Hi Ritvik! The way you've set up the linear model(s), the intercept parameters B0 and B2 will represent the intercept from Tt=0, rather than at the particular time that the interruption happens. So, if you got a large positive change in the B3 slope parameter, you'll probably get a negative change in the B2 parameter (since a steeper line at positive Tt will intercept the yt axis *below* where the original line did! Wouldn't it make more sense to do a shift/translation of the Tt parameter so that it works as if the interruption happens at the yt axis? For example, use a translated variable Ut = (Tt - t). So, if the interruption happens at t = 75, then when Tt = 75, Ut = (Tt - t) = (75 - 75) = 0. And your model can then be: yt = B0 + B1⋅Ut + Dt⋅(B2 + B3⋅Ut) At least this way, the B0 and B2 params will have a much more interpretable meaning. B0 will be the value of yt at Tt = t of the main linear model, and B2 will be the *initial change* in vt at Tt = t. In other words, how much 'immediate effect' did the interruption have; how much of a vertical 'jump' up or down. Granted, to calculate vt at any particular Tt, you'll have to first convert to Ut, but that's not so bad, just a simple shift. And if you really need to find the linear params in terms of Tt, it's fairly easy to just plug in Ut = (Tt - t) and expand out to find the transformed linear params for Tt.
I would add that it's not experimental but quasi experimental because it's not really granting that the effect is due to our treatment. It might be due to a common cause or be causally biased by it so we are also assuming causal sufficiency and a sufficient knowledge of the underlying process. Still we have no way to guarantee that some unobserved cause happened in that time window. Good and clear video, as usual btw, keep going bro °u°
This is indeed a helpful video, I think we all faced a scenario where covid disrupted the time series data as in my scenario with a network of apparel stores, the sales were impacted significantly by covid lockdowns and going forward, projections of future growth based on something like Holt Winters Seasonal no longer have reliable history. I'm trying to figure a way to re-populate what the missing time series values would have been during the lockdowns, given the before and after situation.
i still dont get the part of getting the effect without the experiment.... if i dont have the time series data of my ice cream shop where i decided to advertise as right now im in middle of taking this decision then how can i fit a model which can measure the effects of it?
there's some pretty basic code here scikit-learn.org/stable/auto_examples/gaussian_process/plot_gpr_noisy_targets.html#sphx-glr-auto-examples-gaussian-process-plot-gpr-noisy-targets-py
Any data science text book recommendations?
Pattern Recognition and Machine Learning from Bishop
Introduction to Econometrics
by Jeffrey Wooldridge
@@thekingcraig1 already got that one bro, It was essential reading at uni, thank you though 🙌🏼
@@christusrex334 He has a more advnaced one too
Econometric Analysis of Cross Section and Panel Data
Go on, tell us how to test if and when a change occurred :)
Hi Ritvik! The way you've set up the linear model(s), the intercept parameters B0 and B2 will represent the intercept from Tt=0, rather than at the particular time that the interruption happens. So, if you got a large positive change in the B3 slope parameter, you'll probably get a negative change in the B2 parameter (since a steeper line at positive Tt will intercept the yt axis *below* where the original line did!
Wouldn't it make more sense to do a shift/translation of the Tt parameter so that it works as if the interruption happens at the yt axis? For example, use a translated variable Ut = (Tt - t). So, if the interruption happens at t = 75, then when Tt = 75, Ut = (Tt - t) = (75 - 75) = 0. And your model can then be:
yt = B0 + B1⋅Ut + Dt⋅(B2 + B3⋅Ut)
At least this way, the B0 and B2 params will have a much more interpretable meaning. B0 will be the value of yt at Tt = t of the main linear model, and B2 will be the *initial change* in vt at Tt = t. In other words, how much 'immediate effect' did the interruption have; how much of a vertical 'jump' up or down.
Granted, to calculate vt at any particular Tt, you'll have to first convert to Ut, but that's not so bad, just a simple shift. And if you really need to find the linear params in terms of Tt, it's fairly easy to just plug in Ut = (Tt - t) and expand out to find the transformed linear params for Tt.
I would add that it's not experimental but quasi experimental because it's not really granting that the effect is due to our treatment. It might be due to a common cause or be causally biased by it so we are also assuming causal sufficiency and a sufficient knowledge of the underlying process. Still we have no way to guarantee that some unobserved cause happened in that time window.
Good and clear video, as usual btw, keep going bro °u°
This is indeed a helpful video, I think we all faced a scenario where covid disrupted the time series data as in my scenario with a network of apparel stores, the sales were impacted significantly by covid lockdowns and going forward, projections of future growth based on something like Holt Winters Seasonal no longer have reliable history. I'm trying to figure a way to re-populate what the missing time series values would have been during the lockdowns, given the before and after situation.
Great vid. Please explain how to compute the confidence intervals.
i still dont get the part of getting the effect without the experiment.... if i dont have the time series data of my ice cream shop where i decided to advertise as right now im in middle of taking this decision then how can i fit a model which can measure the effects of it?
So simple… but so powerful!
glad you think so!
thanks, may I ask what do you do for a living?
Pls recommend a book for data science n machine learning
Knocking it out of the park, as usual.
Can you show us the code for this, especially for VAR
there's some pretty basic code here
scikit-learn.org/stable/auto_examples/gaussian_process/plot_gpr_noisy_targets.html#sphx-glr-auto-examples-gaussian-process-plot-gpr-noisy-targets-py
Very clear and helpful. Thank you so much!
You're very welcome!