While Fisher’s challenge has never been even approximately met, the question of whether
headaches arisede novofrom analgesic headache has been investigated and substantiates
Fisher’s claim. It has been conceded that Fisher was correct (Bahraet al.,2003; Lance
et al., 1988).
For a brief history of the ICHD, see Gladstone and Dodick (2004).
Boes and Capobianco (2005) and Ferrariet al.(2007) describe some of the tangled history
as well. It should be noted that the history is a matter of some dispute.
The rate of analgesic use isusuallydefined in terms of treatment days per month, such
that treatment occurs at least two or three days each week, with intake of the drug on at
least ten days per month for at least three months.
From Schuster (2004): “The definition does not apply to headache in women who take
medications for five or six consecutive days for menstrually associated migraines but are
treatment-free the rest of the month, acknowledged Fred D. Sheftell, MD, who partici-
pated in updating the classification. He said that if a woman took medications just four
other days of the month, she would inappropriately meet the 10-days-a-month rule. For
that reason, she must also be taking the drugs at least two to three days each week to
meet the criteria.”
Among the reasons given was the fact that “patients could become chronic due to med-
ication overuse, but this effect might be permanent. In other words, it may not be
reversible after discontinuation of medication overuse. Finally, a system whereby medica-
tion overuse headache became a default diagnosis in all patients with medication overuse
would encourage doctors all over the world to do the right thing, namely, to take patients
off medication overuse as the first step in a treatment plan.”
See, for example, Saper and Lake (2006a) for a proposal to distinguish opioid using MOH
patients from the remaining “less complicated” cases.
See Horowitz and Manski (1998) for a detailed discussion of such bounds. It should be
noted that where such bounds are used, common practice is to report both “best case”
and“worst case” bounds.
Indeed, the notion of “intent-to-treat” can be seen as part of an attempt to test a hypothe-
sisseverelyand not a notion that is an inevitable consequence of adopting “frequentist”
probability notions. See Hollis and Campbell (1999) for a discussion.
He referred to the union premium in wages between otherwise identical workers as the
wage “gap” to distinguish it from what might obtain in a world without unions.
The distinction between ATOT and other estimands is important since it isn’t particularly
meaningful to consider the effect of union status on, say, the CEO of a large multina-
tional, to take a stark example (US law, for example, prohibits this possibility). The paper,
unfortunately, takes a naive approach to characterizes the treatment heterogeneity: in
considering the variation in the effect of union status, it characterizes it by the estimated
probability of being unionized. That is, the effect of unionization is allowed to vary across
workers whose demographic characteristics put them at the same “risk” of being union-
ized. This conflates the treatment effect for workers with extremely low levels of observed
human capital (who typically have very low probabilities of being unionized) with the
treatment effects for those who can’t be unionized (bosses) or those with high levels of
education who are generally hostile to unionization. For an arguably much more sensible
characterization of the heterogeneity in treatment effects, see Card (1992), for example.
Card (ibid.) also deals with the problem of measurement error in union status, which is
ignored in the empirical example in Chib and Hamilton (2002) but has long been an
important issue in non-Bayesian analyses: see Freeman (1984), Jakubson (1991), or Card
(1992) for three examples.
This possibility wasn’t ignored, however. The problem was the lack of data. See, for
example, Abowd and Farber (1982), Freeman and Kleiner (1990, 1999), who take the
possibility quite seriously.
