Between groups designs (2) – outline 1.Block randomization 2.Natural groups designs 3.Subject loss 4.Some unsatisfactory alternatives to true experiments – One group posttest only design – Posttest only with non-equivalent control group – One group pretest-posttest design Between 2
Block Randomization Block randomization (BR) is used to form groups of equal sizes First you create groups (called “blocks”) Then you randomly assign members of a block to your experimental treatments Between 2
Block Randomization # in each block = # of treatments e.g., 4 treatments 4 subjects per block In that case, first 4 subjects to sign up would form Block 1, second 4 subjects to sign up would form Block 2, and so on. Subjects in each block now randomly assigned to treatments Between 2
Block Randomization Block randomization yields treatment groups which all have the same size. – This is important for many statistical tests – Equal ns mean (roughly) equal variances and thus comparable reliability Plus, BR will cause ‘history’ effects to affect all groups equivalently Between 2
Block Randomization BR will eliminate confounding history effects – changes in experimenter – changes in the population (e.g., 1 st vs. 2 nd semester of Psych 020) – actual historic events – imagine if you had run your control group in the week September 3 – 7, 2001 and treatment September 10 – 14, 2001 – block randomization will eliminate such confounds, at the expense of greater error variance Between 2
2. Natural Groups Designs Natural groups designs are those in which individual difference variables are selected rather than manipulated. A simple example is when you use age or sex as an independent variable – you cannot randomly assign people to the conditions “young” or “old,” or to “female” or “male.”
2. Natural Groups Designs We also use natural groups designs when ethical constraints keep us from assigning people to groups E.g., you could assign people to “divorce” and “no divorce” treatments, and perhaps even pay people to get divorced or stay married. But to do so would be unethical Instead we would compare people who have chosen to get divorced to people who have chosen not to – a natural groups design
2. Natural Groups Designs Natural groups designs are useful for: Description – Do divorced people receive psychiatric care at a higher rate than those who are married? Prediction – If so, we can predict that a new set of divorced people is more likely than a new set of married people to need psychiatric care
2. Natural Groups Designs But natural groups designs cannot be used to make inferences about cause! Natural groups designs are correlational studies, not experiments You must NOT draw causal inferences from studies that use natural groups designs (that is, do not offer opinions about what causes any differences on your dependent variable between the groups).
2. Natural Groups Designs Since you did not establish equivalence of your groups at the beginning of your study (you did not randomly assign people to groups), you have not eliminated plausible alternatives to any causal account that you might offer. E.g., do divorced people need more psychiatric care because of the stress of divorce? Or do people who need more psychiatric care place more strain on their relationships or choose a mate unwisely in the first place?
Subject loss For a between-groups experiment to be internally valid, we need the two groups to be equivalent not only at the beginning of the experiment, but also at the end. If more subjects drop out of one group than out of another, the two groups may no longer be comparable. Between 2
Subject loss Two kinds of subject loss: A.Mechanical – subject is lost from the experiment because of equipment failure. – This is probably a random effect – thus, will not produce systematic differences between the two groups. Between 2
Two kinds of subject loss B. Selective – this is when some characteristic of either the subject or the treatment is responsible for the loss – e.g., treatment involves a difficult or unpleasant task, but control condition does not – clinically depressed subjects compared with sub- clinically depressed controls – the most severely depressed subjects in the former group may be the most likely to drop out Between 2
Two kinds of subject loss B. Selective – what can you do? If you notice this loss after the fact, nothing. If you anticipate such loss, you may be able to screen people on some variable that will let you predict loss, and then select subjects on that basis – at a cost to generalizability. Between 2
But what about external validity? Random assignment in Loftus & Burn’s study guaranteed internal validity – the group difference in performance could not have been caused by anything other than the treatment. But what about external validity? – Would the same effects be found with a real-life bank robbery instead of one on film? – Would the same effects be found with people other than young university students? Between 2
But what about external validity? As Stanovich points out, the answer is often, “who cares?” – we often do an experiment to test a particular theory, not to find out what the ordinary person would do in the real world – Often, any kind of subject will do to test our theory, so long as they are competent in our experimental task Between 2
But what about external validity? Of course, sometimes generalizability matters. – if so, then try for representative samples & situations – when you can’t do that, at least use several different types of people, stimuli, and situations – or replicate – partial or complete replication – or use meta-analysis: review of published papers Set criteria for inclusion of papers in your review Select a procedure for amalgamating findings Between 2
Some unsatisfactory alternatives to experiments All of the following fail to control for important threats to the validity of a conclusion: One group posttest only design – Can’t tell if treatment changed behavior if you don’t know what behavior was like to start with. Between 2
Some unsatisfactory alternatives to experiments Posttest only with non-equivalent control group – Control & treatment groups are not equated at the start. – Differences between treatment and control groups could be due to treatment or to other things (since control group is not equivalent). Between 2
Some unsatisfactory alternatives to experiments One group pretest-posttest design – Change in behavior may have been caused by variables other than the one you think produced it. (E.g., maturation, attention, change in the weather…) Between 2