CHP CHEAT SHEET SUMMARY Done by: Grace Lum 1 CONTENTS S/N 2 3 4 5 6 7/8 9 CONTENTS Descriptive studies Case reports, case series, ecological Analytical studies Cross-sectional, cohort, case-control Measures of disease frequency Nominal, ordinal scale Incidence, cumulative incidence Prevalence, point prevalence, period prevalence Incidence vs prevalence Rate Incidence density/incidence rate Cumulative incidence vs incidence density Describing variation in continuous data Mean, median, mode Long ordinal data Range, quartiles Variance, standard deviation Understanding and using interventional studies Uses Basic principles – test and control groups, comparability, unbiased measure Treats to validity – loss to followup, non-compliance, contamination Advantages Disadvantages Ethical issues Phases of trials Community trials Measures of risk and association 2x2 contigency table Relative risk (RR) / Risk ratio (for cohort study) Rate ratio Risk ratio vs rate ratio Attributable risk (AR) Relative risk vs attributable risk Population attributable risk (PAR) Population attributable risk perent (PAR%) Characteristics of cohort/survival study Kaplan-Meier estimate Logrank test Odds and odds ratio (for case-control study) Pearson correlation coefficient (r) Bias Sources of bias Classification of bias Bias, by studies PAGE 4 4 5 5 5 5 5 6 6 7 7 7 7 8 8 9 9 9 9 10 10 11 11 11 11 12 12 12 12 12 13 13 13 13 14 14 15 2 10/11 Assessing sampling errors Sampling distribution of sample means; Standard Error of the Mean Confidence interval Hypothesis testing (p value) Type I and II errors Tests of significance 12 Confounders Confounders – criteria, detection, measures against Effect modifiers Causality and criteria for establishing causality 13 Understanding and using diagnostic and screening tests True positive, false positive, true negative, false negative Sensitivity, specificity, positive predictive value, negative predictive value Sensitive tests vs Specific tests Predictive value Screening tests vs Diagnostic tests Guidelines for screening programmes Outcomes of screening programmes 14 Understanding and using meta-analyses Methods of summarising research findings Meta-analyses Forest plots Heterogeneity Publication bias Pooled analysis Limitations of meta-analysis Tutorial learning points Critical evaluation of study methodology Clinical evaluation of clinical trial Critical evaluation of diagnostic/screening test Clinical decision making (including meta-analysis) Guidelines on evaluation an overview/ systematic review/ meta-analysis 16 16 17 17 17 18 19 19 20 20 20 20 21 21 21 22 22 22 23 23 24 24 25 25 26 26 26 3 2 - DESCRIPTIVE STUDIES Study Details Case reports 1 or 2 patients; describe rare/unusual clinical events Advantages Highly detailed, can use sophisticated methods Disadvantages Cannot estimate frequency; there is bias and chance Case series 10-100 patients Ecological Study unit is population Absence of comparison group Cheap; can generate hypotheses Ecologic fallacy: factor and disease may not be related at individual level 3 - ANALYTICAL STUDIES Analytical studies are used to evaluate (casual) relationships between risk factor and disease. Study Details Advantag es Disadvan tages Cross-sectional At specific point in time Risk factor and disease measured concurrently Simple, rapid, inexpensive Can study wide range of factors Not good for rare conditions Cannot establish cause-effect relationship Length bias (study of survivors) – more likely to catch people who survive Cohort Population no outcome; ask exposure Follow-up and compare outcome between exposed and unexposed Least susceptible to bias Highest level of evidence for association among observational studies Can describe natural history of disease Calculate cumulative incidence/incidence rate Direct measure of risk Can tell cause-effect (temporal sequence) Not good for rare conditions Not feasible when long latent period between exposure and disease Case-control Start with cases and controls from same population base Capture new cases in study period Quick, less resourceintensive Can be used for rare diseases Can be used for diseases with long latent period Prone to bias (recall bias, selection bias) Cannot establish causeeffect relationship Does not measure risk directly, but calculates odds. The odds ratio is an approximation of the relative risk in exposed vs non-exposed. 4 4 – MEASURES OF DISEASE FREQUENCY Categorical variables NOMINAL SCALE o Mutually exclusive with no logical order o Eg. Gender, race ORDINAL SCALE o Mutually exclusive with logical order o Difference between ranks is not equal (subjective) o Eg. Cancer staging, anxiety score Measurement (continuous) variables Ratio – divide one quantity by another; no units Proportion – a ratio in which numerator is included in the denominator; no units Incidence – frequency of new occurrences of a disease, injury or death during period of study. Cumulative incidence – frequency of new cases over a specific period of time. o 𝐼 𝑁 𝐶𝐼 = o o o I = # of new cases during follow-up N = # of disease-free subjects at start of follow-up Most common (direct) way to estimate risk – probability an individual will experience a health status change over a specific follow-up period. o Disadvantages Does not reflect dynamic changing population Does not allow subjects to be followed for different time periods Prevalence – frequency of disease in a population at a point in time. o Cross-sectional study; no follow up Point prevalence – frequency of disease at given point in time. o 𝐶 𝑁 o C = # of observed cases at time t o N = Population size at time t Period prevalence o 𝑃= 𝑃𝑃 = 𝐶+𝐼 𝑁 o C = # of prevalent cases at beginning of time period o I = # of incident cases that develop during that period o N = size of population for same time period Incidence vs prevalence o P = I x D (duration of disease) o If incidence ↓, prevalence may ↑ due to treatment prolonging life. Rate – how rapidly health events are occurring in a population. Needs time (person-years, etc) 5 Incidence Density/ Incidence Rate – individuals in the population at risk studied for varying lengths of time o Incidence density rate = Number who get disease at specific period (I)/ Summation of length each person is at risk (PT) o I = # of new cases during follow-up o PT = total time that disease-free individuals in the cohort are observed over the study period Cumulative incidence Acute diseases with restricted risk period Fixed populations Direct measure of risk Incidence density Chronic disease with extended risk period Dynamic populations Can make comparisons between populations with different observation time Not a direct measure of risk 6 5 – DESCRIBING VARIATION IN CONTINUOUS DATA Measurement (continuous) variables o Infinite number of values; equally spaced o Eg. height, weight, temperature, BP o Ratio scale – 0 has no meaning (eg. height) o Interval sale – 0 is important (eg. Celsius) Measures of central tendency Mean Definition Advantages Disadvantages Average Widely used Median Exactly half Insensitive to outliers Mode Most frequent - Overly sensitive to outliers Less sensitive to actual numerical values Rarely used Long ordinal data o Ordinal data graded on long scale, may be treated as continuous data o Eg. pain on scale of 1-10 o But, not truly continuous data Finite number of values Gaps in continuum Spacing between categories is not numerically equivalent Measures of spread of measurement data Range Quartiles Variance 𝑠2 = ∑(𝑥 − 𝑥̅ )2 𝑛−1 Standard Deviation (s) Details Advantages Disadvantages Difference between highest and lowest observed values Interquartile range – middle 50% of data set; eliminate outliers Combines all values in data set to produce measure of spread Easy to compute Less sensitive to outliers Greatly influenced by outliers Difficult to calculate Square root of variance. Greater spread, greater SD. In a normal distribution, o 68% of scores are within 1 S.D. of mean o 95% of scores are within 2 S.D. of mean Eg. 95% of all the 1000 babies lie within 3.25 +/- (2xSD) kg Describes group you measured. Cannot be used for entire population – use standard error. o 99% of scores are within 3 S.D. of mean Skewness o Positive – scores cluster toward lower end of scale o Negative – scores cluster toward upper end of scale 7 6 – UNDERSTANDING AND USING INTERVENTIONAL STUDIES Uses of interventional studies o o Evaluate efficacy of Therapeutic agents/ surgical approaches Prophylactic agents Preventive measures Test aetiologic hypothesis and provide conclusive evidence of a cause-effect relationship Gold standard Randomised, double-blind, controlled clinical trial Basic principles Test and control groups o Appropriate selection – applicable to real life. Affects generalisability of results to all patients commonly seen in clinical practice. Type of cases (severity, pathology) Contraindications Ability and willingness to comply o Control group To ensure results not due to Natural history of disease Biological variation – ‘regression to the mean’ Effect of being studies – ‘hawthorne effect’ o Compared with respect to outcome Baseline comparability o Through randomisation Remove selection bias Equalises baseline differences – reduce confounders (maximises likelihood that both groups will be similar in characteristics that may influence outcome to treatment) Unbiased measured of outcomes o More than one end-point can be used. Must be scientifically valid, clinically relevant. o Determine side-effects o Masking (blinding); placebo Double-blind is gold standard Patient – prevent placebo effect (expectation causes change in status) Doctor – prevent selection bias, prevent data gathering bias Single blind – subject unaware Double blind – subject and observer unaware Triple blind – data analyst also unaware o Difference observed due to 8 Inherent differences in relevant characteristics of groups Chance variation Difference in way groups were managed, or differences in compliance True difference in effects of treatment Threats to validity Loss to followup o Drop out – affect validity of study and sample size If high dropout rate, problem! People who dropout may be different from people who stay in - bias Minimise: pre-test for compliance, provide incentives, frequent contact, measuring adherence via pill counts etc, not stretching trial for longer than necessary Non-compliance and contamination o Omit medication o Take medicine not intended for them o Intention to treat analysis In real life, will the treatment work? Considering that people may drop out or drop in in real life as well So, compare according to group randomised to, whether they take or not Advantages Known and unknown prognostic factors balanced between groups Blinding can reduce bias Effect of dosage can be examined Disadvantages Expensive Selection criteria may limit generalisability Long follow-up period may be required Compliance not assured Ethical issues Ethical issues Each of treatment options should be equally acceptable under current knowledge o Sufficient evidence and doubt of treatment’s efficacy o No treatment options should be known to be inferior to another based on previous randomised studies No patient should be denied appropriate treatment as a result of participation, or refusal to participate Risk-benefit ratio must be favourable o Potential benefits to subjects and to society justify the risks o Continuous monitoring; stop study if need be 9 Informed consent o Informed they are part of the study, made aware of treatment options, risks and benefits, nature of randomisation, and that they will be selected to receive any of the options. Phases Phase I = safety and dosimetry Phase II = preliminary information on efficacy Phase III = defininte evidence of efficacy Post marketing surveillance = detect uncommon side effects Community trials Entire community serves as experiment/control unit – eg. 3 city study For diseases influenced by wide range of factors like lifestyle, social behaviour, environment Limitations o Expensive o Small sample size o Possible contamination o Communities may not be comparable 10 7/8 – MEASURES OF RISK AND ASSOCIATION We are most interested in associations that are not necessary and not sufficient, but nevertheless associated in an average person. Table showing association between risk and disease occurrence Risk factor present Risk factor absent Total Disease present a (?%) Disease absent b (?%) Total a+b (?%) c (?%) a+c (?%) d (?%) b+d (?%) c+d (?%) a+b+c+d (100%) RELATIVE RISK / RISK RATIO (For Cohort Study) Cumulative incidence of disease among those with risk factor = a/(a+b) o This is the risk of developing disease if risk factor present Cumulative incidence of disease among those without risk factor = c/(c+d) o This is the risk of developing disease if risk factor absent 𝑎/(𝑎+𝑏) Relative risk = risk ratio = 𝑐/(𝑐+𝑑) Eg. Smokers are 2.1 times as likely to die from heart disease than non-smokers RATE RATIO How fast a disease occurs – faster it occurs, higher the risk. Indirect way to measure relative risk. Exposed Unexposed Total New cases I1 I0 I Person-time PT1 PT0 PT Average rate = I/PT o I = number of new cases o PT = person-time over follow-up Rate ratio = (I1/PT1) / (I0/PT0) Eg. Mortality rate among hypertensive is 3.5 times the mortality rate among normotensives RISK RATIO VS RATE RATIO Computes relative risk Cumulative incidence ratio = CIexposed/CIunexposed (RISK RATIO) Incidence density ratio = IDexposed/IDunexposed (RATE RATIO) Relative risk = ratio of risk/rate for one group to risk/rate of another group 11 ATTRIBUTABLE RISK / EXCESS RISK Risk difference o Risk in exposed – risk in non-exposed = a/(a+b) – c/(c+d) Measures excess risk on absolute scale Tells how much reduction in disease there will be if exposure is removed. Qn: How many cases of disease can you attribute to the exposure? Gives idea of health impact on society if you get rid of exposure - depends on how common the disease is Eg. If AR = 0.004, for 1000 smokers, 4 more oral cancers than 1000 non-smokers RELATIVE RISK VS ATTRIBUTABLE RISK Relative risk: measures strength and direction of an association Attributable risk: reflects potential public health consequences of an exposure POPULATION ATTRIBUTABLE RISK (PAR) Risk (total) – Risk (unexposed) = (a+c)/(a+b+c+d) – c/(c+d) Also equals AR x PF (prevalence of exposure in population [ (a+b)/(a+b+c+d) ]) Tells how much reduction in disease there will be if exposure is removed, in the population Takes into account proportion of risk factor in population Qn: How many cases of disease can you attribute to the exposure, in the population? Eg. If smoking is eliminated, the mortality of heart disease decreases by 9 per 100 patients POPULATION ATTRIBUTABLE RISK PERCENT (PAR%) Qn: What % of total risk of disease is due to the risk factor? PAR% = [ Risk (total) – Risk (unexposed) ] / Risk (total) x 100% Also, PAR% = (PF)(RR-1) / [1 + (PF)(RR-1)] x 100% o Where PF = prevalence of exposure factor Eg. If smoking is eliminated, 35% of deaths in population of patients with heart disease can be prevented. COHORT/ SURVIVAL STUDY Characteristics o Well-defined time of entry o Well-defined endpoint o Time in between = survival time 12 o Censored observation = no longer observable. (eg. still alive at study closure, death from other causes, or loss to follow up.) Time = censored survival time Kaplan-Meier estimate / Product-limit estimate o dt = number of events (death) at time t o nt = number of people at risk at time t (remove deaths and censored observations) Probability of survival at that point of time t = Pt = 1 − o Survival function S(t) is the probability of surviving beyond time t 𝑑𝑡 𝑛𝑡 o S(t) = (1 − 𝑑1 ) 𝑛1 (1 − 𝑑2 )... 𝑛2 (1 − 𝑑𝑡 ) 𝑛𝑡 Values should be taken each time an event occur (someone dies) Logrank test o Test of clinical significance o Compare survival distributions between groups of clinical interest o Survival distributions can be displayed using a Kaplan-Meier survival plot o Takes into account information provided by censored observations ODDS (For Case-Control Study) Odds = proportion/ 1-proportion = risk of having something / risk of not having something For case-control, must use odds as cannot find incidence, cannot compute risk or risk ratio Formula o Odds of lung cancer given smoking = (a/a+b)/(b/a+b) = a/b o Odds of lung cancer given non-smoking = (c/c+d)/(d/c+d) = c/d o Odds ratio = (a/b)/(c/d) = ad/bc Odds ratio from case-control study of rare disease is estimate of relative risk of the disease How to write: o Lung cancer patients have 4 times the odds of being smokers than those without lung cancer. The odds of lung cancer in smokers is 4 times the odds in non-smokers. o Fulfills the criteria The cases are incident cases (need to include all new cases over the entire time period; prevalent cases = bias) The controls come from same source in population that cases come from. The disease is rare (low incident rate). o Therefore, odds ratio is a good estimate of the relative risk. o Thus, the risk of lung cancer in smokers is 4 times the risk in non-smokers. PEARSON CORRELATION COEFFICIENT For linear relationship only. r varies from -1 to 1. o >0.8 very good; 0.5-0.8 good; 0.25-0.5 bad; <0.25 very bad Must eyeball data on scatter plot could be sigmoid or otherwise. If skewed measurement data (not normal) or ordinal data (eg. Cancer staging), use Spearman rank correlation coefficient r2 is the proportion of variability in y explained by the variability in x. o The higher the correlation, the greater changes in x can explain changes in y. o Eg. r2=0.25. 25% of variability in y is explained by the variability in x. 13 9 – Bias Any systematic error in design, conduct, or analysis of study that results in distorted estimate of exposure’s effect on risk of disease Any association can be due to o Casual relationship (true result) o Confounders o Chance (random error) o Bias (systematic error) Sources of bias Before study – literature review, study design During study – selection of participants, data collection After study – analysis, publication Classification of bias Bias Details Found in Solutions INFORMATION BIAS – Method for collecting information introduces error Exposure Recall bias Affected persons respond CaseUse cohort study identificat (participants) differently about prior control ion bias exposures compared to study unaffected persons. Affected over-recall Unaffected under-recall Interviewer Interviewers probe for exposure CaseMask interviewer; bias in cases control Train interviewers and tape (interviewers) study interviews; Standardise interviews Disease Surveillanc Some exposures will have more Identical procedures for identificat e bias medical surveillance, thus more identification of diseases ion bias likely to have subclinical disease (standardise doctor visits) diagnosed. Observer Observer who knows exposure Cohort bias status may be biased when study assigning disease outcome (for borderline cases) SELECTION BIAS – Biased estimation of cases or controls (case-control studies), or of exposed or unexposed (cohort study). Method of participant selection distorts exposure-outcome r/ship. Selection bias Case Cases are controls should Cases may not be from comparable represent all individuals control populations; cases should with disease of interest represent all individuals in community with disease; controls Controls may not should represent all represent all individuals individuals without disease without disease of interest in community Selection of cases and controls affect results 14 Admission bias Non-response bias OTHER BIASES Temporal bias Analytic bias Publication bias Admission rates of exposed and unexposed cases and controls differ – eg. more obese, more likely to be admitted People who respond are different from people who don’t respond – reflects attitude Hospital based case control Population-based casecontrol studies; Don’t release hypothesis to emergency doctor Do not know if exposure or outcome occurred first Crosssectional study Cohort study Only articles with positive results published Bias, by studies Case-control studies o Recall bias o Interviewer bias o Selection bias Cohort studies o Observer bias o Lost to follow up bias – individuals lost to follow-up different from those who are not o Non-response bias??? o Problems in measuring exposure Clinical trials (interventional studies) o Disease identification bias – observer bias o Lost to follow up bias – those lost may have different exposure-disease relationship compared to those who remain o Selection bias – if subjects are not randomised o Non-compliance bias – stop taking or take agent assigned to other group o Observer bias 15 10/11 – ASSESSING SAMPLING (RANDOM) ERRORS Truth = observed – errors Systematic errors – have pattern. Can do something to remove it if you can identify pattern. Random (sampling errors) – no pattern Indices of sampling errors Confidence interval (Null:1 for ratio; 0 for difference) o Preferred approach o More informative o Easy to understand Hypothesis testing o Less informative o Many variations/test o Obsession with p<0.05 = statistically significant Sampling distribution of sample means and standard error If systemic error minimal, difference between pop. mean and sample mean = sampling error. Unfortunately, in real life we may not know population mean. Thus, do sampling distribution of sample means – if you do alot of sampling, you’ll get population mean most of the time (normal distribution) Mean of the sampling distribution of sample means is population mean (μ). Standard deviation = measure of degree of sampling variation Standard deviation of sampling distribution of sample means = standard error of the mean (SEM) 𝑆𝐸𝑀 ≅ 𝑆𝐷 √𝑛 o To reduce sampling error, make n (sampling size) as large as possible 68% of all observations lie between μ +/- 1 SEM 95% of all observations lie between μ +/- 2 SEM o Ie. There is a 95% chance that the sample mean is within μ +/- 2 SEM o Eg. if SEM=0.5, sample mean = 3.23, we are 95% confident that μ is between 2.23 and 4.23 (95% confidence interval) 99% of all observations lie between μ +/- 3 SEM SD vs SEM o 95% variation of sample = mean +/- 2 SD o 95% confidence interval of mean of whole universe = mean +/- 2 SEM Confidence interval Eg. RR (95% CI) = 0.75 (0.64-0.87) I am 95% confident that the true population mean lies between 0.64-0.87 At best, there is a 36% reduction in risk At worst, there is a 13% reduction in risk reduction in risk still there! 16 Hypothesis testing (p value) p<0.05 – there is less than 5% chance that the observed difference is due to sampling error. o But, statistically significant does not mean clinically significant and vice versa o Confidence intervals are more informative o Statistical significant difference means that difference is unlikely to be due to random (sampling) error, but there could still be systematic error. Type I and II errors Type I (α) error say there is an association, when there is actually none. (False positive) o Reduce by ↓ p value. Type II (β) error say there is no association, when there is actually one. (False negative) o Reduce by increasing sample size. Tests of significance Comparing means / measurement data that is normally distributed o Comparing 2 unmatched means Unpaired t test o Comparing 2 matched means Paired t test o Comparing 3 or more means One-way analysis of variance Comparing medians / ordinal data, or measurement data that is not normally distributed o Comparing 2 unmatched medians Mann-Whitney U test o Comparing 2 matched medians Wilcoxon’s Sign Rank test o Comparing 3 or more medians Kruscal Wallis test Comparing proportions o Comparing 2 unmatched proportions o Comparing 2 matched proportions o Comparing 3 or more proportions Fisher’s χ2 test McNemar’s χ2 test χ2 test (Pearson’s) *p value lower for measurement data (parametric tests), higher for ordinal data (non-parametric) -- parametric tests compare actual values, non-parametric tests compare ranks (↓ differences) *p value lower for paired test (only considers variance of difference), higher for unpaired test (more variability due to intra-group variance) *p value based on sample size (small size, p value ↑) and variance (small variance, p value ↓) *Matched – the 2 sets of data are matched (eg. same people before and after exposure) Comparing more than 2 means – probably unmatched Do not perform multiple t-tests o Multiple tests each have 5% chance of sampling error 3 tests ↑ p value to 15% Perform post-hoc tests o Eg. bonferroni method – 0.017 per test p value overall still 5% o Other methods: Duncan, Tukey, Scheffe Genomic data: 10-7 17 12 – Confounders Criteria for confounder (X) X is a risk factor for disease X is associated with exposure X must not be in intermediate step in casual path between exposure and disease How to detect confounders Collect data on all possible confounders Evaluate association between possible risk factors (smoking) and confounders (age) o Eg. Is age related to smoking? Smokers are older as 50% of subjects above age 20 are smokers but only 10% of subjects less than age 20 are smokers Evaluate association between confounders (age) and disease (bladder cancer) o Eg. is age related to disease status? Older people have higher risk of bladder cancer as only 38% of individuals less than 20 years have bladder cancer, but 71% of individuals above 20 years have bladder cancer Stratify data by possible confounder (age) o For participants age <20 years, calculate odds ratio o For participants age >20 years, calculate odds ratio o Get Mantel-Haenszel ‘Adjusted Odds Ratio’. If this = 1.0 Thus, after stratification by the confounder (age), there is no relationship between smoking and bladder cancer for individuals within the age strata <20 years and for individuals within the age strata >20 years. Both stratified odds ratios are different from crude odds ratio, thus spurious relationship was explained entirely because of confounding by age. In addition, the Mantel-Haenszel ‘adjusted’ OR is different from the crude OR. Bladder cancer cases were more often smokers not because they were cases, but because they were older. Measures to control confounding Prior to data collection o Restriction – involve only those with certain characteristics (strata) (eg. old persons) o (case-control study) Matching – cases and controls matched by potential confounder, so that age distribution of cases and controls similar o (clinical trial) Randomisation – 2 groups have high chance of being comparable for confounders both known and unknown. After data collection o Stratification – evaluate association between exposure and disease for different strata of potential confounder 18 o o Direct and indirect standardisation – adjust for confounding in 2 or more populations by use of standard population or standard rates Multivariate analysis – sophisticated statistical techniques such as multiple linear regression and logistic regression analysis – allow control of several confounders simultaneously Effect modifiers Relationship between exposure and outcome is modified by a third variable We do not seek to control or eliminate effect modification Occurs when difference in risk/rate between those with and without the risk factor is not homogenous in strata formed by the effect modifier o When stratified by the effect modifier, odds ratios are different Confounder vs effect modifiers o Both involve 3rd factor o Both may be evaluated using stratified analysis o Confounding: stratification removes confounders. Compare crude and stratified ORs. o Effect modification: stratification does not change anything. Compare stratified ORs. o Can be confounder alone, effect modifier alone, or both Causality Approaches to studying etiology of disease o Animal models o In vitro systems o Epidemiologic studies in human populations Henle-Koch’s Postulates (criteria for establishing casual relationships between organisms and disease) o Organism always found with disease o Organism not found with other diseases o The organism, isolated and cultured through several generations, produces disease in experimental animals Criteria for establishing casualty (Hill’s criteria) o Temporal relationship: exposure precedes outcome in time o Biologic plausibility: consistency with existing knowledge from animal studies, etc o Consistency: findings comparable with other studies in different populations/circumstances o Strength of association: strong association (high odds ratio) has less bias o Dose-response relationship: as exposure increases, risk of disease increases o Specificity of association: single exposure leads to single outcome; a lack does not rule out causality o Effects of removing exposure: removal of exposure reduces risk of disease 19 13 – UNDERSTANDING AND USING DIAGNOSTIC AND SCREENING TESTS Test result Disease (defined using the gold standard) Present Absent Positive True positive (a) False positive (b) Negative False negative (c) True negative (d) Diagnostic tests are simpler, less expensive and less invasive than the gold standard test. Characteristics Sensitivity Specificity Positive predictive value Negative predictive value Definition Proportion of disease correctly classified (“positivity in disease”) Proportion of non-diseased correctly classified (“negativity in health”) Likelihood of disease in person with a positive test Likelihood of no disease in person with a negative test Formula 𝑎 (𝑎 + 𝑐) 𝑑 (𝑏 + 𝑑) 𝑎 (𝑎 + 𝑏) ∗ 𝑑 (𝑐 + 𝑑) ∗ Sensitivity vs Specificity How accurate is the test? Trade-off between the two Eg. cut-off point to designate ‘abnormal’ result (eg. blood sugar levels for DM) choice of cut-off determines if test is very sensitive (low cut-off point) or very specific (high cut-off point) More SENSITIVE test Low false negative SNout rule out condition (excluding) Use if benefits of detection/treatment is great Important for screening test (when cost of false negative is high) More SPECIFIC test Low false positive SPin rule in condition (confirming) Use if risks of further treatment is great Important for diagnostic test (when cost of treatment is high eg. chemotherapy) Predictive value Predictive value = what are the implications of results for my patient? Predictive value of test varies according to prevalence of disease Eg. Hospital has higher positive predictive value than Community 20 Screening tests vs Diagnostic tests Characteristic Purpose Costs associated with false negative result Costs associated with false positive result Sensitivity vs specificity Screening test Classify persons into those likely or unlikely to have the disease False reassurance Diagnosis postponed, higher morbidity Labelling/anxiety Risks associated with treatment Sensitive test Diagnostic test Confirm a diagnosis Requires further testing Labelling/ anxiety Risks associated with treatment Specific test Guidelines governing introduction of screening programmes (from tutorial) Natural history of disease o Is there a latent stage? Where asymptomatic, but can detect before critical point o Critical point = point in natural history of disease before which therapy may be more effective Importance of disease o Large numbers; or increase in disease incidence o Serious morbidity or mortality o Great burden of suffering Treatment o Available and acceptable treatment in early stages of disease to slow or stop it. Screening test o Acceptability o Safety o Sensitivity and specificity Cost-benefit ratio o Cost of test o Cost of follow up process if false-positive results found Need for repeated screening o Affected by length of latent phase (shorter length – screen more often) Outcomes of screening programmes (from tutorial) Reduce mortality Reduce case-fatality rate Reduce complications of disease Reduce recurrences or metastases Improve quality of life 21 14 – UNDERSTANDING AND USING META-ANALYSES Methods of summarising research findings Non quantitative o Narrative review – consensus statements, expert reviews Problems: subjective, no formal rules – can have selective inclusion of studies o Systematic review – rules for evaluating quality of article Quantitative o Vote counting Problems: ignores sample size, research design, effect size o Meta-analysis (‘analysis of analyses’) o Pooled analysis – individual data pooled together Meta-analyses Benefit – structured Components o Qualitative – assess quality using predetermined criteria o Quantitative – statistical integration of results o May be integrated – overall results weighted by quality score Steps o Define clinical/research question o Identify relevant articles Specify inclusion criteria – randomised, controlled trial, etc o Evaluate quality of each study according to methodology Criteria for appraisal of study quality – method and adequacy of randomisation, blinding protocols, completeness of follow-up, cointerventions, etc. o Summarise results quantitatively Overall measure of effect May analyse by subgroups if relevant o Discuss possible biases, and implications of results (self-critique) Understanding the findings o Overall effect – forest plots and summary estimates o Are results from various studies too different to be combined? - Heterogeneity o Is there significant publication bias? – Funnel plots Forest plots Provide information on results from each study, and the summary/ combined effect Risk ratio and confidence limits for each study plotted on the same scale. Area of each black square reflects weight of study, usually proportionate to size of the sample 22 Log scale used so that risk ratios of the same magnitude equidistant from 1, and confidence intervals symmetrical about point estimate Last one combines all into a single RR and CI. Heterogeneity Variations in results across trials, beyond the effect of chance Due to o Clinical diversity (participants, intervention, outcomes differ) o Methodological differences (design or conduct of study) o Statistical heterogeneity (excessive or unexpected variation in results) Tests for heterogeneity o Are the individual study results likely to reflect a single underlying effect, or different effects? o Many different statistical tests o Most tests give p value. If p > 0.05, differences assumed to be consequence of sampling chance o Some tests quantify level of heterogeneity across studies Eg. 25% of variation is due to heterogeneity rather than due to chance Dealing with heterogeneity o Fixed effects model Assume variability is exclusively due to random variation Each study weighted by sample size Used when not enough heterogeneity to be a concern o Random effects model Assume different underlying effect to reflect the uncertainty Gives wider confidence intervals Most appropriate if there is substantial heterogeneity between studies Publication bias Problem with identifying relevant articles o Grey literature – unpublished or awaiting publication o Fugitive literature – published but difficult to locate eg. conference proceedings o File-drawer manuscripts – unpublished as expected to be rejected o Foreign language studies Use funnel plots o X axis – increasing treatment effect o Y axis – sample size – a measure of precision o Towards the top (high sample size), results should be less varied o Empty funnel – no publications at 0 treatment effect as these are less likely to be published o Asymmetrical plots Could be that the truth IS this way – not a proof of bias Not very helpful if studies are clinically or methodologically diverse, or if there are few studies. 23 Pooled analysis Aggregates original data from several studies More precise Measurements of exposures and outcomes standardised; adjustment for confounders done consistently Can be very informative when meaningfully stratified eg. compare ethnic groups Improved statistical power Does not overcome biases – garbage in garbage out Limited by availability of data from primary authors Limitations of meta-analysis Publication bias (non-reporting of negative or non-significant study results) Temptation to integrate all studies regardless of quality Cannot change original study (warts and all) Other key findings might be neglected Methodological issues still unresolved eg. how much heterogeneity is allowed 24 TUTORIAL LEARNING POINTS Critical evaluation of study methodology If confidence interval includes 1 not statistically significant Multivariate RR = adjusted for smoking/other factors Trend = dose-dependent response Recruitment of subjects o Cohort study start with group of people without disease o Homogenous population – lifestyle, etc same. Hopefully only 1 exposure different. Need not represent whole world, but those who drink and those who don’t should be otherwise homogenous – to minimise confounders. Results can be generalised to all population, unless some nurse factor interacts with coffee (effect modifier) Baseline characteristics table o Shows you whether there are confounders – where exposed and unexposed differ Interpreting results o Eg. With reference to people who don’t drink coffee, those who drink 1/day have a 0.60 risk of pancreatic cancer. 95% confidence interval is 0.38-0.94, which does not include 1, thus it is statistically significant. o Eg. Trend test downward trend could be by chance as p-value = 0.35 > 0.05 – not statistically significant. o Most RR not significant; no dose-dependent relationship coffee drinking is not a risk factor for pancreatic cancer nd 2 case; interpreting results o Eg. In males, using 0 cups as reference group, those who drink <2, 2-3, and 4+ cups have RR of pancreatic cancer of 1.07, 1.45, 2.08 respectively. However, none are statistically significant because 95% confidence interval includes 1. Clinical evaluation of clinical trial Stratified randomisation o Stratify, then perform randomisation within each strata o Not usually performed as simple randomisation can already mimimise confounders o Use when some confounder is extremely striking Treatment effect 𝐶𝑜𝑛𝑡𝑟𝑜𝑙 𝑒𝑣𝑒𝑛𝑡 𝑟𝑎𝑡𝑒−𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑒𝑣𝑒𝑛𝑡 𝑟𝑎𝑡𝑒 𝐶𝑜𝑛𝑡𝑟𝑜𝑙 𝑒𝑣𝑒𝑛𝑡 𝑟𝑎𝑡𝑒 o 𝑅𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝑟𝑖𝑠𝑘 𝑟𝑒𝑑𝑢𝑐𝑡𝑖𝑜𝑛 (𝑅𝑅𝑅) = o This is more effective number. 𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒 𝑟𝑖𝑠𝑘 𝑟𝑒𝑑𝑢𝑐𝑡𝑖𝑜𝑛 (𝐴𝑅𝑅) = 𝐶𝑜𝑛𝑡𝑟𝑜𝑙 𝑒𝑣𝑒𝑛𝑡 𝑟𝑎𝑡𝑒 − 𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑒𝑣𝑒𝑛𝑡 𝑟𝑎𝑡𝑒 This is an arbitrary number; difficult to interpret o 𝑁𝑢𝑚𝑏𝑒𝑟 𝑛𝑒𝑒𝑑𝑒𝑑 𝑡𝑜 𝑡𝑟𝑒𝑎𝑡 (𝑁𝑇𝑇) = 𝐴𝑅𝑅 1 Number of patients you need to treat before one gets a fully favourable result Precision of estimate of treatment effect o Measured by 95% confidence interval 25 o More narrow; more precise Will results help in caring for my patients? o Similar to patients you see? o Consider side effects, recurrence of symptoms o Cost-benefit analysis Critical evaluation of diagnostic/screening test Why does cardiologist have higher positive predictive value than school doctor? o Higher sensitivity and specificity o Prevalence of disease higher for cardiologist than school doctor o Thus, not good indicator; but most direct advice you can give to patients when interpreting their results o Sensitivity and specificity are more stable indices Clinical decision making (including meta-analysis) Evidence based medicine practice Effectiveness in changing over-prescription of antibiotics o Depends on how well doctor can communicate his point to change: I – ideas C – concerns E – expectations; of the patients o Follow clinical practice guidelines no grounds for patient to sue you Type of healthcare system can affect management of patient’s disease Communication o Diagnosis/ disease + prevention o Prognosis – allay fears or prepare for the worst o Complications – treatment, occupation etc. Pre-empt patients. Guidelines on evaluation an overview/ systematic review/ meta-analysis Are the results of the study valid? o Primary guides Did the overview address a focussed clinical question? Criteria for inclusion of articles appropriate? o Secondary guides Were important, relevant studies missed? Was validity of included studies appraised? Were assessments of studies reproducible? Were results similar from study to study? What are the results? o What are the overall results? o How precise were the results? Will the results help my patients? o Can be applied to my patient care? o Were all clinically important outcomes considered? o Are the benefits worth the harms and costs? 26