BIO361H5 Lecture 5: Lec 5 - Sept 20
![](https://new-preview-html.oneclass.com/Gy340B26qOkoN9bLWp6WQAdXzp1xr5w7/bg1.png)
Sept 20 BIO361: Lec 4
p-value (how many times did we get a value that was extreme and divided it by the # of reps)
larger randomization = the smaller the difference. The more # reps → the more stable the p-value
that you get if you re-run the randomization
feasibility – how much time it takes to run the code
the larger # of randomizations → may take rStudio to take longer
keep changing # of randomizations and see when the p-value will start to stabilize out to see
what # of randomization you use
how to know when results are correct → run different tests, assumptions may not be met when
comparing other statistical tests. Figure out how many different ways there are to divide the 10
numbers (the sample) into the # of groups (2), how many of those have a difference/p-value that
is larger than what was observed.
How good is this p-value?
Empirical p ~ 0.03
Exact p: list all combos of sampling 5 out of 10 mice in treatment
• 252 possibilities
•
• Calculated difference b/t treatment and control fo all 252 combos
• Calculate proportion of differences > 7.6 = exact p-value
• 7 values > 7.6 so p = 7/252 = 0.0278
What other tests could we use?
Two sample t-test, ANOVA, non-parametric tests
Data doesn’t have to be normal for nonparametric tests
Nonparametric uses ranks – the 10 values go from highest to lowest value
• Less accurate and less quality, doesn’t tell you the difference just ranks
• Less powerful than parametric tests
• Always get larger p-values with non-parametric tests
• Parametric tests actually use the measurements
Kruskal wallis – using subset of data just using females (or male)
If assumptions are met use the two-sample t-test (more powerful to detect effect since it is a
parametric test)
If assumptions aren’t met (randomization or non-parametric tests?)
Randomization just deals with data (not ranks)
Can we increase sample size by increasing randomization?