Suppose that the log of the spell is normally distributed with mean lnτij and variance σj . Then,
the survival function is expressed as
Sij(t)=1-Φ
zln t - ln τ?
—-
< σ j √
The hazard rate, the rate at which the spell is completed after duration t, is
'ln t - ln τt '
' —-
φ
hij(t) =
< σ j √
tσ jSj( t )
where φ(∙) is the probability density for the standard normal distribution. Next, we assume that
the mean duration of a spell (lnτij) depends on independent variables zi (gender, marital status,
age, age squared, education, US farm experience, English speaking ability, race, ethnicity,
availability of free housing, task, region (California, Florida, and other), the year of the interview
(after 2001 or not), dummy variable for seasonal workers) so that
lnτij = zi'βj .
Then, the duration can be expressed as lntij = zi'βj + uij where uij ~ N(0,σj ). However, duration
tij is observed only if person i has legal status j. This is a typical case for selection bias. Assuming
ei and uij are bivariately normally distributed with correlation coefficient ρ, the mean of the log
of the duration conditioned on the legal status of person i is corrected as
E[ln tij | ln tij is observed] = zi'βj + ρσjλij
where λij is the correction term for the selection bias which is given as5
5 Correction term is set to zero for native-born citizen.