Design of experiements Flashcards
Increasing replication
Reduces SE
Why are controls necessary
To determine if the sample group is truly effected by what you are changing hence important to control other variables. ie reject null hypothesis that u1 = u2
Size of sample?
Large enough = representative
Appropriate to be ethical and cost effective
Excluding bias
Randomisation, controlling confounding factors, blinding
Power
Probability to reject the null hypothesis
Function of difference in pop means (increase AD = increased power)
SD (Increased = reduced power)
Sample size ( increased sample = increased power)
Larger populations
Need larger sample due to increase variability
Computing the sample size from power
u1-u2)/o
To determine the Power
1st = decide what difference need to be picked up ( This is often difficult to determine but crucially isnβt the difference in means or the latest literature - sometimes it isnβt possible to specify
2nd = Compute the SD 3rd = Decide on the power wanted
(2π^2 γ(π§_(πΌ/2)+π§_π½)γ^2)/γ(π_1βπ_2)γ^2
(2π^2 γ(π§_(πΌ/2)+π§_π½)γ^2)/γ(π_1βπ_2)γ^2
How do randomisation and replication control variance
Replication = o/sqn Randomisation = unbiased treatment estimator
NB there may be problems relying on randomisation alone
Restricted randomisation
Ie blocking/ stratification prevents poorly distributed groups ie in clinical trials equal number of diabetics, ages sexes in both arms
Removes confounding variables as both are present in both populations to equal extent
NB factors must be known
ANOVA
Used for stiuations in between unpaired and paired t-test. 3+ variables. Stands for analysis of variance.
For example
1 All mice in same cage = unpaired t
2 Pairs of mice in single cage x 6 = paired t test
3 Four mice in a cage x 3= ANOVA
Normal distribution
Independence of observations β this is an assumption of the model that simplifies the statistical analysis.
Normality β the distributions of the residuals are normal.
Equality (or βhomogeneityβ) of variances, called homoscedasticity β the variance of data in groups should be the same
Factorial designs
Change of two variables at once ie fluid and salt level
High fluid, high Na
High fluid, low Na
Low fluid, high Na
Low fluid, low Na
Benefits
Efficient, can investigate salt by fluid interaction
Catch of factorial design
Unfeasible for more than 2 variants
Only works if both variants are linked and measure the same thing
Interaction factorial design
Change in Na at low fluid vs change in Na at high fluid
Difference in Na as fluid level changes. Only possible with factorial design
Pseudoreplication
Artificially inflating number of samples or replicates than what occurs naturally - often occurs insidiously and due to lack of independence in replicates
Outcome of pseudoreplication
Underestimate of o hence SE. Hence more likely to get misleading small p-value
Common source of pseudoreplication
Ignoring groupings!
Wt loss between GP practises - Underestimate of o due to same GP having same nurse, same group motivation, enviroment etc not independent
Remove pseudoreplication
Experiment design
