For a specific exposure-outcome combination of interest to you, specify which lifecourse model is likely most appropriate and why you think this is the case. Describe the regression models you could use to test your hypothesis. Are there any possible data sets in which this test could be conducted, and if so, what concerns would you have about interpreting your proposed test of the lifecourse model?
Due before class: April 27, 2015
Do repeat/chronic STH infections in early childhood affect height at age 12?
Early childhood STH Infection --> Height at age 12
I think the accumulation model would make most sense because I am expecting that the effects of repeat infections or chronic infections accumulate over time.
E(Y) = β0 + βT1+ βT2 + βT3
I’m not aware of an existing data set that could be used to measure this, but it could be done in my mentor, Jeremy Kennan’s, upcoming WASH trial where STH infection will be measured every year for 4 years. You could do a follow up study 10 years later and link the data.
** what concerns would you have about interpreting your proposed test of the lifecourse model?
** I don't understand the coding of the model - are T1, T2, and T3 indicators for the timing of assessment or indicators for infection?
** do you intend to let the betas for each time vary? or are all betas in this equation equal?
A clearer model
simple accumulation of risk:
E(height)=b0+b1*(SHT1+STH2+STH3)
Where STHx = STH infection at times 1, 2, and/or 3
Screening, Brief Intervention, and Referral to Treatment (SBIRT)
Study
This RCT study assessed the Specific Substance Involvement Score (SSIS) of HIV-+ substance using men who have sex with men (SUMSM) who completed the WHO’s Alcohol, Smoking and Substance Involvement Screening Test (ASSIST) at baseline and every three months thereafter, a total of four visits (baseline and 3 f/u visits). The participants were randomized into a clinician-based intervention group. or web-based intervention group. The ASSIST asks 10 questions (use ever, last 30 days, regrets, family/friends worried, etc.) for 10 substances such as smoking, alcohol, cannabis, amphetamines, cocaine, etc.
H0
I hypothize that the clinician-based intervention will reduce the SSIS significantly more than the web-based intervention.
Lifecourse model
I think, I would use he risk and protective factors model because resilience may be an important factor of the success of a behavioral intervention over time.
Test
E(Y) = β0 + β1Baseline + β2f/uvisit1 + β3f/uvisit2 + β4f/uvisit3
Data Set
My advisor’s data set (not yet analyzed). Concerns: each drug could be considered a mediator.
Roland: it's really interesting that you brought up "resilience" and remind me to discuss this in class.
By "resilience" do you mean the capacity to recover after injury or the capacity to be unharmed by injury in the first place? If the latter, I don't think this necessarily implies one lifecourse model over another. However, the capacity to recover after inury implies that there is not a strict "critical" period and is more consistent with a cumulative risk model.
Question: Among adults with a history of an alcohol use disorder, is the rate of cognitive decline slower in those with higher years of education than in those with lower years of education? (I’m essentially interested in whether years of education is somehow protective against any effects on cognition that an alcohol use disorder might have had.)
Exposure: alcohol use disorder
Outcome: rate of cognitive decline
Effect modifiers: years of education
Other covariates: age, sex, smoking, marital status, concurrent medical disorders (e.g., hypertension, diabetes, head trauma), concurrent medication use
Dataset: National Alzheimer’s Coordinating Center’s Minimum Data Set (n = 74,397) & Uniform Data Set (n = 31,872), a longitudinal dataset of adults with normal cognition/mild cognitive impairment/dementia diagnosis (https://www.alz.washington.edu/WEB/data_descript.html). This dataset started in 1984 and has neurocognitive testing and substance use variables (e.g., smoking, alcohol abuse, drug abuse).
Lifecourse models: 1) This would probably be an “accumulation of risk” model (based on the Ben-Shlomo and Kub 2002 publication), since the risk of cognitive decline would accumulate with various “insults” – such as alcohol use, smoking, medical disorders, etc.. Some of these “insults” might be correlated; for example, someone with alcohol use might be prone to falling & have head trauma. 2) These research questions would also be looking at life course trajectories, as in the Wills et al. 2011 & Wilson et al. 2009 publications).
Regression models: I don’t fully know the specific formulas, but I would at least include the following in a mixed model. 1) Accumulation of risk model: fixed effects (between-subjects) + random effects (within-subjects) + interaction term (years of education) + a squared/cubic/quadratic term to model for nonlinearity? 2) Lifecourse trajectory model: includes education, education x time, education x time-squared; similarly need to model for nonlinearity
Concerns: Recall bias with substance use variables; baseline IQ might be another variable which might better capture lifecourse trajectory of cognition, but IQ not available in this dataset. This dataset doesn’t account for whether someone is actively in treatment for their alcohol use disorder. It’s only a current/past/never assessment of an alcohol use disorder.
Raj: This is a great example of when defining the causal question is tricky and Tyler's distinction between a question about the causal effects of both exposures versus a question about the causal effect of one exposure within subgroups defined by the 2nd exposure is useful.
Are you envisioning that you want to know the effect of education among people with lots of EtOH abuse? Or about the effect of EtOH abuse among high education indivdiuals? Or about whether the effects of EtOH and education are more than additive (or more than multiplicative?)
In your model specification, what is the thing you describe as "interaction term (years of education)", i.e., what is the interaction with?
Why are you interested in the education*time interactions but not the etoh*time interactions?
What is the role of parental acculturation on risk of incident diabetes?
Lifecourse model:
Critical period
There has been a growing concern of diabetes affecting younger and younger children. Many health campaigns encourage parents to establish healthy habits in their children so that these habits can be maintained throughout life. Given the Hispanic paradox, where Hispanics tend to have better health outcomes than their white counterparts, it is possible that parents with lower acculturation may pass on health behaviors to their children that help reduce diabetes risk. Children often learn to imitate their parents in their early years, so I believe examining the parents’ acculturation level during this critical period (4-8 years) can provide greater insight into diabetes risk among children.
Regression models:
According to the article by Naumova et al, we can assess the risk of diabetes before/after the critical period using mixed models, specifically using logistic regression to determine if the children had diabetes or not.
We can have three separate subject-specific slopes for risk before, during, and after the critical period.
Datasets
Possible dataset to use is MESA (for parental acculturation), but children’s diabetes risk is likely not available. It is quite difficult to obtain the right data for this study question because 1) we would need data collected on parents who had very children during the course of the study (in order to correctly assess parental acculturation during the children’s critical period) and 2) we would need to follow-up on the participant’s children.
Tu My,
A core part of the "critical period" model is that people do not necessarily develop the disease during the critical period. It is "critical" for exposure, but disease itself may develop decades later. This is the fundamental premise of Barker's birthweight model: in utero experiences change development (prioritizing brain development over kidney, lung) in ways that put the person at long term risk. Even with current early onset of T2D, you would still expect the effects of parents' acculturation on kids T2D to be delayed - parents teach their kids what to eat and how to cook - a direct influence that probably wanes by age 15 or so (?), but the consequences of those habits (cooking healthy, traditional (?) meals versus eating an unhealthy western diet) probably accumulate over years and culminate in disease in adulthood or middle age.
Note here we conceptualize the parents' influence on the kid to have a sensitive period in the kids' childhood, but the effect of diet on diabetes to accumulate across time.
How much can you assess a person's acculturation retrospectively? For example, could you base this on date of immigration, on current acculturation measures (diet, language, place of residence)? What are the best measures of acculturation and how could they be modified to be used for your research question? I am not sure about Latinos, but many cohorts now incorporate a substudy of kids (or parents), e.g,. PSID randomly selected households in the late 1960s and follows every person who becomes a family member to a household member (e.g., kids, spouses, temporary spouses, spouses of kids, I think stepkids; Nurses' Health Study enrolled the GUTS cohort (Growing Up Today Study) of nurses' kids; Framingham of course enrolled the next generation - I think there are 3 generations of the original families now I do not know if they followed the Omni cohort kids (Omni was Framingham's effort at enrolling a more racially diverse cohort).
For a specific exposure-outcome combination of interest to you, specify which lifecourse model is likely most appropriate and why you think this is the case. Describe the regression models you could use to test your hypothesis. Are there any possible data sets in which this test could be conducted, and if so, what concerns would you have about interpreting your proposed test of the lifecourse model?
Exposure: infection with TB
For previously uninfected individuals this would correspond to the change from TST negative to TST positive. For previously infected individuals, additional infections can occur but are unobserved.
Outcome: clinical disease (about 90% of those infected do not develop clinical disease)
Lifecourse model: Accumulation of risk + sensitive period
We know that individuals with household contact who are infected are far more likely to develop clinical disease than those who are infected from casual contact. This lends support to an accumulation of risk model. However, there is also reason to believe that there is a sensitive period of about 18 months after an infection during which additional infections confer additional risk beyond the risk conferred by additional infections after 18 months.
Regression models:
Well, this is tricky since reinfection is unobserved. Furthermore, it's unclear how multiple infections might cause disease within this 18 month sensitive period. Perhaps, 2 infections in an 18 month window is worse than 1, and 3 is worse than 2. But perhaps at some point there has been so much repeated exposure that say 5 infections isn't worse than 4, for example.
Also, it's unclear whether multiple infections within an 18 month period increases your lifetime risk or just your risk within about a 1-5 year period following the 18 month period. The latter is more likely to be true, but I will assume the former.
I will also make the following simplifying assumptions as a first stab at the problem:
- Each infection increases the hazard developing active disease
- Each additional infection within an 18 month period increases the hazard of active disease
Proportional Hazard model:
\lambda (t|x) = \lambda_0 (t) exp(beta_1 N + beta_2 \sum_i M_i (L_i - 1) + beta_3 X N )
Where t is the time since most recent infection, \lambda (t|x) is the hazard of disease, N is lifetime number of infections, M_i is the number of 18 month periods during which multiple infections occurred, and L_i is the number of infections accumulated during that period. Here, I would define \lambda_0 (t) to be the hazard for an individual with 1 infection, since infection is a necessary cause of disease.
Notes:
There is no intercept in this model, since you must be infected to get active TB!
Also, I didn't get into what the covariates X are, but you could imagine interaction terms between sex and number of infections--e.g. women who are infected have lower risk of active disease than men who are infected.
Datasets and interpretation
There are many datasets with TST status and active disease. The difficult things about this problem are
1) Multiple infections are unobserved. We could, however, make assumptions about infections given known exposure to active cases.
2) TB is dynamical.
So, I think ultimately a dynamical model would be more productive here--a model that includes partially observed states (i.e. infection) and instead fits an observed outcome like progression to disease using a model with constant risk of infection versus accumulation + sensitive period.
Exposure-Outcome: Is first pregnancy a critical period that affects subsequent levels of physical activity (measured in MET-min/wk)?
Potential confounders: Age, race/ethnicity, pre-pregnancy body mass index (BMI), average self reported hours of pre-pregnancy physical activity, depression during pregnancy, marital status, education, income, employment early in pregnancy, employment change during pregnancy, number of children in the home, and nausea and vomiting during pregnancy, gestational weight gain category, postpartum weight retention
Life course model: Piece wise mixed effects model (as described in Naumova et al 2001).
Regression model: Yij = β0 + β1tijδij + β2tij (1 – δij) + b0i+ b1itijδij + b2itij (1 – δij) + eij
fixed effects: β0, β1, β2 - examples would include age, race/ethnicity, and education
random effects: b0i - intercept, b1i- slope of PA pre pregnancy, b2i - slope of PA post pregnancy
tijδij and tij (1 – δij) are the pre- and post- event times
Possible datasets: The CARDIA or MESA datasets. The concerns would be that it's hard to capture accurate the PA levels pre pregnancy given that the intervals of questionnaires are fixed so each participant will have varying intervals between PA level measurement pre and post exposure. Also, this is not the main focus for CARDIA so a lot of the potential confounders around pre and post pregnancy would not be available (eg post partum depression).
Exposure: Sexual abuse in childhood (binary)
Outcome: Risk behavior for HIV infection in the adult life (measured by mean number of unprotected intercourses in a given period of time)
Lifecourse model: critical period model, with later life risk factors
The idea here is that exposure to sexual abuse in childhood (critical period) may influence the adult sexual risk behavior through increased incidence of depression, lower self-esteem, and reduced self-care.
Regression model: linear model (as defined in the Mishra paper) including a term for the critical period exposure, and additional terms for later life course risk factors, such as gender, age, race, SES.
E(Y) = α+β1X1+β2X2+β3X3+β4X4+β5X5
β1=coefficient for sexual abuse in childhood
β2=coefficient for gender
β3=coefficient for age
β4=coefficient for race
β5=coefficient for SES
Possible dataset: the IPrEx trial and the IPrEx OLE cohort have assessed these variables among men and transgender women. However, they only included persons with high risk for HIV infection, and did not include women. We would still be able to analyze if sexual abuse in childhood is associated with higher risk-behavior levels among men with high risk of HIV acquisition, adjusted for age, race and SES.
Interesting example but I am confused about the exposure here - is the idea to compare women who were pregnant to women who were never pregnant? Or each woman after pregnancy to herself prior to pregnancy?
Classic lifecourse epi formulation would be something like "do exposures during pregnancy have disproportionate impacts on subsequent physical activity, compared to the same exposure when a woman was not pregnant?"
EXPOSURE = air pollution from unconventional natural gas development (UNGD)
OUTCOME = respiratory infection in mothers and later in children
HYPOTHESIS = (1) Women with more exposure to UNGD will have more respiratory infections; (2) pregnancy is a sensitive period where the effect of UNGD will be even stronger, so compared to the time before and after pregnancy women with more exposure to UNGD will have more respiratory infections; and (3) women of low SES will have more respiratory infections at the same levels of exposure as women of middle or high SES.
LIFECOURSE MODEL = Critical period model with effect modifiers (critical period = pregnancy and modifier = SES). This might be a difficult hypothesis to test since women are likely to be changing quite rapidly prior to, during, and after pregnancy. They might have more health-seeking behaviors or less. They might be more health-conscious or extremely sleep deprived and thus more likely to get an infection. I would also be concerned about my ability to appropriately parameterize the longitudinal poisson model with multiple interactions.
DATA SET = Geisinger Health electronic health records from 2005-2013 (n = 16,000 births in a region with unconventional natural gas development beginning in 2005 and accelerating exponentially through 2013)
REGRESSION MODELS = Longitudinal poisson (or negative binomial, depending on dispersion) regression with indicator variables for 0-9 months prior to pregnancy, pregnancy, and 0-9 months post-pregnancy, exposure variable, three-way interaction between pregnancy time variable, exposure, and individual SES and random intercept for mother (multiple observations within mother are correlated).
Exposure/Outcome: Trachoma is the world’s leading cause of preventable blindness and occurs in poverty stricken areas with limited access to water and health care. Repeated ocular Chlamydia infections in children aged 1 month to 12 years can eventually lead to blindness later in life (~40 years old). The disease progresses over years as repeated infections of Chlamydia trachomatis cause scarring on the inside of the eyelid, the eyelashes eventually turn in which causes rubbing on the cornea at the front of the eye (trachiasis).
A proposed regression model could be the Accumulation of risk model where, infection-period represents the “time a child has the infection”. In a model like this we would assume that hose children who have 7 infections would have a greater risk than those who only had 3 over the course of their childhood. We would assume the outcome is independent of when the child had the infection (I would need to talk with a clinician about this assumption in that, it may be true that younger children have better healing capabilities, but based on my knowledge, it seems like all kids heal quickly if they are under 12).
Our model would be defined as:
E(Y)=α + β, where Sj represents the 1, 2, …j infections, β represents the length of the infection and α represents baseline risk of blindness.
Data Set: The proctor foundation has done trachoma studies over the years in the Gurage area of Ethiopia. I am aware of data that was collected on yearly basis from 2003-2006 but there is no person identity linked to the data, children were randomly sampled. Additionally we do not have the outcome data, nor the length of infection. It might be interesting to go back to these villages (36 villages, 50 children randomly sampled annually) and check the prevalence of trachoma and or blindness to estimate alpha and possibly beta, but it might be a reach.
