Upload
bryan-hunt
View
217
Download
0
Embed Size (px)
Citation preview
BC Jung
A Brief Introduction to Epidemiology - XIII
(Critiquing the Research: Statistical Considerations)
Betty C. Jung, RN, MPH, CHES
BC Jung
Learning/Performance Objectives
Quick review – Basics of inferential statistics– Common measures of association
To be able to statistically critique studies – Statistical Caveats– Statistical Issues– Statistical Rules of Thumb
BC Jung
Introduction
Refresh your memory
– Basics of inferential statistics
– Common measures of associations used in epidemiologic studies
BC Jung
Measures of Association &Hypothesis Testing
Test Statistic =Observed Association - Expected Association
Standard Error of the Association Type I Error: Concluding there is an
association when one does not exist Type II Error: Concluding there is no
association when one does exist
BC Jung
Measures of Association Two Main Types of Measures
– Difference Measures (Two Independent Means, Two Independent Proportions, The Attributable Risk)
– Ratio Measures (Relative Risk, Relative Prevalence, Odds Ratio)
BC Jung
Measures of Association:Difference Measures
Two Independent Means Two Independent
Proportions The Attributable Risk
BC Jung
Attributable Risk (AR)
The difference between 2 proportions Quantifies the number of
occurrences of a health outcome that is due to, or can be attributed to, the exposure or risk factor
Used to assess the impact of eliminating a risk factor
BC Jung
Measures of Association:Ratio Measures
Relative Risk (RR) Relative Prevalence (RP) Odds Ratio (OR)
BC Jung
Strength of AssociationRelative Risk;(Prevalence); Odds Ratio Strength of
Association
0.83-1.00 1.0-1.2 None
0.67-0.83 1.2-1.5 Weak
0.33-0.67 1.5-3.0 Moderate
0.10-0.33 3.0-10.00 Strong
<0.01 >10.0 Approaching Infinity
Source: Handler,A, Rosenberg,D., Monahan, C., Kennelly, J. (1998) Analytic Methods in Maternal and Child Health. p. 69.
BC Jung
Caveats about Classifying Data All persons in an epidemiologic study
must be classifiable All study reports should clearly state
criteria used for classifying variables Studies that use different criteria for
defining the presence of any health state are not comparable with respect to reported rates of that health state
BC Jung
Caveats about Quantitative & Categorical Variables
Information on variability between persons is lost when quantitative data are categorized
Collapsing a quantitative variable into a categorical variable with two or more categories may obscure the fact that the underlying variable has a much larger range in one category than in another category
BC Jung
Caveats about Quantitative & Categorical Variables (Continued)
Be careful about comparing ranges because a larger sample will generally have a larger range
Collapsing quantitative variables into categories limits the choices of appropriate statistical tests of significance
Try using commonly used categories (as five- or ten-year age bands) to facilitate comparisons across studies
BC Jung
Berkson’s Fallacy
Associations based on hospital or clinic data are influenced by differential admission rates among groups of people
Similar source of selection bias occur when associations are based on autopsy data
BC Jung
Caveats about P-Values The size of the p-value has no relationship to the
potential practical significance of the findings The P-value reveals nothing about the
magnitude of effect (i.e., how much one group differs from another), or the precision of measurement (i.e., the amount of random error)
The nature of the sample, not the p-value, will determine whether inferences to the population of interest can be made (and the sample must be representative of the population)
BC Jung
Confidence Interval Estimation
Uses the sample mean to construct an interval (range) of numbers to estimate the effect
Provides some indication of how probable it is (e.g., 68%, 90%, 95%), or how “confident” one can be, that the true mean lies within the range of numbers in the interval estimate
BC Jung
Greenhalgh’s Questions to Ask About the Analysis (A)
Have the authors set the scene correctly? Have they determined whether their
groups are comparable, and, if necessary, adjusted for baseline differences?
What sort of data have they got, and have they used appropriate statistical tests?
BC Jung
Greenhalgh’s Questions to Ask About the Analysis (B)
If the authors have used obscure statistical tests, why have they done so and have they referenced them?
Are the data analyzed according to the original protocol?
Were paired tests performed on paired data?
BC Jung
Greenhalgh’s Questions to Ask About the Analysis (C)
Was a two-tailed test performed whenever the effect of an intervention could conceivably be a negative one?
Were “outliers” analyzed with both common sense and appropriate statistical adjustments?
Have assumptions been made about the nature and direction of causality?
BC Jung
Greenhalgh’s Questions to Ask About the Analysis (D)
Have “P values” been calculated and interpreted appropriately?
Have confidence intervals been calculated, and do the authors’ conclusions reflect them?
Have the authors expressed the effects of an intervention in terms of the likely benefit or harm which an individual patient can expect?
BC Jung
Statistical Issues:Epidemiological Studies
Logistic regression for binary outcomes
Cox regression for survival analysis Poisson distribution for disease
incidence or prevalence Odds ratio approximates relative
risk when disease is rare
BC Jung
Statistical Issues: Environmental Studies
Good statistical models are hard to come by
Publication bias can exaggerate excess risk
Odds ratios less than two (or greater than 0.5) can be interesting
BC Jung
Statistical Issues:Environmental Studies
What is the statistical basis for the environmental standard?
Variability vs. uncertainty What’s the quality of the
metadata Biomarkers as surrogates for
clinical outcomes
BC Jung
Statistical Issues:Risk Assessment
Hazard identification Dose-response evaluation Exposure assessment Risk characterization Risk management
BC Jung
Statistical Rules of Thumb
Use a logarithmic formulation to calculate sample size for cohort studies
Use no more than 4 or 5 controls per case for case-control studies
Obtain at least 10 subjects for every variable investigated for logistic regression
BC Jung
Statistical Rules of Thumb
Increase sample size in proportion to dropout rate. If dropout rate is expected to be 20%, then increase n/0.80
If dropout is greater than 20%, review reasons for dropouts
Accept substitutes with caution
BC Jung
Statistical Rules of Thumb
Choosing cutoff points Do not dichotomize unless absolutely
necessary Select an additive or multiplicative
model according to: theoretical justification, practical application, and computer implication
BC Jung
References
For Internet Resources on the topics covered in this lecture, check out my Web site:
http://www.bettycjung.net/ Other lectures in this series:
http://www.bettycjung.net/Bite.htm