Post on 18-Mar-2016
description
Applied Epidemiologic AnalysisFall 2002
Applied Epidemiologic Analysis
Patricia Cohen, Ph.D.
Henian Chen, M.D., Ph. D.
Teaching Assistants
Julie Kranick Sylvia TaylorChelsea Morroni Judith Weissman
Applied Epidemiologic AnalysisFall 2002
Lecture 5
Examination of conditional (interactive) effectsand non-linear effects in multiple regression
Goals:• To examine studies in which interactive effects were examined among categorical or scaled exposures or other predictors of scaled disease indicators•To understand how some non-linear relationshps may be handled by re-scaling while others are intrinsically non-linear
Applied Epidemiologic AnalysisFall 2002
Objectives
• To appreciate two reasons for conducting analyses of interactive and non-linear relationships: – as assumption checks– as primary study questions
• To understand the methods and statistical indices used to investigate interactive and non-linear relationships
Applied Epidemiologic AnalysisFall 2002
Note re: strata and conditional effects
When analyses of relationships between risk and outcome are reported within strata, it may be possible to see conditional effects directly.
Applied Epidemiologic AnalysisFall 2002
Effects expressed as means of exposed and not exposed
• But, watch out for effects of differences in exposure or outcome rates on the magnitude of some effect indices.
.10 .30
.15 .32
Male Female
Young
Middle
.18Older
Not exposed Not exposedExposed Exposed
.40.15
.11
.16.31
.45
.48
Applied Epidemiologic AnalysisFall 2002
First Study : Role of genotype in the cycle of violence in maltreated children
Reference:
Caspi A, McClay J, Moffit TE, Mill J, Martin J, Craig IW, Taylor A, Poulton R, Science, 297, No 2, 851-854, 2002.
Primary goal: Determination of a possible gene – environment interaction
Applied Epidemiologic AnalysisFall 2002
The problem: Role of genotype study
Do maltreated children show a differential tendency to later antisocial behavior as a function of identifiable genetic differences?
Is the effect of child maltreatment conditional on genotype?
Is the effect of genotype modified by environmental conditions (child maltreatment)?
or, put another way
or, put another way
Applied Epidemiologic AnalysisFall 2002
Population studied, study design, and sample size : Role of genotype study
442 non-Maori New Zealand males who were part of a cohort studied from birth to age 26
Applied Epidemiologic AnalysisFall 2002
Measurement issues: Role of genotype study
Details of genotyping of DNA are provided in Supplemental material available on the Science website.
The measure of childhood maltreatment included:• observations of negative affect or behavior toward
child at age 3• harsh discipline at ages 7 or 9 by parent report• 2 or more changes in primary caretaker by age 10• retrospective adult report of physical or sexual child
abuse.
Cumulative index counted the number of these measures that were positive
• 64% = 0• 28% = 1 (called probable)• and 8% = 2 or more.
Applied Epidemiologic AnalysisFall 2002
Measurement issues: Role of genotype study
Genotype measure:• Fewer or more replicates of MAOA allele
Antisocial behavor: • Conduct disorder in adolescence (age 11,13,15, or 18)• Court record of conviction for a violent offence• Self reported age 26 tendency toward aggressive acts• Informant report at age 26 of symptoms of antisocial
personality disorder
Stratifiers: none
Applied Epidemiologic AnalysisFall 2002
The effect being estimated: The conditional effect of maltreatment depending on a polymorphism in the MAOA encoding gene
Basic analysis to answer study questions:Multiple linear regression analyses:
• Predictors: MAOIA, Maltreatment history, MAOIAxMaltreatment product
• Dependent variables: Indicators of antisocial behavior
Applied Epidemiologic AnalysisFall 2002
Selection and inclusion of confounders in the analysis
Examined IQ and family social class as potential confounders.
Applied Epidemiologic AnalysisFall 2002
What would the model be?
IQ and family social class as potential confounders.
MAO-A
IQ
Family socialclass
Child maltreatment
Aggression
Applied Epidemiologic AnalysisFall 2002
Table S2. Results of final regression analyses testing G ラ E interaction effects onantisocial outcomes. The table presents final models with main effects and interactions entered simultaneously. Childhood maltreatment was handled as a single quantitative variable in the regression analyses, ranging from no maltreatment to severe maltreatment.
Applied Epidemiologic AnalysisFall 2002
Composite indexof antisocial behavior (z scores)
Low MAOA activity
High MAOA activity
Applied Epidemiologic AnalysisFall 2002
Conclusion: Role of genotype study
• A functional polymorphism in the MAOA gene moderates the impact of early childhood maltreatment on the development of antisocial behavior in males.
• “Men having the combination of low-activity MAOA genotype and maltreatment were 12% of the male birth cohort by accounted for 44% of the cohort’s violent convictions, yielding an attributable risk fraction (11%) comparable to that of the major risk factors associated with cardiovascular disease. Moreover, 85% of cohort males having (this) genotype who were severely maltreated developed some form of antisocial behavior.”
Applied Epidemiologic AnalysisFall 2002
Second Study : Poverty, family process, and the mental health of immigrant children in Canada
Reference:
Beiser M, Hou F, Hyman I, & Tousignant M. American Journal of Public Health, 92, No 2, 220-227, 2002.
Applied Epidemiologic AnalysisFall 2002
The problem: Child mental health
The apparently good mental health of immigrant children in combination with the poverty that often characterizes immigrant families presents a paradox.
Do family variables account for this protective effect?
Applied Epidemiologic AnalysisFall 2002
Population studied, study design, and sample size : Child mental health
National cross-sectional survey of Canadian children ages 4-11, including:
– 684 immigrant children – 2573 offspring of immigrant parents– 10,092 offspring of non-immigrant parents.
Applied Epidemiologic AnalysisFall 2002
Measurement issues: Child mental health
Stratifiers: • Because
– the large samples being employed made every test of interaction significant as virtually a fore-gone conclusion
– the relationships with and among covariates were also different for the different groups
• Therefore, the 3 groups of children were treated as separate strata and their estimates from the regression equations compared in the text.
Used measures with previously reported reliabilities and validities.
Applied Epidemiologic AnalysisFall 2002
The effect being estimated: a: Differential effects of poverty on mental health indicators in native born children of non-immigrant parents, children of immigrant parents, and immigrant children.
b: Whether these differences are attributable to differences in family functioning.
Basic analysis to answer study questions:Multiple linear regression analyses of each group:• Dependent variables:
– Emotional problems– Behavioral problems in offspring
• Basic IV (“exposure”):– poverty
Applied Epidemiologic AnalysisFall 2002
Selection and inclusion of predictors in the analysis
• Control variables (Set 1) included – age– sex– length of stay in Canada– race/ethnicity– single parent
• Potential mediators (variables standing between or explaining the relationship between the exposure variable and the disease) – ineffective parenting– parental depression– family dysfunction.
Applied Epidemiologic AnalysisFall 2002
Applied Epidemiologic AnalysisFall 2002
Applied Epidemiologic AnalysisFall 2002Note: constant units across samples
A model of these results
PovertyEmotional problems
Single parent
Ineffectiveparenting
Familydysfunction
Parental depression
3.12
2.52
1.56
52.2%
3.20
.07
15.6% .02
.10
.00
.21
.62
.52
.02
.12
.14
.17
1.42
Immigrant childrenChildren of nonimmigrant parents
Applied Epidemiologic AnalysisFall 2002
Models shown in paper
• The models shown in the paper use standardized coefficients to represent causal effects.
• Is this a problem?
Applied Epidemiologic AnalysisFall 2002
Conclusions: Child mental health
Family factors played a relatively weak role with foreign-born children.But for majority-culture children, material deprivation is a less significant threat to mental health than the familial concomitants of poverty.
Applied Epidemiologic AnalysisFall 2002
Third Study : Analysis of 13 P-DNA postlabeling studies on occupational cohorts exposed to air pollution
Reference:
Peluso M, Ceppi A, Munnia A, Runtoni R, Parodi S. American Journal of Epidemiology, 153, No 6, 546-558.
Applied Epidemiologic AnalysisFall 2002
The problem: DNA adducts
Guidelines for exposure to airborne pollutants resulting from combustion of organic matter (as in industry, residential heating, and motor vehicles) are based on quantitative studies of air pollution and DNA adducts in the exposed. An accumulated literature needs to be combined to give the best estimates of the nature of these relationships.
Applied Epidemiologic AnalysisFall 2002
Population studied, study design, and sample size : DNA adducts
A meta-analysis based on 36 sets of exposed-referent pairs of observations in 13 studies.
What is a meta-analysis?
How can we generate dose-response curves from a meta-analysis?
Applied Epidemiologic AnalysisFall 2002
Measurement issues: DNA adducts
Used exposed worker:referent ratios rather than absolute DNA adduct levels, since large inter-laboratory differences may be present in terms of number of adducts.
Measures of pollution using a specific protocol produced relatively comparable across-studies measure of polycyclic aromatic hydrocarbons (PAHs).
Response:
Dose:
Applied Epidemiologic AnalysisFall 2002
Selection and inclusion of confounders in the analysis
Discussed but not included in dose-response analyses (e.g. smoking, qualitatively different sources of pollutants)
Applied Epidemiologic AnalysisFall 2002
The effect being estimated: The dose-response nature of the relationship between measures of air pollution and DNA adducts
Basic analysis to answer study questions:• Multi-level linear regression analyses (taking into account
the cluster-effect of the study from which the subject data came):
• Dependent variables: – DNA adducts in white blood cells (ratio exposed/control)
• Basic IV (“exposure”): – PAH indicator B(a)P
Applied Epidemiologic AnalysisFall 2002
FIGURE 4. Dose-response relation between frequency ratios and external benzo[a]pyrene (B(a)P) concentrations in work environments in ameta-analysis of occupational exposure to air pollution (13–25), as predicted from equation 6. The inset shows an extrapolated dose-responsecurve at low exposure doses, assuming a linear dose-response relation, for B(a)P levels between 0 and 4.5 ng/m3, the lowest value in the database.FRi, frequency ratio for the ith study.
Applied Epidemiologic AnalysisFall 2002
Conclusions: DNA adducts
“levels of DNA adducts in exposed workers, with respect to referents (controls), were associated with B(a)P levels,..and that the relation of the dose-response curve was sublinear in heavily exposed industrial workers.”