Skip to main content
Humanities LibreTexts

15.6.4: Deducing Predictions for Testing

  • Page ID
    36307
  • \( \newcommand{\vecs}[1]{\overset { \scriptstyle \rightharpoonup} {\mathbf{#1}} } \) \( \newcommand{\vecd}[1]{\overset{-\!-\!\rightharpoonup}{\vphantom{a}\smash {#1}}} \)\(\newcommand{\id}{\mathrm{id}}\) \( \newcommand{\Span}{\mathrm{span}}\) \( \newcommand{\kernel}{\mathrm{null}\,}\) \( \newcommand{\range}{\mathrm{range}\,}\) \( \newcommand{\RealPart}{\mathrm{Re}}\) \( \newcommand{\ImaginaryPart}{\mathrm{Im}}\) \( \newcommand{\Argument}{\mathrm{Arg}}\) \( \newcommand{\norm}[1]{\| #1 \|}\) \( \newcommand{\inner}[2]{\langle #1, #2 \rangle}\) \( \newcommand{\Span}{\mathrm{span}}\) \(\newcommand{\id}{\mathrm{id}}\) \( \newcommand{\Span}{\mathrm{span}}\) \( \newcommand{\kernel}{\mathrm{null}\,}\) \( \newcommand{\range}{\mathrm{range}\,}\) \( \newcommand{\RealPart}{\mathrm{Re}}\) \( \newcommand{\ImaginaryPart}{\mathrm{Im}}\) \( \newcommand{\Argument}{\mathrm{Arg}}\) \( \newcommand{\norm}[1]{\| #1 \|}\) \( \newcommand{\inner}[2]{\langle #1, #2 \rangle}\) \( \newcommand{\Span}{\mathrm{span}}\)\(\newcommand{\AA}{\unicode[.8,0]{x212B}}\)

    Condition 1, the deducibility condition, is somewhat more complicated than a first glance might indicate. Suppose you suspect that one of your co-workers named Philbrick has infiltrated your organization to spy on your company's chief scientist, Oppenheimer. To test this claim, you set a trap. Philbrick is in your private office late one afternoon when you walk out declaring that you are going home. You leave a file folder labeled "Confidential: Oppenheimer's Latest Research Proposal” on your desk. You predict that Philbrick will sneak a look at the file. Unknown to him, your office is continually monitored on closed-circuit TV, so you will be able to catch him in the act.

    Let's review this reasoning. Is condition 1 satisfied for your test? It is, if the following reasoning is deductively valid:

    • Philbrick has the opportunity to be alone in your office with the Oppenheimer file folder, (the test's initial conditions).
    • Philbrick is a spy. (the claim to be tested)
    • So, Philbrick will read the Oppenheimer file while in your office, (the prediction)

    This reasoning might or might not be valid depending on a missing premise. It would be valid if a missing premise were the following:

    If Philbrick is a spy, then he will read the Oppenheimer file while in your office if he has the opportunity and believes he won’t be detected doing it. (background assumption)

    Is that premise acceptable? No. You cannot be that sure of how spies will act. The missing premise is more likely to be the following hedge:

    If Philbrick is a spy, then he will probably read the Oppenheimer file while in your office if he has the opportunity and believes he won’t be detected doing it. (new background assumption)

    Although it is more plausible that this new background assumption is the missing premise used in the argument for the original prediction, now the argument isn't deductively valid. That is, the prediction doesn't follow with certainty, and condition 1 fails. Because the prediction follows inductively, it would be fair to say that condition 1 is "almost" satisfied. Nevertheless, it is not satisfied. Practically, though, you cannot expect any better test than this; there is nothing that a spy must do that would decisively reveal the spying. Practically, you can have less than ideal tests about spies or else no tests at all.

    In response to this difficulty with condition 1, should we alter the definition of the condition to say that the prediction should follow either with certainty or probability? No. The reason why we cannot relax condition 1 can be appreciated by supposing that the closed-circuit TV does reveal Philbrick opening the file folder and reading its contents. Caught in the act, right? Your conclusion: Philbrick is a spy. This would be a conclusion many of us would be likely to draw, but it is not one that the test justifies completely. Concluding with total confidence that he is a spy would be drawing a hasty conclusion because there are alternative explanations of the same data. For example, if Philbrick were especially curious, he might read the file contents yet not be a spy. In other words, no matter whether the prediction comes out to be true or false, you cannot be sure the claim is true or false. So, the test is not decisive because its result doesn't settle which of the two alternatives is correct.

    Yet being decisive is the mark of an ideally good test. We would not want to alter condition 1 so that this indecisive test can be called decisive. Doing so would encourage hasty conclusions. So, the definition of condition 1 must stay as it is. However, we can say that if condition 1 is almost satisfied, then when the other two conditions for an ideal test are also satisfied, the test results will tend to show whether the claim is correct. In short, if Philbrick snoops, this tends to show he is a spy. More testing is needed if you want to be surer.

    This problem about how to satisfy condition 1 in the spy situation is analogous to the problem of finding a good test for a non-universal generalization. If you suspect that most cases of malaria can be cured with quinine, then no single malaria case will ensure that you are right or that you are wrong. Finding one case of a person whose malaria wasn't cured by taking quinine doesn't prove your suspicion wrong. You need many cases to adequately test your suspicion.

    The bigger issue here in the philosophy of science is the problem of designing a test for a theory that is probabilistic rather than deterministic. To appreciate this, let’s try another scenario. Suppose your theory of inheritance says that, given the genes of a certain type of blue-eyed father and a certain type of brown-eyed mother, their children will have a 25 percent chance of being blue-eyed. Let's try to create a good test of this probabilistic theory by using it to make a specific prediction about one couple's next child. Predicting that the child will be 25 percent blue-eyed is ridiculous. On the other hand, predicting that the child has a 25 percent chance of being blue-eyed is no specific prediction at all about the next child. Specific predictions about a single event can't contain probabilities. What eye color do you predict the child will have? You should predict it will not be blue-eyed. Suppose you make this prediction, and you are mistaken. Has your theory of inheritance been refuted? No. Why not? Because the test was not decisive. The child's being born blue-eyed is consistent with your theory's being true and also with its being false. The problem is that with a probabilistic theory you cannot make specific predictions about just one child. You can predict only that, if there are many children, then 25 percent of them will have blue eyes and 75 percent won't. A probabilistic theory can be used to make predictions only about groups, not about individuals.

    The analogous problem for the spy in your office is that when you tested your claim that Philbrick is a spy you were actually testing a probabilistic theory because you were testing the combination of that specific claim about Philbrick with the general probabilistic claim that spies probably snoop. They don’t always snoop. Your test with the video camera had the same problem with condition 1 as your test with the eye color. Condition 1 was almost satisfied in both tests, but strictly speaking it wasn't satisfied in either.

    Our previous discussion should now have clarified why condition 1 is somewhat more complicated than a first glance might indicate. Ideally, we would like decisive tests or, as they are also called, crucial tests. Practically, we usually have to settle for tests that only tend to show whether one claim or another is true. The stronger the tendency, the better the test. If we arrive at a belief on the basis of these less than ideal tests, we are always in the mental state of not being absolutely sure. We are in the state of desiring data from more tests of the claim so that we can be surer of our belief, and we always have to worry that someday new data might appear that will require us to change our minds. Such is the human condition. Science cannot do better than this.


    This page titled 15.6.4: Deducing Predictions for Testing is shared under a CC BY-NC-SA license and was authored, remixed, and/or curated by Bradley H. Dowden.

    • Was this article helpful?