Home

GLIMMPSE User Manual

Contents

1. To include clustering in the study click Add clustering and follow the prompts Use the Remove clustering button to remove clustering information Add clustering lt q D A Help 4 Save Design Cancel Figure 17 Clustering Screen If the study design does not include clustering simply click PP to proceed To add clustering click the Add clustering button Three text boxes will appear at the bottom of the screen Clustering In a clustered design the independent sampling unit is a cluster such as a community school or classroom Observations within a cluster are correlated The labels for observations within a cluster must be exchangeable For example child id within classroom can be reassigned arbitrarily In contrast observations across time cannot be reassigned and should not be considered clustered observations Clustering or repeated measures or a combination creates a multilevel design The common correlation between any pair of cluster members is termed the intraclass correlation or intracluster correlation To include clustering in the study click Add clustering and follow the prompts Use the Remove clustering button to remove clustering information Remove clustering Cluster label Number of observations or sub clusters within each cluster of this type Intra cluster correlation Add subgroup Remove subgrou X Cancel Save Design Figure 18 Clustering Options Enter the Cluster name
2. s likely to be more useful 9 Power Total Sample Size lt q bb Help H Save Design Cancel Figure 41 Solving For Screen When Power is selected the inputs will be used for a power analysis The power analysis will produce a value s between O and 1 representing the probability the study will answer the question of interest When Total Sample Size is selected the inputs will be used to calculate the number of individual sampling units also called participants if referring specifically to people needed for the study to achieve the desired power If the number of participants is not set we recommend solving for sample size in order to obtain the appropriate sample size for achieving the goals of your study However if sample size is set due to budgetary or other restrictions a power calculation will indicate the probability that the study will provide a definitive answer to the question of interest On the screen select Power or Total Sample Size by selecting the appropriate radio button Upon completing the selection click PP to proceed 4 1 3 Desired Power if solving for Total Sample Size When solving for sample size the user must enter the desired power for the study Enter the target values as decimals i e 0 95 im the Power Values box and click to add the value to the list Colorado School of UF Heath Outcomes amp Policy College of Medicine PUBLIC HEALTH Page 38 of 58 Y GLIMMPSE User Manual
3. GLIMMPSE has used these values to calculate a covariance matrix which describes the overall variability Changes in variability can dramatically affect power and sample size results It is not possible to know the variability until the experiment is observed To account for this uncertainty it is common to calculate power or sample size for alternative values for variability S By clicking the box below GLIMMPSE will calculate power using the calculated Variability covariance matrix the covariance matrix divided by 2 and the covariance matrix Cena ae ape multiplied by 2 Within Participant PR Yes include power for the covariance matrix the covariance matrix divided by 2 and the covariance matrix multiplied by Sigma Scale Factors 2 q bb 4 Help Save Design X Cancel Figure 33 Flexible Variability When finished click PP to proceed 3 8 Options This screen provides an introduction to the Options section After reading the screen click VP to proceed Options In this selection you may optionally request confidence intervals for power and power curve images For designs with multiple outcomes you may select one or more statistical tests For designs with a baseline covariate you may select from different power methods Options Z Statistical Test Confidence Intervals Y Power Curve Figure 34 Options 3 8 1 Statistical Tests The Statistical Tests screen allows the user to select on
4. Introduction The GLIMMPSE wizard will guide you through several steps to calculate power or sample size Use the forward and back arrows to navigate through the wizard You may save your work at any time by clicking the Save Design link at the lower right of the screen The Cancel link also at the lower right of the screen allows you to cancel your current work and begin a new study design The help manual may be accessed by clicking the Help link General steps for a power analysis are listed on the left hand side of the screen We will ask you to specify e The Typelerror rate e The independent and dependent variables e The primary study hypothesis of interest e Choices for group means e Choices for standard deviations and correlations for study outcomes e The statistical test and additional display options Click the forward arrow to begin pp Help Y Save Design Figure 10 Introduction Screen Health Outcomes amp Policy College of Medicine UF X Cancel Page 13 of 58 GLIMMPSE User Manual Version 2 0 0 Caletlata Would you like to solve for power or sample size l i To begin your calculation please indicate whether you would like to solve for power Start or total sample size Y Solving For If you have a rough idea of the number of research participants you will be able to 4 Desired Power recruit then solving for power may be more beneficial 4 Type Error If you have fewer restr
5. a e The design essence matrix Es X e The between C and within subject U contrast matrices e The null hypothesis matrix Op e Choices for the regression coefficients matrix B e Choices for error covariance Ze and or predictor covariance matrices g Zy Zyg e The statistical test you plan to use and additional display options In GLIMMPSE terminology referring to subjects is used for consistency with statistical tradition In the context of research on human participants subjects should be interpreted as research participants Click the forward arrow to begin lt p 4 Help H Save Design Cancel Figure 40 Introduction Screen for Matriz Mode Click PP to begin entering the details of the analysis 4 1 2 Solving For The Solving For screen allows the user to select either a power or sample size calculation Colorado School of PUBLIC HEALTH Health Outcomes amp Policy UF oven Page 37 of 58 GLIMMPSE User Manual Version 2 0 0 Calculate Would you like to solve for power or sample size To begin your calculation please indicate whether you would like to solve for power Start or total sample size y Solving For lf you have a rough idea of the number of research participants you will be able to g Type 1 Ema recruit then solving for power may be more beneficial If you have fewer restrictions on recruitment and would like to ensure a well powered study then solving for sample size
6. http samplesizeshop org documentation glimmpse glimmpse validation results 2 3 Technical documentation for the software is available at http samplesizeshop org documentation glimmpse 4 GLIMMPSE software modules are available for downloaded from http samplesizeshop org software downloads glimmpse software downloads References Colorado School of UF Heath Outcomes amp Policy PUBLIC HEALTH Page 57 of 58 GLIMMPSE User Manual Version 2 0 0 Apple 2010 Safari Version 5 0 3 Cupertino CA URL http www apple com safari Glueck DH Muller KE 2003 Adjusting Power for a Baseline Covariate in Linear Models Statistics in Medicine 22 2535 2551 Google 2011 Google Chrome Web Browser Version 23 0 1271 95 Mountain View CA URL https www google com intl en chrome browser McGovern J Tyagi S Stevens M Mathew S 2003 Java Web Services Architecture Morgan Kaufmann San Francisco CA Microsoft 2010 Internet Explorer Version 8 Redmond WA URL http www microsoft com windows internet explorer worldwide sites aspx Mozilla 2011 Firefox Web Browser Version 8 6 12 Mountain View CA URL http www mozilla com en US firefox Muller KE Barton CN 1989 Approximate Power for Repeated Measures ANOVA Lacking Sphericity Journal of the American Statistical Association 84 406 549 555 Muller KE Edwards LJ Simpson SL Taylor DJ 2007 Statistical Tests with Accurate Size and Powe
7. lt u 45444 wee es DOr OPUNE sans be eR ew Boe oe Se ES 3 8 1 Statistical Tests 3 8 2 Power Calculation Method 3 8 3 Confidence Intervals 3 8 4 Power Curve Options do Callao esa rra da ES Me G E 4 Matriz Mode Screen by Screen Tour Al e amp 44 Gs oo oe ee A EDER RSS os 4 1 1 Introduction oaoa a 4 1 2 Solving For aaa 4 1 3 Desired Power if solving for Total Sample Size ALA Type LETO se cae sa a Ed ww Alda DESIN nessa rsdd ris RA 4 2 1 Design Essence aooaa a 4 2 2 Covariate a 4 2 3 Smallest Group Size 4 3 Coefficients a a a 4 3 1 Beta Coefficients B Matrix 4 3 2 Beta Scale Factors A Hypothes vo oe ko acaso sae AAA 4 4 1 Between Participant Contrast 4 4 2 Within Participant Contrast 4 4 3 Null Hypothesis A5 VariabDilty sessar oo oe eRe OE EHR ARE 4 5 1 Error Covariance 048 4 5 2 Outcomes Covariance 4 4 5 3 Variance of Covariate Colorado School of UF Heath Outcomes amp Policy College f Medici PUBLIC HEALTH of Medicine Version 2 0 0 Page 2 of 58 GLIMMPSE User Manual Version 2 0 0 4 5 4 Covariance of Outcomes and Covariate oaoa a o a a a a e a a al 45 0 Sigma Scale Factors s s s e es escasear ara ARA 51 TNI daras as
8. 0 0 GLIMMPSE 2 0 0 is an open source online tool for calculating power and sample size GLIMMPSE has been designed so that researchers and scientists with a varying levels of statistical training can have access to reliable power and sample size calculations For optimum usability GLIMMPSE provides two different modes In Guided Mode users receive step by step guided instructions for entering data in order to obtain power and sample size outputs In Matriz Mode users receive less guidance and are assumed to possess in depth statistical training GLIMMPSE can compute power or sample size for univariate and multivariate linear models with Gaussian errors Muller and Stewart 2006 GLIMMPSE supports two main types of study design models designs with only fixed predictors and designs with fixed predictors and a single Gaussian covariate The values of a fixed predictor are set as part of the study design and are known without appreciable error In contrast Gaussian covariates are not observed until data is collected Common designs with only fixed predictors include t tests analysis of variance ANOVA and multivariate analysis of variance MANOVA Common designs that control for a covariate include analysis of covariance ANCOVA and multivariate analysis of covariance MANCOVA Details about power calculations for the general linear multivariate model with Gaussian data and fixed predictors can be found in Muller and Peterson 1984 Mulle
9. 2 5 1 Resizing and Entering Values Into a Matrix In Matrix Mode GLIMMPSE requires the user to define the matrices for the power calculation Figure 9 shows an example of a matrix template in Matriz Mode Sometimes the matrix dimensions are pre determined If not the user can set the matrix dimensions by typing the number of rows into the row text box see 1 in Figure 9 and the number of columns into the column text box see 2 in Figure 9 Fill in the elements of the matrix by entering values into the text boxes within the matrix template see 3 in Figure 9 Categorical Predictors 1 gt 3 3 A 2 Figure 9 Example of entering values into a matrix 3 Using Guided Mode A Screen by Screen Tour In Guided Mode users receive step by step guided instructions when entering inputs for calculating power and Colorado School of UF Heath Outcomes amp Policy College of Medicine PUBLIC HEALTH Page 12 of 58 gt GLIMMPSE User Manual sample size for use in study design 3 1 Start 3 1 1 Introduction Version 2 0 0 The Introduction screen contains a summary of the steps involved in the power or sample size analysis After reading the screen click PP to begin entering the details of the study design Start amp Solving For amp Type Error 3 1 2 Solving For The Solving For screen allows the user to select either a power or sample size calculation Colorado School of PUBLIC HEALTH
10. Select the time location etc from the list s below This will allow you to edit the means at the selected time location etc month 3 ws Ww Help Save Design X Cancel Figure 29 Means Screen For designs with repeated measures the user may enter means at each time place etc The drop down lists below the table of means allow the user to select a specific time place etc in order to edit the means When finished click PP to proceed Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 27 of 58 GLIMMPSE User Manual Version 2 0 0 3 6 3 Flexible Means The Flexible Means screen allows the user to compute power or sample size for the means as specified the mean values divided by 2 and the mean values multiplied by 2 Leave the checkbox blank to compute power or sample size only for the means as entered on the Means screen Flexible Means Power and sample size results will change depending on the mean values specified on the previous screen It is not possible to know exact values for the means until the experiment is observed To account for the uncertainty it is common to calculate power for the mean values as specified the mean values divided by 2 and the mean values multiplied by 2 Yes include power calculations for the mean values as entered the mean values divided by 2 and the mean values multiplied by 2 Means Y Means Scale Factors for Means lt q bp
11. Version 2 0 0 Power Values Enter the desired power values in the list box below Power values are numbers Start between 0 and 1 Higher values correspond to a greater likelihood of rejecting the null hypothesis Common values are 0 8 or 0 9 although 0 9 or higher is usually Solving For preferred Desired Power l l l l a Y Type Error Type each value into the list box and click Add To remove an item highlight the cha value and click the Delete button Desi In oars Power Values TE 0 95 ur TF OCT C e V UL Ll E gt a gt lt q D A Help Save Design X Cancel Figure 42 Desired Power Screen When finished click DP to proceed 4 1 4 Type I Error Enter the target values for Type I Error as decimals i e 0 05 in the Type I Error Values box The user may specify up to five Type I Error values Type Error A Type error occurs when a scientist declares a difference when none is actually Start present The Type error rate is the probability of a Type error occurring and is often referred to as a Type error rates range from 0 to 1 The most commonly used Solving For values are 0 01 0 05 and 0 1 Y Desired Power Enter each Type error value into the text box and click Add You may enter up to 5 C Cm values To remove a value select the value in the list box and click the Delete Design button FE Type I Eror Vales Hypothesis 0 05 e Varia
12. baseline covariate you may select from different power methods Options Statistical Test Y Confidence Intervals Y Power Curve SN Ww Help Y Save Design X Cancel Figure 58 Options Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy cae a Page 52 of 58 GLIMMPSE User Manual Version 2 0 0 4 6 1 Statistical Tests The Statistical Tests screen allows the user to select one or more statistical tests for the power or sample size calculations A tutorial providing guidelines for selecting a test is available from the GLIMMPSE Tutorials page at http samplesizeshop org education tutorials Full theoretical details are available in Muller and Stewart 2006 Select the statistical test s by clicking one or more check boxes For designs with a Gaussian covariate only the Hotelling Lawley trace and the Univariate Approach to Repeated Measures are valid Statistical Tests Select the statistical tests to include in your calculations For study designs with a single outcome power is the same regardless of the test selected Note that only the Hotelling Lawley Trace and the Univariate Approach to Repeated Measures are supported for designs which include a baseline covariate Hotelling Lawley Trace Pillai Bartlett Trace Wilks Likelihood Ratio Univariate Approach to Repeated Measures with Box Correction Options Univariate Approach to Repeated Measures with Geisser Greenhouse Correction yY Stati
13. box After each entry click Lada press on the keyboard or click anywhere on the screen To delete a value select the unwanted value and click to remove the value from the list Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 51 of 58 gt GLIMMPSE User Manual Version 2 0 0 Scale Factors for Covariance Changes in variability can dramatically affect power Scale factors allow you to y compute power for alternative covariance values by scaling Ze For example entering the scale factors 0 5 1 and 2 would compute power for Ze divided by 2 Ze matrix as Desian entered and Ze matrix multiplied by 2 Soefficients Enter each scale factor in the text box below and click Add To use Ze as entered on the previous screen enter a 1 in the list below Fe Matrix Scale Factors Variability Outcomes Covariance Y Variance of Covariate Covariance of Outcomes a and Covariate Sigma Scale Factors _ pt q bb Help Y Save Design X Cancel Figure 57 Covariance Scale Factors When finished click PP to proceed 4 6 Options This screen provides an introduction to the Options section After reading the screen click PP to proceed Caleulita Options In this selection you may optionally request confidence intervals for power and power SR curve images Desian For designs with multiple outcomes you may select one or more statistical tests Coefficients For designs with a
14. you cancel your current work and begin a new study design The help manual may be accessed by clicking the Help link General steps for a power analysis are listed on the left hand side of the screen We will ask you to specify e The Type l error rate e The independent and dependent variables e The primary study hypothesis of interest e Choices for group means e Choices for standard deviations and correlations for study outcomes e The statistical test and additional display options Click next to begin lt q bb Help a Save Design X Cancel Figure 2 Example Start Screen Following the Introduction screen Figure 2 GLIMMPSE will prompt the user to enter the details for the power or sample size calculation The user may enter the details in any order However some screens cannot be accessed unless the user has completed information from previous screens For example if the user has not entered information in the Responses tab then the user will not be able to enter information in the Hypothesis tab If a tab is inaccessible due to missing information in an earlier screen it will be indicated by a red circle with a slash through it Czileulats Hypothesis gt Hypothesis Figure 3 More information is required to access the Hypothesis screen Notice that in Figure 2 above there are pencil icons beside the two available screens in the active Start tab Colorado School of PUBLIC HEALTH UF Health Out
15. 13 14 1 15 110 Reset to Equal Spacing Units position Type Ordinal Number of Measurements 3 Add Level Remove Level Save Design X Cancel Figure 25 Repeated Measures Add Level When finished click DP to proceed 3 5 Hypotheses 3 5 1 Introduction This screen provides an introduction to the Hypothesis section After reading the information on the screen click PP to proceed Gallet Hypothesis amp Hypothesis 3 5 2 Hypotheses Study Hypotheses Introduction Power and sample size calculations are typically based on the primary study hypothesis GLIMMPSE will generate several possible hypotheses based on your study design Click the forward arrow to continue Figure 26 Hypothesis Introduction Screen The Hypotheses screen allows the user to select the primary hypothesis of interest The user first selects the type Colorado School of PUBLIC HEALTH Health Outcomes amp Policy UF loys Page 25 of 58 GLIMMPSE User Manual Version 2 0 0 of hypothesis by clicking the appropriate radio button Additional information will be requested depending on the type of hypothesis Caleulaito Hypotheses The options below show the hypotheses which are available for the current study design Select the hypothesis which most closely resembles your primary study hypothesis This hypothesis will be used to determine power for your study To specify a hypothesis select the radio button match
16. Health Outcomes amp Policy College of Medicine Colorado School of UF PUBLIC HEALTH GLIMMPSE User Manual Y Version 2 0 0 Authors Zacc Coker Dukowitz Brandy Ringham and Sarah Kreidler September 2012 Copyright C 2012 Regents of the University of Colorado Denver GLIMMPSE is released under the GNU Public License version 2 0 GLIMMPSE Version 2 0 0 is funded by NIDCR 1 R01 DE020832 01A1 to the University of Florida Keith E Muller PI Deborah Glueck University of Colorado site PI Previous funding was received from an American Recovery and Re investment Act supplement 3K07CA088811 065 for NCI grant K07CA088811 GLIMMPSE User Manual Version 2 0 0 Contents 1 Introduction 4 1 1 Version Information and Licensing 4 1 2 Welcome to GLIMMPSE 2 0 0 a E A 4 os Mb GCLIMUPSE sacro praias Peas REPARAR 4 2 Using GLIMMPSE 5 2 1 When to Use GLIMMPSE e a 5 22 low o Uee CLINIMPSE aer be owe ee AAA a aa A 5 2 2 1 Initiating the GLIMMPSE Wizard 5 2 2 2 Choosing Between Guided Mode and Matrix Mode 0 6 2 3 Basic Navigation for GLIMMPSE in Both Guided Mode and Matriz Mode 6 Zoek Ly pue INO K DOs ipod ARRE 8 Zo Wena Drop Down Lists s s se se ee oe OE eA A ee ee SH oe Ae E a 9 2 3 3 Radio Buttons and Check Boxes e 9 2 3 4 Results Report as 4 ck wR eG wee eras AAA 10 2 4 Basic Navigation for Guided Mode 11 2 4 1 Enter
17. LIC HEALTH Health Outcomes amp Policy College of Medicine UF Page 56 of 58 GLIMMPSE User Manual Version 2 0 0 Calculate Power Curve 1 04 0 94 0 84 0 74 _ 0 6 3 o 0 54 a 0 44 Options 0 3 Y Statistical Test 0 2 Y Confidence Intervals sind 0 0 r r Y Power Curve 75 100 125 150 Total Sample Size Series Power Results Power Total Sample Size Test Type Error Rate Means Scale Factor Variability Scale Factor 0 608 60 HLT 0 05 1 1 0 801 90 HLT 0 05 1 1 0 908 120 HLT 0 05 1 1 0 960 150 HLT 0 05 1 1 Save to CSV View Matrices q Ww Help 4 Save Design X Cancel Figure 63 Results 5 Additional GLIMMPSE Resources Additional resources for GLIMMPSE are available at http samplesizeshop org The Sample Size Shop project is a collaborative effort between the University of Florida and the University of Colorado Denver The goals of the project are to develop new statistical methods for calculating power and sample size provide user friendly software to perform the power and sample size calculations and educate researchers regarding both the methods and the software The following online resources are available for the GLIMMPSE software 1 Tutorials demonstrating power and sample size calculations with GLIMMPSE for a variety of study designs are available at http samplesizeshop org education tutorials 2 Validation reports showing the accuracy of GLIMMPSE calculations are available at
18. Y GLIMMPSE User Manual Caleta Sampling Unit Y Study Groups Covariate Y Clustering Y Relative Group Sizes Controlling for a single normally distributed predictor A common experimental design is an analysis of covariance which includes one or more fixed predictors and one or more continous control variables the covariates For example one might run an experiment with 10 males and 10 females with an indicator variable for gender as a fixed predictor and age as a covariate Acommon special case uses a series of repeated measurements on a continuous outcome The first measurement observed prior to treatment is used as a baseline covariate The other repeated measurements are outcomes in the general linear multivariate model GLIMMPSE can calculate power for hypotheses concerning the fixed predictors optionally controlling for a single normally distributed predictor If you plan to include a single normally distributed predictor in your model click the check box below At present the GLIMMPSE software does not calculate power for multiple normally distributed predictors nor non normally distributed predictors Control for a single normally distributed predictor lt q bp 4 Help H Save Design Figure 16 Gaussian Predictor Screen When finished click PP to proceed 3 2 4 Clustering X Cancel Version 2 0 0 Clustering is present when research participants are organized into groups Often random
19. a trend in a given factor click the Edit Trend link and select an appropriate trend Between Participant Factors treatment Edittrend None FE Within Participant Factors B Choose multiple options grade Editirend All polynomial trends Figure 7 Example of radio buttons and check boxes 2 3 4 Results Report Power results are displayed in a table with each row representing an individual power calculation If multiple factors have been specified in the study design for example multiple Type I error rates variability scale factors etc then the results table will have multiple rows See Table 1 below for an example of the information displayed for a given results report Every results report for power contains both calculated and desired power values When solving for power these two values are the same When solving for sample size it may not be possible to achieve the exact power value specified by the user In this case nominal power is the default power value the power value specified by the user and actual power is the calculated power for the sample size that best matches the desired power A power curve may also be requested with power on the vertical or Y axis and either the regression coefficient scale factor covariance scale factor or total sample size on the horizontal or X axis Power results can be saved to a comma delimited file so that users can import the data into other statistical packages T
20. ated measures are added via the Add Level button For example consider a design in which a partic ipant s blood pressure is measured every month for six months and at each visit in three different positions for example standing sitting and supine The design would include doubly repeated measures with one level for month and a second nested level for position The user may add up to three levels of repeated measures To add a sub level click the Add level button Three more text boxes will appear Colorado School of PUBLIC HEALTH UF Health Outcomes amp Policy P a g e 4 of 58 Y GLIMMPSE User Manual Casletato Responses Y Response Variables Y Repeated Measures bp Help Version 2 0 0 Repeated Measures Repeated measures are present when a response variable is measured on each research participant on two or more occasions or under two or more conditions If the study includes repeated measurements click Add repeated measures and follow the prompts The text entered in the Units text box indicates the dimension over which measures were taken ex time days locations etc The choice of Type indicates whether the repeated measures are numeric ex time ordinal ex 1st 2nd 3rd or categorical ex arm leg hand You may specify up to 3 levels of repeated measures Remove Repeated Measures Units month Type Numeric Number of Measurements 6 Spacing 112
21. ath Outcomes amp Policy College of Medicine PUBLIC HEALTH Page 14 of 58 Y GLIMMPSE User Manual Caleta Start Solving For 2 Desired Power amp Type Error sampling Unit Power Values Enter the desired power values in the list box below Power values are numbers between 0 and 1 Higher values correspond to a greater likelihood of rejecting the null hypothesis Common values are 0 8 or 0 9 although 0 9 or higher is usually preferred Type each value into the list box and click Add To remove an item highlight the value and click the Delete button Power Values 095 g q D Help E Save Design Figure 12 Desired Power Screen When finished click DP to proceed 3 1 4 Type I Error Version 2 0 0 X Cancel Enter the target values for Type I Error as decimals i e 0 05 in the Type I Error Values box The user may specify up to five Type L Error values Caietato Start Solving For Y Desired Power Type Error sampling Unit Responses Colorado School of PUBLIC HEALTH Type Error A Type error occurs when a scientist declares a difference when none is actually present The Type error rate is the probability of a Type error occurring and is often referred to as a Type error rates range from 0 to 1 The most commonly used values are 0 01 0 05 and 0 1 Enter each Type error value into the text box and click Add You may enter up to 5
22. bility di bb 4 Help Y Save Design X Cancel Figure 43 Type I Error Colorado School of UF Health Outcomes amp Policy PUBLIC HEALTH Page 39 of 58 GLIMMPSE User Manual Version 2 0 0 When finished click PP to proceed 4 2 Design 4 2 1 Design Essence In the Design section the user will define the composition of the study by specifying the number of groups how subjects are divided into groups the size of each group and whether the design will include a Gaussian covariate The Design Essence Matrix In the general linear multivariate model with fixed predictors Y XB E the X matrix represents the study design The same is true for F in the general linear multivariate model with fixed predictors and a Gaussian predictor Glueck and Muller 2003 For simplicity we will only discuss X since the instructions do not change for F In data analysis the X matrix would contain a single row for each subject Since power analysis does not include actual data the design essence matrix Muller and Stewart 2006 is a version of the X matrix that contains a single row for each unique combination of predictors in the study design Note that the essence matrix specifies only the fixed or categorical predictors in the study design For example consider a 2 factor ANOVA design with 2 levels per factor 3 subjects per group and a cell means coding In data analysis the design matrix and corresponding essence matrix
23. comes amp Policy P a g e T of 5 8 GLIMMPSE User Manual Version 2 0 0 The pencil icons indicate screens which require additional information Once the user have entered the required information the pencil will turn into a green check mark Caleuleata Start ri 2 Figure 4 Indication that the Solving for screen is complete Some screens are optional and will already have a check mark beside them 2 3 1 Typing Into a Text Box Several screens in GLIMMPSE will ask you to specify information by typing into a text box To input information in a text box click in the text box and type the requested information To complete the entry you may 1 Click anywhere on the screen 2 Press on your keyboard or 3 Click for text boxes associated with a list of values To delete entries in a list associated with the text box click on the entry so that it becomes highlighted in blue Click to delete the highlighted entry Figure 5 shows three examples of text box entries with the text boxes highlighted in yellow Colorado School of Health Outcomes amp Polic PUBLIC HEALTH UE ci sr i Page 8 of 58 GLIMMPSE User Manual Version 2 0 0 A Were the outcomes measured multiple times on each subject O No measurements were only taken one time Yes measurements were repeated over a single dimension ex days weeks locations etc How many times Mo Over wnat dimension Relative Categorical Predictors Gr
24. ctors allow you to consider alternative values for the regression coefficients by scaling the B matrix For example entering the scale factors 0 5 1 and 2 would compute power for the B Coefficients matrix divided by 2 the B as entered and the B matrix multiplied by 2 Y Beta Coefficients Enter each scale factor in the text box below and click Add To use the B matrix as y aa sia nng specified on the previous screen enter a 1 in the list below B Matrix Scale Factors Add Deleie 0 5 1 2 4 bb Help E Save Design Cancel Figure 48 Beta Matrix When finished entering your values click PP to proceed 4 4 Hypothesis In this section the user defines the contrast matrices in the study The contrast matrices C and U consist of the hypotheses to be tested They are used to calculate the expected hypothesis matrix O CBU 4 4 1 Between Participant Contrast The C matrix consists of the between participant contrasts The between participant contrasts test hypotheses between independent sampling units The number of rows in the C matrix represent the degrees of freedom for the hypothesis test For example suppose an investigator wants to compare the average final exam test scores of students in class A and class B The contrast matrix would be C 1 1 When multiplied by B this becomes the difference in the proposed average test scores between class A and class B Enter the number of rows number
25. dies or clinical experience For example suppose an investigator wants to compare foal birth weight between dams who are given feed formula A feed formula B and standard feed In order to be cost effective the new feed formulas must improve foal birth weight by more than 7 kg The null hypothesis then is Og appearing as shown below Null Hypotheses Oo Matrix For O CBU the general linear hypothesis is stated as Ho O Oo In most cases the Og matrix will contain zeros The number of rows in Og must equal the number of rows in C and the number of Hypothesis yp columns must match the number of columns in U To ensure conforming matrices Between Participant the dimensions of Oy cannot be adjusted on this screen Contrast Within Participant Enter your Og matrix below Contrast 2 1 _ Null Hypothesis Matrix 7 7 44 bb Help Y Save Design X Cancel Figure 52 Non zero Null Hypothesis Matrix Colorado School of Health Outcomes amp Policy PUBLIC HEALTH UF ci Page 48 of 58 GLIMMPSE User Manual Version 2 0 0 O has the same number of rows as C and the same number of columns as U Therefore its size cannot be adjusted on this screen The user need only enter the matrix cell values When finished specifying the O matrix click PP to proceed 4 5 Variability Variability describes how much measurements differ from each other In this section the user defines the covariance of errors and covarianc
26. e 6 because blood pressure was measured every 6 months For numeric repeated measures GLIMMPSE 2 0 auto populates equidistant measurements To change the distance between measures type into the text boxes For example if blood pressure was measured every month for the first three months then every other month for the next six months the user would type 1 2 3 5 7 9 into the text boxes Colorado School of Health Outcomes amp Polic PUBLIC HEALTH UF io i Page 23 of 58 GLIMMPSE User Manual Casietiato Responses Y Response Variables yY Repeated Measures Version 2 0 0 Repeated Measures Repeated measures are present when a response variable is measured on each research participant on two or more occasions or under two or more conditions If the study includes repeated measurements click Add repeated measures and follow the prompts The text entered in the Units text box indicates the dimension over which measures were taken ex time days locations etc The choice of Type indicates whether the repeated measures are numeric ex time ordinal ex 1st 2nd 3rd or categorical ex arm leg hand You may specify up to 3 levels of repeated measures Remove Repeated Measures Units month Type Numeric lt Number of Measurements 6 Spacing 1 2 113 4 15 16 Reset to Equal Spacing Add Level Remove Level vp Help Save Design X Cancel Figure 24 Repeated Measures Nested repe
27. e Commensurate observations share the same measurement scale and units e Multivariate outcomes arise from a single ISU and therefore are not independent and need not be commensurate e Doubly multivariate outcomes include repeated measures of two or more noncommensurate variables Quoted from Muller amp Stewart 2007 definition 6 1 q bb i Help H Save Design X Cancel Figure 14 Sampling Units Introduction Screen After reading the information on the screen click PP to proceed 3 2 2 Study Groups Independent sampling units may be randomized to different treatments or be classified by characteristics such as gender The characteristics divide the sampling units into study groups The Study Groups screen allows the user to define the study groups by specifying fixed predictors Enter the fixed predictors as described in Section 2 4 1 For one sample designs with no fixed predictors leave the table blank Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 16 of 58 gt GLIMMPSE User Manual Calculate Sampling Unit Y y y y Y Study Groups Covariate Clustering Relative Group Sizes Smallest Group Size Predictor Version 2 0 0 Study Groups Describe the predictors which assign independent sampling units into groups such as gender or treatment The choice of study design determines the values of fixed predictors such as drug dose or gender A common example
28. e or more statistical tests for the power or sample size calculations A tutorial providing guidelines for selecting a test is available from the GLIMMPSE Tutorials page Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy cape a Page 31 of 58 GLIMMPSE User Manual Version 2 0 0 at http samplesizeshop org education tutorials Select the statistical test s you wish to use by clicking one or more check boxes Statistical Tests Select the statistical tests to include in your calculations For study designs with a single outcome power is the same regardless of the test selected Note that only the Hotelling Lawley Trace and the Univariate Approach to Repeated Measures are supported for designs which include a baseline covariate Y Hotelling Lawley Trace Pillai Bartlett Trace Wilks Likelihood Ratio Univariate Approach to Repeated Measures with Box Correction Univariate Approach to Repeated Measures with Geisser Greenhouse Correction Options Univariate Approach to Repeated Measures with Huynh Feldt Correction cee A eee Confidence Intervals Y Power Curve q Ww Help b Save Design X Cancel Figure 35 Statistical Tests When finished click PP to proceed 3 8 2 Power Calculation Method For designs with a baseline covariate two different methods are available to calculate power quantile and uncon ditional power For theoretical details please see Glueck and Muller 2003 Se
29. e related to the Gaussian covariate 4 5 1 Error Covariance For each independent sampling unit Xe is the covariance of the random errors conditional on the values of the fixed predictors The Error Covariance screen allows the user to define Xe by directly entering the covariance matrix values To ensure conformance with the B and U matrices the dimensions of the Xe matrix cannot be modifed on this screen Caeiro Covariance of Errors Le Matrix For each independent sampling unit Ze is the covariance of the random errors For univariate designs Ze will be a 1x1 matrix containing the variance of the error term More complex structures may be entered for multivariate or repeated measures designs Values for Ze are typically obtained from pilot data or previous studies Ze is a square symmetric matrix with dimensions equal to the number of columns in B Variability Enter values for Ze in matrix below 4 4 amp Error Covanance 1 02 01 0 05 A lama ale ra s 0 2 1 0 2 0 1 0 1 0 2 1 0 2 0 05 0 1 0 2 1 q Ww Help Save Design X Cancel Figure 53 Error Covariance When finsihed click PP to proceed 4 9 2 Outcomes Covariance For designs with a Gaussian covariate the user must specify the covariance of the outcomes Xy For each independent sampling unit By is the covariance of the outcomes conditional on the fixed predictors One can think of Xy as the error covariance for each independent sampling unit in a model contai
30. eases if the covariate explains some portion of the variance in the outcome The Zyg matrix defines the strength of the association between the covariate and the outcomes For univariate designs Zyg will be a 1x1 matrix containing the covariance between the outcome and the covariate For multivariate and repeated measures designs the matrix will be px1 where p is the number of outcomes and each row contains the covariance between the covariate and the corresponding outcome Variability Enter values for Zyg in the matrix below a Ma Y Outcomes Covariance Y Variance of Covariate 12 Covariance of 12 Y Outcomes and 12 Covariate 12 Y dd py Help H Save Design Cancel Figure 56 Covariance of Outcomes and Covariate When finished click PP to proceed 4 5 5 Sigma Scale Factors GLIMMPSE allows users to specify scale factors for the covariance matrices For the general linear multivariate model with fixed predictors the scale factors are applied to the user specified Xe matrix For the general linear multivariate model with fixed predictors and a Gaussian covariate the scale factors are applied to the Xe matrix which is calculated from My 5 5 and Nyy Since variability can dramatically impact power it is common to calculate power for the proposed value as well as alternative values such as half and twice the proposed value To specify one or more covariance scale factors enter the scale factors in the wig Matrix Scale Factors
31. ectly enter the matrices for the If you have previously saved a study ANOVA ANCOVA and regression general linear model This mode is design from GLIMMPSE you may with guidance from the study design designed for users with advanced upload it here Click browse to select wizard This mode is designed for statistical training your study design file more applied researchers including physicians nurses and other principal investigators Figure 1 GLIMMPSE Start Screen 2 2 2 Choosing Between Guided Mode and Matrix Mode The GLIMMPSE start screen presents three options Guided Mode Matriz Mode and Upload a Study Design In Guided Mode users receive step by step guided instructions when entering inputs for power or sample size calculations To choose Guided Mode click in the Guided Study Design box In Matriz Mode users receive less guidance and are assumed to possess in depth statistical training Matrix Mode allows direct input of all matrices required for a power or sample size calculation To choose Matrix Mode click in the Matrix Study Design box If the user has a study design saved from a previous GLIMMPSE session the user may upload it by clicking Choose File in the Upload a Study Design box GLIMMPSE will open the saved study design and allow the user to continue the power or sample size analysis 2 3 Basic Navigation for GLIMMPSE in Both Guided Mode and Matriz Mode Once a mode of entry has been chosen the steps r
32. ed click PP to proceed 3 8 3 Confidence Intervals Power analysis involves some uncertainty in the choices for means and variability Therefore the Confidence Intervals screen allows the user to request confidence intervals on the power results To include confidence intervals uncheck the checkbox The information on the confidence interval screen describes the data set or publication from which the choices for means and variances were obtained For example if a scientist were calculating power based on the means and variances obtained from pilot data the scientist would enter information describing the pilot data set The following information is required The Assumptions section allows the user to indicate if he or she is uncertain about the variance but reasonably certain of the mean values or uncertain of both the means and variance The Upper and lower tail probabilities define the width of the confidence interval For example a centered 95 confidence interval would have both upper and lower tail probabilities of 0 025 The Total sample size value indicates the number of independent sampling units in the pilot data set or publica tion The Rank of the design matriz describes a property of the predictor matrix used in the pilot data set Please see Muller and Stewart 2006 for details about matrix rank Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy Mase K Page 33 of 58 GLIMMPSE User Manua
33. ee pre specified rows one for each region and two columns one for vitamin D and one for calcium To calculate power the investigator must enter the expected mean vitamin D column 1 and calcium column 2 levels of the children in the rural row 1 suburban 120 4 row 2 and urban row 3 regions The investigator may choose B 60 8 as shown below 45 10 Colorado School of UF Heath Outcomes amp Policy College of Medicine PUBLIC HEALTH Page 43 of 58 GLIMMPSE User Manual Version 2 0 0 Caleulets Regression Coefficients B Matrix The B matrix contains regression coefficients Specify the values you expect to see for these coefficients These values may be determined from pilot data or previous studies We recommend selecting values which represent scientifically meaningful differences At least one of the values in the B matrix should be non zero Coefficients Otherwise power will equal the test size Y Beta Coefficients The number of columns in B indicates the number of outcomes in your study i e the number of columns in Y To adjust the number of outcomes in your study change the column dimension in the text box above the matrix data The number of rows inB must equal the number of columns in Es X so it cannot be adjusted on this screen amp Beta Scale Factors Enter values for the regression coefficients in the matrix below E 2 120 4 60 8 45 10 lt q bb Help E Save Design Cancel Fi
34. equired for GLIMMPSE to calculate power are listed as tabs on the left side of the Introduction screen A white background indicates a tab as active and a blue background designates a tab as inactive Black text designates a page within a tab as active and gray designates a page as inactive Only one page within one tab can be active at a time On the bottom right of any screen in GLIMMPSE is a menu of options enabling users to save their study design by clicking Sedes consult the help library by clicking He or cancel without saving and return to the Start Your Study Design screen by clicking Cancel Each section is broken into one or more sub sections with the title in bold at the top of the page Each screen contains instructions and or areas for user inputs Users navigate through the sections and sub sections by clicking PP to advance or SS to go back The user may also navigate by clicking on the section titles in the left navigation panel Colorado School of UF Heath Outcomes amp Policy College of Medicine PUBLIC HEALTH Page 6 of 58 GLIMMPSE User Manual Version 2 0 0 Introduction The GLIMMPSE wizard will guide you through several steps to perform a power or sample size calculation Use the forward and back arrows to navigate through the wizard You may save your work at any time by clicking the Save Design link at the lower left of the screen The Cancel link also at the lower left of the screen allows
35. g unit The U matrix is most useful for multivariate designs and repeated measures For example suppose an investigator wants to examine whether student test scores improve from their midterm exams to their final exams The investigator would have two measurements per student one for the midterm and one for the final The within participant contrast matrix would be U 1 1 The matrix contrasts two different test scores the midterm and the final for the same student Enter the number of columns or the number of within subject contrasts in the study in the column text box right Press on your keyboard or click anywhere on the screen to resize the blank matrix Fill in the contrasts in the matrix The U matrix must conform to the B matrix so the number of rows cannot be adjusted on this screen Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 46 of 58 GLIMMPSE User Manual Version 2 0 0 Within Subject Contrast U Matrix The U matrix defines the contrasts for within subject effects The matrix is necessary for multivariate and repeated measures designs The number of rows in U must equal the number of columns in B To ensure conforming matrices the row dimension of U cannot be adjusted on this screen Enter your within subject contrast matrix below Hypothesis 9 P s gna Participant 1 Contrast Within Participant Contrast Null Hypothesis Matrix lt q bp i Help E Save Des
36. gn Cancel Figure 44 Type I Error When finished click PP to proceed 4 2 2 Covariate Currently GLIMMPSE only performs power calculations for hypotheses about fixed predictor variables However a single continuous normally distributed predictor variable may be included in the analysis To include such a predictor click the checkbox next to Control for a single normally distributed Gaussian predictor at the bottom of the screen Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy P a g e 41 of 5 8 Y GLIMMPSE User Manual Design Controlling for a single normally distributed predictor A common experimental design is an analysis of covariance which includes one or more fixed predictors and one or more continous control variables the covariates For example one might run an experiment with 10 males and 10 females with an indicator variable for gender as a fixed predictor and age as a covariate A common special case uses a series of repeated measurements on a continuous outcome The first measurement observed prior to treatment is used as a baseline covariate The other repeated measurements are outcomes in the general linear multivariate model GLIMMPSE can calculate power for hypotheses concerning the fixed predictors optionally controlling for a single normally distributed predictor If you plan to include a single normally distributed predictor in your model click the check box bel
37. gure 47 Beta Matrix Enter the number of columns or the number of outcomes in the study in the column text box right in the matrix Press on your keyboard or click anywhere on the screen to resize the blank matrix Enter proposed values of the B coefficients in their corresponding text boxes in the matrix When finished click PP to proceed 4 3 2 Beta Scale Factors GLIMMPSE allows users to specify scale factors for the B matrix in order to generate power or sample size values for different coefficient values Since power is based on proposed regression coefficients it is common to calculate power for the proposed value as well as alternative values such as half and twice the proposed value One or more scale factors for the B matrix may be specified for inclusion in the power calculation For example to calculate power for regression coefficients that are half the values in your B matrix enter 0 5 To use the exact B matrix specified enter 1 After each entry click Lada press Enter or click anywhere on the screen To delete a value select the unwanted value and click to remove the value from the list Colorado School of UF Health Outcomes amp Polic PUBLIC HEALTH ii i Page 44 of 58 GLIMMPSE User Manual Version 2 0 0 Galeulata Scale Factors for Regression Coefficients In power analysis it is not possible to know the exact values of regression coefficients before the experiment is observed Scale fa
38. he user would select a 2 for the male drug group and 1 for the remaining groups Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 20 of 58 GLIMMPSE User Manual Version 2 0 0 Relative Group Sizes Specify whether the study subgroups are of equal or unequal size For equal group sizes select a 1 in the drop down list next to each study Sampling Unit subgroup This is the default study design l id For unequal group sizes specify the ratio of the group sizes For example consider Y Covariate a design with an active drug group and a placebo group If twice as many study Ciusierino participants receive the placebo a value of 2 would be selected for the placebo l group and a value of 1 would be selected for the active drug group Relative Group Sizes Treatment group Gender 1 drug male 1 7 drug female 1 y placebo male 1 7 placebo female qq bb Help 4 Save Design Cancel Figure 20 Relative Group Sizes Screen When finished click PP to proceed 3 3 Smallest Group Size When solving for power the user specifies the total sample size for the design by the relative group sizes and the smallest group size On the Smallest Group Size screen the user may enter one or more values describing the number of participants in the smallest group For example consider a design with a treatment and a placebo group in which three times as many partic
39. i Help 4 Save Design Cancel Figure 30 Scale Factors for Means Screen When finished click PP to proceed 3 7 Variability 3 7 1 Introduction This screen provides an introduction to the Variability section After reading the information on the screen click PP to proceed Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy P a g e 8 of 5 8 GLIMMPSE User Manual Version 2 0 0 Caleulaita Variability Introduction Power analysis requires information about the variability in the study outcomes In this section you will provide standard deviations and correlations for the study outcomes and the covariate if specified Click the forward arrow to continue Variability Yin Participant P Va gma Scale Factors Figure 31 Variability Introduction Screen 3 7 2 Within Participant Variability For a given participant responses may vary across repeated measurements and for different response variables The amount of variability can dramatically impact power and sample size The Within Participant Variability screen allows the user to describe the variability he or she expects to observe for each within participant factor and response variable Separate tabs are presented for each source of correlation in the design The Responses tab allows the user to specify the standard deviations of the response variables and any correlation between them If repeated measures are present a single tab
40. ias RS EEE E E 52 A Olds Statistical ans FRANCIS TR LA we EES SG 52 4 6 2 Power Calculation Method aoao aaa a 02 ce 53 4 6 3 Confidence Intervals 2 oa a a a a 54 4 6 4 Power Curve Options ooo a a a a 55 AT Calle A IEA 56 5 Additional GLIMMPSE Resources 57 Colorado School of UF alh Outcomes amp Policy PUBLIC HEALTH a Page 3 of 58 GLIMMPSE User Manual Version 2 0 0 1 Introduction 1 1 Version Information and Licensing This manual describes version 2 0 0 of the GLIMMPSE software The manual applies to all 2 0 x versions of GLIMMPSE e g 2 0 0 2 0 1 2 0 2 etc GLIMMPSE is released under the GNU Public License version 2 0 The GLIMMPSE program is free software Users can redistribute it and or modify it under the terms of the GNU General Public License as published by the Free Software Foundation using either version 2 of the Li cense or any later version This program is distributed in the hope that it will be useful but WITHOUT ANY WARRANTY without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE See the GNU General Public License for more details You should have received a copy of the GNU General Public License along with this program If you have not received a copy of the GNU General Public License and would like one please write to the Free Software Foundation Inc 51 Franklin Street Fifth Floor Boston MA 02110 1301 USA 1 2 Welcome to GLIMMPSE 2
41. ictions on recruitment and would like to ensure a well powered study then solving for sample size is likely to be more useful Power Total Sample Size q bb Help Save Design X Cancel Figure 11 Solving For Screen When power is selected the inputs will be used for a power analysis The power analysis will produce a value s between O and 1 representing the probability the study will provide an answer to the question of interest When sample size is selected the inputs will be used to calculate the number of individual sampling units also called participants if referring specifically to people needed for the study to achieve the desired power If the number of participants is not set we recommend solving for sample size in order to obtain the appropriate sample size for achieving the goals of your study However if the sample size is set due to budgetary or other restrictions a power calculation will indicate the probability that your study will provide a definitive answer to the question of interest On the screen select Power or Total Sample Size by selecting the appropriate radio button Upon completing the selection click PP to proceed 3 1 3 Desired Power if solving for Total Sample Size When solving for sample size the user must enter the desired power for the study Enter the target values as decimals i e 0 95 in the Power Values box and click to add the value to the list Colorado School of UF He
42. ign X Cancel Figure 50 Within participant Contrast Matrix When finished click PP to proceed 4 4 3 Null Hypothesis The null hypothesis matrix Oo represents the test values the user expects to observe when the null hypothesis is true When performing a power analysis the values for the hypothesis tests are calculated as O CBU and then compared against Og Commonly Oy is a matrix of zeroes For example suppose an investigator wants to compare resting metabolic rate between subjects with HIV lipoat rophy subjects with HIV only and healthy controls The null hypothesis of no difference between the three groups al ob appearing as follows Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 47 of 58 Y GLIMMPSE User Manual Version 2 0 0 Null Hypotheses Oo Matrix For O CBU the general linear hypothesis is stated as Ho O Oo In most cases the Og matrix will contain zeros The number of rows in Og must equal the number of rows in C and the number of Hypothesis i l i columns must match the number of columns in U To ensure conforming matrices Between Participant the dimensions of Og cannot be adjusted on this screen Contrast Within Participant Enter your Og matrix below Contrast x p 4 Null Hypothesis Matrix dl 0 b gt A Help H Save Design X Cancel Figure 51 Null Hypothesis Matrix Sometimes however the null hypothesis is based on previous stu
43. ing Predictor Variables ps m0 5 644 Eds A 11 2 5 Basic Navigation for Matriz Mode 0 12 2 5 1 Resizing and Entering Values Into a Matrix e 12 3 Using Guided Mode A Screen by Screen Tour 12 E E Eee eee eee E 13 SAd bet sxs b L nte 99 5 AE 13 o izo SON DOM ac ceso ooo AA AR AAA 13 3 1 3 Desired Power if solving for Total Sample Size e e 14 S E DEA TOS esa yaa 15 da Damping HER C Dr sere essere aaa 16 Oude MOCHO amp passe LARA ars 16 do DNA TOLDOS capos carros PEDRO REA eee thee es oe ee eee ee 16 OO COMO s wo Seu DE sho LE age SENEGAL LED EAD O Al 17 dae CCO pc ah eee arrasar 18 3 2 5 Relative Group Sizes if solving for Power aoaaa aa a 20 Soop males TOMOS s o R ER ro Issa eee ee es 21 JA PORCO p ara amp poa AAA AA 22 Dl MOCHO y e E E E E E aa E E E das 22 3 4 2 Response Variables a a a 22 3 4 3 Repeated Measures a a 23 Ap Colorado School of UF Health Outcomes amp Policy PUBLIC HEALTH o Page 1 of 58 GLIMMPSE User Manual 3 9 Hypotheses e 3 5 1 Introduction si sentes EM EDS Ou 2s HYPO DICE i a EE E we d 3 0 Means a eee sc aeee ade aaee ani pustida 3 6 1 Introduction 0048 3 6 2 Means 3 6 3 Flexible Means 0 3 7 Variability coso za ooo raro dee eee DES Tele A OQUE arre adas A 3 7 2 Within Participant Variability 3 7 3 Sigma Scale Factors
44. ing the type of hypothesis you wish to test and then enter the requested details Note that trends within an interaction hypothesis may be specified by selecting the Interaction button For more Hypothesis information about the type of hypothesis click the magnifying glass icon 2 Hypothesis Grand mean Main Effect Trend Interaction Select one predictor of interest for the main effect hypothesis Between Participant Factors panees Within Participant Factors month D Help H Save Design Cancel Figure 27 Hypotheses Screen A Grand Mean hypothesis compares the overall mean response in a sample of participants against a known value For example an investigator may wish to determine if body mass index values for participants in a particular state differs from the United States national average After selecting the Grand mean radio button the user will be prompted to enter the known mean value for each response variable A Main effect hypothesis tests for the effect of a single predictor variable averaged across all other factors For example testing whether responses of participants in the treatment group differ on average from participant responses in a placebo group is a common main effect hypothesis After selecting the Main effect radio button the user will be prompted to select one predictor of interest by selecting the appropriate radio button A Trend hypothesis tests whether the effect of a single predict
45. ipants receive the treatment compared to the placebo With a smallest group size of 20 30 or 40 the total sample size for the design would be 80 i e 60 treated participants 20 with placebo 120 and 160 participants respectively Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy po Page 21 of 58 Y GLIMMPSE User Manual Version 2 0 0 Caleta Size of the Smallest Group Enter the number of independent sampling units participants clusters in the smallest group in the study If your group sizes are equal the value is the same for all groups You may enter multiple values for the smallest group size in order to consider a range Sampling Unit of total sample sizes Y Study Groups Enter one or more sample sizes in the text box below and click Add To remove a Y Covariate sample size from the list highlight it and click the Delete button Y Clustering Y Relative Group Sizes Size of the Smallest Group Add baisait Y Smallest Group Size 20 30 40 q Ww Help Y Save Design X Cancel Figure 21 Smallest Group Size Screen When finished click PP to proceed 3 4 Responses 3 4 1 Introduction This screen provides an introduction to the Responses section After reading the information on the screen click PP to proceed Responses Introduction Responses are the quantitative outcomes of interest in a study In this section you will describe the response variables meas
46. ization in a study occurs at the group level rather than by individual research participants The Clustering screen allows the user to enter up to three levels of clustering An example of clustering would be a study design in which the participants are students randomly selected from different schools in an area In this case each school would represent a cluster An example of subgroups within a cluster would be each classroom within a given school Colorado School of PUBLIC HEALTH Health Outcomes amp Policy College of Medicine UF Page 18 of 58 Y GLIMMPSE User Manual Caetilato Sampling Unit Y Study Groups Y Covariate Clustering Relative Group Sizes otart Sampling Unit Study Groups Covariate amp Clustering Y Relative Group Sizes lt q bb Help Version 2 0 0 Clustering In a clustered design the independent sampling unit is a cluster such as a community school or classroom Observations within a cluster are correlated The labels for observations within a cluster must be exchangeable For example child id within classroom can be reassigned arbitrarily In contrast observations across time cannot be reassigned and should not be considered clustered observations Clustering or repeated measures or a combination creates a multilevel design The common correlation between any pair of cluster members is termed the intraclass correlation or intracluster correlation
47. l Version 2 0 0 Matrix Mode allows direct input of all matrices required for a power calculation In Matrix Mode users receive less guidance than in Guided Mode and are assumed to possess in depth statistical training 4 1 Start 4 1 1 Introduction The Introduction screen briefly describes the required matrix inputs for the power or sample size calculation Colorado School of PUBLIC HEALTH UF College of Medicine Health Outcomes amp Policy Page 36 of 58 Y GLIMMPSE User Manual Gzleuleats Start amp Solving For amp Type Error Design Coefficients Hypothesis Variability Options Version 2 0 0 Introduction The GLIMMPSE wizard will guide you through several steps to perform a power or sample size calculation Use the forward and back arrows to navigate through the wizard or click on the desired option in the left navigation panel You may save your work at any time by clicking the Save Design link at the lower right of the screen The Cancel link also at the lower right of the screen allows you cancel your current work and begin a new study design The help manual may be accessed by clicking the Help link General steps for a power calculation for the general linear multivariate model are listed on the left hand side of the screen If you have questions regarding terminology please consult the glossary in the user manual We will ask you to specify e The Type l error rate
48. l Version 2 0 0 Confidence Interval Options If the means B or the error covariance Ze are sample estimates then the power values produced from these matrices will be random quantities To account for this randomness GLIMMPSE can calculate confidence intervals for power values using the techniques described by Taylor and Muller 1995 Gribbin 2007 and Park 2007 Don t include confidence intervals for power Select the assumptions for the confidence intervals Bis fixed but Xe is estimated Both B and Xe are estimated Enter the upper and lower tail probabilities for the confidence intervals We typically recommend the value 0 05 for the lower tail probability and O for the upper tail probability Options Lower tail 0 025 Y Statistical Test Upper tail 0 025 Y Power Method Describe the data from which you obtained the values for B and Ze Confidence Intervals Y Power Curve Total sample size 100 Rank of the design matrix 2 SN Ww Help Y Save Design X Cancel Figure 37 Confidence Intervals When finished Click PP to proceed 3 8 4 Power Curve Options The Power Curve Options screen allows the user to create a power curve A power curve describes the change in power Y axis of the power curve relative to the total sample size regression coefficient scale factor or the variability scale factor all options for the X axis of the power curve To create a power curve the user must 1 uncheck the check b
49. le s also called dependent variables Figure 8 shows an example of entering variable labels To enter a variable label type the label into the text box provided see 1 in Figure 8 After each entry click Lada press on the keyboard or click anywhere on the screen to populate the field below that text box see 2 in Figure 8 For the predictor variables GLIMMPSE also asks the user to specify the categories for each variable For example the predictor variable gender has two categories male and female To specify categories associated with a given predictor select a predictor in the text box on the left see 2 then enter the category labels into the text box on the right see 3 in Figure 8 After each entry click L da press on the keyboard or click anywhere on the screen to populate the category text box see 4 Only category labels associated with the highlighted predictor label are shown To delete predictors or category labels select the unwanted label and click Delete This removes the label from the list If the user removes a predictor the associated categories are automatically deleted Colorado School of Health Outcomes amp Polic PUBLIC HEALTH UF Eme f Page 11 of 58 GLIMMPSE User Manual Version 2 0 0 11 3 Predictor ea Category ya Lada Female Em Gender Male Race Ethnicity 2 4 Figure 8 Example of entering labels 2 5 Basic Navigation for Matriz Mode
50. lect the power methods by clicking the check boxes If quantile power is selected the user must also specify one or more quantile values For example median power would be obtained by selecting Quantile power and entering 0 5 in the quantile list box Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 32 of 58 GLIMMPSE User Manual Version 2 0 0 Power Calculation Method For designs including a baseline covariate two methods are available to calculate power unconditional power and quantile power One can think of the random covariate values as having been sampled from a normal distribution Thus there are many possible realizations of the same experiment and each realization may have a different power The unconditional power is defined as the average of the possible power values Gatsonis and Sampson 1989 Glueck and Muller 2003 The 100 x yn quantile power is the power value chosen so that power as small or smaller occurs in 100 x v percent of all possible realizations of the experiment For a detailed description of unconditional and quantile power please see Gatsonis and Sampson 1989 and Glueck and Muller 2003 Options Select one or more power methods below Y Statistical Test Unconditional v Power Method Y Quantile Confidence Intervals a l Quantiles Add Delete y owe Curve 0 5 lt bp Help Y Save Design X Cancel Figure 36 Statistical Tests When finish
51. nalysis it is necessary to provide mean values for the outcomes The means are often obtained from published results or pilot data To detect a difference between the groups at least two subgroups should have means which differ by a scientifically meaningful amount For example in a study of cholesterol lowering medication we would expect the mean cholesterol level in the active drug group to be lower than the placebo group Means Click next to specify the means you expect to observe Scale Factors for Means Figure 28 Means Introduction Screen 3 6 2 Means The Means screen allows the user to enter the expected mean value for the experiment Expected mean values are typically drawn from the literature or from pilot data Differences between the entered means typically represent the smallest clinically relevant difference The table should contain at least one value that is non zero Means The table below shows the mean values for each outcome within each study subgroup The study subgroups are listed along the left hand side of the table and the outcomes are listed across the top Enter the mean values you expect to observe for each outcome within each study subgroup The table should contain at least one value that is non zero Also at least two subgroups should have means which differ by a scientifically meaningful amount treatment blood pressure Means drug 0 0 7 Means y Y Scale Factors for Means Pare
52. ng decay values between 0 05 and 0 5 Base Correlation 07 Decay Rate 0 05 month 1 month 2 month 3 month 4 month 5 month 6 month 1 1 0 0 7 0 69688 0 69378 0 69069 0 68762 month 2 0 7 1 0 0 7 0 69688 0 69378 0 69069 month 3 0 69688 0 7 1 0 0 7 0 69688 0 69378 month 4 0 69378 0 69688 0 7 1 0 0 7 0 69688 month 5 0 69069 0 69378 0 69688 0 7 1 0 0 7 month 6 0 68762 0 69069 0 69378 0 69688 0 7 1 0 Unstructured correlation Help Figure 32 Within Participant Variability When finished click DP to proceed 3 7 8 Sigma Scale Factors Save Design X Cancel While GLIMMPSE requests standard deviations it actually computes variances when it conducts the power or sample size calculations There may be considerable uncertainty about what standard deviation or variance value to use To account for this uncertainty it is common to calculate power or sample size using alternative values for variability The Flexible Variability screen allows you to compute power for half the variance the variance as specified and twice the variance This generates scale factors of 0 5 1 and 2 for the covariance matrix If you wish to have a range of variances check the checkbox Colorado School of PUBLIC HEALTH Health Outcomes amp Policy College of Medicine UF Page 30 of 58 Y GLIMMPSE User Manual Version 2 0 0 Caleulata Flexible Variability On the previous screens you entered standard deviations and correlations
53. ning only the fixed predictors and excluding the Gaussian covariate The Outcomes Covariance screen allows the user to define My by directly entering the covariance matrix values To ensure conformance with the B and U matrices the dimensions of the Xy matrix cannot be modifed on this screen Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 49 of 58 Y GLIMMPSE User Manual Csietiato Variability 2 Outcomes Covariance Y Variance of Covariate Covariance of Outcomes and Covariate Sigma Scale Factors Covariance of Outcomes ly For each independent sampling unit Zy is the covariance of the outcomes conditional on the fixed predictors One can think of Zy as the error covariance for each independent sampling unit in a model containing only the fixed predictors and excluding the Gaussian covariate For univariate designs Zy will be a 1x1 matrix containing the variance of the outcome conditional on the fixed predictors More complex structures may be entered for multivariate or repeated measures designs Ly is a square symmetric matrix with dimensions equal to the number of columns in B Enter values for Zy in the matrix below x 4 4 1 0 2 0 2 1 0 1 0 2 0 05 0 1 SNN Ww 0 1 0 2 0 2 0 05 0 1 0 2 1 Help Figure 54 Outcomes Covariance When finished click PP to proceed 4 5 3 Variance of Covariate Y Save Design X Cancel Version 2 0 0 For designs wi
54. nk of the design matrix 2 de vv Help Save Design X Cancel Figure 61 Confidence Intervals When finished Click PP to proceed 4 6 4 Power Curve Options The Power Curve Options screen allows the user to create a power curve A power curve describes the change in power Y axis of the power curve relative to the total sample size regression coefficient scale factor or the variability scale factor all options for the X axis of the power curve To create a power curve the user must 1 uncheck the check box 2 select the value to appear on the horizontal axis and 3 add one or more data series Depending on the study design the user may request a large number of power or sample size values in a single request A data series is defined by selecting a subset of the power or sample size values The user creates a data series by selecting values for several study design variables and clicking the button A data series will be displayed as a single line on the power curve plot Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy spree Page 55 of 58 GLIMMPSE User Manual Caleta Options Y Statistical Test Power Method Confidence Intervals q amp 4 Power Curve Power Curve Options You may optionally create a power curve image for your results by unchecking this checkbox Then select the values you would like to display on the power curve by selecting the appropriate options below d
55. o not want to create a power curve Version 2 0 0 1 Select the quantity to display on the horizontal axis of the power curve the vertical axis will display the power value Total Sample Size 2 Add data series to the plot Select values for each variable below Click add to include sample size values matching these criteria as a data series on the plot To remove a data series highlight it in the list box and click Remove data series Data Series Label Series 1 Regression Coefficient Scale Factor 2 0 Variability Scale Factor 1 0 Hotelling Lawey Trace lt Statistical Test Type Error 0 01 Power Method Quantile Power Quantile 05 Data Series Label Regression Coefficient Scale Factor Variability Scale Factor Statistical Test Series 1 2 0 1 0 Hotelling Lawley Tr 0 4 m gt i bp Help Y Save Design X Cancel Figure 62 Power Curve When finished click PP to proceed to the results screen Note that the Power Curve Options screen is the final screen in the GLIMMPSE wizard If the study design is not complete the PP button will be disabled 4 7 Calculate When sufficient information has been entered for your power or sample size calculation the Calculate button will be highlighted green Click MEA to receive the results of your power analysis Example results are shown in Figure 63 For detailed information regarding the Power Results table refer to Table 1 Colorado School of PUB
56. o save the power results click Saveto CSV beneath the table of results For transparency the matrices used in the calculations are accessible on the results screen To view the exact matrices used in the calculations click Matices beneath the table of results This is most useful in Guided Mode where matrix information is largely hidden from the user Colorado School of UF Heath Outcomes amp Policy PUBLIC HEALTH ia Page 10 of 58 GLIMMPSE User Manual Version 2 0 0 Column Name Description Test Name of the statistical test Actual Power Calculated power Total Sample Size Total number of research participants required to achieve the actual power Beta Scale Scale factor applied to the B or Bp matrix Sigma Scale Scale factor applied to the Siz matrix Alpha The Type I error value Nominal Power The desired power Power Method Indicates whether conditional unconditional or quantile power was used Juantile If the current power method is quantile power this indicates the quantile of the distribution of possible powers Otherwise this field is empty Power Lower Lower limit of the 95 confidence interval Power Upper Upper limit of the 95 confidence interval Table 1 Information displayed for each power result 2 4 Basic Navigation for Guided Mode 2 4 1 Entering Predictor Variables In Guided Mode GLIMMPSE requires the user to enter labels for predictor variable s also called independent variables and outcome variab
57. of a fixed predictor is treatment group for which the independent sampling unit is randomized to a placebo or an active drug group For a one sample design leave the table blank To enter fixed predictors 1 Enter the name of each predictor in the left text box and click Add For example one might enter treatment as a predictor 2 Select the predictor from the left text box to display the current list of values associated with the predictor To add a new value enter the value in the Category text box and click Add For example one could select treatment then add the values drug and placebo Each predictor should have at least two values Category Add Delete Delete home based program i delayed program control q bp Help kh Save Design X Cancel Figure 15 Study Groups Screen When finished click PP to proceed 3 2 3 Covariate The Covariate screen allows the user to control for a single normally distributed predictor also known as a normally distributed covariate For example a scientist may wish to examine the effect of a drug when controlling for age In this case age would be the covariate If the study design does not include a normally distributed predictor leave the checkbox blank If the study design does include a covariate check the checkbox Colorado School of PUBLIC HEALTH Health Outcomes amp Policy College of Medicine UF Page 17 of 58
58. of contrasts in the study in the row text box left under C Matrix Press on your keyboard or click anywhere on the screen to resize the blank matrix Fill in the contrasts you wish to test in the matrix The number of rows in the C matrix cannot exceed the number of rows in the essence matrix minus 1 In addition the C matrix must conform to the B matrix so the number of columns cannot be adjusted on this screen Health Outcomes amp Policy Colorado School of UF P A5 of 58 College of Medicine ag e O PUBLIC HEALTH GLIMMPSE User Manual Version 2 0 0 Calera Between Subject Contrast C Matrix The C matrix defines the contrasts for between subject effects The number of rows in the C matrix is at most one fewer than the number of rows in Es X The number of columns in C must equal the number of columns in B To ensure conforming matrices the number of columns of C cannot be adjusted on this screen Enter your between subject contrast matrix below Hypothesis 4 X 9 Between Participant 1 1 Contrast Within Participant Y PA a pa ee Null Hypothesis Matrix bb i Help E Save Design Cancel Figure 49 Between participant Contrast Matrix When you have completed the matrix click PP to proceed 4 4 2 Within Participant Contrast The U matrix consists of the within participants contrasts The within participants contrasts are the hypotheses that compare measurements on the same independent samplin
59. or follows a particular polynomial pattern such as a linear or quadratic trend across different levels of the predictor After selecting the Trend radio button the user will be prompted to select one predictor of interest In addition the user may select from six possible trends no trend change from baseline linear trend quadratic trend cubic trend or all polynomial trends An Interaction hypothesis tests whether the effect of one predictor changes depending on the value of one or more additional predictors An interaction test also can be interpreted as a test of differences as well as a test of parallel trajectories of response For example testing whether the effect of a cholesterol lowering medication on total serum cholesterol differs depending on the participant s gender is an example of an interaction hypothesis After selecting the Interaction radio button the user will be prompted to select one or more factors of interest by clicking the appropriate check boxes In addition the user may specify a trend for given factor by clicking the Edit trend button When finished click PP to proceed 3 6 Means Colorado School of UF Heath Outcomes amp Policy College of Medicine PUBLIC HEALTH Page 26 of 58 GLIMMPSE User Manual Version 2 0 0 3 6 1 Introduction This screen provides an introduction to the Means section After reading the information on the screen click bp to proceed Means Introduction In power a
60. oup Size 3 1 El a w iad RT 05 Figure 5 Examples of text boxes that are used to A collect information on repeated measures B specify the size and contents of a matrix and C specify one or more choices for an item used in the power calculation 2 3 2 Using Drop Down Lists When GLIMMPSE requires you to choose from a defined list of options these options will be presented in a drop down list Figure 3 shows an example of a drop down list To choose an option from a drop down list click on the down arrow see 1 then select your choice from the list of options see 2 1 Relative Gender Group Siz 1 vi Male v Female 2 900 JO OA ORN Figure 6 Example of a drop down menu 2 3 3 Radio Buttons and Check Boxes In some cases you must choose from a list of options by selecting a radio button or checking a box The radio buttons allow you to select only one option The check boxes allow you to select more than one option To select an option click on the radio button or check the box next to that option Figure 7 shows an example of a radio button see A and a check box see B Colorado School of UF Heath Outcomes amp Policy College of Medicine PUBLIC HEALTH Page 9 of 58 GLIMMPSE User Manual Version 2 0 0 Grand mean Main Effect Trend SE Interaction A Choose one option Select two or more predictors to Include in the interaction hypothesis To test for
61. ow At present the GLIMMPSE software does not calculate power for multiple normally distributed predictors nor non normally distributed predictors C Control for a single normally distributed predictor lt q bb Help E Save Design Figure 45 Gaussian Predictor Screen When finished click DP to proceed 4 2 3 Smallest Group Size Version 2 0 0 When solving for power the user specifies the total sample size for the design by the relative number of repeated rows in the design essence matrix and the smallest group size On the Smallest Group Size screen the user may enter one or more values describing the number of participants in the smallest group To enter one or more per group sample size type the sample size in the Per Group Sample Size box After each entry click Lada press on your keyboard or click anywhere on the screen To delete a value select the unwanted value and click to remove the value from the list Colorado School of PUBLIC HEALTH Health Outcomes amp Policy College of Medicine UF Page 42 of 58 GLIMMPSE User Manual Version 2 0 0 Caleukiis Size of the Smallest Group Enter the number of independent sampling units participants clusters in the smallest group in the study If your group sizes are equal the value is the same for all groups You may enter multiple values for the smallest group size in order to Design consider a range of total sample sizes E Enter one or mo
62. ox 2 select the value to appear on the horizontal axis and 3 add one or more data series Depending on the study design the user may request a large number of power or sample size values in a single request A data series is defined by selecting a subset of the power or sample size values The user creates a data series by selecting values for several study design variables and clicking the button A data series will be displayed as a single line on the power curve plot Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 34 of 58 Y GLIMMPSE User Manual Version 2 0 0 Caleulata Power Curve Options You may optionally create a power curve image for your results by unchecking this checkbox Then select the values you would like to display on the power curve by selecting the appropriate options below FL do not want to create a power curve 1 Select the quantity to display on the horizontal axis of the power curve the vertical axis will display the power value Total Sample Size 2 Add data series to the plot Select values for each variable below Click add to include sample size values matching these criteria as a data series on the plot To remove a data series highlight it in the list box and click Remove data series a Data Series Label Series 1 Options Regression Coefficient Scale Factor 1 0 Y Statistical Test Variability Scale Factor 1 0 x Y Confidence Intervals Statistical Tes
63. pant s blood pressure once a month for six months If the design does not have repeated measures click PP to proceed If the design includes repeated measures click Add repeated measures and fill in the requested information Units is a user specified description of the repeated measure For example if the repeated measures are taken once every month the unit could be month Enter a label for the units of the repeated measure Enter the Type of unit For Numeric repeated measures both the distance and ordering between measurements is meaningful Measuring blood pressure every month for 6 months is a numeric repeated measure GLIMMPSE will auto populate an equal distance between repeated numeric measures You can change the distance between the measures by typing into the text boxes For Ordinal repeated measures the ordering of the measurements is meaningful but the distance between measurements is assumed to be equal For example repeated measures of the participant s heart rate taken in the morning afternoon and evening For Categorical repeated measures neither the ordering nor the distance between the measures is meaningful For example repeated measures of breast density using three different instruments Device A B and C Number of Measurements allows you to specify the number of times the repeated measure will be taken For the blood pressure example the Number of Measurements would b
64. ple a centered 95 confidence interval would have both upper and lower tail probabilities of 0 025 The Total sample size value indicates the number of independent sampling units in the pilot data set or publica tion The Rank of the design matriz describes a property of the predictor matrix used in the pilot data set Please see Muller and Stewart 2006 for details about matrix rank Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 54 of 58 GLIMMPSE User Manual Version 2 0 0 Calle tft Confidence Interval Options If the means B or the error covariance Ze are sample estimates then the power values produced from these matrices will be random quantities To account for this randomness GLIMMPSE can calculate confidence intervals for power values using the techniques described by Taylor and Muller 1995 Gribbin 2007 and Park 2007 Don t include confidence intervals for power Select the assumptions for the confidence intervals Bis fixed but Xe is estimated Both B and Xe are estimated Enter the upper and lower tail probabilities for the confidence intervals We typically recommend the value 0 05 for the lower Options tail probability and O for the upper tail probability Y Statistical Test Lower tail 0 025 Y Power Method Upper tail 0 025 2 Confidence Intervals J Powa A Describe the data from which you obtained the values for B and Ze Total sample size 100 Ra
65. r and Barton 1989 Muller Lavange Ramey and Ramey 1992 Muller Edwards Simpson and Taylor 2007 Muller and Stewart 2006 and Muller et al 2007 Details for fixed predictors with a single Gaussian covariate can be found in Glueck and Muller 2003 GLIMMPSE utilizes a Java web services architecture McGovern Tyagi Stevens and Mathew 2003 designed to facilitate future support of additional statistical models The tool is hosted at http glimmpse samplesizeshop Org 1 3 Why GLIMMPSE Other programs such as POWERLIB NQuery and Pass also calculate power and sample size So why use GLIMMPSE GLIMMPSE has several advantages over these other programs because GLIMMPSE 1 Is free GLIMMPSE provides free online power and sample size computing 2 Is user friendly In both Guided Mode and Matrir Mode GLIMMPSE provides a step by step interface to assist researchers in producing accurate power and sample size calculations Colorado School of UF Heath Outcomes amp Policy College f Medici PUBLIC HEALTH pe Page 4 of 58 GLIMMPSE User Manual Version 2 0 0 3 Calculates power and sample size for any univariate or multivariate test for the general linear multivariate model assuming fixed predictors 4 Produces confidence intervals on power estimates for designs with fixed predictors 5 Produces power and sample size calculations for designs with a single Gaussian covariate 6 Supports designs with unequal group si
66. r for Balanced Linear Mixed Models Statistics in Medicine 26 19 3639 3660 Muller KE Lavange LM Ramey SL Ramey CT 1992 Power Calculations for General Linear Multivariate Models Including Repeated Measures Applications Journal of the American Statistical Association 87 420 1209 1226 Muller KE Peterson BL 1984 Practical Methods for Computing Power in Testing the Multivariate General Linear Hypothesis Computational Statistics and Data Analysis 2 143 158 Muller KE Stewart PW 2006 Linear Model Theory Univariate Multivariate and Mixed Models John Wiley and Sons Hoboken New Jersey Colorado School of UF Heath Outcomes amp Policy College f Medici PUBLIC HEALTH pe Page 58 of 58
67. re sample sizes in the text box below and click Add To remove a Y Covariate sample size from the list highlight it and click the Delete button 2 Smallest Group Size Size of the Smallest Group Add Delete lt q bb i Help E Save Design X Cancel Figure 46 Smallest Group Size Screen When finished click PP to proceed 4 3 Coefficients 4 3 1 Beta Coefficients B Matrix This section requires the user to enter choices for values for the hypothesis test O CBU In the general linear multivariate model with fixed predictors Y XB E the B matrix represents the proposed relationship between the predictor variables X and the outcome variables Y The same is true for Br in the General Linear Multivariate Model with Fixed Predictors and a Gaussian Predictor For simplicity we will only discuss B since the instructions do not change for Bp To calculate power enter values for the regression coefficients for each unique combination of predictors in the study design The row dimension of B is determined by the number of columns in the essence matrix Change the column dimension of B to match the intended number of outcomes in the study or the columns of Y in the general linear multivariate model with fixed predictors regression equation For example an investigator may want to compare vitamin D and calcium levels of children who live in three different regions urban suburban and rural The B matrix would have thr
68. servations or sub clusters BV 50 within each cluster of this type Intra cluster correlation 25 Cluster label Classroom Number of observations or sub clusters within each cluster of this type Intra cluster correlation 5 Cluster label Number of observations or sub clusters within each cluster of this type Intra cluster correlation Remove subgrou qq bb Help H Save Design Cancel Figure 19 Clustering Sub Groups Continuing with the above example the subgroup Cluster name would be classroom the Number of observations would be the number of students within each classroom and the ntra cluster correlation would be the expected agreement between students within each classroom To remove a subgroup or remove clustering simply click Remove subgroup or Remove clustering When finished click PP to proceed 3 2 5 Relative Group Sizes if solving for Power For designs with multiple study groups see Section 3 2 2 the user may specify equal or unequal group sizes On the Relative Group Sizes screen the user can select the relative sizes of each group by selecting a value from the drop down list For example consider a design with males and females randomized to receive either an active drug or a placebo For equal group sizes a 1 should be selected for each drop down list as shown in Figure 20 However if there were twice as many males receiving the drug compared to females receiving the drug t
69. specify the Number of observations or sub clusters within each cluster of this type and specify the Intra cluster correlation The Intra cluster correlation is the expected correlation between pairs of observations within the cluster Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy cae a Page 19 of 58 GLIMMPSE User Manual Version 2 0 0 To add a subgroup to the cluster click Add subgroup and fill in the information for that subgroup GLIMMPSE allows one primary cluster and two subgroups Clustering In a clustered design the independent sampling unit is a cluster such as a community school or classroom Observations within a cluster are correlated The labels for observations within a cluster must be exchangeable For example child Sampling Unit id within classroom can be reassigned arbitrarily In contrast observations across i time cannot be reassigned and should not be considered clustered observations Y Study Groups E E l Clustering or repeated measures or a combination creates a multilevel design Y Covariale The common correlation between any pair of cluster members is termed the amp Clustering intraclass correlation or intracluster correlation Y Relative Group Sizes To include clustering in the study click Add clustering and follow the prompts Use the Remove clustering button to remove clustering information Remove clustering Cluster label School Number of ob
70. stical Test Univariate Approach to Repeated Measures with Huynh Feldt Correction Y Confidence Intervals lO Univariate Approach to Repeated Measures uncorrected Y Power Curve q bp Help 4 Save Design Cancel Figure 59 Statistical Tests When finished click PP to proceed 4 6 2 Power Calculation Method For designs with a baseline covariate two different methods are available to calculate power quantile and uncon ditional power For theoretical details please see Glueck and Muller 2003 Select the power methods by clicking the checkboxes If quantile power is selected the user must also specify one or more quantile values For example median power would be obtained by selecting Quantile and entering 0 5 in the Quantiles list box Health Outcomes amp Policy Colorado School of UF P 53 of 58 College of Medicine ag e O PUBLIC HEALTH GLIMMPSE User Manual Version 2 0 0 Caleulata Power Calculation Method For designs including a baseline covariate two methods are available to calculate power unconditional power and quantile power One can think of the random covariate values as having been sampled from a normal distribution Thus there are many possible realizations of the same experiment and each realization may have a different power The unconditional power is defined as the average of the possible power values Gatsonis and Sampson 1989 Glueck and Muller 2003 The 100 x yn quantile power is the power
71. t Hotelling Lawley Trace Power Curve Type Error 0 05 Data Series Label Regression Coefficient Scale Factor Variability Scale Factor Statistical Test 1 0 Hotelling Lawley Tr 0 05 T Series 1 1 0 m Help H Save Design Cancel Figure 38 Power Curve When finished click PP to proceed to the results screen Note that the Power Curve Options screen is the final screen in the GLIMMPSE wizard If the study design is not complete the PP button will be disabled 3 9 Calculate When sufficient information for your power or sample size calculation has been entered the Calculate button will be highlighted green Click to receive the results of your power analysis Example results are shown in Figure 39 For detailed information regarding the Power Results table refer to Table 1 Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy P a g e 35 of 58 GLIMMPSE User Manual Calculate Power Curve AMET Po 75 20 0 12 3 13 0 Total Sample Size Power by Total N Power Results Power Total Sample Size 0 206 6 0 485 5 0 722 10 0 897 12 0 942 14 0 978 16 0 231 18 0 987 20 Save to LSN View Matrices Li Test HLT HLT HLT HLT HLT HLT HLT Type Error Rate 0 05 0 05 0 05 0 05 0 05 0 05 0 05 0 05 Figure 39 Results Mesns Scale Factor Variability Scale Factor Help Y Save Design 4 Matriz Mode Screen by Screen Tour Cance
72. th a Gaussian covariate the covariate is assumed to have a Gaussian distribution with mean zero and variance E The Variance of Covariate screen allows the user to enter a value for o Casiletiato Variability Outcomes Covariance Warance of Covanate Covariance of Outcomes and Covariate Y Sigma Scale Factors Colorado School of PUBLIC HEALTH Variance of the Gaussian Covariate og 2 The Gaussian covariate is assumed to follow a Gaussian distribution with mean zero and variance 07g Enter values for 0 in the matrix below x 1 1 1 Help Figure 55 Variance of Covariate Health Outcomes amp Policy College of Medicine UF Y Save Design q X Cancel Page 50 of 58 GLIMMPSE User Manual Version 2 0 0 When finished click PP to proceed 4 5 4 Covariance of Outcomes and Covariate When controlling for a Gaussian covariate power is typically improved when the covariate explains some portion of the variance in the outcome The covariance matrix between the outcomes and the Gaussian covariate Mya describes the association between the outcomes and the Gaussian covariate The Covariance of Outcomes and Covariate screen allows the user to specify values for Myg To ensure conformance with the Xy matrix the dimensions of the Nyq matrix cannot be modifed on this screen Cajeulata Covariance of Outcomes with Covariate Zyg When controlling for a covariate power incr
73. ured in your study If the response variables were measured on multiple occasions for each participant you may describe the type and timing of the repeated measurements Responses Click the forward arrow to continue Y Repeated Measures Figure 22 Responses Introduction Screen 3 4 2 Response Variables The Response Variables screen allows the user to specify the response or dependent variables for the study For example if expected pain is the desired outcome enter expected pain in the text box Colorado School of UF PUBLIC HEALTH Health Outcomes amp Policy o Page 22 of 58 GLIMMPSE User Manual Version 2 0 0 Calculata Response Variables Enter the response variables in the table below For example in a study investigating cholesterol lowering medication the response variable could be HDL LDL and total cholesterol Note that repeated measurement information will be addressed on the next screen Responses Response Variables Response Variables Add Delete D Repeated Measure Expected pain N bp Help E Save Design X Cancel Figure 23 Response Variables When finished click PP to proceed 3 4 3 Repeated Measures The Repeated Measures screen allows the user to describe repeated measures Repeated measures are present in a study when multiple measurements are taken on each research participant An example of repeated measures would be researchers taking a partici
74. value chosen so that power as small or smaller occurs in 100 x v percent of all possible realizations of the experiment For a detailed description of unconditional and quantile power please see Gatsonis Options and Sampson 1989 and Glueck and Muller 2003 AA H Select one or more power methods below 2 Power Method Unconditional Confidence Intervals o g Quantile Y Power Curve Quantiles Add si 0 5 q bp Help Y Save Design X Cancel Figure 60 Power Method When finished click PP to proceed 4 6 8 Confidence Intervals Power analysis involves some uncertainty in the choices for means and variability Therefore the Confidence Intervals screen allows the user to request confidence intervals on the power results To include confidence intervals uncheck the checkbox The information on the confidence interval screen describes the data set or publication from which the choices for means and variances were obtained For example if a scientist was calculating power based on the means and variances obtained from pilot data the scientist would enter information describing the pilot data set The following information is required The Assumptions section allows the user to indicate if he or she is uncertain about the variance but reasonably certain of the mean values or uncertain of both the means and variance The Upper and lower tail probabilities define the width of the confidence interval For exam
75. values To remove a value select the value in the list box and click the Delete button Type Error Values 01 7 05 bp Help E Save Design Figure 13 Type I Error Health Outcomes amp Policy College of Medicine UF X Cancel Page 15 of 58 GLIMMPSE User Manual Version 2 0 0 When finished click PP to proceed 3 2 Sampling Units 3 2 1 Introduction This screen provides an introduction to the Sampling Units section and defines the concept of an independent sampling unit The sampling unit is typically the study participant For multilevel designs and cluster randomized trials the sampling unit may be a group of participants such as schools or neighborhoods Cele tices Independent Sample Unit Introduction In this section you will describe features which distinguish independent observations Sampling Unit o e Independent observations have values which are statistically independent a id e An independent sampling unit ISU provides one or more observations such Covariate that observations from one unit are statistically independent from any other Y Clustering distinct unit while observations from the same unit may be correlated e The observational unit distinguishes one correlated observation from another within the ISU e Observing the same variable in two or more instances across time space or other dimension within an ISU creates repeated measures
76. will be presented for each level of repeated measures Figure 32 shows an example design in which blood pressure is measured once a month for six months GLIMMPSE will automatically combine the sources of correlation into a final covariance matrix Colorado School of UF Heath Outcomes amp Policy College of Medicine PUBLIC HEALTH Page 29 of 58 gt GLIMMPSE User Manual Cale ulerta Within Participant Variability Version 2 0 0 For a given research participant responses vary across response variables and across repeated measurements The amount of variability can dramatically impact power and sample size Click on each of the tabs below to describe the varibility you expect to observe for the response variables and each within particpant factor month Responses Structured Correlation The Linear Exponential Auto Regressive Model LEAR Simpson et al 2010 Variability The LEAR model describes correlation which monotonely decreases with distance between repeated measurements The model has two correlation parameters the Within Participant Variability Y Sigma Scale Factors base correlation and the decay rate The base correlation describes the correlation between measurements taken 1 unit apart The decay rate describes the rate of decrease in the base correlation as the distance or time between repeated measurements increases Our experience with biological and behavioral data lead us to suggest usi
77. would be Es X O DP eee ee pd pa Al A pel 2 OO O O O o oOo oF FF OOO e e pd pd pi OOO FR KF KF OO O amp O SO O H H O O A O ta GLIMMPSE requires that the design coding is full rank Unequal group sizes may be coded by replicating a row to reflect the relative sizes of the groups After entering the desired dimensions for the matrix in the row and column dimension text boxes click anywhere on the screen for the matrix to be resized Type in the matrix text boxes to enter the matrix information Colorado School of Health Outcomes amp Polic PUBLIC HEALTH UF Eme f Page 40 of 58 Y GLIMMPSE User Manual Version 2 0 0 The Design Essence Matrix In the general linear model Y XB E the X matrix contains predictor and covariate information For power analysis please specify a design essence matrix Es X The Es X matrix contains one and only one copy of each unique row in the Design full design matrix This allows separation of the study design information from overall and relative sample size Design Essence v Covariate Enter the Es X matrix below To change the row dimension of the matrix enter the updated number of rows in the left most textbox above the matrix data To change the column dimension enter the desired number of columns in the right most textbox above the matrix data Please use a full rank coding for this matrix 3 o 1 0 0 0 1 0 0 0 1 lt q bb Help E Save Desi
78. zes and complicated covariance structures 7 Creates basic power curves 2 Using GLIMMPSE 2 1 When to Use GLIMMPSE GLIMMPSE is a tool researchers and scientists can use to calculate reliable values for power and sample size GLIMMPSE calculates power or sample size for designs with normally distributed outcomes and for a variety of multilevel and longitudinal studies GLIMMPSE can calculate power and sample size for common statistical tests and models including e One sample t test e Paired t test e Two sample t test e Analysis of variance ANOVA e Analysis of covariance ANCOVA e Repeated measures analysis of variance e Multivariate analysis of variance MANOVA e Multivariate analysis of covariance MANCOVA 2 2 How to Use GLIMMPSE 2 2 1 Initiating the GLIMMPSE Wizard GLIMMPSE can be accessed with a standard web browser at http glimmpse samplesizeshop org The GLIMMPSE start screen is shown in Figure 1 GLIMMPSE has been tested in Internet Explorer 8 Microsoft 2010 Mozilla Firefox 13 0 1 Mozilla 2011 Google Chrome 23 0 1271 95 Google 2011 and Safari 5 0 3 Apple 2010 Colorado School of Health Outcomes amp Polic PUBLIC HEALTH UE ci sr i Page 5 of 58 GLIMMPSE User Manual Version 2 0 0 Start Your Study Design Select one of the options below to begin your power or sample size estimate Guided Study Design Matrix Study Design Upload a Study Design Build common study designs including Dir

Download Pdf Manuals

Related Search

Related Contents

Bedienungsanleitung herunterladen  Golla Cam Bag - PRO  IDEAS User Manual - IDEAS™ v6.0.0  think forward DIFFRACplus TOPAS  Toshiba Portégé Z10t-A-10M  Rheem XE50M12EC45U0 Instructions / Assembly  Kettler 08451-799 User's Manual  Frigidaire FFRE1533Q1 Energy Guide  HD4469 HD4468 HD4467 HD4408 HD4407  BridgeX GT Series BX5020 Gateway IB to Ethernet User Manual  

Copyright © All rights reserved. Failed to retrieve file