I have two bacterial srains - one is wild type and the other has one gene deleted. In my experiment, I calculate some phenotype for both of these strains. I repeat this experiment 5 times. Now I want to see if there is a difference in this phenotype between the two strains, so I perform a t-test (assuming the data is normally distributed for now). My question is - should I perform a paired t-test or unpaired? I am a bit confused about it since I feel I have some rationale for doing it either ways -
Why I think its paired : in every replicate, I compare the same two strains and I repeat this 5 times
Why I think its unpaired : Its not like I am subjecting the same strain to two different conditions (which often happens in paired tests). Rather, I have two different strains being subjected to the same condition.
Any help would be appreciated!
•
written
by
c_u
2 answers
Assuming your use of the t-test is valid, you should use the unpaired version. Your question is getting at the difference between biological and technical replicates, which is different than paired vs. unpaired measurements.
A paired measurement would be taken across conditions or time (the groups being compared), so for example measuring 5 individuals at two time points and testing the distributions of time1 vs. time2. This is different than what you're doing, which is doing technical replication within each group but comparing those to independent observations in another group (5 replicates in group1 versus 5 replicates in group2).
I think you reasoned through this already in your statement "I think its unpaired."
You would, however, expect there to be less variation under your technically replicated experimental design than if you had used 5 biological replicates. You might mention this in your limitations.
Would the experiment make sense if you did one condition 5 times, and the other 7 times? If the answer is "yes", its unpaired.
If someone had at some point accidently ruined one of your samples, would you have said "okay, gotta throw out some other sample too, it's useless now"?
That's a paired experiment
Is there a right way to make a 1-1 connection between samples from different conditions, and a lot of wrong ways? Or would any pairing be arbitrary? If there is only one right way, that's paired.
Log in to answer this question.
-
Paired vs Unpaired t-test: Compositional analysis in the Human snRNA-seq data
written by joonhong kwon •
-
What are the rules on when it is acceptable to use a paired t-test for lab animals?
written by ASid •
-
Effect of different sample sizes on p-value
written by c_u
-
Comparing two ratios between two experiments.
written by Vanish007 •
-
Forum: Conceptual - Do I need to justify myself if I generate more than three replicates?
written by c_u
-
Align paired and unpaired reads simultaneously using Bowtie2?
written by xiaozhongzhiping •
-
HISAT2 : Differences in paired and unpaired rates
written by gwe3409 •
-
What's the unpaired reads in the raw-data
written by Yunlong JIA •
-
Setup for multiple strains, multiple conditions, and replicates
written by montezuma •
-
Paired end reads DNA strains relationship.
written by arronar
and
You deleted a gene and there is a phenotype that you are able to observe in all 5 replicates. In that case do you really need statistical evidence? Or is the phenotype subtle/variable so you can't say for sure that the gene deletion is causing that? If there is no visible phenotype then how are you quantifying the difference?
Thanks Genomax. The phenotype I measure is exhibited by all relevant bacterial strains, and I am trying to see if there is a significant difference in this phenotype, between the control and the deletion mutant. I'm sorry if I wasn't very clear earlier
A universal advice: use regression modelling instead of t-test =) I think your question will be answered right when you write down the necessary formula.
The answer for the simplest case will be equivalent to a t-test.
Hi, thanks for that comment. Since I comparing two samples, how do you suggest regression will help determine the difference? Because with a t-test it is quite straightforward
Simply code your groups as predictors and run a regression analysis =) but when you start to think about the design matrix - you immediately come to the answer which kind of model is the most suitable
https://lindeloev.github.io/tests-as-linear/
t-test does have a cool stuff such as non-homogenity of variances (Welch correction) but in practice it is not as suitable as possibility to model different confounders and interactions