Replication & Serial Position Curve (Dan Reisberg) ch. 5
In the methods essays so far, we've talked about some of the steps needed to
make sure an individual result from a particular experiment is unambiguous.
We've talked about the need for a precise hypothesis, so that there's no
question about whether the result fits with the hypothesis or not. We've talked
about the advantages of random assignment, to make certain that the result
couldn't be the product of preexisting differences in our comparison groups.
We've discussed the need to remove confounds so that, within the experiment,
there is no ambiguity about what caused the differences we observe.
Notice that all of these points concern the interpretation of individual
results, so that each experiment yields clear and unambiguous findings. It's
important to add, though, that researchers rarely draw conclusions from
individual experiments, no matter how well designed the experiment is. One
reason for this is statistical: A successful replication-a reproduction
of the result in a new experiment-provides assurance that the original result
wasn't just a fluke or a weird accident. Another reason is methodological: If we
can replicate a result with a new experimenter, new participants, and new
stimuli, this tells us there was nothing peculiar about these factors in the
first experiment. This is our guarantee that the result was produced by the
factors deliberately varied in the experiment and was not the chance by-product
of some unnoticed factor in the context.
In addition, researchers generally don't repeat experiments exactly as they were
run the first time. Instead, replications usually introduce new factors into the
design, to ask how these alter the results. Specifically, researchers offer a
hypothesis about the original result and then deduce from this hypothesis
predictions about factors that should alter the data pattern. Testing these
predictions allows them to test the hypothesis.
We gave an example of this method in the textbook: If people are asked to recall
as many words as they can from a list they just heard, the results show a
characteristic U-shaped serial-position curve. This result is easily replicated,
so we know it doesn't depend on idiosyncratic features of the experimental
context. We therefore want to ask, what causes this reliable pattern? One
proposal, of course, is provided by the "modal model," a theoretical account of
memory's basic architecture. But is this model correct?
To address this question, researchers have varied factors in the basic
list-learning experiment that should, if the hypothesis is correct, alter the
results. One factor is speed of list presentation: According to our hypothesis,
if we slow down the presentation, this should increase recall for all but the
last few words on the list. A different factor is distraction right after the
list's end: Our hypothesis predicts that this will decrease the recency effect
but will have no other effects. These predictions both turn out to be right.
Notice, then, that our confidence in our hypothesis rests on many
results-results showing the replicability of the basic finding, and then other
results testing predictions derived from our hypothesis. In the end, it's this
fabric of results, easily explained by our hypothesis and not easily explained
in any other way that convinces us that our hypothesis is indeed correct.