## Time-series data from FIRE shows increasing trend of disinvitations caused by pressure from the Left

I’ve elevated to its own post (here) a comment from reader Ed Kroc on yesterday’s post, “Shenanigans in Illinois. 2: DePaul University bans yet another speaker“. In that post I cited FIRE’s (Foundation for Individual Rights in Education’s) “Disinvitation Database,” which comprised 308 cases of individuals being disinvited from speaking at colleges since 2000. A quick scan of the data suggested that the relative proportion of disinvitations prompted by protests from the left versus from the right had increased in recent years. I asked if a reader might analyze this, and Ed complied.

It looks as if I were right, though there’s no formal statistical analysis yet. (One way is to simply divide the years in half and do a 2 X 2 chi-square analysis of disinvitations from each side in the first versus second eight years. There are, of course, caveats, the most serious of which is noted below:

Ed’s original comment is below:

Here’s that plot you were looking for (should be hosted at this link.

I only used data from actual disinvitations, not from attempted disinvitations. There is no statistical evidence of a nonzero trend for disinvitations from the right. There does appear to be a significant increasing trend for disinvitations from the left, even if you exclude the very influential year 2016.

Caveats: as far as I can tell, this data is essentially self-reported; i.e., if no one reports an incident, then it doesn’t show up in the database. It could be people are paying more attention to censorship from the left recently, which would explain the trend. Or it could be that people are not reporting censorship from the right as much as they used to for some reason. Etc….

1. TJR
Posted August 5, 2016 at 1:00 pm | Permalink

If you fit simple trend lines then I suspect the “red” one will be flat (i.e. slope term within two standard errors of zero).

The “blue” one will almost certainly not be, but as JAC noted in the previous post, the 2016 point is clearly highly influential (high leverage).

(This is not really the right way to analyse a time series of course, but is reasonably sensible in a back of an envelope sort of way, as long as you remember not to extrapolate).

• Posted August 5, 2016 at 1:56 pm | Permalink

You are correct about the simple linear fits. (Of course, you’re also right that this is not really the correct way to analyze a time series given the autocorrelation, but it’s just a rough exploratory indicator.)

2. Geoffrey Howe
Posted August 5, 2016 at 1:24 pm | Permalink

There needs to be more info here. First off, it’s not even apparent if Blue represents disinvitations OF the Left or FROM the Left.

Second off, a glance over the chart (ignoring 2016) looks like a lot of noise, with things going up and down without significant impact. There might be a curve fit that’s applicable here, but I can’t really eyeball that kind of thing.

Lastly, number of disinivations is small enough that noise could explain most of the deviations, so I’m not sure I can draw any real conclusions from this.

2016 is a big deal, but it could also be represented by a reporting bias, or a bias on the part of FIRE. Unfortunately, without more info here, this chart doesn’t seem to really support anything.

• Posted August 5, 2016 at 1:53 pm | Permalink

I suggest you go back and look at the original article and database; that will answer your first question.

• Posted August 5, 2016 at 2:00 pm | Permalink

It’s supposed to be disinvitations FROM the left; the original FIRE database explains this, though I probably should have included the clarification in the graphic.

Reporting bias is a real concern here, and I don’t know how that could be assessed. Assuming it’s not too great though (perhaps a big assumption), there is evidence of something going on with disinvitations from the left.

3. Posted August 5, 2016 at 3:47 pm | Permalink

I took the Disinvitation data, restructured it as a time-to-event dataset in Excel (constructed a variable for “time” including 16 “times” representing the years of analysis and a variable for “censor” that indicates when an entry had a disinivation event), saved it as a .csv file and imported it into R, where I plotted a Kaplan-Meier curve showing the difference in time to disinvitation for the “Left” (solid line) and “Right” (dotted line). While there were more disinvitations among the “Left” (n=61), the risk was greater among the “Right” (n=55) (p-value=0.00027). I’ve provided an image of the Kaplan-Meier curve demonstrating this in my twitter pic below.

Kaplan-Meier curve demonstrating the (unintuitive) greater risk in the “Right”

• Posted August 5, 2016 at 3:51 pm | Permalink

• Posted August 5, 2016 at 4:30 pm | Permalink

Hi Charleen, cool graphic! I don’t quite understand your analysis though. Shouldn’t we be considering a dataset that includes all invited speakers, and then examining the risk of disinvitation relative to this pool? Isn’t the increased risk you’re reporting now just a function of the “disinvited from the left” events being more concentrated in recent years? Or maybe I’m totally misunderstanding something!

• Posted August 5, 2016 at 4:37 pm | Permalink

Yes, we should look at all invited speakers, but I didn’t think the Disinvitation set included everyone who was invited. From what I saw (and I didn’t look at it long), we had data on kerfuffles that lead and didn’t lead to disinvitation but not info about everyone who was invited.

I could rerun this and keep all the people in the dataset were who were almost disinvited and count them as “not censored.” That might show us a very different picture 🙂

As for your question about the concentration of left disinvitations in recent years, no the analysis looked at risk over the entire spread of those years. Those in the right were worse off early on too. But, there are other ways to skin this chicken, and I’m not an expert at this 🙂 If I have time, I’ll replot, but I might not get to it. I was simply smitten by the challenge earlier and couldn’t stop myself.

• Posted August 5, 2016 at 4:54 pm | Permalink

This is the code I used to generate the plot in R. (I had created the variables “Time” and “Censor” in Excel first.)

time.surv <- survfit(Surv(Time, Censor)~ strata(Group), conf.type="none")
summary(time.surv)
plot(time.surv, lty=c(1,3),xlab="Time", ylab="Survival Probability")
survdiff(Surv(Time, Censor) ~ Group, rho=0)

Here is the dataset I used.

• Posted August 5, 2016 at 8:29 pm | Permalink

This plot includes whatever invites were included in the Disinvitation set.

• Posted August 5, 2016 at 5:31 pm | Permalink

On second thought, Ed, I don’t think I can call the calculation I did a risk for exactly the reason you mentioned about not having the set of everyone invited. Given that, I’m not sure how to make sense of what I calculated, but I sure had fun doing it!