Model Selection Methods - Backward Elimination

Authors

Jose Storopoli

Andreas Noack

Joel Owen

As stated in Introduction to Covariate Selection Methods, broadly, there are two main types of Stepwise Covariate Model (SCM):

  1. Forward Selection (FS)
  2. Backward Elimination (BE)

In this tutorial we’ll explore Backward Elimination (BE) in detail.

1 Backward Elimination (BE)

Backward elimination stepwise covariate modeling procedures are very similar to FS procedures. However, instead of beginning with a base model, BE starts with a full model, denoted as \(M_0\), and is the reference model for the models of the first iteration, The first iteration tests all defined covariate-parameter effects in the full model (\(c \in C\)) in univariable runs by removing all available \(c\) from \(M_0\) in separate runs. The model with the smallest increase in criterion value is subject to being removed from the model, and is designated \(M_k\), the reference for the next iteration.

BE proceeds with the subsequent iterations by individually removing all remaining \(c_k\) from \(M_k\), where \(k\) is the iteration number, and computes the selected metric for each model. The \(c_k\) model selected for removal (the worst \(c\) in iteration \(k\)) is removed from the set \(C_{k+1}\) of covariate effects for iteration \(k+1\). BE proceeds with additional iterations until it cannot find \(c_k\) to remove without degrading the fit of the model fit metric or if there aren’t \(c_k\) left to be removed from the model.

Here’s a breakdown of the BE algorithm:

  • Full model with all covariate effects: \(M_{\text{full}}\)
  • Set of covariate effects: \(C\)
  • Data: \(D\)
  • Criterion, a function that takes a model along with data, and outputs a real number (with the assumption that lower is better): \(f: (M, D) \to \mathbb{R}\)
  • Best model with a subset of covariate effects \(C^*\) according to criterion \(f\) evaluated on data \(D\)
  1. define \(M = M_{\text{full}}\)

  2. define \(f_m = f(M, D)\)

  3. for every \(c \in C\):

    1. define \(M_c = M - c\)

    2. define \(f_c = f(M_c, D)\)

    3. if \(f_c < f_m\)

      • define \(M_{\text{best}} = M_c\)
      • define \(c_{\text{best}} = c\) else unset \(c_{\text{best}}\)
  4. if \(c_{\text{best}}\) is defined

    • define \(M = M_{\text{best}}\)
    • define \(C = C \setminus \{ c_{\text{best}} \}\)
    • goto step 2
  5. define \(C^* = C^\complement\)

  6. return \(M\) and \(C^*\)

In the procedures above, we start by defining the initial model \(M\) as the full model with all covariate effects \(M_{\text{full}}\). Then we proceed by defining the initial criterion \(f_m\) as the criterion \(f\) of the initial model \(M\) and data \(D\). Next, we do a for loop on every covariate effect \(c\) in the set of all covariate effects \(C\), where we:

  1. define the current model \(M_c\) as the model \(M\) without the covariate effect \(c\), that is \(M - c\).

  2. define the criterion of the current model \(f_c\) as the criterion \(f\) applied to the current model \(M_c\) and data \(D\).

  3. if the criterion improved, i.e \(f_c < f_m\), then:

    • define the new best model \(M_{\text{best}}\) as the current model \(M_c\).
    • define the new best covariate effect \(c_{\text{best}}\) as the covariate effect \(c\).
  4. if we have a best covariate effect \(c_{\text{best}}\), then:

    • define the new “initial model” \(M\) as the best model \(M_{\text{best}}\).
    • define the new set of covariate effects as the set \(C\) minus best covariate effect \(c_{\text{best}}\)
    • go to the step 2 and resume from there.
  5. define the new set of covariate effects \(C^*\) as the set complement \(C^\complement\) of the set \(C\)

  6. return the final “initial model” \(M\) and the new set of covariate effects \(C^*\)

Note

\(C^\complement\) is the set complement of the set \(C\).

\[C = C^{\complement^\complement}\]

2 The covariate_select function

All stepwise covariate selection technique in Pumas uses the covariate_select function. It has the following function signature:

covariate_select(
    model,
    population,
    param,
    approx;
    control_param,
    criterion = aic,
    method = CovariateSelection.Forward,
    fit_options,
)

There are positional arguments and keyword arguments.

  1. model: a Pumas model
  2. population: a Population that is commonly obtained from read_pumas
  3. param: a set of initial parameter estimates
  4. approx: a likelihood approximation method, e.g. FOCE(), LaplaceI(), etc.

These are exactly the same positional arguments that we pass to the fit function to perform a model fit procedure.

The keyword arguments we have some new concepts:

  • control_param: this is a tuple of parameters that encapsulates unique covariate effects such that, if we restrain their values to \(0\), we effectively eliminate any influence of the covariate effects in our model. control_param can be thought as a list of covariate effects that we want to screen in our covariate selection routine.
  • criterion: Criterion for covariate selection. This needs to be a function that accepts a Pumas model and returns a real value. If you don’t provide a function Pumas will, by default, use the aic function. Note that criterion has the assumption that lower is better. Hence, you’ll need to take that into account when supplying a different function or defining your own user-defined function.
  • method: the desired method to be used for the stepwise covariate selection procedure. The default is CovariateSelection.Forward which stands for Forward Selection (FS). You can also opt for CovariateSelection.Backward if you want to use Backward Elimination (BE)
  • fit_options: a NamedTuple of options that will be used during the model fit procedures that covariate_select will perform. These will be the keyword arguments available for the fit function, e.g. (; constantcoef = (; myparam = 1.0)).

3 Example in Pumas

Let’s showcase an example that builds from the model we used in Covariate Selection Methods - Introduction.

First, let’s import the following packages:

using Pumas
using PharmaDatasets

We are going to use the po_sad_1 dataset from PharmaDatasets:

df = dataset("po_sad_1")
first(df, 5)
5×14 DataFrame
Row id time dv amt evid cmt rate age wt doselevel isPM isfed sex route
Int64 Float64 Float64? Float64? Int64 Int64? Float64 Int64 Int64 Int64 String3 String3 String7 String3
1 1 0.0 missing 30.0 1 1 0.0 51 74 30 no yes male ev
2 1 0.25 35.7636 missing 0 missing 0.0 51 74 30 no yes male ev
3 1 0.5 71.9551 missing 0 missing 0.0 51 74 30 no yes male ev
4 1 0.75 97.3356 missing 0 missing 0.0 51 74 30 no yes male ev
5 1 1.0 128.919 missing 0 missing 0.0 51 74 30 no yes male ev

This is an oral dosing (route = "ev") NMTRAN-formatted dataset. It has 18 subjects, each with 1 dosing event (evid = 1) and 18 measurement events (evid = 0); and the following covariates:

  • age: age in years (continuous)
  • wt: weight in kg (continuous)
  • doselevel: dosing amount, either 30, 60 or 90 milligrams (categorical)
  • isPM: subject is a poor metabolizer (binary)
  • isfed: subject is fed (binary)
  • sex: subject sex (binary)

Let’s parse df into a Population with read_pumas:

population =
    read_pumas(df; observations = [:dv], covariates = [:wt, :isPM, :isfed], route = :route)
Population
  Subjects: 18
  Covariates: wt, isPM, isfed
  Observations: dv

We’ll use the same 2-compartment oral absorption model as in Covariate Selection Methods - Introduction, but now adding additional covariates and covariate effects:

model = @model begin
    @metadata begin
        desc = "full covariate model"
        timeu = u"hr"
    end

    @param begin
        """
        Clearance (L/hr)
        """
        tvcl  RealDomain(; lower = 0)
        """
        Central Volume (L)
        """
        tvvc  RealDomain(; lower = 0)
        """
        Peripheral Volume (L)
        """
        tvvp  RealDomain(; lower = 0)
        """
        Distributional Clearance (L/hr)
        """
        tvq  RealDomain(; lower = 0)
        """
        Fed - Absorption rate constant (h-1)
        """
        tvkafed  RealDomain(; lower = 0)
        """
        Fasted - Absorption rate constant (h-1)
        """
        tvkafasted  RealDomain(; lower = 0)
        """
        Power exponent on weight for Clearance
        """
        dwtcl  RealDomain()
        """
        Proportional change in CL due to PM
        """
        ispmoncl  RealDomain(; lower = -1, upper = 1)
        """
          - ΩCL
          - ΩVc
          - ΩKa
          - ΩVp
          - ΩQ
        """
        Ω  PDiagDomain(5)
        """
        Proportional RUV (SD scale)
        """
        σₚ  RealDomain(; lower = 0)
    end

    @random begin
        η ~ MvNormal(Ω)
    end

    @covariates begin
        """
        Poor Metabolizer
        """
        isPM
        """
        Fed status
        """
        isfed
        """
        Weight (kg)
        """
        wt
    end

    @pre begin
        CL = tvcl * (wt / 70)^dwtcl * (1 + ispmoncl * (isPM == "yes")) * exp(η[1])
        Vc = tvvc * (wt / 70) * exp(η[2])
        Ka = (tvkafed + (isfed == "no") * (tvkafasted)) * exp(η[3])
        Q = tvq * (wt / 70)^0.75 * exp(η[4])
        Vp = tvvp * (wt / 70) * exp(η[5])
    end

    @dynamics Depots1Central1Periph1

    @derived begin
        cp := @. 1_000 * (Central / Vc)
        """
        DrugY Concentration (ng/mL)
        """
        dv ~ @. Normal(cp, cp * σₚ)
    end
end
PumasModel
  Parameters: tvcl, tvvc, tvvp, tvq, tvkafed, tvkafasted, dwtcl, ispmoncl, Ω, σₚ
  Random effects: η
  Covariates: isPM, isfed, wt
  Dynamical system variables: Depot, Central, Peripheral
  Dynamical system type: Closed form
  Derived: dv
  Observed: dv

The underlying Pumas model that covariate_select will use needs to include all of the covariate effects to be tested. This is different than just specifying particular covariates in individual models for a more manual process. Note that one covariate could have more than one effect.

During each iteration of BE, covariate_select will set the covariate effect being evaluated for elimination from the model to zero. Each of the covariate effects should be totally controlled by a respective parameter defined in the @param block. The model should be coded in such way that, if the corresponding parameter value is set to zero, then the covariate effect is completely removed from the model.

In the model above, we are defining the following parameters in the @param block with respect to covariate effects:

  • tvkafasted: typical value of the magnitude of the isfed covariate effect on tvka
  • dwtcl: typical value of the magnitude of the wt covariate efffect on tvcl
  • ispmoncl: typical value of the magnitude of the isPM covariate efffect on tvcl

These will be the parameters that we will be testing for the covariate selection.

Further down in the @pre block we see that if we set any of these parameters value to zero we will completely remove the covariate effect from the model. For the case of tvkafasted, this expression in the @pre block:

Ka = (tvkafed + (isfed == "no") * (tvkafasted)) * exp(η[3])

shows us that if tvkafasted is zero, then the whole expression (isfed == "no")*(tvkafasted) evaluates to zero, and we are only left with:

Ka = tvkafed * exp(η[3])

which removes the effect of tvkafasted from the value of Ka.

For the case of dwtcl and ispmoncl, we have the expression:

CL = tvcl * (wt / 70)^dwtcl * (1 + ispmoncl * (isPM == "yes")) * exp(η[1])

in the @pre block.

  1. if dwtcl is set to zero, then (wt / 70)^dwtcl evaluates to one. This would remove the wt effect on tvcl.
  2. if ispmoncl is set to zero, then ispmoncl * (isPM == "yes") evaluates to zero. This would remove the isPM effect on tvcl.

Note that it is the effect size parameter (i.e., dwtcl or ispmoncl), that is set to zero and effectively removes the covariate effect, whether continuous or categorical.

With the model done, now we define our initial parameters:

iparams = (;
    tvkafed = 0.4,
    tvkafasted = 1.5,
    tvcl = 4.0,
    tvvc = 70.0,
    tvq = 4.0,
    tvvp = 50.0,
    dwtcl = 0.75,
    ispmoncl = -0.7,
    Ω = Diagonal(fill(0.04, 5)),
    σₚ = 0.1,
)
(tvkafed = 0.4,
 tvkafasted = 1.5,
 tvcl = 4.0,
 tvvc = 70.0,
 tvq = 4.0,
 tvvp = 50.0,
 dwtcl = 0.75,
 ispmoncl = -0.7,
 Ω = [0.04 0.0 … 0.0 0.0; 0.0 0.04 … 0.0 0.0; … ; 0.0 0.0 … 0.04 0.0; 0.0 0.0 … 0.0 0.04],
 σₚ = 0.1,)

3.1 BE with Different Metrics

Finally, we can run the covariate selection procedure.

We’ll use FOCE() as our estimation method. For the control_param tuple, we include all of the covariate effects we mentioned above: dwtcl, tvkafasted and ispmoncl. These will be the parameters that we’ll test for covariate selection.

We will now choose the criterion as BIC (bic) and the Backward Elimination (BE) method (CovariateSelection.Backward):

covar_result = covariate_select(
    model,
    population,
    iparams,
    FOCE();
    control_param = (:dwtcl, :tvkafasted, :ispmoncl),
    criterion = bic,
    method = CovariateSelection.Backward,
)
[ Info: fitting model with zero restrictions on: 
[ Info: criterion: 2691.9638537448213
[ Info: fitting model with zero restrictions on: dwtcl
[ Info: criterion: 2714.4999368673707
[ Info: fitting model with zero restrictions on: tvkafasted
[ Info: criterion: 2724.0471785654026
[ Info: fitting model with zero restrictions on: ispmoncl
[ Info: criterion: 2725.077583277103
[ Info: current best model has zero restrictions on dwtcl
[ Info: criterion did not improve, stopping!
┌ Info: final best model summary
│   best_criterion = 2691.9638537448213
└   best_zero_restrictions = ()
Pumas.CovariateSelectionResults
Criterion:                                               bic
Method:                                             Backward

Best model:
  Unrestricted parameters:       dwtcl, tvkafasted, ispmoncl
  Zero restricted parameters:                               
  Criterion value:                                   2691.96

Fitted model summary:
4×4 DataFrame
 Row │ free_parameters                   zero_parameters  fitted_model         ⋯
     │ Tuple…                            Tuple…           FittedPumas…         ⋯
─────┼──────────────────────────────────────────────────────────────────────────
   1 │ (:dwtcl, :tvkafasted, :ispmoncl)  ()               FittedPumasModel(... ⋯
   2 │ (:tvkafasted, :ispmoncl)          (:dwtcl,)        FittedPumasModel(...
   3 │ (:dwtcl, :ispmoncl)               (:tvkafasted,)   FittedPumasModel(...
   4 │ (:dwtcl, :tvkafasted)             (:ispmoncl,)     FittedPumasModel(...
                                                               2 columns omitted

As you can see, covariate_select, upon completion, prints its result.

Now, let’s showcase how you can use a custom criterion.

Here’s an example of a function that takes a fitted Pumas model, m, and outputs the -2LL (minus 2 times log-likelihood) without the constant. This is a model fit measure commonly used by NONMEM users and is is known as OFV: Objective Function Value.

Hence, we will name the function ofv:

ofv(m) = (-2 * loglikelihood(m)) - (nobs(m) * log(2π))
ofv (generic function with 1 method)

Now let’s run covariate_select again but with ofv as a criterion:

covar_result_ofv = covariate_select(
    model,
    population,
    iparams,
    FOCE();
    control_param = (:dwtcl, :tvkafasted, :ispmoncl),
    criterion = ofv,
    method = CovariateSelection.Backward,
)
[ Info: fitting model with zero restrictions on: 
[ Info: criterion: 2117.3591383883204
[ Info: fitting model with zero restrictions on: dwtcl
[ Info: criterion: 2145.4936434698693
[ Info: fitting model with zero restrictions on: tvkafasted
[ Info: criterion: 2155.0408851679003
[ Info: fitting model with zero restrictions on: ispmoncl
[ Info: criterion: 2156.071289879601
[ Info: current best model has zero restrictions on dwtcl
[ Info: criterion did not improve, stopping!
┌ Info: final best model summary
│   best_criterion = 2117.3591383883204
└   best_zero_restrictions = ()
Pumas.CovariateSelectionResults
Criterion:                                               ofv
Method:                                             Backward

Best model:
  Unrestricted parameters:       dwtcl, tvkafasted, ispmoncl
  Zero restricted parameters:                               
  Criterion value:                                   2117.36

Fitted model summary:
4×4 DataFrame
 Row │ free_parameters                   zero_parameters  fitted_model         ⋯
     │ Tuple…                            Tuple…           FittedPumas…         ⋯
─────┼──────────────────────────────────────────────────────────────────────────
   1 │ (:dwtcl, :tvkafasted, :ispmoncl)  ()               FittedPumasModel(... ⋯
   2 │ (:tvkafasted, :ispmoncl)          (:dwtcl,)        FittedPumasModel(...
   3 │ (:dwtcl, :ispmoncl)               (:tvkafasted,)   FittedPumasModel(...
   4 │ (:dwtcl, :tvkafasted)             (:ispmoncl,)     FittedPumasModel(...
                                                               2 columns omitted

As you can see, this works as expected and returns the same best model as the BIC approach.

First, we have the best model that was found with its parameters and criterion value (showcasing the bic case).

Next, we have a DataFrame-rendered table with the iterations that BE tested as the rows, and the following columns:

  • free_parameters: covariate effects that were allowed to influence the model, i.e. were not forced to zero.
  • zero_parameters: covariate effects that were not allowed to influence the model, i.e. were forced to zero.
  • fitted_model: the underlying Pumas’ model fit
  • criterion: criterion value (bic in our case)

We can access the results of the covariate selection procedure by accessing the fits field of the covar_result object, and convert it to a DataFrame in order to perform any data wrangling or plotting.

Here’s an example of how to sort the results by the criterion value:

sort(DataFrame(covar_result.fits), :criterion)
4×4 DataFrame
Row free_parameters zero_parameters fitted_model criterion
Tuple… Tuple… FittedPu… Float64
1 (:dwtcl, :tvkafasted, :ispmoncl) () FittedPumasModel(...) 2691.96
2 (:tvkafasted, :ispmoncl) (:dwtcl,) FittedPumasModel(...) 2714.5
3 (:dwtcl, :ispmoncl) (:tvkafasted,) FittedPumasModel(...) 2724.05
4 (:dwtcl, :tvkafasted) (:ispmoncl,) FittedPumasModel(...) 2725.08

As you can see this puts the “best” model (according to the method and criterion) in the first row.

As you can see, after running the covariate selection procedure, we get a summary of the best model found and the fitted model for each step of the procedure.

You can recover the best model fit by accessing the best_model field of the covar_result object:

covar_result.best_model
FittedPumasModel

Successful minimization:                      true

Likelihood approximation:                     FOCE
Likelihood Optimizer:                         BFGS
Dynamical system type:                 Closed form

Log-likelihood value:                    -1306.793
Number of subjects:                             18
Number of parameters:         Fixed      Optimized
                                  0             14
Observation records:         Active        Missing
    dv:                         270              0
    Total:                      270              0

-------------------------
               Estimate
-------------------------
tvcl            3.6888
tvvc           70.044
tvvp           48.706
tvq             4.3506
tvkafed         0.41114
tvkafasted      1.0894
dwtcl           0.81856
ispmoncl       -0.62382
Ω₁,₁            0.022271
Ω₂,₂            0.050192
Ω₃,₃            0.031194
Ω₄,₄            0.10595
Ω₅,₅            0.035072
σₚ              0.098043
-------------------------