Beruflich Dokumente
Kultur Dokumente
Dawn L. Teele
dawn.teele@yale.edu
April 2008
Abstract
This paper is meant to be a guide to using event history modeling.
Topics include the formulation of the hazard rate and the survival function, parameterization of the model, and common problems such as right
and left censoring and ties. The appendices contain commands for setting
up event history models in Stata.
1.1
When the estimate of interest is the conditional probability of an event occurring, a linear regression specication will not su ce. If left in a linear model, the
explanatory variables will yield predictions for the coe cients that are di cult
to compute and interpret (Allison 1984), and little information will be shed on
the conditions that lead to an event occurring.
There are two statistical techniques that can be used to understand the
eects of the covariates on the response variable: The logit or log-odds of the
1.2
In the following discussion I will make reference to work that I conducted for my
senior thesis to provide a practical example of the way in which event history
analysis can be used (Teele 2006). The basic question I asked is, what causes
countries to ratify the child labor convention with the International Labor Organization? taking ratication as the event to be studied. Bear in mind
1 Formally, the logit or "log-odds" transformation is constructed by: logit_Depvar =
ln (1 DDepvar
: The transformation is necessary to have standard errors that are normally
epvar )
distributed.
2 If all of the covariates used in the regression are constant over a countrys event history,
cross-sectional analysis of the dependent variable in the year of ratication would su ce.
that I use this example here to demonstrate the steps that one should take to
implement an event history model, and am not purporting to have uncovered
un-falsiable truths in this demonstration.
The sample contains 150 countries from 1976-2004, the periods are in years
(discrete time) and the independent variables will be both time-varying and
time-invariant. To give you a sense of the data, Figure 1 shows the ratication
patterns by region. It is clear that there was a surge in ratication during the
later part of the sample (1997-2003), and so there must be a story to tell about
the conditions leading up to this time period. For this reason the event history
methodology seems, conceptually, to t the data.
1.3
Event history analysis uses the duration of time either discrete or continuous
time before an event occurred as the dependent variable (though not the
dependent variable in the conventional sense). In our example, Convention 138
was introduced to the ILO assembly in 1976, so the event history begins in this
year and ends in 2004, the last year that data for this sample are available.
However, there are 47 countries that were not members of the ILO in 1976,
meaning that they have delayed risk in the sample. Practically speaking, a
subject cannot experience an event if it is not part of the relevant sample (we
cannot understand the eect that a blood pressure drug has on increasing the
time between heart attacks if a subject has not in fact had a heart attack), and
so we must be careful when constructing the dataset to deal with these cases
1.4
(1)
Because ratication (or failure) can only occur once for each country, this
is a single spell analysis. However, there may be more than one failure at a
time if countries ratify in the same year; the Breslow method of estimation for
failure ties is accounted for in Stata.
The second component of the hazard, the survival function is dened as:
X
S(t) = P r(T tj ) =
f (tj )
(2)
j=i
The basic hazard rate for discrete-time is simply the ratio of the probability
of failure to the probability of survival, or the conditional probability of survival
given that failure has not already occurred. This relationship can be expressed
in two ways:
h(tj ) =
f (t)
S(t)
(3)
3 I have run across some literature (specically, Bockmann 2001) in which authors disregard
the observations that are not in the entire sample. This seems to be a waste, both because it
lowers the n but also because it devalues any general conclusions that may be inferred from
the results.
h(tj ) = P r(T = tj jT
tj )
(4)
Equation 4 gives the rate at which countries ratify C138 conditional on their
survival until j. The hazard can be augmented to include the eect of a vector
of time-varying explanatory variables xij :
h(tj jxij ) =
f (tj jxit )
S(tj jxit )
(5)
In this specication it is assumed that the only thing that aects the hazard over time are elements of xij , the vector of time-varying covariates, but
qualitative variables can be included in the model just as easily.
Implicit in the hazard estimate is what is known as the baseline hazard,
h0 (t), which is a parameter representing the distribution of the data with respect
to time. The idea is that the hazard is a function of h0 (t), an intercept, and a
vector of explanatory variables, xij .
The eect of the baseline h0 (t) on the hazard ratio depends on the restrictions that are placed on this parameter. Using a Cox proportional hazards
specication is a way to allow the hazard to be estimated without placing any
parametric assumptions on the baseline. Meaning that the regression remains
agnostic as to how time inuences the probability of the event.
Proportional hazards, especially the Cox model, has become increasingly
more popular in he social science literature due to its exibility.4 However, if
the distribution is known (or suspected), a functional form for h0 (t) can, and
should, be specied.
So long as the hazard is a multiplicative of the covariates, the model is classied as proportional hazards, and most parametric models can be written in this
notation. There are substantive dierences in the estimates when other distributions are assumed, and alternatives to the Cox model include the exponential,
Gompertz and Weibull distributions.
1.5
Two preliminary steps in event history analysis entail looking at the components
of the hazard function for dierent sub-groups, and testing to see whether the
dierences are statistically signicant. Given that the division between countries
in the OECD and those who are not members generally corresponds to the level
of development, I will compare these two sub-groups, but others (such as region
of the world) could just as easily be chosen.
The cumulative hazard estimates for OECD and non-OECD countries can
be seen in Figure 2. No parametric assumptions were placed on the estimates of
4 Estimating the hazard rate was rst done within a conditional logit model, and is attributed to Cox (1972). Over time, the Cox Proportional Hazard model has become
more prevalent in empirical social science. For a discussion of the benets of the proportional hazards assumption as opposed to other functional forms of the hazard model see
Box-Steensmeier and Jones (2004) and Yamaguchi (1991).
the cumulative hazard functions, and at rst glance the OECD countries appear
to have a higher hazard ratio based on the higher intercept and steeper function
toward the end of the sample.
Another common way to look at sub-groups within the event history panel
is to calculate a Kaplan-Meier estimate of the survivor function, and then graph
these results for dierent qualitative sub-groups. The Kaplan-Meier estimator
also has no parametric assumptions and is given by:
^ =
S(t)
Y nj
jjtj
rj
nj
(6)
cally decreasing over the sample. Though the intercept for OECD countries is
visibly lower than the intercept for non-OECD countries, a statistical test
the log rank test for the equality of survivor functions demonstrates that
the dierence is not statistically signicant.
Taken together, the Cumulative Hazard and Kaplan-Meier graphs illustrate
that though the hazard functions display some variation across groups, the
survivor functions move very closely to one another. Solely looking at the cumulative hazard functions (or not looking at the log-rank tests) could lead to a
misspecication of the model. Appendix A shows the output for the log-rank
test in the scenario above, and lists a few other statistical tests that can be
used to test the null hypothesis that the survivor functions for two (or more)
sub-groups are the same.
In the literature on ILO convention ratication, the Cox proportional hazards model is the predominant specication, although little rationale other that
its exibility is given (see Boockmann 2001, Chau and Kanbur 2001, and Abu
Sharkh 2002). The Cox model may have advantages when the shape of the
baseline hazard is unknown, however for this data the survivor functions are
decreasing monotonically over time, which leads to a suspicion that the risk of
ratication as represented by the hazard is increasing over time.
1.6
A parametric regression model that provides e cient coe cient estimates for
monotonic functions is the Weibull specication (Allison 1984, Box-Steensmeier
and Jones 2004). The baseline hazard function that the Weibull estimates is
7
given by:
h0 (t) = ptp
exp(a)
(7)
where p is a shape parameter that remains constant, and a is a scale parameter that is estimated by the covariates. When the hazard is monotonically
increasing with respect to time, p > 1. Given a set of covariates xij the Weibull
equation is:
h(tj jxij ) = h0 (t)exp(xit
= ptp
exp(
+ xit
x)
(8)
x)
As can be seen from Equation 8, the covariates that inuence the hazard do
so as a multiple of the baseline hazard. For this reason the Weibull still falls
into the category of a proportional hazards model, but is dierentiated from
the Cox model based on the ancillary parameters a and p. The tted Weibull
model estimates a, p and 0x (Cleves, Gould & Gutierrez 2002).
Estimation Results
Table 1 lists the coe cients and standard errors for Weibull estimates of the
ratication of Convention 138. The signs of the coe cients indicate the direction
of the eect that the variable has on the probability that a country will ratify
C138 given that it has not already. The coe cients themselves do not have
meaningful interpretations, but because the Weibull model still ts into the
family of proportional hazard models, the eect that each covariate has on the
hazard ratio can be found by taking the exponential the coe cients.5
The ve regressions in Table 1 are not nested models, but have been included
to demonstrate two things: rst, as can be seen in regressions (1) and (3), percapita GDP and its quadratic term do not have a statistically signicant eect
on the hazard ratio. This result is consistent with the results found by Chau
and Kanbur (2001), and is intuitive given that many of the richest countries
ratied C138 very late into the sample, if at all.
The eect of international trade on the propensity to ratify C138 is positive and statistically signicant in all tested variations of the model (for other
combinations of covariates see Teele (2006)). The negative coe cient on trades
quadratic term indicates that trade has a positive eect on ratication but at
a decreasing rate. This result predicts that as trade increases, the probability
of ratication increases as well. This result could be motivated by a desire to
5 For hazard ratios the null hypothesis tested is that coe cient has no eect on the hazard
such that exp(Bx ) = 1. This is equivalent to testing whether the non-exponentiated coe cients are equal to zero, as with standard linear regression. For this reason, traditional ways
of assessing signicance apply in this analysis.
0.393
(0.639)
-0.274
(0.689)
GDP squared
-0.0285
(0.0419)
0.00956
(0.0446)
Region
-0.00287
(0.0749)
-0.0262
(0.0536)
0.0494
(0.0795)
Trade, % GDP
2.208
(0.969)
2.422
(1.023)
3.138
(1.120)
3.525
(1.159)
Trade squared
-0.923
(0.477)
-0.974
(0.489)
-1.131
(0.532)
-1.255
(0.542)
-0.00494
(0.00296)
-0.00485
(0.00292)
Child Labor %
1.436
(0.797)
2.440
(1.042)
Constant
ln_p
Constant
Observations
0.128
(0.0893)
-7.427
(2.455)
-7.009
(0.744)
-5.909
(2.555)
-7.908
(0.874)
-8.793
(1.082)
0.581
(0.0979)
3646
0.570
(0.0978)
3559
0.574
(0.0978)
3479
0.616
(0.106)
2779
0.640
(0.105)
2779
signal to trade partners that the country has a similar set of values and restrictions placed on its labor standards a move that is meant to facilitate
trade negotiations.
Per-capita international aid is statistically signicant at the 10 percent level
and is found to have a negative eect on the hazard. This means that countries
that receive more international aid per person are less likely to ratify the child
labor convention. This result is contrary to sociological theories that pressure
from international aid organizations and foreign governments aects ratication
patterns for aid recipients.6 However, it should be noted that countries that have
higher per-capita aid are probably some of the poorest, and this variable may
picking up on other correlates of poverty.
The other result of interest in Table 1, found in Regression (5) is that child
labor has a positive eect on the hazard ratio, and is statistically signicant.7
This result is particularly intriguing given that C138, which is technically the
minimum age to work convention species that a minimum age of 15 be required for economic participation. The regression shows that for higher levels
of child labor specically dened as the percent of children 1014 who are
economically active countries are more likely to ratify the child labor convention.
Regression (5) includes all of the covariates that have a statistically signicant eect on the hazard. Hazard ratios have been calculated for these covariates
and can be found in Table 2. For estimates of the hazard ratio that are larger
than one, the covariate has a positive eect on the probability of ratication.
Covariates whose hazard estimates are less than one have a negative eect on
the probability of ratication. The hazards reported in Table 2 mirror the previous discussion. Both child labor and trade have very large eects on the hazard
ratio.
In the next section I will construct a graph holding child labor constant at the
saple mean and varying the level of trade in order to give a visual interpretation
of the results above.
2.1
Because there are so many steps that go into the data collection even before the
regression command can be hit, it can be tempting to let the regressions speak
for themselves. However, regression coe cients or, in our case above, hazard
rates, are not always easy to understand. The saying goes that pictures are
worth a thousand words, so for the next page or so I will show you a picture
that can help us interpret the results from above.
First, it is important to look at the data to see what a probable range for the
6 It could be argued that it is not aid per-capita but the overall reliance on international
aid that allows countries to be pressured into ratication. For this reason, international aid
as a percent of government expenditures was also tested for its eect on the hazard but was
not statistically signicant.
7 Without the inclusion of regional dummies, child labor is signicant at the 10 percent
level.
10
33.96
(3.04)
Trade squared
0.285
(-2.32)
0.995
(-1.66)
Child Labor %
11.48
(2.34)
Region
1.137
(1.44)
2779
Observations
variables of interest are. Table ?? below shows means and standard deviations
for trade and child labor, the two variables that were statistically signicant
in the section above. Figure 4 presents the tted values for Regression (5)
evaluated at the average level of child labor, 17 percent, for dierent levels of
trade. The groups with lower trade (one standard deviation below the mean)
as a percent of GDP are less likely to ratify the convention than those with
above-average trade (one standard deviation above the mean).
The slopes of the tted value curves show that for higher levels of trade, the
hazard rate is much higher. The direct interpretation of this is: holding child
labor constant, countries with higher levels of trade as a percent of GDP are
much more likely to ratify the ILO convention banning child labor.
This result supports the hypothesis that countries may sign conventions to
enhance their reputational capital within the global market place. If questions of labor standards arise in trade negotiations, countries can point to the
convention as a law that they uphold, while knowing full well that the ILO has
very little coercive power to punish their actions should they be discovered.
Conclusion
Finally we come to the end of this foray into event history modeling. The
Appendices that follow are meant to help you get started with setting up your
event history panel in Stata, and I have also included the commands that were
11
12
As a brief introduction to some of the survival time commands that are necessary
to set up an event history panel:
snapspan panelid year dcnv138, gen(date0) replace
/* Convert snapshot data to time-span data */
rename year date1
stset date1, id(panelid) failure(dcnv138) origin(ismember==1) exit(iloflake==1)
stdes /* describes st set */
stvary
note: obtain K-M survival estimate
sts generate kmS=s
label var kmS "K-M"
note: obtained N-A cumulative hazard estimate
sts generate naH=na\qquad
label var naH "N-A"
note: calculate N-A survivor estimate
g naS=exp(-naH)
label var naS "N-A"
note: calcualte K_M cumulative hazard estimate
g kmH=-log(kmS)
label var kmH "K-M"
13
|
Events
Events
OECD | observed
expected
------+------------------------0
|
91
89.33
1
|
18
19.67
------+------------------------Total |
109
109.00
chi2(1) =
Pr>chi2 =
*or
sts
sts
sts
0.18
0.6710
can use any of the following to test for the equality of survivor functions:
test OECD, wilcoxon /* Wilcoxon-Breslow Test */
test OECD, tware\qquad /* Tarone-Ware Test */
test OECD, peto /* Peto-Peto Test */
The following are the stata commands used to construct the graphs and tables
above.
For the bar chart in Figure 1:
graph bar (asis) Rat138 if Year!=1975,
over(Year, label(angle(forty_five)
labsize(small))) stack
ytitle(, size(small)) ylabel(, angle(horizontal) labsize(small))
title(Ratification of ILO Convention 138)
subtitle(by region and year)
note(Source:
International Labour Organization)
For the hazard rate estimates in Figure 2:
sts graph, hazard by(OECD)
ylabel(,angle(horizontal))
ytitle(Conditional Probability of Ratification)
xtitle(Years since C138 introduced, margin(medium))
legend(order(1 "non-OECD" 2 "OECD") size(small))
xlabel(0 "1976" 10 "1986" 20 "1996" 30 "2006", valuelabel)
title(Cumulative Hazard Estimates)
subtitle(OECD and non-OECD countries)
note(Source: International Labour Organization)
14
15
References
[1] Allison, Paul D. 1984. Event History Analysis: regression for longitudinal
event data. Newbury Park, California: Sage Publications.
[2] Artecona,
Raquel;
Martn
Rama.
A Database of Labor Market Indicators. World Bank.
2002.
[3] Beck, Thorsten; George Clarke; Alberto Gro; Phillip Keefer; Patrick
Walsh. 2001. The Database of Political Institutions. World Bank.
[4] Boockmann, Bernhard. 2003. Mixed Motives: An empirical analysis of ILO
roll-call voting. Constitutional Political Economy 14(4), December 2003,
pp.263285.
[5] . 2001. The Ratication of ILO Conventions: A hazard rate analysis.
Economics and Politics 13(3): 281309.
[6] Box-Steensmeier, Janet; Bradsford Jones. 2004. Event History Modeling:
a guide for social scientists. Cambridge, United Kingdom: Cambridge University Press.
[7] Chau, Nancy; Ravi Kanbur. 2002. The Adoption of International Labor
Standards Conventions: Who, when and why? Revised version published
in Brookings Trade Forum, 2002.
[8] Cleves, Mario; William Gould; Roberto Gutierrez. 2002. An Introduction
to Survival Analysis in Stata. College Station, Texas: Stata Press.
[9] Jaggers, Keith; Monty Marshall. 2002. Polity IV Project. University of
Maryland. www.cidcm.umd.edu/inscr/polity.
[10] Kopka, Helmut; Patrick Daly. 1999. A Guide to LATEX. Dorchester, England: Dorset Press.
[11] StataCorp. 2005. Statistical Software: Release 9.0. College Station, Texas:
Stata Corporation.
[12] Teele, Dawn. 2006. Child Labor and the Minimum Age to Work Convention.
Reed College Thesis.
16