Any selection based solely on fold change is arbitrary and there is no right nor wrong threshold. You pick the threshold you feel is best, be it 1.5 fold, 2 fold or whatever.

Ordinarily, it would be optimal to simultaneously filter differential expressed genes using both a statistical significance threshold (e.g. an FDR < 0.05) and a Fold Change threshold (e.g. 1.5 fold).

Usually it is 1.5 fold change. but again it also depends upon your disease.

Sorry to say I am not interested in disease but gene expression overall. What will be the reference in case one decide on the basis of fold change

Any selection based solely on fold change is arbitrary and there is no right nor wrong threshold. You pick the threshold you feel is best, be it 1.5 fold, 2 fold or whatever.

Ordinarily, it would be optimal to simultaneously filter differential expressed genes using both a statistical significance threshold (e.g. an FDR < 0.05) and a Fold Change threshold (e.g. 1.5 fold).

I'm not the king of mcroarray, but I'm more interested in genes that are consistently differentally expressed. Unlike in prokaryotes, huge fold changes are not necessary to produce large changes in phenotype, especially when those changes affect long-term architectural changes such as in aging.

If you have multiple arrays for the comparison you can calculate a p-value for the change and de an oversampling correction (the FDR mentioned by Michael. Simplest, but often loosing much real information is Bonferoni). You can combine the p-value for the change with a minimal fold change since you will want things that are both significant and biologically meaningful. But... There is also your biological question. You might not be really interested in individual changing genes but more in the processes affected. In that case you will probably want to use rather arbitrary selection criteria for p-value, overrepresentation correction and fold change and do an gene ontology analysis (eg using GO-Elite ) or a pathway analysis (eg using PathVisio) or a GSEA (which you could do in Bioconductor, or also in PathVisio). (Disclaimer I am involved in the development of both GO-Elite and PathVisio, you might want to want different tools).

URLs:

GO-Elite (GO Analysis): http://www.genmapp.org/go_elite/

paper: http://dx.doi.org/10.1093/bioinformatics/bts366

PathVisio: http://www.pathvisio.org

paper: http://dx.doi.org/10.1186/1471-2105-9-399

BioConductor (GSEA): http://www.bioconductor.org

Assuming your arrays are Affymetrix you might first want to run our array quality control and normalisation pipeline on http://www.arrayanalysis.org

Anything equal or greater than 2 folds at p value 0.05, you can treat it as over expressed.

I have done many microarray analysis. The cut off whether 1.5 fold , 2 fold is arbitrary.

I decided on giving more weight to the p value, depending on the number of samples and genes, usually a P value of 0.001 or less. I also used known genes as a measure, e.g. in looking at the intreferon effect I would take into account genes known to be induced by this drug. I also would also look at variation in housekeeping genes.

you can go with out side fold change cut off >=2.0 by taking significant p value 0.05 and FDR test (multiple correction of p-value). after this analysis you can select the highly up/down regulated genes. In next step you can map the identified highly up/down gene list in DAVID tool for the pathway and functional annotation of genes.

It depends. I have taken two-fold change. You can refer my paper http://article.sapub.org/10.5923.j.ajbe.20120205.03.html

I would say, the most stringent cutoff is 2 fold change, p

If you're interested in the relative fold-change among the three species, I would recommend that you run the observed ratios through a Bayesian Analysis of Gene Expression Levels (BAGEL), sort by P < 0.05, then within those significant differences by the maximum difference between any pair. That will give you a statistically significant candidate list ordered by the size of difference.

BAGEL is at

http://www.yale.edu/townsend/software.html

Townsend, J.P., and D.L. Hartl. 2002. Bayesian analysis of gene expression levels: statistical quantification of relative mRNA level across multiple strains or treatments. Genome Biology 3 (12): research0071.1-0071.16.

Townsend, J.P. 2004. Resolution of large and small differences in gene expression using models for the Bayesian analysis of gene expression levels and spotted DNA microarrays. BMC Bioinformatics 5: 54.

Fold change is arbitrary and needs some statics to back it up. So combining corrected p values and FC is the right way to go about it. People tend to use 1.5 at first and increase the stringency to 2FC.