Correlation or Causation

Correlation or causation measurement for medical based studies • A correlation between variables does not automatically mean the change in one variable is the cause of the change in the other variable • Causation indicates one event is the result of the occurrence of the other event; i.e. there is a causal relationship between the two events • Cause: an event, condition or characteristic (or a combination of these) that has an important role in bringing about a disease or other outcome Causation and causal inference: sufficient-component cause model • Kenneth Rothman’s (1976) sufficient-component cause model • A sufficient cause is a minimum set of factors that will produce disease (a whole pie) • The pie is made up of components (slices) • There may be different types of pies consisting of different components for the same disease, i.e. different mechanisms that lead to a heart attack • A necessary component is one that has to be present for the disease to occur Causal pie 56 Bradford Hill ‘considerations’ • Another way to think about causality • Austin Bradford Hill 1965: An association is more likely to be causal when: 1. Strength 🡪 there is a strong association (as measured by the size of the risk ratio) o A weak association can still be causal while a strong association may not 2. Consistency 🡪 a relationship is observed repeatedly e.g. in different study designs o Different populations may have different responses and other factors may change the effect size o Size of an effect varies between populations and over time 3. Specificity 🡪 a single cause produces a single effect o Absence of specificity does not negate causal effect, many diseases have multiple risk factors. E.g. an exposure such as smoking can have multiple disease outcomes 4. Temporality 🡪 intuitive that a cause should precede an effect o Not always easy to measure in practice, especially if a disease develops slowly E.g. Vitamin D and cancer 5. Biological gradient 🡪 dose-response relationship exists between the variables, i.e. y (the outcome) increases as a function of x (the exposure) but could be nonlinear or a threshold effect o E.g. ambient temperature and mortality 6. Plausibility 🡪 association can be explained in a way that is consistent with prior knowledge (e.g. biological explanation that makes sense) o Apparently implausible associations may in time prove plausible e.g. leaded petrol and crime rate 7. Coherence 🡪 no fundamental contradiction with present knowledge o New knowledge is ‘new’ and may challenge accepted theories. E.g. Malaria 8. Experiment 🡪 more likely to be causal if findings are from experiments o Usually interpreted now to include studies that address bias and confounding variables 9. Analogy 🡪 similar exposures may have similar effects What makes a good study? • Whether the study type used is appropriate for the clinical question 🡪 need to know: o Various study designs o Indication of each design o Inherent strengths and weaknesses of each design o The ideal implementation of each design Randomised controlled trial (RCT) • Chance alone determines exposure to the study factor • Outcome is measured later • Major strength: no difference between the exposed and unexposed group EXCEPT for the study factor 57 Cohort study • Exposure to the study factor is determined by subjects • Investigators measure the extent of the exposure • Outcome is measure later • MAJOR STRENGTHS: Clearer indication of temporal sequence as participants are disease free at the start of the study & can examine multiple effects of a single exposure • Subjects selected on presence [cases] or absence [controls] of the outcome factor • Then exposure factor is measured in cases and controls (after the outcome is known) • Relative frequency of the exposure in cases and controls is compared • MAJOR STRENGTH: good for examining causality in rare diseases Type of question informs the research study • Risk factor question o Is physical inactivity (risk factor) associated with depression (outcome) – or vice-versa? o Best answered by a case-control or a cohort study o What would be an advantage of a case-control study? o What would be an advantage of a cohort study? • Prognosis question o I have lung cancer (condition) – what are my chances of survival (prognosis)? o Best answered by a cohort study (inception cohort study) • Intervention question o Does taking a combination of calcium plus vitamin D work better to prevent fractures in elderly women than taking calcium alone? o Best answered by an RCT Critical appraisal • The systematic assessment of the quality of evidence presented in the literature, its relevance and applicability, in order to assist in making good decisions • Critical appraisal of the evidence (i.e. epidemiological studies) enables us to make informed, evidence-based decisions 🡪 answer questions about risk factors, prognosis, and intervention (treatment and prevention) • How do we know the results are useful? o Just because it is published doesn’t make it true o Sources of bias are different depending on study type – INTERNAL VALIDITY o Just because it is true doesn’t make it useful o Clinical relevance – EXTERNAL VALIDITY • General strategy: o Be systematic o All studies will have strengths and weaknesses o Look for key issues for the particular study type o Those aspects done well will be important strengths o Those aspects done poorly will be important weaknesses o Try to predict the size and direction of any bias • What to look at: o Research question (exposure & outcome) o Study type o Selection o Measurement (exposure, outcomes, confounders) o Findings o Generalisability/applicability to your patient ▪ NB: There is no such thing as a perfect study (practicality trumps perfection) - what their impacts are of the compromises made Research question • Did the study address a clearly focused question? • What is the study factor? • What is the outcome factor? • Is the question relevant your patient? Study type • What is the study type – is it appropriate to answer the study question? • Some issues more specific or relevant to particular study types, • Some relevant to all study designs Selection issues at the start • RCT and cohort – are the groups similar at the start of the study o Randomisation process (RCT) o Selection process (cohort) • Well defined study base (case control) o Cases representative of all cases (selected appropriately)? o Did all cases take part? Selection issues at the end • RCT and cohort are the groups similar at the end except for the exposure o Losses and intention to treat (RCT) o Losses (cohort) • Case control o Do the controls come from the same study base? o Are the controls representative of all controls (selected in an appropriate way)? o Did all selected controls take part? Population and sample • Study population • Various sampling methods (e.g. simple/stratified/cluster random, convenience, purposive, quota, snowball) • Selection bias, representativeness, generalisability/ external validity • Sample size, study power, sample size and power calculation, confidence level • Response rate Measurement • Exposure – measured accurately to minimize bias? • Outcome – measured accurately to minimize bias • Confounders: o all important confounders considered? o were confounders considered in design? • Key principles: o Was measurement blinded? o Objective, standardised, validated approaches used o Measured using the same person(s)/equipment/approach, or distribute subjects from different study groups equally between the various people/equipment/approaches • Look at: o The tool (questionnaire, blood analyser, microscope, assays, medical records) o The procedure (interviews, blood sampling, biopsy, medical records extraction) o The personnel (training, multiple collectors) o 🡪 Are they good enough for you to trust the study findings? Data collection • Validation (either in current or past study), internal validity, face and construct validity, accuracy, precision, gold standard, sensitivity, specificity, positive & negative predictive values, Likelihood Ratio • Reliability: standardisation, inter-rater, intra-rater, agreement, Kappa • Measurement bias, classification error, random error, systematic error Findings • What are the findings? • Are there confidence intervals or p values? • How precise are the results (are the confidence intervals too wide)? • Do you believe the results? • Were the methods appropriate? • Have the main potential confounders been controlled in the analysis? • Anticipate numerous statistical and qualitative jargon! • Often the issue is NOT the choice of procedure, but how well it was undertaken • When in doubt, consult an academic or biostatistician Data analysis • Measures of central tendency (mean, median, modus) and dispersion (range, SD, quartiles) • Statistical/significance tests: bivariate (Chi-squared, ANOVA, T-test and their variations and alternatives), multivariate (multiple linear/logistic regression). Less often: survival analysis, factor analysis. • Controlling/adjusting for confounders Main and secondary findings • Formulate your own one-sentence summary of the study findings, to avoid losing the wood for the trees • Use this one-sentence summary as the basis for your conclusion and patient education Significance of findings • Statistical significance: o Based on p values 🡪 the probability that the finding was purely by chance o In clinical research, generally agreed that we want there to be only a 5% or less probability that the treatment results, risk factor, or diagnostic results could be due to chance alone o Anything <0.05 statistically significant • Clinical significance: o Are the results big enough to matter to a patient? o 95% confidence interval expresses a range of values in the population that you are 95% confident to contain the “true” population value o Clinical significance may differ between the lower and upper limits of 95%CI! o Context really matters! Are the findings applicable to the patient? • What are the results, can they be applied locally? • To what extent is the study population similar (and different) to your patient? • Best evidence not always the most appropriate for all cases • Consider the wider context of o The healthcare setting o The patient • Do results of the study fit with other available evidence? No matter how good a study seems, remember: • One observational study rarely provides sufficiently robust evidence to recommend changes to clinical practice or within health policy decision making • Sometimes, however, they provide the only evidence • Findings are always stronger when supported by multiple sources of evidence • Even the most well designed and rigorously carried out study may not produce results that are important (even if significant statistically) or may not provide information that is relevant to your practice