class: top, right, inverse, title-slide # STT 2820 Chapter 12 ### Alan Arnholt ### updated: 2019-02-14 --- class: inverse, center, middle # Experiments and Observational Studies --- # Observational Studies-researchers don't assign choices, they observe them. -- 1. Retrospective---look back at past choices -- 2. Prospective---identify subjects and follow them through time. --- # Experiments and vocabulary An **experiment** is the application of treatments to randomly assigned subjects and the comparison of outcomes. -- In experiments, our explanatory variables are called **factors** and the outcomes are measured by **response variables**. The different values a factor can have are called **levels**. **levels `\(\times\)` levels = total treatments**. -- **Subject** or **participants** are people on whom an experiment is performed. **Experimental units** are non-people on whom an experiment is performed, say rats, petri dishes, test tubes, etc. --- class: inverse, center, middle # The Four Principles of Experimental Design --- # The Four Principles of Experimental Design 1. Control - make conditions as similar as possible for all treatment groups (control variation) -- 2. Randomize - equalize effects of unknown or uncontrollable variation (reduce bias & combat systematic error) -- 3. Replicate - apply treatment to multiple individuals, apply treatment at different times, locations, etc. (if something works on only one person, it is a story, not data) -- 4. Block - group similar individuals and randomize their group (removes identifiable variation) --- class: inverse, center, middle # Good Experiments --- # The best experiments are: -- 1. Randomized -- 2. Comparative -- 3. Double-blind -- 4. Placebo-controlled -- **Double blind** means that neither the subject nor the person applying the treatment know which treatment is being received. A **placebo** is a treatment that should not have an effect (like a sugar pill to cure cancer) --- # Example 12.1 You have 60 female volunteers who are willing to try anything to get rid of a headache. You want to test two pain medicines at 500 and 1000 mg doses. You also want to know whether resting or activity is more helpful is ending pain. Design a completely randomized, double-blind, placebo-controlled experiment. --- # Solution * Factors: Medicine and Rest -- * Medicine has three levels (placebo pill, 500mg, and 1000mg), Rest has two levels (rest, no rest) -- * Treatments: there are 3 `\(\times\)` 2 = 6 treatments. -- * 10 women are randomly assigned to each treatment. -- * All pills look the same, one person puts the pills into vials 1-30 and 31-60 with 10 of each kind -- * Subjects are randomly assigned numbers. -- * 1-30 rest, 31-60 no rest -- * Compare headache relief after 2 hours. --- class: inverse, center, middle # Interpretation Issues --- * A **lurking** variable is typically found in an observational study. -- * **Confounding** is a problem when the effects of different treatments can't be separated. Confounding is typically a problem for experiments. -- * **Statistically significant** means not random. We have a statistically significant result when differences are too large to be explained by natural variation.