2016-08-24 Followup VI

[andkov]

Discussion topics

1. Dealing with estimation errors when using misspecification matrix. When estimating models B and C as defined here, produces error:

 Error in optim(method = "BFGS", control = list(trace = 0, REPORT = 1,  : 
  initial value in 'vmmin' is not finite 

[andkov]

• 2. After recoding education from years of education into tertiles:

ds_clean$educat <- car::Recode(ds_clean$edu,
" 1:9   = '-1'; 
  10:11 = '0';
  12:30 = '1';
")

my models involving this covariate will have a harder time converging, at times failing to. Any similar experiences?

[andkov]

• 3. Specification of the models to fit could be unified by sharing the same object with model specification matrices Q and E. An example would be:

# Specification of the fitted models

# common value for all models
q <- .01

# ---- model-A --------------

# transition matrix
Q <- rbind( c(0, q, q, q), 
            c(q, 0, q, q),
            c(0, q, 0, q), 
            c(0, 0, 0, 0)) 
# misclassification matrix
E <- rbind( c(0, 0, 0, 0),  
            c(0, 0, 0, 0), 
            c(0, 0, 0, 0),
            c(0, 0, 0, 0) )
# transition names
qnames = c(
  "Healthy - Mild",  # q12
  "Health - Severe", # q13
  "Healthy - Dead",  # q14
  "Mild - Healthy",  # q21  
  "Mild - Severe",   # q23
  "Mild - Dead",     # q24
  "Severe - Mild",   # q32
  "Severe - Dead"    # q34
)
A <- list(q=q, Q=Q, E=E, qnames = qnames)
#
# add definition for B, C, and D
#
model_specification <- list(A=A, B=B, C=C, D=D)
# use in estimation:
# define specification matrices
(Q      <- model_specification[["A"]][["Q"]])
(E      <- model_specification[["A"]][["E"]])
(qnames <- model_specification[["A"]][["qnames"]])
# compile model objects with msm() call
mA1 <-     estimate_multistate(ds, Q, E, qnames,cov_names = "age")
mA2 <-     estimate_multistate(ds, Q, E, qnames,cov_names = "age + age_bl")
mA3 <-     estimate_multistate(ds, Q, E, qnames,cov_names = "age + age_bl + male")
mA4 <-     estimate_multistate(ds, Q, E, qnames,cov_names = "age + age_bl + male + educat")
mA4_edu <- estimate_multistate(ds, Q, E, qnames,cov_names = "age + age_bl + male + edu")

# save models estimated by msm() in a external object for faster access in the future
models_A <- list("age"    = mA1, 
                 "age_bl" = mA2, 
                 "male"   = mA3, 
                 "educat" = mA4, # estimated with only 200 ids
                 "edu"    = mA4_edu) # turn of after estimation
saveRDS(models,     "./data/shared/derived/models/version-2/models_A.rds")           # turn of after estimation
models_A <- readRDS("./data/shared/derived/models/version-2/models_A.rds")
lapply(models_A, names)

[andkov]

Minutes

Andrey

• when education is compressed

Robson

• use rounded values from edu into Q matrix of the model that used categorized education

Graciela = or use the starting values from the model that did not include education at

• unfolding strategy : use estimated models from a smimpler model as the starting values in the next, incrementally more complex model

Annie

• models too long to estimate,

Emiel

• need more time, work on personal computer becuase University environment does not allow installing any additional software

Johan

• can't estimate model with misspecification
• worked when excluded individuals with illegal transitions

Andrey

• I can't run misspecified models either.

Robson

• the error is associated with illegal transions

Emiel

• are we doing sex specific tertile of the lifestyle variables (as Marioni writes)
• we centered age AND divied by 20

Graciela

• not, we are not using sex specific
• not using social engagement at this moment

Annie

• but we use occupation, wait on social engagement
• I centered age at 75

Centering age: study specific age. Use whatever work for your estimation case.

Robson

• transformation of age variable should not affect the resultant inferences.

Andrey

• I centered age at 80

Judith

• just starting with Whitehall

....

Graciela

• let's keep to to the smallest number of states possible

Next meeting is scheduled for Friday, 2016-09-02-08:00 PDT (16:00 UCT)