Inference

In Pumas, inference is performed after model fitting to obtain uncertainty estimates on the model's population-level parameters (fixed effects), which can then be used to derive standard errors and confidence intervals. This is accomplished by calling the infer function on the object returned by fit:

Pumas.inferFunction
infer(fpm::FittedPumasModel; level=0.95, rethrow_error::Bool=false, sandwich_estimator::Bool=true) -> FittedPumasModelInference

Compute the vcov matrix and return a struct used for inference based on the fitted model fpm. The confidence intervals are calculated as the (1-level)/2 and (1+level)/2 quantiles of the estimated parameters. sandwich_estimator is a boolean that switches on or off the sandwich estimator. 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.

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

Perform bootstrapping by resampling the Subjects from the Population stored in fpm. The confidence intervals are calculated as the (1-level)/2 and (1+level)/2 quantiles of the estimated parameters. The number of samples used in the bootstrapping is bts.samples. bts.ensemblealg specifies the ensemblealg used here. If ensemblealg is EnsembleSerial(), a single thread will be used. If it is EnsembleThreads() (the default value), multiple threads will be used. See the documentation of Bootstrap for more details on constructing an instance of Bootstrap.

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

Perform sampling importance re-sampling for the Subjects from the Population stored in fpm. The confidence intervals are calculated as the (1-level)/2 and (1+level)/2quantiles of the estimated parameters.ensemblealgcan beEnsembleThreads()(the default value) to use multi-threading orEnsembleSerial() to use a single thread.

The function can be specialized to behave differently by setting the second (optional) argument type to different values. Details of each specialization are given below.

The infer function accepts the following arguments:

• fpm: a fitted Pumas model, typically returned by fit.
• type: the second (optional) argument determines the inferential approach used. Details are given below.
• level: for confidence intervals, the confidence level to use. Typically, 0.95, corresponding to 95% confidence intervals.
• ...: further arguments, the meaning of which is dependent on the type of inference. Details are given below.

The function returns an object of type FittedPumasModelInference. Further functions can be called on this object, such as stderror to obtain the standard errors.

We now describe in more detail the different inferential approaches that can be used by specifying the type argument. The default is to use the asymptotic variance-covariance matrix obtained from the "robust" (also called "sandwich") estimator [8]. Other options include:

Inference from asymptotic variance-covariance matrices

This inferential approach is the default when no type is specified in infer:

infer(fpm)

• rethrow_error: whether to throw an error when it is not possible to compute the variance-covariance matrix (e.g., because the observed information matrix is not numerically positive definite). The default is to return a failed FittedPumasModelInference object but not throw an error.

• sandwich_estimator: whether to use the "robust" variance-covariance estimator [8]. true to default. If false, then the inverse of the observed Fisher information (i.e., negative inverse Hessian of the log-likelihood) is used.

Note that the variance-covariance matrix can also be obtained directly by calling the vcov function directly on a fitted Pumas model (returned by fit).

Here is an example of obtaining confidence intervals by setting sandwich_estimator to false in order to not use the robust estimator:

my_infer = infer(fpm; sandwich_estimator = false)

The FittedPumasModelInference returned by infer has a show method that prints the estimates, standard errors and confidence intervals in the REPL. To extract this information in a DataFrame (e.g. for markdown reporting), the coeftable function can be used:

StatsAPI.coeftableFunction
coeftable(fpm::FittedPumasModel) -> DataFrame

Construct a DataFrame of parameter names and estimates from fpm.

coeftable(cfpm::Vector{<:FittedPumasModel}) -> DataFrame

Construct a DataFrame of parameter names and estimates and their standard deviation from vector of fitted single-subject models vfpm.

coeftable(pmi::FittedPumasModelInference; level = pmi.level) -> DataFrame

Construct a DataFrame of parameter names, estimates, standard error, and confidence interval with confidence level level. The default value of level is the one that was passed in infer.

Additionally, you can also use the correlation_diagnostic to print a list of parameter pairs with high, medium or low correlations:

Pumas.correlation_diagnosticFunction
correlation_diagnostic(
pmi::FittedPumasModelInference;
high_cor_threshold = 0.7,
medium_cor_threshold = high_cor_threshold / 2,
)

A function which accepts the output of the infer function as its argument and prints the lists of parameter pairs with high, medium or low correlations. Parameter pairs with high correlation are those with a correlation higher than high_cor_threshold. Parameter pairs with a low correlation are those with a correlation lower than medium_cor_threshold. The remaining parameter pairs have a medium correlation. Assume fpm is the output of fit. The following is an example of how to use correlation_diagnostic:

fpmi = infer(fpm)
correlation_diagnostic(fpmi)

Inference using the robust variance-covariance estimator is consistent under heteroscedasticity of residuals, but relies on asymptotic normality of the maximum likelihood estimator. While this is often a reasonable approximation for the parameters associated with the mean of the response variable, it is often a poor approximation for the variance parameters. Furthermore, confidence intervals derived using this approach do not necessarily respects bounds on the parameters, and it is not unlikely for a parameter that by definition is strictly positive (such as a clearance, volume or variance) to have an associated confidence interval that contains negative values. For that reason, inference based on other approaches, such as Bootstrap and SIR, may be preferable in some situations.

Bootstrap-based inference

Bootstrap-based inference is obtained by calling the function Bootstrap and passing the returned object of type Bootstrap as the type argument of infer. The resulting function signature is:

infer(fpm, Bootstrap())

There are no additional keyword arguments to the infer function, but Bootstrap has a number of optional arguments that can be used to control the bootstrap procedure. It has the following signature:

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

Constructs a Bootstrap object that allows the user to use bootstrapping with the supplies setting in infer function calls. See ?infer and https://docs.pumas.ai/stable/analysis/inference/ for more information about obtaining inference results.

The keyword arguments are as follows:

• rng: used to select a pseudo-random number generator to beused for the random resamples of the datasets.
• samples: controls the number of resampled Populations.
• resamples: the number of model parameter samples from the original sample distribution to re-sample (by weighed re-sampling without replacement). The number of resamples must be less than that of the original sample. An original sample-to-resample ratio of at least 5:1 is suggested.
• stratify_by control the stratification used in the re-sampling scheme. Should be a set to a Symbol with the name of a covariate it is possible to stratify by a covariate with a finite number of possible values (e.g :study for a covariate named study).
• ensemblealg controls the parallization across fit calls. Each fit is fitted using the ensemblealg that was used in the original fit.

The type of bootstrap procedure used is nonparametric bootstrap, also called paired bootstrap [9]. Bootstrap datasets are obtained by resampling individuals from the original dataset with replacement (all records associated with the sampled individual are included), after stratifying by any covariates specified. The stratification ensures that the number of individuals in each stratum remains the same in each resampled dataset, which can help with the stability of the estimator. For example, if an analysis consists of two studies, a small study with rich sampling, and a large study with sparse sampling, stratifying on the study will ensure that each resampled dataset has the same number of subjects with rich sampling. The fitting algorithm is applied to each resampled dataset to obtain $N$ estimates for each parameter, where $N$ is the number of replicates specified in samples. This collection is treated as a random draw from the sampling distribution of the estimator for the parameters, and a nonparametric 95% confidence interval is obtained from the 2.5th and 97.5th quantiles of this sample.

Inference based on the nonparametric bootstrap approach is often considered more reliable than using the asymptotic variance-covariance matrix, but is computationally intensive. Unlike the latter approach, bounds on the parameters are respected (e.g. the confidence interval for a strictly positive parameter cannot contain negative values).

Here's an example with all options specified:

using Random

my_bootstrap = infer(
fpm,
Bootstrap(;
rn = MersenneTwister(1234),
sample = 500,
stratify_by = [:sex, :trt],
),
)

The coeftable function will work on the returned object, and the confidence interval and standard errors will be bootstrap-based.

The bootstrap requires a FittedPumasModel; FittedPumasEMModel support is coming in a future version of Pumas.

Sampling importance resampling (SIR)

As an alternative to the nonparametric bootstrap procedures (which can be computationally intensive or unfeasible with small sample sizes), [10] suggests the sampling importance resampling (SIR) procedure for estimating the uncertainty. This can be obtained by calling the function SIR and passing the returned object of type SIR as the type argument (second optional positional argument) of infer:

Pumas.SIRType
SIR(; rng=Random.default_rng, samples, resamples)

Constructs an SIR object that allows the user to use sampling importance re-sampling (SIR) with the supplies settings in infer function calls. See ?infer and https://docs.pumas.ai/stable/analysis/inference/ for more information about obtaining inference results.

The keyword arguments are as follows: • rng: used to select a pseudo-random number generator to beused within the SIR procedure. Requires loading Random package (using Random) • samples: the number of model parameter samples simulated from a truncated multivariate normal distribution (proposal distribution) based on the maximum-likelihood estimates and robust variance-covariance matrix. This constitutes the original sample (uncertainty) distribution. • resamples: the number of model parameter samples from the original sample distribution to re-sample (by weighed re-sampling without replacement). The number of resamples must be less than that of the original sample. An original sample-to-resample ratio of at least 5:1 is suggested.

Both samples and resamples must be supplied, as there are no defaults.

Note that the proposal distribution for the SIR procedure is based on the robust variance-covariance matrix and will therefore fail whenever the computation of the robust variance-covariance matrix fails. Confidence intervals based on SIR respect bounds on the parameters (e.g. the confidence interval for a strictly positive parameter cannot contain negative values).

Confidence intervals based on SIR respect bounds on the parameters (e.g. the confidence interval for a strictly positive parameter cannot contain negative values).

Here is an example of using how to use SIR in Pumas:

my_sir = infer(my_fit, SIR(; samples = 200, resamples = 70))

The coeftable function will work on the returned object, and the confidence interval and standard errors will be based on SIR.

Note that if the robust variance-covariance matrix has already been computed by a call to infer, it is possible to avoid recomputing it by calling infer again with the output of the first call to infer as the first argument, and the result from SIR as the second argument. For example:

my_inference = infer(fpm)
my_sir = infer(my_inference, SIR(; samples = 300, resamples = 100))

SIR requires a FittedPumasModel FittedPumasEMModel support is coming in a future version of Pumas.

Calculating mean values and variance-covariance matrices from Bootstrap and SIR

With the output from an infer call using Bootstrap or SIR in hand, it is possible to get mean values of the sampled parameters and a variance-covariance matrix based on the same. If the output of the infer call was named infer_output, these quantities are simply produced using the following code:

mean(infer_output)
vcov(infer_output)

Extracting standard errors (stderror)

The stderror function is used to extract standard errors from either a FittedPumasModel object returned by fit or a FittedPumasModelInference object returned by infer:

StatsAPI.stderrorFunction
stderror(f::AbstractFittedPumasModel) -> NamedTuple

Compute the standard errors of the population parameters and return the result as a NamedTuple matching the NamedTuple of population parameters.

When calling stderror on a infer object, the type of standard errors returned will depend on the options passed to infer. For instance, if nonparametric bootstrap inference was performed, then the standard errors will be bootstrap-based.

Calling stderror directly on a fitted Pumas model object will result in using the default arguments to infer, i.e. standard errors based on the robust variance-covariance matrix. They can be inefficient because it may result in the variance-covariance matrix being re-calculated unnecessarily it is therefore recommended to not call stderror directly on a fitted Pumas model object, but to call infer first and store the object, which is then passed to stderror`.