May be Odds Ratio

Hi Jacobo,

I read this post several times and I am still confused about the details of what and how you are measuring these values. This needs clarification before I can provide a reasonably coherent answer.

More specifically, what are these variables and how (and when) are they measured? It is not clear if these are the same individuals measured 3 times (start, mid and end), or if these are different people that are measured at these time points, and if they are, or not,  measured at all 3 time points. 

Additionally, it is not clear what you are trying to determine here. What is your hypothesis?

Ariel

Hi Ariel,

I din't want to give much detail to avoid confusion, but it seems I oversimplified it.

During the experiment I record several biosignals from the subject, once the experiment is over I use those signal to try to predict three parameters. For doing so I am trying different ways of processing and classification of the signals. Each method has three different steps "Start", "Mid" and "End" (one of each). For each combination I calculate how good each method predicts the parameters. So basically for each subject I have 44 different results that comes from the 44 different ways of processing such signal.

What I want to study it can be divided into several points:

1. Which is the best combination of "Start", "Mid" and "End". Probably, since the data is very variable and the number of subjects is low (lower than 20) is not going to be possible to find one that is better than the others (statistically speaking)

2. For the three categories "Start","Mid" and "End" discover which is the best one. For this one I think it is possible to do it with a one-way ANOVA as for the previous one.

3. Finally, and this is the analysis that I don't know how to do it. I want to study what is the effect of each category to the final result. For example, let's say that the best "Start" is S2, the best "Mid" is M1 and the best "End" is E4, when calculated separated (calculated in 2) . Nevertheless the best combination is S1M1E3 (calculated in 1). In other words, the interaction between the different categories.

I actually don't know if this is possible, but I think it should. Thank you for your answer, and sorry if I didn't explain me correctly.

Hi Jacobo,

I am still having a hard time understanding your data layout and what it is you are trying to do. What specifically are your data, what are the variables, how do you determine what is the start, mid, and end - or are these different variables? Does each subject have all variables or levels of the variables?

I suggest that you attach a spreadsheet with the layout of your data and a few observations. Then, if you describe what you are hoping to accomplish, it would make it easier to respond to your post.

Ariel

It seems I 'm not very good at explaining me, sorry.  As you suggest here is the file with the data.

Nevertheless I will try to explain again. I want to study which is the best method to process a signal. For processing the signal there are 3 steps: "Start", "Mid" and "End". In total I have 44 combinations of those three conditions for each subject represented by for example S1M3E4 which means Start 1, Mid 3 and End 4.

"Start" means which is the source of the signal: EEG, EMG, EEG+EMG.

"Mid" means the classifier that I am going to use: ANN, SVM.

"End" means the output that I want to classify: Position, Angles.

For each subject I can calculate all the combinations.

For each combination I calculate a score (between 0 and 1) and that score is the value that represent that combination for that subject. 

I already made an analysis grouping them by "Start", "Mid" and "End" and now want to study the interaction between them, if possible.

Thank you for your pattience.

 "During the experiment I record several biosignals from the subject, once the experiment is over I use those signal to try to predict three parameters."

If this is your research question, ANOVA is not the method to use. ANOVA is used to see if there are differences between groups. To predict outcomes based on parameters that is multiple regression territory. And if you have 44 possible combinations you need a whole lot more participants. I am adding a link to power and sample size:

http://powerandsamplesize.com/

Hi Gloria,

I don't want to use ANOVA to predict outcomes. I use different methods to predict outcomes, I want to study which method is the best one, for that I use ANOVA.

It is true that if I want to compare the 44 combinations I would need an incredible number of subjects, but that is out of the question. That is why I divided the comparasion into trhee parts, to be able to compare the differents elements that create each method. But I don't know if comparing each part separately is correct, of if I should also analyze the effects of each part with the other two.

I am sorry it's taken me a couple days to respond. I just looked at the data layout but it is still not clear to me how these data will help you get to what you want? There are a lot of variables, and you indicate that you are looking for the best combination - but you state that you have already made predictions, so I am confused as to what you are trying to accomplish? ANOVA is for determining whether there are differences between groups, but you state that you are looking for the best combination?

I wish I could help, but I just don't understand what you need. Please clarify, since it is obvious we are miscommunicating.

Ariel

Thank you Ariel for taking so much time trying to solve my problem.

I think the miscommunication may arise because we come from different field and the way we use words. It is obvious that I am not statistician anyway I will try to explain in the most correct way the question. Before that I will try to make a simple example as model of my experiment.

The example is not going to make much sense, but is to clarify. Let's imagine that I have 4 different kinds of seeds (let call them s1 s2 s3 and s4). I want to see which kind of seed is the one that grow the most, so I plant 10 (this would be the number of subjects) seeds of each in different fields. The fields are located at 4 different heights (h1, h2, h3, h4). Also at each height there are five fields so I use a different compost for each field (c1 c2 c3 c4 c5). 

After 1 year I measure the size of each plant taking into consideration what kind of seed is, the height of the field, and the compost used in that field. For each one of those combinations. There are 80 combinations of kind of seeds,height and compost, and I planted only 10 seed of each kind, so probably I am not going to be able to say which is the best (with best I mean highest) combination of seed AND height AND compost. At least I want to see if I can say which is the best seed OR height OR compost.

In my experiment seed, height and compost are "Start", "Mid" and "End".  Also we need to imagine that I have only 10 of each kind of seeds (the number of subjects)  and that I clone them for each field, so I can consider that I am using the same seeds over and over. Also in the simplify case there are 80 combinations instead of 44 because I didn't take into consideration some restrictions.

Now I will try to explain it in the most precise way using the "correct" words. There are 80 different conditions. The dependent variable is the size of the plant. For each condition I have 10 samples. As Gloriam said, it is not enough data to have a significant "best" combination (i.e. I will not be able to say the highest plant are using seed s1, at height h3 using the compost c4). Nevertheless each condition is modulated(?) by three factors (seed, height, compost). Then for each factor I would like to know which is the best value. Which is the best compost? OR Which is the best kind of seed? OR Which is the best height? Finally I would like to know if there is any interaction(?) between the three  factors.

The results data it has the data for each fubject (rows) and conditions (columns). I hope this clarify the problem. Thank you.

You don't have enough data to answer your question. Reduce de number of variables. That is the only solution. Look for factor analysis.

Thanks for that clarification, Jacobo. I agree with Gloriam that factor analysis is an appropriate approach, however you still have a problem with small sample size vs many variables. Another approach to consider is cluster analysis. This would generate (clusters) groups of observations that share similar levels of the various conditions you are interested in.