Lecture Notes

Focus of this course is identification of causes. Conceptual Framework: Philosophy to FactsCounterfactual definition of cause- 1. Elaborated Causal Theory2. Gather Data to Confront Theory3. Compare Theory with Data4. Acknowledge Uncertainty Cohen: Statistical Power Subway MoralDo not use epidemiological methods unless you know the underlying foundation towards it. Obesity Epidemic: Epidemiology provides the evidence and to each of the links to the arrows in a systems maps. We are going to look at TV, Food Advertisements, and Obesity- Turned obesity into a disease, but is obesity a disease?- Disease= internal disfunction? health related condition? directly/indirectly which questions this definition. There's been a lot of push back because there's a lot of stigma involve, feminism. Pharmaceutical involvement is also or could be involved. -Disease labels are social constructs that are negotiated by stakeholders. Defining a CAUSE: Avoid using the word of cause, but then you make a policy suggestion. If something is not a cause, you can't technically intervene. We should take acknowledge of Voltaire, define terms. We use counterfactual all the time.  Causal Scenario Examples: Scenario 2 about the dog is a causal scene. The reason is we asked: if the phone didn't ring, would the teakettle still whistle? Not necessarily. But in Scenario 2, if the stranger didn't approach house, would dog bark? No, so causal. Definition: X and Y occurred and, within a causal field, in the circumstances, Y would not have occurred if X had not at least not when and how it did. (Mackie)1. In the circumstances...To isolated the circumstances to get away from other possibility as to why Y occurred.2. Causal field of interest: a lot of times when you ask causal question, you bracket them. You are asking within this particular contexts. The reality of the world that you are taking into consideration for identification. Ex. things in body or characteristics of people. Causal field is limited by your purpose. 3. At least not when and how it did. Rothman's Definition: "A cause of a specific disease event is an antecedent event, condition, or characteristics that was not necessarily for disease at the moment it occurred, given that other conditions are fixed." OR "A cause of a disease event is an event, condition, or characteristics that preceded the disease event and without which the disease event would not have occurred at all or would not have occurred until some later time."Rothman's Model of Casual Pies: Assumption is that there are many different routes to get disease. 1. Every individual have their own route of getting disease. Epidemiology is trying to generalize from unique to multiples but there are more than one way to get disease. 2. No cause work by self, but work with other causes. A sufficient cause is not one cause but constellation of causes that work together towards disease. Any two or more factor in pie is called causal partners. Mackie's INUS Causes: The kind of causes we are interested in Epi is...Insufficient but Necessary components of Unnecessary but Sufficient causes. Each sufficient cause is unnecessary because there's many ways to get to disease. INUS is each component of a pie. There are different types of INUS causes. 1. Interaction is ubiquitous. All causes work with other causes.  2. A cause is a strong cause dependent on presence of causal partners. How strong a risk factor/etc/punitive cause is dependent on prevalence of its causal partners. 3. You can get rid of disease but rid of its manipulable causal partners. So when when you get rid of one piece, then it will decrease the rate of disease because it eliminates sufficient causes/pie for some people, but it won't necessary eliminate all of the disease. The order of which cause is arbitrary. A cause will always have another cause. The order is arbitrary based on who you are and what you can do and goal or what you want to achieve. INUS causes are all necessary for their causal pie to happen so there can't be redundancy in the pies. Ex. R and A can't be in same causal pie. When R is a cause of A. Necessary cause is when you have INUS cause in all sufficient causes of disease. Example: Does A cause disease?Is there in fact a sufficient cause that has A as a component? Are there individuals who have A, got disease, who would not have gotten disease A?- Scenario 1: The causal effect of A: The difference between a fact and a counterfactual. If you subtract fact and counterfactual and there is a difference, then yes there is a causal effect of A. - Scenario 2: In absence of A but because have X then have disease. Then there is no difference and so A doesn't have causal effect. - In population, we average individuals' causal effects: Frequency of disease in a group of exposed people should be higher than the frequency of this same group if they were not exposed. Major Problem in Detecting Causes: Fundamental problem of cause inference- facts can be observed but counterfactuals can't. So causes can't be seen so only be inferred. Dr. Meditative: throughout class to see the truth and compare to what we get on the data. We try to look for proxy for the counterfactual. We in study use non-smokers as proxy for "if people didn't smoke." However, if we have a poor proxy for the counterfactual, then the study is off and can give a wrong answer. All of epidemiology is: under what condition would disease in the unexposed be a good proxy for counterfactual? This is when exposed and unexposed are like in all causes of disease except for the exposure or have full comparability. All of epidemiology is about counterfactuals and the proxy substituted. Full comparability is never known and causation is inferred. So we want to make causal inference as strong as possible. Measures of Disease frequency: Risk, rate, odds, prevalenceMetrics of Comparison: rates and differenceRisk= the proportion of a people in a fixed cohort who start out disease free, who develop disease over specific time period. The risk is the only measure of disease. Rate=instantaneous force of disease; measure of how quickly people in a population change status from non-disease to disease. Uses person-time. Number of people who get disease divided of the total amount of time they are in the study. We care about when people get sick. People contribute to denominator until they get sick. Risk and rate have same numerator but different denominator. Odds=Probability/Complement=Probability of Disease/Probability of Not Disease=risk/1-risk.Prevalence=the proportion of people at any given moment of time to get disease. We don't care if they got better or when. It is just how many people have disease at a certain time. Take home message and Mystery of the Week:1. A cause is defined by a causal contrast2. A causal contrast compares a fact with a counterfactual (non-observable)3. Counterfactual can't be seen and is a fundamental problem of causal inference. 4. Our studies are only as good as our proxies for the counterfactual. Mystery of the week:1. Why do we need more than one measure2. How does theory help ID causes3. Why do we to ID causes to improve the world?

Last time:- Counterfactual Causal - Cause is shown by the contrast of supposed the same population but one with and one with no exposure. A difference show it is cause. No difference show it is not a cause. A proxy for causal contrast is to use a control group, but use different groups of people. -Quantification: 1. Risk: Start with non-disease of interest, follow over time and observe for outcome. Proportion of people in fixed cohort who develop the disease over a specific time period. This is most intuitive and fundamental but hard to obtain.2. Rate: Got disease/person time; (would incidence rates be considered a rate even though it is not a true rate?)3. Odds: A probability divided by its complement4. Prevalence: We want proportion of people any moment in time. This can capture people who recovered. Basic Comparison Used in Epidemiology: Risk and DifferenceA. Ratio: This is used more in epidemiology. 1. Risk ratio: a/(a+b)/c/(c+d) 2. Rate ratio: a/(a+b)-c/(c+d)3. Odds ratio: ad/bcB. Difference:1. Risk difference2. Rate difference3. Odds differenceFleiss, 1981Rate of lung cancer for smoker and nonsmokers; rates of CAD for smoker and nonsmoker. Lung cancer rate is 11 (smoke vs. nonsmoker). CAD is 2. Looking at this, rate ratio would let you think it is important for lung cancer. But then the rate difference is greater, because CAD is more common. -Ratio is more used because: chronic disease shift. There was a problem in chronic disease because the way causes for chronic disease is different than what was there for infectious disease. There was a question whether smoking was cause for lung cancer. Because the ratio look more impressive, it was pushed to be used a paradigm. There were social historical constraints that made risk more used. Shows:1. Epidemiological method are developed for population in need at particular time.2. Counterfactual definition of the cause suggest difference measure might have more validity especially when we talk about interaction.Both get you causal contrast but tell you different things. Health DisparitiesRisk ratio of poor vs rich is 3 in Time 1. Then in Time 2, risk of poor vs poor is 8. Risk difference in T1 is 0.4. But then in T2, difference is 0.007. Reason:1. 60% of people in T2 had disease. 2. But then the intervention lowered for both, but comparison between the rich and poor is still different after getting better. The poor population just still need more work. Both are fine and acceptable but they can give a different feeling for the same data. Article will be posted on the webpage. Risk in Exposed vs. Risk due to the ExposureExample: We had identified 3 casual pies. We have E and G in Pie 1. Pie 2 has L. Pie 3 has P. Any pie can lead to disease.Population 1 all have E. Over time, people who get pie will get disease. Person 1,3 got the disease due to the exposure because if they didn't have E, then they would not have had disease. Person 2 got the disease because they got L so they didn't get the disease from the exposure. W/o E they still out have gotten disease.Person 5 also did not get disease from exposure because they got from exposure pie P. 4 people got disease, but only 2 got because of exposure. With these people you can do a risk ratio with people who got disease from exposed vs. unexposed. Whatever the INUS cause is, we need to compare the rate of disease with or with no INUS cause. Our comparison is only has good as our proxy for our counterfactual. Therefore we can only infer causation.Goal of Epidemiology Studies is to make causal inference as strong as possible. At each step of the counterfactual to inference, we want to make it as strong as possible. Step 1: Make question strong for causal inference: develop elaborated causal theory. Two Approaches to Causal Inference1. Induction: Look at specific facts in real world. From facts, you develop a theory/generalized principle. Ex. See white swan today. See white swam tomorrow. Induce that all swans are white. 2. Deduction Start with theory and make prediction on specific incidence. Ex. All men are mortal. Sophocles is a man. Therefore Sophocles is mortal. Stress causes disease. Individuals under stress will get depression.Danger of Induction: Torture the data and it will confess. No logical limitation to your explanation. Whatever you see, you can explain the trend no matter what you see. Ex. Fat Jeans- they fit anything. All you get is association but you never get causation. Problem is that you can't intervene on association and only causation. So we can't be satisfied with association. In opposition to address this problem from induction, people think about eliminate inadequate theories. 1. We expect our data results under theory and then see if the results match or no. We eliminate the inadequate or not logical reason in the first place. 2. We can also eliminate inadequate theories by falsifying the data. We need to protect self from having a bias about theory which leads us to interpret data a certain way. Having a frame of what will falsify theory will help make your theory. More falsify you make your theory, when it doesn't get knocked down at the end of the data means your theory was correct. Problem: we never see facts. We want to see how much disease in exposure group, unless we measure the whole population we don't get a fact. There are constraints on what you can imagine also. Ex. In 1950, ulcers were thought to be caused personality. While data show it was due to infection, but because paradigm was there so people couldn't see it. Ex. Fertilization was changed when women went in field. Egg was waiting for fertilization, but in real egg does things. Facts are refracted by the way we think. Our attempt to falsification is fallible. When we going through study, we want to make sure we are not careless and things are falsify-able because things can go wrong in study. The question is if we should start with induction or deduction. Develop Causal Theory1. Be open; get ideas from everywhere. Dissertation can come from anywhere. Once you have idea, then it makes it seem like everyone is talking about your idea. 2. Know your stuff: read broadly. Read in particular literature of people who don't agree with you. 3. What's bugging you? Keep diary of what you are reading and your reaction towards it. Keep diary or notes can allow you what your really thinking about or bothering you: things that are contrasting facts. You have idea and then develop theory with idea.Hills's Causal Criteria:A. Begin with Hill's Criteria help you in the development of your theory. Things to consider:- Temporality: example women more prone to depression due to menopause but depression starts from teenage years. - Plausibility: why would exposure and disease be linked in the first place?- Coherence: can you make a good story out of your theory. For instance, you didn't do hw for a class. You are asked why did you do your hw? Logic thing so have one thing to explain. - Analogy: Susser and schizophrenia about whether it is a neurodevelopment disorder? Or is it a neurodegenerative disorder? This hypothesis came because of neural tube defects. They found neurodevelopment problems due to nutrient defects. So it was thought that in utero that cause other neurodevelopment disorder. So look at schizophrenia. B. Evidence for theories:- Consistency- Experimental evidence.C. Evidence regarding specific corollaries- Strength- Biological gradient- Specificity: one bacteria, one disease. But then environmental exposure, we don't find this. Direct cause/relationship so in study we need to ask ourselves to whether we should expect specificity. Elaboration of Causal Theories (Shadish, Cook, and Campbell): These are all defined different ways. Causal Identification: Isolate the cause of interest from other causes. Two component:A. Association (main effect)B. Eliminate Competing hypothesis (internal validity-eliminate competing hypothesis): confoundingCausal Explanation: Explain how and when the cause of interest acts.C. Explain how exposure works (construct validity): mediationD. Explain when exposure works (external validity): effect modificationExample: Independent risk factors=Smoking is INUS cause of CAD and Genetic cause CAD. Each are a cause. Confounding= Competing explanation: Hormones and social roles are cause of depression. Mediation=Low SES leads to exposure to environmental toxin and then cancer; chain of events.Effect modification=together work together to cause diseaseCausal Identification: footprint that causes leave behind. You can never see causes. -Association beyond chance-Temporal order-Sole plausibility There are only 5 reason why causes lead to disease1. Chance (association challenges this)2. Reverse causation (temporality challenges this)3. Bias (sole plausibility challenge this)4. Confounding  (sole plausibility challenge this)5. Direct causeSo, we want to eliminate all other factors, and the one which remains must be true. DAGS: Diagrams and arrows mean causality. X-->D; X is INUS cause of D. X--> DE--> DE is cause of interest. X is whole pie and is independent. Confounding:X--> DX-->EX is a INUS cause of D. But X is also a cause of E. So it makes E and D associated as a result of common cause. But E does not cause D.Mediation:E-->X--> DEffect Modification(X and E)-->DThe presence of each other makes disease. Venn Diagram E and X is one way of showing effect modification.Developing a Causal Theory and Hypothesis about ObesityTV watching is INUS cause of obesity=theory. Main effect: Obesity is caused by increase caloric intake.Knock out confounding: not because of couch potatoes effect (no exercise)Mediation: Increase TV lead to increase one eating food. Effect Modification: When does it work: when at least one parent is obese. These are components but mediators might not happen. Goal: To have a coherence and parsimony theory. Specificity and complexity in your predictions. Example:Depression is associated with loss of significant relationship, with no replacement. Then if that is the case, it is parsimonious and falsifiable, you would see many of the other studies. The question is when to abandon or when is it to change theory.How do you defend your viability of your "casual theory"? Aka Literature ReviewBackground and Significance as Principled Argument (Craft of Research) Talks about Principled Argument. Claim, Evidence, Warrant (what are the implications as a result of the evidence to the claim) that links Claim and EvidenceThen what are the qualification?Thesis: Not all Claim are Acceptable. Need to beSubstantive (about something important)Contestable (challenges a belief)Strong Evidence has to beAccurate (no careless errors and fully recognizes limitations of the evidence)Authoritative (what counts as evidence in epidemiology)Strong Warrants:What general belief must I have before I can agree that your accurate evidence supports your claim?Strong Qualification:Without Concessions--Raise objections and alternatives you considered and rejected-Anticipate objections-Anticipate alternatives. (couple more points missing)Mystery of the Week:Noncomparability....

1-28-13

2-4-13