Methodological foundations of psychiatric epidemiology

Sandro Galea, MD, MPH, DrPH Chair, Department of EpidemiologyAnna Cheskis

Gelman and Murray Charles Gelman Professor of Epidemiology

July 24, 2014

Intended Audience & Learning Objectives

This lecture will be most informative for someone with a beginning-to-intermediate knowledge of the topic. With this in mind, by the end of this lecture, users will be able to:

• Define the concepts of outcomes, causes, exposures, and risk factors

• Recognize different theories of causation• Recognize and define different types of epidemiologic study

designs• Define odds ratios, relative risk, incidence rates, and population

attributable proportion

How do we assess mental health?

• One of the objectives of the study of epidemiology is to measure a population’s health state and to explore potential reasons, or causes, of these states of health or individual events

• Valid statistical methods are used whenever possible in epidemiology to provide robust measures of both the occurrence or distribution of outcomes and the association of these outcomes with potential exposures

What is an exposure? • A possible cause of disease or outcome

that is being investigated

````

Social and Economic Policies

Neighborhoods and Communities

Institutions

Living Conditions

Social Relationships

Individual Risk FactorsGenetic/Constitutional Factors

Environme

ntIndividual/Population Health

Pathophysiologic pathways

Life

cours

e

Exposures

Outcome

What is a risk factor?

• An exposure or other variable, either at an individual-level or population-level, that is associated with an increased risk of the outcome

• A risk factor might also be called a determinant, but it is not necessarily causal, even if there is an observed association with the outcome

What is an outcome?

• The disease, event, or state of health that an epidemiologist is trying to understand or predict using potential risk factors

• Example: depression in the past 12 months

Job loss

DepressionGenetic

susceptibility

How are exposures related to outcomes?

ExposedUnexposed

Some people are exposed

Both exposed and unexposed can have the outcome/disease

ExposedUnexposed

What is a cause?

• An antecedent event, condition, or characteristic that is necessary for the occurrence of the outcome

• Example: a potentially traumatic event is necessary for the occurrence of post-traumatic stress disorder

Essential properties of a cause

• Association: the co-occurrence of the exposure and disease

• Temporal priority: The exposure is present before the disease in order

• Sole plausible explanation: only possible explanation after examining alternatives (confounding, misclassification, etc.)

Sufficient vs. necessary cause theory

• Casual partners: two or more risk factors that are involved in a “causal pathway”

• Sufficient: not necessary to get the disease but can still cause the disease

• Necessary: Any risk factor that is a causal partner in all sufficient causes; the disease will not occur without it

Sufficient vs. necessary cause theory

Causal pies: each pie is sufficient for the outcome. “A” is a necessary cause because it is present it every pie; the rest

are not

Observed

Counterfactual (parallel universe)

The counterfactual approach

Outcome

Observed

Counterfactual (parallel universe)

The counterfactual approach

Outcome

Observed

Counterfactual (parallel universe)

The counterfactual approach

Outcome

• Mental health example: if childhood abuse was a risk factor being investigated for adult depression, what would happen if we were to remove childhood abuse in someone’s life? Would depression still occur?

• If depression would not occur without this exposure, then the exposure might be considered a cause

The counterfactual approach

• The chronic disease era necessitates more complicated frameworks than the infectious disease era

The web of causation

• Confounders: variables that co-occur with the exposure of interest, and contribute to the non-comparability of the exposed and unexposed

• Mediators: risk factors that link the exposure of interest to the disease

• Causal partners/interactions: one cause combining with another cause to create the outcome

How do we assess joint effects of multiple exposures?

How do we assess joint effects of multiple exposures?

• Example of a confounder

Study designs: How do we use all of this information to study

causes? Take a sample of the population

Use sample to conduct analyses

Non-diseased

Diseased Total

UnexposedA B A+B

ExposedC D C+D

TotalA+C B+D N

Study designs: How do we use all of this information to study

causes?

• Longitudinal studies:• Interventions• Cohort studies

• Case-control studies/ cross-sectional prevalence studies

Study designs: How do we use this information to study

causes?

Longitudinal studies

Wave 1 sample

Wave 2 sample

Attrition:Subjects may drop

out of study or become

unreachable

Types of longitudinal studies

• Randomized controlled trials (subjects who are comparable on many different factors are randomly assigned to an exposure group and then followed to assess outcomes)

• Positives: Standard for good comparability• Best scenario: double-blind trials (neither

the researchers nor the participants know whether they are assigned to an exposure or not, to decrease the chance of bias)

• Negatives: expensive, time-consuming

Interventions

• Example: the World Health Organization’s Health-Promoting School framework

• Cluster-randomized controlled trials: Randomization at the level of school, district, or other geographical area

• Interventions included input to the curriculum; changes to the school's ethos or environment or both; and engagement with families/communities (these are considered exposures)

• This intervention was compared against schools that implemented either no intervention or continued with their usual practice (no exposure; considered controls)

• Conclusion: Intervention effects were generally small but have the potential to produce public health benefits at the population level

Interventions

• Considered a type of observational study; the investigator does not have control over who is exposed vs. not exposed, but rather observes the outcome of a series of events and gathers information on exposures and risk factors

• Positives: temporality is still relatively easy to distinguish using incidence, a measure of new onset of disease, since study subjects are followed over time

• Negatives: attrition; subjects may drop out

Cohort Studies

Cohort Studies• Example: Millennium Cohort Study: U.S.

soldiers assessed for health status every three years

• Also observational, but not longitudinal; a “snapshot” of a moment in time for measuring prevalence of certain outcomes or states of health

• Negatives: achieving a representative sample of the population is difficult; simple random sampling is preferred

• Problematic for causal inference – not usually used to investigate specific risk factors

• Positives: relatively cheap and easy to conduct

Cross-sectional studies

• Example: a population-based epidemiological study of Singapore residents taken at one time

• Aims are to assess the distribution of different types of mental illnesses across different ethnic groups, and to develop and validate a tool for the assessment of positive mental wellbeing for the Singapore population

Cross-sectional studies

• An observational study in which two groups of subjects are chosen on the basis of their outcome: a group who have certain outcome (cases) and a group of people who do not (controls)

• Positives: Require less time and smaller sample sizes than cohort studies

• Negatives: more difficult to simulate full comparability and to discern temporal order

Case-control studies

• Example: A study assessing characteristics associated with a history of suicide attempts among psychiatric outpatients

• Subjects: 154 suicide attempters (cases) and 122 patients without suicide attempt history (controls) who attended the two public hospitals in Durango City, Mexico

• Socio-demographic, clinical and behavioral characteristics were obtained retrospectively from all patients and compared in relation to the presence or absence of suicide attempt history

Case-control studies

Some key concepts

• For cohort studies:• Magnitude of an association between an

exposure and an outcome• Likelihood of developing the outcome in

the exposed group compared to the non-exposed group

• Incidence of the outcome in the exposed group divided by incidence of disease in the non-exposed group

What is relative risk?

• An estimate of the relative risk for case-control studies

• Ratio of the odds of exposure among cases compared to that among controls

• Example: in the case-control study example above, living in an urban residence was associated with a 2.3 times greater likelihood of having attempted suicide

What is an odds ratio?

• The excess rate of disease/outcome in the study population of both exposed and non-exposed individuals that is attributable to the exposure

• Helps to determine which exposures have the most relevance to the health of an overall population

• Rate of disease in the entire sample minus the rate in the unexposed group

What is population attributable risk?

• Example: a study of sleep disturbances in Japan estimated that the population attributable risk percent of suicide associated with sleep disturbances and mental disorders respectively were 56.4% and 35.3%

What is population attributable risk?

• Must specify desired values for the probabilities of type I error (rejecting a null hypothesis when it is actually true) and type II error (failing to reject the null when it is not true)

• The power is the probability of rejecting the null and concluding a statistically significant difference between the study groups, where one actually exists (1 – type II error)

Estimating the power of a study

• Internal validity: Is the observed association valid? Could an alternative explanation come from chance, bias, or confounding?

• Generalizability: are results applicable to populations outside of the study sample?

• Generalizability is not possible without validity

Extrapolating from studies

• If a study does not have a truly representative sample of the population, there is a potential for selection bias

• For example, if selected subjects in a case-control study do not accurately represent the distribution of potential risk factors of the population they are chosen from, the prevalence of exposure among one group may be artificially higher than that of the other group

Challenges in Psychiatric Epidemiology

• Disease or exposure misclassification

• Example: study subjects may under- or over-report outcomes or outcomes (ex: recall bias), or clinicians might search harder for an outcome among exposed subjects, whether they are aware of the bias or not

• Therefore, it is best for the assessment of an outcome to be done “blind” if possible (see slide 26)

Challenges in Psychiatric Epidemiology

• Attrition in longitudinal studies (see slide 24)

• Attrition can be either differential (for example, if depressed people drop out of a study more often, one may see lower levels of mental health problems in follow-up sample than there are in the population they are representing), or random, if the attrition is not at all associated with either the exposure or outcome

• Sensitivity analysis can help one determine whether attrition has a significant effect on the findings

Challenges in Psychiatric Epidemiology

Conclusions

• An exposure is a possible cause of an outcome that is being investigated

• Determining whether a factor is truly causal requires multiple studies and approaches

• Two main types of epidemiological studies are longitudinal (from which one can calculate relative risk) and cross-sectional (from which one can calculate odds ratios)

• Methodological challenges include selection bias, misclassification, and attrition

Helpful references

• Evans AS. Causation and disease: A chronological journey. Springer 1993

• Galea S, Riddle M, Kaplan G. Causal thinking and complex system approaches in epidemiology. Int J of Epidemiol 2010

• Hennekens CH, Buring JE. Epidemiology in Medicine. Lippincott Williams & Wilkins, 1987

• Krieger N. Epidemiology and the web of causation: Has anyone seen the spider? Soc Sci Med 1994

• McMichael AJ. Prisoners of the proximate: Loosening the constraints on epidemiology in an age of change.” American Journal of Epidemiology 1999

• Rothman KJ, Greenland S, Lash TL. Modern Epidemiology: Third edition. Lippincott Williams & Wilkins, 2012

• Rockhill B. Theorizing about cases at the individual level while estimating effects at the population level. Epidemiology & Society 2005

• Susser E, Schwartz S, Morabia A, Bromet EJ. Psychiatric Epidemiology: Searching for the causes of mental disorders. Oxford University Press, 2006

• Susser M. What is a cause and how do we know one? A grammar for pragmatic epidemiology. American Journal of Epidemiology 1991