BookmarkSubscribeRSS Feed
ddesire
Calcite | Level 5

Dear SAS Community members.

I ran a two-way factorial analysis involving Temperature (T) and Cultivar C). 

Below are the p-values

T           0.4308

C          0.0355

T x C  0.0199

So, as you can see, there is a significant T x C interaction. However, when I open the interaction to look at the simple effects of cultivar at each T levels, or of the temperature at each cultivar level,  there was no significant difference.

My question is: what should I report, the main effects or the non-significant simple effects?

Thank you in advance for your help.

Best regards,

11 REPLIES 11
Rick_SAS
SAS Super FREQ

Even among experts, there is some disagreement on this question. Some statisticians advocate that you should go back and run the model that has only main effects. If both main effects are significant, you should include that information in the discussion. 

 

Here's a blog post that I found useful and well-written.  It suggests creating an interaction plot. In PROC GLM, use the STORE statement to save the model to an item store. Then use PROC PLM  and the EFFECTPLOT statement create the interaction plot. For an example, see the article "Use the EFFECTPLOT statement to visualize regression models in SAS."

PaigeMiller
Diamond | Level 26

However, when I open the interaction to look at the simple effects of cultivar at each T levels, or of the temperature at each cultivar level,  there was no significant difference.

Maybe I misunderstand your words (or maybe not), but this doesn't seem to me to be the proper way to evaluate an interaction. Perhaps you could actually show us what you did that you described above.

 

The interaction plots (mentioned by @Rick_SAS) are a much better way to see the interaction (and also a much better way to explain the interaction to people). 

 

My question is: what should I report, the main effects or the non-significant simple effects?

 

I believe you report them all.

--
Paige Miller
sld
Rhodochrosite | Level 12 sld
Rhodochrosite | Level 12

If the overall test of interaction is significant, but the simple effects (as obtained by, say, the SLICE option on the LSMEANS statement) are not, then I typically take this as an indication of appreciable uncertainty about the existence of interaction in this particular set of data--in other words, I may not be convinced that interaction exists, or I might think interaction probably exists but I don't have much as much evidence as I would like for that conclusion.

 

By removing the interaction term from the model, I am explicitly assuming that there is no interaction. Maybe I'm comfortable with that assumption, or maybe I'm not. It depends on the scenario.

 

 

 

ddesire
Calcite | Level 5

Thank to you all for the suggestions.

As Rick suggested, I produced the interaction graphs which seem to suggest some interactions but the means comparison don't show it.

What would be the rationale to remove the interaction term from the model?

As Paige requested, here is in attach the sas code I ran. Maybe I'm doing something wrong. I'm open to suggestions.

Best regards,

sld
Rhodochrosite | Level 12 sld
Rhodochrosite | Level 12

Your RANDOM statement looks odd to me; it does not correspond to any experimental design that I would consider to be likely. I think it probably is not correct but I would need details about your experimental design to say for sure.

 

ddesire
Calcite | Level 5

This was a growth chamber experiment with temperature regime assigned to different chambers (3 different chambers). Within each chamber/temperature, there were six cultivars blocked in 3 blocks. Each temperature x cultivar was repeated two times (2 runs).

Thank you,

sld
Rhodochrosite | Level 12 sld
Rhodochrosite | Level 12

You have 4 levels of temperature and 3 growth chambers, so you clearly did not assign the same temperature level twice to the same growth chamber; if you did to some extent, then the models I propose below are not consistent with your design.

 

If we assume that there is no systematic chamber effect on the mean response and that there is no systematic temporal effect on the mean response, then we can think of your design as having 8 growth chambers, randomly assigned to 4 temperatures. (This obviously is not what you did, but might be a good-enough model if the assumptions are valid.) The 3 blocks within each chamber are not true replications, rather they are subsamples because they are not independent realizations of a particular temperature level. You can fit a model with the full set of 144 observations:

 

proc glimmix data=wue
    plots=(studentpanel boxplot(fixed student));
    class run block temp var;
    model inswue = temp var temp*var / ddfm=kr2 ;
    random  run(temp)
            block*run(temp)
            var*run(temp);
    lsmeans temp var / diff adjust=tukey lines cl;
    lsmeans temp*var / plot=meanplot(sliceby=var join cl) slice=(temp var);
    run;

or you can compute the means over the 3 blocks within each chamber for each variety and fit a model with 48 observations:

 

proc sort data=wue out=wue_srt;
    by run temp var;
proc means data=wue_srt noprint;
    by run temp var;
    var inswue;
    output out=wue_means mean= ;
    run;
proc glimmix data=wue_means
    plots=(studentpanel boxplot(fixed student));
    class run temp var;
    model inswue = temp var temp*var / ddfm=kr2 ;
    random  run(temp);
    lsmeans temp var / diff adjust=tukey lines cl;
    lsmeans temp*var / plot=meanplot(sliceby=var join cl) slice=(temp var);
    run;

This second model is a two-way factorial in a split-plot design; the first model is a form of split-plot with subsamples. Because you have the same number of blocks in each chamber, the test results are the same for both models.

 

I think these models are, more or less, valid and certainly more correct than your original model, if I correctly understand your design. There is no evidence of interaction, so (unfortunately) your original question is now moot.

 

I hope this helps.

 

 

 

ddesire
Calcite | Level 5

Hi sld,

Thank you very much for these changes. With this code, nothing is significant.

Here is one detail about the treatment structure. Given the particularity of one of the temperature regime (35/25) which is ramping unlike others that are constant and for the fact that only one growth chamber has that capability, the 35/25 was assigned to this one chamber during both runs, but with different randomization each run. Would your code still apply?

Best regards,

sld
Rhodochrosite | Level 12 sld
Rhodochrosite | Level 12

Growth chamber studies are challenging to design, largely because no one has a whole bunch of growth chambers to work with. Or sometimes even more than one.

 

I'd probably use the model I suggested and be sure to make the experimental design clear in Methods. The risk you run with this design is that if there was an effect associated with the 35/25 treatment, then it could be due to the treatment per se or it could be due to that particular growth chamber. Or both. You have no way to know for sure.

 

ddesire
Calcite | Level 5

Thank you sld,

It is very complicated indeed. I will re-run the whole analysis with your suggested code.

Best regards,

PaigeMiller
Diamond | Level 26

I didn't really want your SAS code, I wanted to see the results. Your SAS code is kind of useless without the data. The results ought to make things clear about what you did, and whether or not the interaction or main effects are real.

--
Paige Miller

sas-innovate-2024.png

Available on demand!

Missed SAS Innovate Las Vegas? Watch all the action for free! View the keynotes, general sessions and 22 breakouts on demand.

 

Register now!

What is ANOVA?

ANOVA, or Analysis Of Variance, is used to compare the averages or means of two or more populations to better understand how they differ. Watch this tutorial for more.

Find more tutorials on the SAS Users YouTube channel.

Discussion stats
  • 11 replies
  • 1709 views
  • 6 likes
  • 4 in conversation