DOCSLIB.ORG

False Discovery Rate Control Is a Recommended Alternative to Bonferroni-Type Adjustments in Health Studies Mark E

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
False Discovery Rate Control Is a Recommended Alternative to Bonferroni-Type Adjustments in Health Studies Mark E Author's personal copy Journal of Clinical Epidemiology 67 (2014) 850e857 REVIEW ARTICLES False discovery rate control is a recommended alternative to Bonferroni-type adjustments in health studies Mark E. Glickmana,b,*, Sowmya R. Raoa,c, Mark R. Schultza aCenter for Health care Organization and Implementation Research, Bedford VA Medical Center, 200 Springs Road (152), Bedford, MA 01730, USA bDepartment of Health Policy and Management, Boston University School of Public Health, 715 Albany Street, Talbot Building, Boston, MA 02118, USA cDepartment of Quantitative Health Sciences, University of Massachusetts Medical School, 55 N. Lake Avenue, Worcester, MA 01655, USA Accepted 12 March 2014; Published online 13 May 2014 Abstract Objectives: Procedures for controlling the false positive rate when performing many hypothesis tests are commonplace in health and medical studies. Such procedures, most notably the Bonferroni adjustment, suffer from the problem that error rate control cannot be local- ized to individual tests, and that these procedures do not distinguish between exploratory and/or data-driven testing vs. hypothesis-driven testing. Instead, procedures derived from limiting false discovery rates may be a more appealing method to control error rates in multiple tests. Study Design and Setting: Controlling the false positive rate can lead to philosophical inconsistencies that can negatively impact the practice of reporting statistically significant findings. We demonstrate that the false discovery rate approach can overcome these inconsis- tencies and illustrate its benefit through an application to two recent health studies. Results: The false discovery rate approach is more powerful than methods like the Bonferroni procedure that control false positive rates. Controlling the false discovery rate in a study that arguably consisted of scientifically driven hypotheses found nearly as many sig- nificant results as without any adjustment, whereas the Bonferroni procedure found no significant results. Conclusion: Although still unfamiliar to many health researchers, the use of false discovery rate control in the context of multiple testing can provide a solid basis for drawing conclusions about statistical significance. Ó 2014 Elsevier Inc. All rights reserved. Keywords: False discovery rate; False positive rate; FWER; Multiple tests; P-value; Study-wide error rate 1. Introduction made, with plenty of advocates on each side of the argu- ment. It appears doubtful that researchers will coalesce We are now in an age of scientific inquiry where health behind a unified point of view any time soon. and medical studies are routinely collecting large amounts Significance level adjustments that control study-wide of data. These studies typically involve the researcher at- error rates are still common in peer-reviewed health studies. tempting to draw many inferential conclusions through An examination of recent issues of several highly cited numerous hypothesis tests. Researchers are typically medical and health journals (Journal of the American Med- advised to perform some type of significance-level adjust- ical Association, New England Journal of Medicine, Annals ment to account for the increased probability of reporting of Internal Medicine, and Medical Care) reveals an abun- false positive results through multiple tests. Such adjust- dant use of multiple-test adjustments that control study- ments are designed to control study-wide error rates and wide error rates: We found 191 articles published in 2012 lower the probability of falsely rejecting true null hypothe- to 2013 making some adjustment for multiple testing, with ses. The most commonly understood downside of these 102 (53.4%) performing the Bonferroni or another study- procedures is the loss of power to detect real effects. Argu- wide error adjustment. Some other studies reported explic- ments have been put forth over the years whether adjust- itly, and almost apologetically, that they had not performed ments for controlling study-wide error rates should be an adjustment, and some even reported consequences of not having adjusted for multiple tests. Despite the continued popularity of multiple test adjust- Conflict of interest: None. Funding: None. ments in health studies that control false positive error * Corresponding author. Tel.: 781-687-2875; fax: 781-687-3106. rates, we argue that controlling the false discovery rate E-mail address: [email protected] (M.E. Glickman). [1] is an attractive alternative. The false discovery rate 0895-4356/$ - see front matter Ó 2014 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.jclinepi.2014.03.012 Author's personal copy M.E. Glickman et al. / Journal of Clinical Epidemiology 67 (2014) 850e857 851 in-roads into more general health studies. Of the 191 arti- What is new? cles we found in highly cited journals that mention adjust- ments for multiple tests, only 14 (7.3%) include false Controlling the false positive rate to address multi- discovery rate adjustments. plicity of tests in health studies can result in logical This article is intended to remind readers of the funda- inconsistencies and opportunities for abuse. mental challenges of multiple-test adjustment procedures Errors in hypothesis test conclusions depend on the that control study-wide error rates and explain why false frequency of the truth of null hypotheses being discovery rate control may be an appealing alternative for tested. drawing statistical inferences in health studies. In doing so, we distinguish between tests that are exploratory and False discovery rate control procedures do not suf- those that are hypothesis driven. The explanations we pre- fer from the philosophical challenges evident with sent to discourage use of adjustments based on study-wide Bonferroni-type procedures. error rate control are not newdthe case has been made Health researchers may benefit from relying on strongly over the past 10 to 20 years [4e9]. Arguments false discovery rate control in studies with multiple in favor of using false discovery rate control have been tests. made based on power considerations [10e13], but we are unaware of explanations based on the distinction between exploratory and hypothesis-driven testing. is the expected fraction of tests declared statistically signif- 2. Adjustment for multiple testing through false posi- icant in which the null hypothesis is actually true. The false tive rate control discovery rate can be contrasted with the false positive rate, which is the expected fraction of tests with true null The usual argument to convince researchers that adjust- hypotheses that are mistakenly declared statistically ments are necessary when multiple tests are performed is to significant. In other words, the false positive rate is the point out that, without adjustments, the probability of at least probability of rejecting a null hypothesis given that it is one null hypothesis being rejected is larger than acceptable true, while the false discovery rate is the probability that levels. Suppose, for example, that a researcher performs 100 a null hypothesis is true given that the null hypothesis has tests at the a 5 0.05 significance level in which the null been rejected. hypothesis is true in every case. If all the tests are indepen- Table 1 illustrates the distinction between the false pos- dent, then the probability that at least one test would be incor- itive and false discovery rates. Suppose that a set of tests rectly rejected is 1 À (1 À 0.05)100 5 0.9941, or 99.41%. In can be cross-classified into a 2 Â 2 table according to truth most studies, tests are not independent (eg, when tests share of the hypotheses (whether the null hypothesis is true or the same data), in which case, the probability of at least one not), and the decision made based on the data (whether to incorrect rejection would not be quite so large, though likely reject the null hypothesis or not). Let a be the fraction of large enough to be of some concern. Recognizing that the tests with true null hypotheses that are not rejected, b be probability of at least one false positive may be unacceptably the fraction of tests with true null hypotheses that are large, a common strategy is to adjust the significance level as a mistakenly rejected, c be the fraction of tests with false null function of the number of tests performed. hypotheses that are mistakenly not rejected, and d be the One of the simplest and most commonly used ap- fraction of tests with false null hypotheses that are rejected. proaches to adjusting significance levels is the Bonferroni Assuming these fractions can be viewed as long-run rates, procedure [14,15]. Letting n be the number of tests per- the false positive rate is computed as b/(a þ b), whereas formed, and a the significance level one would normally the false discovery rate is computed as b/(b þ d ). Although use if performing only one test, the Bonferroni procedure the numerators of these fractions are the same, the denom- involves rejecting null hypotheses whose P-values are less inator of the false positive rate is the rate of encountering than a/n rather than a. For example, if a study involves 100 true null hypotheses, and the denominator of the false dis- hypothesis tests and the researcher would ordinarily use covery rate is the overall rate of rejecting null hypotheses. a 5 0.05 as the significance level for a single test, then Conventional hypothesis testing, along with procedures the Bonferroni procedure requires the researcher to to control study-wide error rates, are set up to limit false compare each of the 100 P-values to a/n 5 0.05/ positive rates, but not false discovery rates. False discovery 100 5 0.0005. By invoking this procedure, the researcher rate control has become increasingly standard practice in is guaranteed that the probability of at least one false pos- genomic studies and the analysis of micro-array data where itive, regardless of the dependence among the tests, is no an abundance of testing occurs.
Load more

Recommended publications
  • Underestimation of Type I Errors in Scientific Analysis
    Open Access Journal of Mathematical and Theoretical Physics Research Article Open Access Underestimation of type I errors in scientific analysis Abstract Volume 2 Issue 1 - 2019 Statistical significance is the hallmark, the main empirical thesis and the main argument Roman Anton to quantitatively verify or falsify scientific hypothesis in science. The correct use of Department of Theoretical Sciences, The University of Truth and statistics assures and increases the reproducibility of the results and findings, and Common Sense, Germany there is recently a widely wrong use of statistics and a reproducibility crisis in science. Today, most scientists can’t replicate the results of their peers and peer-review hinders Correspondence: Roman Anton, Department of Theoretical them to reveal it. Besides an estimated 80% of general misconduct in the practices, Sciences, The University of Truth and Common Sense, Germany, and above 90% of non-reproducible results, there is also a high proportion of generally Email false statistical assessment using statistical significance tests like the otherwise valuable t-test. The 5% confidence interval, the widely used significance threshold Received: October 29, 2018 | Published: January 04, 2019 of the t-test, is regularly and very frequently used without assuming the natural rate of false positives within the test system, referred to as the stochastic alpha or type I error. Due to the global need of a clear-cut clarification, this statistical research paper didactically shows for the scientific fields that are using the t-test, how to prevent the intrinsically occurring pitfalls of p-values and advises random testing of alpha, beta, SPM and data characteristics to determine the right p-level, statistics, and testing.
    [Show full text]
  • The Reproducibility of Research and the Misinterpretation of P Values
    bioRxiv preprint doi: https://doi.org/10.1101/144337; this version posted August 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC 4.0 International license. 1 The reproducibility of research and the misinterpretation of P values David Colquhoun University College London Email: [email protected] Revision 6 Abstract We wish to answer this question If you observe a “significant” P value after doing a single unbiased experiment, what is the probability that your result is a false positive?. The weak evidence provided by P values between 0.01 and 0.05 is explored by exact calculations of false positive rates. When you observe P = 0.05, the odds in favour of there being a real effect (given by the likelihood ratio) are about 3:1. This is far weaker evidence than the odds of 19 to 1 that might, wrongly, be inferred from the P value. And if you want to limit the false positive rate to 5 %, you would have to assume that you were 87% sure that there was a real effect before the experiment was done. If you observe P = 0.001 in a well-powered experiment, it gives a likelihood ratio of almost 100:1 odds on there being a real effect. That would usually be regarded as conclusive, But the false positive rate would still be 8% if the prior probability of a real effect was only 0.1.
    [Show full text]
  • From P-Value to FDR
    From P-value To FDR Jie Yang, Ph.D. Associate Professor Department of Family, Population and Preventive Medicine Director Biostatistical Consulting Core In collaboration with Clinical Translational Science Center (CTSC) and the Biostatistics and Bioinformatics Shared Resource (BB-SR), Stony Brook Cancer Center (SBCC). OUTLINE P-values - What is p-value? - How important is a p-value? - Misinterpretation of p-values Multiple Testing Adjustment - Why, How, When? - Bonferroni: What and How? - FDR: What and How? STATISTICAL MODELS Statistical model is a mathematical representation of data variability, ideally catching all sources of such variability. All methods of statistical inference have assumptions about • How data were collected • How data were analyzed • How the analysis results were selected for presentation Assumptions are often simple to express mathematically, but difficult to satisfy and verify in practice. Hypothesis test is the predominant approach to statistical inference on effect sizes which describe the magnitude of a quantitative relationship between variables (such as standardized differences in means, odds ratios, correlations etc). BASIC STEPS IN HYPOTHESIS TEST 1. State null (H0) and alternative (H1) hypotheses 2. Choose a significance level, α (usually 0.05) 3. Based on the sample, calculate the test statistic and calculate p-value based on a theoretical distribution of the test statistic 4. Compare p-value with the significance level α 5. Make a decision, and state the conclusion HISTORY OF P-VALUES P-values have been in use for nearly a century. The p-value was first formally introduced by Karl Pearson, in his Pearson's chi-squared test and popularized by Ronald Fisher.
    [Show full text]
  • Statistical Significance for Genome-Wide Experiments
    Statistical Significance for Genome-Wide Experiments John D. Storey∗ and Robert Tibshirani† January 2003 Abstract: With the increase in genome-wide experiments and the sequencing of multi- ple genomes, the analysis of large data sets has become commonplace in biology. It is often the case that thousands of features in a genome-wide data set are tested against some null hypothesis, where many features are expected to be significant. Here we propose an approach to statistical significance in the analysis of genome-wide data sets, based on the concept of the false discovery rate. This approach offers a sensible balance between the number of true findings and the number of false positives that is automatically calibrated and easily interpreted. In doing so, a measure of statistical significance called the q-value is associated with each tested feature in addition to the traditional p-value. Our approach avoids a flood of false positive results, while offering a more liberal criterion than what has been used in genome scans for linkage. Keywords: false discovery rates, genomics, multiple hypothesis testing, p-values, q-values ∗Department of Statistics, University of California, Berkeley CA 94720. Email: [email protected]. †Department of Health Research & Policy and Department of Statistics, Stanford University, Stanford CA 94305. Email: [email protected]. 1 Introduction Some of the earliest genome-wide analyses involved testing for linkage at loci spanning a large portion of the genome. Since a separate statistical test is performed at each locus, traditional p- value cut-offs of 0.01 or 0.05 had to be made stricter to avoid an abundance of false positive results.
    [Show full text]
  • Cochrane Handbook for Systematic Reviews of Diagnostic Test Accuracy
    Cochrane Handbook for Systematic Reviews of Diagnostic Test Accuracy Chapter 11 Interpreting results and drawing conclusions Patrick Bossuyt, Clare Davenport, Jon Deeks, Chris Hyde, Mariska Leeflang, Rob Scholten. Version 0.9 Released December 13th 2013. ©The Cochrane Collaboration Please cite this version as: Bossuyt P, Davenport C, Deeks J, Hyde C, Leeflang M, Scholten R. Chapter 11:Interpreting results and drawing conclusions. In: Deeks JJ, Bossuyt PM, Gatsonis C (editors), Cochrane Handbook for Systematic Reviews of Diagnostic Test Accuracy Version 0.9. The Cochrane Collaboration, 2013. Available from: http://srdta.cochrane.org/. Saved date and time 13/12/2013 10:32 Jon Deeks 1 | P a g e Contents 11.1 Key points ................................................................................................................ 3 11.2 Introduction ............................................................................................................. 3 11.3 Summary of main results ......................................................................................... 4 11.4 Summarising statistical findings .............................................................................. 4 11.4.1 Paired summary statistics ..................................................................................... 5 11.4.2 Global measures of test accuracy ....................................................................... 10 11.4.3 Interpretation of summary statistics comparing index tests ............................... 12 11.4.4 Expressing
    [Show full text]
  • Redefine Statistical Significance We Propose to Change the Default P-Value Threshold for Statistical Significance from 0.05 to 0.005 for Claims of New Discoveries
    comment Redefine statistical significance We propose to change the default P-value threshold for statistical significance from 0.05 to 0.005 for claims of new discoveries. Daniel J. Benjamin, James O. Berger, Magnus Johannesson, Brian A. Nosek, E.-J. Wagenmakers, Richard Berk, Kenneth A. Bollen, Björn Brembs, Lawrence Brown, Colin Camerer, David Cesarini, Christopher D. Chambers, Merlise Clyde, Thomas D. Cook, Paul De Boeck, Zoltan Dienes, Anna Dreber, Kenny Easwaran, Charles Efferson, Ernst Fehr, Fiona Fidler, Andy P. Field, Malcolm Forster, Edward I. George, Richard Gonzalez, Steven Goodman, Edwin Green, Donald P. Green, Anthony Greenwald, Jarrod D. Hadfield, Larry V. Hedges, Leonhard Held, Teck Hua Ho, Herbert Hoijtink, Daniel J. Hruschka, Kosuke Imai, Guido Imbens, John P. A. Ioannidis, Minjeong Jeon, James Holland Jones, Michael Kirchler, David Laibson, John List, Roderick Little, Arthur Lupia, Edouard Machery, Scott E. Maxwell, Michael McCarthy, Don Moore, Stephen L. Morgan, Marcus Munafó, Shinichi Nakagawa, Brendan Nyhan, Timothy H. Parker, Luis Pericchi, Marco Perugini, Jeff Rouder, Judith Rousseau, Victoria Savalei, Felix D. Schönbrodt, Thomas Sellke, Betsy Sinclair, Dustin Tingley, Trisha Van Zandt, Simine Vazire, Duncan J. Watts, Christopher Winship, Robert L. Wolpert, Yu Xie, Cristobal Young, Jonathan Zinman and Valen E. Johnson he lack of reproducibility of scientific not address the appropriate threshold for probabilities. By Bayes’ rule, this ratio may studies has caused growing concern confirmatory or contradictory replications be written as: over the credibility of claims of new of existing claims. We also do not advocate T Pr Hx discoveries based on ‘statistically significant’ changes to discovery thresholds in fields ()1obs findings.
    [Show full text]
  • Benjamini-Hochberg Method
    P-values and statistical tests 7. Multiple test corrections Marek Gierliński Division of Computational Biology Hand-outs available at http://is.gd/statlec 2 Lets perform a test � times False positives True positives H0 true H0 false Total Number of Significant �� �� � discoveries Not �� �� � − � significant Total �( �) � Number of tests True negatives False negatives 3 Family-wise error rate ���� = Pr (�� ≥ 1) 4 Probabilities of independent events multiply 1 1 � � and � = �(�)×�(�) 2 4 1 2 1 2 Toss two coins 1 2 1 2 1 2 5 Probabilities of either event is 1 − 1 − � = 1 1 � � or � = ? 2 4 1 � � = 1 − �(�) 2 1 1 2 4 � � and � = � � ×� � = Toss two = 1 − � � coins 1 1 � � or � = 1 − � � and � 2 4 1 = 2 � � or � = 1 − 1 − � � 1 1 2 4 1 = 3 �(� or �) = 1 − 1 − = 2 4 6 False positive probability H0: no effect Set � = 0.05 One test Probability of having a false positive �) = � Two independent tests Probability of having at least one false positive in either test = �= = 1 − 1 − � � independent tests Probability of having at least one false positive in any test F �F = 1 − 1 − � 7 Family-wise error rate (FWER) Probability of having at least one false positive among � tests; � = 0.05 F �F = 1 − 1 − � Jelly beans test, � = 20, � = 0.05 FWER =( �=( = 1 − 1 − 0.05 = 0.64 � 8 Bonferroni limit – to control FWER Probability of having at least one false positive Controlling FWER among � tests; � = 0.05 We want to make sure that F �F = 1 − 1 − � ���� ≤ �J. Then, the FWER is controlled at level �′. Bonferroni limit FWER � �J = � � F � = 1 − 1 − ≈ � F � 9 Test
    [Show full text]
  • Hypothesis Testing
    Hypothesis Testing Wolfgang Huber, EMBL Karl Popper (1902-1994) Logical asymmetry between verification and falsifiability. ! No number of positive outcomes at the level of experimental testing can confirm a scientific theory, but a single counterexample is logically decisive: it shows the theory is false. The four steps of hypothesis testing Step 1: Set up a model of reality: null hypothesis, H0 Step 2: Do an experiment, collect data Step 3: Compute the probability of the data in this model Step 4: Make a decision: reject model if the computed probability is deemed to small H0: a model of reality that lets us make specific predictions of how the data should look like. The model is stated using the mathematical theory of probability. Examples of null hypotheses: • The coin is fair • The new drug is no better or worse than a placebo • The observed CellTitreGlo signal for my RNAi-treated cells is no different from that of the negative controls Example Toss a coin a certain number of times ⇒ If the coin is fair, then heads should appear half of the time (roughly). ! But what is “roughly”? We use combinatorics / probability theory to quantify this. ! For example, in 12 tosses with success rate p, the probability of seeing exactly 8 heads is Binomial Distribution H0 here: p = 0.5. Distribution of number of heads: 0.20 0.15 p 0.10 0.05 0.00 0 1 2 3 4 5 6 7 8 9 10 11 12 P(Heads ≤ 2) = 0.0193 n P(Heads ≥ 10) = 0.0193 Significance Level If H0 is true and the coin is fair (p=0.5), it is improbable to observe extreme events such as more than 9 heads 0.0193 = P(Heads ≥ 10 | H0 ) = “p-value” If we observe 10 heads in a trial, the null hypothesis is likely to be false.
    [Show full text]
  • 8. Multiple Test Corrections
    P-values and statistical tests 8. Multiple test corrections Marek Gierliński Division of Computational Biology Hand-outs available at http://is.gd/statlec 2 Gene expression experiment Differential expression: compare gene expression in two conditions (e.g. by t-test) gene_id p.value 1 GENE00001 0.040503700 1 in 20 chance of a false positive 2 GENE00002 0.086340732 3 GENE00003 0.552768467 4 GENE00004 0.379025917 5 GENE00005 0.990102618 6 GENE00006 0.182729903 7 GENE00007 0.923285031 8 GENE00008 0.938615285 9 GENE00009 0.431912336 10 GENE00010 0.822697032 10,000 genes 11 GENE00011 0.004911421 12 GENE00012 0.873463918 13 GENE00013 0.481156679 ~500 genes are “significant” by chance 14 GENE00014 0.442047456 15 GENE00015 0.794117108 16 GENE00016 0.214535451 17 GENE00017 0.231943488 18 GENE00018 0.980911106 19 GENE00019 0.422162464 20 GENE00020 0.915841637 ... 3 Even unlikely result will eventually happen if you repeat your test many times 4 Lets perform a test ! times False positives True positives Reality No Effect Total effect Number of Significant "# $# % discoveries Not $& "& ! − % significant Test result Total !( !) ! Number of tests True negatives False negatives 5 Family-wise error rate !"#$ = Pr(!) ≥ 1) 6 Probability of winning at least once Play lottery Probability of winning is ! Events Probability ! ! " 1 − ! Events Probability !! !×! !" !×(1 − !) "! (1 − !)×! "" 1 − ! ' Probability of not winning at all Probability of winning at least once 1 − 1 − ! ' 7 False positive probability H0: no effect Set ! = 0.05 One test Probability of having a false positive &' = ! Two independent tests Probability of having at least one false positive in either test ( &( = 1 − 1 − ! + independent tests Probability of having at least one false positive in any test , &, = 1 − 1 − ! 8 Family-wise error rate (FWER) Probability of having at least one false positive among ! tests; " = 0.05 ( '( = 1 − 1 − " Jelly beans test, ! = 20, " = 0.05 -.
    [Show full text]
  • Hypothesis Testing (1/22/13)
    STA613/CBB540: Statistical methods in computational biology Hypothesis Testing (1/22/13) Lecturer: Barbara Engelhardt Scribe: Lisa Cervia 1. Hypothesis Testing 1.1. Classical. In classical hypothesis testing we are concerned with evaluating the probability of the data given that a hypothesis is true: P (DjH = t) In Bayesian hypothesis testing, we instead consider the probability that the hypothesis is true given a set of data: P (H = tjD) Throughout both paradigms, we compare the null hypothesis, or one specific model of the data, with an alternative hypothesis, a second possible model of the data. We are interested in detecting data points for which the alternative model is a significantly better at explaining the data than the null model: Null hypothesis: H0 Alternative hypothesis: HA So, as an example, consider the hypothesis that, for two nucleotide sequences of length n, that they are somehow are homologous at a higher rate than expected by chance (assuming the probability 1 of a nucleotide match in the null hypothesis is 4 .) 1 Null, H : X ∼ Binomial(n; p = ) 0 4 1 One sided test, H : X ∼ Binomial(n; p < ) A 4 1 Two sided test, H : X ∼ Binomial(n; p 6= ). A 4 1.2. Test Statistic. Let t(x) be the threshold of a test statistic evaluated on m data points, or features, where x = fx1; :::; xmg, and let m be the number of tests that are performed. For those values that are more extreme than the threshold t(x), we are going to call those feature significant, and for them we will say that we reject the null hypothesis in favor of the alternative hypothesis.
    [Show full text]
  • How Should We Use Hypothesis Tests to Guide the Next Experiment?
    Many tests, many strata: How should we use hypothesis tests to guide the next experiment? Jake Bowers Nuole Chen University of Illinois @ Urbana-Champaign University of Illinois @ Urbana-Champaign Departments of Political Science & Department of Political Science Statistics Associate Fellow, Office of Evaluation Fellow, Office of Evaluation Sciences, GSA Sciences, GSA Methods Director, EGAP https://publish.illinois.edu/nchen3 Fellow, CASBS http://jakebowers.org 25 January 2019 The Plan Motivation and Application: An experiment with many blocks raises questions about where to focus the next experiment Background: Hypothesis testing and error for many tests Adjustment based methods for controlling error rate. Limitations and Advantages of Different Approaches 1/23 Motivation and Application: An experiment with many blocks raises questions about where to focus the next experiment A (Fake) Field Experiment ”...We successfully randomized the college application communication to people within each of the 100 HUD buildings. The estimated effect of the new policy is 5 (p < .001).” ● 120 ● ● ● ● ● 100 ● ● ● ● ● ● ● ● ● ● ● ● ● ● 80 ● ● ● ● ● ● 60 40 20 0 2/23 0 1 people in just a few buildings? Policy Expert: ...Great! But this seems small. Could it be mainly effecting A (Fake) Field Experiment 60 50 80 50 40 40 60 30 30 40 20 20 20 10 10 0 1 0 1 0 1 ● ● 70 70 ● 80 ● 60 60 50 50 60 40 40 40 30 30 20 20 20 10 10 0 0 0 1 0 1 0 1 ● ● 80 100 80 ● 80 60 60 ● 60 40 40 40 20 20 20 0 0 1 0 1 0 1 3/23 Jake: This is a natural and important question both for the development of new theory and better policy.
    [Show full text]
  • Multiple Comparisons: Problem and Solutions
    Multiple comparisons: problem and solutions Gareth Barnes Wellcome Centre for Human Neuroimaging University College London SPM M/EEG Course By the end of this talk Understand what the multiple comparisons problem is. Be familiar with some common approaches. Be able to explain Random field theory. Contrast c General Pre- Statistical Linear processing Inference Model Is this real ? Need to avoid cherry-picking. i.e. need an objective threshold. Need to adjust this threshold depending on how many independent tests you do. (e.g. if you do 20 tests with a false positive rate of 1/20 .. then expect one peak just due to chance). T test on a single voxel / time point T value Test only this 1.66 Task A T distribution Task B P=0.05 u Threshold Multiple tests in space u u u u u If we have 100,000 voxels, u α=0.05 5,000 false positive voxels. This is clearly undesirable; to correct t for this we can define a null hypothesis t t for a collection of tests. t t t signal Use of ‘uncorrected’ p-value, α =0.1 11.3% 11.3% 12.5% 10.8% 11.5% 10.0% 10.7% 11.2% 10.2% 9.5% Percentage of Null Pixels that are False Positives Bonferroni correction Set the test-wise error rate (α) to be a the ideal Family-Wise Error rate (FWER) (αFWE) divided by the number of tests. e.g. for five tests choose p<0.01 for each test to control family at p<0.05 This correction does not require the tests to be independent but becomes very stringent if they are not.
    [Show full text]