Study designs, randomization, bias errors, power, p-value, sample size
Download as PPTX, PDF15 likes1,409 views
This document provides an overview of key concepts in research study design, including randomization, bias/errors, power, p-values, and sample size. It discusses the importance of random sampling and random allocation in analytical studies with more than one group. Types of biases are described along with ways to reduce systematic error such as proper study design, randomization, and blinding. The relationship between random errors, power, and p-values is explained. Finally, sample size calculation and common effect sizes for different statistical tests are briefly covered.
Study designs, randomization, bias errors, power, p-value, sample size
- 1. Study designs, Randomization, Bias/Errors, Power, P-value, Sample Size Dhritiman Chakrabarti Assistant Professor, Dept of Neuroanaesthesiology and Neurocritical Care, NIMHANS, Bangalore
- 2. Why to do studies? • To answer research questions. • Starts with a deficiency or need of knowledge Literature review for gaining contemporary knowledge Identify a research question See novelty, feasibility, scientific validity of the question Formulate a study plan Study designs
- 5. Past Present Future Cross-sectional design Exposure Outcome Prospective Cohort design Exposure Outcome Retrospective Cohort design Exposure Outcome Case-Control design Exposure Outcome Ambispective Cohort design Exposure ≈ Outcome
- 7. Random Sampling and Random Allocation • For Descriptive studies and Analytical single group studies – Random sampling • For Analytical studies >1 group – Random allocation • Random Sampling methods: 1. Simple random sampling 2. Systematic sampling 3. Stratified sampling 4. Cluster sampling 5. Multistage sampling
- 8. • Random allocation methods: • Simple Randomization– Unequal groups • Block randomization – Equal groups • Stratified randomization – To match confounding variables.
- 9. Random sampling • Simple random – Its just random. • Systematic – Selecting cases after every sampling interval K. K = Sampling frame/Sample size. • Stratified sampling – Simple random sampling within defined “strata”. Stratification is dividing population based on one or more confounding variables. • Cluster sampling – Division of sampling frame into clusters Simple random sampling of clusters from list of all clusters Sampling all cases within each cluster. • Multistage sampling – Combination of above said methods at various stages.
- 10. Randomization/Random allocation • Block randomization: Sample size required (N); Number of groups(G); Block size(B); Number of blocks. • Requirements – Sample size has to be a multiple of block size; block size has to be a multiple of number of groups; Number of blocks = Sample size/block size. • Number of permutations of group order in blocks = B!/G!*(B-G)! • For G = 2, B=4, No. of combination of blocks = 4!/2!*(4-2)! = 6 = AABB, BBAA, ABAB, ABBA, BABA, and BAAB • Method: Start with G (suppose 2) Compute block size B as either G*2,G*3 (4,6) etc. Compute number of blocks required by N/B Compute number of combinations and line them up Randomize cases by the combination line on first come first serve basis.
- 11. Bias/Errors • Error Anything which causes result to deviate from true state. • Types Random (cause unknown) and Systematic (cause known). • Systematic error is Bias Any error which favors one outcome over other. • Many many types of biases.. • To reduce random error or error due to chance Improve sample size and representativeness of the sample to the population of interest. • To reduce systematic error Proper study design/protocol, randomization, blinding, proper data analysis.
- 12. Blinding • Open label study – No blinding. • Single blinding – Case is blinded to group allocation. • Double blinding – Case and Data collector is blinded to group allocation. • Triple blinding – Case, Data collector and data analyst is blinded to group allocation
- 13. Hypothesis • Based on contradiction type of experimental theory. • Your research question helps form the alternate hypothesis. • Opposite of the research question forms the null hypothesis. • Your aim by your experiment is to try and reject the null hypothesis and prove this rejection is not due to chance (or random error). • Suppose an experiment to find difference of outcome between two interventions. • Alternate There is a difference of outcome. • Null There is no difference of outcome. • Now imagine there are two levels of this scenario – one is reality and the other is your result.
- 14. Random errors, Power of study and P-value • Alpha error or Type I error – Finding a difference (in your result) when its not there (in reality) – Should be pre-specified as P- value and interpretation of results depends on it. • Beta Error or Type II error – Not finding a difference when it is there. • P-value – Probability that the difference found is due to random error or chance. • Power – The probability of avoiding Beta error – i.e. Finding a difference when it is there.
- 15. Sample size • Helps to reduce random error – which here is called sampling error – which helps increase power of study. • Sample size cannot be too high- wastage of resources and danger to patients; or too low- danger of not getting interpretable results Needs to be determined a priori. • Estimation is based on required “effect size” of primary outcome – based on previous studies, pilot data or clinically significant difference from an established parameter or statistic. • Also required are the Power of study and Alpha error level we want to avoid.
- 16. Effect sizes for different tests η2 Anova Omega-squared Anova Multivariate eta-squared one-way MANOVA Cohen's f one-way an(c)ova (regression) η2 Multiple regression Cohen's f Multiple Regression Cohen's d t-tests Cohen's ω chi-square Odds Ratios 2 by 2 tables Average Spearman rho Friedman test Pearson's r Pearson's Correlation
- 17. Softwares for Sample size calculation • G-Power • StatCalc by EpiInfo