1 / 26

Principles of Experimentation

Principles of Experimentation. ST711 Fall 2014. Learning objectives. Differentiate an observational study from a comparative experiment Identify treatments and blocks Differentiate experimental units (EUs) from observational units

connie
Download Presentation

Principles of Experimentation

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Principles of Experimentation ST711 Fall 2014

  2. Learning objectives • Differentiate an observational study from a comparative experiment • Identify treatments and blocks • Differentiate experimental units (EUs) from observational units • List four principles of experimentation and their impact on statistical analysis

  3. Correlation does not imply causation • Hilarious Examples • Observational studies • Observe and record environment without interrupting it • Find relationships but limited in explaining why they’re there • Lurking variables • Unobserved/unrecorded events that may be behind relationship • Firefighter example • Confounded variables • Recorded variables that could also explain a relationship

  4. Can smoking cause cancer?Study 1 • 50 smokers/nonsmokers each (age 55-65) • Compare rates of cancer in two groups • Rate of cancer for smokers was higher than for nonsmokers • Potential lurking/confounding variables • Smokers may be genetically predisposed to cancer • Smokers made other poor health choices (diet and exercise) • Non-smokers may be more likely to visit doctor • When they started smoking and how much could affect rate

  5. Does smoking cause cancer?Study 2 • 50 similar mice • Randomly assign 25 mice to be subjected to same daily amount of cigarette smoke for fixed time • Otherwise given nearly identical treatment (exercise, diet, etc.) • Advantage: each mouse has same chance of being a smoker (small chance for confounding/lurking variable) • Disadvantage: we care about humans smoking, not mice!

  6. Sources of variation • Source of variation: anything that causes two responses to be different from each other • Identify potential sources of variation prior to data collection • Goal: Determine whether identified, controllable sources explain the majority of variation • Need to reduce impact of unknown or uninteresting sources to increase sensitivity of tests • Control of major sources of variation allows us to make causal inferences

  7. Elements of comparative experiments • Comparative Experiment: a study in which an experimenter • Assigns a delineated set of conditions to a group of subjects • Compares conditions’ effects on a response variable • Treatment: a possible condition that may be assigned • Different treatments are always considered to be potential major sources of variation • Other sources lead to experimental variability • Experimental unit (EU): whatever we assign a treatment to • Why isn’t Study 1 a comparative experiment? • Identify the treatments and EU’s in Study 2.

  8. Does smoking cause cancer?Study 3.1 • Conduct Study 3 this way: • Break up 50 mice into 10 chambers with 5 mice in each • Randomly assign 5 of 10 chambers to receive smoke • Mice are put in the same chamber each day • Claim: the individual mice are not the EU’s • What are the EU’s?

  9. Observation units and pseudo-replication • Observational unit (OU): “part” of EU on which response is measured • Study 3: each mouse is both an EU and OU • Study 3.1: mouse is OU, chamber is EU • Different EU’s implies different treatments or different re-creations of the same treatment • Why important: two OU’s from the same EU are correlated since they received the same treatment re-creation • Pseudo-replication: Treating OU’s as EU’s will underestimate experimental variability (increase Type I error)

  10. Design and analysis of experiments • Design implies a certain analysis • Goals of analysis should impact design • Design questions: • Sample size: how many EU’s and OU’s to use? • How to make inference as broad as possible? • How to handle inherent variability among EU’s? • How to allocate treatments to EU’s? • How do I collect/record data? • Potential analysis goals: • Different treatments lead to change in response(s) • Conditions lead to maximum or minimum response • Build statistical model

  11. Key principles of design • Representativeness • Replication of treatments • Error control (blocking and covariates) • Randomization

  12. representativeness • Applies to all studies, not just experiments • Population of EU’s should be representative of the population we want to make inferences on • Study 3: extend results to humans? • Homogeneous EU’s will reduce experimental variability at the cost of representativeness • Design techniques exist that broaden pool of EU’s

  13. Replication • Replication: a re-creation or copy of the same treatment applied to a different EU • In order to be certain of a treatment’s effect, it must be observed repeatedly. • Increasing replication leads to • Better estimate of experimental variability • Increased precision of treatment comparisons • Assurance against aberrant results due to random chance • Increase in cost • Ideal sample sizes will minimize cost without sacrificing benefits of increased replication • How many times do we replicate treatments in Study 3.1?

  14. Error control – Blocking • Techniques to reduce experimental variability without sacrificing representativeness • Block: group of EU’s that are more similar than other EU’s • Idea: make treatment comparisons within each block and pool results together • Confounding: If we assign the same treatment to every EU in a block we can’t separate block and treatment effects

  15. Error control - Blocking • Example: Three different crop management practices are to be compared on a vineyard. • Have 3 vines, each vine is broken up into 3 sections Each section of a vine is EU Sections of one vine more similar than those of another Assign each treatment to a section of every vine Comparing treatments within each vine “cancels out” the characteristic of that vine Vine 1 Vine 2 Vine 3

  16. Error control - Blocking

  17. Error control - Blocking

  18. Error control - Blocking • Common blocking factors • Time • Proximity (closer EU’s more similar) • Physical characteristics (age, sex, etc.) • Useful to generalize experiments (not afraid of different EU’s) • If blocks are not a major source of variation we could increase experimental variability • Cannot assign block characteristics, otherwise they would be treatments

  19. Error control - covariates • Covariates: additional explanatory variables (categorical or continuous) that • Are measured just prior to or during treatment assignment • Could significantly reduce experimental variability • Requires assumptions about how the covariates and response are related • Could even be an interaction between covariates and treatment • Example: interested in multiple nutrition and workout regimes to help people lose weight • Potential weight loss likely restricted by starting weight • Measure weight prior to treatment

  20. Error control - covariates

  21. Error control - covariates

  22. Error control - covariates

  23. randomization • Given a fixed number of replications for each treatment, determine allowable assignments of treatments to EU’s (depends on experiment) • A design has been properly randomized when all allowable assignments are equally likely to be used • Example 1: 3 treatments, 9 EU’s, 3 replications per treatment. There are possible randomizations

  24. randomization • Example 2: Let’s say we group the EU’s into 3 blocks each with 3 EU’s. Each treatment appears in each block once (still 3 replications each). • In order to maintain that each treatment appears in each block, we have to randomize within each block • There are 3!=3*2*1=6 allowable randomizations in each block so there are a total of allowable randomizations • Block designs will always require separate randomizations

  25. randomization • Haphazard assignment does not equal proper randomization! • Avoids bias and accusations of it • Reduces chance of observing a treatment effect due to random chance alone (confounding) • Justifies statistical models and analysis in some cases (randomization tests) • Foundation for causal inference

  26. Ethical considerations • Obvious ethical considerations that need to be considered (e.g. human testing) • Clever experimental designs can maximize information using minimal resources and reduce impact on environment and animals • Ethical considerations/constraints can often lead to interesting design problems • Crossover designs • Clinical trials

More Related