Twitter API suspension

[g4brielvs]

What is the purpose of this Pull Request?

This is an analysis (take 1) to start the conversation to understand how Twitter API suspension might have impacted Rosie's level of engagement.

What was done to achieve this purpose?

I used time series analysis, particularly an autoregressive model.

How to test if it really works?

An overview of the methodology would be a good start.

Who can help reviewing it?

@cuducos @jtemporal

TODO

[jtemporal]

Question: Can we reliably use the Quadratic regression on the before 2018 data? I ask this because we have a huge gap in information (which I think is due to Rosie's sabbatical).

Other than the data itself, I wonder if this is leading to an inconclusive result since the linear regression shows one thing and the quadratic another. Am I missing some mathematical/statistical concept here?

[g4brielvs]

@jtemporal thank you for the feedback. I am with you. I pointed out that a polynomial regression might not be the best approach here, specially because we have a reason to think that that time series is not a stationary process. That is why I used an autoregressive integrated moving average (ARIMA) model instead.

Have you had the chance to look at my notebook?

[cuducos]

Can we reliably use the Quadratic regression on the before 2018 data?

Probably not, but that was may naïve approach just to get started. As the mathematician who really adds values in the analysis is @g4brielvs, what about git rm my notebooks (which were merely warmups for his analysis)? We can also checkout to my commits to see what I've tried.

[jtemporal]

hi @g4brielvs Just started looking at yourt notebook. Bellow I'll write down some changes I think would be good to have:

• [ ] I noticed some cells are run out of order, maybe you could re-run all of them and guarantee they all work if run sequentially? That would be a major improvement.
• [ ] Remove @cuducos notebook. In that way we can focus on the "real" analysis =P and as he said it was just a warm up.
• [ ] As I mentioned before add a .describe() after the .shape so we can have statistics on numeric features documented for anyone reading the notebook without running it.
• [ ] I think there's a plot missing on the engagement part, maybe it wasn't displayed in time.

That is why I used an autoregressive integrated moving average (ARIMA) model instead.

I like that <3 I think is a better approach to the matter at hand

Note that the negative trend apparently started before and has been accentuated after the block. ... Between Fev/2019 and Apr/2019 - right after the block - the slope has higher negative value and continuously stabilizes, but in a negative trend.

<3 null hypothesis validated: block = bad

[g4brielvs]

@jtemporal thank you! I made those changes

[g4brielvs]

@jtemporal Hey! I just wanted to check if this PR is still relevant. If more changes are needed, I'd be happy to work on those.

[jtemporal]

Hi @g4brielvs I think we need to check with @sergiomario on this 😉