1 Depressivity over the Life Course and Across Generations – A Work in Progress Elizabeth CookseyCurtis EberweinRosella GardeckiPamela IngRandall J. Olsen
2 Why Is This an Interesting Topic for Longview? Depression is probably responsible for more lost days of work and lost productivity than any other cause (short of death)Mental health problems typically manifest themselves in adolescence and young adulthoodBoth environmental and genetic factors are likely involved, both of which may be illuminated with longitudinal data
3 Longitudinal Data We Use National Longitudinal Surveys of Labor Market Experience – project dates from 1965Mature Women in 1966Young Women in 19681979 Youth Cohort born (NLSY79)Young Adult Studybiological children of female respondents in NLSY79Data on perinatal health + social, emotional and cognitive development through age 14 thence “conventional” interviews every 2 years
4 Key Depression Measures Committee on Epidemiological Studies Depression scale (CES-D) is due to Radloff (1977) and the most widely used measure of depression in the U.S.Has 20 items that solicit symptom frequency over the past weekrarely or none of the time (less than once a week)some or a little of the time (1 to 2 days of the week)occasionally or a moderate amount of the time (3 to 4 days a week)most or all of the time (5-7 days of the week)Answers scored 0-3Ex: “I felt that I could not shake off the blues even with help from my family or friends.”
5 Open Questions about CES-D Depression is supposed to be a long-run, not short term condition; questions ask about the last 7 daysIs the CES-D flawed because of the near-term horizon of the questions? Having repeated measures over several years will help answer that questionThe original focus of the CES-D was on young adults; does it perform adequately over the life cycle?
6 Psychometrics in Omnibus Social Science Surveys We often hear that to do a good job of measuring a psychological construct, we need to ask a seemingly interminable sequence of scale itemsHaving all these answers creates a question in the minds of other social scientists as to how to use the dataTypically a simple sum of scores used; is this really the right approach no matter how questions may vary?Is there a reasonable way we can use the results of these scales as explanatory variables for investigations that center on economic or sociological applications?
7 Item Response Theory – A Quick Summary; Apologies to the Cogniscenti It assumes there is a latent distribution for a trait that is, or can be represented as, standard normalIt presumes the scale items measure a unidimensional traitItems differ in their ability to discriminate in where a person lies on this continuum and, if the response set is Likert-like (as opposed to correct/incorrect in achievement or cogntivie testing), moving through the response set may have different implications for where the person is on the unidimensional scaleThe outcome of the testing protocol is a z-score for where a particular person is on the continuum
8 Rasch Model-TestingRespondent i has an innate capability qi drawn from a standard normal distribution. Item j has a difficulty level bj and the probability that respondent i will answer item j correctly is F[a(qi - bj)], where F is the logistic cumulative distribution function. The parameter bj is referred to as the “threshold” as it is the limit point at which persons with larger values of qi are more likely to endorse the item than not. When a is same for all items, the number correct is a sufficient statistic for qi and results will agree with Classical Testing Theory (i.e., scaling based on sum correct). Our data reject Rasch.
9 Two Parameter ModelWe let the a parameter differ across items so we haveF[aj (qi - bj)]a captures the responsiveness to the item of a change in q. If q is above the threshold bj and aj is large, probability of endorsement is large; if a person “endorses” an item with a large aj and a large bj, it means their qi is large, so they are more likely to endorse other items with similar or smaller thresholds.Estimation requires maximum likelihood; simultaneously estimate qi and the item parameters aj and bj. Items need not be the same across waves.
10 Data We UseYoung Women and Mature Women CES-D in 1995, 1997, 1999, 2001 & 2003 –about 4800 observations at each point; used a 7 item scaleNLSY79 in 1992 (20 items), 1994 and at age 40 (7 items for last two) – about 8400 observations each;Young Adults with varying numbers of cases; they were asked 7 items scales with 2 more items added in 2008.Taken as a group, we have tens of thousands of observations ranging from the teens to the 80’s.
11 Correlation over Time in Depressivity for Women’s Cohorts: Takeaway – Substantial Stability in qi Half of Variation is Measurement Error199519971999200120031.0000.5260.4790.4620.4250.5470.5220.4680.5450.511
12 Correlation over Time in Depressivity for Young Adults: Takeaway – Substantial Stability in qi Half of Variance is Measurement Error199419961998200020022004200620081.0000.4070.3480.388nsd0.2940.2490.3170.4120.3790.3680.3200.2570.2700.2910.3120.2990.3030.2980.3460.2860.4080.3070.3160.4270.3980.429
13 Variation of Mean of q Over the Life Course AgeMeanS.E. meanSD of qS.E.14-210.000.00 (baseline)1.0022-29-0.010.020.930.0130-37-0.131.0340-47-0.561.2148-55-0.621.2756-63-1.230.031.3564-71-1.621.2072-79-1.511.1780-84-188.8.131.52.08
14 Intergenerational Effects – How Does Mother’s qi Affect Her Daughters and Sons? (correcting for measurement error)qYA,M = qM + e( ) ( )qYA,F = qM + e( ) ( )
15 Preliminary ResultsCES-D captures a great deal of the variation in depressivity over the life courseAbout half of the variation is measurment error – the rest is due to high order auto-correlation or fixed effectsDepressivity falls from adolescence into the 60’s and then starts to riseDepressivity does pass down the generations, but it is unclear whether it is environmental, taught or geneticWe have work is underway relating measureable factors to depressivity – this may influence our estimates of intergenerational transmission