The Long-Term Effects of Family Circumstances and Adversity on the

The Long-Term Effects of Family Circumstances and Adversity on the
Incidence of Posttraumatic Stress Disorder:
The Case of Vietnam-Era Veterans
Elizabeth Savoca
Smith College
Northampton, MA 01063
The All-UC Group in Economic History Conference
April 30-May 2, 2010
Berkeley, CA
Acknowledgements: This research was supported by the U.S. National Institute of Mental
Health (R03 MH067598-02). I thank Joan Repaal, Chris Rohlfs, and Robert Rosenheck
for their assistance in gaining access to the data. I thank participants in the Eighth
Workshop on Costs and Assessments in Psychiatry, sponsored by the International Center
of Mental Health Policy and Economics and participants in the 16th International Meeting
of the Psychometric Society for their helpful comments and suggestions on earlier drafts.
And I thank Shivani Aryal for her most able research assistance.
Introduction
There is much controversy today about whether Posttraumatic Stress Disorder
(PTSD) is over-diagnosed in U. S. veterans (McNally 2003, 2007). This issue is the focus
of intense policy debates as the current U. S. government grapples with the fiscal
challenges it faces in providing sufficient support to those veterans who need and seek
psychiatric treatment and disability compensation (U.S. Department of Veteran Affairs
2005). At the same time, a long-standing question in academic and professional circles is
why relatively few people exposed to intensely traumatic events develop the full-blown
symptoms of the disorder (Bowman and Yehuda 2004). This paper provides econometric
evidence that contributes to both debates for the answer to the first question may have
serious implications for the answer to the second question.
The analysis is based on the male veterans who participated in the National
Survey of the Vietnam Generation, a landmark study, sponsored by the U.S. Veterans
Affairs Administration in the late 1980s (Kulka et al. 1988). This is a uniquely rich crosssectional survey containing detailed information on the veteran’s current psychiatric
health as well as retrospective data on his traumatic experiences and on the family
dynamics and the structure of the household in which the veteran was raised. In addition
to the usual indicators of parents’ socioeconomic status, it paints a vivid picture of their
emotional and behavioral stability; it indicates a veteran’s birth order as well as the
number of siblings; and it contains diagnostic data designed to simulate a clinical
assessment of PTSD.
This paper contributes to two major literatures. For one, it adds to the growing
research in household economics on the effects of the family rearing environment on a
1
person’s development and achievements. Economic models of family investments of both
time and money in their children predict variations in investments not only with
household income but also with birth rank, number of siblings, parents’ educational
attainment and partnering arrangements. Most empirical work has focused on the effects
of these family characteristics on children’s educational outcomes and subsequent
success in the workplace (Booth and Kee 2009; Caceres-Delpiano 2006; Kantarevic and
Mechoulan 2006; Conley and Glauber 2006; Black et al. 2005; Ejrnaes and Portner 2004;
Behrman and Taubman 1986; Lindhert 1977; Astone and McLanahan 1991; Krein and
Beller 1988; Manski et al. 1992; Lang and Zagorsky 2001; Ginther and Pollak 2004). Far
fewer studies have considered their impact on health capital formation. The amount of
time, energy, and financial resources that parents provide to their children, though, may
also have significant consequences for the child’s emotional stability and physical health
later in life. Indeed, recent work by Argys et al. (2006) finds that birth order is
significantly correlated with risky health-related behavior among adolescents: first-born
teens are less likely to be sexually active, smoke, use drugs, and drink alcohol—
behaviors which often persist into adulthood and may have long-lasting consequences for
health and well-being.
Second, this paper considers two sources of bias in measures of association
between survey indicators of PTSD and its covariates, endogeneity bias and measurement
error bias. The endogeneity bias may arise from the possibility that combat exposure, a
necessary condition for the diagnosis of combat-related PTSD, is endogenously
determined. In recent studies of the effects of military service on long-term health,
Angrist et al. (2009), Conley and Heerwig (2009) and Dobkin and Shabani (2009),
2
recognize that self-selection into service may confound health comparisons between
civilians and veterans. This paper, which compares veterans with PTSD to those without,
allows for the possibility that the unobserved variables which affect how the veteran
reacts to combat stress may also influence the likelihood of exposure to combat.
Measurement error bias may arise from misclassifications in the veteran’s PTSD
status. A large body of research has documented substantial errors in survey
classifications of both physical and mental impairments (Bound et al. 2001; Baker et al.
2004; Thomas and Frankenberg 2000; and Savoca 1992, 1995), even among survey
instruments that are designed to simulate clinical appraisals (Helzer et al. 1985; Anthony
et al. 1985; and Kessler et al. 1998, 2005). Scholars in the social sciences have long been
mindful of the implications of classification errors for prevalence estimates and for
regression coefficients when the error-ridden variable is a regressor (Aigner 1973). Few
studies, however, have recognized the complications for estimates of measures of
association caused by classification errors in categorical outcome variables, particularly
in health-related applications.1 This paper illustrates a method of accounting for
classification errors in outcome variables, proposed by Hausman et al. (1998) and
Hausman (2001), which has broad applications to analyses of health survey data.
Posttraumatic Stress Disorder
As far back as the U.S. Civil War, military physicians documented a constellation
of symptoms—shortness of breath, chest pain, fatigue, heart palpitations, dizziness,
disturbed sleep patterns—they found difficult to ascribe to physiological causes (Hyams
et al. 1996). In 1864, Dr. Henry Hartshorne applied the diagnostic label, “muscular
exhaustion of the heart”; in 1871 Dr. Jacob Da Costa called it “irritable heart syndrome”
3
(Paul 1987). The notion that these symptoms could be precipitated by terrifying events
was first suggested by Dr. Jean-Martin Charcot in several case studies of veterans of
the1870-1871 Franco-Prussian War (Jones and Wessley 2007).
During World War I, a similar set of symptoms occurred on an epidemic scale.
Hospitals in England were deluged with servicemen who were sent back for treatment for
what was variously labeled, ‘shell shock,’ ‘effort syndrome,’ and ‘war neuroses’
(Shephard 2001). In Britain, these syndromes were the third most common reason for
disability compensation claims filed by World War I veterans (Hyams et al. 1996).
Psychiatric disability claims paid to U.S. veterans average $30,000 per claim for a total
cost of nearly $42 billion by 1940 (Shephard 2001). Research programs for the study of
the determinants of these disorders and effective treatments were given high priority by
the British War Office and fostered many collaborative efforts by American and British
medical scientists. (Shephard 2001). By the start of the Second World War, the dominant
view of the medical communities in Britain and America was that their origins were
psychological rather than physiological (Hyams et al. 1996; Paul 1987).
In World War II, the U.S. military committed to the psychiatric screening of
potential recruits. In their study of Air Force personnel, Grinker and Spiegel wrote that
“the goal of successful psychiatric selection is the weeding out of candidates who,
although capable of learning to fly, will readily succumb emotionally to the stresses of
…the dangers of combat.” They argued that military psychologists had “perfected a
battery of tests” designed to select those fit for service, tests which could be administered
1
But see Kenkel et al. (2004).
4
“with great speed.” 2 During the first year of the war, this program screened out onefourth of the white draft registrants. Nonetheless, high rates of psychiatric breakdowns
were reported, in both Europe and the Pacific theater (Shephard 2001).
In 1952, the American Psychiatric Association (APA) published its first edition of
the Diagnostic and Statistical Manual of Mental Disorders (DSM-I), its official
classification of mental diseases. Military psychiatrists played a key role in establishing a
diagnostic category for the current diagnosis of PTSD (Institute of Medicine 2008). It
was labeled, “Gross Stress Reactions,” under the category, “Transient Situational
Personality Disorders.” This nomenclature reflected the view that even the most
emotionally stable personalities could break under unusually extreme, “intolerable”
stress, in particular, stress related to combat or “civilian catastrophe.” However, once the
danger has subsided only those with pre-disposing conditions would continue to present
symptoms. For the latter “a more definitive diagnosis” would need to be established.
(APA 1952). Officially, the disorder was considered acute not chronic.
Sixteen years later, the APA published its second edition of the DSM. Although
psychiatric casualty rates during the Korean War were high (Hyams et al. 1996), DSM-II
omitted any specific category for combat-related disorders but, instead, provided, as
examples of triggers for a new disorder labeled, “adjustment reaction to adult life”, a set
of rather disparate events, including the “fear associated with military combat” along
with the “resentment… associated with an unwanted pregnancy.” (APA 1968, page 49).
Retrospective studies of the evolution of PTSD could find no official explanation for this
diagnostic shift. Various explanations, though, have emerged: the authors of DSM-II had
2
They did admit, however, that the tests had not yet been given “scientific validity.” (Grinker and
Spiegel 1945, page 10).
5
little military experience; the military had developed highly successful front-line
treatments so that the rate of psychiatric breakdowns in the first two years of the Vietnam
War (1965-1967) was so negligible that the diagnosis was hardly worth its own category;
the U.S. Veterans Administration wanted to the reduce the financial burdens on its
compensation system (Shephard 2001; Scott 1990; Wilson 1994).
In the sixties and seventies, the medical profession was generally dissatisfied with
psychiatric diagnostic methods. Many argued that they lacked reliability and scientific
validity. They used, as a case in point, the wide regional variation in rejection rates for
psychiatric disorders by draft boards during World War II, variations that could not
plausibly reflect regional differences in disease prevalence (Spiegal 2005). In the early
1970s, the APA appointed Dr. Robert Spitzer to head the task force assigned to the third
DSM revision. Spitzer’s mission was to construct an evidenced-based set of disorders. At
this time, psychiatrists were encountering Vietnam veterans with what appeared to be
delayed onset of the symptoms associated with acute combat disorders, to which they
assigned the label, “Post-Vietnam Syndrome.” They lobbied Spitzer for the opportunity
to present to members of the DSM-III taskforce systematic empirical evidence to support
a separate diagnostic category for combat-related disorders (Scott 1990). They succeeded
in their efforts; included in DSM-III was the new diagnosis of Posttraumatic Stress
Disorder. (APA 1980). The DSM defined two types: acute, where symptoms begin soon
after the trauma and end within six months, for which “the prognosis for remission is
good,” and chronic, which lasts longer than six months (APA 1980, page 237).
6
The National Survey of the Vietnam Generation
Population based surveys are essential for studying questions related to the
prevalence, causes, correlates and consequences of psychiatric disorders. This is so,
because psychiatric diseases often go unnoticed by physicians and surgeons in the general
medical sector (Saravey et al. 1991). And also because it is widely believed that a
relatively small fraction of the population at-risk seeks treatment in the specialty mental
health sector (Robins and Regier 1991). Recognizing this need, in the 1980s and 1990s,
the U.S. government sponsored many large-scale surveys of the U.S. population in order
to estimate community-based prevalence rates and treatment needs and to identify high
risk groups. The National Survey of the Vietnam Generation (NSVG), completed in the
late 1980s, was one such survey.
Because the cost of conducting clinical psychiatric evaluations of each survey
participant would be prohibitive, the surveys included structured diagnostic interviews
that could be self-administered or administered by lay interviewers with no more than
one or two weeks of training. An algorithm would then process the responses to the
interview questions to arrive at psychiatric diagnoses that followed the professional
diagnostic criteria detailed various versions of the DSM.
The principal investigators for the NSVG specified as one of its major goals the
collection of information about PTSD that would be credible to the scientific community
(Kulka et al. 1988). Consequently, the first step in the survey design tested six candidate
PTSD survey measures on clinical subjects, that is, veterans undergoing psychiatric
treatment. Diagnosis based on the Mississippi Scale for Combat-Related Stress, a 35 item
self-reported questionnaire, was found to have the highest concordance with direct
7
clinical appraisals in the clinical setting. A follow-up validation study selected a sample
of NSVG respondents for more thorough clinical assessments conducted by experienced
mental health professionals.
In Table 1 we see that the accuracy of the Mississippi scale is lower in the
community sample but still satisfies some professional standards for a valid screening
instrument (Rogan and Gladen 1978). The probability that it detects the disorder among
true cases (77.3%)3 is far greater than the fraction that it detects among true noncases
(17.2%). Its accuracy is comparable to the screening instrument in current use by the U.
S. Department of Defense to assess PTSD among veterans returning from deployment to
Iraq (Tanielian and Jaycox 2008).4 Moreover, its accuracy exceeds that of other widely
used survey screening instruments for psychiatric disorders in the general population
(Anthony et al. 1985; Helzer et al. 1985; Kessler et al. 1998). For example, validation
studies of the diagnostic screening instruments in the U.S. National Comorbidity Survey,
the first large-scale epidemiological survey of the mental health of the general U.S.
population, found K values in excess of 0.60 only for two disorders, (agoraphobia and
social phobia); K values in the range of 0.50 to 0.60 for only four disorders (major
depressive episode, mania, simple phobia, and alcoholism). The remainder fell below
0.50, in particular, PTSD had a K value of only 0.39 (Kessler et al. 1998).
3
The proportion of true cases receiving a positive survey diagnosis equals one minus the false
negative rate. (See Table 1).
4
The U. S. Department of the Defense assesses PTSD with a screening instrument developed for
use in the primary care setting. A test of its validity among veterans under patient care found a K
value of 0.61 (Prins et al. 2004). The K value or Kappa score compares the actual agreement
between the two diagnoses to the chance agreement we would observe if the diagnoses were
completely independent. In Table 1, for the post-survey study, the 95% confidence interval in
parentheses beneath the Kappa score (author’s calculations) falls far above zero, strongly
suggesting that the agreement between the two screening methods is more than what would occur
8
The 35 items in the Mississippi Scale were chosen to cover the diagnostic criteria
established in the third (revised) version of the DSM (APA 1987). In that version, a
necessary condition for a positive diagnosis is the experience of at least one traumatic
event “that is outside the range of usual human experience and would be markedly
distressing to almost anyone” (APA 1987, pg. 247).5 Once it is established that such an
event has occurred, diagnosis is based on the intensity of the subject’s reactions to the
event. Symptoms include flashbacks, excessive avoidance of situations or people that
might lead the subject to recall the event, and physiological reactions to events that
remind the subject of the traumatic event (increased heart rate and sweating, e.g.).
(Appendix A lists the full DSM-III-R diagnostic criteria). The subject rates each of the 35
items on a 5-point Likert scale according to symptom severity. The Mississippi Scale is
the total of the 35 item scores, thus ranging from 35 to 175. The NSVG study team chose
a cutoff of 89 or higher for its PTSD diagnosis. These criteria resulted in a 31% lifetime
prevalence rate; 15% were assessed as currently meeting full diagnostic criteria (Kulka et
al. 1990).
Many scholars regarded the NSVG prevalence rates for PTSD as implausibly
high. Critics pointed to the fact that only about 15% of Vietnam veterans had combatrelated occupational specialties; others questioned the reliability of the veteran’s
recollection of war-related traumatic events, events which, for the NSVG participants,
occurred well over a decade prior to the interview (McNally 2003). Still others wondered
whether the diagnosis was a result of help-seeking and compensation-seeking behavior
by chance. The typical rule of thumb is that a score in excess of .75 implies excellent agreement;
between .4 and .75, fair to good (Kulka et al. 1988).
5
The subject need not be the direct victim of the event. For example, a soldier who witnesses
wartime atrocities against civilians and enemy soldiers meets the DSM-III-R exposure criteria.
9
(Frueh 2005). Over the past decade dozens of articles have debated this issue; McNally
(2007) summarizes the current view in psychiatry:
Outright malingering [is]…unlikely to account for many of the PTSD-respondents in the
[NSVG]…One factor likely contributing to the prevalence of PTSD is the number of individuals
who did not have a combat MOS [military occupational specialty], but who were nevertheless
exposed to horrific, life-threatening events (e.g., medics, ambushed truck drivers)…Finally, some
individuals experiencing psychological, interpersonal, and occupational problems might have
attributed their difficulties to war-related events despite the true causes lying elsewhere. The
Vietnam-PTSD narrative would [have] been one way for some veterans to make sense of their
lives since the war.” (page 198).
Empirical Analysis of Patterns in Stress Reactions to Traumatic Events
I begin with an analysis of the determinants of a veteran’s overall score on the
Mississippi Scale for Combat-related Stress, a continuous measure of PTSD, which can
be interpreted as representing a veteran’s propensity to develop a diagnosis of PTSD.
This approach provides a convenient framework for testing the exogeneity of trauma
exposure. The true propensity, Y*, is a function of observed exogenous pre-disposing
conditions, X, as well as trauma exposure, Trauma:
(1) Y
*
= β X + γTrauma + ε
Table 2 reports summary statistics on the covariates selected to represent the
underlying source of variation in the veterans’ responses to trauma exposure. Many
capture aspects of the family structure and the family rearing environment during
childhood, which have been shown, in the general population, to influence success and
both physical and emotional well-being later in life
In their comprehensive review of the empirical research on the determinants of
children’s attainment, Haveman and Wolfe (1995) note four parental choice variables that
have been studied extensively by economists: parents’ education, marital stability,
10
income, and mother’s labor force participation decisions. These variables reflect multiple
aspects of the family rearing environment: the time and economic resources available for
investments in children, the quality of these investments, the cultural and intellectual
environment of the home, and the genetic endowment that the parents pass on to their
offspring. In Table 2 we see that 34% percent of veterans answered “poor” to the
question, “Was your family well-to-do, average, or poor?” The fraction was higher
among positive PTSD cases than among negative cases. Positive cases were also more
likely to have parents who did not graduate from high school, especially fathers. Seventyseven percent of veterans grew up in an ‘intact’ family, that is, they lived with both
biological parents until age sixteen. That fraction is lower for positive PTSD cases while
the fraction who reported that their mothers worked outside the home during their
childhood was higher.
Models with a psychosocial orientation suggest that potentially disruptive
episodic events can also permanently alter a child’s development regardless of the
presence of chronic childhood adversity (Haveman and Wolfe 1995; Haas 2008). Table 2
shows that veterans suffering from PTSD were more likely to report that their fathers
experienced at least one episode of unemployment and that they switched schools at least
once because of a family location move.
In recent decades much has been written on the relationship between fertility
decisions and children’s outcomes. The initial work focused on the effect of family size
on children’s attainment, addressing Becker’s “quantity-quality tradeoff’ hypothesis
(Becker and Lewis 1973; Becker and Tomes 1976). Later work explored the effects of
birth order. The theoretical underpinnings for the relationship between educational
11
attainment and birth order suggest that the relationship is not monotonic. The time spent
by parents with a child is an important factor in increasing his capacities. An only child
or a first-born child, during the critical early years, presumably faces less competition for
his parents’ time and resources than do later born children. While last-born children,
particularly in large families, may be at a disadvantage in their early years, they have a
‘monopoly’ on their parents’ time during their teenage years as their siblings leave home.
They also gain the attentions of older siblings throughout their childhood.6 Indeed,
Lindert (1977) finds that the total amount of time devoted to the care of a child until age
eighteen by any person in the household is highest for the first- and the last-born in
families of three or more children. Keller and Zach (2002) find that first-born infants
receive significantly more attention from their parents, particularly from their fathers than
do later-born.
Much has been written on the empirical challenges of disentangling family size
6
More nuanced arguments predict that the relationship between birth order and parental time and
money inputs depends on the parents’ preferences. Parents who are “nondiscriminatory” will
allocate their resources equally among their offspring. Any observed birth order differences in
achievements, then, will reflect relationships between birth order and innate talents. There is a
widespread belief that first-born have richer genetic endowments. Parents who are “achievement
maximizers” will allocate their resources disproportionately to the child who will provide the
highest return on their investment. They are strategic not only in how they allocate their resources
but also in when they decide to stop having children. Each child is a random draw from the gene
pool; parents stop having children when they have the one who exceeds their expectations. These
models predict that the last born will have the greatest innate abilities, will receive the most
attention from their families, and, consequently, will be the most successful later in life. Parents
who are adverse to inequalities in their children’s well-being and success will devote the greatest
resources to the least able child. This “overcompensation” should minimize correlation between
birth order and success. See Ejrnaes and Portner (2004), Behrman and Taubman (1986), and
Hanushek (1992) for formal expositions of these models. In other disciplines the theories relating
birth order to intelligence rely on biological factors (the high levels of maternal antibodies in
higher order births may cause fetal brain damage) or emphasize the how the intellectual
environment of the family varies with family structure (Kristensen and Bjerkedal 2007). The
relationship between birth order and personality traits is still hotly debated (The Independent
2007).
12
effects from birth-order effects.7 The potential endogeneity of family size to unobservable
family factors that might also affect a person’s observed development may contaminate
the estimate effects of other control variables that are highly correlated with family size,
such as birth order. To mitigate this problem I use an index of relative birth order,
proposed by Ejrnaes and Portner (2004), which equals (p-1)/(n-1), where p is the
respondent’s birth order and n is the number of children in the family into which he was
born. It is easy to interpret; it ranges from 0 for first-born to 1 for last-born. And it is
virtually free of correlation with number of siblings. As we see in the bivariate
comparisons in Table 2, there is little difference in either the average relative birth order
or the average number of siblings between positive and negative PTSD cases.
Many scholars have argued that the longstanding empirical finding that parental
divorce adversely affects the development and achievements of their offspring may
reflect the effects of parental characteristics that are correlated with marital discord but
are often left out of the analysis (Manski et al. 1992; Cherlin 1999; Lang and Zagorsky
2001). The NSVG contains detailed indicators of the family rearing environment which
allow us to control for other potentially disruptive parental behavior. These variables are
based on the respondent’s recollection and assessment of whether either of his parents
had mental or nervous conditions, had a serious drinking or drug problem, or had been
arrested or jailed during the respondent’s childhood. Veterans who received positive
PTSD diagnoses reported a much higher incidence of substance abuse, mental health
problems and potentially criminal behavior among their parents.
7
For recent discussions, see Black et al. (2005), Conley and Glauber (2006), and Booth and Lee
(2009).
13
In addition to the self-administered questionnaire, which forms the basis of the
outcome variable of this study, the NSVG gave the respondents the opportunity to discuss
at length with the interviewer up to thirty traumatic events in their lives, both war-related
and not war-related, and the degree of stress each event may have caused. Trained coders
then independently reviewed the survey responses and rated each event according to how
closely it corresponded to the DSM definition of trauma: “unlikely to be traumatic,”
“probably traumatic,” “definitely traumatic,” and “severely traumatic.” These variables
serve as indicators of the events which triggered the stress reactions. Studies have shown
that the likelihood of developing PTSD varies by type of stressor (Kessler et al. 1999;
Ilkin et al. 2005). Therefore, two trauma indicators are included in the set of regressors,
one war-related and one not related to war. Each takes on a value of one only if the
NSVG study team coded the event as definitely meeting clinically defined trauma.
OLS Estimation
Table 3 reports OLS estimates of equation (1). Among the standard set of
characteristics, parents’ education has no significant relationship, either practically or
statistically, to a veteran’s propensity to development symptoms of the disorder, nor does
the veteran’s report of family income. This result stands in sharp contrast to studies of
parental education and a child’s intellectual development but is somewhat consistent with
the findings of Argys et al. (2006). They report very weak relationships between parents’
education and income (broadly defined) and alcohol and drug use among adolescents;
most estimates are statistically insignificant. These adolescent behaviors are strong
predictors of adult substance use problems (Mullahy and Sindelar 1989), which, in turn,
occur at dramatically higher rates among veterans classified as positive for PTSD (Kulka
14
et al. 1990). Marital stability and the mother’s work decision do matter. Veterans reared
in what were then regarded as traditional families, those whose parents stayed together
and whose mothers did not work outside the home, are predicted to score 8 points less on
the Mississippi Scale. In ‘non-intact’ families, the mother’s work decision has no
statistically significant effect on the score, while the benefits of marital stability disappear
if the mother works outside the home (95% confidence interval: (-3.89, 3.00)). This
finding is consistent with Hayward and Gorman’s study of the effect of childhood
circumstances on adult male mortality using the U.S. National Longitudinal Survey of
Older Men, 1966-1990 (2004); middle-age men raised in traditional households had the
lowest death rate.
Veterans reared in families with at least one parent who has a history of mental
illness exhibit much higher scores on the Mississippi Scale. This result is consistent with
the findings of DeWit et al. (2005) in their study of social phobia in a population sample
and with Kessler et al. (1999) in their study of the risk factors for PTSD in the general
population. These findings may reflect genetic factors in the transmission of
psychopathology. They may also reflect learned behavior and the parents’ ability to
adequately develop a nurturing, supportive relationship with the child.
The advantages of birth order accrue in equal measure to the first and last-born.
The middle child scores roughly 5 points higher than the first-born and the last-born
(95% confidence intervals: (2.48, 8.60) and (1.99, 8.69), respectively).8 This finding is
consistent with some of the literature on birth order and educational achievement.
Hanushek (1992) finds that in large families the first- and last-born perform better in
8
Recall that the index of relative birth order is set to zero for first-borns, one for last-borns, and
one-half for the middle child.
15
elementary school on reading comprehension and vocabulary tests than do middle-born
children. The last-born, however, has a distinct advantage over the first-born. Lindert
(1977) finds the same pattern in his study of the relationship between birth order and
educational attainment. Kanterevic and Mechoulan (2006), though, find that the
estimated educational advantage of the last-born disappears when they control for
mother’s age at birth. Black et al. (2005) and Booth and Kee (2009) also find a declining
monotonic effect, for Norway and Britain, respectively, with the first-born achieving the
highest level of schooling. Argys et al. (2006) find that later-born children are much more
prone to risky health-related behavior; they do not explore nonlinearities in this
relationship. Number of siblings, however, was found to have no relationship to the
PTSD score. Younger veterans, non-white veterans and veterans whose fathers
experienced disruptions in employment scored higher on the Mississippi scale.
Instrumental Variables Estimation
It is reasonable to assume that the Mississippi Scale (Y) measures the true illness
propensity (Y*) with error: Y = Y
*
+ ν . Then the estimated regression model becomes:
(2) Y = β X + γTrauma + ε + ν
The OLS coefficients are consistent estimates of
and as long as the regressors are
uncorrelated with the regression error, ε + . If the measurement error, , were purely
random, then OLS estimates would remain consistent. However, many have speculated
that there may be a systematic bias in self-reports of the symptoms arising from trauma
exposure, so that COV(Trauma, )
0. Furthermore, True and Lyons (1999) and others
have argued that genetically determined personality traits and the family rearing
environment may influence a person’s taste or tolerance for risk. Individuals with higher
16
tolerance may enter hazardous occupations, such as fire-fighting and demolition, which
put them at greater risk of exposure to highly stressful situations. In military life, such
people may enter branches of the armed forces which are more likely to engage in
combat or they may be more likely to volunteer for dangerous combat missions. If there
is some correlation between the unobservable factors affecting reactions to stress with the
occurrence of traumatic stress, then the COV(Trauma, ε)
0 as well, thus introducing
endogeneity bias into the OLS coefficient.
One solution is to use an instrumental variables strategy. Following Rohlfs (2007)
and Dobkin and Shabani (2009), as a measure of exogenous trauma risk, I use the number
of combat casualties per 1000 males in the veteran’s birth year. (See Appendix B for
details). Table 4 reports first stage reduced-form estimates relating the occurrence of any
trauma to the exogenous risk factors. The first column lists the probit estimates of the
probability of trauma exposure; the second the linear probability estimates. Veterans from
cohorts that had high rates of combat deaths were much more likely to have been exposed
to trauma themselves.
We also see that the probability of exposure to intensely traumatic events is much
higher for veterans who were raised in families in which at least one parent had a history
of drug and alcohol problems and/or mental illness. The effects of parental
psychopathology on trauma exposure could reflect both genetic factors and the impact of
early childhood environment. The data are not sufficiently rich to differentiate between
the two explanations or to identify the underlying causal effects. However, it is worth
noting that in their study of genetic risk factors for PTSD using the Vietnam Era Twin
Registry, True and Lyon (1999) find that aspects of their shared family rearing
17
environment contributed nothing toward explaining differences in combat exposure while
genetic factors could account for 47% of the variance in exposure.
Persons raised in disadvantaged environments may have limited educational
opportunities, thus, denying them access to positions that are insulated from physical risk.
Individuals who grow up in families with parents who have serious mental and
behavioral problems may be at greater risk of being victims of abuse or of witnessing
physical harm to close relations. Bromet et al. (1998) found that the men and women in
the U. S. National Comorbidity Survey, who reported incidents of neglect or abuse during
their childhood, were more likely to report a family history of mental disorders and
substance abuse and were less likely to have been raised in an intact family.
Table 5 reports the second stage estimates relating the PTSD Mississippi Score to
the instrument for trauma and the other demographic and family background
characteristics. Column (1) uses the instrument for the probability of trauma exposure
constructed from the first stage probit estimates in Table 4; Column (2) uses the
instrument constructed from the linear probability model of Table 4. The third column in
Table 5 also reports the OLS estimates, which assume exogeneity. The instrumental
variable estimates of the effect of Trauma are somewhat larger than the OLS estimates,
implying that COV(Trauma, + ) < 0. This result is consistent with many hypotheses.
The veterans who were exposed to intensely traumatic events may be more resilient,
COV(Trauma, ) < 0, or they may be more reluctant to acknowledge the psychological
toll of the experience, COV(Trauma, )< 0. That COV(Trauma, ) may be negative also
implies that veterans who were not exposed to high levels of combat stress may have
exaggerated their symptoms. However, a Hausman test finds that the difference between
18
the OLS and two-stage least squares estimates are statistically insignificant (p-value =
0.80).
Empirical Analysis of Risk Factors for Diagnosed Cases of PTSD
Here I consider whether family background can help distinguish between veterans
who reach the threshold level of symptoms for a positive diagnosis from veterans who do
not. The outcome variable is a binary variable indicating the presence or absence of the
diagnosis. Although the instrumental variables estimation and the Hausman test suggest
that trauma exposure may be treated as an exogenous risk factor, even purely random
misclassification in PTSD status, uncorrelated with any of the exogenous variables, can
bias regression coefficients.
The population model specifies that the probability of a positive diagnosis is a
function of a set of observed exogenous background characteristics and events that are
assumed to be measured without error:
(3) P ( y = 1 | x ) = F ( xβ )
where y represents the true disease status; it takes on a value of 0 if the diagnosis is
negative, 1 if positive. The vector, x = ( x1 x 2 ,
and β = ( β1 , β 2 ,
x k ) , consists of k exogenous covariates
β k ) is a k -dimensional parameter vector.
The variable Y is the survey disease classification with misclassification
parameters: r0 = P (Y = 1 | y = 0) and r1 = P (Y = 0 | y = 1) . The standard
misclassification model assumes that the misclassification mechanism is nondifferential,
that is, given the true diagnosis, no other characteristic (z) of the respondent provides
additional information about the observed disease classification: P (Y | y, z ) = P (Y | y ) .
19
The relationship between Y and y can be expressed as:
P (Y = 1 | x) = r0 + (1 − r0 − r1 ) P ( y = 1 | x). 9 This structure allows us to specify an
estimable relationship between the observed disease classification and the observed
covariates:
(4)
P (Y = 1 | x) = r0 + (1 − r0 − r1 )F (xβ )
In equation (4) it becomes apparent that the marginal effect of a covariate on the
probability of a positive survey diagnosis equals its effect on the probability of truly
having the disease only if the survey classification is error free.
The standard estimation strategy for equation (4) assumes either the logit or probit
specification for F and maximizes the likelihood function or minimizes the regression
sum of squares over β , setting the values of r0 and r1 to estimates derived from a
validation study (Savoca 2004; Poterba and Summers 1995). As long as the estimated
error rates are consistent estimates of the error rates in the study population, then the
estimates of β will be consistent as well.
There are many reasons, however, why one might suspect that the error rates from
the validity study contain bias. For one, selection into the validation sample was not
purely random, but was based, in part, on the survey diagnosis (Kulka et al. 1988). The
validity study oversampled veterans who screened positive on the survey instrument.
Even if selection into the validation sample were entirely random, the assessment of
validity calls for an independent assessment under identical conditions (Bound et al.
9
P(Y = 1| x) = P(Y = 1, y = 1| x) + P(Y = 1, y = 0| x) = P(Y = 1| y = 1, x)P(y = 1| x) + P(Y = 1| y =
0, x)P(y = 0| x) = P(Y = 1|y = 1)P(y = 1| x) + P(Y = 1|y = 0)P(y = 0| x) = (1 – r1)P(y = 1| x) +
r0(1 – P(y = 1| x) = r0 + (1 – r0 – r1)P(y = 1 |x).
20
2001). In this validity study, independent assessments were unlikely to have occurred
since the respondents selected into the validity sample were probably affected by the first
assessment. Some respondents may have deliberately attempted to be consistent across
the two interviews. Others may have reflected on their responses to the initial interview
and used the re-interview as an opportunity to correct recall errors or to cast themselves
in a different light. Furthermore, clinical conditions may have changed between the initial
survey interview and the clinical re-interview.
Instead of the standard approach, I adopt an alternative strategy proposed by
Hausman et al. 1998. Here the error rates are estimated simultaneously with the
parameter vector, β . The regression sum of squares for equation (4) is minimized over
β , r , and r1 .10
0
Results
I estimate two versions of equation (4) via nonlinear least squares; both assume
the probit specification for F as well as for the misclassification parameters. Model 1
makes the standard assumption that the error rates are constant across the respondents.
10
See Savoca (2010) for a detailed discussion of the conditions for identification.
21
Model 2 allows the error rates to vary with race and age.11
Table 6 reports estimates of the misclassification parameters. The top panel
reports the parameter estimates. A series of F-tests established that the false positive rates
differ by age but not race, while the false negative rates are unrelated to race and age.
Model 3 presents estimates that impose this latter constraint. The bottom panel uses the
parameter estimates to predict the errors rates for selected cohorts.
Overall, the estimated false positive rates are relatively low. This is consistent
with studies of classification errors in other widely used psychiatric screening
instruments in general population surveys. Estimated false positive rates for the diseases
covered by the Composite International Diagnostic Interview (CIDI), used in the U.S.
National Comorbidity Survey, range from a low of .8% for mania to a high of only 7.8%
for major depression (Savoca 2004, 2005). Analysis of the Diagnostic Interview Schedule
(DIS), the pre-curser to the CIDI, shows, with the exception of alcohol use disorders,
false positive rates no higher than 4% (Savoca 2004). The estimates in Table 6 are also
comparable to error rates found in self-reports of physical diseases. In a large scale study
of the general Canadian population, Baker et al. (2004) compare self assessments to
11
There is a vast empirical literature in cognitive and social psychology on the circumstances
under which survey responses are most likely to contain errors (Bound et al. 2001). Much of it
focuses on the respondent'
s cognitive processes: his comprehension of the survey questions, his
retrieval of information from memory, and his communication skills. Many empirical studies
have also looked at the relationship between response errors and the length of the recall period.
Both literatures suggest that the time since discharge from the military and the education level of
the respondent might also be a source of variation in the error rates. Both variables, however, are
potentially endogenous. A veteran’s post-military educational advancement may be directly
impeded by the readjustment difficulties caused by PTSD or may be influenced by unobservable
characteristics that also make him more susceptible to PTSD. Studies have shown that soldiers
suffering mental disorders diagnosed during active duty have much higher service attrition rates
than other soldiers (Hoge et al. 2005). Since identification of models which depart from the
nondifferential error assumption has been established only for the case where the source of
variation is exogenous (Lewbel 2000), I restricted the specification of Model 2 to exogenous
characteristics of the respondent.
22
administrative records. With very few exceptions, false positive rates for a wide range of
chronic illnesses fall well within a range of 0% to 18%.
For all three models, we also see that the estimated false positive rates are lower
than the estimated false negative rates. This, too, is a pattern that shows up in all major
survey measures of physical and psychiatric diseases. The estimated false negative rates
reported here, however, are somewhat lower than the false negative rates for other
commonly used survey indicators of health in the general population. Savoca (2004,
2005) finds false negative rates ranging from 32.8% for agoraphobia to 79% for manic
disorders. Baker et al. (2004) find that an overwhelming majority of self-reports of
physical ailments have false negative rates well above 30%. Many, such as cancer and
bronchitis, exceed 50%. Nonwhites are estimated to have higher false positive rates and
lower false negative rates than those of whites, although the differences are not
statistically significant. Both rates decline with age; this pattern in the false positive rates
is statistically significant.
Table 7 reports estimates of the effects of the control variables on true disease
status, that is, it reports estimates of the parameters in F(x ) in equation (4). Overall, the
coefficient estimates across models generally agree in sign and statistical significance.
However, the magnitudes of the coefficients in the error-adjusted regressions are
substantially larger than the estimates from the naïve model, which does not account for
classification errors. These differences imply substantial differences in the estimated
effects of key variables on the true versus the survey probability of a positive PTSD
diagnosis. For example, with the other regressors set to sample means, both the naïve
model and the error-corrected model predict that combat exposure versus no exposure
23
will raise the probability of a positive PTSD diagnosis; but the naïve model predicts an
increase of 33.5 percentage points while the error-corrected models predict increases
slightly higher than 50 percentage points. Similarly, all models predict that veterans
reared in families with parental histories of mental illness are much more likely to
develop PTSD. The predicted increases in the error-corrected models fall in the range of
35 to 38 percentage points; the naïve model predicts a 20 percentage point rise in the
probability of a positive diagnosis.
To assess the implications of misclassification errors on prevalence estimates, we
can compare the sum of the predicted probabilities of a positive diagnosis, F(xβ ) , using
the error corrected coefficient estimates from Model 3, to the sample prevalence estimate,
Y / n . The sample prevalence overstates true prevalence by 6 percentage points (22%
versus 16%), or 27%.12 Proportionately, this closely matches the results of Dohrenwend
et al. (2006) in their study of a subsample of the veterans who participated in the NSVG;
their prevalence estimate drops from 12% to 9% (25%).
Conclusions
Recent empirical studies of the accuracy of PTSD diagnoses in the National
Survey of the Vietnam Generation (NSVG), an influential study of veterans from the
12
Given that the false negative rates are higher than the false positive rates, a reader might be
puzzled by these results. Equation (4) implies that the survey prevalence exceeds the ‘true’
prevalence, i.e., P(Y = 1) > P(y = 1) whenever r0 (1 - P(y = 1)) > r1P(y = 1). Therefore, if the
disease-free population is relatively large, a small false positive rate can still imply that a large
number of respondents were incorrectly classified as positive cases. Conversely, a large false
negative rate applied to relatively small number of true positives is consistent with having a small
number of positive cases misclassified as negative cases.
24
Vietnam-Era, suggest that its PTSD prevalence estimates may somewhat overstate true
prevalence. These findings are based largely on re-examinations of the professional
criteria used to arrive at the diagnosis and on efforts to independently verify the veterans’
accounts of traumatic war-time events. This study provides new statistical evidence on
the extent of misclassification in the NSVG indicators of combat-related PTSD. The
approach here is indirect. Guided by economic and psychosocial models of the childhood
antecedents of adult success and well-being, I use variables that reflect the veteran’s
child-rearing environment to help predict which are likely to develop PTSD and then
allow the data to re-classify veterans to achieve the best fit.
My findings are consistent with the literature. When adjusted for errors in
diagnoses, my sample PTSD prevalence estimate falls from 22% to 16%, a drop of 6.0
percentage points, or 27%. Proportionately, this closely matches the results of
contemporary analyses of the same data. Furthermore, the false-positive and falsenegative rates implied by the data follow a pattern found in virtually all validation studies
of epidemiological surveys of population health -- relatively high false-negative rates
and relatively low false positive rates.
There is little evidence that the measurement error in the PTSD screening
instrument is endogenous to the veteran’s combat experience. A Hausman test supports
the null hypothesis of the exogeneity of trauma exposure. Veteran’s exposed to high
levels of combat stress are no more or less likely to misreport their symptoms.
Classification errors, however, do vary with exogenous characteristics; old veterans have
lower false positive rates.
These classification errors have serious consequences not only for PTSD
25
prevalence estimates but also for inferences about measures of association. Failure to
correct for misclassification in PTSD may dramatically understate the effects of risk
factors, both war-related and not war-related. Veterans who were reared in families with
parental histories of mental illness are more likely to be exposed to trauma and to have
reactions severe enough to lead to a positive diagnosis. Similarly, growing up in what
was then (1950s and earlier) considered a traditional household, where both parents are
present and where the mother does not work outside the home, substantially reduces
vulnerability to the disorder. These findings corroborate numerous studies in psychiatric
epidemiology. What this paper shows, however, is that the magnitude of their effects are
substantially larger when misclassification bias is taken into account. Hence,
conventional risk factors have a greater ability to differentiate true cases from noncases
than is widely believed.
26
Appendix A.
Diagnostic Criteria for Posttraumatic Stress Disorder
(Source: APA, 1987, page 248.)
A. The person has experienced an event that is outside the range of usual human experience
and that would be markedly distressing to almost anyone, e.g., serious threat to one’s life
or physical integrity; serious threat or harm to one’s children, spouse, or other close
relatives and friends; sudden destruction of one’s home or community; or seeing another
person who has recently been, or is being, seriously injured or killed as a result of an
accident or physical violence.
B. The traumatic event is persistently reexperienced in at least one of the following ways:
(1) recurrent and intrusive distressing recollections of the event (in young children,
repetitive play in which themes or aspects of the trauma are expressed)
(2) recurrent distressing dreams of the event
(3) sudden acting or feeling as if the traumatic event were recurring (includes a sense
of reliving the experience, illusions, hallucinations, and dissociative [flashback]
episodes, even those that occur upon awakening or when intoxicated)
(4) intense psychological distress at exposure to events that symbolize or resemble
an aspect of the traumatic event, including anniversaries of the trauma
C. Persistent avoidance of stimuli associated with the trauma or numbing of general
responsiveness (not present before the trauma), as indicated by at least three of the
following:
(1)
(2)
(3)
(4)
efforts to avoid thoughts or feelings associated with the trauma
efforts to avoid activities or situations that arouse recollections of the trauma
inability to recall an important aspect of the trauma (psychogenic amnesia)
markedly diminished interest in significant activities (in young children, loss of
recently acquired developmental skills such as toilet training or language skills)
(5) feelings of detachment or estrangement from others
(6) restricted range of affect, e.g., unable to have loving feelings
(7) sense of a foreshortened future, e.g., does not expect to have a career, marriage,
or children, or a long life
D. Persistent symptoms of increase arousal (not present before the trauma), as indicated by
at least two of the following:
(1)
(2)
(3)
(4)
(5)
(6)
difficulty falling asleep or staying awake
irritability or outbursts of anger
difficulty concentrating
hypervigilance
exaggerated startle response
physiologic reactivity upon exposure to events that symbolize or resemble an
aspect of the traumatic event (e.g., a woman who was raped in an elevator breaks
out in a sweat when entering any elevator)
27
Appendix B. Construction of the Instrument for Combat Risk
Total male casualties by year were obtained from the Defense Casualty Analysis System File
from the National Archives online database: http://aad.archives/gov/aad.
Total births by year were derived from Historical Statistics, Statistical Abstract of the United
States, 2003, Table No. HS-13. Live Births, Deaths, Infant Deaths, and Maternal Deaths, 19002001.
Yearly male-female sex ratios were derived from Birth, Stillbirth, and Infant Mortality Statistics
for the Continental United States, the Territory of Hawaii, and the Virgin Island. U.S.
Department of Commerce, 1936 and from “Trend Analysis of the Sex Ratio in the United States,”
Mathews and Hamilton (2005), Table A.
28
References
Aigner, D. 1973. “Regression With a Binary Independent Variable Subject to Errors of
Observation.” Journal of Econometrics, Vol. 1, pages 49-59.
American Psychiatric Association. 1952. Diagnostic and Statistical Manual of Mental
Disorders, first edition. Washington, D. C.: American Psychiatric Association.
American Psychiatric Association. 1968. Diagnostic and Statistical Manual of Mental
Disorders, second edition. Washington, D. C.: American Psychiatric Association.
American Psychiatric Association. 1980. Diagnostic and Statistical Manual of Mental
Disorders, third edition. Washington, D. C.: American Psychiatric Association
American Psychiatric Association. 1987. Diagnostic and Statistical Manual of Mental
Disorders, third edition, revised. Washington, D. C.: American Psychiatric Association.
Angrist, J. D., Chen, S. H., and B. R. Frandsen. 2009. “Did Vietnam Veterans Get Sicker
in the 1990s? The Complicated Effects of Military Service on Self-Reported
Health” National Bureau of Economic Research, Working Paper No. 14781.
Anthony, J. C., Folstein, M. Romanoski, A. J., Von Korff, M. R., Nestadt G. R., Chahal,
R., Merchant, A., Brown, H., Shapiro, S., Kramer, M., Gruenberg, E. M.. 1985.
"Comparisons of the DIS and a Standardize Psychiatric Diagnosis." Archives of
General Psychiatry, Vol. 42, pages 667-675.
Argys, L. M., Rees, D. I., Averett, S. L., and B. Witoonchart. 2006. “Birth Order and
Risky Adolescent Behavior.” Economic Inquiry, Vol. 44, No. 2, pages 215-233.
Astone, N. M. and S. McLanahan. 1991. "Family Structure, Parental Practices, and High
School Completion." American Sociological Review, Vol. 56, pp. 309-320.
Baker, M., Stabile, M. and Deri, C. 2004. “What Do Self-Reported, Objective, Measures
29
of Health Measure?” The Journal of Human Resources, Vol. 39, No. 4, pages 10671093.
Becker, G. and G. Lewis. 1973. “On the Interaction between the Quantity and Quality of
Children.” Journal of Political Economy, Vol. 81. No. 2, Part 2, pages S279-S288.
Becker, G. and N. Tomes. 1976. “Child Endowments, and the Quantity and Quality of
Children.” Journal of Political Economy, Vol. 84, No. 4, pages S143-S162.
Behrman, J. R., and P. Taubman. 1986. "Birth Order, Schooling, and Earnings." Journal
of Labor Economics, Vol. 4, No. 3, Part 2, pp. S121-S145.
Black, S. E., Devereux, P. G., and K. G. Salvanes. 2005. "The More the Merrier? The
Effect of Family Size and Birth Order on Children'
s Education." Quarterly Journal of
Economics, Vol. CXX, No. 2, pp. 669-700.
Booth, A. L. and H. J. Kee. 2009. “Birth Order Matters: The Effect of Family Size and
Birth Order on Educational Attainment.” Journal of Population Economics, Vol. 22,
No. 2, pages 367-397.
Bound, J., Brown, C., and N. Mathiowetz. 2001. “Measurement Error in Survey Data.” In
Heckman, J., J., and E., Leamer, eds, Handbook of Econometrics, Vol., 5.
Amsterdam: Elsevier Science, B. V.
Bowman, M. L. and R. Yehuda. 2004. “Risk Factors and the Adversity-Stress Model.” In
Rosen, G. M., editor, Posttraumatic Stress Disorder: Issues and Controversies.
Chicester, England: John Wiley & Sons.
Bromet E., Sonnaga, A. and R. C. Kessler. 1998. "Risk-Factors for DSM-III-R
Posttraumatic Stress Disorder: Findings from the National Comorbidity Survey."
American Journal of Epidemiology, Vol. 147, pages 353-361.
30
Caceres-Delpiano, J. 2006. “The Impacts of Family Size on Investment in Child Quality.”
Journal of Human Resources, Vol. 41, No. 4, pages, 738-754.
Cherlin, A. 1999. “Going to Extremes: Family Structure, Children’s Well-Being, and
Social Science.” Demography, Vol. 26, No. 4, pages 421-428.
Conley, D. and R. Glauber. 2006. “Parental Educational Investment and Children’s
Academic Risk: Estimates of the Impact of Sibship Size and Birth Order from
Exogenous Variation in Fertility.” Journal of Human Resources, Vol. 41, No. 4,
pages 722-737.
Conley, D. and J. A. Heerwig. 2009. “The Long-Term Effects of Military Conscription
on Mortality: Estimates from the Vietnam-Era Draft Lottery.” National Bureau of
Economic Research, Working Paper No. 15105.
DeWit, D. J., Chandler-Coutts, M., Offord, D. R., King, G., McDougall, J., Specht, J.,
and S. Stewart. 2005. “Gender Differences in the Effects of Family Adversity on the
Risk of Onset of DSM-III-R Social Phobia.” Journal of Anxiety Disorders. Vol. 19,
No. 5, pages 479-502.
Dobkin, C. and Shabani, R. 2009. “The Health Effects of Military Service: Evidence
from the Vietnam Draft.” Economic Inquiry, Vo. 47, No. 1, pages 69-80.
Dohrenwend, B. P., Turner, J. B., Turse, N. A., Adams, B. G., Koenan, K.C., and
Marshall, R. 2006. “The Psychological Risks of Vietnam for U.S. Veterans: A Revisit
with New Data and Methods.” 2006. Science, Vol. 313, No. 5789, pages 979-982.
Ejrnaes, J. and C. C. Portner. 2004. "Birth Order and Intrahousehold Allocation of Time
and Education." The Review of Economics and Statistics, Vol., 86, No. 4, pp. 10091019.
31
Frueh, B. C., J. D. Elhai, A. L. Grubaugh, J. Monnier, T. B. Kashdan, J. A. Sauvageot, M.
B. Hamner, B. G. Burkett and G. W. Arana. 2005. “Documented Combat Exposure of
U. S. Veterans Seeking Treatment for Combat-Related Posttraumatic Stress Disoder.”
British Journal of Psychiatry, Vol. 186, pages 467-472.
Ginther, D. K. and R. A. Pollack. 2004. “Family Structure and Children’s Educational
Outcomes: Blended Families, Stylized Facts, and Descriptive Regression.”
Demography, Vol. 41, No. 4, pages 671-696.
Grinker, R.R and J. P. Spiegel. 1945. Men Under Stress. Philadephia: The Blakeston
Company.
Haas, S. 2008. “Trajectories of Functional Health: The ‘Long-Arm’ of Childhood Health
and Socioeconomic Factors.” Social Science and Medicine, Vol. 66, No. 4, pages
849-861.
Hanushek, E. 1992. “The Trade-off Between Child Quantity and Quality.” The Journal of
Political Economy, Vol. 100, No. 1, pages 84-117.
Hausman, J. A., Abrevaya, J., Scott-Morton, F. M. 1998. “Misclassification of the
Dependent Variable in a Discrete Response Setting.” Journal of Econometrics, Vol.
87, pages 239-269.
Hausman, J. A. 2001. “Mismeasured Variables in Econometric Analysis: Problems from
the Right and Problems from the Left.” Journal of Economic Perspectives, Vol. 15,
No. 4, pages 57-68.
Haveman, R. and B. Wolfe. 1995. “The Determinants of Children’s Attainments: A
Review of Methods and Findings.” The Journal of Economic Literature, Vol. 33, No.
4, pages 1829-1878.
32
Hayward, M. D. and B. Gorman. 2004. “The Long Arm of Childhood: The Influence of
Early-Life Social Conditions on Men’s Mortality.” Demography, Vol. 41, No. 1,
pages 87-107.
Helzer, J. Robin, L. N., McEvoy, L. R., Spitnagel, El. L., Stoltzman, R. K., Farmer, A.,
and I. F. Brockington. 1985. "A Comparison of Clinical and Diagnostic Interview
Schedule Diagnoses." Archives of General Psychiatry, Col. 42, pages 657-666.
Hoge, C. W., Toboni, H. E., Messer, S. C., Bell, N., Amoroso, P., Orman, D. T. 2005.
“The Occupational Burden of Mental Disorders in the U.S. Military: Psychiatric
Hospitalizations, Involuntary Separations, and Disability.” American Journal of
Psychiatry, Vol. 162, pages 585-591.
Hyams, C. H., Wignall, F. S., Roberts, R. 1996. “War Syndrome and Their Evaluation:
From the U.S. Civil War to the Persian Gulf War.” Annals of Internal Medicine, Vol.
125, No. 5, pages 398-405.
Ilkin, J. F., McKenzie, D. P., Creamer, M. C., McFarlane, A. C., Kelsall, H. L., Glass, D.
C., Forbes, A. B., Horsley, K. W. A., Harrex, W. K., Sim, M. R. 2005.“War Zone
Stress Without Direct Combat: The Australian Naval Experience of the Gulf War.”
Journal of Traumatic Stress, Vol. 18, No. 3, pages 193-204.
Institute of Medicine. 2008. Gulf War and Health: Physiologic, Psychologic, and
Psychosocial Effects of Deployment-Related Stress. Washington D.C.: The National
Academies Press.
The Independent. 2007. “Are the Family Clichés True?”, November 27.
Jones, E. and Wessley, S. 2007. “A Paradigm Shift in the Conceptualization of
Psychological Trauma in the 20th Century.” Journal of Anxiety Disorders, Vol. 21,
33
pages 164-175.
Kantarevic, J. and S. Mechoulan. 2006. “Birth Order, Educational Attainment, and
Earnings: An Investigation Using PSID.” Journal of Human Resources, Vol. 41, No.
4, pages 755-778.
Keller, H. and U. Zach. 2002. “Gender and Birth Order as Determinants of Parental
Behavior.” International Journal of Behavioral Development, Vol. 26, No. 2, pages
177-184.
Kenkel, D. S., Lillard, D. R., and A. D. Mathios. 2004. “Accounting for Misclassification
Error in Retrospective Smoking Data.” Health Economics, Vol. 13, pages 1031-1044.
Kessler, R.C., Berglund, P., Demler, O., Jin R., Merikangas,, K. R. , PhD, and Walters E.
E. 2005. “Lifetime Prevalence and Age-of-Onset Distributions of DSM-IV
Disorders in the National Comorbidity Survey Replication.” Archives of General
Psychaatry, Vol. 62, pages 593-602.
Kessler, R. C., Sonnega, A., Bromet, E., Hughes, M., Nelson, C., and N. Breslau. 1999.
"Epidemiological Risk Factors for Trauma and PTSD." In Yehuda, R., editor, Risk
Factors for Posttraumatic Stress Disorder. Washington D.C.: American Psychiatric
Press, Inc.
Kessler, R. C., Wittchen, H., Abelson, McGonable, J. M., Schwarz, N., Kendler, K. S.,
Knauper, B., and S. Zhao. 1998. “Methodological Studies of the Composite
International Diagnostic Interview (CIDI) in the U.S. National Comorbidity Survey
(NCS). International Journal of Methods in Psychiatric Research, Vol. 7, No. 1,
pages 33-35.
Krein, S. F. and A. H. Beller. 1988. "Educational Attainment of Children from Single-
34
Parent Families: Differences by Exposure, Gender, and Race." Demography, Vol. 25,
pp. 679-690.
Kristensen, P. and Bjerkedal T. 2007. “Explaining the Relation Between Birth Order and
Intelligence.” Science, Vol. 316, No. 5832, page 1717.
Kulka, R.A,, Schlenger, W.E., Fairbank, J. A., Hough, R. L., Jordan, B. K., Marmar, C.
R., and D. S. Weiss. 1988. Contractual Report of the Findings from the National
Vietnam Veterans Readjustment Study. Volume 1: Executive Summary, Description of
Findings, and Technical Appendices. Research Triangle Park, NC: Research Triangle
Institute.
Kulka, R.A,, Schlenger, W.E., Fairbank, J. A., Hough, R. L., Jordan, B. K., Marmar, C.
R., and D. S. Weiss. 1990. Trauma and the Vietnam War Generation: Report of
Findings From the National Vietnam Veterans Readjustment Study. New York:
Brunner/Mazel.
Lang, K. and J.L. Zagorsky. 2001. “Does Growing Up With a Single Parent Really
Hurt?” Journal of Human Resources, Vol. 36, No. 2, pages 253-273.
Lindert, P. H. 1977. "Sibling Position and Achievement." Journal of Human Resources,
Vol. 12, No. 1, pp. 198-219.
Lewbel, A. 2000. “Identification of the Binary Choice Model with Misclassification.”
Econometric Theory, Vol. 16, pages 603-609.
Manski, C. F., G. D. Sandfur, S. McLanahan, and D. Powers. 1992. "Alternative
Estimates of the Effect of Family Structure During Adolescence on High School
Graduation." Journal of the American Statistical Association, Vol. 87, No. 417, pp.
25-37.
35
Mathews T.J., and Hamilton, B.E. 2005. “Trend analysis of the sex ratio at birth in the
United States.” National Vital Statistics Reports, Vol. 53, No 20.
McNally, Richard. 2003. “Progress and Controversy in the Study of Posttraumatic Stress
Disorder.” Annual Review of Psychology, Vol. 54. pp. 229-252.
McNally, R. 2007. “Can We Solve the Mysteries of the National Vietnam Veterans
Readjustment Study?” Journal of Anxiety Disorders, Vol. 21, pages 192-200.
McNally, Richard. 2004. “Conceptual Problems with the DSM-IV Criteria for
Posttraumatic Stress Disorder.” In Rosen, G. M., editor, Posttraumatic Stress
Disorder: Issues and Controversies. Chicester, England: John Wiley & Sons.
Mullahy, J. and J. Sindelar. 1989. “Life-cycle Effects of Alcoholism on Education,
Earnings, and Occupation.” Inquiry, Vol. 26, pages 272-282.
Paul, O. 1987. “Da Costa Syndrome or Neurocirculatory Asthenia.” British Heart
Journal, Vol. 58, pages 306-315.
Poterba, J. M., and L. Summers. 2005. “Unemployment Benefits and Labor Market
Transitions: A Multinomial Logit Model With Errors in Classification.” The Review
of Economics and Statistics, Vol. 77, No. 2, pages 207-216.
Prins, A., P. Ouimette, R. Kimmerling, R. P. Cameron, D. S. Hugelshofer, J. ShawHegwer, A. Thrailkill, F. D. Gusman, and J. I. Sheikh. “The Primary Care PTSD
Screen (PC-PTSD): Development and Operating Characteristics. Primary Care
Psychiatry, Vol. 9, No. 1, pages 9-14.
Robins, L. and Regier, D. 1991. Psychiatric Disorders in America. Free Press: New
York.
Rogan, W. J. and B. Gladen. 1978. “Estimating Prevalence from the Results of a
36
Screening Test.” American Journal of Epidemiology, Vol. 107, No. 1, pages 71-76.
Rohlfs, C. 2007. "Does Combat Exposure Make You a More Violent or Criminal
Person?: Evidence from the Pre-Lottery Vietnam Draft." Manuscript.
Saravey, S. M., Steinberg, M. D., Weinschel, B., Pollack, S. and Alovis, N. 1991.
“Psychological Comorbidity and Length of Stay in the General Hospital.”
American Journal of Psychiatry, Vol. 148, pages 423-329.
Savoca, E. 1992. “Measurement Error in Self-Evaluations of Mental Health: Implications
for Labor Market Analysis.” In Frank, R. G., and W. G. Manning, editors, Economics
and Mental Health. Baltimore: The Johns Hopkins University Press.
Savoca, E. 1995. “Controlling for Mental Health in Earnings Equations: What Do We
Gain and What Do We Lose?” Health Economics, Vol. 4, pages 399-410.
Savoca, E. 2000. “Measurement Error in Binary Regressors: An Application to
Measuring the Effects of Specific Psychiatric Diseases on Earnings.” Health Services
and Outcomes Research Methodology, Vol. 1, No. 2, pages 149-164.
Savoca, E. 2004. “Sociodemographic Correlates of Psychiatric Diseases: Accounting for
Misclassification in Survey Diagnoses of Major Depression, Alcohol and Drug Use
Disorders.” Health Services and Outcomes Research Methodology, Vol. 5, pages 175191.
Savoca, E. 2005. “Revised Estimates of the Risk Factors for Psychiatric Diseases in the
United States.” In 2005 Proceedings of the Health Policy Statistics Section,
American Statistical Association. Alexandria, VA: American Statistical
Association.
37
Savoca E. 2010. “Accounting for Misclassification Bias in Binary Outcome Measures of
Illness: The Case of Posttraumatic Stress Disorder in Male Veterans.” Manuscript.
Scott, W. J. 1990. “PTSD in DSM-III: A Case in the Politics of Diagnosis and Disease.”
Social Problems, Vol. 37, No. 3, pages 294-310.
Shephard, B. 2001. A War of Nerves: Soldiers and Psychiatrists in the Twentieth
Century. Cambridge, MA: Harvard University Press.
Spiegal A. 2005. “The Dictionary of Disorder: How One Man Revolutionized
Psychiatry.” The New Yorker, January 3rd.
Tanielian T. and Jaycox, L. H. 2008. Invisible Wounds of War: Psychological and
Cognitive Injuries, Their Consequences, and Services to Assist Recovery. Santa
Monica, CA: Rand Corporation.
Thomas, D. and E. Frankenberg. 2000. “The Measurement and Interpretation of Health in
Social Surveys.” Rand Corporation, Labor and Population P rogram, Working
Paper Series, No. 01-06.
True, W. R. and M. Lyons. 1999. "Genetic Risk Factors for PTSD: A Twin Study." In
Yehuda, R., editor, Risk Factors for Posttraumatic Stress Disorder. Washington
D.C.: American Psychiatric Press, Inc.
U. S. Census Bureau. 2003. Historical Statistics, Statistical Abstract of the United States,
2003. Washington D.C.
U.S. Department of Commerce. 1936. Birth, Stillbirth, and Infant Mortality Statistics for
the Continental United States, the Territory of Hawaii, and the Virgin Islands.
U. S. Department of Veteran Affairs. 2005. “Review of State Variances in VA Disability
Compensation Payments.” Report No. 05-00765-137.
38
Wilson, J. P. 1994. “The Historical Evolution of PTSD Diagnostic Criteria: From Freud
to DSM-IV.” Journal of Traumatic Stress, Vol. 7, No. 4, pages 681-698.
39
Table 1. Diagnostic Accuracy of the Mississippi Scale for Combat-Related Stress
Diagnostic Accuracy in a Clinical Sample (Pre-Survey):
Percent Correctly Classified
Kappa
88.9%
0.753
False
Negative
Rate
6.0%
False Positive
Rate
20.3%
Diagnostic Accuracy in a Validation Study of NSVG Participants (Post-Survey):
Percent Correctly Classified
81.6%
Kappa
False
Negative
Rate
22.7%
False Positive
Rate
0.528
17.2%
(0.332, 0.724)
Source: Kulka et al. (1988), Appendix D, Exhibit D-3 and Exhibit D-8.
Notes: Percent correctly classified is the percent of veterans whose survey and clinical
diagnoses were in agreement. The false negative rate is the percent of veterans with a positive
clinical diagnosis who received a negative survey diagnosis. The false positive rate is the
percent of veterans with a negative clinical diagnosis who received a positive survey diagnosis.
The Kappa score compares the actual agreement between the diagnoses to the chance
agreement we would observe if the diagnoses were completely independent, with a zero score
corresponding to pure chance agreement.
40
Table 2. Description of Variables
Mean (Standard Deviation)
Full Sample
Positive PTSD
Cases
Negative
PTSD Cases
Dependent Variables
Indicators of PTSD
Diagnosis (positive = 1; negative = 0)
Mississippi Score (min=35, max = 175)
0.22 (0.41)
72.50 (21.77)
1.00 (0.00)
105.41 (15.90)
0.00 (0.00)
63.40 (12.27)
Independent Variables
Indicators of Trauma Exposure (At least one
definite or severely traumatic event = 1;
otherwise = 0)
Combat-related Trauma
Trauma, not related to war
0.367 (0.482)
0.190 (0.392)
0.691 (0.463)
0.373 (0.485)
0.277 (0.448)
0.138 (0.346)
Family Characteristics
Number of Siblings
Intact Family
Poor
Relative Birth Order
Location Move Involving School Switch
3.859 (2.986)
0.772 (0.420)
0.337 (0 .473)
0.398 (0.383)
0.414 (0.493)
4.000 (2.957)
0.712 (0.454)
0.399 (0.491)
0.403 (0 .370)
0.451 (0.499)
3.819 (2.995)
0.788 (0.409)
0.320 (0.467)
0.397 (0.386)
0.403 (0.491)
0.057 (0.233)
0.154 (0.361)
0.120 (0.326)
0.219 (0.414)
0.040 (0.195)
0.136 (0.343)
0.043 (0.203)
0.605 (0.489)
0.069 (0.253)
0.670 (0.471)
0.036 (0.187)
0.587 (0.493)
0.488 (0.500)
0.511 (0.501)
0.481 (0.500)
0.602 (0.490)
0.678 (0.468)
0.580 (0.494)
0.210 (0.407)
0.279 (0.449)
0.190 (0.393)
41.197 (5.165)
0.539 (0.499)
39.820 (3.579)
0.476 (0.501)
41.584 (5.468)
0.556 (0.497)
1064
233
831
Parental Characteristics
History of Mental Illness
History of Problems with Alcohol or
Drugs
History of Criminal Behavior
Father Did Not Graduate from High
School
Mother Did Not Graduate from High
School
Mother Worked During Respondent’s
Childhood
Father experienced at least one spell of
Unemployment during the
respondent’s childhood
Characteristics of the Respondent
Age
White
Sample Size
41
Table 3. Ordinary Regression Estimates of Intensity of Stress Reactions To Traumatic Events
Dependent Variable
Independent Variables:
Indicators of Trauma Exposure (At least one
definite or severely traumatic event = 1;
otherwise = 0)
Combat-related Trauma
Trauma, not related to war
Score on Mississippi Scale
(1)
(2)
Log of Score on
Mississippi Scale
(3)
15.972a (1.516)
5.491a (2.006)
0.354
0.099
0.240 a (0.023)
0.068a (0.025)
-7.736 b (3.025)
-3.622 (3.265)
7.292 b (3.488)
0.220 (1.400)
-0.008 (0.219)
21.963a (5.833)
-21.760a (5.807)
-0.699 (1.203)
-0.149
-0.081
0.166
0.005
-0.001
0.386
-0.381
-0.016
-0.092b (0.035)
-0.037 (0.040)
0.095 (0.049)
0.006 (0.018)
0.000 (0.003)
0.291a (0.076)
-0.280a (0.076)
-0.009 (0.016)
10.001a (3.150)
3.372c (1.986)
3.794 (3.045)
1.078 (1.429)
-0.586 (1.485)
3.247 b (1.548)
0.107
0.056
0.035
0.024
-0.013
0.061
0.114a
0.040c
0.061
0.011
-0.005
0.042b
Family Characteristics
Intact Family
Mother Worked During Respondent’s Childhood
Interaction Effect (mother worked, family intact)
Poor
Number of Siblings
Relative Birth Order
Relative Birth Order Squared
Location Move Involving School Switch
Parental Characteristics (Yes = 1; No = 0)
History of Mental Illness
History of Problems with Alcohol or Drugs
History of Criminal Behavior
Father Did Not Graduate from High School
Mother Did Not Graduate from High School
Father experienced at least one spell of
Unemployment during the
respondent’s childhood
(0.040)
(0.025)
(0.041)
(0.019)
(0.019)
(0.020)
Characteristics of the Respondent
-0.159
-0.009a (0.001)
Age
-0.668a (0.098)
a
White
-3.851 (1.328)
-0.088
-0.051a (0.016)
Notes: Standard errors, in parentheses next to the coefficient estimates, are heteroscadastic consistent. Column (2)
reports the beta coefficients for the estimates in column (1). For the discrete covariates in the log specification of
Column (3), exp( β̂ ) -1 and its standard error is reported as a measure of the impact of a categorical change on the
proportionate change in the Mississippi score. Sample Size = 1064.
a
p-value .01; b .01< p-value .05; c .05< p-value .10.
42
Table 4. Estmates of the Probability of Exposure to Trauma
Dependent Variable :Exposure to Any Type of Trauma
(Yes = 1; No = 0)
Independent Variables
Cohort Combat Casualty Rate (deaths per 1000)
Probit Model
Linear
Probability
Model
0.102a (0.025)
0.038a (0.009)
0.062
0.031
0.010
0.136
-0.018
-0.352
0.294
0.089
0.023
0.011
0.003
0.051
-0.006
-0.129
0.107
0.033
Family Characteristics
Intact Family
Mother Worked During Respondent’s Childhood
Interaction (mother worked and family intact)
Poor
Number of Siblings
Relative Birth Order
Relative Birth Order Squared
Location Move Involving School Switch
Parental Characteristics (Yes = 1; No = 0)
History of Mental Illness
History of Problems with Alcohol or Drugs
History of Criminal Behavior
Father Did Not Graduate from High School
Mother Did Not Graduate from High School
Father experienced at least one spell of
Unemployment during the
respondent’s childhood
(0.185)
(0.197)
(0.216)
(0.093)
(0.016)
(0.402)
(0.403)
(0.083)
0.361b (0.172)
-0.233a (0.207)
0.318 (0.113)
0.062 (0.096)
-0.012 (0.098)
-0.010 (0.103)
(0.067)
(0.072)
(0.079)
(0.035)
(0.006)
(0.150)
(0.150)
(0.031)
0.138b (0.066)
0.122a (0.044)
-0.089 (0.075)
0.023 (0.036)
-0.003 (0.037)
-0.037 (0.038)
Characteristics of the Respondent
Age
0.004 (0.009)
0.001 (0.003)
White
0.002 (0.090)
0.001 (0.034)
Notes: Standard errors, in parentheses next to the coefficient estimates, are heteroscadastic consistent.
Sample Size = 1064.
a
p-value .01; b .01< p-value
.05.
43
Table 5. Instrumental Variables Estimates of the Intensity of Stress To Traumatic Events
Dependent Variable: Score on the Mississippi Scale
Instrumental Variables
(1)
(2)
OLS
(3)
Independent Variables:
Exposure to Traumatic Event
(Yes = 1; No = 0)
Family Characteristics
Intact Family
Mother Worked During Respondent’s Childhood
Interaction (mother worked and family intact)
Poor
Number of Siblings
Relative Birth Order
Relative Birth Order Squared
Location Move Involving School Switch
Parental Characteristics
History of Mental Illness
History of Problems with Alcohol or Drugs
History of Criminal Behavior
Father Did Not Graduate from High School
Mother Did Not Graduate from High School
Father experienced at least one spell of
Unemployment during the
respondent’s childhood
23.444b (10.192)
20.911b (9.773)
18.477 a (1.242)
-8.029a (2.861)
-4.092 (2.663)
7.294b (3.120)
-0.064 (1.218)
-0.008 (0.224)
22.972a (6.899)
-22.695a (6.673)
-0.738 (1.234)
-7.968a (3.098)
-4.084 (3.327)
7.326b (3.532)
0.060 (1.469)
-0.026 (0.226)
22.584a (6.108)
-22.367a (6.021)
-0.659 (1.245)
-8.012a (3.021)
-4.178 (3.274)
7.455b (3.491)
0.169 (1.406)
-0.044 (0.220)
22.322a (5.891)
-22.158a (5.848)
-0.568 (1.203)
9.681a
2.644
4.422
1.110
-0.978
3.634b
10.024a (3.446)
2.950 (2.390)
4.179 (3.258)
1.174 (1.451)
-0.985 (1.491)
3.559b (1.556)
10.347a
3.235
3.945
1.256
-1.005
3.486b
(2.763)
(2.193)
(3.061)
(1.811)
(1.715)
(1.735)
(3.221)
(1.995)
(3.097)
(1.430)
(1.490)
(1.548)
Characteristics of the Respondent
Age
-0.652a (0.086)
-0.659a (0.102)
-0.668a (0.097)
a
a
-4.158 (1.340)
-4.156a (1.336)
White
-4.151 (1.376)
Notes: Robust standard errors are reported in parentheses next to the coefficient estimates. Column (1)
uses the instrument for trauma constructed from the probit model in Table 4; Column (2) uses the
instrument for trauma constructed from the linear probability model in Table 4. Sample size = 1064.
a
p-value .01; b .01< p-value .05.
44
Table 6. Nonlinear Weighted Least Squares Estimates of Probit Models
of the Errors in PTSD Diagnoses
Parameter Estimates
Model 1
False Positive Rate
Constant
White
Age
False Negative Rate
Constant
White
Age
Model 2
Model 3
-1.228a (0.064)
0.638 (0.689)
-0.230c (0.136)
-0.043a (0.017)
0.374 (0.643)
---0.039b (0.016)
-0.729a (0.258)
0.559 (2.516)
0.218 (0.425)
-0.035 (0.064)
-0.740a (0.261)
-----
Predicted Error Rates for Selected Cohortsd
False Positive Rates
Age 35
White
Nonwhite
Age 40
White
Nonwhite
Age 45
White
Nonwhite
Age 50
White
Nonwhite
0.110a (0.012)
0.136a (0.028)
0.192a (0.036)
a
0.094 (0.016)
0.139a (0.021)
0.063a (0.015)
0.096a (0.020)
a
0.040 (0.017)
0.065a (0.023)
0.162a (0.027)
0.119a (0.013)
0.084a (0.015)
0.058a (0.019)
0.230a (0.079)
False Negative Rates
0.233a (0.079)
Age 35
White
0.326b (0.134)
Nonwhite
0.252c (0.148)
Age 40
White
0.266a (0.095)
Nonwhite
0.200c (0.115)
Age 45
White
0.211 (0.140)
Nonwhite
0.154 (0.136)
Age 50
White
0.164 (0.191)
Nonwhite
0.116 (0.163)
Notes: Standard errors, in parentheses next to the coefficient estimates, are heteroscedastic-consistent.
Sample Size = 1064. Model 1 specifies constant error rates. In Model 2 the error rates vary by age and
race. In Model 3, only the false positive rate varies across observations, and only by age.
a
p-value 0.01; b0.01< p-value 0.05; c 0.05< p-value 0.10. dStandard errors computed via the delta
method.
45
Table 7. Nonlinear Weighted Least Squares Estimates of Probit Models
of the Determinants of the Probability of a Positive PTSD Diagnosis
Without Error
Corrections
With Error Corrections
Model 1
Model 2
Model 3
0.944a (0.112)
0.263b (0.128)
4.237a (1.287)
0.457 (0.304)
4.095a (1.372)
0.459 (0.298)
4.194a (1.379)
0.533c (0.318)
-0.509b (0.205)
-0.111 (0.218)
0.377 (0.244)
-0.002 (0.109)
-0.006 (0.018)
1.522a (0.472)
-1.428a (0.477)
-0.045 (0.099)
-1.858b (0.733)
-0.697 (0.613)
1.931b (0.839)
0.574c (0.312)
0.055 (0.055)
3.700b (1.640)
-3.657b (1.646)
0.585c (0.308)
-1.768b (0.710)
-0.724 (0.607)
1.827b (0.814)
0.535c (0.308)
0.051 (0.054)
3.302 b (1.558)
-3.222b (1.563)
0.587c (0.307)
-1.948a
-0.743
2.035b
0.629c
0.061
3.679b
-3.659b
0.626c
(0.753)
(0.621)
(0.869)
(0.328)
(0.056)
(1.683)
(1.695)
(0.321)
0.496a (0.180)
0.128 (0.135)
0.296 (0.225)
0.146 (0.115)
-0.081 (0.115)
0.250b (0.114)
2.445b (0.979)
-0.337 (0.347)
1.282c (0.755)
0.493 (0.319)
-0.491 (0.332)
0.643c (0.342)
2.254b
-0.256
1.187c
0.578c
-0.423
0.570c
2.347b
-0.341
1.196
0.549c
-0.521
0.648c
(1.011)
(0.360)
(0.743)
(0.332)
(0.344)
(0.344)
-0.051a (0.011)
-0.185c (0.108)
-0.129b (0.052)
-0.287 (0.303)
-0.109c (0.056)
0.057 (0.401)
Independent Variables
Trauma Exposure (At least one definite or severely
traumatic event = 1; otherwise = 0)
Combat-related trauma
Trauma, not related to war
Family Characteristics
Intact Family
Mother Worked During Respondent’s Childhood
Interaction (mother worked and family intact)
Poor
Number of Siblings
Relative Birth Order
Relative Birth Order Squared
Location Move Involving School Switch
Characteristics of Parents (Yes = 1; No = 0)
History of Mental Illness
History of Problems with Alcohol or Drugs
History of Criminal Behavior
Father Did Not Graduate from High School
Mother Did Not Graduate from High School
Father experienced at least one spell of
Unemployment during the
respondent’s childhood
Characteristics of the Respondent
Age
White
Notes: See Table 6.
a
p-value 0.01; b0.01< p-value
0.05; c 0.05< p-value
(1.037)
(0.353)
(0.711)
(0.326)
(0.321)
(0.326)
-0.102b (0.048)
-0.256 (0.311)
0.10
46