I have a data , like this
Genotype condition time
A uninfected 0day
A uninfected 0 day
A uninfected 0 day
A mock 2day
A mock 2day
A mock 2 day
A infetced 2day
A infected 2 day
A infected 2 day
A mock 4day
A mock 4 day
A mock 4day
A infected 4day
A infected 4day
A infected 4 day
B uninfected 0 day
B uninfected 0 day
B uninfected 0 day
B mock 2day
B mock 2day
B mock 2day
B infected 2day
B infected 2day
B infected 2day
B mock 4day
B mock 4day
B mock 4day
B infected 4day
B infected 4day
B infected 4day
I have to apply a model such that i can compare the effect of treatment over time of two different genotypes.
I have used the design formula Genotype+condition+time+condition:time+Genotype:time
and then LRT test using Genotype+condition+time
Due to only one uninfected condition in each genotype it is saying
Model matrix is not full rank
1.Which model can I use for this type of data for time series comparision.
2.Can i add uninfected again before 4 day (as this is the same). Will it be statistically correct ?
•
written
by
hamsasekar02 •
Hi,
First of all, if think that there are typos in your conditions as sometimes you write (un)infetced (instead of infected). Your design is not full rank because all "uninfected" are on "day0". A possible solution (that may or may not be good depending on your experimental settings) could be to merge the mock and uninfected levels :
Genotype condition time
A uninfected 0
A uninfected 2
A infected 2
A uninfected 4
A infected 4
B uninfected 0
B uninfected 2
B infected 2
B uninfected 4
B infected 4
Thank you for the quick reply We need to actually compare uninfected with mock and infected, so merging will be a little difficult option
Ok, but as I said, in your design, sample uninfected and sample on Day0 are basically the same, that is why u get the error. As long as you include the interaction between "condition" and "time", the error will be there.
In my proposed workaround, even if you can't compare uninfected with mock directly, u can compare uninfected samples on day0 with uninfected samples on day2/day4, which is basically the same thing.
I'd suggest you make a column for each of mock_time and infected_time; then fit ~ -1 + Genotype + mock_time + infected_time. This only fits linearly with time (but it can be extended to include quadratic time if needed)
To do that, you just need to encode the design differently.
treatments <- data.frame(
genotype = please_name_your_data$Genotype,
# - Assuming the `time` column in the posted dataframe is actually numeric,
# rather than the strings that you've copied into your question...
mock_time = with(please_name_your_data, time * condition == "mock"),
infected_time = with(please_name_your_data, time * condition == "infected")
)
design <- model.matrix(~ -1 + genotype + mock_time + infected_time)
# etcetera
I'm not sure how to test the value of including the Genotype/Time interaction (personally, I wouldn't have considered it when coding the design).
Log in to answer this question.
You have the same problem as discussed at Error in differential analysis for samples with different time points. In short: no biological replicates.
The uninfected samples aren't useful, you can remove them. Also, as you have no replicates you will either have to ignore the effect of Genotype, condition, or time. A better solution would be to sequence replicates (ideally 6 of each condition, but at the minimum 3 of each).
Thank you for the reply Sorry, I forgot to write the point that i have 3 biological replicates for each . I want to compare uninfected with mock and with infected according to my work plan.
Why aren't the uninfected samples useful? They provide a T_0 baseline for the rest of the experiment
yes , we are also taking it as the baseline to compare with mock and infected
They have no corresponding samples with which to compare, so it's unclear if any change is actually due to time or just "doing something" (i.e., mock treatment).
On the contrary, if they didn't have the T_0 sample, their mock-infection placebo could have profound time-dependent effects but they'd have no baseline sample against which to compare those effects and wouldn't be able to identify them. This is a well designed experiment and I really don't understand the criticism
Their baseline for comparison is currently day 2, which it is regardless of whether the uninfected samples are present or not. One can't meaningfully use day 0 as a baseline because it can't be distinguished from the lack of treatment. There was likely a good biological reason for this, but as we're naive to that what I mentioned is the most we can say. This also corresponds to Carlo's answer, where he correctly notes that
mockanduninfectedcan't be compared if we think there's a change simply due to time at day 2.On the contrary, if they didn't have the T_0 sample, their mock-infection placebo could have profound time-dependent effects but they'd have no baseline sample against which to compare those effects and wouldn't be able to identify them This is exactly what we want to know and we have two genotypes which are known to have variation in the levels of genes even in the uninfected level (without any mock and infected). So, considering this we took the uninfected level