design matrix is rank deficient

[pmeiners]

Would it be possible to include a function for factorial designs? Right now, only a single ATE can be calculated by a call of conduct_ri.
This would probably require a different syntax for the 'formula' argument.

[acoppock]

Can you say more about what test you want to conduct randomization inference for? There's an example in the help file for an interaction.

[pmeiners]

Thank you for your answer and thank you for the great packages you provided in the declaredesign project. Maybe I misunderstood something about the purpose of the nested model testing. Testing the interaction does not seem to work for my data since I only get the error message "coefficients not defined because the design matrix is rank deficient".
It's very strange, since I just literally used the vignette code. The only difference is that the variables used are not truly numeric.
But using the simulated data approach from your vignette works fine.

Here is my data:
data_ri2.csv.zip

packageVersion("ri2")
# [1] ‘0.2.0’

dec <- declare_ra(N=849, prob = 0.50)
ate_hat <- coef(lm(opinion ~ Z, data = res_ri))[2]
ate_hat

ri2_out <-
  conduct_ri(
    model_1 = opinion ~ Z + X, # restricted model
    model_2 = opinion ~ Z + X + Z*X, # unrestricted model
    declaration = dec,
    sharp_hypothesis = ate_hat,
    data = res_ri
  )

Edit: This happens to me with all model comparisons, not just with this one...

[eamanj]

This seems to be a problem only if your treatments are encoded as "character" type. I had a similar problem and kept getting "coefficients not defined because the design matrix is rank deficient" with a simple binary treatment and ATE as the statistic.
But once I changed the treatment variable to a factor, it worked fine.

[pmeiners]

I figured it out. The data has to be either numeric or integer for the example code to work.

[josh-ashkinaze]

Hi am still getting this error with my simulation and I changed the treatment variable to factor, tried different formats etc. Wondering if anyone has any ideas?

library(purrr)
library(ri2)
library(dplyr)

d_to_log_odds<- function(d){d*1.81} 
p_to_log_odds <- function(p){log(p/(1-p))}

N = 8000

# Define coefficients
b.lib <- d_to_log_odds(0.2)
b.recip <- d_to_log_odds(0.2)
b.treat <- d_to_log_odds(0.5)
b.treat.lib <- d_to_log_odds(0.3)
b.0 <- p_to_log_odds(0.05)

# Create evenly sized blocks of treatment, liberal recip
blocks <- list(
  treat = c(0,1), 
  lib = c(0,1),
  recip = c(0,1)
)  %>% cross_df()
X <- blocks[rep(seq_len(nrow(blocks)), each = N/nrow(blocks)), ]

# Linear predictions
X$lin_pred <- 
            b.treat * X$treat +
            b.lib * X$lib + 
            b.recip * X$recip + 
            b.treat.lib * X$lib * X$treat

# Logit transform 
X$pr <-1/(1+exp(-X$lin_pred))

# Make realizations
X$y = rbinom(N,1,X$pr)



# RI2
# TEST FOR LIB X TREAT INTERACTION

################


declaration <- 
  with(X,{
    declare_ra(
      blocks = interaction(X$lib, X$recip, sep="_"),
      prob=0.5)
  })

ate_hat <- coef(lm(y ~ treat, data=X))[2]

# Trying as factors
X$treat <- as.factor(X$treat)
X$lib <- as.factor(X$lib)
X$recip <- as.factor(X$recip)
conduct_ri(
  model_1 = "y ~ treat + recip + lib",
  model_2 = "y ~ treat + recip + lib + treat*lib",
  assignment = "treat",
  outcome = "y",
  declaration = declaration,
  sharp_hypothesis = ate_hat,
  data = X,
)


# Trying as numeric
X$treat <- as.numeric(X$treat)
X$lib <- as.numeric(X$lib)
X$recip <- as.numeric(X$recip)
conduct_ri(
  model_1 = "y ~ treat + recip + lib",
  model_2 = "y ~ treat + recip + lib + treat*lib",
  assignment = "treat",
  outcome = "y",
  declaration = declaration,
  sharp_hypothesis = ate_hat,
  data = X,
)