Calculating Parameter Uncertainty

Author

Patrick Kofod Mogensen

1 Introduction

A typical workflow for fitting a Pumas model and deriving parameter precision typically involves:

  1. Preparing the data and the model.
  2. Checking model-data compatibility.
  3. Obtaining initial parameter estimates.
  4. Fitting the model via a chosen estimation method.
  5. Interpreting the fit results.
  6. Computing parameter precision.
  7. (Optionally) proceeding with more advanced techniques like bootstrapping or SIR for robust uncertainty quantification.

The following sections will walk through these steps using a one-compartment PK model for Warfarin, focusing on the PK aspects only. Exploratory data analysis (EDA), although extremely important, is out of scope here. Readers interested in EDA are encouraged to consult other tutorials.

2 Model and Data

2.1 Model Definition

Below is the PK model, named warfarin_pk_model, defined in Pumas. This model contains:

  • Fixed effects (population parameters): pop_CL, pop_Vc, pop_tabs, pop_lag
  • Inter-individual variability (IIV) components: pk_Ω
  • Residual error model parameters: σ_prop, σ_add
  • Covariates for scaling: FSZCL and FSZV
  • Differential equations describing the PK behavior in the compartments.
using Pumas
using PharmaDatasets
using DataFramesMeta
using PumasUtilities
using StatsBase # for cov2cor
warfarin_pk_model = @model begin
    @metadata begin
        desc = "Warfarin 1-compartment PK model (PD removed)"
        timeu = u"hr"
    end

    @param begin
        # PK parameters
        """
        Clearance (L/hr)
        """
        pop_CL  RealDomain(lower = 0.0, upper = 13.4, init = 0.134)
        """
        Central volume (L)
        """
        pop_Vc  RealDomain(lower = 0.0, upper = 81.1, init = 8.11)
        """
        Absorption lag time (hr)
        """
        pop_tabs  RealDomain(lower = 0.0, upper = 5.23, init = 0.523)
        """
        Lag time (hr)
        """
        pop_lag  RealDomain(lower = 0.0, upper = 5.0, init = 0.1)

        # Inter-individual variability
        """
          - ΩCL: Clearance
        """
        pk_Ω  PDiagDomain([0.01])

        # Residual variability
        """
        σ_prop: Proportional error
        """
        σ_prop  RealDomain(lower = 0.0, init = 0.00752)
        """
        σ_add: Additive error
        """
        σ_add  RealDomain(lower = 0.0, init = 0.0661)
    end

    @random begin
        pk_η ~ MvNormal(pk_Ω)    # mean = 0, covariance = pk_Ω
    end

    @covariates begin
        """
        FSZCL: Clearance scaling factor
        """
        FSZCL
        """
        FSZV: Volume scaling factor
        """
        FSZV
    end

    @pre begin
        CL = FSZCL * pop_CL
        Vc = FSZV * pop_Vc * exp(pk_η[1])

        tabs = pop_tabs
        Ka = log(2) / tabs
    end

    @dosecontrol begin
        lags = (Depot = pop_lag,)
    end

    @vars begin
        cp := Central / Vc
    end

    @dynamics begin
        Depot' = -Ka * Depot
        Central' = Ka * Depot - (CL / Vc) * Central
    end

    @derived begin
        """
        Concentration (ng/mL)
        """
        conc ~ @. Normal(cp, sqrt((σ_prop * cp)^2 + σ_add^2))
    end
end
PumasModel
  Parameters: pop_CL, pop_Vc, pop_tabs, pop_lag, pk_Ω, σ_prop, σ_add
  Random effects: pk_η
  Covariates: FSZCL, FSZV
  Dynamical system variables: Depot, Central
  Dynamical system type: Matrix exponential
  Derived: conc
  Observed: conc

2.2 Data Preparation

The Warfarin data used in this tutorial is pulled from PharmaDatasets for demonstration purposes. Note how the code reshapes and prepares the data in “wide” format for reading into Pumas. Only the conc column is treated as observations for the PK model.


warfarin_data = dataset("paganz2024/warfarin_long")

# Step 2: Fix Duplicate Time Points
# -------------------------------
# Some subjects have duplicate time points for DVID = 1
# For this dataset, the triple (ID, TIME, DVID) should define
# a row uniquely, but
nrow(warfarin_data)
nrow(unique(warfarin_data, ["ID", "TIME", "DVID"]))

# We can identify the problematic rows by grouping on the index variables
@chain warfarin_data begin
    @groupby :ID :TIME :DVID
    @transform :tmp = length(:ID)
    @rsubset :tmp > 1
end

# It is important to understand the reason for the duplicate values.
# Sometimes the duplication is caused by recording errors, sometimes
# it is a data processing error, e.g. when joining tables, or it can
# be genuine records, e.g. when samples have been analyzed in multiple
# labs. The next step depends on which of the causes are behind the
# duplications.
#
# In this case, we will assume that both values are informative and
# we will therefore just adjust the time stamp a bit for the second
# observation.
warfarin_data_processed = @chain warfarin_data begin
    @groupby :ID :TIME :DVID
    @transform :tmp = 1:length(:ID)
    @rtransform :TIME = :tmp == 1 ? :TIME : :TIME + 1e-6
    @select Not(:tmp)
end

# Transform the data in a single chain of operations
warfarin_data_wide = @chain warfarin_data_processed begin
    @rsubset !contains(:ID, "#")
    @rtransform begin
        # Scaling factors
        :FSZV = :WEIGHT / 70            # volume scaling
        :FSZCL = (:WEIGHT / 70)^0.75     # clearance scaling (allometric)
        # Column name for the DV
        :DVNAME = "DV$(:DVID)"
        # Dosing indicator columns
        :CMT = ismissing(:AMOUNT) ? missing : 1
        :EVID = ismissing(:AMOUNT) ? 0 : 1
    end
    unstack(Not([:DVID, :DVNAME, :DV]), :DVNAME, :DV)
    rename!(:DV1 => :conc, :DV2 => :pca)
end
317×13 DataFrame
292 rows omitted
Row ID TIME WEIGHT AGE SEX AMOUNT FSZV FSZCL CMT EVID DV0 pca conc
String3 Float64 Float64 Int64 Int64 Float64? Float64 Float64 Int64? Int64 Float64? Float64? Float64?
1 1 0.0 66.7 50 1 100.0 0.952857 0.96443 1 1 missing missing missing
2 1 0.0 66.7 50 1 missing 0.952857 0.96443 missing 0 missing 100.0 missing
3 1 24.0 66.7 50 1 missing 0.952857 0.96443 missing 0 missing 49.0 9.2
4 1 36.0 66.7 50 1 missing 0.952857 0.96443 missing 0 missing 32.0 8.5
5 1 48.0 66.7 50 1 missing 0.952857 0.96443 missing 0 missing 26.0 6.4
6 1 72.0 66.7 50 1 missing 0.952857 0.96443 missing 0 missing 22.0 4.8
7 1 96.0 66.7 50 1 missing 0.952857 0.96443 missing 0 missing 28.0 3.1
8 1 120.0 66.7 50 1 missing 0.952857 0.96443 missing 0 missing 33.0 2.5
9 2 0.0 66.7 31 1 100.0 0.952857 0.96443 1 1 missing missing missing
10 2 0.0 66.7 31 1 missing 0.952857 0.96443 missing 0 missing 100.0 missing
11 2 0.5 66.7 31 1 missing 0.952857 0.96443 missing 0 missing missing 0.0
12 2 2.0 66.7 31 1 missing 0.952857 0.96443 missing 0 missing missing 8.4
13 2 3.0 66.7 31 1 missing 0.952857 0.96443 missing 0 missing missing 9.7
306 31 48.0 83.3 24 1 missing 1.19 1.13936 missing 0 missing 24.0 6.4
307 31 72.0 83.3 24 1 missing 1.19 1.13936 missing 0 missing 22.0 4.5
308 31 96.0 83.3 24 1 missing 1.19 1.13936 missing 0 missing 28.0 3.4
309 31 120.0 83.3 24 1 missing 1.19 1.13936 missing 0 missing 42.0 2.5
310 32 0.0 62.0 21 1 93.0 0.885714 0.912999 1 1 missing missing missing
311 32 0.0 62.0 21 1 missing 0.885714 0.912999 missing 0 missing 100.0 missing
312 32 24.0 62.0 21 1 missing 0.885714 0.912999 missing 0 missing 36.0 8.9
313 32 36.0 62.0 21 1 missing 0.885714 0.912999 missing 0 missing 27.0 7.7
314 32 48.0 62.0 21 1 missing 0.885714 0.912999 missing 0 missing 24.0 6.9
315 32 72.0 62.0 21 1 missing 0.885714 0.912999 missing 0 missing 23.0 4.4
316 32 96.0 62.0 21 1 missing 0.885714 0.912999 missing 0 missing 20.0 3.5
317 32 120.0 62.0 21 1 missing 0.885714 0.912999 missing 0 missing 22.0 2.5

2.3 Creating a Pumas Population

Below is the creation of a population object in Pumas using read_pumas. Only the conc data are treated as the observation variable:

pop = read_pumas(
    warfarin_data_wide;
    id = :ID,
    time = :TIME,
    amt = :AMOUNT,
    cmt = :CMT,
    evid = :EVID,
    covariates = [:SEX, :WEIGHT, :FSZV, :FSZCL],
    observations = [:conc],
)
Population
  Subjects: 31
  Covariates: SEX, WEIGHT, FSZV, FSZCL
  Observations: conc
Note

The same data can contain multiple endpoints or PD observations. In this tutorial, the focus is solely on PK fitting. PKPD modeling on this warfarin dataset will be introduced later.

Also note, that parameter inference can be expensive and for that reason we simplified the model for this tutorial to decrease overall runtime.

2.4 Checking Model-Data Compatibility

Before performing any fit, it is recommended to verify whether the defined model can handle the provided dataset. Pumas offers functions such as loglikelihood and findinfluential for these checks.

2.4.1 The loglikelihood Function

The loglikelihood function computes the log-likelihood of the model given data and parameters. In Pumas, the signature typically looks like:

loglikelihood(model, population, params, approx)

where:

  • model: The Pumas model definition (e.g., warfarin_pk_model).
  • population: A Pumas population object (e.g., pop).
  • params: A named tuple or dictionary containing parameter values.
  • approx: The approximation method to use. Common options include FOCE(), FO(), LaplaceI(), etc.

For example, one might write:

# A named tuple of parameter values
param_vals = (
    pop_CL = 0.12,
    pop_Vc = 7.3,
    pop_tabs = 0.523,
    pop_lag = 0.5,
    pk_Ω = Diagonal([0.01]),
    σ_prop = 0.00752,
    σ_add = 0.0661,
)
foce_fit = fit(warfarin_pk_model, pop, param_vals, FOCE())
[ Info: Checking the initial parameter values.
[ Info: The initial negative log likelihood and its gradient are finite. Check passed.
Iter     Function value   Gradient norm 
     0     2.754619e+04     3.345035e+04
 * time: 0.027840137481689453
     1     5.618146e+03     1.386157e+04
 * time: 1.0565710067749023
     2     2.806727e+03     6.164132e+03
 * time: 1.0785679817199707
     3     1.814316e+03     3.717848e+03
 * time: 1.0991380214691162
     4     9.458168e+02     1.261428e+03
 * time: 1.1184430122375488
     5     6.506311e+02     6.846058e+02
 * time: 1.137132167816162
     6     5.067237e+02     3.407504e+02
 * time: 1.155717134475708
     7     4.543150e+02     1.615765e+02
 * time: 1.17336106300354
     8     4.390128e+02     1.001202e+02
 * time: 1.1904079914093018
     9     4.359539e+02     1.004875e+02
 * time: 1.2070610523223877
    10     4.350899e+02     9.448070e+01
 * time: 1.2233130931854248
    11     4.340047e+02     8.066223e+01
 * time: 1.2399821281433105
    12     4.322893e+02     5.164762e+01
 * time: 1.256864070892334
    13     4.307161e+02     2.264796e+01
 * time: 1.2734739780426025
    14     4.302010e+02     1.080854e+01
 * time: 1.3584530353546143
    15     4.301247e+02     1.195746e+01
 * time: 1.3738760948181152
    16     4.301086e+02     9.445190e+00
 * time: 1.3892710208892822
    17     4.300988e+02     7.960007e+00
 * time: 1.4039571285247803
    18     4.300587e+02     7.716936e+00
 * time: 1.4188330173492432
    19     4.300051e+02     7.117247e+00
 * time: 1.4342751502990723
    20     4.299453e+02     7.208415e+00
 * time: 1.4500961303710938
    21     4.299169e+02     7.127430e+00
 * time: 1.4655671119689941
    22     4.299077e+02     7.022170e+00
 * time: 1.4808759689331055
    23     4.298994e+02     7.386201e+00
 * time: 1.495575189590454
    24     4.298766e+02     8.313301e+00
 * time: 1.510688066482544
    25     4.298227e+02     9.481826e+00
 * time: 1.5260210037231445
    26     4.296924e+02     1.085355e+01
 * time: 1.5415141582489014
    27     4.294405e+02     1.706433e+01
 * time: 1.5572900772094727
    28     4.291427e+02     1.848276e+01
 * time: 1.5729291439056396
    29     4.289303e+02     1.058023e+01
 * time: 1.5888621807098389
    30     4.288371e+02     9.815220e+00
 * time: 1.6050000190734863
    31     4.288319e+02     9.597527e+00
 * time: 1.6197099685668945
    32     4.288091e+02     8.868287e+00
 * time: 1.6356561183929443
    33     4.287649e+02     7.800905e+00
 * time: 1.6514301300048828
    34     4.286492e+02     8.240185e+00
 * time: 1.6670801639556885
    35     4.284193e+02     1.118489e+01
 * time: 1.6830780506134033
    36     4.280494e+02     1.142225e+01
 * time: 1.6994011402130127
    37     4.277503e+02     6.478284e+00
 * time: 1.7163550853729248
    38     4.276843e+02     3.593549e+00
 * time: 1.7330620288848877
    39     4.276750e+02     2.901808e+00
 * time: 1.7490391731262207
    40     4.276737e+02     2.925309e+00
 * time: 1.7646770477294922
    41     4.276724e+02     2.942991e+00
 * time: 1.7801051139831543
    42     4.276678e+02     2.975008e+00
 * time: 1.8380701541900635
    43     4.276570e+02     3.010425e+00
 * time: 1.8546221256256104
    44     4.276278e+02     3.489417e+00
 * time: 1.8717281818389893
    45     4.275555e+02     4.423545e+00
 * time: 1.8890111446380615
    46     4.273850e+02     5.582714e+00
 * time: 1.9069180488586426
    47     4.270496e+02     6.557308e+00
 * time: 1.925084114074707
    48     4.266090e+02     6.537839e+00
 * time: 1.943507194519043
    49     4.263506e+02     3.908843e+00
 * time: 1.9616141319274902
    50     4.262968e+02     2.885389e+00
 * time: 1.9801239967346191
    51     4.262907e+02     2.842439e+00
 * time: 1.9983041286468506
    52     4.262900e+02     2.821199e+00
 * time: 2.0167109966278076
    53     4.262894e+02     2.809832e+00
 * time: 2.034550189971924
    54     4.262879e+02     2.794691e+00
 * time: 2.053194046020508
    55     4.262842e+02     2.781348e+00
 * time: 2.072111129760742
    56     4.262743e+02     3.009687e+00
 * time: 2.0911591053009033
    57     4.262477e+02     5.430536e+00
 * time: 2.110504150390625
    58     4.261705e+02     9.974222e+00
 * time: 2.130488157272339
    59     4.258983e+02     2.055347e+01
 * time: 2.151236057281494
    60     4.256807e+02     2.645976e+01
 * time: 2.1848690509796143
    61     4.254650e+02     3.195354e+01
 * time: 2.2188611030578613
    62     4.250777e+02     4.156365e+01
 * time: 2.314378023147583
    63     4.250303e+02     5.260486e+01
 * time: 2.35090708732605
    64     4.240447e+02     4.793859e+01
 * time: 2.370526075363159
    65     4.222991e+02     3.988756e+01
 * time: 2.390158176422119
    66     4.175763e+02     1.012857e+01
 * time: 2.410703182220459
    67     4.156021e+02     1.547767e+01
 * time: 2.4396111965179443
    68     4.143567e+02     2.010443e+01
 * time: 2.4695851802825928
    69     4.130607e+02     1.768446e+01
 * time: 2.489320993423462
    70     4.116886e+02     8.431093e+00
 * time: 2.5101330280303955
    71     4.112484e+02     7.131364e+00
 * time: 2.5309641361236572
    72     4.109556e+02     4.884649e+00
 * time: 2.5518441200256348
    73     4.106945e+02     2.263126e+00
 * time: 2.5730910301208496
    74     4.105934e+02     3.304597e+00
 * time: 2.593473196029663
    75     4.105660e+02     9.482860e-01
 * time: 2.613939046859741
    76     4.105529e+02     3.362095e-01
 * time: 2.635197162628174
    77     4.105426e+02     3.015047e-01
 * time: 2.6567859649658203
    78     4.105395e+02     1.804386e-01
 * time: 2.6769001483917236
    79     4.105392e+02     3.818348e-02
 * time: 2.694594144821167
    80     4.105391e+02     4.591111e-02
 * time: 2.711292028427124
    81     4.105391e+02     8.641460e-03
 * time: 2.7278430461883545
    82     4.105391e+02     2.258521e-03
 * time: 2.743338108062744
    83     4.105391e+02     2.645649e-04
 * time: 2.7566750049591064
FittedPumasModel

Successful minimization:                      true

Likelihood approximation:                     FOCE
Likelihood Optimizer:                         BFGS
Dynamical system type:          Matrix exponential

Log-likelihood value:                   -410.53908
Number of subjects:                             31
Number of parameters:         Fixed      Optimized
                                  0              7
Observation records:         Active        Missing
    conc:                       239             47
    Total:                      239             47

-----------------------
             Estimate
-----------------------
pop_CL        0.12758
pop_Vc        8.3876
pop_tabs      0.48983
pop_lag       0.72171
pk_Ω₁,₁       0.010524
σ_prop        0.19957
σ_add         0.55153
-----------------------

2.5 Computing Parameter Precision with infer

The infer function in Pumas estimates the uncertainty (precision) of parameter estimates. Depending on the chosen method, infer can provide standard errors, confidence intervals, and correlation matrices.

The signature for infer often looks like:

infer(
    fpm::FittedPumasModel;
    level = 0.95,
    rethrow_error::Bool = false,
    sandwich_estimator::Bool = true,
)

where:

  • fpm::FittedPumasModel: The result of fit (e.g., foce_fit).
  • level: The confidence interval level (e.g., 0.95). The confidence intervals are calculated as the (1-level)/2 and (1+level)/2 quantiles of the estimated parameters
  • rethrow_error: If rethrow_error is false (the default value), no error will be thrown if the variance-covariance matrix estimator fails. If it is true, an error will be thrown if the estimator fails.
  • sandwich_estimator: Whether to use the sandwich estimator. If set to true (the default value), the sandwich estimator will be used. If set to false, the standard error will be calculated using the inverse of the Hessian matrix.

An example usage:

inference_results = infer(foce_fit; level = 0.95)
[ Info: Calculating: variance-covariance matrix.
[ Info: Done.
Asymptotic inference results using sandwich estimator

Successful minimization:                      true

Likelihood approximation:                     FOCE
Likelihood Optimizer:                         BFGS
Dynamical system type:          Matrix exponential

Log-likelihood value:                   -410.53908
Number of subjects:                             31
Number of parameters:         Fixed      Optimized
                                  0              7
Observation records:         Active        Missing
    conc:                       239             47
    Total:                      239             47

---------------------------------------------------------------------
            Estimate           SE                      95.0% C.I.
---------------------------------------------------------------------
pop_CL       0.12758         0.0046779        [ 0.11842  ; 0.13675 ]
pop_Vc       8.3876          0.22464          [ 7.9474   ; 8.8279  ]
pop_tabs     0.48983         0.17922          [ 0.13856  ; 0.8411  ]
pop_lag      0.72171         0.16464          [ 0.39902  ; 1.0444  ]
pk_Ω₁,₁      0.010524        0.0071909        [-0.0035698; 0.024618]
σ_prop       0.19957         0.044554         [ 0.11225  ; 0.28689 ]
σ_add        0.55153         0.2842           [-0.0054881; 1.1085  ]
---------------------------------------------------------------------

This is the usual asymptotic variance-covariance estimator and we already saw this previous tutorials.

To get a matrix representation of this, use vcov()

vcov(inference_results)
7×7 Symmetric{Float64, Matrix{Float64}}:
  2.1883e-5     0.000332493   5.45952e-6   …   3.30925e-5    0.000379372
  0.000332493   0.050461      0.00233924       0.00117512   -0.00391389
  5.45952e-6    0.00233924    0.032121        -0.000413843   0.00462368
  0.00019032   -0.0071644    -0.0160153       -0.00259753    0.0226855
 -1.83205e-6   -9.94307e-5   -0.000353466     -0.000234415   0.00119
  3.30925e-5    0.00117512   -0.000413843  …   0.00198502   -0.0102392
  0.000379372  -0.00391389    0.00462368      -0.0102392     0.0807681

and to get the condition number of the correlation matrix implied by vcov use

cond(inference_results)
44.47139507390104

Some users request the condition number of the covariance matrix itself and even if the use is often misguided it can be calculated as well.

cond(inference_results; correlation = false)
10854.252604987605

It is also possible to calculate the correlation matrix from the vcov output using the cov2cor function from StatsBase

cor_from_cov = cov2cor(Matrix(vcov(inference_results)))
7×7 Matrix{Float64}:
  1.0          0.31641     0.00651188  …  -0.0544625   0.158779    0.285359
  0.31641      1.0         0.0581035      -0.0615543   0.117415   -0.0613072
  0.00651188   0.0581035   1.0            -0.274263   -0.0518272   0.0907765
  0.247112    -0.193717   -0.542756        0.550551   -0.354113    0.484835
 -0.0544625   -0.0615543  -0.274263        1.0        -0.731675    0.582294
  0.158779     0.117415   -0.0518272   …  -0.731675    1.0        -0.808653
  0.285359    -0.0613072   0.0907765       0.582294   -0.808653    1.0

And we see that the cond call above matches the condition number of the correlation matrix

cond(cor_from_cov)
44.47139507390104

2.5.1 Failure of the asymptotic variance-covariance matrix

It is well-known that the asymptotic variance-covariance matrix can sometimes fail to compute. This can happen for a variety of reasons including:

  1. There are parameters very close to a bound (often 0)
  2. The parameter vector does not represent a local minimum (optimization failed)
  3. The parameter vector does represent a local minimum but it’s not the global solution

The first one is often easy to check. The problematic parameters can be ones than have lower or upper bounds set. Often this will be a variance of standard deviation that has moved very close to the lower boundary. Consider removing the associated random effect if the problematic parameter is a variance in its specification or error model component if a combined additive and proportional error model is used and a standard deviation is close to zero.

It is also possible that the parameters do not represent a local minimum. In other words, they come from a failed fit. In this case, it can often be hard to perform the associated calculations in a stable way, but most importantly the results would not be interpretable even if they can be calculated in this case. The formulas are only valid for parameters that represent the actual (maximum likelihood) estimates. Please try to restart the optimization at different starting points in this case.

If you have reasons to believe that the model should indeed be a combined error model or if the random effect should be present it is also possible that the model converged to a local minimum that is not the global minimum. If the optimization happened to move to a region of the parameter state space that is hard to get out of you will often have to restart the fit at different parameter values. It is not possible to verify if the minimum is global in general, but it is always advised to try out more than one set of initial parameters when fitting models.

2.5.2 Bootstrap

Sometimes it is appropriate to use a different method to calculate estimate the uncertainty of the estimated parameters. Bootstrapping is a very popular approach that is simply but can often come a quite significant computational cost. Researchers often perform a bootstrapping step if their computational budget allows it or if the asymptotic variance-covariance estimator fails. Bootstrapping is advantageous because it does not rely on any invertability of matrices and it cannot produce negative variance confidence intervals because the resampled estimator respects the bounds of the model.

The signature for bootstrapping in infer looks as follows.

infer(fpm::FittedPumasModel, bts::Bootstrap; level = 0.95)

This does not help much before also looking at the interface for Bootstrap itself.

Bootstrap(;
    rng = Random.default_rng,
    samples = 200,
    stratify_by = nothing,
    ensemblealg = EnsembleThreads(),
)

Bootstrap accepts a random number generator rng, the number of resampled datasets to produce samples, if sampling should be stratified according to the covariates in stratify_by, and finally the ensemble algorithm to control parallelization across fits. On the JuliaHub platform this can be used together with distributed computing to perform many resampled estimations in a short time.

bootstrap_results = infer(foce_fit, Bootstrap(samples = 50); level = 0.95)
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Warning: Terminated early due to NaN in gradient.
└ @ Optim ~/run/_work/PumasTutorials.jl/PumasTutorials.jl/custom_julia_depot/packages/Optim/ZhuZN/src/multivariate/optimize/optimize.jl:98
┌ Info: Bootstrap inference finished.
│   Total resampled fits = 50
│   Success rate = 1.0
└   Unique resampled populations = 50
Bootstrap inference results

Successful minimization:                      true

Likelihood approximation:                     FOCE
Likelihood Optimizer:                         BFGS
Dynamical system type:          Matrix exponential

Log-likelihood value:                   -410.53908
Number of subjects:                             31
Number of parameters:         Fixed      Optimized
                                  0              7
Observation records:         Active        Missing
    conc:                       239             47
    Total:                      239             47

--------------------------------------------------------------------
            Estimate          SE                     95.0% C.I.
--------------------------------------------------------------------
pop_CL       0.12758        0.049707       [0.10943    ;  0.14334 ]
pop_Vc       8.3876         1.571          [6.9928     ; 10.244   ]
pop_tabs     0.48983        0.27768        [5.3359e-7  ;  0.86574 ]
pop_lag      0.72171        0.1839         [0.4968     ;  1.0     ]
pk_Ω₁,₁      0.010524       0.011808       [5.215e-79  ;  0.036166]
σ_prop       0.19957        1.769          [0.075233   ;  0.28436 ]
σ_add        0.55153        0.30582        [2.9579e-116;  0.91403 ]
--------------------------------------------------------------------
Successful fits: 50 out of 50
Unique resampled populations: 50
No stratification.

Again, we can calculate a covariance matrix based on the samples with vcov

vcov(bootstrap_results)
7×7 Matrix{Float64}:
  0.00247081   0.0744889    0.00713803   …   0.0869423  -0.00302846
  0.0744889    2.46798      0.204042         2.57599    -0.0730363
  0.00713803   0.204042     0.0771041        0.234532    0.0348872
 -0.00109644  -0.0269498   -0.0336948       -0.0258234  -0.0212816
 -7.65084e-5  -0.00110434   0.000187027     -0.0031925   0.00230502
  0.0869423    2.57599      0.234532     …   3.12943    -0.143379
 -0.00302846  -0.0730363    0.0348872       -0.143379    0.0935244

and we can even get a DataFrame that includes all the estimated parameters from the sampled population fits

DataFrame(bootstrap_results.vcov)
50×7 DataFrame
25 rows omitted
Row pop_CL pop_Vc pop_tabs pop_lag pk_Ω₁,₁ σ_prop σ_add
Float64 Float64 Float64 Float64 Float64 Float64 Float64
1 0.135175 8.39811 0.564577 0.6634 0.00582785 0.225835 0.37769
2 0.133637 8.57208 0.549971 0.951618 0.0254604 0.165505 0.687631
3 0.124322 8.4453 0.455475 0.620348 0.0137868 0.167735 0.583125
4 0.135073 8.43253 0.503238 0.690185 0.0158018 0.147854 0.821073
5 0.124444 8.17174 0.805164 0.488214 2.23345e-10 0.179057 0.682639
6 0.133152 8.22381 0.390986 0.849976 0.0109805 0.185165 0.75736
7 0.132996 8.31309 0.622861 0.75064 0.00689214 0.153262 0.837761
8 0.144336 10.6512 0.121373 0.916338 1.28494e-106 0.173561 2.7241e-129
9 0.124389 8.25699 0.661485 0.691952 7.4399e-10 0.164051 0.756688
10 0.114851 8.26133 0.0943829 1.4947 0.0284981 0.0742399 0.642479
11 0.133158 8.19032 0.584724 0.881794 0.0103163 0.182199 0.654715
12 0.133837 8.50587 0.465672 0.676962 0.00638902 0.182697 0.535889
13 0.127928 8.2168 0.656253 0.800147 0.0275024 0.134609 0.79723
39 0.128647 8.21614 0.470363 0.684001 0.0155254 0.163668 0.788704
40 0.134979 8.58739 0.364157 0.881497 0.0369369 0.124792 0.935501
41 0.130251 8.13023 0.46527 0.717524 0.0157143 0.162017 0.610404
42 0.124927 8.7531 0.525946 0.561603 0.0164374 0.158832 0.705722
43 0.122152 8.49019 4.15059e-9 0.999996 6.77691e-135 0.241843 3.07323e-83
44 0.12862 8.50102 0.569731 0.715056 6.66497e-10 0.188884 0.58755
45 0.132123 8.19804 0.391365 0.777406 0.013773 0.182285 0.763217
46 0.125933 8.22744 0.432347 0.586395 0.0292002 0.120257 0.834049
47 0.131828 8.42803 0.488984 0.720136 0.00754987 0.215658 0.687501
48 0.12945 8.0876 0.237585 0.760896 0.00726597 0.186104 0.628473
49 0.129102 8.61346 0.572642 0.690742 0.00539275 0.17699 0.471174
50 0.126611 8.39068 0.490131 0.663827 0.0100943 0.115705 0.652958

This is very useful for histogram plotting of parameter distributions.

2.5.3 Sampling Importance Re-sampling

Pumas has support for inference through Sampling Importance Re-sampling through the SIR() input to infer. The signature for SIR in infer looks as follows.

infer(fpm::FittedPumasModel, sir::SIR; level = 0.95, ensemblealg = EnsembleThreads())

This performs sampling importance re-sampling for the population in fpm. The confidence intervals are calculated as the (1-level)/2 and (1+level)/2 quantiles of the sampled parameters. ensemblealg can be EnsembleThreads() (the default value) to use multi-threading or EnsembleSerial() to use a single thread.

The signature for the SIR specification is

SIR(; rng, samples, resamples)

SIR accepts a random number generator rng, the number of samples from the proposal, samples, can be set and to complete the specification the resample has to be set. It is suggested that samples is at least 5 times larger than resamples in practice to have sufficient samples to resample from.

sir_results = infer(foce_fit, SIR(samples = 1000, resamples = 200); level = 0.95)
[ Info: Calculating: variance-covariance matrix.
[ Info: Done.
[ Info: Running SIR.
[ Info: Resampling.
Simulated inference results

Successful minimization:                      true

Likelihood approximation:                     FOCE
Likelihood Optimizer:                         BFGS
Dynamical system type:          Matrix exponential

Log-likelihood value:                   -410.53908
Number of subjects:                             31
Number of parameters:         Fixed      Optimized
                                  0              7
Observation records:         Active        Missing
    conc:                       239             47
    Total:                      239             47

-------------------------------------------------------------------
            Estimate           SE                     95.0% C.I.
-------------------------------------------------------------------
pop_CL       0.12758         0.0028963        [0.12197  ; 0.1333 ]
pop_Vc       8.3876          0.24066          [7.9332   ; 8.8405 ]
pop_tabs     0.48983         0.13264          [0.24441  ; 0.74641]
pop_lag      0.72171         0.091714         [0.50355  ; 0.84321]
pk_Ω₁,₁      0.010524        0.0053638        [0.0012933; 0.02113]
σ_prop       0.19957         0.028962         [0.16231  ; 0.27523]
σ_add        0.55153         0.20917          [0.039908 ; 0.86033]
-------------------------------------------------------------------

Notice, that SIR bases its first samples number of samples from a truncated multivariate normal distribution with mean of the maximum likelihood population level parameters and covariance matrix that is the asymptotic matrix calculated by infer(fpm). This means that to use SIR the matrix is question has to be successfully calculated by infer(fpm) under the hood.

The methods for vcov and DataFrame(sir_results.vcov) that we saw for Bootstrap also applies here

vcov(sir_results)
7×7 Matrix{Float64}:
 8.38868e-6    0.000132666   8.81323e-8  …   5.7814e-7     0.000100732
 0.000132666   0.0579191    -0.00505429     -0.000203157  -0.00385386
 8.81323e-8   -0.00505429    0.0175931      -0.00107094    0.00727587
 3.68159e-5    0.000465673  -0.0080209      -0.000299421   0.00310756
 2.09475e-6    0.000235207   2.033e-5       -8.62468e-5    0.000535516
 5.7814e-7    -0.000203157  -0.00107094  …   0.000838786  -0.00521685
 0.000100732  -0.00385386    0.00727587     -0.00521685    0.0437531

and

DataFrame(sir_results.vcov)
200×7 DataFrame
175 rows omitted
Row pop_CL pop_Vc pop_tabs pop_lag pk_Ω₁,₁ σ_prop σ_add
Float64 Float64 Float64 Float64 Float64 Float64 Float64
1 0.127571 8.65869 0.19149 0.808368 0.00674417 0.260674 0.0424837
2 0.126884 8.31607 0.532419 0.619936 0.00300365 0.255596 0.1466
3 0.12869 8.32958 0.638063 0.588227 0.0035923 0.260376 0.0255513
4 0.128182 7.91908 0.569104 0.761485 0.016481 0.1901 0.684522
5 0.128051 8.61148 0.290952 0.870127 0.0200761 0.181501 0.61098
6 0.128252 8.45581 0.319999 0.763224 0.0021016 0.263796 0.0400501
7 0.125095 8.51985 0.478599 0.657732 0.0115755 0.181019 0.557215
8 0.126083 8.31458 0.509982 0.576583 0.00873988 0.223252 0.302577
9 0.126005 8.5038 0.456142 0.703952 0.00602465 0.24008 0.252261
10 0.131828 8.49869 0.445267 0.680906 0.00753632 0.208978 0.627915
11 0.126417 8.66168 0.673703 0.541912 0.0112181 0.191789 0.570906
12 0.12767 8.86345 0.451972 0.494212 0.00532104 0.227752 0.399223
13 0.126427 8.56067 0.47573 0.522138 0.0190835 0.201154 0.507841
189 0.128254 8.38625 0.503374 0.674967 0.00822299 0.210554 0.332138
190 0.128723 8.46931 0.746122 0.491684 0.00894756 0.203937 0.393635
191 0.129779 8.34794 0.504504 0.839448 0.0220383 0.160892 0.760316
192 0.131045 8.55697 0.546378 0.707979 0.0132994 0.152927 0.822215
193 0.122134 8.23774 0.489023 0.529574 0.00290233 0.236042 0.0681814
194 0.124065 8.23777 0.757667 0.494929 0.0106029 0.189046 0.614687
195 0.126841 8.26098 0.694462 0.703826 0.015406 0.165526 0.793722
196 0.124526 7.89228 0.506947 0.717785 0.00558366 0.268672 0.332431
197 0.124948 8.28127 0.379557 0.777262 0.00774241 0.186265 0.597754
198 0.129879 8.05742 0.616092 0.775751 0.0142323 0.166636 0.906107
199 0.128529 8.30409 0.62777 0.612318 0.00453657 0.247357 0.331663
200 0.129403 8.20053 0.740329 0.608549 0.00480835 0.207861 0.684651

2.5.4 Marginal MCMC

An alternative to Bootstrap and SIR is to simply use the MarginalMCMC sampler which is a Hamiltonian Monte Carlo (HMC) No-U-Turn Sampler (NUTS) that will sample from the marginal loglikelihood. This means that individual effects are marginalized out and then we sample the population level parameters. This does not resample populations like Bootstrap so inference may be more stable if many resampled populations lead to extreme estimates and it differs from SIR in that it does not need the asymptotic covariance matrix to be calculated and sampled from.

This method requires slightly more understanding from the user when setting the options that can be found through the docstring of MarginalMCMC. Some knowledge of Bayesian inference is advised.

inference_results = infer(foce_fit, MarginalMCMC(); level = 0.95)

As sampling based inference can be computationally intensive we exclude the actual invocation of this method from this tutorial.

3 Concluding Remarks

This tutorial showcased a typical Pumas workflow for parameter inference in Pumas models. We showed the different methods supported by Pumas for calculating parameter uncertainty.