This package should respect lazyeval expected behaviour.

This is necessary to interface with shiny in the future.

Mostly useless and dangerous. The BEFORE workaround is confusing for the user. Automatic sorting of parameter or state values definition will be a mess.

Now that the package is more or less complete, vignettes should be reorganized to be more to the point.

I want to believe.

Right now it's the bare minimum.

Usually, we measure number of people being in a state.
So we should not start with only one person.

Investigate.

magrittr is a good starting point. Silly pipe pictures not allowed though.

Should be based on the newdata framework.

Because people are lazy.

Heterogeneity, interpretation and incertitude.

Functions from dplyr and ggplot2 use non-standard evaluation. This causes R CMD check to complain. globalVariables() are used to avoid that problem, but I'd rather keep those to a minimum.

Calls to globalVariables() must be audited and alternative solutions must be found whenever possible.

Separate help?

Use invoke_rows() instead.

The parallelisation should mostly occur where models are run iteratively in run_model_newdata().

Should include:

• Binomial and multinomial variables.
• Beta and gamme density.
• Bounded density (e.g. non negative).
• Empirical density.
• RR, OR, HR...

Bottlenecks significantly reduce the speed of iterative functions for heterogeneity and probabilistic analysis. They should be identified and improvements should be suggested.

Should be avoided and that step should be inside define_model; and simplified like that

mod_standard <- define_model(
  parameters = param_standard,
  transition_matrix = mat_trans,
  state A,
  state B,
  state C,
  ...
)

non homogeneous

1. explain age <- floor(age/10-.5)*10+5
2. replace merge(data.frame(age = age, sex = sex), death_prob)$mr by death_prob[death_prob$age == age & death_prob$sex == sex, "mr"] (more intuitive)
3. standardRR should be renamed
4. define markov_cycle variable
5. should set the variables standard treatment costs 394 and NP1 costs 579 somewhere in the text and in the code (not at the end)

homogeneous

And move it to a new repo.

Because it's easier to visualize.

Right now the code is pretty messy and unoptimized.

Need the following steps:

1. Ability to run model with new data.
2. Generate new data from arbitrary distributions with correlation structure.
3. Provide interface to specify parameter distribution and correlation structure.
4. Provide convenience functions to specify distribution density for all kind of parameters.

Should be replaced by its abstract value (although being an equation probably difficult to solve).
For instance
for this line : C 0.202 * rr 0.067 * rr 0.01 * rr
C should be output like 0.721rr

Mainly for check_matrix.

The code of define_resample() works, but is incredibly messy and probably too complex for what it does.

This is a recipe for bugs and headaches. This part of the package should be refactored.

No named state -> generating names: we should show this warning at define_matrix() function and not afte

One dataframe instead of a list of dataframes.

Follow classic file hierarchy style. See magrittr or dplyr github repo for reference.

The values do not depend on state and are added to state values for all individuals.

Should support:

• Beginning of cycle;
• End of cycle;
• Half cycle correction;
• Life-year correction.

Which method should be left by default is open to debate.

summary function shouldn't output .cost, .effect and .icer when there is no comparison between several strategies. Or replace values with NA
Same remark for the less costly strategy

• Compute ICER

An unevaluated state with 2 values: The term "unevaluated" leads to confusion and I think we should remove it.

It seems to be a step which could be inside run_model.

run_model(
  transition_matrix = mat_trans,
  state_A,
  state_B,
  cycles = 10,
  cost = cost,
  effect = utility
)

All parameters and state value names starting by . should be forbidden.

res_mod_1

## 1 Markov model, run for 10 cycles.
## 
## Model name:
## 
## A

Perhaps we could replace that output by the summary function output?

Could be fixed by tweaking diagram::plotmat() parameters.

Writing 1-(standardRR+mr) could be avoided, as it is the complement of the other states; I propose to use an alias for that (could be letter C)