Hi dear friends,
We have a study with just 5 samples, and we wish to implement a model on them. Since our sample amount is very few, which kinds of modeling may be suitable for them.
I had made a logit linear regression in R (gml), but it had not acceptable result.
Many thanks.
I do not think that there is a statistical procedure for your situation just by looking at your sample size.
It depends on the number of independent parameters you select. Even 5 observations may be useful if this number is too small. You can use linear regression techniques.
A model is a model is a model.
It should be sensible and serve the purpose (that is: useful).
Even in the absence of any data is the model based only on theoretical considerations. The knowledge of some data can be used to modify our believes about the values of the model coefficients. With enourmous amounts of data our believes about these values will converge, almost independent of what we thought about them initially. Noteably, having a large sample does not make a bad model good.
The point is that in small samples initial believes about the coefficients based on theoretical considerations matter more and should be accounted for. This is done by a Bayesian analysis.
https://www.stat.auckland.ac.nz/~brewer/stats331.pdf
http://blog.efpsa.org/2015/08/03/bayesian-statistics-why-and-how/
https://alexanderetz.com/2015/07/25/understanding-bayes-updating-priors-via-the-likelihood/
Also worth to read:
http://plato.stanford.edu/entries/probability-interpret/
algomagic.s3.amazonaws.com/algomagic_717c_knuth_knuth_me08_final.pdf
if is mean for quantitative data use t test ,if greater that 30 samples use z test ,
Dear Ali,
A statistical test which works well with small samples is Fisher's exact test. It related to observed and expected frequencies in a 2 by 2 grid and is a special case of the two way chi-squared test.
Dear Ali,
From my experience, the Generalized Additive Modeling is a very good algorithm for modeling and it is better the glm models.
You can implement it in R through gam and mgcv packages.
Dear Ali,
Could you show please the head of your data?
You may use:
head(sp2) to show the first six lines of data.
Many thanks for your time and help
> head(sp2)
results jamiat metraj.vahede.kasb tanavo.kala sandogh
1 0 0.321055 0.117204 0.714286 0.189189
2 1 0.211710 0.270860 0.050000 0.162162
3 1 0.167895 0.162934 0.057143 0.081081
4 0 0.086468 0.053059 0.107143 0.405405
5 0 0.212873 0.395943 0.071429 0.162162
Since my samples are so few (5 or 6 samples), I was in hope to use Fisher Exact Test method for implementing the model. Do you think it be possible to do that in R? Is there a specific function similar to glm / gam for Fisher Exact Test method? Or any other methods which are suitable for very few sample data.
Best Regards
Ali Madad
Dear Ali,
I have tried the GAM algorithm for your data, but it didn't work as it requires more data rows.
Therefore, I would try the classification algorithms, such as Bayes Classifier, Support Vector Machine, Discriminant Analysis, etc:
Naive Bayes Classifier:
require(e1071)
library(e1071)
model
Dear Watheq J. Al-Mudhafar,
I checked the "Naive Bayes Classifier" as the first method. It was great. Many thanks for your lead.
In fact I am working on my daughter`s MSc thesis. Her samples were very few and linear regression was not suitable for it. And we had not logical results. But in Bayes the results became much much more acceptable.
After she implements this model to all her data, we will proceed to the second and third methods which you mentioned.
Your hints made us so happy. We pray for your success.
Best Regards
Dear Ali,
I'm so glad for the progress of your daughter's research.
There are some necessary figures that you can create in R for the NB Classifier. You can find the related codes in my paper link.
Hope you all the best,
Watheq
https://www.onepetro.org/conference-paper/OTC-26947-MS?sort=&start=0&q=October-26947&from_year=&peer_reviewed=&published_between=&fromSearchResults=true&to_year=&rows=10#
- Badges
- Science topic