Availability Bias & Sensitivity Analysis Problems with bad data.

Slides:



Advertisements
Similar presentations
Properties of Least Squares Regression Coefficients
Advertisements

Multiple Regression Analysis
Regression and correlation methods
EVAL 6970: Meta-Analysis Vote Counting, The Sign Test, Power, Publication Bias, and Outliers Dr. Chris L. S. Coryn Spring 2011.
Statistical Issues in Research Planning and Evaluation
4.3 Confidence Intervals -Using our CLM assumptions, we can construct CONFIDENCE INTERVALS or CONFIDENCE INTERVAL ESTIMATES of the form: -Given a significance.
LECTURE 3 Introduction to Linear Regression and Correlation Analysis
Chapter 19 Confidence Intervals for Proportions.
Multiple regression analysis
2.5 Variances of the OLS Estimators
Chapter 12 Simple Regression
BA 555 Practical Business Analysis
Econ Prof. Buckles1 Multiple Regression Analysis y =  0 +  1 x 1 +  2 x  k x k + u 1. Estimation.
Introduction to Inference Estimating with Confidence Chapter 6.1.
Evaluating Hypotheses
Heterogeneity in Hedges. Fixed Effects Borenstein et al., 2009, pp
Linear Regression 2 Sociology 5811 Lecture 21 Copyright © 2005 by Evan Schofer Do not copy or distribute without permission.
1 1 Slide © 2008 Thomson South-Western. All Rights Reserved Slides by JOHN LOUCKS & Updated by SPIROS VELIANITIS.
Standard Error of the Mean
Inference for regression - Simple linear regression
Statistics for Managers Using Microsoft Excel, 4e © 2004 Prentice-Hall, Inc. Chap 12-1 Chapter 12 Simple Linear Regression Statistics for Managers Using.
Overview of Meta-Analytic Data Analysis
Inference in practice BPS chapter 16 © 2006 W.H. Freeman and Company.
14. Introduction to inference
Daniel Acuña Outline What is it? Statistical significance, sample size, hypothesis support and publication Evidence for publication bias: Due.
CORRELATION & REGRESSION
1 Least squares procedure Inference for least squares lines Simple Linear Regression.
Inferences for Regression
Inference for a Single Population Proportion (p).
Inference for Linear Regression Conditions for Regression Inference: Suppose we have n observations on an explanatory variable x and a response variable.
Moderators. Definition Moderator - A third variable that conditions the relations of two other variables Example: SAT-Quant and math grades in school.
Review of Statistical Models and Linear Regression Concepts STAT E-150 Statistical Methods.
Statistical Applications for Meta-Analysis Robert M. Bernard Centre for the Study of Learning and Performance and CanKnow Concordia University December.
1 1 Slide © 2014 Cengage Learning. All Rights Reserved. May not be scanned, copied or duplicated, or posted to a publicly accessible website, in whole.
1 Psych 5500/6500 Standard Deviations, Standard Scores, and Areas Under the Normal Curve Fall, 2008.
Lecture 8 Simple Linear Regression (cont.). Section Objectives: Statistical model for linear regression Data for simple linear regression Estimation.
1 Multiple Regression Analysis y =  0 +  1 x 1 +  2 x  k x k + u.
Chapters 8 & 9 Linear Regression & Regression Wisdom.
The Campbell Collaborationwww.campbellcollaboration.org C2 Training: May 9 – 10, 2011 Introduction to meta-analysis.
Review of Building Multiple Regression Models Generalization of univariate linear regression models. One unit of data with a value of dependent variable.
1 Introduction Description of the video series Meta-analysis in R with Metafor.
Chapter 5 Parameter estimation. What is sample inference? Distinguish between managerial & financial accounting. Understand how managers can use accounting.
Section 10.1 Confidence Intervals
Chapter 8 Delving Into The Use of Inference 8.1 Estimating with Confidence 8.2 Use and Abuse of Tests.
Correlation Assume you have two measurements, x and y, on a set of objects, and would like to know if x and y are related. If they are directly related,
Issues concerning the interpretation of statistical significance tests.
Chapter 14: Inference for Regression. A brief review of chapter 4... (Regression Analysis: Exploring Association BetweenVariables )  Bi-variate data.
1 Prof. Dr. Rainer Stachuletz Multiple Regression Analysis y =  0 +  1 x 1 +  2 x  k x k + u 1. Estimation.
Marshall University School of Medicine Department of Biochemistry and Microbiology BMS 617 Lecture 11: Models Marshall University Genomics Core Facility.
Fixed- v. Random-Effects. Fixed and Random Effects 1 All conditions of interest – Fixed. Sample of interest – Random. Both fixed and random-effects meta-analyses.
Introduction to Multiple Regression Lecture 11. The Multiple Regression Model Idea: Examine the linear relationship between 1 dependent (Y) & 2 or more.
SAMPLING DISTRIBUTION OF MEANS & PROPORTIONS. SAMPLING AND SAMPLING VARIATION Sample Knowledge of students No. of red blood cells in a person Length of.
SAMPLING DISTRIBUTION OF MEANS & PROPORTIONS. SAMPLING AND SAMPLING VARIATION Sample Knowledge of students No. of red blood cells in a person Length of.
Chapter 9: Introduction to the t statistic. The t Statistic The t statistic allows researchers to use sample data to test hypotheses about an unknown.
Regression Analysis Presentation 13. Regression In Chapter 15, we looked at associations between two categorical variables. We will now focus on relationships.
Statistics for Business and Economics Module 1:Probability Theory and Statistical Inference Spring 2010 Lecture 4: Estimating parameters with confidence.
Chapter 10 Confidence Intervals for Proportions © 2010 Pearson Education 1.
Meta-analysis Overview
I-squared Conceptually, I-squared is the proportion of total variation due to ‘true’ differences between studies. Proportion of total variance due to.
Inference for Least Squares Lines
Testing for moderators
CHAPTER 12 More About Regression
9: Basic Confidence and Prediction Intervals
Chapter 8 Part 2 Linear Regression
Statistical Methods For Engineers
CHAPTER 29: Multiple Regression*
CHAPTER 26: Inference for Regression
CHAPTER 12 More About Regression
CHAPTER 12 More About Regression
Statistical inference for the slope and intercept in SLR
Presentation transcript:

Availability Bias & Sensitivity Analysis Problems with bad data

Models and Specification We let an equation ( aka model) stand for the data If the model is a good representation of the data, then our inferences are likely to be good. A model is said to be misspecified if the terms in the equation do not correspond correctly to the phenomenon we are representing In meta-analysis, there may be many unknown moderators that are missing from the equation, or there may be moderators included in the model that should be absent. Finding an adequate model is important.

Availability Problem Missing at random is okay for inference if the model is properly specified Nonrandom is a problem Sources of nonrandom samples of studies Publication bias – significant, small N Language bias (English; more stat sig) Availability to the researcher (friends help) Familiarity-stuff in one’s own field Duplication- significant study appears more than once in the lit Citation – significant studies more likely cited Dissertations; other filedrawer hard to find Deliberate misrepresentation for financial reasons

Bias Questions Evidence of any bias? Does Bias nullify the effect? If not, what is the impact and best estimate of the effect? How would we know? Registries of studies yet to be conducted From observed data themselves It’s hard to be conclusive about bias unless someone admits fraud

Summary Effects (Registry) Each dot is from a meta-analysis that contained both published studies and dissertations. Summary effects for both for the same content. Source: Ferguson & Brannick, 2012, Psych Methods

Forest Plot Drift to the right by precision? Effects of second hand smoke on cancer incidence. Source: Borenstein, Hedges, Higgins & Rothstein, 2009, p. 282

Precision Forest Plot in R Sort your effect sizes by V i or SE i before (or after) importing to R. Them run the forest plot.  install.packages("metafor")  library(metafor)  install.packages("xlsx")  library(xlsx)  dataRocks <- read.xlsx("/Users/michaelbrannick/Desktop/RocksLMX_AC2. xlsx",sheetName="Sheet1”)

Forest Plot Results I added the brown line. Though the funnel plot looks suspicious, this doesn’t look bad to me. LMX and Affective Commitment

Trim & Fill Creates symmetry; Adjusts summary ES Source: Borenstein, Hedges, Higgins & Rothstein, 2009, p. 287 Trim-and-fill is a kind of sensitivity analysis. Do not consider the results of the symmetrical (augmented) data as more valid than the original. More a ‘what if’ analysis.

Trim-and-fill in R Ordinary funnel

Trim-and-Fill Results Note the syntax required to get the plot. Print the results (e.g., resRocks2) to see the adjusted mean. Oddly enough, the program inferred that there were too many studies on the low side, not the high side. Therefore, the best estimate is probably the original, unadjusted one.

Forest Sleep & AD Sorted by ES. Note how the small studies are mostly all on the bottom right. This suggests availability bias.

Trim & Fill Sleep n AD This is one of the worst cases of APPARENT publication bias (funnel asymmetry) that I’ve seen.

Exercise Upload the dataset TnFEx.xlsx Compute a trim-and-fill analysis for the overall effect size. What is the overall ES before and after? What is your conclusion about publication bias and its implication for the overall effect size?

Cumulative Forest Source: Borenstein, Hedges, Higgins & Rothstein, 2009, p. 288 This sort of plot is being used to study the sociology of publication. You might expect that studies would follow the mean over time, departing randomly from that value. But no…

Cumulative Forest in Metafor If you have coded date of publication or some other marker for order of entry (date of publication is not always a good marker of when the study was done), you can use rma to compute this. In the object dataRocks (which I imported from Excel), there is a variable labeled ‘Year’ that I used to tell metafor (cumul program) the variable to use for sorting the effect sizes.

I added the brown line. It is clear from the analysis that the significance (rho = 0) was never an issue. This kind of analysis is common in medicine, where there are often very small effect sizes. There are some jumps in the progression. Notice how the confidence interval expands when there is a jump early on. This is because the discrepant study increases the REVC while decreasing the sampling variance. This is a random (varying) effects analysis. Cumulative Forest for LMX – AC Data

Egger’s Regression T i = effect size; v i =sampling variance of ES Should be flat (centered) if no bias (beta1 is zero). This shows small studies have higher values. Significant negative slope (beta1) is concerning. Source: Sutton (2009). In Cooper, Hedges, & Valentine (Eds) Handbook fo Research Synthesis Methods p. 441

Funnel Asymmetry Test Metafor offers several different tests for funnel plot asymmetry (variations on Egger’s test; see Viechtbaur 2010 for a list of references). The default test is shown below for the LMX data. The slope appears quite flat, and we have about 90 ES, so there is some power for the test. Thus funnel plot asymmetry indicative of publication bias does not appear likely in this example.

Radial Plot Crude approximate explanation: This plot shows how the effect size (y-axis) relates to the weight (x-axis). Line is drawn from zero on y-axis. Not a typical regression. Forced zero intercept. Assumed no effect of treatment. LMX – Affective Commitment Data >radial(resRocks) More technically accurate, but harder to remember: the plot shows the standardized ES as a function of the precision (square root of the weight).

Funnel SleepDat (reminder)

Radial (SleepRes)

Precision Forest (SleepDat)

Exercise – Funnel Test Recall the TnF dataset Run a test of funnel asymmetry on these data (use the standard RE meta-analysis model). What is your result? What is your interpretation of the finding?

Problems with Diagnostics Small study effects Small studies may actually have larger effects (not bias). What characteristics are associated with larger vs. smaller N studies? Heterogeneity Funnel, trim-and-fill designed for common-effects model Really should be residuals after removing moderators

What Else Can We Do? Maybe there is bias in the sample Maybe there is bad data Does it matter?

Sensitivity Analysis Outliers. Run twice. Share both results with your reader. Source: Greenhouse & Iyengar (2009). In Cooper, Hedges, & Valentine (Eds) Handbook fo Research Synthesis Methods p. 422

Finding Outliers (1) Externally standardized residuals are compared to a distribution of residuals from a model in which that observation is excluded. Metafor uses rstudent() to find these. Most meaningful in the context of meta-regression (moderator analysis).

Finding Outliers (2) This is the result of an analysis with no moderators – it is the overall result for the varying (random) effects analysis. Equivalent to a regression with intercept only.

Finding Outliers (3) This analysis considers the study sample size and the REVC in addition to the raw distance to the mean. Look for ‘large’ values – the z refers to the unit normal, so |2| is large, but I would probably start with 2.5 or 3. Your call though. I would export and sort with Excel (there are 92 residuals in this analysis). LMX data

Outliers 4 Here I have inserted a categorical independent variable into the model. The residuals are now studentized, deleted residuals that remain after accounting for differences in culture (vertical integration and collective orientation). As you can see, adding the moderator made very little difference for observation 12 (-2.08 to -2.05). For an important moderator, there could be a large difference.

If you show that the impact of outliers and nuisance variables (and moderators) is minimal or at least not a threat to your inferences, then your conclusions will be more credible.

Leave One Out In this sensitivity analysis, every study is removed, 1 by 1, and the overall ES is re-estimated. Helpful to judge the impact of influential studies (sometimes outlying ES, sometimes very large N) Can be run in rma if the model has no moderators (i.e., for overall ES or a subset of studies) Two uses for this Find and remove problematic studies Share summary of findings with your reader

Partial Results Because there are 92 ES in the file, I would write the file to Excel and then sort by estimate, I-square, whatever. But you can see that for the 20 ES shown, there is little impact from dropping each study.

Exercise Using the Devine data, run Outlier analysis (with publication vs. dissertation included in the model) Leave-one-out analysis (without the moderator) Delete any problematic studies (can use Excel) Rerun the analysis without problematic studies What is your new result? Does your conclusion or inference change?

Fail-safe N How many studies do we need to make the result n.s.? (Rosenthal) [this is an older analysis – I don’t recommend it.] How many studies do we need to make the summary effect small enough to ignore? (Orwin) kfs = failsafe studies kobt = studies in meta dobt = summary ES dc = desired lower boutnd dfs= studies with this (eg 0 or less) size needed to lower ES Corwin, R. G. (1983). A fail-safe N for effect size in meta-analysis. Journal of Educational Statistics, 8,

Favors Ground Favors Online Meta-analysis of 71 Samples (studies) comparing classroom instruction versus online instruction on tests of declarative knowledge (college classes and corporate training). Red line at bottom shows expected outcomes. Average difference near zero, but SUBSTANTIAL variability. Begs us to ask why the study outcomes are so different. Red is random effects. Orange is fixed effects. Add moderators.

Individual Studies Previous methods all concern the mean of studies or overall effect size. What if we are interested in the effect size for an individual study? If common effects, the effect size for all studies is the mean, so we have nothing to estimate. If varying effects, there is uncertainty about the true location of the study’s effect size. Bayesian thinking moves individual studies toward the mean. Can also get a confidence interval for the revised estimate.

Sensitivity Analysis -Tau Varying levels of tau-squared This graph shows the effect of the magnitude of tau-squared on the posterior estimates. As tau-squared increases, the effects spread out and there are some ordinal changes, too.