Jeremy Howard's Blog, page 4

Avoiding the smoke - how to breath clean air

If you���re in western USA (like us) at the moment, you might be finding it hard to breath. Breathing air that contains the fallout from fires can make you feel pretty awful, and it can be bad for long-term health as well. Wildfire smoke contains fine particulate matter, known as ���PM2.5���, which can be inhaled deep into the lungs. The ���2.5��� here refers to the size of the particles ��� they are 2.5 microns or smaller. To see the air quality in your area, check out this AirNow map. Once it���s orange, you might find you start feeling the effects. If it���s red or purple, you almost certainly will. (Sometimes it can appear smokey outside, but the air quality can be OK, because the smoke might be higher in the atmosphere.)

The good news is that there���s a lot you can do to make the air you breathe a lot better. You might be wondering why a data scientist like me is commenting on air filtration��� The reason is that I was a leader of the Masks4All movement, including writing the first and most comprehensive scientific paper on the topic, which meant I studied filtration very closely for months. In fact, the size of particles we want to block for wildfires is very similar to the size of particles we want to block for covid-19!

Contents


Masks
Filtering your home air
Adding a filter to your central air
Adding filters to fans and portable A/C


Acknowledgements

The three ways that you can breathe cleaner air are to use a mask, filter your home central airconditioner or heater, and use fans with filters. I���ll show you the details below. (There���s quite a few links to places you can buy products in this post; I don���t get any commission or anything from them, they���re just things that I���ve personally found helpful.)

Masks

Therefore, you won���t be surprised to learn that one of the most effective things that you can do is to wear a mask. To block most PM2.5 particles you���ll want a mask that���s well-fitted and uses a good filter material. I���ve already prepared advice on that topic for COVID-19, and pretty much all of it is exactly the same for wildfire PM2.5, so go read this now. One bit that���s less of an issue is the ���Sanitation��� section ��� wildfire PM2.5 particles aren���t bearing disease, so you only have to worry about sanitation if your mask is actually getting dirty (or if you���ve been out in public with it on).

Personally, I like the O2 nano mask, or any well-fitted mask that you can insert a Filti filter in to. Recent aerosol science tests show that a neck gaiter folded to create two layers works well too (but make sure you add a nose clip to remove gaps around your nose). Check out Etsy for lots of mask designs that include a filter pocket and nose clip.

Choose from thousands of mask designs with a filter pocket



Filtering your home air

To clean the air in your home, the basic idea is to have it getting continually pushed through a filter. A filter is simply a piece of material which air can get through, but PM2.5 particles can���t. No filter is perfect, but there are readily-available options which work very well. Filters have a MERV rating, which tells you how many small particles they remove. For wildfire, you generally want MERV 13.

Don���t just buy the highest rating filter you can find. Filters with higher ratings have smaller holes (generally speaking), which means they also don���t let air through as fast. Remember, we want your home air going through the filter quickly, to ensure all your air is getting cleaned, so we don���t want the filter to negatively impact air-flow too much. I recommend Filtrete��� Healthy Living Air Filters. These have good air flow even for the MERV 13 spec.

Adding a filter to your central air

If you���ve got central heating or air conditioning, then you���re in luck. That will have strong fans, covering all of your rooms. The trick is to filter the air coming in to the system. Nearly all home systems simply pull their air in through a large vent inside your home. Some units have a filter slot in the unit itself, whereas for some the input vent is in a totally separate location in the house. Note that air conditioners blow air out to outside the house, but they don���t suck air in from outside the house (except, generally, for more fancy commercial building HVAC systems).

Once you���ve found the inlet vent that your central air is pulling in from, add a filter to it. If there���s already one there, make sure it���s MERV 13 or 14. You should change it every 3 months or so (depending on the brand). A vent with a filter installed looks like this:

An inlet vent, showing filter underneath

NB: Most filters have an arrow on the side showing the direction of airflow. So make sure you put it the right way around! Also, make sure you buy the right size. Measure the size of your vent, and buy a filter that is at least big enough to cover the hole. If there are gaps, the air will go through them, instead of your filter!

If there���s not a obvious place to add a filter to your vent, you���ll need to get creative. It might not look pretty, but you could always just remove the vent cover and fasten the filter straight over the top, using tape, poster tack, etc.

Once you���ve got your filter in place, the most important thing is to set your central air settings such that it has the fan running all the time. Most systems have an ���auto��� setting , which only turns the fan on when heating or cooling. You don���t want that! Set the fan to ���on���, not to ���auto���. That way, you���re getting as much air through that filter as possible.

Adding filters to fans and portable A/C

I recommend having an air purifier in every room. Most air purifiers don���t really do that much, because they���re normally quiet and small (which means they don���t move much air). There are extra large purifiers for sale, but they���re very expensive, and often sold out at the moment.

But we can create our own air purifier that works as well or better than the big expensive ones. An air purifier is simply a fan blowing air through a filter. So if we use a big fan and a good filter, then we have a good air purifier! The trick is to buy a 20 inch ���box fan��� (which is just a fan in a 20 inch square box), and stick a 20 inch filter in front of it. We pick 20 inches because that���s pretty big, and a bigger fan and bigger filter means more filtration can happen in a given time.

I bought a few of these box fans: PELONIS 3-Speed Box Fan. I���m not saying this one is any better or worse than any other ��� just buy whatever you can get your hands on. You want one that has a high speed setting, to push lots of air through.

For filters, anything of the right size and MERV 13 or 14 spec should be fine. I bought this pack of 6 20 inch Filtrete filters. Generally, higher quality filters will allow better air flow. Also thicker filters can increase airflow too; e.g. instead of the 20x20x1 filters I got, you could try 20x20x4 (4 inch thick) filters.

The fans I bought have the on/off/speed switch on the front, so I first turned that to the maximum speed setting, since once I attached the filter I couldn���t access the switch any more. Then I stuck some of this adhesive foam all the way around the front face of the fan, trying to leave no gaps. The idea is that when I then stick the fan on top of this, there will be as few gaps as possible. It would probably work just as well to stick a long piece of poster tack all around the front face. Finally, I stuck the filter to the front of the fan by using a generous quantity of high quality packing tape.

The completed DIY air purifier

These things are pretty noisy! But it���s a lot better than having a smoky house. They���re also pretty good for helping keep COVID-19 at bay, so if you have a shop or business, sprinkle a few of these around the place if you don���t have good filtered HVAC with a high change rate.

Another approach I���ve found useful is to buy a compact portable air conditioner. These come with a hose that blows hot air out through your window, and sucks air in through the front or back of the unit. You can stick a filter in front of where it sucks air in, using a similar approach to the fan discussed above.

Acknowledgements

Many thanks to Jim Rosenthal of Tex-Air Filters, and to Richard Corsi for the home-made air purifier idea. Jim has a fancier version for those with the budget. Thanks also to Jose-Luis Jimenez, Linsey Marr, Vladimir Zdimal, Adriaan Bax, and Kimberly Prather for many discussions that have helped me improve my (still limited!) understanding of aerosol science.

fast.ai releases new deep learning course, four libraries, and 600-page book

fast.ai is a self-funded research, software development, and teaching lab, focused on making deep learning more accessible. We make all of our software, research papers, and courses freely available with no ads. We pay all of our costs out of our own pockets, and take no grants or donations, so you can be sure we���re truly independent.

Today is fast.ai���s biggest day in our four year history. We are releasing:

fastai v2: A complete rewrite of fastai which is faster, easier, and more flexible, implementing new approaches to deep learning framework design, as discussed in the peer reviewed fastai academic paper
fastcore, fastscript, and fastgpu: Foundational libraries used in fastai v2, and useful for many programmers and data scientists
Practical Deep Learning for Coders (2020 course, part 1): Incorporating both an introduction to machine learning, and deep learning, and production and deployment of data products
Deep Learning for Coders with fastai and PyTorch: AI Applications Without a PhD: A book from O���Reilly, which covers the same material as the course (including the content planned for part 2 of the course)

Also, in case you missed it, earlier this week we released the Practical Data Ethics course, which focuses on topics that are both urgent and practical.

Contents


fastai v2
Practical Deep Learning for Coders, the course
Deep Learning for Coders with fastai and PyTorch, the book
fastcore, fastscript, and fastgpu
fastcore
fastscript
fastgpu


Acknowledgements


fastai v2

fastai is a deep learning library which provides practitioners with high-level components that can quickly and easily provide state-of-the-art results in standard deep learning domains, and provides researchers with low-level components that can be mixed and matched to build new approaches. It aims to do both things without substantial compromises in ease of use, flexibility, or performance. This is possible thanks to a carefully layered architecture, which expresses common underlying patterns of many deep learning and data processing techniques in terms of decoupled abstractions. These abstractions can be expressed concisely and clearly by leveraging the dynamism of the underlying Python language and the flexibility of the PyTorch library. fastai includes:

A new type dispatch system for Python along with a semantic type hierarchy for tensors
A GPU-optimized computer vision library which can be extended in pure Python
An optimizer which refactors out the common functionality of modern optimizers into two basic pieces, allowing optimization algorithms to be implemented in 45 lines of code
A novel 2-way callback system that can access any part of the data, model, or optimizer and change it at any point during training
A new data block API
And much more���





fastai's layered architecture

fastai is organized around two main design goals: to be approachable and rapidly productive, while also being deeply hackable and configurable. It is built on top of a hierarchy of lower-level APIs which provide composable building blocks. This way, a user wanting to rewrite part of the high-level API or add particular behavior to suit their needs does not have to learn how to use the lowest level.

To see what���s possible with fastai, take a look at the Quick Start, which shows how to use around 5 lines of code to build an image classifier, an image segmentation model, a text sentiment model, a recommendation system, and a tabular model. For each of the applications, the code is much the same.

Example of using fastai for image segmentation

Read through the Tutorials to learn how to train your own models on your own datasets. Use the navigation sidebar to look through the fastai documentation. Every class, function, and method is documented here. To learn about the design and motivation of the library, read the peer reviewed paper, or watch this presentation summarizing some of the key design points.

All fast.ai projects, including fastai, are built with nbdev, which is a full literate programming environment built on Jupyter Notebooks. That means that every piece of documentation can be accessed as interactive Jupyter notebooks, and every documentation page includes a link to open it directly on Google Colab to allow for experimentation and customization.

It���s very easy to migrate from plain PyTorch, Ignite, or any other PyTorch-based library, or even to use fastai in conjunction with other libraries. Generally, you���ll be able to use all your existing data processing code, but will be able to reduce the amount of code you require for training, and more easily take advantage of modern best practices. Here are migration guides from some popular libraries to help you on your way: Plain PyTorch; Ignite; Lightning; Catalyst. And because it���s easy to combine and part of the fastai framework with your existing code and libraries, you can just pick the bits you want. For instance, you could use fastai���s GPU-accelerated computer vision library, along with your own training loop.

fastai includes many modules that add functionality, generally through callbacks. Thanks to the flexible infrastructure, these all work together, so you can pick and choose what you need (and add your own), including: mixup and cutout augmentation, a uniquely flexible GAN training framework, a range of schedulers (many of which aren���t available in any other framework) including support for fine tuning following the approach described in ULMFiT, mixed precision, gradient accumulation, support for a range of logging frameworks like Tensorboard (with particularly strong support for Weights and Biases, as demonstrated here), medical imaging, and much more. Other functionality is added through the fastai ecosystem, such as support for HuggingFace Transformers (which can also be done manually, as shown in this tutorial), audio, accelerated inference, and so forth.

Medical imaging in fastai

There���s already some great learning material made available for fastai v2 by the community, such as the ���Zero to Hero��� series by Zach Mueller: part 1; part 2.

Practical Deep Learning for Coders, the course

Previous fast.ai courses have been studied by hundreds of thousands of students, from all walks of life, from all parts of the world. Many students have told us about how they���ve become multiple gold medal winners of international machine learning competitions, received offers from top companies, and having research papers published. For instance, Isaac Dimitrovsky told us that he had ���been playing around with ML for a couple of years without really grokking it��� [then] went through the fast.ai part 1 course late last year, and it clicked for me���. He went on to achieve first place in the prestigious international RA2-DREAM Challenge competition! He developed a multistage deep learning method for scoring radiographic hand and foot joint damage in rheumatoid arthritis, taking advantage of the fastai library.

This year���s course takes things even further. It incorporates both machine learning and deep learning in a single course, covering topics like random forests, gradient boosting, test and validation sets, and p values, which previously were in a separate machine learning course. In addition, production and deployment are also covered, including material on developing a web-based GUI for our own deep learning powered apps. The only prerequisite is high-school math, and a year of coding experience (preferably in Python). The course was recorded live, in conjunction with the Data Institute at the University of San Francisco.

After finishing this course you will know:

How to train models that achieve state-of-the-art results in:

Computer vision, including image classification (e.g.,classifying pet photos by breed), and image localization and detection (e.g.,finding where the animals in an image are)
Natural language processing (NLP), including document classification (e.g.,movie review sentiment analysis) and language modeling
Tabular data (e.g.,sales prediction) with categorical data, continuous data, and mixed data, including time series
Collaborative filtering (e.g.,movie recommendation)


How to turn your models into web applications, and deploy them
Why and how deep learning models work, and how to use that knowledge to improve the accuracy, speed, and reliability of your models
The latest deep learning techniques that really matter in practice
How to implement stochastic gradient descent and a complete training loop from scratch
How to think about the ethical implications of your work, to help ensure that you���re making the world a better place and that your work isn���t misused for harm

We care a lot about teaching, using a whole game approach. In this course, we start by showing how to use a complete, working, very usable, state-of-the-art deep learning network to solve real-world problems, using simple, expressive tools. And then we gradually dig deeper and deeper into understanding how those tools are made, and how the tools that make those tools are made, and so on. We always teach through examples. We ensure that there is a context and a purpose that you can understand intuitively, rather than starting with algebraic symbol manipulation. We also dive right into the details, showing you how to build all the components of a deep learning model from scratch, including discussing performance and optimization details.

The whole course can be completed for free without any installation, by taking advantage of the guides for the Colab and Gradient platforms, which provide free, GPU-powered Notebooks.

Deep Learning for Coders with fastai and PyTorch, the book

To understand what the new book is about, and who it���s for, let���s see what others have said about it��� Soumith Chintala, the co-creator of PyTorch, said in the foreword to Deep Learning for Coders with fastai and PyTorch:

But unlike me, Jeremy and Sylvain selflessly put a huge amount of energy into making sure others don���t have to take the painful path that they took. They built a great course called fast.ai that makes cutting-edge deep learning techniques accessible to people who know basic programming. It has graduated hundreds of thousands of eager learners who have become great practitioners.

In this book, which is another tireless product, Jeremy and Sylvain have constructed a magical journey through deep learning. They use simple words and introduce every concept. They bring cutting-edge deep learning and state-of-the-art research to you, yet make it very accessible.

You are taken through the latest advances in computer vision, dive into natural language processing, and learn some foundational math in a 500-page delightful ride. And the ride doesn���t stop at fun, as they take you through shipping your ideas to production. You can treat the fast.ai community, thousands of practitioners online, as your extended family, where individuals like you are available to talk and ideate small and big solutions, whatever the problem may be.

Peter Norvig, Director of Research at Google (and author of the definitive text on AI) said:

���Deep Learning is for everyone��� we see in Chapter 1, Section 1 of this book, and while other books may make similar claims, this book delivers on the claim. The authors have extensive knowledge of the field but are able to describe it in a way that is perfectly suited for a reader with experience in programming but not in machine learning. The book shows examples first, and only covers theory in the context of concrete examples. For most people, this is the best way to learn.The book does an impressive job of covering the key applications of deep learning in computer vision, natural language processing, and tabular data processing, but also covers key topics like data ethics that some other books miss. Altogether, this is one of the best sources for a programmer to become proficient in deep learning.

Curtis Langlotz, Director, Center for Artificial Intelligence in Medicine and Imaging at Stanford University said:

Gugger and Howard have created an ideal resource for anyone who has ever done even a little bit of coding. This book, and the fast.ai courses that go with it, simply and practically demystify deep learning using a hands on approach, with pre-written code that you can explore and re-use. No more slogging through theorems and proofs about abstract concepts. In Chapter 1 you will build your first deep learning model, and by the end of the book you will know how to read and understand the Methods section of any deep learning paper.

fastcore, fastscript, and fastgpu

fastcore

Python is a powerful, dynamic language. Rather than bake everything into the language, it lets the programmer customize it to make it work for them. fastcore uses this flexibility to add to Python features inspired by other languages we���ve loved, like multiple dispatch from Julia, mixins from Ruby, and currying, binding, and more from Haskell. It also adds some ���missing features��� and cleans up some rough edges in the Python standard library, such as simplifying parallel processing, and bringing ideas from NumPy over to Python���s list type.

fastcore contains many features. See the docs for all the details, which cover the modules provided:

test: Simple testing functions
foundation: Mixins, delegation, composition, and more
utils: Utility functions to help with functional-style programming, parallel processing, and more
dispatch: Multiple dispatch methods
transform: Pipelines of composed partially reversible transformations


fastscript

Sometimes, you want to create a quick script, either for yourself, or for others. But in Python, that involves a whole lot of boilerplate and ceremony, especially if you want to support command line arguments, provide help, and other niceties. You can use argparse for this purpose, which comes with Python, but it���s complex and verbose. fastscript makes life easier. In fact, this is a complete, working command-line application (no need for any of the usual boilerplate Python requires such as if __name__=='main'):

from fastscript import *
@call_parse
def main(msg:Param("The message", str),
upper:Param("Convert to uppercase?", bool_arg)=False):
print(msg.upper() if upper else msg)

When you run this script, you���ll see:

$ python examples/test_fastscript.py
usage: test_fastscript.py [-h] [--upper UPPER] msg
test_fastscript.py: error: the following arguments are required: msg


fastgpu

fastgpu provides a single command, fastgpu_poll, which polls a directory to check for scripts to run, and then runs them on the first available GPU. If no GPUs are available, it waits until one is. If more than one GPU is available, multiple scripts are run in parallel, one per GPU. It is the easiest way we���ve found to run ablation studies that take advantage of all of your GPUs, result in no parallel processing overhead, and require no manual intervention.

Acknowledgements

Many thanks to everyone who helped bring these projects to fruition, most especially to Sylvain Gugger, who worked closely with me over the last two years at fast.ai. Thanks also to all the support from the Data Institute at the University of San Francisco, and to Rachel Thomas, co-founder of fast.ai, who (amongst other things) taught the data ethics lesson and developed much of the data ethics material in the book. Thank you to everyone from the fast.ai community for all your wonderful contributions.

Forward from the 'Deep Learning for Coders' Book

To celebrate the release of fast.ai���s new course, book, and software libraries, we���re making available the foreword that Soumith Chintala (the co-creator of PyTorch) wrote for the book. To learn more, see the release announcement.

Your browser does not support the video element.

In a very short time, deep learning has become a widely useful technique, solving and automating problems in computer vision, robotics, healthcare, physics, biology, and beyond. One of the delightful things about deep learning is its relative simplicity. Powerful deep learning software has been built to make getting started fast and easy. In a few weeks, you can understand the basics and get comfortable with the techniques.

This opens up a world of creativity. You start applying it to problems that have data at hand, and you feel wonderful seeing a machine solving problems for you. However, you slowly feel yourself getting closer to a giant barrier. You built a deep learning model, but it doesn���t work as well as you had hoped. This is when you enter the next stage, finding and reading state-of-the-art research on deep learning.

However, there���s a voluminous body of knowledge on deep learning, with three decades of theory, techniques, and tooling behind it. As you read through some of this research, you realize that humans can explain simple things in really complicated ways. Scientists use words and mathematical notation in these papers that appear foreign, and no textbook or blog post seems to cover the necessary background that you need in accessible ways. Engineers and programmers assume you know how GPUs work and have knowledge about obscure tools.

This is when you wish you had a mentor or a friend that you could talk to. Someone who was in your shoes before, who knows the tooling and the math���someone who could guide you through the best research, state-of-the-art techniques, and advanced engineering, and make it comically simple. I was in your shoes a decade ago, when I was breaking into the field of machine learning. For years, I struggled to understand papers that had a little bit of math in them. I had good mentors around me, which helped me greatly, but it took me many years to get comfortable with machine learning and deep learning. That motivated me to coauthor PyTorch, a software framework to make deep learning accessible.

Jeremy Howard and Sylvain Gugger were also in your shoes. They wanted to learn and apply deep learning, without any previous formal training as ML scientists or engineers. Like me, Jeremy and Sylvain learned gradually over the years and eventually became experts and leaders. But unlike me, Jeremy and Sylvain selflessly put a huge amount of energy into making sure others don���t have to take the painful path that they took. They built a great course called fast.ai that makes cutting-edge deep learning techniques accessible to people who know basic programming. It has graduated hundreds of thousands of eager learners who have become great practitioners.

In this book, which is another tireless product, Jeremy and Sylvain have constructed a magical journey through deep learning. They use simple words and introduce every concept. They bring cutting-edge deep learning and state-of-the-art research to you, yet make it very accessible.

You are taken through the latest advances in computer vision, dive into natural language processing, and learn some foundational math in a 500-page delightful ride. And the ride doesn���t stop at fun, as they take you through shipping your ideas to production. You can treat the fast.ai community, thousands of practitioners online, as your extended family, where individuals like you are available to talk and ideate small and big solutions, whatever the problem may be.

I am very glad you���ve found this book, and I hope it inspires you to put deep learning to good use, regardless of the nature of the problem.

Soumith Chintala, co-creator of PyTorch

Applied Data Ethics, a new free course, is essential for all working in tech

Today we are releasing a free, online course on Applied Data Ethics, which contains essential knowledge for anyone working in data science or impacted by technology. The course focus is on topics that are both urgent and practical, causing real harm right now. In keeping with the fast.ai teaching philosophy, we will begin with two active, real-world areas (disinformation and bias) to provide context and motivation, before stepping back in Lesson 3 to dig into foundations of data ethics and practical tools. From there we will move on to additional subject areas: privacy & surveillance, the role of the Silicon Valley ecosystem (including metrics, venture growth, & hypergrowth), and algorithmic colonialism.

If you are ready to get started now, check out the syllabus and reading list or watch the videos here. Otherwise, read on for more details!

Issues related to data ethics make headlines daily, as real people are harmed by misuse

There are no prerequisites for the course. It is not intended to be exhaustive, but hopefully will provide useful context about how data misuse is impacting society, as well as practice in critical thinking skills and questions to ask. This class was originally taught in-person at the University of San Francisco Data Institute in January-February 2020, for a diverse mix of working professionals from a range of backgrounds (as an evening certificate courses).

About Data Ethics Syllabi

Data ethics covers an incredibly broad range of topics, many of which are urgent, making headlines daily, and causing harm to real people right now. A meta-analysis of over 100 syllabi on tech ethics, titled ���What do we teach when we teach tech ethics?��� found that there was huge variation in which topics are covered across tech ethics courses (law & policy, privacy & surveillance, philosophy, justice & human rights, environmental impact, civic responsibility, robots, disinformation, work & labor, design, cybersecurity, research ethics, and more��� far more than any one course could cover). These courses were taught by professors from a variety of fields. The area where there was more unity was in outcomes, with abilities to critique, spot issues, and make arguments being some of the most common desired outcomes for tech ethics course.

There is a ton of great research and writing on the topics covered in the course, and it was very tough for me to cut the reading list down to a ���reasonable��� length. There are many more fantastic articles, papers, essays, and books on these topics that are not included here. Check out my syllabus and reading list here.

A note about the fastai video browser

There is an icon near the top left of the video browser that opens up a menu of all the lesson. An icon near the top right opens up the course notes and a transcript search feature.

Use the icons on the top left and right of the video browser to collapse/expand a menu and course notes/transcript search



Topics covered

Lesson 1: Disinformation

From deepfakes being used to harass women, widespread misinformation about coronavirus (labeled an ���infodemic��� by the WHO), fears about the role disinformation could play in the 2020 election, and news of extensive foreign influence operations, disinformation is in the news frequently and is an urgent issue. It is also indicative of the complexity and interdisciplinary nature of so many data ethics issues: disinformation involves tech design choices, bad actors, human psychology, misaligned financial incentives, and more.

Watch the Lesson 1 video here.

Lesson 2: Bias & Fairness

Unjust bias is an increasingly discussed issue in machine learning and has even spawned its own field as the primary focus of Fairness, Accountability, and Transparency (FAccT). We will go beyond a surface-level discussion and cover questions of how fairness is defined, different types of bias, steps towards mitigating it, and complicating factors.

Watch the Lesson 2 video here.

Lesson 3: Ethical Foundations & Practical Tools

Now that we���ve seen a number of concrete, real world examples of ethical issues that arise with data, we will step back and learn about some ethical philosophies and lenses to evaluate ethics through, as well as considering how ethical questions are chosen. We will also cover the Markkula Center���s Tech Ethics Toolkit, a set of concrete practices to be implemented in the workplace.

Watch the Lesson 3 video here.

Lesson 4: Privacy and surveillance

Huge amounts of data are being collected about us: apps on our phones track our location, dating sites sell intimate details, facial recognition in schools records students, and police use large, unregulated databases of faces. Here, we discuss real-world examples of how our data is collected, sold, and used. There are also concerning patterns of how surveillance is used to suppress dissent and to further harm those who are already marginalized.

Watch the Lesson 4 video here.

Lesson 5: How did we get here? Our Ecosystem

News stories understandably often focus on one instance of a particular ethics issue at a particular company. Here, I want us to step back and consider some of the broader trends and factors that have resulted in the types of issues we are seeing. These include our over-emphasis on metrics, the inherent design of many of the platforms, venture capital���s focus on hypergrowth, and more.

Watch the Lesson 5 video here.

Lesson 6: Algorithmic Colonialism, and Next Steps

When corporations from one country develop and deploy technology in many other countries, extracting data and profits, often with little awareness of local cultural issues, a number of ethical issues can arise. Here we will explore algorithmic colonialism. We will also consider next steps for how students can continue to engage around data ethics and take what they���ve learned back to their workplaces.

Watch the Lesson 6 video here.

For the applied data ethics course, you can find the homepage here, the syllabus and reading list and watch the videos here.

Essential Work-From-Home Advice: Cheap and Easy Ergonomic Setups

You weren���t expecting to spend 2020 working from home. You can���t afford a fancy standing desk. You don���t have a home office, or even much spare space, in your apartment. Your neck is getting a permanent crick from hunching over your laptop on the couch. While those of us who are able to work from home are privileged to have this option, we still don���t want to permanently damage our backs, necks, or arms from a bad ergonomic setup.

This is not a post for ergonomic aficionados (the setups I share could all be further optimized). This is a post for folks who don���t know where to get started, have a limited budget, and are willing to try simple, scrappy approaches. Key takeway: for 34 dollars (21 for a good mouse, and 13 for a cheap keyboard), as well as some household items, you can create an ergonomic setup like the one below. I will show many other options throughout the post, for both sitting and standing, as well as approaches you can easily assemble/disassemble (if you are using the family dinner table and need to clear it off each evening).

While visiting family, I created an ergonomic setup on a counter



You can permanently damage your body with bad ergonomics

You can permanently damage your back, neck, and wrists from working without an ergonomic setup. Almost two decades ago, my partner Jeremy suffered from repetitive stress injury due to working without an ergonomic setup. At the time, his arms were paralyzed and he had to take months off from work. Even now and after years filled with good ergonomics and yoga, this still impacts his life, severely limiting how much time he can spend in cars or on planes, and creating painful flare-ups. Please take this issue seriously.

Key advice: Have a separate keyboard and mouse

The most important thing to know is that you want your screen approximately at eye height, and your elbows at approximately right angles to your torso as they type and use the mouse. This is the case whether you are sitting or standing. If you are using a laptop, this will be impossible with the built-in keyboard and trackpad (no matter how nice they are). It is essential to have a separate keyboard and mouse. If you only do one thing to address ergonomics, obtain a separate keyboard and mouse.

If you can���t afford an external monitor, no worries, you can just elevate your laptop. Over the years, I have used cardboard boxes, drinking glasses, bottles of soda, board games, and stacks of books to elevate my laptop. I will recommend some keyboards and mice that I like below, but anything is better than using the ones built into your laptop (since that forces you to keep your screen at the wrong height). For example, the picture in the intro is of a set-up I created while visiting a family member���s apartment in 2014, using books and a cardboard box to elevate my keyboard, mouse, and laptop to the appropriate heights.

For the deep learning study group, I routinely used a brown cardboard box. Bonus: I could store everything in the box when we had the clear out of that room each night.

Above is a picture from the deep learning study group, which meets 5 days a week, for 7 weeks, every time we run the deep learning course. I use a brown cardboard box to elevate my keyboard. We have to clear out of that conference room each evening, and it is simple for me to put my items in the box. This sort of solution could work if you don���t have a dedicated office space in your home, and need to be able to set up/take down your workstation regularly.

I rarely worked in coffee shops pre-pandemic (and never do now), but when I had to I would still try to create an ergonomic setup (and go to a coffeeshop where there was enough space!). Here, I���ve stacked my laptop on top of my rolled-up backpack. Ideally, my screen would be higher, but this is still better than having it at table level. Don���t let the perfect be the enemy of the good. Every step you take towards a more ergonomic setup is helpful.

When working at a coffee shop (pre-pandemic), I brought an external keyboard and mouse, and used my rolled-up backpack to raise the height of my laptop screen



About standing desks

If you have a regular desk (or even just a table) at home and want a standing desk, one option is to convert it using the $22 standing desk approach, which involves an Ikea side table and shelf. I had a previous job in which this was quite popular. Here is a photo of my work desk from that time.

In a previous job, many of us set up $22 standing desks using Ikea side tables

Standing on a hard floor can be difficult for your back. I have a GelPro mat, which I love. If you can���t afford a GelPro mat, standing on a folded-up yoga mat works great too.

Note that standing desks are not a cure-all. I���ve often seen people with expensive standing-desk converters (also known as desktop risers) that still have their monitor way too low. Even if you have an external monitor and desktop riser, makes sure your monitor is at an appropriate height. It is likely you will still need to stack it on top of something. If you don���t like the aesthetics of using books or other household items, you can buy a monitor stand, such as this one.

Using a standing desk with poor posture is not very ergonomic, so be cognizant of when you start feeling fatigued. I prefer to switch between standing and sitting throughout the day, as my energy fluctuates.

Budget Recommendations

My ���budget recommendation��� would be to get an Anker vertical mouse for $21 and literally any keyboard. If you have to choose, I���ve found that having a good mouse is way more important than a good keyboard. It is important that you get some keyboard though, so that you can elevate your laptop screen. In the setup below, I���m using a lightweight travel keyboard that isn���t particularly ergonomic, but it works fine.

The barista at this coffee shop kindly let me use 2 plastic tubs to prop up my laptop.

I realize that at a time when many Americans do not have enough to eat, that you may not have 34 dollars to spare (21 dollars for a mouse and 13 dollars for a cheap keyboard). However, if this is an option for you, it is well worth the cost. If you permanently damage your back, neck, or arms, no amount of money may be enough to heal them later.

Other products I like

My favorite mouse is the Logitech wireless trackball mouse. I have also used and liked the Anker vertical mouse. For keyboards, I like Goldtouch (I use an older version of this one) or the Microsoft Ergonomic Keyboard. And if you are looking for a compact, lightweight travel keyboard, I like the iClever foldup keyboard.

As mentioned above, GelPro mats are great if you are going to be standing, and a folded-up yoga mat is a cheaper alternative.

I have a Roost portable, lightweight laptop stand, which is great, although I can���t use it since I switched from a Macbook Air to a Microsoft Surface Pro. None of the links in this post are affiliate links; I���m just recommending what I���ve personally used and like.

For more about home office set-ups, Jeremy recently posted a twitter thread about his preferred computer set-up (which includes some pricier options). It���s also worth noting that his desk has a small footprint, and fits in the corner of our living room.

I couldn't be happier with my little standing desk setup. I have tried far to many products over the years, and here's what I highly recommend:
1/ pic.twitter.com/lMagQPLys1

— Jeremy Howard (@jeremyphoward) July 22, 2020

Cloth masks can protect the wearer

As we all know now, the science shows that DIY masks are particularly good at protecting those around you, in case you���re infected with COVID-19. But that doesn���t mean that you can���t do a lot to protect yourself too.

Unfortunately, many public health bodies still incorrectly claim that there is no evidence that DIY masks are useful at protecting the wearer. There���s actually plenty of evidence they can. Effective protection for the wearer of a mask depends on three critical things:

Material : does the mask filter particles of the appropriate sizes?
Fit : do particles squeeze in through the gaps of your mask?
Sanitation : can you clean and re-use the mask?

Let���s look at each in turn.

Material

The droplets that you need to filter out to protect yourself when wearing a mask are smaller than those that you have to filter out to protect those around you. That���s because they evaporate rapidly to become around 5x smaller after they���re ejected from your mouth. It���s unlikely that particles smaller than 1 micron can contain the virus, and particles can be up to 100 microns, so that���s the size range that ideally we want to filter. However, I haven���t seen any studies yet that look at that size range. Nearly all studies mainly look at much smaller particles, since that���s what the official NIOSH standard requires. The good news is that anything that does well on those tests will almost certainly do even better for larger particles, so here we���ll focus on NIOSH standard tests.

After looking at dozens of academic papers and websites, by far the best information I���ve found is in a table from maskfaq.com based on testing from TSI. I���ve extracted the best performing materials into the table below, sorted by quality, and color-coded by efficiency.

Table of filtration for highest quality DIY materials

A higher efficiency is better���it shows the percentage of particles that were filtered (remember, this is with much smaller particles at a much higher flow rate than we see in practice). A lower resistance is better���it is a measure of how hard it is to breath through. The ���Q��� column shows the filter quality factor, which combines efficiency and resistance. For materials with high Q but low efficiency, you can use more layers to increase the efficiency (although doubling the number of layers won���t necessarily mean doubling the efficiency).

Based on this table, the clear winner appears to be Filtrete 1900. It���s over 85% effective, and has an astonishingly low resistance. There are instructions available for creating masks with this material. One piece of filter material makes hundreds of mask filters, so you can get together with your community to make lots from a single order. However, be aware of three key issues:

It is not approved for use in masks by 3M. My guess is they just haven���t tested it and want to avoid liability; there isn���t any fiberglass or similar substances in it that might be problematic.
It can���t be washed, and I don���t know if there are other ways to sanitize it. However, it lasts three months as an air conditioner filter.
It might only filter small droplets, since it relies on electrostatic attraction for filtration. So it���s probably best to combine it with cotton, such as in a cloth mask with a filter pocket





Choose from hundreds of thousands of mask designs with a filter pocket

Personally, I prefer to use Filti, which is a nanofiber material specifically designed for face masks. It has the highest efficiency of any DIY material I���ve seen tested. You can buy masks with Filti pockets at Amazon or Etsy. Filti can be sanitized with heat and re-used. You can buy pre-cut material for 20 masks for around $20, which makes it better value if you���re not making many masks.

An even more economical option are shop towels. They���re not anywhere near as effective as Filti or Filtrete for wearer protection, but at around $20 for 200 towels, which you can fold to create two layers, you can stape rubber bands to them and give them away to anyone that needs them.

Fit

For wearer protection, fit is particularly important because, as you inhale, you will be sucking air (and floating particles in the air) straight through any gaps. The main places you are likely to have gaps are:

Around your nose
The sides of your mouth
At the bottom of your mask.

The bottom of your mouth is easy to handle: just make sure your mask is large enough to cover well past your chin, and is nice and wide at the bottom, and you should find that that creates a good seal where your chin is.

To fit around your nose properly, use a moldable nose piece. This is the thing that sits over the bridge of your nose and you mold to follow you face. The cheapest way to make one is to cut out a piece of aluminum foil, and fold it five times, to create a strip. You can see how in the video below.

Alternatively, you can use pipe cleaner, soft wire tie, or just buy adhesive nose strips.

To close the gaps at the sides of your mouth (and also helpful for your nose) you can use a mask brace. There are two great approaches explained at fixthemask.com. The first approach just uses three rubber bands, and is shown in the video above. This works well, and has been tested and shown to be capable of passing the NIOSH N95 fit test. However, it can be a bit awkward and uncomfortable, so I prefer the rubber sheet brace shown here:

A rubber mask brace

The only tricky bit is finding the material. I managed to find the correct type of rubber for around $20 from Amazon. One piece will make ten braces. I found that I could easily print the design on my printer and then use it as a stencil for cutting the rubber sheet with scissors. I���m not very crafty, so if I can do it, anyone can���

Another alternative to improving fit is adding a nylon stocking layer. I haven���t tried this myself, but researchers at Northeastern University have tested it and found it works well.

One tip that helps: get a larger mask with straps that tie all the way around the back, rather than just going over your ears. These can often have a much better fit. A thoroughly documented design with extensive tests is available at diymask.site.

If you have a 3D printer, there are some very thoughtful rigid designs in section IV of this paper, as well as some great fabric designs. If you have a heat sealer, there���s an excellent series of videos showing how to quickly create a mask that passes the N95 fit test. Many of these designs are available to purchase from hobbyists, crafters, and non-profits, often for no more than the cost of the materials. For instance, here is a rigid mask for just US$2.

Sanitation

For basic cloth masks, you can simply throw them in the wash. Anything involving soap will destroy the virus���s protective lipid layer. I believe that shop towels should be fine to wash too.

Most specialized filter material, including Filti, can���t be washed. Instead, put the filter material in a ziploc bag, and put it in a 160F oven for 30 minutes. (I asked one of the Stanford researchers that wrote these guidelines for tips on how to do it at home, and they suggested the ziploc bag trick.) I don���t know if Filtrete can handle these temperatures however, so you are probably best off simply disposing of Filtrete inserts when they���re dirty.

Generally you���ll probably be using specialized filter material as inserts in a cloth mask���s pocket. In these cases, take the insert out before you wash the cloth mask. If you forget, throw the insert away and get a new one ��� seriously, I mean it; Filti, for instance, loses about half its filtration after washing!

Recommendation

I suggest buying a large cloth mask with a Filti insert, moldable nose piece, and straps that go around the head, from Amazon or Etsy. When you need to sanitize it, put the cloth mask in the wash and sanitize the insert in the oven as described above.

If you use that kind of mask, or follow the other approaches described on this page, you should be able to achieve good protection when you go out. As well as wearing a mask, wear goggles or glasses (including sun glasses) too, since the virus can also enter through the eyes.

Particle sizes for mask filtration

Summary: SARS-CoV-2 particles do not float freely in the air. They are expelled as relatively large droplets, which research shows are easily caught by a simple cloth or paper mask. If an infected person doesn���t wear a mask, their droplets quickly evaporate into smaller droplet nuclei, which are harder to filter with a cloth mask. However there are some cloth mask designs which can do a very good job of this too.

I���ve seen a lot of confusion about the efficacy of mask filtration, and the impact of masks on re-breathing CO2. In each case, part of the problem is based on a failure to understand the relevant particles, and particle sizes. So let���s see if we can resolve some of the confusion!

Here are some basic parameters (all approximate measurements):

The SARS-CoV-2 virus particle is 100nm (nanometers) in diameter.
A CO2 molecule is 0.33nm diameter.
When we speak we produce droplets between 20 and 2000��m (micrometers) in diameter. Note that a micrometer is a thousand times larger than a nanometer!
Larger droplets fall to the ground fairly quickly. Smaller droplets evaporate in (at most) a few seconds to a droplet nuclei of around 1��m.
A 27��m droplet would carry 1 virion on average, and would evaporate to 5��m in a few seconds.
Small particles do not fly straight through materials, but instead follow brownian motion, resulting in them coming in contact with a material even when the material weave is larger than the particle.
Many materials, such as paper towel, have a complex weave which make it very difficult for particles to fully penetrate.
Materials like chiffon and silk also have electrostatic effects that result in charge transfer with nanoscale aerosol particles, making them particularly effective (considering their sheerness) at excluding particles in the nanoscale regime (<���100 nm).

So, the first thing to note is that CO2 is going to flow through any mask without any trouble. There is no known mask material that will filter 0.33nm molecules. If it did, you wouldn���t be able to breathe at all!

The size of the virus particle itself is not relevant to any discussion of mask filtration. This is because virus particles never float freely in the air, but are always at least suspended in a droplet nuclei ten times larger than the virus itself. A droplet containing a single particle will on average start out 270 times larger than the virion, and will evaporate to nuclei of 50 times larger than the virion.

The size of the weave of the fabric is also not directly comparable to the size of the droplets or droplet nuclei, due to the three dimensional nature of many types of material, the indirect route taken by small particles in brownian motion, and the electrostatic effects in many materials. So if you���ve seen those claims that masks can���t possibly stop COVID-19 because the virus is too small, now you know why they���re totally wrong.

A popular meme created by someone that doesn't understand aerosol science

Therefore, the only way to really understand the efficacy of a mask is to actually test it in practice. Because the size of droplets that are ejected are much larger than those that remain in the air (due to evaporation), testing must be done separately for source control (protecting others from the wearer) vs PPE (protecting the wearer from others).

Source control efficacy

There are two main ways to physically test a mask:

Have someone wearing it breathe, talk, cough, and so forth, or
Synthetically simulate these actions using a spray mechanism, such as a nebulizer.

Because actual human actions are complex and hard to simulate correctly, the first approach is preferred where possible. Generally, we are most interested in speech droplets, because people that are coughing and sneezing should stay home, so those are of less importance to community transmission, and breathing is not believed to contain significant concentrations of SARS-CoV-2 particles.

There is a study that looked at the protective effect of a simple cloth mask for speech droplet source control, that found that approximately 99% of the forward-facing droplets visible in a laser chamber were blocked. The cloth mask in the study was moist, in order to avoid dust contamination of the equipment; a followup experiment from the same group, pictured here (but not published yet), found that a dry paper towel had the same results.

There are no studies that have directly measured the filtration of smaller or lateral particles in this setting, although using Schlieren imaging it has been shown that all kinds of masks greatly limit the spread of the droplet cloud, consistent with a fluid dynamic simulation that estimated this filtration level at 90%.

Fluid dynamics simulation of droplet cloud with vs without mask

Another approach to studying source control efficacy tested viral shedding in respiratory droplet samples and aerosol samples. In this study, seasonal coronavirus was tested, which is in the same genus as SARS-CoV-2. Cloth masks were not tested; neither were speech droplets ��� only breathing and coughing were studied. An unfitted surgical mask was 100% effective at blocking coronavirus particles.

In a pair of studies from 50 years ago, a portable isolation box, provided with a filtered air supply and a means of access for a test subject���s head, was attached to an Andersen Sampler and used to measure orally expelled bacterial contaminants before and after masking. In one of the studies, during talking, unmasked subjects expelled more than 5,000 bacterial contaminants per 5 cubic feet; 7.2% of the contaminants were associated with particles less than 4��m in diameter. Masked subjects (using a cotton muslin and flannel blend) expelled an average of 19 contaminants per 5 cubic feet; 63% were less than 4��m in diameter. So overall, over 99% of contaminants were filtered. The second study used the same experimental setup, but studied a wider range of mask designs, including a 4-ply cotton mask. For each mask design, over 97% contaminant filtration was observed.

PPE efficacy

Protection of the wearer (PPE) is much more challenging that source control, since, as discussed, the particles are much smaller (although not as small as a free virion). It���s also much harder to directly test mask efficacy for PPE using a human subject, so instead simulations must be used. There are two considerations when looking at efficacy:

The filtration of the material
The fit of the design.

There are many standards around the world for both of these issues, such as the U.S. National Institute for Occupational Safety and Health (NIOSH) N95 classification. The ���N95��� designation means that when subjected to careful testing, the respirator blocks at least 95 percent of very small (0.3 micron) test particles. These are much smaller than virus-carrying droplets or droplet nuclei, which means that masks that do not meet this standard may be effective as PPE in the community.

One recent study looked at the aerosol filtration efficiency of common fabrics used in respiratory cloth masks, finding that efficacy varied widely, from 12% to 99.9%. Underlining the importance of fit for specialized medical masks, an unfitted N95 respirator had the worst efficacy. Many materials had >=96% filtration efficacy for particles >0.3 microns, including 600 TPI cotton, cotton quilt, and cotton layered with chiffon, silk, or flannel. These findings support studies reported in 1924 by Wu Lien Teh, which described that a silk face covering with flannel added over the mouth and nose was highly effective against pneumonic plague.

Many studies use very small aerosol particles at very high flow rates, such as a study that was used as the basis for a table in WHO���s Advice on the use of masks in the context of COVID-19. In this study, tiny 78nm aerosol particles were blasted through cloth at a rate of 95 liters per minute. Only N95 and equivalent masks were able to stand up to this torrent of aerosol, which would never be seen in practice in any normal community setting. The machines used for these studies are specifically designed for looking at masks that hold their shape (respirators) which are glued or attached with beeswax firmly to the testing plate. Flexible masks such as cloth and surgical masks can get pulled into the hole in the testing plate.

There are many designs of cloth masks, with widely varying levels of fit. There have been few tests of different designs. One study looked at unfitted surgical masks, and used three rubber bands and a paper clip to improve their fit. All eleven subjects in the test passed the N95 fit test using this approach. A simple mask cut from a t shirt achieved a fit score of 67, not up to the 100 level required for N95, but this mask offered substantial protection from the challenge aerosol and showed good fit with minimal leakage. Wu Lien Teh noted that a rubber support could provide good fit, although he recommended that a silk covering for the whole head (and flannel sewed over nose and mouth areas), with holes for the eyes, tucked into the shirt, is a more comfortable approach that can provide good protection for a whole day.

Overall, it appears that cloth face covers can provide good fit and filtration for PPE, but results will vary a lot depending on material and design.

Conclusion

Overall, there is evidence that simple cloth face masks will generally provide good protection to those around the wearer (source control). This is possible because droplets expelled during speech are much larger than the droplet nuclei they later turn into through evaporation.

There are also some combinations of material and design which can provide good protection to the wearer as well (PPE). However, many cloth masks do not have the materials or design necessary to achieve the highest level of protection.

Introducing the first cohort of USF CADE Data Ethics Research Fellows

The University of San Francisco is welcoming three Data Ethics research fellows (one started in January, and the other two are beginning this month) for year-long, full-time fellowships. We are so excited to have them join our community. They bring expertise in an interdisciplinary range of fields, inlcuding bioethics, public policy, anthropology, computer science, data privacy, and political philosophy. We had many fantastic applicants for the program, and we wish we had been able to offer a larger number of fellowships. We hope to be able to expand this program in the future. Without further ado, here is our first cohort of data ethics research fellows: Ali Alkhatib, Razvan Amironesei, and Nana Young.

from left to right: Ali Alkhatib, Razvan Amironesei, Nana Young



Ali Alkhatib

Ali Alkhatib is a social computing researcher trained in Computer Science and Anthropology. His research explores how people relate to artificial intelligence and technology broadly, and attempts to situate those relationships in historical backdrops and ontological foundations using scholarship from the social sciences.

His paper Street-Level Algorithms: A Theory at the Gaps Between Policy and Decisions won the best paper award in 2019 at CHI Conference on Human Factors in Computing Systems, the premier international conference of Human-Computer Interaction, and his paper Examining Crowd Work and Gig Work Through The Historical Lens of Piecework won honorable mention at CHI 2017. Ali wrote the powerful essay Anthropological/Artificial Intelligence & the HAI.

Ali studied Computer Science at Stanford for a few years while pursuing a PhD with Michael Bernstein as his advisor. He earned my B.A. in Anthropology & B.S. in Informatics, specializing in human-computer interaction, both from UC Irvine in 2014, with an honors thesis on the Culture of Quantified Self, working under Tom Boellstorff.

Razvan Amironesei

Razvan Amironesei, PhD, was most recently a visiting scholar in the Department of Philosophy at the University of California, San Diego, where he chairs a multicampus faculty research group on algorithms and politics. He conducts interdisciplinary research (1) on the genealogy of datasets in collaboration with Google researchers by showing the constitution of algorithmic bias and its relation to harm as a historical, ethical, and technical problem and (2) on specific issues related to privacy practices of data related to human rights and questions regarding cybersecurity in the tech sector.

Over his past 8 years with UCSD, Razvan has written and received three grants that he used to organize events on the political and ethical dimensions of algorithms at UC San Diego, UC Berkeley, and UCLA. His Ph.D. dissertation in philosophy was devoted to the relationship between biopower and the concept of life, where he engaged with a sociological and theoretical analysis of Human-computer interaction technologies, in particular the question of brain surveillance. In his MA, he worked on questions around surveillance and privacy via a historical analysis of disciplinary technologies. Razvan has previously taught many undergraduate and graduate level courses in political philosophy and ethics including: ���Ethics and Healthcare,��� ���The Ethics of Human Cloning,��� and ���Politics, Power, Violence.���

Nana Young

Nana Young is a global health bioethicist with domestic and international experience conducting independent health disparities research in low- and middle-income settings. Her research interests include ethical implications of disruptive technology, cyberharms and vulnerable populations, algorithmic justice, and harnessing the power of artificial intelligence to drive ethical, sustainable development in low and middle-income countries.

Nana Young earned her MA in Bioethics & Science Policy at Duke University, with a thesis on ���When Private Bodies Deliver Public Goods: Why the Expectation of Private Altruism to Substitute for State Public Good Delivery is a Desertion of Government Responsibility that Places the Poor at Heightened Risk.��� While at Duke, she helped design a course on race, genomics, emerging technologies, and society. She earned a BA in sociology at Princeton, with a thesis on a ���Qualitative Study of the Socio-Cultural Sources of Mental Illness Stigma in Ghana, West Africa.���

Previously, Nana worked on strategic initiatives at a non-profit to shape strategic engagement with industry, civil society, academic and government actors towards the promotion and implementation of policy and health systems solutions of pressing global health issues including NCDs, disease epidemics, climate change, tobacco control, mental health, maternal mortality, and more.

Please join me in welcoming these data ethics research fellows to the University of San Francisco Center for Applied Data Ethics!

Masks - FAQ for Skeptics

A bit of skepticism is healthy, and it���s especially reasonable given how much the official guidance on masks has varied over time and across regions. But of course, a good skeptic reads the evidence, and makes an informed decision based on that. So here���s some frequently asked questions I���ve been seeing from curious skeptics, and answers (with citations).

Contents


Why should most people wear masks?
Shouldn���t only sick people wear masks?
Shouldn���t we just follow WHO���s guidelines?
Is there a randomized controlled trial (RCT) for the impact of masks on community transmission of respiratory infections in a pandemic?
Shouldn���t we wait for an RCT before doing something?
Aren���t there RCTs that show no effect of mask usage?
Doesn���t a mask need to be 100% effective to be useful?
Do we really know if the virus is transmitted through the air?
If it���s spread through the air, can a cloth mask really stop it? Isn���t the virus too small?
I heard a doctor say that masks don���t help. Is that true?
Won���t wearing masks make people just be less careful about physical distancing?
Mightn���t people handle their masks wrong and make things worse?
What if people touch their face more and infect themselves in the process?
Where am I going to get a mask anyway?
Won���t this make people take masks away from healthcare workers?
What about the article ���Masks-for-all for COVID-19 not based on sound data���?
Isn���t wearing a mask a personal choice?
Mightn���t wearing a mask cause people of color to get harassed?
Isn���t wearing a mask something that only Asian cultures do?


Why should most people wear masks?

Wearing a mask decreases the number of people infected by an infectious mask wearer (���source control���), because it reduces by around 99% the number of droplets that are ejected during speech. It also probably somewhat decreases the likelihood of an uninfected wearer getting infected, although it���s less effective for this, since many of the droplets quickly evaporate into small droplet nuclei that are hard to block. Reducing the number of people infected has an exponential impact, because it decreases the effective reproduction rate, R.

About half of infections are from people that aren���t showing symptoms ��� so people that don���t know they���re sick are infecting others. Because masks are far more effective at blocking infection at the source, that means we need everyone to wear a mask in public, since otherwise unmasked people put those around them at risk.

Shouldn���t only sick people wear masks?

Patients without symptoms pose a risk of infecting others, so it���s not enough to wait until you have symptoms to wear a mask. Four recent studies show that nearly half of patients are infected by people who do not themselves have symptoms���thus they aren���t even coughing or sneezing yet, but they can spread the disease just by talking in close proximity to someone else.

Shouldn���t we just follow WHO���s guidelines?

WHO says ���if you are healthy, you only need to wear a mask if you are taking care of a person with COVID-19���. WHO also says that ���Studies of influenza, influenza-like illness, and human coronaviruses provide evidence that the use of a medical mask can prevent the spread of infectious droplets from an infected person to someone else and potential contamination of the environment by these droplets.��� Remember, you don���t know if you���re healthy, and you don���t know if the people that you���re with are healthy either. So to follow WHO���s guidelines, you really need to be wearing a mask when around others.

Many countries have been clear about this. The U.S. CDC (Center for Disease Control) recommends wearing cloth face coverings in public settings��� because ���a significant portion of individuals with coronavirus lack symptoms��� and they can be contagious spreaders of the virus. Other countries that are officially recommending mask use include China, Japan, France, India, South Korea, Canada, Germany, Brazil, Spain, Indonesia, Israel, the Czech Republic, Singapore, South Africa, Slovenia, Bulgaria, Slovakia, Austria, Bosnia, Mongolia, Taiwan, Colombia, Philippines, Ukraine, Uzbekistan, Vietnam, Cuba, Turkey, Chile, Zambia, Rwanda, Luxembourg, Panama, Malaysia, Poland, Ecuador, Singapore, Morocco, Kenya, Venezuela, Rwanda, Nigeria, Ethiopia, Guinea, Honduras, Hong Kong, Bulgaria, Benin, Cyprus.

Many countries have gone further, and mandated mask use in most public settings, including Indonesia, Israel, the Czech Republic, Slovenia, Bulgaria, Slovakia, Austria, Bosnia, Mongolia, Taiwan, Singapore, Colombia, Poland, Panama, Philippines, Uzbekistan, Ukraine, Vietnam, Cuba, Morocco, Turkey, Kenya, Zambia, Luxembourg, Ecuador, Chile, Venezuela, Honduras, Ethiopia, Rwanda, Benin, Guinea, Parts of China, and Parts of USA (including New York, New Jersey, Maryland, Pennsylvania, Connecticut, Puerto Rico, Los Angeles, Miami, Washington DC, San Antonio, Most of Hawaii, and San Francisco).

Hopefully, WHO will update their guidelines to be clearer in the future. Their most recent guidelines say that ���WHO is collaborating with research and development partners to better understand the effectiveness and efficiency of nonmedical masks. WHO is also strongly encouraging countries that issue recommendations for the use of masks in healthy people in the community to conduct research on this critical topic. WHO will update its guidance when new evidence becomes available.���

Is there a randomized controlled trial (RCT) for the impact of masks on community transmission of respiratory infections in a pandemic?

A randomized controlled trial (RCT) is sometimes considered the ���gold standard��� for assessing evidence to see whether a medical intervention actually works. It���s mainly used for assessing new drugs. In an RCT, a representative sample is selected, and randomly split into two groups, one of which receives the medical intervention (e.g. the drug), and one which doesn���t (normally that one gets a placebo). This can, when things go well, show clearly whether the drug made a difference. Generally, a ���p value��� is calculated, which is the probability that effect seen in the data would be observed by chance. If that p value is less than some number (often 0.05) the RCT is considered to be ���statistically significant���. Without an RCT, it can be harder to distinguish whether two groups differ because of the intervention, or because of some other difference between the groups.

There has never been, and will never be, an RCT for the impact of masks, or hand-washing, or social distancing on community transmission of respiratory infections in a pandemic. The reason is that the following steps would be needed:

Select 100 or so communities that are representative, and do not have significant population interaction (i.e people don���t move from one region to another)
Select at random 50 communities where everyone must all wear masks in public; the populations of the other 50 must never wear masks in public
Wait a few months
See how many people died in each set of communities

Because we have such a strong prior expectation that masks are likely to be effective, there are probably no jurisdictions where it would be considered ethical to run such a study. It would also be very challenging to ensure compliance. A smaller and simpler trial that does not look at whole communities, but only individuals, would face similar ethical problems, and would also not be able to actually answer the question of whether community transmission is impacted.

The American Statistical Association (ASA) has released a ���Statement on Statistical Significance and P-Values��� with six principles underlying the proper use and interpretation of the p-value. In particular, note the following principles:

P-values do not measure the probability that the studied hypothesis is true, or the probability that the data were produced by random chance alone.
Scientific conclusions and business or policy decisions should not be based only on whether a p-value passes a specific threshold.
A p-value, or statistical significance, does not measure the size of an effect or the importance of a result.

So what should policy decisions be based on? They should be based on an assessment of the potential upsides and downside of an intervention along with their probabilities, versus the potential cost, in order to come up with an approximate expected value (ideally, a probability distribution) if the intervention is used, versus not used.

Shouldn���t we wait for an RCT before doing something?

No. Even if we ignore the impossibility of running such an RCT, a British Medical Journal paper points out that ���there is a moral argument that the public should be given the opportunity to change their behavior in line with the precautionary principle, even when direct experimental evidence for benefit is not clear cut���. The precautionary principle is (from Wikipedia) ���a strategy for approaching issues of potential harm when extensive scientific knowledge on the matter is lacking.��� Most nations have agreed, via UNICEF, to act in compliance with this principle.

No jurisdiction has waited for an RCT before recommending hand washing or physical distancing. Many jurisdictions have enforced extreme physical distancing through mandated lockdowns or shelter-in-place orders, despite the lack of an RCT showing their effectiveness at reducing community transmission of COVID-19, and despite this being a far more expensive intervention.

Aren���t there RCTs that show no effect of mask usage?

For an RCT to show no effect, we would need to observe two groups that are very similar, with enough data to make us confident that the effect is small enough to be not practically useful. There are no RCTs that have found this for any use of masks for any type of coronavirus.

The closest thing we have, perhaps, to a relevant RCT is the paper The First Randomized, Controlled Clinical Trial of Mask Use in Households to Prevent Respiratory Virus Transmission: This was an Australian study for influenza control in the community, but not during a pandemic, and without any enforcement of compliance (such as would be provided by a mask mandate). It stated that ���observational epidemiologic data suggest that transmission of viral respiratory infection was significantly reduced during the SARS epidemic with the use of face masks as well as other infection control measures��� and ���in an adjusted analysis of compliant subjects, masks as a group had protective efficacy in excess of 80% against clinical influenza-like illness.��� However, the authors noted that ���we found compliance to be low, but compliance is affected by perception of risk. In a pandemic, we would expect compliance to improve. In compliant users, masks were highly efficacious.���

There is some evidence that basic masks are more effective for coronavirus than influenza. There is a lot of evidence that compliance with mask wearing can be high during the COVID-19 pandemic, since many communities are already at well over 80% compliance (including many at close to 100%, due to mask mandates).

Some other RCTs that are frequently discussed with regards to public mask use, none of which study impact on community transmission, include:

Cluster randomised controlled trial to examine medical mask use as source control for people with respiratory illness: in this study, the relative risk for laboratory confirmed viral infections was 0.06 to 15.54. In other words, the study was too underpowered to tell whether the impact of masks was extremely effective, or extremely ineffective. The study looked at the impact of continuously wearing a mask whilst in a hospital. This is a very different situation to using a mask for short periods of time whilst shopping, in transit, and so forth.
Facemasks and hand hygiene to prevent influenza transmission in households: a cluster randomized trial: This was a Hong Kong study for influenza control in the community. It was an underpowered study, however it concluded ���Hand hygiene and facemasks seemed to prevent household transmission of influenza virus when implemented within 36 hours of index patient symptom onset. These findings suggest that nonpharmaceutical interventions are important for mitigation of pandemic and interpandemic influenza.���
A cluster randomised trial of cloth masks compared with medical masks in healthcare workers: This did not test the impact of wearing masks on source control and did not have a control group that did not wear masks. The study found that cloth masks were not effective as PPE for rhinovirus. This is consistent with a more recent study that found that simple masks were not effective for blocking rhinovirus; however, it found them 100% effective at blocking seasonal coronavirus.


Doesn���t a mask need to be 100% effective to be useful?

No. No mask is 100% effective. However, fewer virus particles mean a better chance of avoiding infection, and research shows that, if you are infected, the lower your viral exposure load, the better your chance of only a mild illness.

Do we really know if the virus is transmitted through the air?

I claimed earlier that a mask works ���because it reduces by around 99% the amount of droplets that are ejected during speech���. That���s only useful if we know that the virus is actually transmitted through the air. We have to be careful here, because many people (including the WHO) use a definition of ���airborne��� that does not include transmission caused by droplets traveling through the air, but only tiny particles that float freely in the air for hours. We don���t know for sure whether the virus is ���airborne��� under this definition. However, it does seem clear that a key transmission route of COVID-19 is via droplets that fly out of our mouths. It has been known since 1934 (and studied in hundreds of papers since) that respiratory infections are transmitted through these droplets, and that the smaller ones quickly evaporate. We���ve known since 1946 (and studied in hundreds of papers since) that this creates tiny particles that are extremely hard to stop. This mode of respiratory infection is well understood, and consistent with the transmission of SARS.

Unfortunately, WHO makes things rather confusing, through their publication Modes of transmission of virus causing COVID-19. This document claims that ���According to current evidence, COVID-19 virus is primarily transmitted between people through respiratory droplets and contact routes.��� Five references provided to support this assertion (the WHO page currently lists 6, but it appears that the reference numbers are off by one, since the following reference does not relate to the correct document). However, a review of the references shows that none of them provide evidence supporting any particular transmission route. These are the references provided:

Community Transmission of Severe Acute Respiratory Syndrome Coronavirus 2, Shenzhen, China, 2020: ���We suspect that community transmission and intrafamily transmission have potentially become the new transmission modes in the city. Also, nosocomial infection and transmission might pose a potential risk for COVID-19 control.��� No evidence of whether transmission was through droplets, Aerosol, or fomites is presented.
A familial cluster of pneumonia associated with the 2019 novel coronavirus indicating person-to-person transmission: The focus of this study was to identify whether the key issue to analyze and control is super-spreader events, or whether other types of transmission could be a problem. ���Our study showed that person-to-person transmission in family homes or hospitals, and intercity spread of this novel coronavirus are possible���. No analysis of transmission methods was done, however there is some evidence here that wearing a mask is effective, because the only person that was not infected in the family in this case was the only one to wear a mask.
Early Transmission Dynamics in Wuhan, China, of Novel Coronavirus���Infected Pneumonia: This study only looked at ���infected persons, relatives, close contacts, and health care workers���. It did not attempt to identify the method of transmission or find out whether transmission to more remote people occurred.
Clinical features of patients infected with 2019 novel coronavirus in Wuhan, China: this study looks at ���the epidemiological, clinical, laboratory, and radiological characteristics and treatment and clinical outcomes��� of patients. It does not look at or comment on transmission method in any way.
Active Monitoring of Persons Exposed to Patients with Confirmed COVID-19: ���despite intensive follow-up, no sustained person-to-person transmission of symptomatic SARS-CoV-2 was observed in the United States among the close contacts of the first 10 persons with diagnosed travel-related COVID-19. Analyses of timing of exposure during each patient���s illness as well as the type and duration of exposures will provide information on potential risk factors for transmission.��� This analysis was not done in this paper, so therefore it does not provide any information about transmission methods.

However, there are cases reported in the literature which support the conclusion that the virus is transmitted through the air. In particular, the paper COVID-19 Outbreak Associated with Air Conditioning in Restaurant, Guangzhou, China, 2020, notes that ���Virus transmission in this outbreak cannot be explained by droplet transmission alone. Larger respiratory droplets (>5 ��m) remain in the air for only a short time and travel only short distances, generally <1 m���. Indirect Virus Transmission in Cluster of COVID-19 Cases, Wenzhou, China, 2020 notes that ���the rapid spread of SARS-CoV-2 in our study could have resulted from spread via fomites (e.g., elevator buttons or restroom taps) or virus aerosolization in a confined public space (e.g., restrooms or elevators). All case-patients other than those on floor 7 were female, including a restroom cleaner, so common restroom use could have been the infection source. For case-patients who were customers in the shopping mall but did not report using the restroom, the source of infection could have been the elevators. The Guangzhou Center for Disease Control and Prevention detected the nucleic acid of SARS-CoV-2 on a doorknob at a patient���s house (5), but Wenzhou Center for Disease Control and Prevention test results for an environmental sample from the surface of a mall elevator wall and button were negative���. Other cases not yet reported in the scientific literature include a Seattle choir rehearsal where 45 have been diagnosed with COVID-19 or ill with the symptoms, despite all attendees required to use hand sanitizer, and a conference in Boston where 77 people became infected. Locations that had sporting events and festivals, where there tends to be singing and shouting, which eject larger and more droplets further distances, have had large COVID-19 outbreaks.

Perhaps the most interesting case for understanding transmission routes is a call center building in Guro-gu, Seoul, There are a total of 163 confirmed cases since 8 March. Of the 163 confirmed cases, 97 are persons who worked in the building (11th floor = 94; 10th floor = 2; 9th floor = 1), and 66 are their contacts.��� So nearly all the cases in the building were on a single floor. This strongly suggests that transmission must be mainly through the air, otherwise, if transmission was primarily through touching surfaces, elevator buttons and front doors etc would have caused substantial transmission to other floors.

I have only been able to find a single case that seems likely to be caused by infection through surface contact. In this case, a person in Singapore who sat in a church pew that was earlier used by an infected patient became infected themselves. It���s possible that the virus was transmitted through the pew, although it���s also possible that since the people went to the same church, they came into contact before or after the service.

If it���s spread through the air, can a cloth mask really stop it? Isn���t the virus too small?

Coronavirus particles are so small that they can fit through the weave of most household cloth materials. Medical masks, such as N95 respirators, use special materials that create difficult to navigate pathways in the fabric that make it very hard for these tiny particles to get through the material. They also are specially fitted to the face of each healthcare worker to minimize gaps that these particles can get through.

Many commentators have been distracted by this, not realizing that the droplets that are ejected from an infected mask wearer are far bigger than the virus particles, and are easily blocked with around 99% efficacy, as shown in this recent NEJM paper that used laser light scattering to explore the effect. (The paper includes videos that make it easy to see for yourself what���s going on.)

We don���t know for sure yet whether droplets ejected during breathing are also an important transmission path. These droplets are much smaller, and have a lower total viral load, compared to droplets from speech, but I haven���t found any studies that directly study the impact of this on COVID-19 transmission.

The good news, however, is that we do have a study that shows the impact of wearing an unfitted mask on seasonal coronavirus transmission, based on the amount of virus particles found in droplets ejected during breathing. In this study, the unfitted mask was 100% effective in blocking these for seasonal coronavirus.

There���s even a study that tested the efficacy of a cloth mask at blocking COVID-19. Unfortunately, there are some problems with the study:

Symptomatic people should stay home, so testing speech would have been more helpful than testing coughs
The test was done at 8 inches, which is much closer than people should be if following physical distancing guidelines
Only 4 patients were tested
The data analysis was not done correctly.

We can���t fix the first three problems, but we can fix the fourth. When we do, we find that over 95% of the viral load was blocked by the cloth masks.

To improve protection for the wearer, a coffee filter or paper towel can be easily inserted into the mask to improve filtering. For instance, the Hong Kong Consumer Council recommended design includes a paper towel, after scientists ���scanned kitchen paper towel under electronic microscope and revealed that the fabric size, gaps and layout of kitchen paper towel are similar to that of the middle layer of surgical mask���.

I heard a doctor say that masks don���t help. Is that true?

No. In fact, many of the top doctors around the world are speaking up and telling the public to use masks, including an open letter from 100 doctors in Britain, Canada���s chief public health officer, and the US Surgeon General. The official position of the peak medical bodies in 11 of the top 12 (by GDP) countries is that masks can be used to reduce transmission (the outlier being the UK, which is expected to change its advice in the coming days).

Some doctors are taking longer to change their ways. Unfortunately, Western doctors don���t have a great track record of accepting the science of public health hygiene. The scientist who discovered the importance of hand-washing, Ignaz Semmelweis, was mocked and ignored by doctors at the time, and for decades afterwards. Before introducing hand washing, ���puerpual fever��� was killing hundreds of mothers a year in his unit. Afterwards, says Carl Zimmer, he ���brought the death rate pretty much to zero. I mean, he couldn���t completely eliminate it, but he got pretty close. There were some months where like no women died at all. None.���

In the same way, Western doctors didn���t believe the young Malaysian-Chinese doctor, Wu Lien-teh, who realized that the 1910 Manchurian Plague was transmitted through the air, and that a simple cotton mask could reduce transmission. The podcast 99% Invisible explains: ���Wu believed that all of his medical staff, as well as the general public, should wear masks, but other doctors wouldn���t listen to him possibly because of his young age and race. One French doctor named G��rald Mesny openly antagonized him and refused to wear his mask. Soon after, Mesny caught the plague and died vindicating Wu���s theory. After Mesny���s death, everyone began wearing Wu���s mask. People began photographing it and it became a symbol of medical success and the plague ended after 7 months. The government implemented a lot of epidemic practices that we still see today���wearing masks, quarantining patients, and cutting off travel to limit exposure.���

Won���t wearing masks make people just be less careful about physical distancing?

There is no evidence that mask use reduces compliance with other recommended strategies, such as physical distancing. Anecdotal evidence suggests that wearing masks is a useful reminder of the gravity of the situation, and may remind others to keep their distance. Historically, public health initiatives such as seat belts, condoms, and motorbike helmets are generally associated with concerns about negative outcomes due to increasing risky behavior, but overall population results have not borne out these concerns in practice.

As our paper explains (section B.2; see paper for citation details):

���One concern around public health messaging promoting the use of face-covering has been that members of the public may use risk compensation behavior and neglect physical distancing based on overvaluing the protection a surgical mask may offer due to an exaggerated or false sense of security (49). Similar arguments have previously been made for HIV prevention strategies (50) (51) and other safety devices and mandates such as motorcycle helmet laws (52) and seat-belts (53). However, research on these topics finds no such increase in adverse outcomes at the population level but rather improvements in safety and well-being, suggesting that even if risk compensation occurs in some individuals, that effect is dwarfed by the increased safety at the population level (53, 54). Further, even for deliberately high-risk recreational activities such as alpine skiing and snowboarding, wearing a helmet was generally associated with risk reduction oriented behavior (55), suggesting safety devices are both compatible with and perhaps encourage safety-oriented behavior. Even for high-risk recreational activities like alpine skiing and snowboarding, helmet use has greatly reduced injury rates (56).

In general, various forms of risk compensation theories have been proposed for many different safety innovations, but have not been found to have empirical support (57) at the population level. These findings strongly suggest that, instead of withholding a preventative tool, accompanying it with accurate messaging that combines different preventative measures would display trust in the general public���s ability to act responsibly and empower citizens, and risk compensation is unlikely to undo the positive benefits at the population level (58).���

Mightn���t people handle their masks wrong and make things worse?

As discussed, it appears that transmission of COVID-19 through surfaces is very rare. There are no reported cases that I���ve found that show transmission through an infected mask. Since there are now hundreds of millions of people around the world required to wear masks in public, we would expect to have seen examples of this by now.

The idea that wearing masks could increase risk due to touching it doesn���t stand up to scrutiny. If your mask has virus particles in it, there���s two possibilities for how they got there:

You���re already infected, in which case this isn���t really an issue, or
You���re not already infected, and the virus particles came from someone else, which means that your mask stopped them from going into your mouth.

COVID-19 is transmitted through the inside of the mouth, nose, or the eyes. If a mask stops virus particles from entering your mouth, then it���s done its job. People should be told to wash their mask when they get home, to minimize the chance that they get infected through an infected surface.

What if people touch their face more and infect themselves in the process?

As discussed, COVID-19 is transmitted through the inside of the mouth, nose, or the eyes. A mask covers the mouth and nose, making it much harder to accidentally touch them. People should be encouraged to avoid touching their mouth and nose regardless of whether or not they are wearing a mask.

Where am I going to get a mask anyway?

Masks can be made by cutting the ends off a sock, stapling rubber bands to a piece of kitchen towel, cutting the arms of a t-shirt, folding a handkerchief over hair-ties or rubber bands, by using a scarf or bandana, and so forth. There is no evidence to suggest that any of these masks are not effective at blocking droplets from an infected person.

Won���t this make people take masks away from healthcare workers?

Simple homemade cloth masks made of cut up cotton t shirts, paper towels, a handkerchief, etc are very effective for source control, so there���s no need to take medical masks away from healthcare workers. In regions that have mandated mask usage, most people are wearing DIY masks, not medical masks.

What about the article ���Masks-for-all for COVID-19 not based on sound data���?

On April 1st, a retired professor, Lisa Brosseau, and Margaret Sietsema, an Assistant Professor of Environmental and Occupational Health Sciences, wrote an online commentary titled Masks-for-all for COVID-19 not based on sound data. The article is full of uncited claims, falsehoods, and misunderstandings, and wouldn���t normally be something that would be taken seriously and need to be discussed in any detail. However, it has been pushed heavily though social media and medical mailing lists. Therefore I���ll look at its claims here in detail.

The first section ���Data lacking to recommend broad mask use��� claims that ���Sweeping mask recommendations��� will not reduce SARS-CoV-2 transmission, as evidenced by the widespread practice of wearing such masks in Hubei province���. No references or data are provided to back up this claim. Looking at the actual data, however, shows that it supports the opposite conclusion - that masks may have been critical in controlling the Hubei outbreak. A report from Guo Yi in HK01 pointed out that up until Jan 22 most people in Wuhan were not wearing a mask. The next day, the government started requiring masks in public. Wuhan had their peak number of cases on Feb 4, and since then case numbers have been decreasing. Clearly, the evidence here does not support the conclusion that masks were ineffective. Such a broad claim as ���mask recommendations will not reduce transmission���, made without any caveats, on the basis of a single location, made without data or references, where the actual data shows the opposite of the claimed result, suggests that this piece of writing may not have been carefully researched or reviewed.

The next claim made is that ���Our review of relevant studies indicates that cloth masks will be ineffective at preventing SARS-CoV-2 transmission, whether worn as source control or as PPE���. This statement is made without any citations. The article actually presents no studies that provide evidence that cloth masks will be ineffective at preventing SARS-CoV-2 transmission.

The next claim made is ���respirators, though, are the only option that can ensure protection for frontline workers dealing with COVID-19 cases���. This is incorrect. Respirators can not ensure protection. In fact, nothing can ensure protection. However, it would be accurate to say that respirators provide the best practical protection for frontline workers. However this bears no relation to the topic of their article, which purports to be about ���masks for all���, not protection for frontline workers. Unfortunately, there is a shortage of respirators at present, which is why people are looking at what other options might be useful, rather than making idealistic gestures about would is best. (This point is mentioned in the article.) Furthermore, many respirators have a valve which makes them useless for source control, and therefore greatly reduces their effectiveness at reducing transmission.

The next section is titled ���Filter efficiency and fit are key for masks, respirators���. However, this is only accurate when a mask is used for protecting the wearer (PPE) rather than those around the wearer (source control). This section makes many uncited claims, which I won���t bother discussing since no data or research is provided to support them. Focusing instead on those claims with references, various studies are presented which look at how much salt and aerosol particles flow through various fabrics at various pressures and particle sizes. There is no evidence that these simulations have any relationship to actual COVID-19 transmission. The same is true of the studies of fit that are presented. The claims are entirely unrelated to efficacy for source control, since droplets do not evaporate into droplet nuclei before hitting the cloth in the mask, due to the humid environment created by the mask. In practice, a cloth mask has around 99% efficacy at blocking droplets.

The most important section is next: ���We found no well-designed studies of cloth masks as source control in household or healthcare settings���. Even if the failure of the authors was due to a lack of studies, rather than a research failure on their part, this does not support their contention that ���cloth masks will be ineffective at preventing SARS-CoV-2 transmission���. Indeed, their own research here shows that they do not know whether cloth masks will be effective. In the next paragraph they make a similar error, claiming ���may also have very limited utility as source control or PPE in households���, despite not having any data or research to support this claim. In fact, as we���ve seen, the research actually suggests that both cloth masks and surgical masks might be highly effective.

The next section ���Surgical masks as source control��� claims ���Household studies find very limited effectiveness of surgical masks at reducing respiratory illness in other household members���, and cites 4 references. However, none of the references make this claim or show data to support this contention. Their reference 22, for instance, is a meta analysis in which section 3.4 lists the results of each analysis they looked at, and concludes ���If the randomized control trial and cohort study were pooled with the case���control studies, heterogeneity decreased and a significant protective effect was found���. However, most of the studies were underpowered and were unable to distinguish between large, small, or negative protection. There was no study that found limited effectiveness (note that ���not significant��� is a statistical measure related to the amount of data in the study���it doesn���t imply that effectiveness was limited). Reference 23 finds ���There is some evidence to support the wearing of masks or respirators during illness to protect others���. Reference 24 again finds many under-powered studies, but finds ���Eight of nine retrospective observational studies found that mask and/or respirator use was independently associated with a reduced risk of severe acute respiratory syndrome���. Reference 25 does not relate to household use, despite their use of it as a reference here. The authors concludes ���In sum, wearing surgical masks in households appears to have very little impact on transmission of respiratory disease���, however, none of the references provided support this claim.

In the section ���Cloth masks as PPE���, the authors claim ���A randomized trial comparing the effect of medical and cloth masks on healthcare worker illness found that those wearing cloth masks were 13 times more likely to experience influenza-like illness than those wearing medical masks.��� However, this assertion is incorrect. The setting was actually 90% rhinovirus, which it has been found is ineffective for filtering with cloth masks. However, COVID-19 is not rhinovirus, and unlike rhinovirus is actually filtered effectively with cloth. In addition, the study referenced did not just compare surgical masks with cloth masks, but compared a regular supply of 2 new medical masks per day, with just 5 masks for a 4 week period. This is clearly inappropriate in the hot, busy, healthcare setting that was studied here.

Isn���t wearing a mask a personal choice?

The Republican governor of Maryland, Larry Hogan, said ���Some people have said that covering their faces infringes on their rights, but this isn���t just about your rights or protecting yourself; it���s about protecting your neighbors. And the best science that we have shows that people might not know that they���re carriers of the virus, through no fault of their own, and they could infect other people. Spreading this disease infringes on your neighbors��� rights.���

Making a ���personal choice��� to not wear a mask can put those around you at risk.

(Some people genuinely can���t safely wear a mask, of course, but that is a separate issue.)

Mightn���t wearing a mask cause people of color to get harassed?

Mandating universal mask wearing, rather than just recommending mask use, may have additional benefits such as reducing stigma. From our paper (section B.2; see paper for citation details):

For many infectious diseases, including, for example, tuberculosis, health authorities recommend masks only for those infected or people who are taking care of someone infected. However, research shows that many sick people are reluctant to wear a mask if it identifies them as sick, and thus end up not wearing them at all in an effort to avoid the stigma of illness (60, 61). Stigma is a powerful force in human societies, and many illnesses come with stigma for the sick as well as fear of them, and managing the stigma is an important part of the process of controlling epidemics as stigma also leads to people avoiding treatment as well as preventive measures that would ���out��� their illness (62). Many health authorities have recommended wearing masks for COVID-19 only if people are sick; however, reports of people wearing masks being attacked, shunned and stigmatized have already been observed (63). Having masks worn only by the suspected/confirmed infected also has led to employers in high-risk environments like grocery stores and prisons, and even hospitals, banning employees from wearing one sometimes with the idea that it would scare the customer or the patients (64, 65). Further, in many countries, minorities suffer additional stigma and assumptions of criminality (66). In that vein, black people in the United States have reported that they were reluctant to wear masks in public during this pandemic for fear of being mistaken as criminals (67, 68).

Isn���t wearing a mask something that only Asian cultures do?

Many Western regions have now mandated mask wearing in many public places, including many parts of the USA and many countries in Europe. There is no sign that Westerners are unable or unwilling to wear masks.

Masks for all? The science says yes.

*Update: Jeremy has now written an article about masks in The Conversation lengthy FAQ. Trisha and Jeremy are two of over hundred of the world���s top academics who released an open letter to all U.S. governors asking that ���officials require cloth masks to be worn in all public places, such as stores, transportation systems, and public buildings.��� *

Confused about mask wearing? Sure, it���s complicated. But not as complicated as some people imply. We���ve been looking at the science (see our papers Face Masks Against COVID-19: An Evidence Review ��� with 104 references! ��� and Masks for the public: laying straw men to rest). Here���s a summary of the different streams of evidence, and our take on what it all means.

Translations

We���d love your help translating this article! Please at-mention @jeremyphoward on Twitter with your translation, and I���ll retweet it and add it here.

Italian (published in the journal Evidence)
Spanish (from @tyoc213)
German (from Andr�� Calero Valdez)
French (from GitHub user Bnech)
Greek (from @a8inea)
Swedish (from Christoffer Bj��rkskog)
Dutch (from David Smeijer)
Finnish (from Thomas Brand)
Portuguese (from Leandro Gomide)
Romanian (from Emilian Bold)
Turkish (from Zikri Bayraktar)
Estonian (from Kaur Maran)
Hungarian (from Ago Lajko)
Chinese (Simplified) (from @HFProjectOrg)
Chinese (Traditional) (from @HFProjectOrg)
Japanese (from perapera.ai)
Indonesian (from Lailaturrahmi)
Hindi (from GitHub user alephthoughts)
Nepalese (from Vidya Sagar Gurung)
Hebrew (from Motti Haimi)
Arabic (from Ali Alsahlanee)


The epidemiology of disease spread

You���ve probably seen the videos of closely-packed dominos and mousetraps, where a single item fires off a huge cascade. The closer the dominos (or mousetraps), the more chaos gets generated. Every infectious disease has a transmission rate (R0). A disease with an R0 of 1.0 means that every infected person, on average, infects one other person. A disease whose R0 is less than 1.0 will die out. The strain of flu that caused the 1918 pandemic had an R0 of 1.8. The R0 of the virus which causes COVID-19 was estimated at 2.4 by Imperial College researchers, although some research suggests it could be as high as 5.7. This means that without containment measures, COVID-19 will spread far and fast. Importantly, COVID-19 patients are most infectious in the early days of the disease (To et al. 2020; Zou et al. 2020; Bai et al. 2020; Zhang et al. 2020; Doremalen et al. 2020; Wei 2020), during which they generally have few or no symptoms.

The physics of droplets and aerosols

When you speak, tiny micro droplets are ejected from your mouth. If you���re infectious, these contain virus particles. Only the very largest droplets end up surviving more than 0.1 s before drying out and turning into droplet nuclei (Wells 1934; Duguid 1946; Morawska et al. 2009) that are 3-5 times smaller than the original droplet itself, but still contain some virus.

That means that it���s much easier to block droplets just as they come out of your mouth, when they���re much larger, compared to blocking them as they approach the face of a non-infected person who is on the receiving end of those droplets. But this isn���t what most researchers have been looking at���

The material science of masks

Debates about the effectiveness of masks often assume that the purpose of the mask is to protect the wearer, since this is what all doctors learn about in medical school. Cloth masks are relatively poor (though not entirely ineffective) at this. For 100% protection, the wearer needs a properly fitted medical respirator (such as an N95). But cloth masks, worn by an infected person are highly effective at protecting the people around them. This is known as ���source control���. And it is source control that matters in the debate about whether the public should wear masks.

If you have COVID-19 and cough on someone from 8 inches away, wearing a cotton mask will reduce the amount of virus you transmit to that person by over 90%, and is even more effective than a surgical mask. The researchers who did this experiment described the reduction as ���ineffective���, partially based on an inappropriate analysis in which the patients where the cotton masks were perfectly effective were deleted. We disagree with their conclusion. It means you���ll transmit less than 1/10th of the amount of virus you would otherwise have done, decreasing the viral load, which is likely to lead to a lower probability of infection, and fewer symptoms if infected.

The mathematics of transmission

Mathematical modeling by our team, supported by other research (Yan et al. 2019), suggests that if most people wear a mask in public, the transmission rate (���effective R���) can go beneath 1.0, entirely stopping the spread of the disease. The mask doesn���t have to block every single viral particle, but the more particles it blocks, the lower the effective R.

Modelled impact of mask use on reproduction rate

Just how effective mask-wearing is depends on three things illustrated in the diagram: how well the mask blocks the virus (���efficacy���: horizontal axis), what proportion of the public wear masks (���adherence���: vertical axis), and the transmission rate of the disease (R0: the black lines on the graph). The blue area of the graph indicates an R0 below 1.0, which is what we need to achieve to wipe out the disease. If the mask blocks 100% of particles (the far right of the graph), even low adherence rates will lead to containment of the disease. Even if masks block a much lower proportion of viral particles, the disease could still be contained ��� but only if most or all people wear masks.

The political science of mask-wearing

How do you get all or most people to wear masks? Well, you can educate them and try to persuade them, but a more effective approach is to require them to wear a mask, either in specific settings such as public transportation or grocery stores or even at all times outside the home. Research on vaccination (Bradford and Mandich 2015) shows that jurisdictions which set a higher bar for vaccine exemptions have higher vaccination rates. The same approach is now being used to increase mask wearing compliance, and early results (Leffler et al. 2020) suggest that these laws are effective at increasing compliance and slowing or stopping the spread of COVID-19.

Mask-wearing experiments: artificial and natural

An artificial experiment is when a researcher allocates people (usually at random ��� hence the term ���randomized controlled trial��� or RCT) to either wearing a mask or not wearing a mask (the control group). There have been no RCTs of mask-wearing by members of the public in COVID-19. RCTs of mask-wearing to prevent other diseases (such as influenza or tuberculosis) have tended to show a small effect which in many studies was not statistically significant. In most such studies, people assigned to the mask-wearing group didn���t always wear their masks.

A natural experiment is when we study something that is really happening ��� for example when a country introduces a policy of wearing masks. South Korea, for example, had rapid community spread that tracked the trajectory in Italy in the initial weeks. Then, in late February 2020, the government provided a regular supply of masks to every citizen. From that point, everything changed. As Italy���s death count accelerated to horrific levels, South Korea���s actually started decreasing. Here���s South Korea���s number of reported cases (red), and Italy���s (blue); take a close look at what happened in early March, as the impact of the mask distribution kicked in (this South Korean analysis is thanks to Hyokon Zhiang and visualization by Reshama Shaikh:

Comparison of COVID-19 cases between Korea and Italy

Natural experiments are scientifically imperfect, because there is no direct control group so we can���t be sure that any change is due to the masks. In some countries that introduced mask-wearing, other measures such as strict social distancing, school closures, and cancellation of public events happened at around the same time. Even in these cases, we can find relevant comparisons. For instance, European neighbors Austria and Czechia introduced social distancing requirements on the same date, but Czechia also introduced mandatory mask wearing. The Austrian case rate continued its upward trajectory, whilst Czechia���s flattened out. It wasn���t until Austria also introduced mask laws weeks later that the two counties returned to similar trajectories.

Comparison of COVID-19 cases between Czechia and Austria

Importantly, in every country and every time period where mask usage has been encouraged through laws, or where masks were provided to citizens, case and death rates have fallen.

The behavioral science of mask wearing

Some have claimed that making (or strongly encouraging) people to wear masks will encourage risky behavior (Brosseau et al. 2020) (for example, going out more, washing hands less), with a net negative result, and this effect was seen in some experimental trials of masks. Similar arguments have previously been made for HIV prevention strategies (Cassell et al. 2006; Rojas Castro, Delabre, and Molina 2019) and motorcycle helmet laws (Ouellet 2011). However, real-world research on these topics found that even though some individuals responded with risky behavior, at a population level there was an overall improvement in safety and well-being (Peng et al. 2017; Houston and Richardson 2007).

The economics of mask-wearing

Economic analyses consider how much it costs to provide masks with how much value (both financial and non-financial) might be created ��� and, potentially, lost ��� if they are provided. Such economic studies (Abaluck et al. 2020) indicate that each mask worn by one person (which costs almost nothing) could generate economic benefits of thousands of dollars and save many lives.

The anthropology of mask-wearing

Mask-wearing by the public has been normalized in many Asian countries, partly for individual reasons (to protect against pollution) and partly for collective ones (as a result of recent MERS and SARS epidemics). My mask protects you; yours protects me. However, in most of these countries the norm has been to only wear a mask if you have symptoms; it���s only in recent weeks, as awareness of asymptomatic spread has become better understood, that mask wearing regardless of symptoms has become common.

Conclusion

Whilst not every piece of scientific evidence supports mask-wearing, most of it points in the same direction. Our assessment of this evidence leads us to a clear conclusion: keep your droplets to yourself ��� wear a mask.

You can make one at home, from a t-shirt, handkerchief, or paper towel, or even just wrap a scarf or bandana around your face. Ideally, use tightly woven fabric that you can still breathe through. Researchers recommend including a layer of paper towel as a disposable filter; you can simply slide it between two layers of cloth. There is no evidence that your mask needs to be made with any particular expertise or care to be effective for source control. You can put a cloth mask in the laundry and reuse it, just like you re-use a t-shirt.

If it turns out that you���re incubating COVID-19, the people you care about will be glad you wore a mask.

Epilogue: Jeremy���s Illustration of Source Control

Here���s a little illustration of source control from Jeremy!

References


Abaluck, Jason, Judith A. Chevalier, Nicholas A. Christakis, Howard Paul Forman, Edward H. Kaplan, Albert Ko, and Sten H. Vermund. 2020. ���The Case for Universal Cloth Mask Adoption and Policies to Increase Supply of Medical Masks for Health Workers.��� SSRN Scholarly Paper ID 3567438. Rochester, NY: Social Science Research Network. https://papers.ssrn.com/abstract=3567438.
Bai, Yan, Lingsheng Yao, Tao Wei, Fei Tian, Dong-Yan Jin, Lijuan Chen, and Meiyun Wang. 2020. ���Presumed Asymptomatic Carrier Transmission of Covid-19.��� Jama.
Bradford, W David, and Anne Mandich. 2015. ���Some State Vaccination Laws Contribute to Greater Exemption Rates and Disease Outbreaks in the United States.��� Health Affairs 34 (8): 1383���90.
Brosseau, Lisa M., ScD, Margaret Sietsema, PhD Apr 01, and 2020. 2020. ���COMMENTARY: Masks-for-All for COVID-19 Not Based on Sound Data.��� CIDRAP. https://www.cidrap.umn.edu/news-perspective/2020/04/commentary-masks-all-covid-19-not-based-sound-data.
Cassell, Michael M, Daniel T Halperin, James D Shelton, and David Stanton. 2006. ���Risk Compensation: The Achilles��� Heel of Innovations in Hiv Prevention?��� Bmj 332 (7541): 605���7.
Doremalen, Neeltje van, Trenton Bushmaker, Dylan H. Morris, Myndi G. Holbrook, Amandine Gamble, Brandi N. Williamson, Azaibi Tamin, et al. 2020. ���Aerosol and Surface Stability of SARS-CoV-2 as Compared with SARS-CoV-1.��� New England Journal of Medicine 0 (0): null. https://doi.org/10.1056/NEJMc2004973.
Duguid, JP. 1946. ���The Size and the Duration of Air-Carriage of Respiratory Droplets and Droplet-Nuclei.��� Epidemiology & Infection 44 (6): 471���79.
Houston, David J, and Lilliard E Richardson. 2007. ���Risk Compensation or Risk Reduction? Seatbelts, State Laws, and Traffic Fatalities.��� Social Science Quarterly 88 (4): 913���36.
Leffler, Christopher, Edsel Ing, Craig A. McKeown, Dennis Pratt, and Andrzej Grzybowski. 2020. ���Country-Wide Mortality from the Novel Coronavirus (COVID-19) Pandemic and Notes Regarding Mask Usage by the Public.���
Morawska, LJGR, GR Johnson, ZD Ristovski, Megan Hargreaves, K Mengersen, Steve Corbett, Christopher Yu Hang Chao, Yuguo Li, and David Katoshevski. 2009. ���Size Distribution and Sites of Origin of Droplets Expelled from the Human Respiratory Tract During Expiratory Activities.��� Journal of Aerosol Science 40 (3): 256���69.
Ouellet, James V. 2011. ���Helmet Use and Risk Compensation in Motorcycle Accidents.��� Traffic Injury Prevention 12 (1): 71���81.
Peng, Yinan, Namita Vaidya, Ramona Finnie, Jeffrey Reynolds, Cristian Dumitru, Gibril Njie, Randy Elder, et al. 2017. ���Universal Motorcycle Helmet Laws to Reduce Injuries: A Community Guide Systematic Review.��� American Journal of Preventive Medicine 52 (6): 820���32.
Rojas Castro, Daniela, Rosemary M Delabre, and Jean-Michel Molina. 2019. ���Give Prep a Chance: Moving on from the ���Risk Compensation��� Concept.��� Journal of the International AIDS Society 22: e25351.
To, Kelvin Kai-Wang, Owen Tak-Yin Tsang, Wai-Shing Leung, Anthony Raymond Tam, Tak-Chiu Wu, David Christopher Lung, Cyril Chik-Yan Yip, et al. 2020. ���Temporal profiles of viral load in posterior oropharyngeal saliva samples and serum antibody responses during infection by SARS-CoV-2: an observational cohort study.��� Lancet Infect. Dis. 0 (0). https://doi.org/10.1016/S1473-3099(20)30196-1.
Wei, Wycliffe E. 2020. ���Presymptomatic Transmission of SARS-CoV-2 �������� Singapore, January 23��������March 16, 2020.��� MMWR. Morbidity and Mortality Weekly Report 69. https://doi.org/10.15585/mmwr.mm6914e1.
Wells, WF. 1934. ���On Air-Borne Infection: Study Ii. Droplets and Droplet Nuclei.��� American Journal of Epidemiology 20 (3): 611���18.
Yan, Jing, Suvajyoti Guha, Prasanna Hariharan, and Matthew Myers. 2019. ���Modeling the Effectiveness of Respiratory Protective Devices in Reducing Influenza Outbreak.��� Risk Analysis 39 (3): 647���61. https://doi.org/10.1111/risa.13181.
Zhang, Juanjuan, Maria Litvinova, Wei Wang, Yan Wang, Xiaowei Deng, Xinghui Chen, Mei Li, et al. 2020. ���Evolving Epidemiology and Transmission Dynamics of Coronavirus Disease 2019 Outside Hubei Province, China: A Descriptive and Modelling Study.��� The Lancet Infectious Diseases 0 (0). https://doi.org/10.1016/S1473-3099(20)30230-9.
Zou, Lirong, Feng Ruan, Mingxing Huang, Lijun Liang, Huitao Huang, Zhongsi Hong, Jianxiang Yu, et al. 2020. ���SARS-CoV-2 Viral Load in Upper Respiratory Specimens of Infected Patients.��� New England Journal of Medicine 382 (12): 1177���9. https://doi.org/10.1056/NEJMc2001737.