Measures of Disease Frequency Counts = number of people with a disease Rates--account for the denominator, or size of the population, and imply a period of time Cumulative Incidence (most commonly used as synonymous with "incidence") Synonymous with "attack rate", "risk of disease", and "probability of getting disease" = number of new cases of a disease occurring in a specified time period number of people initially at risk Incidence density Estimates the instantaneous rate of occurrence of disease per unit of time relative to the size of the population at risk. = number of new cases of a disease occurring in a specified time period total amount of "person-time" at risk contributed during the time period Prevalence A "snapshot" view of disease frequency in a population at a single point in time, important for planning and allocation of resources. = number of existing cases of a disease at a specified time size of base population at that time "Point prevalence" refers to the prevalence at a single point in time. "Period prevalence" refers to the prevalence measured over a specific time interval. Mortality = number of people dying of a disease in a specified time period average number of people alive during that period of time Case fatality Measures disease prognosis rather than a population’s risk of dying from the disease = number of people who die of a disease total number of people with the disease Proportional mortality Can be misleading if mortality rates for other causes are unusually high or low in a group. = number of deaths due to a disease in a specified time period total deaths during that time period Mortality = Incidence X Case Fatality Survival Rate = 1 – Case Fatality Neonatal death rate = annual number of deaths in the first 28 days of life annual number of live births (in thousands) Infant death rate annual number of deaths in the first year of life = annual number of live births (in thousands) Years of Potential Life Lost (YPLL) A measure of premature mortality YPLL takes into account the age of death Calculated by multiplying the number of cause-specific deaths in an age group by the difference between the midpoint of the age group and the average age at death (often assumed to be age 75) Descriptive epidemiology Practical uses of descriptive epidemiology o To provide clues to etiology and prevention o To help target prevention or screening efforts o To aid in the planning of health services o To provide baseline data for a program Sources of numerator data for descriptive epidemiology o Vital records (birth and death certificates) o Disease reports (for example, reportable diseases, tumor registries) o Medical records o Surveys Sources of denominator data o Census/vital statistics records o Enrollment records (HMO's, industry or union records, alumni rosters, etc.) What kinds of people get the disease? o Age (immunity, diseases with long latency periods, environmental exposures that vary with age) o Gender (anatomic and physiologic differences, many differences in life style, environmental exposures) o Race/ethnicity (genetic differences in susceptibility, differences in exposures) o Socio-economic status (related nutritional or environmental factors, health care access) o Other factors Where is the disease common or rare, and what are the characteristics of those places? o Geopolitical units o Physical environment o Man-made exposures that vary by place How does the frequency of the disease change over time? o What historical factors appear to correlate with those changes? o Short-term trends; disease outbreaks o Secular (long-term ) trends o Cyclical variations Reasons for an association between a factor and a disease Bias in the sampling of subjects Bias in the measurement of the factor Confounding by another factor Chance Transposition of cause and effect o (disease causes the factor) Causal o (factor causes the disease) Yes No Yes (A) (B) No Exposed Diseased (C) (D) Relative Risk (RR) The ratio of the incidence of a condition in the group of individuals with a specific characteristic (a "risk factor") to the incidence in the group of individuals without the risk factor. Ratio of Two Rates A R(e) A+B Relative Risk = = C R(u) C+D Where: R(e) = rate of disease among the exposed R(u) = rate of disease among the unexposed Units cancel out AKA: risk ratio, rate ratio Example: 3.5x more likely to have cancer if you smoke Attributable Risk (AR) The difference between the incidence of the disease in individuals with a risk factor and in those without Accounts for the baseline incidence of disease and gives the absolute amount of excess risk an individual Ratio of Two Rates Attributable Risk = R(e) – R(u) Where: R(e) = rate of disease among the exposed R(u) = rate of disease among the unexposed Units do NOT cancel out AKA: risk difference, rate difference Example: In population X, 500 per 100,000 die due to smoking Number needed to treat (NNT) NNT Represents a more interpretable transformation of the attributable risk Can be the number needed to screen, number needed to treat, or number needed to harm = 1 Attributable Risk “You’d have to treat (NNT) people to gain one additional outcome” Population Attributable Risk Percent (PAR%) The proportion of cases of disease that is due to a given risk factor The PAR% is the amount of disease that would be prevented if the risk factor could be eliminated from the population. PAR% PAR% = = Total Incidence – Incidence in Unexposed Total Incidence A+C A+B+C+D C C+D – A+C A+B+C+D X 100% Odds Ratio Odds Ratio = A C B D = AD BC Research Designs 1) Descriptive (hypothesis generating) a. Describe Something i. Case Report ii. Case Series b. Describe how something varies (by time, place) i. Descriptive Epidemiology Study 2) Analytical (hypothesis testing) a. Observational Studies (non-experimental) i. Before-After Study BIAS: unclear what would have happened without the intervention ii. Ecological Study / Correlational Study Measurement of Risk: Relative Risk or Attributable Risk Compare same population over time Compare different populations at same time The key feature of this study design is that comparisons are made at the group level (not individuals) Susceptible to the “ecological fallacy” BIAS: the individuals with the exposure aren’t necessarily the ones with the outcome (the ecological fallacy) iii. Cross-Sectional (prevalence study, survey) Measurement of Risk: Relative Risk or Attributable Risk Typically surveys Can include exams or questionnaires Sample should be representative Can produce estimates of prevalence Uncertainty of separating cause vs effect is a serious limitation BIAS: can be difficult to establish temporal relationship between exposure and outcome iv. Case-Control Study (Retrospective Study) Measurement of Risk: Odds Ratio (as an RR estimate) Sample of cases Sample of people from same population without the condition (controls) Susceptible to two types of bias: sampling and measurement Therefore, this design is now giving-way to cohort studies in the medical literature BIAS: a. Selection Bias i. Representativeness of the case group ii. Appropriateness of the control group (especially if study is not population-based) iii. Detection bias (unmasking bias)--results when the identification of cases varies with exposure status b. Bias in Case Selection i. Cases should be incident cases ii. Including prevalent cases of long duration could lead to identification of factors that influence prognosis or survival rather than etiologic factors iii. Other selection problems (such as hospital-only cases) could choose more or less severe cases c. Bias in Control Selection i. Controls should have the same experience with the exposure as those non-diseased individuals in the population that gave rise to the cases d. Information Bias i. Recall bias ii. Non-blinded ascertainment of exposure status iii. Misclassification 1. Wrong exposure or outcome status 2. Inclusion of heterogeneous outcomes (can lead to diluting true findings of association) v. Cohort Study (Prospective Study, Follow-up Study, Longitudinal Study) 1. Prospective (concurrent, futuristic cohort study) 2. Retrospective (historical, retrospective-prospective cohort study) Measurement of Risk: Relative Risk or Attributable Risk Requires assembly of cohort and follow-up over time Limited utility for very rare diseases or very long latencies BIAS: a. Selection Bias i. Selection bias can occur in the formation of exposure groups 1. Those choosing to be exposed may be different than those who don’t 2. There may be other factors related to outcome that determined why one group was exposed and the other wasn't ii. Completeness of follow-up is a major source of potential bias, especially if loss to follow-up is unequal in exposure groups b. Information Bias i. Ascertainment bias can occur, especially if method of determining outcome does not include blinding to exposure status ii. Misclassification is also a potential problem BIAS IN OBSERVATIONAL STUDIES Misclassification Exposed cases as controls Exposed controls as cases Exposed cases as unexposed Exposed controls as unexposed Unexposed cases as controls Unexposed controls as cases Unexposed cases as exposed Unexposed controls as exposed Effect on OR underestimate overestimate underestimate overestimate overestimate underestimate overestimate underestimate b. Intervention Studies AKA “trials” Exposure is controlled Best for short-term outcomes for exposures that can be blinded i. Randomized Controlled Trial (true experiment, clinical trial) Measurement of Risk: Relative Risk or Attributable Risk Essentially the same as a cohort study, except the investigator decides who gets the exposure, using random assignment Strongest study design of all, maximizing internal validity (usually at the expense of external validity) Maximizes internal validity by promoting the equal distribution of potential confounders into exposed and unexposed groups Phase III of Clinical Trial (see below) Steps in RCT a. Define the hypothesis b. Select study subjects i. Requires informed consent and strict inclusion and exclusion criteria ii. Pre-randomization visits are used to help insure successful participation iii. These design elements likely make the study population nonrepresentative c. Randomly allocate subjects to intervention groups i. Assignment must be random ii. Quasi-random techniques must produce assignments that can produce no sources of bias d. “Blind” subjects whenever possible i. Subjects should be blinded (e.g. with placebo treatment) to which study group (experimental vs. control) they are in whenever possible to guard against placebo effects and crossover e. “Blind” investigators whenever possible f. Follow and ascertain all relevant outcomes, monitor for adverse effects and stopping rules i. Those ascertaining outcomes should be blinded to the subject's study group whenever possible to guard against investigator bias (double-blind trial) ii. Consider all relevant outcomes iii. Safety monitoring for adverse effects iv. Stopping rules (the point at which the results become statistically significant and ethically the trial should be ended) v. RCTs must be analyzed using the "intent-to treat" approach: subjects must be analyzed as belonging to the group to which they were first randomized, even if they cross-over to the other group vi. Re-assigning cross-overs invalidates the RCT, as those that cross-over are likely to be quite different from those who don't. vii. Confounding variables must be compared across groups, as errors of randomization can lead to imbalances in the distribution of these factors. viii. Often RCT's compare continuous variables between groups: recognize that in general, it takes a much smaller difference to be statistically significant with continuous variables than with categorical variables Sources of Bias a. Bias in randomization b. Ascertainment of outcomes c. Lost to follow-up d. Inclusion of all relevant outcomes e. Cross-overs, intent to treat analysis f. Selection bias (now relates to external validity/generalization) g. Errors of randomization and confounding h. Selection Bias i. Comparability of study groups (randomization, allocation concealment) ii. Maintenance of comparability of study groups (loss to followup) iii. Selection bias now also relates to external validity/generalization i. Information Bias i. Adherence, contamination, cross-over, intent-to-treat analysis ii. Blinding of outcome ascertainment iii. Inclusion of all relevant outcomes Issues in RCTs a. For drug studies, dosage, route and timing determined from phase I and II; still need to monitor toxicity b. Don’t deviate from protocol c. Include in design approach to non-compliance, protocol deviation (toxicity, crossover, etc.) d. Determine sample size based on the size of the difference that you want to detect among the intervention arms, and the amount of variation that exists in your measurements e. In general, lengthy and very expensive f. Ethical and legal issues are important g. Blinding can be difficult, cross-overs may be common, and drop-outs and lost to follow-up are major problems h. Strong internal validity is strong is achieved at the expense of external validity; your study groups may not generalize to any other population i. Power issues become important when effect sizes are inadequate to reach statistical significance j. Still, the gold standard of research design ii. Natural Experiments Researcher does not determine the group receiving the intervention, which occurs "naturally" or under control of some other process Can address problems where RCT interventions are unrealistic as to what can be implemented and sustained in natural setting Can be used for rapid evaluations of innovative, expensive, or complex interventions or policy changes in natural settings Problems with internal validity May have limited generalizability (but better than RCT) iii. Group Randomization Trials (GRTs) Study design where the unit of assignment is an identifiable group, allocated to different exposures Randomization of the groups is intended to distribute potential sources of bias fairly distributed across the study exposures There are relationships or associations between members of the groups that may introduce bias (interclass correlation) This is worse when the number of groups is small iv. Quasi-Experiment Measurement of Risk: Relative Risk or Attributable Risk These studies are those in which the researcher cannot or does not assign interventions randomly to participants, but may have some control over: a. who gets the intervention b. when the intervention is given, and/or c. when measurement occurs Study design needs to match the question under study: There are instances where an RCT will not be the best design (when you cannot address generalizablility), when an RCT is not feasible, and when an RCT is not appropriate There are two major concerns with quasi-experimental designs: a. additional sources of bias not controlled for by random assignment (especially selection bias) b. interclass correlation that may be responsible for the observed effect, unrelated to the intervention (confounding) Non-equivalent control group design a. Study groups are assembled in a non-randomized fashion intended to minimize unequal distribution of important confounders, and researcher decides which group(s) gets the intervention (X) Group A Ol X O2 Group B Ol O2 Time series design (interrupted times series, with or without non-equivalent control group; with a control group = multiple times-series design). a. Represents a refinement of the pre-post study design in that multiple measurements over time give more of a sense of whether postintervention differences can be assigned to the intervention Study group Ol O2 O3 O4 X O5 O6 O7 O8 Control group Ol O2 O3 O4 O5 O6 O7 O8 3) Summary or Extrapolation Studies a. Meta-analysis b. Decision analysis c. Modeling Phases of Clinical Trials Phase I – Safety Trials o Designed to establish toxicities (if any) and: Subtoxic dose Minimal toxic dose Therapeutic dose Maximum tolerated dose o Preceded by literature review, animal studies, and pharmacological studies (effective period, clearance period, toxic period, effective dosing schedule, follow up period o Trial period must be long enough to demonstrate toxicity over a period of time long enough for the drug to show effective action o Compromise precision in order to expose the fewest individuals to a drug with unknown effectiveness in humans o Some knowledge of effective dose and route needs to be determined before Phase II (effectiveness) trials can be mounted o Since toxicity is observed in all trials, this seems like a reasonable compromise Phase II—Efficacy Trials o Not intended to test the causal relationship between treatment and removal of symptoms o Only to establish the possibility that the treatment could be effective o The trial period must be long enough for the drug to be effective o Outcomes need to be carefully defined: measurable, clear, unambiguous and achievable in a relatively short time o Bias should be addressed with complete follow-up patients and blinded, independent, objective assessment of response o Subjects should be as homogeneous as possible, and the best possible candidates for showing that the treatment works Phase III-Comparative trials o The randomized controlled trial Phase IV – Post Marketing Surveillance Trial Research design and strength of evidence (evidence grading) Grade I--randomized controlled trials Grade II-1--quasi-experimental controlled studies Grade II-2--cohort studies, case-control studies Grade III--descriptive studies, expert opinion, expert panel consensus Causal Relationships Steps in determining cause: o Investigate the statistical and temporal association o Eliminate alternatives through research Often epidemiology must be satisfied with determining “causal association” Causal Association Causation has its roots in the Koch-Henle Postulates Koch-Henle covers diseases for which the cause is necessary and sufficient; usually restricted to infectious agents. These postulates are not appropriate when disease entities are multifactorial, such as most chronic diseases Diseases not caused by infectious agents usually require considering etiologic factors that are sufficient but perhaps not necessary, or even not sufficient but contributory. RULES: o The association should be strong o The association should be consistent o The association should be strongest when you expect it to (biological gradient) o The exposure should precede the outcome o The association should be biologically plausible, including supportive data from other sources Screening for Disease Phases of Disease Prevention Criteria for targeting a disease for screening The disease is important. The disease has a recognizable pre-symptomatic stage. Reliable screening tests exist for the pre-symptomatic stage. Treatment of the disease during the pre-symptomatic stage results in improvement in outcomes. Sufficient resources exist for diagnosis and treatment of disease in the population with positive screening tests. Yes No Yes (A) True + (B) False + No Test Result The "Truth" (C) False – (D) True – Sensitivity Describes how often a screening test detects a disease when it is indeed present = TP/(TP+FN), or true positives over the total with disease. Specificity Describes how often a screening test detects the absence of disease when it is indeed absent = TN/(TN+FP), or true negatives over the total without disease Positive predictive value Describes how often individuals with positive tests actually have the disease = TP/(TP+FP), or true positives over all positives Negative predictive value Describes how often individuals with negative tests are actually disease-free = TN/(TN+FN), or true negatives over all negatives Effect of prevalence Sensitivity and specificity remain the same for the test regardless of the prevalence of disease in the population. Predictive value depends on sensitivity, specificity, as well as disease prevalence. As the test is applied to populations with lower disease prevalence, positive predictive value drops and negative predictive value increases Conversely, as disease prevalence increases, positive predictive value increases and negative predictive value drops Screening Test Biases In the evaluation of a screening procedure or program, at least two sources of bias must be considered: o Lead time bias Lead time bias occurs because screening tests detect illness at earlier points in time. If you developed a screening test that diagnoses a disease one year earlier, but early treatment had no effect on survival, you would know the person had disease for a year longer than those detected without screening o Length time bias If the prevalence pool (those with the disease in the population) are made up of individuals with these differences, then a screening test would be more likely to pick up individuals with longer asymptomatic phases Hence, those detected by screening would be more likely to have longer disease regardless of whether it is detected by screening or not These sources of bias can result in finding improvements in disease mortality even if the early treatment of the disease in no way alters the outcome. Two Methods for combining the results of multiple screening tests in an overall strategy: Parallel testing: the overall screening result is positive if any one test is positive. This strategy increases sensitivity at the expense of specificity Series testing: the overall screening result is positive only if all tests are positive. This strategy increases specificity at the expense of sensitivity Clinical Decision Making Diagnosis Most often, patients present with symptom complexes, not diagnoses We use subjective and objective measures to provide a diagnosis with some degree of certainty Diagnostic tests suffer from the same sources of variation and bias as do screening tests Diagnostic Tests Tests will have sensitivity, specificity and predictive values (just like screening tests) Performance of tests (certainty of diagnosis) will vary with test characteristics and probability of disease given the patient’s presentation The chance that a patient has a disease, given the presentation but before tests are run, is called the pre-test probability The chance or certainty that a patient has a disease given the prior probability and the test result is called the post-test probability Odds = Probability Probability 1 – Probability = Odds Odds + 1 Likelihood ratio (LR) The Likelihood Ratio (LR) is the likelihood that a given test result would be expected in a patient with the target disorder compared to the likelihood that that same result would be expected in a patient without the target disorder (true positive tests vs. false positive tests) The LR is used to assess how good a diagnostic test is and to help in selecting an appropriate diagnostic test(s) or sequence of tests LRs have advantages over predictive values because they: o Are less likely to change with the prevalence of the disorder o Can be calculated for several levels of the symptom/sign or test o Can be used to combine the results of multiple diagnostic tests o Can be used to calculate post-test probability for a target disorder A LR greater than 1 produces a post-test probability which is higher than the pre-test probability An LR less than 1 produces a post-test probability which is lower than the pre-test probability When the pre-test probability lies between 30 and 70 per cent, test results with a very high LR (say, above 10) rule in disease Test results with a very low LR (say, below 0.1) virtually rule out the chance that the patient has the disease Test Res ult Yes The "Truth" Yes No (A) True + (B) False + No (C) False – (D) True – Sensitivity LR+ = = 1 – Specificity 1 – Sensitivity LR- = = Specificity Pre-Test Odds = TP TP+FN FP FP+TN FN TP+FN TN FP+TN Pre-Test Probability 1 – Pre-test Probability = Prevalence 1 – Prevalence Post-Test Odds = Pre-Test Odds X Likelihood Ratio Post-Test Probability = Post-Test Odds Post-Test Odds + 1 Decision Analysis Used to assess the comparative outcomes of two or more procedures Based on probabilities of outcomes based on the literature or estimates/assumptions Applying costs provides the ability to calculate the cost per success Can be used to compare benefits and harms in deciding net benefit of an intervention Can be used to compare costs and cost effectiveness of different interventions Decision Tree Statistics in the Medical Literature Statistics Sample Mean ( ) Variance (s2) Standard Deviation (s) Parameters Population µ Σ σ Types of Variables Qualitative o “nominal” (eg, gender, hair color, city) o amenable to categorical analysis only Quantitative o Continuous Any value between two other values is possible: between values 1 and 2 the values 1.1, 1.01, 1.001, etc. is a possible value Examples: weight, blood pressure, IQ, serum levels of electrolytes, proteins, etc. Likert Scale (1, 2, 3, 4, 5) acts like continuous data o Categorical (Ordinal) “Dichotomous” if only 2 categories Only two discrete categories are possible Can exist without hierarchy: o male/female Can exist with inferred hierarchy: o No disease/disease o Dead/alive o Fail/pass o Exposed (smoker)/not exposed (non-smoker) “Ordinal” if hierarchy implied Nominal data with more than two possible states and an existing hierarchy: A. B, C, D, F grading system Education level (grade school, high school, some college, college graduate, postgraduate school Cancer state I, II, III, IV (in-situ, local, regional, distant metastases) 10-year age categories Mode the most commonly observed value Median the middle observation in a data set arranged lowest to highest Mean the arithmetic average (sum of observations divided by the number of observations) Mean = sum (x) /n Range highest value minus lowest value Variance Variance = sum (x-mean)2 /(n-1) A standardized measure of the sum of the differences between each value and the mean value Standard deviation the square root of the variance, which has special properties when describing a Normal Distribution curve Frequency Distributions Conceptually identical to a pile of anything o the center of the pile is the central tendency o the total width of the pile is the range o the amount of scatter is the variance Frequency distribution common in nature o mean = median o the % of subjects with various values can be estimated by the standard deviations Normal Distribution Mean = median approximately 67% of values lie within a 1 standard deviation distance from the mean approximately 95% of values lie within a 2 standard deviation distance from the mean Kaplan-Meier Plot Hypothesis Testing Null hypothesis o no difference o no association o sameness, equality o statistically testable Alternative hypothesis o opposite of null (hence, more interesting) o not statistically testable The "Truth" (Real Difference) Yes (A) √ No Study Findings (Significant difference) Yes (C) Type II Error β=chance of type II error Power=(1–β)•100% No (B) Type I Error p=likelihood result occurred by chance α=p-value cutoff for significance (D) √ Type I error (false positive): When the investigator concludes there is a difference, when the difference seen occurred by chance or bias Measured by the alpha level, which is the same as the p-value cutoff Type II error--(false negative): When the investigator concludes there is no difference, when there is a true difference obscured by chance (ie, inadequate sample size) or bias Measured by the beta level, reported as “power” = 1-beta If data are continuous and means are to be compared t test o t = difference/SE of difference If data are categorical and proportions are to be compared ….. Chi Square test o X2 = sum across all cells of (obs - exp)2/exp In both cases, compute a test statistic, then convert (by tables or computer) to p value Statistical Confidence Expressions P value o the probability that chance alone caused the observed association o .05 is only a conventional threshold Confidence Interval o the range of values within which you are x% confident that the true value lies o 95% CI is a common convention o mathematics behind 95% CI identical to .05 p value Statistical vs. clinical significance A small observed difference, even if statistically significant, may not translate to a clinically significant difference An observed difference that would be clinically significant but is not statistically significant has two reasons for not being statistically significant: o there is no difference o the power was too low (too small a sample size) for the true difference to reach the cutoff for statistical significance Variability and Bias Could the findings reported be due to? C hance B ias C onfounding Variability Variation exists in all biologic systems Variability exists in all measurement systems Variability can be random or non-random Clinical research is designed to identify and assign non-random variability to the different factors under study Validity Defined as the degree to which a measurement or study reaches a correct conclusion o Internal validity--the extent to which the results of an investigation reflect the true situation of the study population o External validity--the extent to which the results of a study are applicable to other populations Bias The goal of science is to be accurate in the discovery, description, and measurement of the truth Bias is a systematic deviation of study measurement, results or inferences from the truth The internal validity of a study relates to the minimization of bias so that the study result can most confidently be assigned to the factors under study Three major categories: o Selection bias Factors in selecting subjects include the design of the study, the setting of the study and the exposure and outcome of interest Selection bias refers to bias that results because of (usually) pre-existing factors in study subjects that influence their outcome independent of exposure o Information (misclassification) bias Exists when there is random or systematic inaccuracy in measurement Types of information bias: Non-differential misclassification (non-systematic) Differential (systematic) misclassification: o ascertainment bias o recall bias o Confounding A distortion of the true relationship between an given exposure and a given outcome, resulting from a mutual relationship with one or more extraneous factors The effect of the extraneous factor(s) can account for all or part of the observed relationship between exposure and outcome, or mask an underlying relationship Effect Modification Criterias for Confounding The potential confounder must be associated with the outcome of interest: o the confounder is an actual risk factor for the outcome o the confounder affects the likelihood of recognizing the outcome The potential confounder must be associated with the exposure of interest but not be a result of the exposure. When is a factor NOT a confounder? If an individual's status regarding the confounder is a result of the exposure under study, or the confounder is in the "causal pathway" between exposure and outcome If an individual's status regarding the confounder is a result of the disease under study If the confounder is essentially measuring the same thing as the exposure If the association between the confounder and the outcome of interest is thought to be due to chance How do you detect confounding? Determine if the potential confounder is associated with both exposure and outcome Adjust for the potential confounder in analyzing the data. If there is a difference between the adjusted and unadjusted estimate of the effect of the exposure, then potential confounder is a true confounder How do you account for or control for confounding? Prevent confounding through the study design: o Restriction: study only the subjects in a given category. o Matching: match individuals for comparison on the basis of their status regarding the confounder (and use a matched analysis) o Use a RCT design: randomly allocate subjects to exposed and unexposed groups (note that confounding can still happen by chance) Remove effects of confounding in the analysis: o Report stratum- specific rates: list the effect of the exposure on the outcome for each level of the confounder o Use statistical techniques to account for the confounder: rate adjustment Mantel-Haenszel methods to account for the confounder Regression analysis (especially attractive if you need to consider several coincident confounders) Occurs when the effect of a risk factor on an outcome is different at different levels of a third factor; the third factor is known as an effect modifier Note that compared to the definition of confounding, effect modification says nothing about the relationship between the outcome and the effect modifier The most common effect modifiers seen are age and gender It is usually not appropriate to summarize over the strata of an effect modifier (if the difference is clinically significant) Rate Adjustment A technique used to provide a single number (adjusted rate) for each of two or more comparison groups which summarizes the experience of the group but is not influenced by a confounding variable when the comparison is made Rate adjustment is most often used to account for the affect of age (age adjustment); also commonly used to adjust for gender; could be used to adjust for any confounder or group of confounders. In rate adjustment, you essentially estimate how the comparison would turn out if the groups had the same distribution of the confounder (e.g., the same age distribution) The most common type of rate adjustment is called “direct rate adjustment” and produces “ageadjusted rates” or “age- and gender-adjusted rates Used in vital statistics and disease reports (allows comparisons among states “adjusted to the US population”)