mutate and predict.gam results Na/NaN/inf problem

Question

I have an issue with some data, and I simply cannot understand why.

I'm trying to estimate var4 from var3 using a GAM.

Here is the dataset I'm using to obtain my model:

for_model <- read.csv("https://raw.githubusercontent.com/fredlm/mockup/master/for_model.csv")

And the dataset in which I want to estimate var4 :

for_est <- read.csv("https://raw.githubusercontent.com/fredlm/mockup/master/for_est.csv")

What I've done, simply:

for_est <- for_est %>%
mutate(var4 = ifelse(!var3 == 0, predict.gam(gam(var4 ~ s(log(var3)), data = for_model), newdata = .), NA))

It returns the following error:

Error: Problem with mutate() column var4 . var4 = predict.gam(gam(var4 ~ s(log(var3)), data = for_model), newdata = .) . x NA/NaN/Inf in foreign function call (arg 1)

Despite a thorough research on the web and a few hours spent on my data, I can't find how to fix this...

However, when I plot the GAM, things work great:

ggplot(data = for_model,
       aes(x = var3,
           y = var4)) +
  geom_point() +
  geom_smooth(method = "gam",
              formula = y~s(log(x)))

Any idea how to fix this? I've looked for NaN or Inf values but there are none. Also, when I'm trying to estimate var4 from var2 — which is VERY similar to var3 - things work well...

for_est <- for_est %>%
  mutate(var4 = ifelse(!var2 == 0, predict.gam(gam(var4 ~ s(log(var2)), data = for_model), newdata = .), NA))

Thanks a lot!

ps: my apologies for the rather large files, but given that I don't understand the problem, I thought it might make more sense to provide all of them... :)

Answer 1

When you use ifelse to keep away from var3 == 0, you need to restrict the for_est input data the same way. (I split up the model solving from the predicting, just to make testing faster, that doesn't matter)

gamfit <- gam(var4 ~ s(log(var3)), data = for_model)
for_est <- for_est %>%
  mutate(var4 = ifelse(var3 != 0, predict(gamfit, newdata = .[var3 != 0, ]), NA_real_))