Hi everyone,
I have a colony of bats where individuals tend to cluster. I have genotypes for a reasonable number of individuals from the colony and I know in which groups they have been caught.
I would like to test if relatedness within groups is significantly different from relatedness of the colony as an all. My specific hypothesis would be either Ha) individuals tend to cluster with relatives; or Hb) individuals tend to cluster with less related individuals.
I am using the ‘related’ R package (https://frasierlab.files.wordpress.com/) which is mostly an R implementation of Coancestry (Wang, J. 2011). The package includes a ‘grouprel’ function, which
“calculate the average relatedness within each of the specified groups. It will then generate a distribution of ‘expected’ relatedness values by randomly shuffling individuals between groups, while keeping each group size constant, and calculating the average relatedness within each group for each randomization step.”.
Contrary to the ‘Test of difference between groups’ implemented in Coancestry, it only allows to calculate “Relatedness Within Groups”, but the output files show the randomization for each group and to the overall data-set (considering group sample size).
To test my hypothesis my reasoning would be to compare the observed relatedness within each group with the expected relatedness within the colony as an all (is this correct?). Let’s take group 1 (g1) as an example (check Figure 1 attached). As you can see, relatedness within ‘g1’ is lower than expected considering the all colony.
According to the ‘related’ tutorial, the P-values from ‘grouprel’ function (shown in Figure 1) measure if relatedness is HIGHER than expected. What about if it’s lower? Would it be correct to calculate P in an inverse way? Returning to my example, if I calculate P to compare g1 and the remaining colony as ‘the percentage of randomized iterations where the expected values (for the all colony) were greater than or equal to the observed value (of g1)’ I will get p < 0.99. This means that 99% of the times, the expected relatedness is higher than the observed. Would it than be correct to state that only 1% of the times, expected relatedness of the all colony is lower than relatedness in g1 and therefore individuals in group g1 tend to cluster with less related individuals?
Of course I understand this is always dependent on sample size.
Thank you for your contribution!
- Francisco Amorim
It may be better to compare 'within groups' with 'between groups' rather than 'all colony'. But your figure shows higher relatedness for G1 than for the colony, doesn't it? The blue line extends further than the red, indicating higher relatedness on the x axis? Yes, I think it is better to inverse your p values :)
Dear Andrew,
Thank you for your reply.
Regarding your first comment, I don't know if I fully understand. Do you mean compare G1 with G2? I haven't done this yet, but if I find that there's no difference between groups, what does this tell me about the colony 'arrangement'. Contrary, if I find that group relatedness is consistently different from that of the colony that would suggest that they are not clustering randomly.
As for the second point, this is done with 500 iterations, in each individuals from the complete data set are shuffled according to group size. This is a way to see what would the (expected) relatedness level if they grouped randomly. The blue line extends to both ends and this probably as to do with lower sample size in G1. The point is, with that sample size you have relatedness observations falling to the left of the red line (observed relatedness in G1), and for that reason we conclude that observed relatedness in G1 does not differ from the expected. However, for the pink line, which represents expected relatedness of all the individuals in that colony, you have very few values to the right of the observed relatedness, meaning there are differences...
Am I missing something here?
Hi Francisco, I'll see if I can answer this tonight. I'm in New Zealand :)
Yes, a GLM would be good, comparing 'within groups', 'between groups', and 'whole population'. Then you can see all possible interactions. Don't worry about what the outcome will be, you just have to report and explain it, even if it's a surprise!
I think the mode (the value x at which its probability mass function takes its maximum value) is the same for 'expected colony' and 'expected g1' values and does actually differ from the 'observed g1' value. You just have to accept it and try to explain it. That's what research is about, and anomalies make it interesting! But I think you need to do more observations (probably 30 i.e. g2...g30) to see if there is a pattern :)
Hi Andrew,
Sorry for taking so long to reply!
Thank you once again for your reply. What would be the workflow for a GLM? You're suggesting using the groups as predictors to explain relatedness?
Regarding your second point I totally agree and yes, my goal is to do it for multiple groups and see if there's any pattern coming out.
Thank you