Statistical Issues in Data Collection and Study Design For Community Programs and Research October...
-
Upload
alfred-norris -
Category
Documents
-
view
215 -
download
0
Transcript of Statistical Issues in Data Collection and Study Design For Community Programs and Research October...
Statistical Issues in Data Collection and Study Design For
Community Programs and Research
October 11, 2001
Elizabeth Garrett
Division of Biostatistics
Department of Oncology
Overview
• Goals of Data Collection and Study Design
• Key concepts– Reliability – Validity
• “Latent” Constructs
• Study Designs
• Potential Biases
Goals of Data Collection
Two broad goals*
– evaluation of intervention• controlled
• uncontrolled
– summary of population• demographics
• attitudes
The Goal of Study Design
To devise a model for some complex etiologic or clinical process that
gives valid and precise inference.
Issue Specific to Interventions
• Outcomes tend to be “soft”
• Not always an easily quantified response
• We often use one or more “surrogates” to measure outcomes.
Key Concepts• Reliability: Is the data that you are collecting a reliable or
reproducible measure?• Has to do with how closely your measure correlates with the
underlying construct you want to measure.
• Truth = Observed Data + Error– If you collected the same data tomorrow, would you get the same
answer?– If you ask a related question, will the two questions have
correlated answers?– If two different “raters” (i.e. collectors) evaluate the same
individual, do they get the same data?
Validity
• Validity: Is what you are collecting measuring what you want to measure?– Face validity– Construct validity– Criterion validity– Etc.
Valid
Invalid
Reliable(but still invalid!)
Valid and Reliable
Validity
• Internal Validity:– Valid for the population from which you
sampled
• External Validity– Generalizable to a broader population
Latent Constructs
• Definition: A latent variable is a variable that cannot be directly measured.
• Examples:– Quality of Life– Socio-economic status– Distress– Depression
Latent Constructs
• Need to be measured using multiple variables• Variables, taken together, should “define” the construct• Methods should be decided upon ahead of time and
data collection needs to be considered.• Examples: latent class analysis, factor analysis• Coding is important
– likert scale: “On a scale of 1 to 7…..”– binary: yes/no, present/absent– continuous: age, income
Some Study Design Types• Cross-sectional, No Intervention
– Attributes• quantify community
• “summarize” attitudes, demographics, etc.
• descriptive statistics– means, medians, standard deviations
– “pictures” of the sample: histograms, boxplots
• “hypothesis generating”, and NOT “hypothesis testing”
• simplest conceptually
• Cross-sectional, No Intervention (cont.)– Issues to think about
• sampling– Who?
– When?
– Where?
• Data (this is general to ALL study designs)– format?
– Binary versus continuous versus ordinal versus categorical?
– open-ended?
• Pre-post Design, One group (uncontrolled)– Was intervention successful?– Attributes:
• Compare baseline to follow-up
• simplest when only two time points are collected.
• Convenient that each individual serves as his/her own control
• Hypothesis testing: – Ho: intervention worked
– Ha: intervention did not work
• Some methods: binomial tests, signed rank test, paired t-test, regression methods
– Issues to think about• when should “success” be measured?
– 1 week? 1 month? Both?
– What if effect at 1 month but “washed out” by 6 months?
• How is success measured?
– Yes/no? Continuous change in score?
• Learning effect
– bias of questionnaires
– is this the most appropriate design if there is a potential learning effect?
• “Placebo” effect could play a role.
• Adherence!
– Is the effect of intervention smaller than anticipated because some study participants did not adhere?
• Confounders and effect modifiers!
– Are there some individuals that respond more strongly to the intervention than others in such a way that is predictable (e.g. age? weight? political views?)
• Pre-Post, Two Groups (Controlled)– Does intervention group improve more than the
control group?– Attributes
• similar to pre-post, one group
• can quantify placebo and learning effects (caveat)
• hypothesis testing:– Ho: effect in control group = effect in treatment group
– Ha: effect in control group effect in treatment group
• Some methods: 2 sample t-test, rank sum test, fisher’s exact test, regression methods
– Issues to think about• We have a measure of placebo effect
• blinding or masking? Is it possible?
• Randomization– Balance?
– Stratification necessary?
– Matched?
• ITT versus Treatment received?
• Drop out
• Adherence
Other Study Designs
• Case-Control Studies
• Cohort Studies (aka Prospective Study)
• Ecologic Study
Potential Biases to Keep in Mind• Selection Bias (IV)
– individuals who join the study are not representative of the population in a way that affects the outcome.
• Information Bias (IV)– measures tend to be biased in one direction
• Confounding (IV)– Mixing of effects leads to wrong inference
• Effect Modification (IV)– effect of treatment depends on another factor (e.g. age)