I had a quick look at your data; I do not understand well what "animal" represents, seems to be a number but coding for the genotype?

My guess would be to use a two-way, nested ANOVA and see what it gives before trying to select replicates :

essai v

summary( v )

anova( essai )

Residuals seem to be fairly gaussien, so outliners are may be not so evident. However, equality of variances assumptions may be wrong (highest is about 15 times the smallest).

Nevertheless, ANOVA table suggests that there is no treatment effect on speed, but that it may vary from one mouse to another.

After that... Since speed does not seem to depend on the treatment, you may use it as a covariate in further analyses? For selecting replicates, I am not sure it is a good idea unless based on biological grounds, but if needed I would instead select the first quartile of runs for each mouse and then do the previous analysis of variance. But if speed is really different between mice, I do not see how it would solve the problem, since the first quartile should still differ. With your version, if I understand weel (pool all runs, select the results of mice with runs in the first quartile of these pooled runs), however, it may be homogeneous, but if treatment has an effect on speed, it will introduce biases. And why the first quartile and not the first half?

Hope this help a little...

Thank you very much Emmanuel. Nice way the linear model to check influence of sample repeat, I did not think about it.

The variable "Animal" is the ID of the animal of interest (name of the sample which is repeated).

One more comment, in the distribution of the data I am used to find two kinds of samples:

1/ In the bottom tail of the data set: some animals are not interested to visit the beam and have few runs with low speed. In this case, it would make sens to exclude them because their over-care behavior is hiding their normal gait pattern.

2/ At the bottom tail of the data set: some animals which are stress by the new environment and run like hell on the beam. It would make sens to exclude them because the running pattern is really far from a usual gait pattern.

So I am looking for an elegant way to identify them in a reproducible manner on my data set. I was thinking that an individual analysis (in the sens of analyzing independently every animal) seems more neutral that working per group of animals.

Any suggestion?

Not familiar with such experiments... If you see the animals during it, can you "feel" such cases (ideally blindly of the treatment received) and annotate such animals? Then, you may try analysis with or without annotated animals and see if it change the results [this last point, whatever the way to remove some animals anyway].

I am not sure you can derive purely number-based rules to identify such animals, unless you have knowledge from previous experiment of what is a "normal case" value...

Extending your idea, I would use not the first quartile, but the 50 % around the median (the box part in the boxplot), or the 95% around the median or... to remove extreme values on both side.

Question for animals id : you have "9.4" = tg, "9.5" = ctl: is it the same animal ("9") that is tried in the two conditions ("own control", one run = 4, one run = 5), or different animals and coincidence of the numbering scheme?

In the first case, you may need to consider the order in which you apply the treatment and the control and make some kind of cross-over plans.

In the second case, using completely different numbering between the two groups may avoid some surprises in the final analysis of the data (inadvertently using crossed effects instead of nested, for instance).

PS: note also that if a side effect of the treatment is to change the stress state of the animal, trying to remove such animals will introduce biases in the analysis of the treatment effect... So you may want to check that the repartition of removed animals in the same in the two groups, and can be related to chance and not the treatment.

For numbering of animal, 9.5 mean that the animal is coming from the cage 9 ear mark number 5. It is not correlated to the treatment in order to let experimenter blind regarding the treatment.

Moreover, noticing stress or lack of interest is really difficult on this test, it could be that the same animal is stress the 2 first minutes in the beam and after 5 minutes is already bored to cross it.

We are exploratory behavior and stress via other methods (if you are interested on animal behavior I will be happy to share some really nice articles about it).

But in general stress is for this experiment a noise that we would like to remove because we are looking for coordination, and coordination is reduces with stress:

http://www.sciencedirect.com/science/article/pii/S0006899305001265

http://www.pnas.org/content/102/11/4170.abstract

Concerning the 50% around the median, it will give a try to it. I think I need to find a method to graph the correlation between speed and my other parameters in order to convince that the exclusion of "stress" or "not interested" runs is required to calculate accurate coordination parameter.

I am going to plot that and let you know!

Thank you again Emmanuel.

Walking speed is the most important confounder in the interpretation of gait data. There are large postural and kinematic differences between a subject walking 15 cm/s compared to 20 cm/s. It is important to ensure that all of your subjects are walking the same speed for many of the key metrics. Stride length, for example, used widely as a metric of neurodegeneration or other movement disorder is dependent on speed. Since 10% differences are potentially significant 0 in a mouse, say 5 cm vs. 5.5 cm - the apparent wide spread you have in your speeds is going to obfuscate any speed-independent differences you might have in stride length.

Thanks Thomas.

I agree that gait parameters are dependent of the speed, so that it is important to keep it similar in each group.

How would you handle data coming from a transgenic mouse, or mouse treated with a drug, have a really reduced walking speed because of obvious walking abnormality?

Would you then got a look on parameters less sensitive to walking speed?

not just similar, Casadei, but the same=equal. Many aspects of your gait are different when you walk down the hall to get a coffee vs. run down sidewalk to catch the metro...stride length, stride frequency, stance width, angle of foot placement, braking, propulsion, etc. Which parameters do you believe to be less sensitive to walking speed? You would need to purposefully construct a data set of these metrics collected over the [a] range of known speeds....and thereby know which are not sensitive to speed.

Thanks for insisting on it :)

Nice idea to have a look on several walking speed instead of focusing on one. Do you have a reference in mind where it is nicely made?

Else what you you suggest as an acceptable difference of speed between animals? I was looking via one way ANOVA that the speed between groups was the same but i guess I can find something more elegant.

From your excel summary it appears you have a range of speeds from about 10-25 cm/s if I am reading that correctly...that is a place to start. The range of walking speeds fo healthy mice is about 14 t0 45 cm/s....you could look at papers from Clarke and Still from the UK in the early 1990s. Below 14 is more of a foraging behavior than a walk, and above 50 cm/s you will start getting hopping [e.g. see Crone SA et al. J Neurosci. 2009 May 27;29(21):7098-109].