RHMS - lecture 8 - risk factors versus prediction of outcome

NEED TO KNOW;

• basics observation studies

• measures of association

  • risk difference (absolute risk reduction)

  • relative risk

  • odds ratio

  • number needed to treat

  • attributable proportion in the exposed (APe)

  • attributable proportion in the (total ) population (APt)

• selection bias and information bias, condounding and effect modification EXAM MATERIAL

• difference between studying risk factors versus the development of prediction models (main part this lecture)

case control study

• researcher looks backward in time to see how frequently the exposure to a potential risk factor occurred in each group

• cases → individuals who already have the disease or outcome of interest

• controls → individuals who do not have the disease, but are otherwise similar to the cases.

• Usually retrospective, but can be prospective in some designs

• measure of association → odds ratio, often not risk ratio

information bias

• systemic errors in the way data are collected, measured, recorded 

• common types of information bias

  • recall bias → participants remembering past events differently 

  • interviewer bias → interviewers tone, wording, behaviour influence response 

  • response / social desirability bias → participant answer in desirable way

pros cons case control studies

pros;

• Rare ‘disease’ (occurrence of interest)

• Efficiency

• Relative small N (With same precision of effect estimate as in cohort study)

cons;

• Validity comprised and not easy established

  • Selection bias (wrong control group or wrong selection cases), information bias

  • Confounding & Effect modification

cohort studies

• follow group of people over time to see how exposure to a risk factor affects the occurence of a specific outcome

• start with exposure → classify participants based on whether they are exposed or not exposed → follow them over time → compare incidence

• two types of cohort study

  • prospective → healthy population, collect exposure data, follow participants

  • retrospective → use existing records to classify participants by past exposure → look at outcomes that have already occured

• summary → cohort studis assess cause - effect relationships. but often the causality as such is difficult to establish therefore epidemiologists often use association and cause-effect relationships.

pros cohort study

• measurement at the individual level

• follows the natural sequence (exposure → outcome) that is why it is longitudinal and prospective

• more outcomes in one study (can study more diseases like diabetes and cardiovascular disease)

cons cohort study

• costly and long duration

• potential pre-selection due to early exposure (healthy worker effect)

• difficult when incidence in outcome is low or if it has a long preclinical stadium

• clear insight into relevant risk factors needed. 

survivor bias

• survivor bias happens when we only study those who remain

• loss to follow up is a big cause of survivor bias → if only participants who remain in the study are analyzed, the results may be biased.

• people lost to follow up may differ from those who stayed.

• because of this you can never ignore missing data

limitations patient registries

• data quality (clinical registries vary in how accurately and consistently data are entered can lead to information bias)

• missing values → mainly confounders

• reasons;

  • these registries were not designed for scientific research

  • in routine clinical practice everything that is additional to direct patient care should be short and simple.

aim cohort study

• cohort study → observational, longitudinal study where a group of individuals is followed over time to observe whether they develop a specific outcome. 

• cohort study aims to examine the association between a (set of) main determinants and a future outcome.

• one wants to have an unbiased association so control for confounding is important.

effect modification

• occurs when the strength or direction of the association between an exposure and an outcome changes depending on the level of a third variable (the effect modifier).

• when the relationship / effect between an exposure and an outcome differs across levels of a third variable

• e.g. difference between man and women because they are this certain sex.

cohort studies for prediction models

• which variables predict the outcome

• to predict an occurrence of interest in the future we need longitudinal data

  • longitudinal data → consists of repeated measurements on the same subjects over multiple time points. 

• so, cohort studies are appropriate designs

validation

• To determine to what extent the predictions / model is feasible or (too?) optimistic

• Internal validation:

  • Validate in the setting / patients used to make the model

• External validation

  • Validate in a new setting / new patients

why is prediction important

• patients and clinicians want to know the prognosis 

• prognostic stratification 

  • guides further clinical management 

  • more specific diagnostic procedures

  • treatment or not 

• causality is not the main issue 

how to develop prediction models

• pre selection of potential predictors

  • literature / previous research 

  • expert opinion 

• collect data in longitudinal study 

  • cohort 

  • registry data 

• selection of predictors statistically 

  • many discussions on how to do this 

explain versus predict

what is the different between an OR of 1 between explanation and predictor;

• in prediction confounder is not important anymore (causality). not interested in causality.

• in the top one when the OR is 1 than there is no association → important finding!!

• in the predictor when the OR is one than it does not predict anything (so you can delete it out of your study).

TRIPOD statement

• Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD)