Speaker

Biography

Abstract

Caroline Caudan

Caroline is a Statistician with 10 years of experience. She started her career working on the manufacturing and control part of the pharmaceutical industry in order to improve production and control processes.

For the last 4 years, she has been working as a biostatistician on the clinical part of the product development. Since she joined Keyrus Biopharma 3 years ago, she has been involved in the control and analysis of biomarker data, which led her to take a close interest to graphical representation and data visualization. Following this, Caroline started to work, in collaboration with Roche, on the development of a R Shiny application as a support to statistical monitoring.

Interactive statistical monitoring to optimize review of potential study issue with R-Shiny

Background: Statistical Monitoring involves the review of prospective study data collected in participating site to detect inconsistencies between patients and between sites in term of trends.

Method: A Phase IV study (PRO-MSACTIVE) is currently evaluating ocrelizumab in active relapsing Multiple Sclerosis patients in France. Specific statistical methods (volcano plots, mahalanobis distance, funnel plot …), described in a statistical monitoring plan, have been applied to SDTM database to detect potential issues (duplicate records, under-reporting of AEs or PROs, outliers, patients with similar characteristics…). An application has been developed using R-Shiny to generate an interactive web app to ease the identification of site and/or patient during the statistical data review meeting.

Results: The PRO-MSACTIVE study enrolled 422 patients by 46 sites between July 2018 and August 2019. The 3^rd data review meeting was held on the 17^th of october 2019 and 18 standard and planned tests were run on baseline characteristics and follow-up data, with a total of 15 (32.6%) sites identified as needing review/investigation.

Conclusion: Statistical monitoring is useful to identify unusual or clustered data patterns that might be revealing issues that could impact the data integrity and/or may potentially impact patient’s safety. With appropriate interactive data visualization, the important findings can easily be identified/reviewed by study team and appropriate actions be set up and assigned to the most appropriate function for a close follow-up. Interactive statistical monitoring is time consuming to initiate using R-Shiny, but time saving from the 1^st data review as long as analysis performed at each meeting remains similar.

Paolo Eusebi

Paolo Eusebi is a Consultant Statistician and Statistical Programmer. He is working as Contract Statistician for UCB. He is also adjunct Professor of Medical Statistics at the Epidemiology Department of Perugia University. His main research interests are the application of mixture models in meta-analysis, the use of machine learning techniques for subgroup identification and the display of uncertainty in scientific communication. He is statistical reviewer for Lancet Neurology and Associate Editor of BMC Medical Research Methodology. He is an active member of PSI and Data Visualization Society.

Effective visualization of uncertainty – Where we are and where to go

Statistics is all about uncertainty and uncertainty shows up in different ways in our research. If we want to answer questions about the development of the symptoms of an individual patient, we explore the overall distribution of the patients. For questions about the differences between e.g. treatment groups in studies, we are interested on the precision of the treatment effect. And as statisticians we face uncertainty by applying models with limited knowledge about the data generating processes.

The communication about uncertainty in its different forms plays a central role in statistics, yet it’s not well done in high profile medical journals. A random sample (n=50) was obtained from 777 RCTs papers published in BMJ, JAMA, Lancet and NEJM from November 2017 to October 2019.

Overall, uncertainty was not even considered in most of the plots. Those displaying uncertainty predominantly used whiskers or bands for confidence intervals. However, confidence intervals only poorly give a perception of a distribution.

The presentation will showcase different ways to better display uncertainty as inspired from other fields, including politics and weather forecast. These examples will cover also interactive and dynamic plots, which can enrich data visualization experience in electronic media.

Applications of these techniques for clinical trials and other medical data sources will be shown in addition to demonstrate how communication of uncertainty can be improved in medical statistics. These applications will make use of both frequentist and Bayesian approaches.

Click here to view the presentation.

Michael O’Kelly

With colleagues Michael O’Kelly has developed new methods for missing data that are now widely used in clinical trials. His book authored with Bohdana Ratitch, “Clinical trials with missing data: a guide for practitioners”, was published in 2014 by Wiley. He has given a PSI course on missing data, and co-authored a Best Practice proposal for projects involving Modelling and Simulation, which was adopted by the PSI board in 2017. He received the RSS/PSI award for Excellence in Pharmaceutical Statistics in 2017; he is Senior Director with IQVIA Advisory Services Analytics.

Subgroup analysis: a look at the SEAMOS approach (Standardised Effects Adjusted for Multiple Overlapping Subgroups)

In 2018, the PSI/EFSPI Working Group on Subgroup Analysis issued a White Paper in Pharmaceutical Statistics, which noted the usefulness of SEAMOS, a forest-plot based approach, as well as a number of other approaches. SEAMOS resamples from the data, using as its criterion the most extreme estimate of treatment effect, compared to the overall estimate of treatment effect. This presentation explores the SEAMOS approach further, using an idea due to PSI/EFSPI Working Group member Tom Parke, where the candidate subgroups of the forest plot are assessed collectively, using as a criterion the overlap of the multiple confidence intervals of standardised treatment effect estimates, where the measure of difference in subgroups is the confidence level required to preserve overlap of confidence intervals within a set of categories (e.g. preserve overlap between male and female; preserve overlap across regions). SEAMOS and its variants give rise to plots that may help in understanding the true significance of subgroup differences; furthermore, some multivariate parametric assessment of the “extremeness” of the subgroups in a forest plot is possible, as well as the resampling-based approach described in the source paper. This presentation looks at what can work best in practice, and when.

Click here to view the presentation.

Speaker

Biography

Abstract

Caroline Caudan

Caroline is a Statistician with 10 years of experience. She started her career working on the manufacturing and control part of the pharmaceutical industry in order to improve production and control processes.

For the last 4 years, she has been working as a biostatistician on the clinical part of the product development. Since she joined Keyrus Biopharma 3 years ago, she has been involved in the control and analysis of biomarker data, which led her to take a close interest to graphical representation and data visualization. Following this, Caroline started to work, in collaboration with Roche, on the development of a R Shiny application as a support to statistical monitoring.

Interactive statistical monitoring to optimize review of potential study issue with R-Shiny

Background: Statistical Monitoring involves the review of prospective study data collected in participating site to detect inconsistencies between patients and between sites in term of trends.

Method: A Phase IV study (PRO-MSACTIVE) is currently evaluating ocrelizumab in active relapsing Multiple Sclerosis patients in France. Specific statistical methods (volcano plots, mahalanobis distance, funnel plot …), described in a statistical monitoring plan, have been applied to SDTM database to detect potential issues (duplicate records, under-reporting of AEs or PROs, outliers, patients with similar characteristics…). An application has been developed using R-Shiny to generate an interactive web app to ease the identification of site and/or patient during the statistical data review meeting.

Results: The PRO-MSACTIVE study enrolled 422 patients by 46 sites between July 2018 and August 2019. The 3^rd data review meeting was held on the 17^th of october 2019 and 18 standard and planned tests were run on baseline characteristics and follow-up data, with a total of 15 (32.6%) sites identified as needing review/investigation.

Conclusion: Statistical monitoring is useful to identify unusual or clustered data patterns that might be revealing issues that could impact the data integrity and/or may potentially impact patient’s safety. With appropriate interactive data visualization, the important findings can easily be identified/reviewed by study team and appropriate actions be set up and assigned to the most appropriate function for a close follow-up. Interactive statistical monitoring is time consuming to initiate using R-Shiny, but time saving from the 1^st data review as long as analysis performed at each meeting remains similar.

Paolo Eusebi

Paolo Eusebi is a Consultant Statistician and Statistical Programmer. He is working as Contract Statistician for UCB. He is also adjunct Professor of Medical Statistics at the Epidemiology Department of Perugia University. His main research interests are the application of mixture models in meta-analysis, the use of machine learning techniques for subgroup identification and the display of uncertainty in scientific communication. He is statistical reviewer for Lancet Neurology and Associate Editor of BMC Medical Research Methodology. He is an active member of PSI and Data Visualization Society.

Effective visualization of uncertainty – Where we are and where to go

Statistics is all about uncertainty and uncertainty shows up in different ways in our research. If we want to answer questions about the development of the symptoms of an individual patient, we explore the overall distribution of the patients. For questions about the differences between e.g. treatment groups in studies, we are interested on the precision of the treatment effect. And as statisticians we face uncertainty by applying models with limited knowledge about the data generating processes.

The communication about uncertainty in its different forms plays a central role in statistics, yet it’s not well done in high profile medical journals. A random sample (n=50) was obtained from 777 RCTs papers published in BMJ, JAMA, Lancet and NEJM from November 2017 to October 2019.

Overall, uncertainty was not even considered in most of the plots. Those displaying uncertainty predominantly used whiskers or bands for confidence intervals. However, confidence intervals only poorly give a perception of a distribution.

The presentation will showcase different ways to better display uncertainty as inspired from other fields, including politics and weather forecast. These examples will cover also interactive and dynamic plots, which can enrich data visualization experience in electronic media.

Applications of these techniques for clinical trials and other medical data sources will be shown in addition to demonstrate how communication of uncertainty can be improved in medical statistics. These applications will make use of both frequentist and Bayesian approaches.

Click here to view the presentation.

Michael O’Kelly

With colleagues Michael O’Kelly has developed new methods for missing data that are now widely used in clinical trials. His book authored with Bohdana Ratitch, “Clinical trials with missing data: a guide for practitioners”, was published in 2014 by Wiley. He has given a PSI course on missing data, and co-authored a Best Practice proposal for projects involving Modelling and Simulation, which was adopted by the PSI board in 2017. He received the RSS/PSI award for Excellence in Pharmaceutical Statistics in 2017; he is Senior Director with IQVIA Advisory Services Analytics.

Subgroup analysis: a look at the SEAMOS approach (Standardised Effects Adjusted for Multiple Overlapping Subgroups)

In 2018, the PSI/EFSPI Working Group on Subgroup Analysis issued a White Paper in Pharmaceutical Statistics, which noted the usefulness of SEAMOS, a forest-plot based approach, as well as a number of other approaches. SEAMOS resamples from the data, using as its criterion the most extreme estimate of treatment effect, compared to the overall estimate of treatment effect. This presentation explores the SEAMOS approach further, using an idea due to PSI/EFSPI Working Group member Tom Parke, where the candidate subgroups of the forest plot are assessed collectively, using as a criterion the overlap of the multiple confidence intervals of standardised treatment effect estimates, where the measure of difference in subgroups is the confidence level required to preserve overlap of confidence intervals within a set of categories (e.g. preserve overlap between male and female; preserve overlap across regions). SEAMOS and its variants give rise to plots that may help in understanding the true significance of subgroup differences; furthermore, some multivariate parametric assessment of the “extremeness” of the subgroups in a forest plot is possible, as well as the resampling-based approach described in the source paper. This presentation looks at what can work best in practice, and when.

Click here to view the presentation.

Speaker

Biography

Abstract

Caroline Caudan

Caroline is a Statistician with 10 years of experience. She started her career working on the manufacturing and control part of the pharmaceutical industry in order to improve production and control processes.

For the last 4 years, she has been working as a biostatistician on the clinical part of the product development. Since she joined Keyrus Biopharma 3 years ago, she has been involved in the control and analysis of biomarker data, which led her to take a close interest to graphical representation and data visualization. Following this, Caroline started to work, in collaboration with Roche, on the development of a R Shiny application as a support to statistical monitoring.

Interactive statistical monitoring to optimize review of potential study issue with R-Shiny

Background: Statistical Monitoring involves the review of prospective study data collected in participating site to detect inconsistencies between patients and between sites in term of trends.

Method: A Phase IV study (PRO-MSACTIVE) is currently evaluating ocrelizumab in active relapsing Multiple Sclerosis patients in France. Specific statistical methods (volcano plots, mahalanobis distance, funnel plot …), described in a statistical monitoring plan, have been applied to SDTM database to detect potential issues (duplicate records, under-reporting of AEs or PROs, outliers, patients with similar characteristics…). An application has been developed using R-Shiny to generate an interactive web app to ease the identification of site and/or patient during the statistical data review meeting.

Results: The PRO-MSACTIVE study enrolled 422 patients by 46 sites between July 2018 and August 2019. The 3^rd data review meeting was held on the 17^th of october 2019 and 18 standard and planned tests were run on baseline characteristics and follow-up data, with a total of 15 (32.6%) sites identified as needing review/investigation.

Conclusion: Statistical monitoring is useful to identify unusual or clustered data patterns that might be revealing issues that could impact the data integrity and/or may potentially impact patient’s safety. With appropriate interactive data visualization, the important findings can easily be identified/reviewed by study team and appropriate actions be set up and assigned to the most appropriate function for a close follow-up. Interactive statistical monitoring is time consuming to initiate using R-Shiny, but time saving from the 1^st data review as long as analysis performed at each meeting remains similar.

Paolo Eusebi

Paolo Eusebi is a Consultant Statistician and Statistical Programmer. He is working as Contract Statistician for UCB. He is also adjunct Professor of Medical Statistics at the Epidemiology Department of Perugia University. His main research interests are the application of mixture models in meta-analysis, the use of machine learning techniques for subgroup identification and the display of uncertainty in scientific communication. He is statistical reviewer for Lancet Neurology and Associate Editor of BMC Medical Research Methodology. He is an active member of PSI and Data Visualization Society.

Effective visualization of uncertainty – Where we are and where to go

Statistics is all about uncertainty and uncertainty shows up in different ways in our research. If we want to answer questions about the development of the symptoms of an individual patient, we explore the overall distribution of the patients. For questions about the differences between e.g. treatment groups in studies, we are interested on the precision of the treatment effect. And as statisticians we face uncertainty by applying models with limited knowledge about the data generating processes.

The communication about uncertainty in its different forms plays a central role in statistics, yet it’s not well done in high profile medical journals. A random sample (n=50) was obtained from 777 RCTs papers published in BMJ, JAMA, Lancet and NEJM from November 2017 to October 2019.

Overall, uncertainty was not even considered in most of the plots. Those displaying uncertainty predominantly used whiskers or bands for confidence intervals. However, confidence intervals only poorly give a perception of a distribution.

The presentation will showcase different ways to better display uncertainty as inspired from other fields, including politics and weather forecast. These examples will cover also interactive and dynamic plots, which can enrich data visualization experience in electronic media.

Applications of these techniques for clinical trials and other medical data sources will be shown in addition to demonstrate how communication of uncertainty can be improved in medical statistics. These applications will make use of both frequentist and Bayesian approaches.

Click here to view the presentation.

Michael O’Kelly

With colleagues Michael O’Kelly has developed new methods for missing data that are now widely used in clinical trials. His book authored with Bohdana Ratitch, “Clinical trials with missing data: a guide for practitioners”, was published in 2014 by Wiley. He has given a PSI course on missing data, and co-authored a Best Practice proposal for projects involving Modelling and Simulation, which was adopted by the PSI board in 2017. He received the RSS/PSI award for Excellence in Pharmaceutical Statistics in 2017; he is Senior Director with IQVIA Advisory Services Analytics.

Subgroup analysis: a look at the SEAMOS approach (Standardised Effects Adjusted for Multiple Overlapping Subgroups)

In 2018, the PSI/EFSPI Working Group on Subgroup Analysis issued a White Paper in Pharmaceutical Statistics, which noted the usefulness of SEAMOS, a forest-plot based approach, as well as a number of other approaches. SEAMOS resamples from the data, using as its criterion the most extreme estimate of treatment effect, compared to the overall estimate of treatment effect. This presentation explores the SEAMOS approach further, using an idea due to PSI/EFSPI Working Group member Tom Parke, where the candidate subgroups of the forest plot are assessed collectively, using as a criterion the overlap of the multiple confidence intervals of standardised treatment effect estimates, where the measure of difference in subgroups is the confidence level required to preserve overlap of confidence intervals within a set of categories (e.g. preserve overlap between male and female; preserve overlap across regions). SEAMOS and its variants give rise to plots that may help in understanding the true significance of subgroup differences; furthermore, some multivariate parametric assessment of the “extremeness” of the subgroups in a forest plot is possible, as well as the resampling-based approach described in the source paper. This presentation looks at what can work best in practice, and when.

Click here to view the presentation.

Speaker

Biography

Abstract

Caroline Caudan

Caroline is a Statistician with 10 years of experience. She started her career working on the manufacturing and control part of the pharmaceutical industry in order to improve production and control processes.

For the last 4 years, she has been working as a biostatistician on the clinical part of the product development. Since she joined Keyrus Biopharma 3 years ago, she has been involved in the control and analysis of biomarker data, which led her to take a close interest to graphical representation and data visualization. Following this, Caroline started to work, in collaboration with Roche, on the development of a R Shiny application as a support to statistical monitoring.

Interactive statistical monitoring to optimize review of potential study issue with R-Shiny

Background: Statistical Monitoring involves the review of prospective study data collected in participating site to detect inconsistencies between patients and between sites in term of trends.

Method: A Phase IV study (PRO-MSACTIVE) is currently evaluating ocrelizumab in active relapsing Multiple Sclerosis patients in France. Specific statistical methods (volcano plots, mahalanobis distance, funnel plot …), described in a statistical monitoring plan, have been applied to SDTM database to detect potential issues (duplicate records, under-reporting of AEs or PROs, outliers, patients with similar characteristics…). An application has been developed using R-Shiny to generate an interactive web app to ease the identification of site and/or patient during the statistical data review meeting.

Results: The PRO-MSACTIVE study enrolled 422 patients by 46 sites between July 2018 and August 2019. The 3^rd data review meeting was held on the 17^th of october 2019 and 18 standard and planned tests were run on baseline characteristics and follow-up data, with a total of 15 (32.6%) sites identified as needing review/investigation.

Conclusion: Statistical monitoring is useful to identify unusual or clustered data patterns that might be revealing issues that could impact the data integrity and/or may potentially impact patient’s safety. With appropriate interactive data visualization, the important findings can easily be identified/reviewed by study team and appropriate actions be set up and assigned to the most appropriate function for a close follow-up. Interactive statistical monitoring is time consuming to initiate using R-Shiny, but time saving from the 1^st data review as long as analysis performed at each meeting remains similar.

Paolo Eusebi

Paolo Eusebi is a Consultant Statistician and Statistical Programmer. He is working as Contract Statistician for UCB. He is also adjunct Professor of Medical Statistics at the Epidemiology Department of Perugia University. His main research interests are the application of mixture models in meta-analysis, the use of machine learning techniques for subgroup identification and the display of uncertainty in scientific communication. He is statistical reviewer for Lancet Neurology and Associate Editor of BMC Medical Research Methodology. He is an active member of PSI and Data Visualization Society.

Effective visualization of uncertainty – Where we are and where to go

Statistics is all about uncertainty and uncertainty shows up in different ways in our research. If we want to answer questions about the development of the symptoms of an individual patient, we explore the overall distribution of the patients. For questions about the differences between e.g. treatment groups in studies, we are interested on the precision of the treatment effect. And as statisticians we face uncertainty by applying models with limited knowledge about the data generating processes.

The communication about uncertainty in its different forms plays a central role in statistics, yet it’s not well done in high profile medical journals. A random sample (n=50) was obtained from 777 RCTs papers published in BMJ, JAMA, Lancet and NEJM from November 2017 to October 2019.

Overall, uncertainty was not even considered in most of the plots. Those displaying uncertainty predominantly used whiskers or bands for confidence intervals. However, confidence intervals only poorly give a perception of a distribution.

The presentation will showcase different ways to better display uncertainty as inspired from other fields, including politics and weather forecast. These examples will cover also interactive and dynamic plots, which can enrich data visualization experience in electronic media.

Applications of these techniques for clinical trials and other medical data sources will be shown in addition to demonstrate how communication of uncertainty can be improved in medical statistics. These applications will make use of both frequentist and Bayesian approaches.

Click here to view the presentation.

Michael O’Kelly

With colleagues Michael O’Kelly has developed new methods for missing data that are now widely used in clinical trials. His book authored with Bohdana Ratitch, “Clinical trials with missing data: a guide for practitioners”, was published in 2014 by Wiley. He has given a PSI course on missing data, and co-authored a Best Practice proposal for projects involving Modelling and Simulation, which was adopted by the PSI board in 2017. He received the RSS/PSI award for Excellence in Pharmaceutical Statistics in 2017; he is Senior Director with IQVIA Advisory Services Analytics.

Subgroup analysis: a look at the SEAMOS approach (Standardised Effects Adjusted for Multiple Overlapping Subgroups)

In 2018, the PSI/EFSPI Working Group on Subgroup Analysis issued a White Paper in Pharmaceutical Statistics, which noted the usefulness of SEAMOS, a forest-plot based approach, as well as a number of other approaches. SEAMOS resamples from the data, using as its criterion the most extreme estimate of treatment effect, compared to the overall estimate of treatment effect. This presentation explores the SEAMOS approach further, using an idea due to PSI/EFSPI Working Group member Tom Parke, where the candidate subgroups of the forest plot are assessed collectively, using as a criterion the overlap of the multiple confidence intervals of standardised treatment effect estimates, where the measure of difference in subgroups is the confidence level required to preserve overlap of confidence intervals within a set of categories (e.g. preserve overlap between male and female; preserve overlap across regions). SEAMOS and its variants give rise to plots that may help in understanding the true significance of subgroup differences; furthermore, some multivariate parametric assessment of the “extremeness” of the subgroups in a forest plot is possible, as well as the resampling-based approach described in the source paper. This presentation looks at what can work best in practice, and when.

Click here to view the presentation.

Speaker

Biography

Abstract

Caroline Caudan

Caroline is a Statistician with 10 years of experience. She started her career working on the manufacturing and control part of the pharmaceutical industry in order to improve production and control processes.

For the last 4 years, she has been working as a biostatistician on the clinical part of the product development. Since she joined Keyrus Biopharma 3 years ago, she has been involved in the control and analysis of biomarker data, which led her to take a close interest to graphical representation and data visualization. Following this, Caroline started to work, in collaboration with Roche, on the development of a R Shiny application as a support to statistical monitoring.

Interactive statistical monitoring to optimize review of potential study issue with R-Shiny

Background: Statistical Monitoring involves the review of prospective study data collected in participating site to detect inconsistencies between patients and between sites in term of trends.

Method: A Phase IV study (PRO-MSACTIVE) is currently evaluating ocrelizumab in active relapsing Multiple Sclerosis patients in France. Specific statistical methods (volcano plots, mahalanobis distance, funnel plot …), described in a statistical monitoring plan, have been applied to SDTM database to detect potential issues (duplicate records, under-reporting of AEs or PROs, outliers, patients with similar characteristics…). An application has been developed using R-Shiny to generate an interactive web app to ease the identification of site and/or patient during the statistical data review meeting.

Results: The PRO-MSACTIVE study enrolled 422 patients by 46 sites between July 2018 and August 2019. The 3^rd data review meeting was held on the 17^th of october 2019 and 18 standard and planned tests were run on baseline characteristics and follow-up data, with a total of 15 (32.6%) sites identified as needing review/investigation.

Conclusion: Statistical monitoring is useful to identify unusual or clustered data patterns that might be revealing issues that could impact the data integrity and/or may potentially impact patient’s safety. With appropriate interactive data visualization, the important findings can easily be identified/reviewed by study team and appropriate actions be set up and assigned to the most appropriate function for a close follow-up. Interactive statistical monitoring is time consuming to initiate using R-Shiny, but time saving from the 1^st data review as long as analysis performed at each meeting remains similar.

Paolo Eusebi

Paolo Eusebi is a Consultant Statistician and Statistical Programmer. He is working as Contract Statistician for UCB. He is also adjunct Professor of Medical Statistics at the Epidemiology Department of Perugia University. His main research interests are the application of mixture models in meta-analysis, the use of machine learning techniques for subgroup identification and the display of uncertainty in scientific communication. He is statistical reviewer for Lancet Neurology and Associate Editor of BMC Medical Research Methodology. He is an active member of PSI and Data Visualization Society.

Effective visualization of uncertainty – Where we are and where to go

Statistics is all about uncertainty and uncertainty shows up in different ways in our research. If we want to answer questions about the development of the symptoms of an individual patient, we explore the overall distribution of the patients. For questions about the differences between e.g. treatment groups in studies, we are interested on the precision of the treatment effect. And as statisticians we face uncertainty by applying models with limited knowledge about the data generating processes.

The communication about uncertainty in its different forms plays a central role in statistics, yet it’s not well done in high profile medical journals. A random sample (n=50) was obtained from 777 RCTs papers published in BMJ, JAMA, Lancet and NEJM from November 2017 to October 2019.

Overall, uncertainty was not even considered in most of the plots. Those displaying uncertainty predominantly used whiskers or bands for confidence intervals. However, confidence intervals only poorly give a perception of a distribution.

The presentation will showcase different ways to better display uncertainty as inspired from other fields, including politics and weather forecast. These examples will cover also interactive and dynamic plots, which can enrich data visualization experience in electronic media.

Applications of these techniques for clinical trials and other medical data sources will be shown in addition to demonstrate how communication of uncertainty can be improved in medical statistics. These applications will make use of both frequentist and Bayesian approaches.

Click here to view the presentation.

Michael O’Kelly

With colleagues Michael O’Kelly has developed new methods for missing data that are now widely used in clinical trials. His book authored with Bohdana Ratitch, “Clinical trials with missing data: a guide for practitioners”, was published in 2014 by Wiley. He has given a PSI course on missing data, and co-authored a Best Practice proposal for projects involving Modelling and Simulation, which was adopted by the PSI board in 2017. He received the RSS/PSI award for Excellence in Pharmaceutical Statistics in 2017; he is Senior Director with IQVIA Advisory Services Analytics.

Subgroup analysis: a look at the SEAMOS approach (Standardised Effects Adjusted for Multiple Overlapping Subgroups)

In 2018, the PSI/EFSPI Working Group on Subgroup Analysis issued a White Paper in Pharmaceutical Statistics, which noted the usefulness of SEAMOS, a forest-plot based approach, as well as a number of other approaches. SEAMOS resamples from the data, using as its criterion the most extreme estimate of treatment effect, compared to the overall estimate of treatment effect. This presentation explores the SEAMOS approach further, using an idea due to PSI/EFSPI Working Group member Tom Parke, where the candidate subgroups of the forest plot are assessed collectively, using as a criterion the overlap of the multiple confidence intervals of standardised treatment effect estimates, where the measure of difference in subgroups is the confidence level required to preserve overlap of confidence intervals within a set of categories (e.g. preserve overlap between male and female; preserve overlap across regions). SEAMOS and its variants give rise to plots that may help in understanding the true significance of subgroup differences; furthermore, some multivariate parametric assessment of the “extremeness” of the subgroups in a forest plot is possible, as well as the resampling-based approach described in the source paper. This presentation looks at what can work best in practice, and when.

Click here to view the presentation.

Speaker

Biography

Abstract

Caroline Caudan

Caroline is a Statistician with 10 years of experience. She started her career working on the manufacturing and control part of the pharmaceutical industry in order to improve production and control processes.

For the last 4 years, she has been working as a biostatistician on the clinical part of the product development. Since she joined Keyrus Biopharma 3 years ago, she has been involved in the control and analysis of biomarker data, which led her to take a close interest to graphical representation and data visualization. Following this, Caroline started to work, in collaboration with Roche, on the development of a R Shiny application as a support to statistical monitoring.

Interactive statistical monitoring to optimize review of potential study issue with R-Shiny

Background: Statistical Monitoring involves the review of prospective study data collected in participating site to detect inconsistencies between patients and between sites in term of trends.

Method: A Phase IV study (PRO-MSACTIVE) is currently evaluating ocrelizumab in active relapsing Multiple Sclerosis patients in France. Specific statistical methods (volcano plots, mahalanobis distance, funnel plot …), described in a statistical monitoring plan, have been applied to SDTM database to detect potential issues (duplicate records, under-reporting of AEs or PROs, outliers, patients with similar characteristics…). An application has been developed using R-Shiny to generate an interactive web app to ease the identification of site and/or patient during the statistical data review meeting.

Results: The PRO-MSACTIVE study enrolled 422 patients by 46 sites between July 2018 and August 2019. The 3^rd data review meeting was held on the 17^th of october 2019 and 18 standard and planned tests were run on baseline characteristics and follow-up data, with a total of 15 (32.6%) sites identified as needing review/investigation.

Conclusion: Statistical monitoring is useful to identify unusual or clustered data patterns that might be revealing issues that could impact the data integrity and/or may potentially impact patient’s safety. With appropriate interactive data visualization, the important findings can easily be identified/reviewed by study team and appropriate actions be set up and assigned to the most appropriate function for a close follow-up. Interactive statistical monitoring is time consuming to initiate using R-Shiny, but time saving from the 1^st data review as long as analysis performed at each meeting remains similar.

Paolo Eusebi

Paolo Eusebi is a Consultant Statistician and Statistical Programmer. He is working as Contract Statistician for UCB. He is also adjunct Professor of Medical Statistics at the Epidemiology Department of Perugia University. His main research interests are the application of mixture models in meta-analysis, the use of machine learning techniques for subgroup identification and the display of uncertainty in scientific communication. He is statistical reviewer for Lancet Neurology and Associate Editor of BMC Medical Research Methodology. He is an active member of PSI and Data Visualization Society.

Effective visualization of uncertainty – Where we are and where to go

Statistics is all about uncertainty and uncertainty shows up in different ways in our research. If we want to answer questions about the development of the symptoms of an individual patient, we explore the overall distribution of the patients. For questions about the differences between e.g. treatment groups in studies, we are interested on the precision of the treatment effect. And as statisticians we face uncertainty by applying models with limited knowledge about the data generating processes.

The communication about uncertainty in its different forms plays a central role in statistics, yet it’s not well done in high profile medical journals. A random sample (n=50) was obtained from 777 RCTs papers published in BMJ, JAMA, Lancet and NEJM from November 2017 to October 2019.

Overall, uncertainty was not even considered in most of the plots. Those displaying uncertainty predominantly used whiskers or bands for confidence intervals. However, confidence intervals only poorly give a perception of a distribution.

The presentation will showcase different ways to better display uncertainty as inspired from other fields, including politics and weather forecast. These examples will cover also interactive and dynamic plots, which can enrich data visualization experience in electronic media.

Applications of these techniques for clinical trials and other medical data sources will be shown in addition to demonstrate how communication of uncertainty can be improved in medical statistics. These applications will make use of both frequentist and Bayesian approaches.

Click here to view the presentation.

Michael O’Kelly

With colleagues Michael O’Kelly has developed new methods for missing data that are now widely used in clinical trials. His book authored with Bohdana Ratitch, “Clinical trials with missing data: a guide for practitioners”, was published in 2014 by Wiley. He has given a PSI course on missing data, and co-authored a Best Practice proposal for projects involving Modelling and Simulation, which was adopted by the PSI board in 2017. He received the RSS/PSI award for Excellence in Pharmaceutical Statistics in 2017; he is Senior Director with IQVIA Advisory Services Analytics.

Subgroup analysis: a look at the SEAMOS approach (Standardised Effects Adjusted for Multiple Overlapping Subgroups)

In 2018, the PSI/EFSPI Working Group on Subgroup Analysis issued a White Paper in Pharmaceutical Statistics, which noted the usefulness of SEAMOS, a forest-plot based approach, as well as a number of other approaches. SEAMOS resamples from the data, using as its criterion the most extreme estimate of treatment effect, compared to the overall estimate of treatment effect. This presentation explores the SEAMOS approach further, using an idea due to PSI/EFSPI Working Group member Tom Parke, where the candidate subgroups of the forest plot are assessed collectively, using as a criterion the overlap of the multiple confidence intervals of standardised treatment effect estimates, where the measure of difference in subgroups is the confidence level required to preserve overlap of confidence intervals within a set of categories (e.g. preserve overlap between male and female; preserve overlap across regions). SEAMOS and its variants give rise to plots that may help in understanding the true significance of subgroup differences; furthermore, some multivariate parametric assessment of the “extremeness” of the subgroups in a forest plot is possible, as well as the resampling-based approach described in the source paper. This presentation looks at what can work best in practice, and when.

Click here to view the presentation.