d limb, even of suicides, crimes, or injuries
by particular means: every year in a town nearly the same number of
children stray from their parents and are restored by the police:
every year nearly the same number of persons post letters without
putting an address on them.

This maintenance of averages is simple matter of observation, a datum
of experience, an empirical law. Once an average for any kind of event
has been noted, we may count upon its continuance as we count upon
the continuance of any other kind of observed uniformity. Insurance
companies proceed upon such empirical laws of average in length of
life and immunity from injurious accidents by sea or land: their
prosperity is a practical proof of the correctness and completeness
of the observed facts and the soundness of their inference to the
continuance of the average.

The constancy of averages is thus a guide in practice. But in
reasoning upon them in investigations of cause, we make a further
assumption than continued uniformity. We assume that the maintenance
of the average is due to the permanence of the producing causes. We
regard the average as the result of the operation of a limited sum
of forces and conditions, incalculable as regards their particular
incidence, but always pressing into action, and thus likely to operate
a certain number of times within a limited period.

Assuming the correctness of this explanation, it would follow that
_any change in the average is due to some change in the producing
conditions_; and this derivative law is applied as a help in the
observation and explanation of social facts. Statistics are collected
and classified: averages are struck: and changes in the average are
referred to changes in the concomitant conditions.

With the help of this law, we may make a near approach to the
precision of the Method of Difference. A multitude of unknown or
unmeasured agents may be at work on a situation, but we may accept the
average as the result of their joint operation. If then a new agency
is introduced or one of the known agents is changed in degree, and
this is at once followed by a change in the average, we may with
fair probability refer the change in the result to the change in the
antecedents.

The difficulty is to find a situation where only one antecedent has
been changed before the appearance of the effect. This difficulty may
be diminished in practice by eliminating changes that we have reason
to know cou