Lecture 4a Treskova PDF

Lecture 4: Study designs and bias sources in epidemiology Marina Treskova, PhD Head of Research Group Eco-Epidemiology Heidelberg Institute...

Lecture 4: Study designs and bias sources in epidemiology Marina Treskova, PhD Head of Research Group Eco-Epidemiology Heidelberg Institute of Global Health & Interdisciplinary Centre for Scientific Computing Heidelberg University www.hei-planet.com 01/16/2025 Outline Study designs in Epidemiology : Observational studies Bias sources Regression analyses 01/16/2025 Study design is important Epidemiology: The study of the distribution and determinants of health or disease in populations; and the application of this knowledge to control health problems. For example: What is causing a particular disease in a population? What is the (average) risk of disease? Who is at most risk? How can risk be reduced? Epidemiological Questions: major goal is to explain patterns of disease occurrence and causation (etiology) Is there an association between an exposure and outcome? Is that exposure really the cause of the outcome/disease? How much does the exposure increase the risk of the outcome/ disease? 01/16/2025 Measurement Measurements in epidemiological research usually aim to quantify: Exposure Outcome Confounder Once quantified – we can explore associations Different ways of evaluating the association between an exposure and an outcome: these are the different study designs 01/16/2025 Main types of study design Observational: observe and compare i. Ecological ii. Cross-sectional iii. Case-control iv. Cohort Intervention/experimental: intervene and compare v. Randomised controlled trials Intervention/non-experimental: intervene and compare Difference-in-Difference, IV, ITS 01/16/2025 Null Hypothesis The null hypothesis (H0​) is a statement in statistical testing that assumes there is no effect, association, or difference between groups or variables. It serves as the default assumption and is tested to determine whether evidence supports rejecting it. https://medium.com/@andersongimino/differences-between-the-null-and-alternative- hypotheses-6b2e794543f6 01/16/2025 Type 1 and Type 2 Errors Type 1 Error (False Positive): This occurs when we conclude that there is an effect or association when none exists. Example: Concluding that a new drug is effective in lowering blood pressure when it actually has no effect. Significance Level (α\alphaα): The probability of making a Type 1 error is controlled by the significance level (commonly set at 0.05). https://www.simplypsychology.org/type_i_and_type_ii_errors.html 01/16/2025 Type 1 and Type 2 Errors Type 2 Error (False Negative): This happens when we fail to detect an effect or association that is truly present. Example: Concluding that a new drug does not lower blood pressure when it actually does. Power (1−β): The probability of avoiding a Type 2 error is related to the statistical power of the study. Higher sample sizes and effect sizes increase power. https://www.simplypsychology.org/type_i_and_type_ii_errors.html 01/16/2025 Study designs 01/16/2025 Bias Adapted from: Maclure, M, Schneeweis S. Epidemiology 2001;12:114-122 01/16/2025 Explanation of Biases 1. Random Error Definition: Random error arises from chance and affects the precision of the results. It leads to variability in the data but does not systematically favor one outcome over another. Example: Measuring the same individual’s blood pressure multiple times with slightly different results each time due to instrument variability or human error. https://www.open.edu/openlearncreate/mod/oucontent/view.php?id= 172093&section=6.1 Mitigation: Increasing sample size, averaging multiple measurements. 01/16/2025 Confounding 2. Confounding Definition: Confounding occurs when an extraneous variable is associated with both the exposure and the outcome, distorting the apparent relationship. Example: Studying the relationship between alcohol consumption and lung cancer without accounting for smoking, which is associated with both alcohol use and lung cancer. Mitigation: Randomization, matching, or statistical adjustments like stratification or multivariable analysis. https://catalogofbias.org/biases/confounding/ 01/16/2025 Information Bias 3. Information Bias Definition: Information bias results from errors in measuring exposure, outcome, or other study variables, leading to misclassification. Example: A recall bias in a case- control study where individuals with a disease are more likely to remember past exposures than those without the disease. Mitigation: Using objective measures (e.g., lab tests) and ensuring blinding during data https://www.scribbr.com/research-bias/information-bias/ collection. 01/16/2025 Selection bias 4. Selection Bias Definition: Selection bias occurs when the participants included in the study are not representative of the target population due to the way they were selected. Example: Recruiting participants for a study on obesity and heart disease only from fitness centers, excluding less active individuals. Mitigation: Ensuring random https://mostly.ai/blog/data-bias-types sampling and minimizing loss to follow-up. 01/16/2025 Bias of Inference 5. Bias of Inference Definition: This bias arises when incorrect conclusions are drawn from the results of a study, often due to overgeneralization or ignoring limitations. Example: Concluding that a drug is effective for all populations based solely on results from a small, homogeneous sample. Mitigation: Carefully interpreting results and acknowledging limitations in study design. Confirmation bias: the tendency to search for, interpret, or give more weight to information that confirms one's preexisting beliefs or hypotheses, while ignoring or downplaying evidence that contradicts them. 01/16/2025 Reporting Bias 6. Reporting Bias Definition: Reporting bias occurs when certain results are selectively reported, typically those that are statistically significant or favorable. Example: A trial that reports only the positive outcomes of a new drug and omits the negative or null results. Mitigation: Preregistration of studies, requiring full reporting of all outcomes. 01/16/2025 Publication Bias 7. Publication Bias Definition: Publication bias arises when studies with positive or significant findings are more likely to be published than those with negative or null results. Example: Meta-analyses may overestimate treatment effects if only published studies with significant results are included. Mitigation: Encouraging the publication of all study results, including null and negative findings, and using trial registries to identify unpublished data. 01/16/2025 Ecological studies Compare overall population or group statistics of exposures and outcomes, which can’t be linked to individuals. -> Aggregated level Exposure prevalence → Outcome prevalence e.g. Average diet and life expectancy across multiple countries. Breast cancer incidence data and average per caput daily consumption of a wide range of foods for 24 countries were extracted from routinely collected data sources (Armstrong & Mann, 1985). 01/16/2025 © LSTHM Ecological studies (cont) Advantages: Mostly use existing data – therefore cheap and quick Can generate hypotheses Disadvantages populations can differ by many factors other than the exposure of interest (confounding) it is difficult draw conclusions at the individual level (ecological fallacy) 01/16/2025 © LSTHM Cross-sectional studies All factors (exposure, outcome, confounders) are measured at one time point They are often used to measure the prevalence of health outcomes, understand determinants of health, and describe features of a population. Unlike other types of observational studies, cross-sectional studies do not follow individuals up over time. 01/16/2025 © LSTHM Cross-sectional study: example 01/16/2025 Cross-sectional study: example https://www.thelancet.com/journals/langlo/article/PIIS2214- 109X(21)00176-5/fulltext#fig3 01/16/2025 Cross sectional studies (cont) Advantages Quick to conduct Descriptive - Can investigate many exposure and outcome variables hypotheses generating Disadvantages Time sequence of events difficult to ascertain Can only measure prevalence not incidence Large surveys required for rare diseases or exposures 01/16/2025 © LSTHM Case-control studies Identify people with the health outcome (cases) and select appropriate controls Compare the proportion exposed between the two groups. E.g. (Many) vaccine effectiveness studies 01/16/2025 © LSTHM Case-control studies Advantages Can investigate multiple exposures (only one outcome) Can be quick (and can be relatively small) Good for rare outcomes e.g. cancers Disadvantages Limited to one outcome Selection bias if the controls are not from a comparable population to the cases Time sequence of events may be difficult to ascertain (unless you have records) Not good for rare exposures Cannot be used to estimate prevalence or incidence of disease 01/16/2025 © LSTHM Case-control study example The PERCH Study Group, 2019. Causes of severe pneumonia in children from Cases: pneumonia Controls: Children of Africa and Asia: the PERCH multi- patients admitted the same age as cases, country case-control study to hospital, HIV in the community, HIV negative negative Causality difficult to determine Difficult to know whether cases and controls are from comparable populations Sociodemographic and environmental factors; Samples from the nose for laboratory analysis E.g. would the controls have gone to hospital if they had become sick? Depends on parental health seeking behaviour (difficult to measure… ) Impact: Viruses e.g. RSV appeared more Recall bias - recall may differ when common in cases than expected. interviewing parents of sick children in hospital vs. parents of health children at home 01/16/2025 © LSTHM Cohort Studies Cohort studies are observational studies in which the starting point is the selection of a study population, or cohort. Information is obtained to determine which members of this cohort are exposed to the factor of interest. Follow up over time of a group of people according to their exposure status - who develops the health outcome(s) of interest? e.g. A large pooled analysis of 252,745 women looked at the association between use of talcom powder and (later) ovarian cancer 01/16/2025 © LSTHM Cohort studies Advantages Incidence can be measured Time sequence ascertained Rare exposures can be investigated (if cohort groups appropriately selected) multiple exposures and outcomes Disadvantages Often time-consuming and expensive Losses to follow-up can cause selection bias The ascertainment of the outcome may be influenced by knowledge of the exposure status Classification of individuals (exposure or outcome status) can be affected by changes in diagnostic procedures, natural changes over time etc. For rare outcomes a large cohort and/or long period of follow up is required 01/16/2025 © LSTHM Cohort study example Freeman et al., 2006, AIDS, 20(1), 73-83 Outcome: Incidence of HIV infection Male and female sex workers enrolled and followed up for multiple years Selected Exposure: HSV-2 infection Impact: women and men with HSV-2 had 3 At baseline and follow up visits: tests times higher risk of HIV acquisition for STIs and HIV, interviews on sexual compared to those not infected with HSV behavior, partner characteristics, Limitations: There could be unmeasured contraceptive use confounders that account for some or all of the association e.g. the association could be HIV Positive HIV negative confounded by sexual behaviour. A randomised controlled trial of STI treatment for the prevention of HIV acquisition was then conducted. 01/16/2025 © LSTHM Summary Study design is important to come to the right conclusions One study is rarely enough, it often takes a wide range of studies (and many years!) to accumulate evidence to influence policy Many studies (and published papers) have flaws Bad study designs are difficult to interpret Different studies addressing the same questions can differ in their conclusions due to variations in study design: definition/ measurement of the exposure/ outcome Biases in the way the study selected, enrolled and/or retained participants in the study Differences in the way the studies measured confounders and/or adjusted for confounders in the analysis (NB. Some behavioral factors are very difficult to measure and adjust for) Differences in the populations studied Interpret research findings in relation to the study design 01/16/2025 © LSTHM  Questions? www.hei-planet.com