Thursday, June 2, 2016

"Helping" women in CS with impostor syndrome is missing the forest for the trees

Alexis Hancock recently wrote an article on impostor syndrome that has been on my mind ever since, as it adds so nicely to a blog post I wrote several months ago. I wanted to try and explain why so many women have impostor syndrome in CS:
Sociologists like to use performance as a metaphor for everyday life. Erving Goffman in particular championed the metaphor, bringing to light how our social interactions take place on various stages according to various scripts. And when people don't follow the right script on the right stage, social punishment ensues (e.g. stigma).  [...]

Since not following the script/game is costly for individuals, we're trained from a young age to be on the lookout for cues about what stage/arena we're on and what role we should be playing. [...]

Impostor syndrome is the sense that you're the wrong person to be playing the role you're in. You're acting a role that you've been trained in and hired for -- but your brain is picking up on cues that signal that you're not right for the role.

When [people] go on to play roles [they haven't been raised for], they still sometimes encounter social cues indicating they're in the wrong role. Impostor syndrome results.

Impostor syndrome is thought to be quite common amongst women in science. In this light I don't think it's surprising: there are so many cues in society that we are not what a 'scientist' is supposed to look or act like. We don't fit the stereotypes.

I'm far from the first person to argue that impostor syndrome comes from environmental cues. What Hancock's article does is point out the contradiction: impostor syndrome has environmental causes, but is talked about as being an individual's personal problem.

[While struggling with impostor syndrome] I became consumed with proving myself. Still, all the advice I received came in the form of a pep talk to “believe in myself” again. This common response to the struggles of women in tech reinforces the idea that imposter syndrome is the ONLY lens to view and cope… but the truth is, our negative experiences in tech are usually outside of our control. The overwhelming focus on imposter syndrome doesn’t provide a space to process the power dynamics affecting you; you get gaslighted into thinking it’s you causing all the problems.

Similarly, Cate Hudson writes that:
Yet imposter syndrome is treated as a personal problem to be overcome, a distortion in processing rather than a realistic reflection of the hostility, discrimination, and stereotyping that pervades tech culture. [...] Assuming that it’s just irrational self-doubt denies potentially useful support or training. Most of all, chalking up myriad factors to such an umbrella term belies the need to explore where these concerns arise from and how they can be addressed or mitigated. Subtle or not-so-subtle undermining behavior by colleagues? Gendered feedback? Lack of support or mentorship? [...] We pretend imposter syndrome is some kind of personal failing of marginalized groups, rather than an inevitability and a reflection of a broken and discriminatory tech culture.

So many well-intentioned diversity efforts in computer science focus on impostor syndrome and try to help women cope with it. But that discourse treats the women who have impostor syndrome as though they have an individual problem. The effect can silence women: instead of seeing their negative environment as a structural issue, they blame themselves.

Those of us who want to get more women into CS need to stop telling women that they suffer from impostor syndrome and instead help them see environment they're in. The social cues that are affecting them need to be identified and mitigated. And we need to stop teaching women to blame themselves for the sexism around them.

Monday, March 21, 2016

A Seven-Step Primer on Soft Systems Methodology

I'm currently TAing for CSC2720H Systems Thinking for Global Problems, a graduate-level course on systems thinking. In class today we talked about soft systems thinking (SSM), an approach which uses systems thinking to tackle what are called "wicked problems". I thought I'd outline one approach to SSM, as it's useful to CS education research.


Step 1: Identify the domain of interest

Before you can research something, you should first decide what your domain is. What topic? What system are you studying? For example, "teaching computer science" could be your starting point, as could "climate change".

Chances are you're looking at a wicked problem. Conklin's definition of wicked problems are that:
  1. The problem is not understood until after the formulation of a solution.
  2. Wicked problems have no stopping rule.
  3. Solutions to wicked problems are not right or wrong.
  4. Every wicked problem is essentially novel and unique.
  5. Every solution to a wicked problem is a 'one shot operation.'
  6. Wicked problems have no given alternative solutions.
Because you're looking at a domain which doesn't have a clear definition or boundaries, you'll first want to immerse yourself in the domain. One trick is to draw "rich pictures", which are essentially visualized streams of consciousness.

You should also think about what perspectives you bring into this domain. What biases and privileges do you have going into this? Why are you interested in this domain? What do you have to gain or lose here?

Monday, March 14, 2016

"'Women in Computing' As Problematic": A Summary

I've long been interested in why, despite so much organized effort, there percentage of women in CS has been so stagnant. One hypothesis I had for some time was that the efforts themselves were unintentionally counter-productive: that they reinforced the gender subtyping of "female computer scientist" being separate from unmarked "computer scientists".

I was excited earlier this week when Siobhan Stevenson alerted me to this unpublished thesis from OISE: "Women in Computing as Problematic" by Susan Michele Sturman (2009).

In 2005-6, Sturman conducted an institutional ethnography of the graduate CS programmes at two research-intensive universities in Ontario. In institutional ethnography, one starts by "reading up": identifying those who have the least power and interviewing them about their everyday experiences. From what the interviews reveal, the researcher then goes on to interview those identified as having power over the initial participants.

Interested in studying graduate-level computer science education, she started with female graduate students. This led her to the women in computing lunches and events, interviewing faculty members and administrators at those two universities. She also attended the Grace Hopper Celebration of Women in Computing (GHC) and analysed the texts and experiences she had there. Her goal was to understand the "women in computing" culture.

In the style of science studies scholars like Bruno Latour, Sturman comes to the organized women in computing culture as an outsider. As a social scientist, she sees things differently: "Women in the field wonder what it is about women and women's lives that keeps them from doing science, and feminists ask what it is about science that leads to social exclusion for women and other marginalized groups" 

Friday, February 19, 2016

Getting Fedora 23 working on an Asus Zenbook UX305CA (Intel Skylake)

I recently acquired a shiny new Asus Zenbook UX305CA to replace my old UX32A which had been dying a slow death for the past year.

Excitedly, I put the latest Fedora release (23) on the computer, using the Cinnamon spin. While the computer ran Fedora, the screen resolution was set at 800x600 with no other options.

The issue? The Intel Skylake chip in the computer wasn't supported by the kernel that Fedora 23 ships with (kernel version 2.3). Like many linux users with new laptops I've found myself in a bit of an adventure with the new skylake chip. I thought I'd write up how I eventually got Fedora 23 working on this computer for the sake of those following the same path.

To get linux working with kernel 2.3, I found the Arch Wiki invaluable:
  • I needed the kernel boot argument: i915.preliminary_hw_support=1
  • And then you set xorg.conf as described in the Arch Wiki

Once both of those were done my computer was working, but without hardware acceleration. The next step was to install kernel 4.4, which supports Skylake.
  • You'll want to add the repository where Fedora keeps the latest kernel versions: I found 4.4 in kernel-vanilla-stable (see instructions here)
  • Then, once I tried booting with kernel-4.4, I got an error at boot: "double free at 0x(address) Aborted. Press any key to exit". To get rid of the error, I found I had to temporarily disable the validation steps of the new kernel as described in comment 18 on the bugzilla report
  • The mokutil utility will ask you to set a password for altering safe boot. Write it down. When you reboot it will ask for the password on a character by character basis, where the order of the characters is random. I wound up failing this the first time because I assumed the password should be 0-indexed; it's actually 1-indexed.
  • Once I had insecure boot turned on, I could successfully boot kernel-4.4! But cinnamon informed me that software rendering was still on. To solve this, I had to undo what I'd done to make kernel-4.2 work: take out the i915.preliminary_hw_support=1 and set xorg.conf to what is recommended for Intel graphics in general rather than the Skylake bandaid (you just take out the options line).

Once all that was done, the computer's working quite nicely!

Thursday, January 28, 2016

On Paulo Freire, and seeing computing as literacy

Paulo Freire was a Brazilian educator, best known for his book Pedagogy of the Oppressed. Indeed, it's the most commonly assigned reading in education classes which isn't a textbook. His ideas have been used for teaching many topics, such as health and African American studies. And yet, most people in CS education circles aren't familiar with Freire. In this post I'll provide a short introduction to Freire and why his work is relevant to computing education.

To Freire, education is an inherently political act. Education can be a tool of empowerment, and it can also be a tool of oppression. Freire refered to traditional education as the "banking model": the teacher deposits coins of knowledge into the bank accounts of the students. "Instead of communicating, the teacher issues communiques and makes deposits which the students patiently receive, memorize, and repeat. This is the "banking" concept of education, in which the scope of action allowed to students extends only as far as receiving, filing, and storing the deposits." (Freire, 1968)

This model ignores what the student already may know. It fails to give the students a sense of ownership over their knowledge, and fails to stimulate critical thinking. He argued this reinforces oppression. For education to be empowering, students need to be active agents in their own learning.

Wednesday, January 27, 2016

Impostor syndrome viewed through the lens of social theory

Sociologists like to use performance as a metaphor for everyday life. Erving Goffman in particular championed the metaphor, bringing to light how our social interactions take place on various stages according to various scripts. And when people don't follow the right script on the right stage, social punishment ensues (e.g. stigma).

Pierre Bourdieu rather similarly described social interactions as taking place in arenas, seeing them more like games than plays. (Sometimes champs is translated as 'field' rather than arena; it's worth noting Bourdieu intended for it to have a connation of sport/war.) Rather than a script, people get a sense for the rules of the game. And when people don't follow the rules of the game, social punishment ensues.

Whether one is failing at a social game or performance, social punishment can take many forms. For example, sexual harassment is most reported by those who go against gender roles. Powerful women are more likely to be harassed than less powerful women. Women in male-dominated fields are more likely to be harassed. Men who are effeminate, gay, or champions of feminism, are more likely to be harassed. Harassers act to keep people "in their place".

Since not following the script/game is costly for individuals, we're trained from a young age to be on the lookout for cues about what stage/arena we're on and what role we should be playing. Looking for and responding to cues is something we do automatically most of the time. Kahneman would see it as an example of System 1 thinking.

Impostor syndrome is the sense that you're the wrong person to be playing the role you're in. You're acting a role that you've been trained in and hired for -- but your brain is picking up on cues that signal that you're not right for the role.

Thursday, January 21, 2016

CS grades: probably more normal than you think they are

It's commonly said that computer science grades are bimodal. And people in the CS education community have spent a lot of time speculating and exploring why that could be. A few years back, I sat through a special session at ICER on that very topic, and it occurred to me: has anybody actually tested if the grades are bimodal?

From what I've seen, people (myself included) will take a quick visual look at their grade distributions, and then if they see two peaks, they say it's bimodal. I've done it.

Here's the thing: eyeballing a distribution is unreliable. If you gave me some graphs of real-world data, I wouldn't be able to tell on a quick glance whether they're, say, Gaussian or Poissonian. And if I expected it to be one of the two, confirmation bias and System 1 Thinking would probably result in me concluding that it looks like my expectation.

Two peaks on real world data don't necessarily mean you have a bimodal distribution, particularly when the two peaks are close together. A bimodal distribution means you have two different normal distributions added together (because you're sampling two different populations at the same time).

It's quite common for normal distributions to have two "peaks", due to noise in the data. Or the way the data was binned. Indeed, the Wikipedia article on Normal distribution has this histogram of real world data that is considered normal -- but has two peaks:
And since this graph looks in all honesty like a lot of the grades distributions I've seen, I decided I'd statistically test whether CS grades distributions are normal vs. bimodal. I got my hands on the final grades distributions of all the undergraduate CS classes at the University of British Columbia (UBC), from 1996 to 2013. That came out to 778 different lecture sections, containing a total of 30,214 final grades (average class size: 75).

How do you test for normality vs bimodality?

There are a bunch of ways to test whether some data are consistent with a particular statistical distribution.

One way is to fit your data to whatever formula describes that distribution. You can then eyeball whether your resulting curve matches the data, or you could look at the residuals, or even do a goodness of fit test. (It's worth noting that you could fit a normal distribution as bimodal -- the two sub-distributions would be extremely close together! If you can fit a normal distribution to it, this is a simpler explanation of the data -- Occam's razor and all.)

Another way is to use a pre-established statistical test which will allow you to reject/accept a null hypothesis about the nature of your data. I went this route, for the ease of checking hundreds of different distributions and comparing them.

There are a large variety of tests for whether a distribution is normal. I chose Shapiro-Wilk, since it has the highest statistical power.

There aren't as many tests for whether a distribution is bimodal. Most of them work more or less by trying to capture the difference in means in the two distributions that are in the bimodal model, and testing whether the means are sufficiently separate. I used Hartigan's Dip Test, because it was the only one that I could get working in R #OverlyHonestMethods.

I also computed the kurtosis for every distribution, because I had read that a necessary but not sufficient condition for bimodality is that kurtosis < 3. When you do thousands of statistical tests, you're gonna have a lot of false positives. To minimize false positives, I only used Hartigan's Dip Test on distributions where the kurtosis was less than 3. I set my alpha value at 0.05, so I expect a false positive rate of 5%.

Test results

Starting with kurtosis: 323 of the 778 lecture sections had a kurtosis less than 3. This means that 455 (58%) of the classes were definitely not bimodal, and that at most 323 (42%) classes could be bimodal.

Next I applied Hartigan's Dip Test to the 323 classes which had a kurtosis less than 3. For this test, the null hypothesis is that the population is unimodal. As a result, if p is less than alpha, then we have a multimodal distribution. This was the case for 45 classes (10% of those tested, 5.8% of all the classes).

For the Shapiro-Wilk test, the null-hypothesis is that the population is normally-distributed. So, if the p value is less than the alpha value, we can say the population is not normally distributed. This was the case for 106 classes.

44 of the 45 classes which were previously determined to be multimodal were amongst the 106 classes which the Shapiro-Wilk test indicated weren't normally-distributed. In short, 13.6% of the classes weren't normal, many of which are known to be multimodal.

For the 86.4% of classes where we failed to reject the null hypothesis, we can expect but not guarantee due to type II error that they are normal. I've got a large sample size, and good statistical power. From bootstrapping a likely beta value, I estimate my false negative rate is around 1.48%.

Bottom line: An estimated 85.1% of the final grades in UBC's undergrad CS classes are normally-distributed. 5.8% of the classes tested as being bimodal, which isn't a whole lot more than the false positive rate I'd expect to see (5%).

Discussion

I've only analyzed distributions from one institution, so you might be thinking "maybe UBC is special". And maybe UBC is special.

I couldn't get my hands on a similar quantity of data from my home institution (U of Toronto). But every U of T class I could test was normally-distributed (n=5). Including classes that I'd taught, where I'd eyeballed the grades, and then told my colleagues/TAs/students that my grades were bimodal. Oops.

Since I thought CS classes were bimodal, when I looked at my noisy grades distributions, I saw bimodality. Good old System 1 Thinking. Had I taken the time to fit my data, or statistically test it, I would have instead concluded it was normally-distributed.

I'm currently reading Stephen Jay Gould's The Mismeasure of Man, and this part stuck out for me: "Statisticians are trained to be suspicious of distributions with multiple modes." Where you see multiple modes, you're likely either looking at a lot of noise -- or two populations are improperly being sampled together.

Why are CS distributions so noisy? My colleague Nick Falkner recently did a series of blog posts on assessments in CS classes, and how they're so truly ugly. And my colleagues Daniel Zingaro, Andrew Petersen and Michelle Craig have written a couple of lovely articles which together paint a story that if you ask students a bunch of incremental small concept questions, rather than one giant all-encompassing code-writing question, you get grades distributions which look more normal. How we assess our students affects what sort of distribution we get.

Perhaps once we as CS educators figure out better ways to assess our students, our grades distributions won't be quite so noisy -- and prone to miscategorization?