object not found in mediation model

Question

I have to plan a simple mediation model - https://cran.r-project.org/package=mediation

library(mediation)

mediation_model = lm(as.numeric(Age) ~ as.numeric(condition), data = GRA)
tab_model(mediation_model)

model = lm(as.numeric(note) ~ as.numeric(condition) + as.numeric(Age), data = GRA)
tab_model(model)

result <- mediate(mediation_model,model,sims = 120,mediator="as.numeric(Age)",treat="as.numeric(condition)",boot = TRUE)

And it generates this error:

Running nonparametric bootstrap

Error in eval(predvars, data, env) : object 'Age' not found

I taped 'manes(GRA)' and I found 'Age'. The lm works well, I don't know where the mistake is.

Not familiar with "mediation" package, but try to convert to numeric before running models, before lm: GRA$Age <- as.numeric(GRA$Age) and same for "condition", then run lms, then mediate, with mediator as simply "Age". — zx8754
– zx8754, Commented Nov 8, 2023 at 9:40
I made all my code with this non numeric variable without problem. It's why I not expected that — killiann
– killiann, Commented Nov 8, 2023 at 9:53
lm is able to operate because you explicitly told it to convert variables to numeric-class. Anything downstream that wants to work on the model expects that the data then has these same variables already as numeric, there is no attempt to recreate the steps taken on the data. (To be clear ... in ML and some forecasting work in R, there is "feature engineering" that does something akin to recording the data-munging steps, and when the model is used on new data, the steps are recreated/rerun on the new data. That's a great notion, but it is not part of lm and lm-derived summaries.) — r2evans
– r2evans, Commented Nov 8, 2023 at 13:01

r2evans · Accepted Answer · 2023-11-08 13:56:19Z

I'll formalize the comments.

Any steps you perform on the data within the formula are not carried forward for any post-model analysis. This means that when you run mediate, the versions of note, condition, and Age are their original class within your GRA data; the coefficients in model suggest that they should be numeric, but number-like operations are not working.

As a counter-point, in many ML and forecasting (in R) pipelines, the "feature engineering" or pre-conditioning steps are recorded in a sense, so that when you apply that ML/forecasting model to new data, the steps are repeated on the new data. This is a novel and great way to handle modeling, but unfortunately lm and friends are not based on that architecture.

(Disclaimer: I don't know mediate and don't have it installed, I'm not speaking to the correctness of your regression or its use.)

Convert the columns into a frame (either overwriting GRA or saving to a new frame, GRA2 below), and then don't worry about as.numeric(.) in either of your calls to lm(.) or mediate(.) (or any other model-using function):

GRA2 <- GRA |>
  transform(
    note = as.numeric(note),
    condition = as.numeric(condition),
    Age = as.numeric(Age)
  )
mediation_model = lm(Age ~ condition, data = GRA2)
model = lm(note ~ condition + Age, data = GRA2)
result <- mediate(mediation_model, model, sims = 120,
                  mediator="Age", treat="condition", boot = TRUE)

Also, there are other ways to convert those three columns, this is just one. My point is not that you must use transform, feel free to use whatever R code you want to pre-convert the three columns into a frame, and then use that frame in calls to lm and mediate.

Collectives™ on Stack Overflow

object not found in mediation model

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related