# higher intercept value for the outcome

[wenyanyy]

Hi Team,

Sorry for having too many questions. I used my own data to run Robyn code. But as the result, the intercept is too high almost 100% and each variable has 0% coefficient. I tried 3 different data sets and almost have same questions. Could you please tell me how to fix it? For example, maybe change the parameters or lambda in function R file?

Best,

[gufengzhou]

Hi, are you saying when using our simulated data, the waterfall plot looks normal. but with your own data it doesn't? this definitely sounds strange. How do you set hyperparameters? (set_hyperBoundLocal)

[wenyanyy]

Hi,
you're right. When I used your simulated data, the results looks perfect. However, when i used my own data, it looks very strange as the attachment you saw. I didn't change the hyperparameters bounds. my data is monthly data so there is only 14 rows. do you think it is a reason which result a very high intercept? Also for the result ,we don't want to see any variables that have 0 coefficient. Do you think how we can avoid having 0 coefficient for any media channel spending(variables)? Thank you for your time and looking forward for your reply.

[gufengzhou]

Hey, 14 rows are bit toooo sparse :) There's a rule of thumb about the n * p relationship of your input data dimensions (n = num of rows, p = num of columns), which is n should be about multiplier 7-10 of p. For example if you have 10 variables, rule of thumb would be to have 70-100 rows, and indeed we recommend to go for multiplier 10. With only 14 observations, I don't believe any technique can give you meaningful results. Therefore, I'd recommend you to break the time unit down to at least weekly, so that you'll have 14 * 7 = 98 rows at least.
Regarding avoiding 0 coef: Robyn uses ridge regression that is reducing overfitting by shrinking beta coefs. Predictors that have very weak correlation to your dependent variable will probably be reduced to 0. What we recommend is:

1. increase hyperparameter ranges for 0-coef channels on theta (max.reco. c(0, 0.9) ) and gamma (max.reco. c(0.1, 1) ) to give Robyn more freedom
2. split media into sub-channels, and/or aggregate similar channels, and/or introduce other media
3. increase trials to get more samples

However, with only 14 rows I am afraid none of these will help. No data, no magic;)

[wenyanyy]

Hi, Thank you so much for the explanation. Really appreciate Best, Get Outlook for iOS<https://aka.ms/o0ukef>

…

________________________________ From: gufengzhou ***@***.***> Sent: Friday, April 30, 2021 6:17:22 PM To: facebookexperimental/Robyn ***@***.***> Cc: wenyanyy ***@***.***>; Author ***@***.***> Subject: Re: [facebookexperimental/Robyn] # higher intercept value for the outcome (#73) Hey, 14 rows are bit toooo sparse :) There's a rule of thumb about the np relationship of your input data dimensions (n = num of rows, p = num of columns), which is n should be about multiplier 7-10 of p. For example if you have 10 variables, rule of thumb would be to have 70-100 rows, and indeed we recommend to go for multiplier 10. With only 14 observations, I don't believe any technique can give you meaningful results. Therefore, I'd recommend you to break the time unit down to at least weekly, so that you'll have 147 = 98 rows at least. Regarding avoiding 0 coef: Robyn uses ridge regression that is reducing overfitting by shrinking beta coefs. Predictors that have very weak correlation to your dependent variable will probably be reduced to 0. What we recommend is: 1. increase hyperparameter ranges for 0-coef channels on theta (max.reco. c(0, 0.9) ) and gamma (max.reco. c(0.1, 1) ) to give Robyn more freedom 2. split media into sub-channels, and/or aggregate similar channels, and/or introduce other media 3. increase trials to get more samples However, with only 14 rows I am afraid none of these will help. No data, no magic;) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#73 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ATVLBYFUIIUB3TXPKCJOWVDTLJ7TFANCNFSM43ZLPHAA>.