I am doing differential gene expression analysis between control and disease tissue. It is a skin disorder. So the normal tissue and lesion tissue are taken from same persons in the microarray analysis. the normal tissue is considered as control and the lesion is considered as disease. My doubt is to whether to use paired or unpaired t-test. I think that since the normal and disease tissue are from the same person the samples have pairing relationship and should go for paired test. Am i right ? Or should I go for unpaired t test ?
Using a paired analysis could be beneficial. It depends on amount of inter-individual variation. If this is small, a paired analysis just eats up half of the degrees of freedom. This pays of when the inter-individual variance is large.
Paired t-test would be more suitable, as you say, the samples are not independent (they are pairs of control-disease observations).
Moreover, you could consider using Wilcoxon signed rank test, the non-parametric version of paired t-test, that doesn't assume your data is normally-distributed, as might be your case.
You should consider a linear mixed effects model if your data are real numbers and you assume a normal distribution. Furthermore, the structure of the model should include a relationship between the patients i.e. pairing, which can be added as a patient id at the random effects level.
Otherwise, a paired t-test is recommended at least, as other posters have mentioned before me.
First questions to ask!
What sort of data do I have? a) Normal distributions ( Differences between groups - Relationships between sets of data). b) Data that is NOT normally distributed ( Differences between groups - Relationships between sets of data). c) Categorical data ( Differences: 'Goodness of Fit' tests - Relationships: associations).
2. What type of question are you asking? a) Differences between groups of data b) Relationships between sets of data
Its all depends upon your data being normally or not normally distributed. if data is normally distributed use parametric test and if not then use non parametric test.
The most useful test is probably the robust rank test. This is non parametic and more powerful than the Wilcoxon signed rank test.
The data is likely to be very non normally distributed. While the t test is moderately robust to non normality, it is questionable that this data would be sufficinetly normal to use it
Another alternative would be to normalise the data before using the t test. I would suggest Haldane's transformation to do this.
Another possibility would be to use a permutation test. With current computers these are quite doable if the sample sizes are not massive. The results are exact and non parametic.
It should be possible to design a bootstrap test for the data also. I would have to know a bit more about to be able to do that.
Similarly a Bayesian test could be used if the data distribution is reasonably well known. This distribution could be determined with a genetic curve fitting programme.
There may be a few other methods.
Although a linear mixed effects model has been suggested this is a little complex for what seems to be a fairly simple problem.
There are some other matters to consider.
The t test compares MEANS only. Other paperamters shows as the variance, kurtosis and skewness may also differ and these cannot be tested with a t test.
A Kolmogorov Smirnov test might also be useful here
paried t-test works well if you are looking few number of genes. but if you are working in large number of genes, data is multivariate in distribution, there are sophisticated techniques out there like random forest, support vector machine and neural network.
Hotelling test is the usual multivariate equivalent of a t test.
Random forest is more usually used to cluster data into groups. It may not provide an easy answer to the question posed.
Neural nets are very powerful but the results can also be very difficult to decipher.
I'm not sure if a vector support machine would be suitable for this question. I have very little experince with this last technique so I could be wrong on this point.
due to the high cost of gene expression chips, it is often with limited number of samples to compare across groups. That's why bootstrap is usually used for gene expression comparison.
- Badges
- Science topic
- Similar topics
- Biological Science
- Molecular Cell Biology
- Culture Cells
- Cell Line