A Taxonomy of Major Premises and Implications for Falsification and Verification

ABSTRACT

Both naïve and sophisticated falsification arguments depend upon using the logic of modus tollens to employ empirical defeats to conclude that the theory is not true (naïve falsification) or that the conjunction of theory and auxiliary assumptions is not true (sophisticated falsification). However, both types of modus tollens syllogisms depend strongly on the precise form of the major premise. The present goal is to introduce a taxonomy of four types of major premises: positive frame, negative frame, positive-negative frame, and negative-positive frame. For two frames, falsification really is logically superior to verification, but for two other frames, verification is logically superior to falsification. Thus, there is no overall logical basis for preferring falsification to verification or for preferring verification to falsification. Nor are there convincing arguments from Bayesian statistical, frequentist statistical, or empirical standpoints. Thus, the attraction of falsification, whether naïve or sophisticated, comes down, in the end, to something other than an objective philosophical account.

Acknowledgements

I thank Michael Hyman, Timothy Ketelaar, Uli Widmaier, a set of anonymous reviewers, and Editor Vasso Kindi for their helpful comments in the preparation of the present article. Any errors are my responsibility alone.

Notes

1 Popper’s main arguments pertaining to falsification are in Popper (1959), which is a translation of an earlier work and in Popper (1963). Popper advocated for sophisticated falsification. Maxwell (2002) provides an accessible review.

2 An unsophisticated verification argument that does not consider auxiliary assumption is also invalid and similarly commits the fallacy of affirming the consequent: $T\to E$, $E$, $\therefore T$. Also, even a weakened sophisticated verification argument concluding that the theory is true or that the set of auxiliary assumptions is true, is invalid: $\left(T\mathrm{\&}A\right)\to E$, $E$, $\therefore T\mathrm{v}A$.

3 I thank an anonymous reviewer for suggesting this counterargument.

4 I thank an anonymous reviewer for pointing out this issue.

5 Rohrer, Pashler, and Harris (2019) found additional discrepancies between their own findings and those that had been reported by Vohs, Mead, and Goode (2006).

6 There is a large and varied literature indicating confirmatory bias (see Fine 2006; Friedrich 1993; Nickerson 1998 for reviews). It is possible that the high frequency of positive framing in science is connected to this bias (see Hergovich, Schott, and Burger 2010; Klayman and Ha 1987; MacCoun 1998 for reviews specifically relating confirmation bias to scientific hypotheses).

7 The null model includes the null hypothesis but also includes a variety of additional assumptions. Bradley and Brand (2016) and Trafimow (2019a) provided taxonomies of such assumptions. Thus, there is no such thing as a pure test of the null hypothesis.

8 This is not to advocate for Bayesian statistics, as there are problems with respect to the issue of prior probabilities reviewed in an excellent book by Gillies (2000).

9 One reason, not of much relevance here, is that more advanced measurement theories do make assertions about reality that might be wrong (Lord and Novick 1968). For example, item response theory assumes the existence of an underlying trait that causes responses on the scale items and that assumption might be wrong (Hulin, Drasgow, and Parsons 1983; Lord and Novick 1968). That said, however, few researchers other than Lord and Novick (1968) made a point of this. For example, Gulliksen (1987) did not make this point in his famous and extensive review but did emphasize the attenuation equation. Therefore, it is unlikely that the issue of possible wrong assumptions has mattered much historically, though perhaps it should have.