Redhead 1980 Bayesian Recons

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

M. L. G.

REDHEAD
A BAYESIAN RECONSTRUCTION OF THE METHODOLOGY OF SCIENTIFIC RESEARCH PROGRAMMES*
IN A RECENT paper Dorling has discussed a Bayesian approach to the Duhem problem of attributing refutations to particular component parts of a theory and has indicated the relevance of these ideas to a rational reconstruction of Lakatoss methodology of scientific research programmes (MSRP). In a particular example Dorling has demonstrated an asymmetry in the effect of refutation on the posterior probabilities of a hard core (7J and auxiliary hypothesis (H) belonging to the protective belt of the research programme which would justify retaining T and abandoning H. Dorlings result relies on assigning particular values to various conditional probabilities and it is not entirely transparent how critically the resulting asymmetry depends on the values chosen for the illustrative example. The purpose of the present note is to express Dorlings approach in a way which makes its impact more easily assessable and then to consider in more detail how one can model the life and death of a research programme using a special case of the general formalism. Denoting by e the result of an experimental test of T.H we introduce the following five parameters
X

Y k, k, k,

P(T) = P(H) = p(e, T.mH) = p(e,-T.H)


=

= p(e,-T.-H)

and follow Dorling in assuming the probabilistic independence of T and H. If we define enhancement ratios

*An earlier version of the present work formed part of a paper read in Professor Watkinss seminar at the London School of Economics in January 1976. 1 am grateful to Jon Darling for explaining his ideas to me. J. Do&g, Bayesian Personalism, The Methodology of Scientific Research Programmes, and Duhems Problem, Stud. Hist. Phil. Sci. 10 (1979). 177. In the case of refutation these ratios will of course turn out to be less than instead of greater than unity. Stud. Hisf. Phi/. Sci., Vol. 11, (1980), No. 4, pp. 341 - 347. Pergamon Press Ltd., Printed in Great Britain. 341

342

Studies in History and Philosophy of Science

yT = p( T,e)/p( T), yH = p(H,e)/p(H), then elementary algebraic manipulation of Dorlings use of Bayess theorem yields the following general results: If e confirms T-H

(1)

YT

=
xy +

y + k1(1-Y)

k,x(l-y)

+ k&l-X)

+ kJ(l-XN1-Y)

(2)

YH

= xy + k&l-y)

x + k?(l_X)

+ kg(l-x)

+ k,(l-x)(1-y)

while if e refutes T-H then =

kl(l-y)
k&(1-y) + k&l-x) + Ml-X)(1-Y)

(3)

YT

(4)

YH

=
k,x(l-y)

Ml-xl
+ k&l-x) + kdl-xX1-y)

If we now define an asymmetry ratio by A = yT/yH we get:

For confirmation
Y

(5)

+ k(l-y)

x + kz(l-x)

and for refutation


=---* k, k, 1-Y p l-x

(6)

Discussion

343

Notice that the asymmetry ratio does not depend on ks at all. In his example Dorling takes k, = k&l and k& k,, kJ. This gives a large asymmetry ratio for refutation since :; %1. But we can also get a large value for this ratio if x is much closer to unity than Y even if k, = k2. Indeed, the simplest model of a MSRP situation is obtained by taking k, = k, = k, = k, say, where k4 1 and% $1. In this case our formulae reduce to the following:

For confirmation = Y + 41-Y) XY + k(l-xy)

(7)

Yr

(8)

YH

x + &l-x)

xy + k(l-xy)

(9)

Y + 41-Y) xy + k(l-x)

and for refutation (10) yr = I-xy


1-Y

(11)

yH

l-x = 1-xy

(12)

=-,1-Y
l-x

Notice how for refutation the enhancement ratios and asymmetry ratio do not depend on the value assumed for k. For the case of confirmation, however, for k = 1, yT = yn = 1 so there is no enhancement; while for k4xy, yP 1 and X

344

Studies in History and Philosophy of Science as we would expect. can now be modelled using the of successful we illustrate followedby and the a run for the figure

yHa i or alternatively p(T,e)^l and p(H,e)^l, The life and death of a research programme formulae unsuccessful following of failures, (7), (8), (10) and predictions. sequence: (11) to allow In the

for the effects and failures,

accompanying successes as parameter

two alternating

where we have assumed

values x = 0.9, y = 0.5 and

k = O- 1. The probability

pT for the hard core, and the probability

current articulation of the protective belt pH are both shown. After each refutation we take pH = 0.5 again for the new auxiliary hypothesis introduced to account for the anomaly. We are assuming here the simple situation in which each articulation of the protective belt is triggered by an anomaly.3

SFSFFFFFFFF

Sequence ----p p, H

of predictions predictlon

S = Successful F = Failed

predlction

Lakatos of course stressed that realistic research programmes can progress in an ocean of anomalies which are ignored, the articulation of the protective belt being carried along by the internal momentum of the positive heuristic of the programme. Our rational reconstruction of Lakatoss methodology, like all rational reconstructions, is a caricature.

Discussion

345

Suppose after the fourth failure in the degeneration sequence where pT has fallen to O-21 we had got a success, pT would then recover to the value O-60. Even with pT down to O-06 at the sixth failure in the degeneration sequence a success would lift pr to 0.26. At the crossover point of the pT and p,, curves (marked Q in the figure) it would become rational to change the hard core rather than the protective belt, i.e. we now have an internal rational criterion for abandoning a research programme, independent of any competing programme. At first sight we might suppose that if we had a run of successful predictions followed by a failure then the asymmetry in the effect of failure on p7 and pH which was large initially would be eliminated as p+ and pH were both now increased to values very close to unity. We shall now show that this rather natural assumption is unwarranted and that in our model the asymmetric effect of failure on pT and pH is preserved however many successful predictions precede the failure. What we shall prove is the following. After n successful predictions let pT increase from pT (0) to pT (n) and pH from p(o) to p(n). Let the effect of a refutation on pT and pH after n successful predictions be denoted by v,(n) and v(n) respectively.

Then

~ b(n )

vi(n)

= -

YlXO)
YH(O)

1-PM(O)
1 -P,(O)

a result which is independent of n. We have at once the result

v,(n)

1-p&l
1 -P,(n)

(see equation (12)).

y(n)

Now we use equation (7) to obtain the result

p&2+1)

p(n) ph)*pdn)

+ W -pdn)) PAn). + k(l -pdn)*p&)) *

346 So,

Studies in History and Philosophy of Science

Hl -P,(n))
1 -pT(n+

1)

= pr(n)*Mn) + k(l -Pr(n)*P(n))


.

Similarly we can show Ul -P(n)) 1 -p&r+ 1) = p,(n).p(n) + k(l -P&)*P&))


*

Therefore I -p&r
1 -pAn+

+ 1)
1)

I --p&r)
1 -PM

which applied recursively yields the result

1-P,(n)
1 -P,(n)

1 -PM(O) 1 -P)r(O)

= -

1-Y

in our old notation,

l-x

which is independent of n. Hence we get for refutation

after n successes an asymmetry ratio

Yr@) -=yH(n)

1-Y

l-x

In our example x = 0.9, y = O-5, so ~ v(n)

= 5

and the effect of refutation on H is five times as high as on T for all values of n, although of course it is also true that p,,(n) and p&z) both tend rapidly

Discussion

347

towards unity with increasing n. We stress that in this model the asymmetric effect of refutation on the protective belt and the hard core is directly accounted for at all stages by the initial value of pr being higher (closer to unity) than we assume for pH. It is hoped that the above analysis may serve to illustrate and in some respects to clarify Dorlings important contribution to the discussion of Duhems problem.

You might also like