Chapter 14 Experimental Design

replication

- application of every treatment to multiple, independent experimental units. avoid pseudoreplication or confounding variables that arise due to experimental design.

Experimental Studies

- assigns treatment randomly to individuals.

Association v. Causation

- association: other variables affect both X and Y at the same time. - causation - changes in one variable directly causes change in another.

Observational Studies

- environmental variables are not controlled by the researcher.

Sources of Sampling Error

- errors of chance - bias - samples of convenience - pseudoreplication - occurs when sampling units are treated as though they are independent, but are not.

Confounding Variables

- factors that mask or distort the casual relationships between the explanatory and response variables in a study. ex) unknown (genetics, physiology) ex) uncontrolled (temperature, time) ex) experimentally introduced artifacts (stress, behavioral changes) - with random assignment, no cofounding variables will be associated with treatment except by chance.

Blocking

- grouping of experimental units that have similar properties. - breaks the association of a confounding variable, thus reducing noise.

reduce sampling error

- increase precision and power by improving signal to noise ratio. 1. Sample size 2. Replication 3. Strong treatments 4. blocking

strong treatments

- increasing differences between control and experimental samples. improves signal to noise and power. - greater differences Y1-Y2. - consider potential moneary costs.

Factorial Design

- investigates all treatment combinations of two or more variables - examines effect of 2 or more x on a y. - interaction means that effect of one variable depends on the 2nd variable.

sample size

- larger sample size --> smaller SE --> smaller noise. - balanced: all treatments have equal sample sizes. when n1=n2, standard error is smallest and thus precision is highest. - balanced n --> more significant t tests.

Planning for Precision and Power

- planning for precision involves choosing n that yields CI of expected width. n=8( sd/ margin of error)^2 ^sample size for a given margin of error n= 16 (sd/ D)^2 where D is minimum diff. of interest ^ n for 80% power.

Properties of a Good Sample

- precise (low variation) and accurate (near true value)

reduce/eliminate bias

-minimize effects of confounding variables - ways to reduce bias: 1) control - negative/positive 2) randomization - random assignment of treatments. averages out the effects of confounding variables and breaks their assoc. with the explanatory variables 3) blinding - unblinded studies tend to find much larger effects or positive effects at a higher rate. single-blind/double-blind.

bias in clinical trials

1. natural recovery 2. placebo effect 3. regression toward the mean - even w/o treatment the second measure will tend toward the mean

signal to noise

improves power - ability to distinguish differences in means.