# Lesson 8 — Review of Study Design Concepts (v3 expanded)

*Companion-podcast transcript • Sarah & Kiffer* 
*~5980 words • ~32 min audio*

---

**Sarah:** Welcome back to Office Hours. I'm Sarah.

**Kiffer:** And I'm Kiffer. Today we're working through Lesson 8, Review of Study Design Concepts. And the framing matters a lot for this one. This is a checkpoint lesson. We've spent the first seven lessons of this material building specific tools. Causal logic in Lesson 1. Surveillance and outbreak investigation in Lesson 2. Sampling in Lesson 3. Questionnaire design in Lesson 4. Measures of disease frequency in Lesson 5. Screening and diagnostic tests in Lesson 6. And measures of association in Lesson 7. Lesson 8 stops to consolidate.

**Sarah:** Consolidate what, exactly? I want to be clear with students about why we're pausing here.

**Kiffer:** We consolidate the observational study designs they first met earlier in this series, and then explicitly map each design to the measures of association we built in Lesson 7. So the question for today is not, what is a cohort study? Students saw cohort studies earlier. The question is, given a study design, which measure of association comes out of it, and what can you actually claim from that number?

**Sarah:** And the framing the lesson uses is that this is the foundation for the second half of the course. Lessons 9 and 10 will introduce hybrid designs and controlled designs that combine or extend these basics.

**Kiffer:** Right. So if you're a beginning epidemiology student, the goal for today is that by the end of the episode, you can hear the words cross-sectional study, cohort study, case-control study, ecological study, and systematic review, and have a clean picture of what each one is, what number it produces, and where it lives on the evidence hierarchy.

**Sarah:** There are three sections in the lesson. Cross-sectional studies in Section one. Cohort and case-control in Section two. And ecological studies plus an introduction to evidence synthesis in Section three. Let's take them in order.

**Kiffer:** Section one. Cross-sectional studies. But before we get into cross-sectional specifically, the lesson zooms out for the highest-level classification in epidemiology. Observational versus experimental.

**Sarah:** And the line between those two is sharp and worth saying carefully. Who controls the exposure assignment?

**Kiffer:** Exactly. In an experimental study, sometimes called an intervention study or a controlled trial, the researcher deliberately allocates exposure. They decide who gets the new drug and who gets the placebo. They decide which classroom gets the curriculum and which one is the control. In an observational study, the researcher does not assign exposure. They just observe and measure exposures and outcomes as they occur naturally in the world.

**Sarah:** And just to ground this with examples. A randomized controlled trial of a vaccine is an experimental study. The investigators flip a coin, metaphorically, and assign each person to vaccine or placebo. A study following nurses for thirty years and asking whether the ones who happened to smoke got more lung cancer is observational. The investigators didn't assign anyone to smoke.

**Kiffer:** And this distinction matters enormously for causal inference. Experimental designs, especially randomized controlled trials, can establish what we call exchangeability through random assignment. Quick definition. Exchangeability means the two groups, exposed and unexposed, are interchangeable. They look the same on average on every characteristic that matters, because randomization balanced everything by chance. So any difference in outcome you observe must be due to the exposure. There's no other variable left to blame.

**Sarah:** Whereas observational studies have to fight that fight by hand. Exposed and unexposed groups in the real world can differ in lots of ways that also affect the outcome. Smokers and non-smokers differ on income, occupation, stress, education, geography. To say smoking causes lung cancer specifically, you have to disentangle smoking from all those other things. That's the confounding problem.

**Kiffer:** And students might naturally ask, then why don't we just always do experiments? They sound cleaner.

**Sarah:** And the answer is the lesson's bridge into why observational studies remain essential. Most public health questions can't be studied experimentally, and there are three reasons. Ethical, practical, and the long-latency problem.

**Kiffer:** Walk us through them one at a time.

**Sarah:** First, ethical. You can't randomize people to be smokers. You can't randomize them to live in poverty. You can't randomize them to be exposed to industrial pollution or to live in a particular neighborhood. For most of the exposures public health cares about, randomizing people to receive them would be unethical.

**Kiffer:** Second, practical. Even when randomization isn't unethical, it's often impractical. You can't randomize entire populations to a national policy change. You can't randomize a city's water supply. Some exposures only exist at the group level.

**Sarah:** And third, the long-latency problem. Many of the exposure-disease relationships we care about play out over decades. Smoking and lung cancer. Childhood diet and adult heart disease. Asbestos and mesothelioma. You'd need a randomized trial running for forty years before you could see the outcome. That's not feasible.

**Kiffer:** Which is why most epidemiologic research is observational. Observational designs are the backbone of evidence for most health questions. Experiments are the strongest evidence when you can do them. Observation is what's available the rest of the time, which is most of the time.

**Sarah:** Okay. Then within observational, the next split. Descriptive versus analytic.

**Kiffer:** Descriptive studies characterize the distribution of disease in a population. They answer who, what, where, and when. Common descriptive designs include case reports, case-series reports, and surveys that estimate disease prevalence or describe the demographic characteristics of affected populations.

**Sarah:** And the lesson gives a beautiful example of why descriptive matters. The first formal description of what we now call HIV/AIDS was a case-series report in 1981 in the Morbidity and Mortality Weekly Report, published by the United States Centers for Disease Control and Prevention. Five young men in Los Angeles with an unusual pneumonia caused by a fungus called pneumocystis. That report didn't test a hypothesis. It just described an unusual cluster. But it was the signal that something new was happening.

**Kiffer:** And descriptive studies are often the first step in epidemiologic investigation. They generate the hypotheses that the analytic designs in later lessons then test. They don't formally test causal associations themselves.

**Sarah:** Analytic studies do try to test those hypotheses. They include a comparison group. Exposed versus unexposed, or cases versus controls. And they use statistical methods to estimate the strength and direction of associations between exposures and outcomes. Cohort, case-control, and cross-sectional analytic studies are the major types.

**Kiffer:** And the boundary isn't rigid. A cross-sectional survey that just reports prevalence is descriptive. The same survey becomes analytic when it compares prevalence across exposure groups and estimates a measure of association. A purely descriptive ecological study is descriptive. The same ecological data becomes analytic when you start computing rate ratios across regions.

**Sarah:** Okay. So now we can talk about cross-sectional studies specifically, with all that vocabulary in place.

**Kiffer:** Yes. A cross-sectional study measures exposure and outcome status simultaneously in a defined population at a single point in time, or over a short period. It provides a snapshot of the population. And four key features define it.

**Sarah:** Let me list them out for everyone.

**Kiffer:** Go ahead.

**Sarah:** First. Timing. Exposure and outcome are assessed at the same time. There is no follow-up period. Second. Measure of disease frequency. Prevalence, not incidence. Because you're measuring how many people currently have the condition, not how many new cases arose over time. Third. Measure of association. The prevalence ratio or the prevalence odds ratio. Fourth. Sampling. Participants are typically sampled from a defined population without regard to their exposure or disease status.

**Kiffer:** And the worked example from the lesson is great for grounding all of that.

**Sarah:** Walk through it slowly, because I want to push you on the punchline.

**Kiffer:** Researchers survey five thousand adults in a city. They measure two things at the same time. Current physical activity level. Are you sedentary or active? And whether the person has been diagnosed with type two diabetes. They find that twelve percent of sedentary adults have diabetes, compared to four percent of active adults. The prevalence ratio is twelve over four, which equals three. So sedentary adults are three times as likely to currently have diabetes as active adults.

**Sarah:** Okay. Push back on the obvious causal interpretation. Can we conclude that being sedentary caused diabetes?

**Kiffer:** Not necessarily. And this is the core limitation of cross-sectional designs. The story you want to tell is, sedentary lifestyle leads to diabetes. The exposure caused the outcome. But there's another story that's just as consistent with the data. Some people may have become sedentary because of their diabetes. They have joint pain. They have fatigue. They have neuropathy. They've been told to take it easy. Their diabetes caused their sedentary lifestyle. The arrow runs the other way.

**Sarah:** And the data we have can't distinguish those two stories. Because we measured exposure and outcome at the same moment. We don't know which came first. There's no temporal sequence to anchor the causal claim.

**Kiffer:** Right. And this is the defining limitation of cross-sectional studies. The inability to establish temporal sequence. We met this earlier with the dog ownership and blood pressure example. Same logical problem.

**Sarah:** Even with that limitation, the design still has real strengths. List them for us.

**Kiffer:** Strengths. Relatively quick and inexpensive to conduct. You don't have to wait for follow-up. You measure once and you're done. You can study multiple exposures and multiple outcomes simultaneously, because you're collecting a big snapshot of variables. They're useful for estimating disease prevalence, which matters for planning health services. You need to know how many hospital beds, how much medication, how many specialists. They're good for generating hypotheses that analytic designs can later test. And they can often be based on existing data sources, like national health surveys.

**Sarah:** And limitations. The big one we already covered. They can't establish temporal sequence. But there are a couple more worth naming.

**Kiffer:** Yes. Second limitation. Cross-sectional studies measure prevalence, not incidence. And prevalence is influenced by both incidence and duration. So results are distorted by how long disease lasts. Conditions that last longer get over-represented in cross-sectional snapshots. Conditions that resolve quickly or kill quickly get under-represented.

**Sarah:** Walk through that with an example, because it's subtle.

**Kiffer:** Imagine two diseases. Disease A has high incidence. Lots of new cases each year. But it resolves quickly. People recover within a week. Disease B has low incidence. Few new cases. But it lasts forever. Once you have it, you have it for life. If you do a cross-sectional survey today, you'll find lots of people walking around with disease B and few with disease A. Even though disease A actually arose more often. Prevalence at any given moment captures a snapshot biased toward long-duration conditions.

**Sarah:** And the practical implication is that cross-sectional studies are poor for risk-factor research. Because the prevalence association you observe could be driven by what causes new cases, what affects how long people stay sick, what affects who survives long enough to be sampled, or some combination. You can't separate them.

**Kiffer:** And that gets a special name in the lesson. Prevalence-incidence bias. Sometimes called Neyman bias or survival bias. People with rapidly fatal conditions die before they can be sampled. People with quickly resolved conditions get better before they can be sampled. So they're missing from your snapshot.

**Sarah:** Which is why, for studying causes of disease onset, designs that measure incidence directly are preferred. And that's the bridge into Section two.

**Kiffer:** Right. Section two. Cohort and case-control studies. The two analytic designs that can support causal claims about disease onset, because they unambiguously establish temporal sequence.

**Sarah:** Let's take cohort first.

**Kiffer:** A cohort study begins by identifying a group of individuals, the cohort, who are free of the outcome of interest. Then we classify them by exposure status. Then we follow them over time to observe whether the outcome develops. The direction of inquiry moves from exposure to outcome. The same temporal direction as causation itself.

**Sarah:** And there are two flavors. Prospective and retrospective.

**Kiffer:** Prospective cohort, sometimes called concurrent cohort. Participants are enrolled in the present. Their exposure status is assessed now. And then they're followed forward in time to observe the development of outcomes.

**Sarah:** And the textbook example is the Framingham Heart Study.

**Kiffer:** The Framingham Heart Study is the canonical prospective cohort. It's a study based in Framingham, Massachusetts, a small town outside Boston. It began enrolling participants in 1948. The original cohort was about five thousand adult residents who were free of cardiovascular disease at enrollment. Researchers measured everything they could think of. Blood pressure, cholesterol, weight, smoking, diet, occupation. And they followed those people forward to see who developed heart disease.

**Sarah:** And just to be clear about how foundational this is. Pretty much every risk factor for cardiovascular disease that you've ever heard of, the very phrase risk factor itself, came out of Framingham. The study has now followed three generations of participants. Children of the original cohort enrolled in the seventies. Their grandchildren enrolled in the early two thousands. It's still running.

**Kiffer:** And the advantage of prospective cohort is what makes Framingham work. Exposure is measured before the outcome occurs. Which means recall bias is minimized. Nobody's sitting around in 1948 thinking, I might get heart disease in twenty years, let me misremember my smoking history. Smoking was just measured at the time, accurately. And temporal sequence is unambiguous. We know smoking came before the heart attack, because we measured smoking decades earlier.

**Sarah:** The disadvantage is right there too. It's expensive and time-consuming. Especially for diseases with long latency periods or low incidence. You're paying for follow-up of thousands of people for decades.

**Kiffer:** Right. Then retrospective cohort. The investigator uses historical records, employment records, medical charts, registry data, to reconstruct a cohort whose exposure status was determined in the past. The outcomes may have already occurred, or they can be ascertained in the present.

**Sarah:** Quick example. Say you want to study whether asbestos exposure causes mesothelioma in shipyard workers. You go to a shipyard's employment records from 1960. You identify everyone who worked there and their job titles, which tell you about their asbestos exposure. Then you check provincial death registries today to see who developed mesothelioma. The cohort logic is intact, exposure precedes outcome, you're following from exposure to outcome. But you're not actually waiting. The follow-up has already happened.

**Kiffer:** Advantage. Faster and less expensive than prospective, because the waiting period for outcome development has already elapsed. Disadvantage. It relies on the quality and completeness of existing records. Important variables may not have been measured at all. Or they may be inconsistent.

**Sarah:** Now the measures of association from cohort studies. This is where Lesson 7 connects in directly.

**Kiffer:** Two measures. Both based on incidence, because cohort studies follow people forward and observe new cases. First, the risk ratio, sometimes called the relative risk. Both phrases refer to the same number. The risk ratio is the ratio of cumulative incidence in the exposed group to cumulative incidence in the unexposed group.

**Sarah:** Cumulative incidence is the proportion of people who developed disease over the follow-up period. So if you followed two hundred exposed people and twenty got disease, cumulative incidence in the exposed is ten percent. If you followed two hundred unexposed people and four got disease, cumulative incidence in the unexposed is two percent. The risk ratio is ten over two, which equals five. The exposed group has five times the risk.

**Kiffer:** Right. And the risk ratio works well for closed populations with short follow-up. Closed population means no one enters or leaves over the study period. Short follow-up means everyone is observed for roughly the same amount of time.

**Sarah:** When the population is open, where people enter and leave at different times, or when follow-up time varies, you switch to the second measure. The incidence rate ratio.

**Kiffer:** The incidence rate ratio is the ratio of incidence rates between exposed and unexposed groups. Incidence rate uses person-time in the denominator instead of just the number of people. Person-time accounts for variable observation. Someone followed for ten years contributes ten person-years. Someone followed for two years contributes two person-years. So even if some people drop out or join late, you can still compute a meaningful rate.

**Sarah:** And the cohort design is powerful because it gives you both measures. You can get a risk ratio. You can get an incidence rate ratio. You can compute a risk difference, which is the absolute difference between groups. All of those flow naturally from the design.

**Kiffer:** Now strengths and limitations of cohort.

**Sarah:** Strengths. It establishes temporal sequence. It estimates incidence directly. It can examine multiple outcomes from a single exposure, since you're following people for everything that happens to them. And it's less subject to recall bias, because exposure is recorded at baseline before the outcome occurs.

**Kiffer:** Limitations. Expensive and slow, especially for prospective designs. Inefficient for rare outcomes, because you'd need to follow huge numbers of people for a long time to see enough cases. And it's vulnerable to differential loss to follow-up, where dropout is related to both the exposure and the outcome in a way that biases the results.

**Sarah:** And differential loss to follow-up is sneaky. If sick exposed people drop out of the study at higher rates than sick unexposed people, you'll undercount disease in the exposed group, and your risk ratio will be biased toward the null. Toward looking like there's no effect when there really is one.

**Kiffer:** Right. So that's cohort. Now case-control.

**Sarah:** Case-control inverts the direction of inquiry. You start with people who have the outcome already, the cases. And a comparable group who don't, the controls. Then you look backward in time at exposure histories.

**Kiffer:** And the thing to internalize is that this is the same epidemiologic question, asked from the other end. Cohort goes from exposure to outcome. Case-control goes from outcome to exposure. Both are trying to figure out whether exposure and outcome are linked. They just sample on different things.

**Sarah:** And the strengths of case-control are essentially the mirror image of the limitations of cohort.

**Kiffer:** Right. Strengths. It's efficient for rare diseases. You don't need to follow huge populations forward and wait for enough cases. You can go to a cancer registry and pull two hundred cases of liver cancer, and you've already got more cases than a cohort of fifty thousand would have produced in a decade.

**Sarah:** You can study multiple exposures from a single case series. You have your two hundred liver cancer cases. You can ask them about their occupation, their alcohol use, their hepatitis history, their diet, all at once. So case-control is good when you have one rare outcome and many possible exposures.

**Kiffer:** And it's fast and relatively cheap, because the cases already exist and you don't have to follow anyone forward.

**Sarah:** Limitations. It's vulnerable to selection bias in control selection. The lesson is sharp on this. Controls should come from the same source population that gave rise to the cases. They should be people who, had they developed the disease, would have been identified as cases in this study. If they're not, your comparison is broken.

**Kiffer:** And that's often the most scrutinized aspect of any case-control study. Where did the controls come from? Hospital-based controls, patients admitted for other conditions, are convenient but may share risk factors with the cases. Population-based controls, random samples from the community, are more rigorous but harder to recruit.

**Sarah:** Second limitation. Recall bias. People with the disease often remember their exposures differently from people without it. A liver cancer patient looking back at their occupation may scrutinize every solvent they ever touched. A healthy control may not bother. So the cases report more exposure than the controls, even if their actual exposures were the same. That biases the association upward.

**Kiffer:** And third limitation. You cannot directly compute disease incidence or risk. Because you sampled on the outcome, not on exposure. You decided in advance that you'd have two hundred cases and four hundred controls. The ratio of diseased to non-diseased in your study tells you nothing about the population's actual disease frequency. So a risk ratio out of a case-control study is meaningless.

**Sarah:** Which is why the only valid measure of association from a case-control design is the odds ratio.

**Kiffer:** Right. Let me spell it out. The odds ratio compares the odds of exposure among cases to the odds of exposure among controls. If you arrange the four counts in a two-by-two table, exposed cases, unexposed cases, exposed controls, unexposed controls, the odds ratio is the cross-product. Exposed cases times unexposed controls, divided by unexposed cases times exposed controls.

**Sarah:** And the odds ratio is interpretable on its own as a measure of association. But under certain conditions, it also approximates the risk ratio.

**Kiffer:** Yes. The rare disease assumption. When the disease is rare in the source population, typically below five percent or so, the odds ratio closely approximates the risk ratio. The math works out because when prevalence is low, the odds of disease is approximately equal to the probability of disease. So the odds ratio is approximately the risk ratio.

**Sarah:** And there's a related but stronger result we should mention. Under what's called incidence density sampling, the odds ratio approximates the incidence rate ratio directly, without needing the rare disease assumption.

**Kiffer:** Right. Incidence density sampling means you select controls from people who are still at risk in the source population at the moment each case arises. Under that sampling scheme, the odds ratio is an unbiased estimator of the incidence rate ratio, regardless of how common the disease is. We'll come back to incidence density sampling when we cover hybrid designs in Lesson 9, and nested case-control studies specifically.

**Sarah:** So to summarize the cohort versus case-control comparison. Cohort goes exposure to outcome, samples on exposure, measures incidence directly, gives you a risk ratio or incidence rate ratio, is best for rare exposures and multiple outcomes, but is expensive and slow.

**Kiffer:** Case-control goes outcome to exposure, samples on disease, can't measure incidence directly, gives you an odds ratio that approximates the risk ratio under the rare disease assumption, is best for rare diseases and multiple exposures, fast and cheap, but vulnerable to recall and selection bias.

**Sarah:** And the lesson uses a nice practical example to drive the choice home.

**Kiffer:** Yes. Suppose you want to study whether exposure to a specific industrial solvent increases the risk of a rare liver cancer. A prospective cohort study would require following thousands of exposed and unexposed workers for decades. Extremely expensive. Extremely slow. A case-control study, by contrast, could identify two hundred liver cancer cases from a cancer registry, select four hundred matched controls, and assess past occupational exposure through interviews and employment records. Producing results in months rather than years.

**Sarah:** For rare diseases, case-control is usually the design of choice. You can't afford to do it any other way.

**Kiffer:** Section three. Ecological studies and an introduction to evidence synthesis.

**Sarah:** Sections one and two were all individual-level designs. We measured exposure and outcome on people. Now we move up one level.

**Kiffer:** Right. An ecological study, also called a group-level or aggregate study, uses groups, rather than individuals, as the unit of analysis. Instead of measuring exposure and outcome in each person, the investigator compares exposure levels and disease rates across defined populations. Countries, regions, provinces, cities, time periods.

**Sarah:** And the strengths are real. It's cheap and fast, because you're using existing aggregate data, often from national statistics or surveillance systems. It's useful for studying group-level exposures that don't even exist at the individual level. Things like a public health policy, a tax, a regulation, the structure of a healthcare system.

**Kiffer:** And it's often the only feasible design when individual-level data are unavailable. Comparing disease rates across countries using national statistics is a perfectly reasonable thing to do. There's no individual-level alternative.

**Sarah:** And it's worth pausing here on the Canadian data infrastructure side too, because the lesson highlights it. Most ecological and cohort work in Canada doesn't involve enrolling fresh participants. It reuses existing data.

**Kiffer:** Right. Three pieces of infrastructure students should know by name. First, Population Data BC, sometimes shortened to PopData BC. It links de-identified individual-level administrative data across the BC Ministry of Health, Vital Statistics, PharmaNet, the BC Cancer Registry, physician billings, and hospital discharges. Researchers obtain a study-specific extract under a data access agreement.

**Sarah:** Second, the Health Data Research Network Canada. That's a federation of provincial data centres, including PopData BC, ICES in Ontario, MCHP in Manitoba, that supports multi-jurisdictional studies under a single application. Each centre runs the analysis behind its own firewall and only summary results are shared. So individual-level data never crosses provincial lines.

**Kiffer:** And third, CANUE. The Canadian Urban Environmental Health Research Consortium. CANUE is a national repository of standardized environmental exposures at the postal code level. Air pollution, greenness, walkability, noise, climate, neighborhood socioeconomic indices. You can link CANUE indicators to any cohort with postal codes, including PopData BC extracts.

**Sarah:** And the example the lesson walks through is exactly the kind of study you can build from this stack. To estimate the effect of long-term fine particulate matter exposure on cardiovascular disease in BC adults, you can define your cohort from PopData BC, link each person's postal code to CANUE air pollution estimates, follow forward through Vital Stats and hospital discharges, and run a Cox regression. Three different data stewards, no new participant recruited.

**Kiffer:** Which is a nice illustration of how a retrospective cohort with an environmental exposure sits right at the boundary of cohort and ecological designs. Okay. Back to the ecological fallacy, which is the big conceptual warning for this section.

**Sarah:** Yes. Ecological studies have a famous limitation that students need to understand cold.

**Kiffer:** The ecological fallacy. The lesson states it clearly. The ecological fallacy occurs when an association observed at the group level is incorrectly assumed to hold at the individual level.

**Sarah:** And there's a really fun example that illustrates this. The chocolate consumption and Nobel Prize winners study.

**Kiffer:** Tell us the story.

**Sarah:** A few years ago a researcher published a paper looking at countries' per-capita chocolate consumption and their per-capita rate of Nobel Prize laureates. He plotted them and found a striking positive correlation. Countries that consume more chocolate per person have more Nobel laureates per person. Switzerland is high on both. So is Sweden. So is Germany.

**Kiffer:** And the headline-friendly interpretation is, eating chocolate makes you smarter. Eat your way to a Nobel.

**Sarah:** But that's the ecological fallacy in its purest form. Just because countries with higher per-capita chocolate consumption have more Nobel laureates per capita does not mean that the individuals within those countries who eat the most chocolate are the ones who win Nobels. Within a country, the chocolate eaters and the Nobel winners are probably different people entirely.

**Kiffer:** And the actual explanation is straightforward. Both per-capita chocolate consumption and per-capita Nobel laureates are effects of being a wealthy country with strong research infrastructure. Wealthy countries can afford luxury chocolate. Wealthy countries can afford the universities and research institutions that produce Nobel laureates. The two variables are correlated because they share a common cause at the country level. Not because chocolate causes Nobels at the individual level.

**Sarah:** And the lesson uses a more serious historical example too. Emile Durkheim, the founding French sociologist, observed in the late eighteen hundreds that regions of Europe with higher proportions of Protestant residents had higher suicide rates than predominantly Catholic regions.

**Kiffer:** And the careful interpretation is, this is an association at the regional level. It does not mean that Protestant individuals were more likely than Catholic individuals to commit suicide. It is possible the social environment of predominantly Protestant regions, perhaps characterized by greater individualism or weaker social networks, affected everyone living there. Including Catholics. Attributing the group-level finding to individual Protestants would be the ecological fallacy.

**Sarah:** And here's the principle the lesson is sharp on. Ecological studies are appropriate when the inference is at the group level. The fallacy only enters when group-level findings get translated into individual-level claims.

**Kiffer:** Say more about that distinction. It matters.

**Sarah:** Sure. If you're a public health policymaker comparing vaccination rates across countries to evaluate which national policy approaches work best, ecological data are exactly the right data. Your unit of decision is the country. Your unit of analysis is the country. The inference stays at the group level. No fallacy.

**Kiffer:** But if you take that same country-level finding and turn it into a claim about individuals, like, Swiss people who eat more chocolate are smarter, you've crossed the line. You've inferred from group level to individual level, and the data don't support that move.

**Sarah:** Despite the fallacy risk, ecological studies are still valuable. They're inexpensive and quick. They're often the only feasible design for group-level exposures. And they're excellent for hypothesis generation. The key is to interpret them cautiously and seek corroboration from individual-level studies.

**Kiffer:** Then the lesson takes a step back from individual study designs and asks the natural follow-up question. No single study, regardless of design, definitively establishes a causal relationship. So how does the field aggregate evidence across many studies?

**Sarah:** Which is the introduction to systematic reviews and meta-analyses. Evidence synthesis.

**Kiffer:** A systematic review uses a pre-specified, transparent, and reproducible protocol to identify, appraise, and synthesize all available evidence on a specific research question. Unlike a narrative review, where the author picks studies subjectively and writes a story, a systematic review is structured.

**Sarah:** And the structure typically involves five steps. First, define a precise research question, often using the Population, Intervention, Comparator, Outcome framework. Sometimes shortened to PICO.

**Kiffer:** Right. So Population, Intervention, Comparator, Outcome. PICO. You specify each one before you start. Who is the population you care about? What is the intervention or exposure? What's the comparator group? And what's the outcome you're measuring?

**Sarah:** Second, conduct comprehensive systematic searches of multiple databases. PubMed, Embase, Cochrane, and so on. Third, apply explicit pre-specified inclusion and exclusion criteria to the studies you find. Fourth, critically appraise the quality and risk of bias of each included study. Fifth, synthesize findings, either qualitatively in a narrative summary, or quantitatively through meta-analysis.

**Kiffer:** And meta-analysis is the quantitative synthesis. When studies are sufficiently similar, their results can be statistically pooled to produce a single combined estimate of effect. The pooled estimate has more statistical power than any single study, gives you a more precise estimate, and lets you explore heterogeneity. Whether results differ by study design, population, or exposure definition.

**Sarah:** And the standard visualization for a meta-analysis is the forest plot. Walk through it, because it shows up everywhere.

**Kiffer:** A forest plot lists each included study on its own line. For each study you see a point, usually a square, that represents the effect estimate, like the risk ratio or odds ratio for that study. And a horizontal line through the point that represents the confidence interval. The size of the square typically reflects the weight of that study in the meta-analysis. Bigger studies get bigger squares.

**Sarah:** And at the bottom of the plot is a diamond, where the center of the diamond represents the pooled estimate across all studies, and the width of the diamond represents the confidence interval around the pooled estimate. So at a glance you can see whether all the individual studies point in the same direction, how much they vary, and what the combined answer is.

**Kiffer:** And there's a reporting standard for systematic reviews. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses. Shortened to PRISMA. Just like STROBE is the reporting standard for observational studies. PRISMA is a checklist of items that should be present in a systematic review report.

**Sarah:** And PRISMA is the EQUATOR Network's guideline for systematic reviews. We met the EQUATOR Network earlier, the big international consortium that maintains reporting standards across study types. Students will also recall that an earlier lesson was an entire lesson on systematic reviews and meta-analysis. So today's segment is really just connecting that earlier material back to the design hierarchy.

**Kiffer:** Right. The point for now is that systematic reviews and meta-analyses sit at the top of the traditional evidence hierarchy because they synthesize across studies using transparent, reproducible methods. Below them are randomized controlled trials, then cohort studies, then case-control studies, then cross-sectional studies, then ecological studies, then case reports and expert opinion.

**Sarah:** And as we said earlier, the hierarchy is a guideline, not an absolute rule. A well-designed cohort study may provide stronger evidence than a poorly conducted randomized trial. A well-designed cross-sectional study can be more useful than a sloppy cohort. The hierarchy describes typical causal weight, not absolute quality.

**Kiffer:** Then the lesson closes with a summary table mapping designs to questions, which is worth saying out loud.

**Sarah:** Cross-sectional designs answer prevalence questions. How common is this condition right now?

**Kiffer:** Cohort designs answer questions about incidence and risk over time. Among exposed people, how often does the disease arise?

**Sarah:** Case-control designs answer questions about rare outcomes efficiently. Among people who already have this rare disease, what was different about their past exposures?

**Kiffer:** Ecological designs answer group-level questions, or generate hypotheses for individual-level work. How do disease rates vary across countries that have different policies, environments, or population structures?

**Sarah:** And systematic reviews answer questions about the totality of evidence. Across everything we know, what does the literature as a whole tell us about this exposure and this outcome?

**Kiffer:** And the choice of design depends on the question. There is no universally best design. There is a best design for a particular question, given particular constraints.

**Sarah:** Okay. Let me try to pull the takeaways together for this checkpoint lesson.

**Kiffer:** Yeah. Seven I'd want a student to leave with.

**Sarah:** First. Observational versus experimental is the highest-level split in epidemiology. The investigator's control over exposure assignment is what distinguishes them. Most epidemiology is observational, because most exposures can't be ethically or practically randomized, and many outcomes have latencies that exceed the patience of any randomized trial.

**Kiffer:** Second. Within observational, the descriptive versus analytic split. Descriptive characterizes patterns and generates hypotheses. Analytic includes a comparison group and tests hypotheses by estimating measures of association.

**Sarah:** Third. Cross-sectional studies measure exposure and outcome simultaneously. They give you prevalence and prevalence ratios. They cannot establish temporal sequence, which means they're vulnerable to reverse causation. And because prevalence is influenced by both incidence and disease duration, they're poor for risk-factor research.

**Kiffer:** Fourth. Cohort studies follow exposed and unexposed groups forward in time. The direction of inquiry runs from exposure to outcome. They measure incidence directly. They give you a risk ratio for closed populations with short follow-up, or an incidence rate ratio for open populations with variable time at risk. The Framingham Heart Study, since 1948, is the textbook prospective cohort. Strengths are clean temporal sequence and direct incidence. Limitations are cost, time, inefficiency for rare outcomes, and vulnerability to differential loss to follow-up.

**Sarah:** Fifth. Case-control studies invert the direction. Outcome to exposure. They sample on disease status. They cannot directly compute incidence or risk, so the only valid measure of association is the odds ratio. The odds ratio approximates the risk ratio when the disease is rare. They're efficient for rare diseases and for studying multiple exposures from one case series. They're vulnerable to selection bias in control selection and to recall bias.

**Kiffer:** Sixth. Ecological studies use group-level data. The unit of analysis is a population, not an individual. They're cheap, fast, and useful for group-level exposures and group-level questions. Their famous limitation is the ecological fallacy. The chocolate and Nobel Prize example is a clean illustration. Ecological designs are appropriate when the inference stays at the group level. The fallacy enters when you translate group-level findings into individual-level claims.

**Sarah:** Seventh. Systematic reviews and meta-analyses synthesize evidence across studies using a transparent, pre-specified protocol. Forest plots are the standard visualization, with a square and confidence interval for each study and a diamond for the pooled estimate. PRISMA is the reporting guideline. And an earlier lesson is the deeper treatment of systematic reviews if students want to revisit the details.

**Kiffer:** And one more meta-takeaway. This is a checkpoint lesson. The reason we stopped to consolidate is that the second half of this material is going to push beyond these four observational designs. Lesson 9 is hybrid designs. Nested case-control studies built within cohorts. Case-cohort designs. Case-crossover designs that use each case as their own control. Lesson 10 is controlled studies, including randomized field trials. Each of those is a variant on the four standard designs you just consolidated. If the four standard designs are clear, the hybrids will make sense. If they're not, the hybrids will feel arbitrary.

**Sarah:** And the practical study advice for students. Take the summary table from the lesson. The one mapping designs to research questions and to measures of association. Try to reproduce it on a blank piece of paper without looking. If you can write down, for each design, the direction of inquiry, the measure of disease frequency, the measure of association, and the main strength and main limitation, you've got the foundation.

**Kiffer:** And for each design, sketch the two-by-two table. We met that table back earlier with case-control. It's the same table for cohort, case-control, and cross-sectional. What changes is what's fixed by sampling and what gets computed from the cells. Cohort fixes the row totals because you sample on exposure. Case-control fixes the column totals because you sample on disease. Cross-sectional fixes the grand total because you sample on neither. The arithmetic of every measure of association we've covered flows from those constraints.

**Sarah:** Next up is Lesson 9. Hybrid Study Designs. Where we move beyond the basic four observational designs into combinations and extensions.

**Kiffer:** Take care, everyone.

**Sarah:** See you there.