random vectors for multivariate explanatory variables

[akgandhi]

There does not seem to be an obvious mechanism to define random vectors in a DAG, for example, to use in a multivariate regression. Is there an option/ability to apply dag_plate to index over the X_i's in a multivariate random vector (where the underlying data object would then be a design matrix).

[flyaflya]

Thanks Amit for the note. There is some support for this already, but it is non-obvious and under-supported with documentation at this moment. Maybe you can help guide my development with some use cases you would like to see?

In regards to current functionality, you can see a demonstration of design matrices in a DAG using the dev version of causact:

remotes::install_github("flyaflya/causact")

and looking at examples 3, 4, and 5 of this R-script: https://raw.githubusercontent.com/flyaflya/causact/master/gretaExamples.R

The DAG for example #3 is shown below.
Example DAG is shown below:

Let me know what you think. I can make a release from the DEV version so this capability is available to your students via CRAN.

[akgandhi]

Thanks a ton - let me give this a spin to see how it works and will let you know tomorrow. Btw am placing to just RStudio cloud for teaching so all three configuration of packages can be managed by me on the cluster- saves the headache of being a tech support desk!! Get Outlook for Android<https://aka.ms/ghei36>

…

________________________________ From: Adam Fleischhacker <notifications@github.com> Sent: Tuesday, December 1, 2020 5:01:06 PM To: flyaflya/causact <causact@noreply.github.com> Cc: Gandhi, Amit K <akgandhi@wharton.upenn.edu>; Author <author@noreply.github.com> Subject: Re: [flyaflya/causact] random vectors for multivariate explanatory variables (#36) Thanks Amit for the note. There is some support for this already, but it is non-obvious and under-supported with documentation at this moment. Maybe you can help guide my development with some use cases you would like to see? In regards to current functionality, you can see a demonstration of design matrices in a DAG using the dev version of causact: remotes::install_github("flyaflya/causact") and looking at examples 3, 4, and 5 of this R-script: https://raw.githubusercontent.com/flyaflya/causact/master/gretaExamples.R The DAG for example #3<#3> is shown below. Example DAG is shown below: [image]<https://user-images.githubusercontent.com/17184441/100801456-0263d300-33f6-11eb-86c2-6f0a17b9f093.png> Let me know what you think. I can make a release from the DEV version so this capability is available to your students via CRAN. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#36 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AKMZXWBQOMXAQ4OEPAMCJ7LSSVRSFANCNFSM4UCNOWKQ>.

[akgandhi]

Hi
I am trying the code from "greta example #3" in the script provided above with the developer version of causact you recommended. In particular:

packageVersion("causact")
[1] ‘0.3.3.9000’

The graph is processed and generates a visual DAG (the same one you show above). However when trying to run the actual simulation the following error arises:

drawsDF = graph %>% dag_greta()
Error: objects of class function cannot be coerced to greta arrays

Does this error arise for you?

[flyaflya]

My bad. The "int" node needs to be defined in the DAG. Without the "int" node, the code thinks the int in int + design %*% coefs is the function int. If you add the node, it works:

library(greta)
library(tidyverse)
library(causact)
### greta example #3
graph = dag_create() %>%
  dag_node("Rating","y",
           data = attitude$rating,
           rhs = normal(mu,sd)) %>%
  dag_node("Exp. Rating","mu",
           rhs = int + design %*% coefs,
           child = "y") %>%
  dag_node("Intercept","int",
           rhs = normal(0,10),
           child = "mu") %>%
  dag_node("Std Dev of Ratings","sd",
           rhs = cauchy(0,3, truncation = c(0,Inf)),
           child = "y") %>%
  dag_node("Predictor Coefficients","coefs",
           rhs = normal(0,10),
           child = "mu",
           extract = FALSE) %>%
  dag_node("Predictors","design",
           data = attitude[, 2:7],
           keepAsDF = TRUE,
           child = "mu") %>%
  dag_plate(descr = "Predictors",label = "j",
            nodeLabels = c("coefs"),
            data = names(attitude[, 2:7])) %>%
  dag_plate(descr = "Observation", label = "i",
            nodeLabels = c("mu","y"))

graph %>% dag_render()
graph %>% dag_render(shortLabel = TRUE)
drawsDF = graph %>% dag_greta()
drawsDF %>% dagp_plot()

I updated the gretaExamples file on github to reflect the change. Thanks for finding this bug.

[akgandhi]

This works! Thank you. Two quick follow up questions (happy to start a new issue if you think it more appropriate).

1. are there any examples where the extract = TRUE option would arise (vs passing the index variable as a separate parent node to the common child).

2. I see here you use the plate iterator to define a vector greta array for "coeffs". Could we equivalently have used an array valued greta distribution (e.g., multivariate normal) - will test out myself.

[flyaflya]

1. Most of my multi-level examples use extract=TRUE which is the default. This is not readily clear as I often add the index variable as a parent in these cases using dag_plate(..., addDataNode = TRUE). The dotted line in this DAG from the top of Chapter 25 (causact.com) indicates that extract=TRUE:
   

other examples with extract = TRUE are shown in the readme file from CRAN: https://cran.r-project.org/web/packages/causact/readme/README.html

2. not sure about this off the top of my head. let me know your findings :-)

[akgandhi]

Hi Adam Was planning to start to dive into your package with grad students next week. Out of curiosity have you experimented with specifying time series models with a graphical representation via causact. It's an area I am actively researching but wanted to touch base to see if you had any experience with causact or greta in this regard. Thanks!! Get Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: Adam Fleischhacker ***@***.***> Sent: Saturday, December 5, 2020 4:57:58 PM To: flyaflya/causact ***@***.***> Cc: Gandhi, Amit K ***@***.***>; Author ***@***.***> Subject: Re: [flyaflya/causact] random vectors for multivariate explanatory variables (#36) 1. Most of my multi-level examples use extract=TRUE which is the default. This is not readily clear as I often add the index variable as a parent in these cases using dag_plate(..., addDataNode = TRUE). The dotted line in this DAG from the top of Chapter 25 (causact.com) indicates that extract=TRUE: [image]<https://user-images.githubusercontent.com/17184441/101266115-0a838180-371a-11eb-882d-11163b7a4a5e.png> 2. not sure about this off the top of my head. let me know your findings :-) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#36 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AKMZXWAQHIT5JSAD2NQU3W3STKUGNANCNFSM4UCNOWKQ>.

[flyaflya]

Hi Amit: Great that you will be digging into greta/causact ☺ Let me know of any issues or suggestions as you work with causact, I’d be happy to help. I would like to represent time series data compactly using plate notation, but have not thought much about it recently. If you have ideas for an implementation, I would be happy to code it up. Best, Adam From: Amit Gandhi ***@***.*** Sent: Thursday, April 1, 2021 9:54 AM To: flyaflya/causact ***@***.***> Cc: Fleischhacker, Adam ***@***.***>; Comment ***@***.***> Subject: Re: [flyaflya/causact] random vectors for multivariate explanatory variables (#36) Hi Adam Was planning to start to dive into your package with grad students next week. Out of curiosity have you experimented with specifying time series models with a graphical representation via causact. It's an area I am actively researching but wanted to touch base to see if you had any experience with causact or greta in this regard. Thanks!! Get Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: Adam Fleischhacker ***@***.***<mailto:***@***.***>> Sent: Saturday, December 5, 2020 4:57:58 PM To: flyaflya/causact ***@***.***<mailto:***@***.***>> Cc: Gandhi, Amit K ***@***.***<mailto:***@***.***>>; Author ***@***.***<mailto:***@***.***>> Subject: Re: [flyaflya/causact] random vectors for multivariate explanatory variables (#36) 1. Most of my multi-level examples use extract=TRUE which is the default. This is not readily clear as I often add the index variable as a parent in these cases using dag_plate(..., addDataNode = TRUE). The dotted line in this DAG from the top of Chapter 25 (causact.com) indicates that extract=TRUE: [image]<https://user-images.githubusercontent.com/17184441/101266115-0a838180-371a-11eb-882d-11163b7a4a5e.png> 2. not sure about this off the top of my head. let me know your findings :-) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#36 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AKMZXWAQHIT5JSAD2NQU3W3STKUGNANCNFSM4UCNOWKQ>. — You are receiving this because you commented. Reply to this email directly, view it on GitHub<#36 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AEDDNOKLZCA4UMES52F2RHLTGR3GVANCNFSM4UCNOWKQ>.