Page:Sm all cc.pdf/23

 are of the form ‘A probably is related to B, because within my sample of N specimens I observe that the Ai are correlated with Bi.’

Replication and Confirmation
The terms replicatability and reproducibility are often used to refer to the similarity of replicate measurements; in this sense they are dependent only on the precision of the measurements. Sometimes replicatability is used in the same sense as replication, describing the ability to repeat an entire experiment and obtain substantially the same results. An experiment can fail to replicate because of a technical error in one of the experiments. More often, an unknown variable has different values in the two experiments, affecting them differently. In either case the failure to replicate transforms conclusions. Identifying and characterizing a previously unrecognized variable may even eclipse the original purpose of the experiments.

Replication does not imply duplication of the original experiment’s precision and accuracy. Indeed, usually the second experiment diverges from the original in design, in an attempt to achieve higher precision, greater accuracy, or better isolation of variables. Some [e.g., Wilson, 1952] say that one should not replicate an experiment under exactly the same conditions, because such experiments have minor incremental information value. Exact replication also is less exciting and less fundamental than novel experiments.

If the substantive results (not the exact data values but their implications) or conclusions of the second experiment are the same as in the first experiment, then they are confirmed. Confirmation does not mean proved; it means strengthened. Successful replication of an experiment is a confirmation. Much stronger confirmation is provided by an experiment that makes different assumptions and different kinds of measurements than the first experiment, yet leads to similar interpretations and conclusions.

In summary, precision is higher than accuracy, because accuracy is affected by both precision and systematic biases. Accuracy is higher than reliability, because reliability is affected not just by measurement accuracy but also by the validity of assumptions, simplifications, and possibly generalizations. Reliability is increased if other experiments confirm the results.

Probability
Probability is a concern throughout science, particularly in most social sciences, quantum physics, genetics, and analysis of experiments. Probability has a more specific meaning for mathematicians and scientists than for other people. Given a large number of experiments, or trials, with different possible outcomes, probability is the proportion of trials that will have one type of outcome. Thus the sum of probabilities of all possible outcomes is one.

Greater probability means less uncertainty, and one objective of science is to decrease uncertainty, through successful prediction and the recognition of orderly patterns. Induction (Chapter 3), which is a foundation of science, is entirely focused on determining what is probably true. Only by considering probability can we evaluate whether a result could have occurred by chance, and how much confidence to place in that result.

“Looking backwards, any particular outcome is always highly improbable” [Calvin, 1986]. For example, that I am alive implies an incredibly improbable winning streak of birth then reproduction