Chapter 10 Design
- 1. Chapter 10 Statistical Principles of Experimental Design
- 2. The aims of research design: To get a reliable result in the lowest cost of manpower, time and money; To estimate the random error within the observed data; To promote the efficiency of the research There are two kinds of researches : Laboratory experiment and Clinical trial Medical survey
- 3. 10.1 Principles of research design 1. Control 2. Balance 3. Randomization 4. Replication
- 4. 1. Control If there is no control, then there is no comparison. Others Effect of others Effect of treatment Treatment Subject Others Control Effect of control Subject Effect of others
- 5. Some forms of control: (1) Empty control (2) Placebo control (for animal, Experimental control) (3) Mutual control (4) Self control (5) Standard control (6) Historical control
- 6. (1) Empty control Others Effect of others Effect of treatment Treatment Subject A Others Subject B Effect of others
- 7. (2) Placebo control Others Effect of others Effect of treatment Treatment Subject A Others Placebo Effect of placebo Subject B Effect of others
- 8. (3) Mutual control Others Effect of others Effect of Treatment A Treatment I Subject A Others Treatment II Effect of Treatment B Subject B Effect of others
- 9. (4) Self control Others Effect of others Effect of treatment Treatment Subject Others Control Effect of control Subject Effect of others
- 10. (5) Standard control There is no control group, but compare with certain “standard” Others Effect of others Effect of treatment Treatment Subject
- 11. (6) Historical control There is no control group, but compare with “historical result” Others Effect of others Effect of treatment Treatment Subject
- 12. 2. Balance: The experimental group and control group are almost the same in all aspects except the treatment. Others Effect of others Effect of treatment Treatment Subject Others Control Effect of control Subject Effect of others
- 13. 3. Randomization Many factors, we know that they may influence the results, but they are slight and very difficult to deal with – Randomization is the best choice! Example To improve the homogeneity of subjects, collect a number of students with same age and gender; randomly arrange them into two groups to make them comparable in height and weight.
- 14. Randomization is the prerequisite of statistical inference. Randomization Casual Randomization means that all subjects in population have same probability to be sampled out for research.
- 15. 4. Replication One meaning of replication : The results can be reproduced in different labs and by different researchers. Another meaning of replication : The study should be performed in a big enough sample. Altman & Dore checked 90 papers: 39% mentioned their sample size and why. Sample sizes of 27% papers were too small to make a conclusion.
- 16. How to estimate sample size in the design stage? Four parameters are needed: (1) : the maximal probability of type error is allowed (2) : the maximal probability of type I error is allowed (3) : the minimal difference between two means is allowed (4) : the standard deviation among subjects in the same group
- 17. Calculate by Example 13-1 A pilot study shows, the sample mean and standard deviation of pulse among male patients with lead poisoning are 67/min and 5.97/min respectively. To test whether the population mean of this kind of patients is lower than the mean of normal males (one-side test), how many cases are needed? Given , take 1. Comparison between the mean and a given number
- 18. 2. Comparison between two means of two independent samples Calculate by Example 13-2 Comparing mean reductions of blood sedimentation between drug A and B. To test whether the effects of two drugs are different (two-side), how many cases are needed? A pilot study shows, . Take , .
- 19. 3. Comparison between two frequencies of two independent samples Calculate by Example 13-3 Comparing two chemotherapies for lymphoma, how many cases are needed? The pilot study shows, the remission rates are . Given .
- 20. 10.2 Experimental design Why? To plan and arrange subject selection, treatment assignment, data collection and statistical analysis To make sure validity, reproducibility and economy. 2. Types of research Experiment: animal experiment, clinical trial, community intervention trial Survey Both need well design !
- 21. Subject: Subject could be: gene, protein, cell, tissue, animal, patient or healthy population. Subject should be clearly defined, and homogeneous. Eligible subjects Treatment effects Control 3. Three elements of experimental design
- 22. 2) Treatment: A measure used to intervene a life process. Treatment factor: drug Levels of treatment factor: dose A, dose B Treatment factor and non-treatment factors should be identified. Non-treatment factors : age, gender, disease status, weather, environment … Treatment should be standardized: 3) Effect: Change caused by treatment. Effect measurement should be objective, accurate, precise, specific and sensitive.
- 23. (1) Completely random design Randomly allocate the subjects into two or more groups Independently sampling from two or more populations Example 13-4 Randomly allocate 10 animals into two groups. (Using Table 16, read 0-9 only) A: 1, 4, 6, 8, 9 B: 2, 3, 5, 7, 10 4.Commonly used experimental designs
- 24. Example 13-5 Randomly allocate 15 animals into three groups. A: 4, 6, 8, 11, 15 B: 3, 5, 9, 12, 14 C: 1, 2, 7, 10, 13 Data analysis: t test/ANOVA; rank sum test Advantage: Simple; stable Disadvantage: If there are many confounders, they may not be balanced by randomization only.
- 25. (2) Paired design Example 13-6 Randomly allocate 8 pairs of subjects into two groups (read 0-7 only, odd for AB) A: 1-1, 2-2, 3-2, 4-1, 5-2, 6-1, 7-2, 8-1; B: the rest Data analysis: Paired t test; signed rank sum test Advantage: Well control the non-treatment factors, Save sample size Disadvantage: Difficult to perform in practice Used for short period studies
- 26. (3) Random block design Example 13-7 Four groups of subjects; four treatments: A, B, C, D.
- 27. Data analysis: Analysis of variance for random block design Advantage: The individuals in same block are quite similar so that the comparison within block is more sensitive than that between completely randomized groups. Disadvantage: if any one individual observation is missed, the data analysis in the block will be very difficult.
- 28. (4) Cross-over design 1st period wash-out 2nd period Group 1 A none B Group 2 B none A Example 13-8 If there are 16 patients, then randomly allocate them into group 1 and 2 (same as Completely randomized design). If there are 8 pairs of patients, then randomly allocate two subjects within each pair into group 1 and 2 (same as paired design).
- 29. Data analysis: Analysis of variance for cross-over design Advantage: Well control the non-treatment factors, save sample size; Everyone receives both treatments – Equity Disadvantage: Assume the subjects keep the same in 1 st period and 2 nd period -- it can only be used for chronic diseases; Wash-out period is required – the treatment should be stopped during wash-out period
- 30. Blinding and Placebo Necessary for clinical trial. Blinding (Masking) -- To reduce the bias caused by psychological effect of knowing the treatment Single blinding: Any patient does not know what treatment is taken. Double blinding: Both patient and physician do not know the treatment. Placebo – To ensure blinding Everything should be the same to the treatment group, except that it does not contain any effective components of the treatment. Special skills are needed for blinding and placebo.
- 31. 1. Survey Observe the existing process Without intervention Well design Example for surveys: Health condition survey Epidemiological survey Etiologic survey Clinical follow up survey Sanitary survey …… . 10.3 Survey design
- 32. 2. Design (1) The purpose of survey , clear (2) The population concerned , well defined . (3) Space, time and sample size , specified (4) Observed unit , well decided individual? Family? Class? (5) The questionnaire , well designed Items , carefully chosen Language , clear and specific Possible answers , well coded in advance
- 33. (6) Data collection Direct observation Measurement, observation, test, count … Interview Form filling Group meeting Telephone or internet ( the response rate ?)
- 34. (1) Overall survey (Complete survey) Example: National Census To get the population parameters directly There is no sampling error, but non-sampling error is relatively high (2) Sampling survey Sampling – Observation – Statistical inference – knowledge about population Advantage: Efficient Disadvantage: Complicated in design, implementation and analysis 3. Classification of survey
- 35. Simple random sampling Every individual has same probability to be sampled. Suitable for small population. Systematic sampling According to the sequence of individuals, to sample subjects in a fixed interval. When population has a sequence number ( such as ID code), it is convenient to carry out. Methods of random sampling
- 36. Stratified sampling The population is stratified according some factors that may influence the results of study and then the individuals in strata were randomly sampled. Some important confounding factors could be controlled by stratified sampling. Cluster sampling If the individuals belong to certain unit ( such as community, school, class, city, county), we may directly sample the unit other than individuals.
- 37. Stratified sampling < Systematic sampling < Simple random sampling < Cluster sampling Sampling error
- 38. (3) Typical survey (case survey) Advantage: Only a few typical subjects are observed, which well reflect the main characteristics of same kind of subjects. Disadvantage: There is no ground for any statistical inference.
- 39. (4) Case-control survey (retrospective study) Good for rare diseases : Outcome Exposure (Cause?) “ Case” : Patients “ Control” : Non patients, main confounders are similar to the patients Comparing the frequency (or intensity) of exposure 1:1 case-control; 1:m case-control group – group case control
- 40. (5) Cohort survey (Follow-up study, prospective study) Exposure ( Cause?) Outcome Different groups of exposure Comparing the outcomes of different groups: Does different condition of exposure leads to different outcome?
- 41. Summary 1. Commonly used Experimental Designs Completely random design Paired design Random block design Cross-over design
- 42. 2. Content of Survey Design Purpose Population concerned Space, time and sample size Observed unit Questionnaire Data collection
- 43. 3.Classification of survey Overall survey Sampling survey Sampling error Stratified sampling <Systematic sampling <Simple random sampling <Cluster sampling Typical survey Case-control survey Cohort survey