows:

 

Warnings

An inline graph specification was expected but not found.

Execution of this command stops.

 

I appreciate any help.

 Eins

* Sample dataset (processed data )*.
DATA LIST LIST/ trial(F4) year(A5) study(A10) measure cilow ciup percwi(4 F8.3).
BEGIN DATA
 1 "1989" "Hodnett  " .502 .262  .962   4.940
 2 "1991" "Kennell  " .352 .216  .575   8.694
 3 "1992" "Bréart-Fr" .785 .483 1.276   8.878
 4 "1992" "Bréart-Bg" .811 .653 1.007  44.598
 5 "1997" "Gagnon   " .867 .573 1.311  12.237
 6 "1998" "Langer   " .280 .203  .384  20.654
 7 "    " "Total    " .594 .514  .687 100.000
END DATA. 

*/ Assuming you have SPSS 14:

SORT CASES BY trial(D).
STRING YearAndStudy(A30).
COMPUTE YearAndStudy=CONCAT(RTRIM(year)," ",study).
COMPUTE RefLine=1.

GGRAPH
  /GRAPHDATASET NAME="graphdataset" VARIABLES=YearAndStudy ciup cilow measure percwi RefLine
  MISSING=LISTWISE REPORTMISSING=NO
 /GRAPHSPEC SOURCE=INLINE.

* If the effect you are measuring is OR or RR, then this syntax can
 do the task *.

BEGIN GPL
 SOURCE: s=userSource(id("graphdataset"))
 DATA:   YearAndStudy=col(source(s),   name("YearAndStudy"), unit.category())
 DATA:   ciup=col(source(s),    name("ciup"))
 DATA:   cilow=col(source(s),   name("cilow"))
 DATA:   measure=col(source(s), name("measure"))
 DATA:   percwi=col(source(s),  name("percwi"))
 DATA:   RefLine=col(source(s), name("RefLine"))
 COORD:  transpose(rect(dim(1,2), transpose()))
 GUIDE:  axis(dim(2), label(" Favours treatment                 Favours Control"))
 SCALE:  cat(dim(1))
 SCALE:  log(dim(2))
 ELEMENT: interval(position(region.spread.range(YearAndStudy*(cilow+ciup))),
  shape(shape.line), color(color.black))
 ELEMENT: point(position(YearAndStudy*measure), shape(shape.square),
  size(percwi), color.interior(color.black))
 ELEMENT: line(position(YearAndStudy*RefLine), shape(shape.line))
END GPL.

[David Marso]

ro wrote

Hi all,

IMNSHO,
Using Length as a "Covariate" in this case is *WRONG*.  It is correlated with the treatments (diet).
I should think of this perhaps as a Bivariate response (use both weight and length as dependent variables).
OTOH, I don't know WTF I am talking about...Sounds like the OP is fishing!!!
HTH, David

Gene, Longcms is the lenght, this was the variate that have good correlation with weight, in a GLM analysis i can add a covariate...for me this is an ancova analysis, i am right?...

my data was checked twice for 2 different people, but iam gonna review again...

Art, i plot data and lines are not parallel, interaction exist!, but i am not agree with a point (mean), i will check again.

anyway, can i add interaction to a lineal model (remember that i am try to find the best lineal model that adjust to my data), i only thought that fixed effects and covariates are main effects to my model...

Thanks Gene, Rich and Art for you support and you spend time, i will check all data (previously was checked but i am gonna do again with original data...).
any news i ll stay in touch...