Bayesian Multiple Imputation for Functional Data

arXiv:2605.08018 · stat.ME · Submitted 2026-05-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Bayesian Multiple Imputation for Functional Data".

Jane: Here's the gist of the paper. Bayesian multiple imputation for functional data develops BAMIFun to address the limitations of single-imputation methods by explicitly incorporating estimation uncertainty into downstream analyses The BAMIFun framework provides a principled approach to reconstructing incomplete trajectories while quantifying the uncertainty arising from the imputation Framework Development The BAMIFun algorithm is a Bayesian multiple imputation framework for functional data that embeds penalized splines within a Bayesian framework to enforce trajectory smoothness It performs multiple imputation for single-level functional data via a low-rank representation and achieves valid downstream inference by pooling estimates using Rubin’s rules Furthermore, it extends the framework to multiway functional data using Functional Tensor Singular Value Decomposition (FTSVD) model, enabling Bayesian multiple imputation in settings not supported by existing methods Single-Level Data Modeling For single-level functional data, the model assumes a low-rank structure X = VUT + E where V is an N×R matrix, U is a K×R matrix, and E is an N×K matrix of i.i.d.

Tom: Gaussian error terms The functional data X is assumed to be smooth along the functional domain I by the low-rank factor matrix U, represented as UT = BΘ where Θ is an L × K matrix of spline basis functions The additional smoothness constraint on U can be induced by adding a penalty on the integral of the squared second derivative R(u''r(t)) squared dt Inference and Uncertainty Quantification For downstream analysis, BAMIFun constructs the 95% credible interval for functional observations as

Xˆ L i(tk), Xˆ U i(tk): where Xˆ U i(tk) is the 0.975 quantile of the posterior samples The point estimate β̂(t) for the functional coefficient can be obtained by pooling estimates from separate SoFR models using Rubin’s rule to construct the total variance VˆT(t) = VˆW(t) + (1+ 1/S)VˆB(t) Multiway Data Extension For multiway functional data, BAMIFun adopts the FTSVD model which approximates the multiway functional data through the decomposition X = UΛ(W ⊙ V) T + E The Gibbs sampler for multiway BAMIFun under (7) is presented in the Supplementary Material Section S5.2 Simulation and Validation Simulation studies show that BAMIFun achieves substantially improved coverage and more reliable downstream inference compared to existing methods, while maintaining similar imputation accuracy Case studies using a physical activity dataset and an infant gut microbiome dataset further demonstrate the practical advantages of our proposed methods under severe missingness The results show that BAMIFun consistently outperforms BAMITA in terms of imputation accuracy across all settings and maintains coverage close to the nominal level throughout (93.3%–94.0% at N = 100 and 96.8%–98.8% at N = 300) Computational Cost BAMIFun is more expensive than frequentist FPCA imputation, ranging from roughly 40 seconds at N = 100 to about 7 minutes at N = 1,000 in the single-level simulation In the multiway simulation it ranges from roughly 2 minutes at N = 100 to about 18 minutes at N = 1,000 reflecting the additional cost of its penalized-spline smoothness updates within the Gibbs sampler Case Study Results In the NHANES data case study, BAMIFun provides empirical coverage close to the nominal level for the estimated 95% credible intervals across all levels of missingness The pooled β̄(t) closely tracks the oracle coefficient fit to the unmasked profiles even when 90–97.5% of each subject’s curve is unobserved Discussion BAMIFun gains substantially improved coverage and valid downstream inference, as confirmed in our simulations and applications For multiway functional data, the performance instead depends on the structure, as BAMIFun is more accurate when a low-rank structure is present It attains close-to-nominal coverage and remains applicable under extreme missingness, whereas existing MFPCA software cannot run References Chiou, J.-M., Chen, Y.-T.

Jane: and Yang, Y.-F.

Tom: (2014), ‘Multivariate functional principal component analysis: A normalization approach’, Statistica Sinica pp.

Title and authors: Tom: So we're looking at this paper called Bayesian Multiple Imputation for Functional Data by Ziren Jiang, Lei Xuan, Eric F. Lock, and Erjia Cui. Jane, what’s the main thing this title tells us right off the bat?

Jane: It’s basically saying they are using a Bayesian approach to fill in missing data when that data is functional—that means it’s a curve or a time series—and they are doing multiple imputation, which is filling in those gaps.

Lu: And I think the key word here is "Bayesian." That means instead of just guessing the best fit, they’re incorporating uncertainty into their guesses right from the start.

Meng: Uncertainty sounds nice in theory, but for production AI, we need to know what breaks when we use those guesses. Is this method stable at scale?

Lalam: I think it addresses a big problem because existing methods often rely on single imputation, which just gives you one possible answer and makes everything else look too good.

Tom: Exactly. So, they are tackling the limitations of those older single imputation methods by building something new to handle missing trajectories in functional data.

The paper's summary: Jane: Okay, so what exactly is BAMIFun? It sounds like a framework designed specifically for this problem. Can you explain the core idea behind the BAMIFun framework in simple terms?

Lu: Well, it sets up a Bayesian low-rank model for single-level functional data and puts penalized splines inside that Bayesian structure to force smoothness on the trajectory.

Tom: For someone listening, that just means they’re not letting the imputed curve be wiggly or jagged; they are using mathematical constraints to keep the shape of the missing part realistic.

Meng: So, what about how many different curves can this handle? Because if we have complex data, we need something more than just one level.

Lalam: They actually extend it beyond single-level data to multiway functional data by using something called Functional Tensor Singular Value Decomposition, or FTSVD model.

Jane: FTSVD is the tool they use for multiway data, which means if you have observations indexed over multiple things—like subject and visit time points—this method can handle it where other tools fail.

Tom: It’s a big jump from just one curve to handling entire data structures that are much more complex than simple single trajectories.

The paper's improvements: Lu: The main improvement they point out is how they handle uncertainty in the final analysis, which is crucial. They construct a ninety-five percent credible interval for those functional observations themselves.

Jane: That’s smart because it means you don't just get one line; you get a range of possibilities for that curve, which gives much more reliable inference later on.

Meng: So, if I use these results to train a prediction model, the uncertainty quantification has to be solid. How do they make sure those confidence intervals are actually accurate?

Tom: They achieve this by pooling estimates from different imputation runs using Rubin’s rules, which gives you a total variance that accounts for both the missing data within one imputation and the differences between the imputations themselves.

Lalam: That total variance formula they use is V sub hat t equals V sub hat w t plus one over one plus S times V sub hat b t. It’s designed to properly combine those two sources of uncertainty into a single, sensible measure.

Jane: So, the improvement isn't just filling in the blanks; it’s giving you mathematically sound confidence intervals for whatever you want to do next with that data.

Conclusion: Tom: So we’ve seen how BAMIFun uses Bayesian methods and penalized splines to handle missing functional data and how they use Rubin's rules to give us proper uncertainty quantification. Jane, what’s the final word on what this paper means for people who don't read statistics?

Jane: It means that when you are working with missing curves, you can move past the overly optimistic guesses of single imputation and get inferences that are much more reliable because they properly account for all the ways uncertainty enters your analysis.

Lu: I think what’s really interesting is how it scales. They show it can handle multiway data using FTSVD, which opens up possibilities for complex datasets in fields like genomics that have multiple dimensions of observation.

Meng: From a deployment standpoint, the computational cost is something we have to watch. The simulation showed that the cost goes up pretty fast with sample size N, reaching about eighteen minutes at one thousand samples for multiway data. That’s a factor we need to consider when thinking about running this on live traffic.

Lalam: I just want to say that they show empirical validation too. In their case studies on physical activity and gut microbiome data, BAMIFun outperformed another method called BAMITA in terms of accuracy across all settings and kept the coverage close to what was expected, hitting between ninety-three point three percent and ninety-four point zero percent at one hundred samples.

Tom: So it’s a solid piece of work that shows a principled way to reconstruct incomplete trajectories while quantifying the uncertainty behind those reconstructions. That’s what this paper is all about with Bayesian multiple imputation for functional data.

University of Minnesota

stat.ME

Submitted: 2026-05-08

Updated: 2026-10-30

Comments: 2 Tables, 3 Figures

Code: https://github.com/ZirenJiang/BAMIFun

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: Bayesian multiple imputation for functional data develops BAMIFun to address the limitations of single-imputation methods by explicitly incorporating estimation uncertainty into downstream analyses…

Key concepts

Bayesian Multiple Imputation
This technique fills in missing data multiple times using a probabilistic framework. Instead of one guess, it generates several plausible datasets based on prior beliefs and then pools the results to account for the uncertainty inherent in those guesses.
Penalized Splines
These are mathematical functions used to model the smooth shape of functional data. By adding a penalty term related to the second derivative, the method ensures that imputed trajectories are not only plausible but also adhere to a desired level of smoothness, preventing unrealistic jumps in the data.
Functional Tensor Singular Value Decomposition (FTSVD)
This is an advanced mathematical model used to handle complex multiway functional data. It decomposes the high-dimensional data into simpler components, allowing the imputation process to work effectively even when observations are indexed across multiple dimensions, such as different subjects and time points.

Terminology

Summary

Bayesian multiple imputation for functional data develops BAMIFun to address the limitations of single-imputation methods by explicitly incorporating estimation uncertainty into downstream analyses The BAMIFun framework provides a principled approach to reconstructing incomplete trajectories while quantifying the uncertainty arising from the imputation

Framework Development

The BAMIFun algorithm is a Bayesian multiple imputation framework for functional data that embeds penalized splines within a Bayesian framework to enforce trajectory smoothness It performs multiple imputation for single-level functional data via a low-rank representation and achieves valid downstream inference by pooling estimates using Rubin’s rules Furthermore, it extends the framework to multiway functional data using Functional Tensor Singular Value Decomposition (FTSVD) model, enabling Bayesian multiple imputation in settings not supported by existing methods

Single-Level Data Modeling

For single-level functional data, the model assumes a low-rank structure X = VUT + E where V is an N×R matrix, U is a K×R matrix, and E is an N×K matrix of i.i.d. Gaussian error terms The functional data X is assumed to be smooth along the functional domain I by the low-rank factor matrix U, represented as UT = BΘ where Θ is an L × K matrix of spline basis functions The additional smoothness constraint on U can be induced by adding a penalty on the integral of the squared second derivative R(u''r(t)) squared dt

Inference and Uncertainty Quantification

For downstream analysis, BAMIFun constructs the 95% credible interval for functional observations as [Xˆ L i(tk), Xˆ U i(tk)] where Xˆ U i(tk) is the 0.975 quantile of the posterior samples The point estimate β̂(t) for the functional coefficient can be obtained by pooling estimates from separate SoFR models using Rubin’s rule to construct the total variance VˆT(t) = VˆW(t) + (1+ 1/S)VˆB(t)

Multiway Data Extension

For multiway functional data, BAMIFun adopts the FTSVD model which approximates the multiway functional data through the decomposition X = UΛ(W ⊙ V) T + E The Gibbs sampler for multiway BAMIFun under (7) is presented in the Supplementary Material Section S5.2

Simulation and Validation

Simulation studies show that BAMIFun achieves substantially improved coverage and more reliable downstream inference compared to existing methods, while maintaining similar imputation accuracy Case studies using a physical activity dataset and an infant gut microbiome dataset further demonstrate the practical advantages of our proposed methods under severe missingness The results show that BAMIFun consistently outperforms BAMITA in terms of imputation accuracy across all settings and maintains coverage close to the nominal level throughout (93.3%–94.0% at N = 100 and 96.8%–98.8% at N = 300)

Computational Cost

BAMIFun is more expensive than frequentist FPCA imputation, ranging from roughly 40 seconds at N = 100 to about 7 minutes at N = 1,000 in the single-level simulation In the multiway simulation it ranges from roughly 2 minutes at N = 100 to about 18 minutes at N = 1,000 reflecting the additional cost of its penalized-spline smoothness updates within the Gibbs sampler

Case Study Results

In the NHANES data case study, BAMIFun provides empirical coverage close to the nominal level for the estimated 95% credible intervals across all levels of missingness The pooled β̄(t) closely tracks the oracle coefficient fit to the unmasked profiles even when 90–97.5% of each subject’s curve is unobserved

Discussion

BAMIFun gains substantially improved coverage and valid downstream inference, as confirmed in our simulations and applications For multiway functional data, the performance instead depends on the structure, as BAMIFun is more accurate when a low-rank structure is present It attains close-to-nominal coverage and remains applicable under extreme missingness, whereas existing MFPCA software cannot run

References

Chiou, J.-M., Chen, Y.-T. and Yang, Y.-F. (2014), ‘Multivariate functional principal component analysis: A normalization approach’, Statistica Sinica pp. p. 0

Where the paper says it

Lock, Erjia Cui∗ Division of Biostatistics, University of Minnesota Abstract Missing data are pervasive in modern functional datasets, where trajectories are often sparsely or irregularly observed. Although Functional Principal Component Analysis (FPCA) is widely used to reconstruct incomplete curves, existing approaches typically employ single imputation, leading to overly optimistic inferences in downstream analyses.

1571–1596

Cong, X., Judge, M., Xu, W., Diallo, A., Janton, S., Brownell, E. A., Maas, K. and Graf, J. (2017), ‘Influence of feeding type on gut microbiome development in hospitalized preterm infants’, Nursing research 66(2), 123–133

Crainiceanu, C. M., Goldsmith, J., Leroux, A. and Cui, E. p. 0

Where the paper says it

BAMIFun: Bayesian Multiple Imputation for Functional Data Ziren Jiang, Lei Xuan, Eric F. Lock, Erjia Cui∗ Division of Biostatistics, University of Minnesota Abstract Missing data are pervasive in modern functional datasets, where trajectories are often sparsely or irregularly observed.

p. 0
Where the paper says it

BAMIFun: Bayesian Multiple Imputation for Functional Data Ziren Jiang, Lei Xuan, Eric F. Lock, Erjia Cui∗ Division of Biostatistics, University of Minnesota Abstract Missing data are pervasive in modern functional datasets, where trajectories are often sparsely or irregularly observed.

(2024), Functional data analysis with R, Chapman and Hall/CRC

Cui, E., Crainiceanu, C. p. 5

Where the paper says it

Under the spline basis expansion UT = BΘ, we have Z (u ′′ r (t))2 dt = Z X L l=1 BrlΘ′′ l· (t) 2 dt = Z X L l1=1 X L l2=1 Brl1Θ′′ l1· (t) Θ′′ l2· (t) Brl2 dt = B T r· Z Θ′′(t)[Θ′′(t)]T dt Br· = B T r·PBr· , (2) where Br· is the r-th row of matrix B, Θl·(t) is the l-th spline basis, and P is the L×L penalty matrix with the element of l1-th row and l2-th column being Pl1l2 = R [Θ′′ l1· (t)]T [Θ′′ l2· (t)]dt. As shown in Jiang, Crainiceanu and Cui (2025), incorporating the penalty R (u ′′ r (t))2dt into the log-likelihood is equivalent to imposing a multivariate normal prior p(Br·) ∝ exp − BT r·PBr· 2σ 2 B on each row of the spline coefficients matrix B, where σ 2 B is the smoothing parameter.

p. 5
Where the paper says it

Under the spline basis expansion UT = BΘ, we have Z (u ′′ r (t))2 dt = Z X L l=1 BrlΘ′′ l· (t) 2 dt = Z X L l1=1 X L l2=1 Brl1Θ′′ l1· (t) Θ′′ l2· (t) Brl2 dt = B T r· Z Θ′′(t)[Θ′′(t)]T dt Br· = B T r·PBr· , (2) where Br· is the r-th row of matrix B, Θl·(t) is the l-th spline basis, and P is the L×L penalty matrix with the element of l1-th row and l2-th column being Pl1l2 = R [Θ′′ l1· (t)]T [Θ′′ l2· (t)]dt. As shown in Jiang, Crainiceanu and Cui (2025), incorporating the penalty R (u ′′ r (t))2dt into the log-likelihood is equivalent to imposing a multivariate normal prior p(Br·) ∝ exp − BT r·PBr· 2σ 2 B on each row of the spline coefficients matrix B, where σ 2 B is the smoothing parameter.

p. 5
Where the paper says it

Under the spline basis expansion UT = BΘ, we have Z (u ′′ r (t))2 dt = Z X L l=1 BrlΘ′′ l· (t) 2 dt = Z X L l1=1 X L l2=1 Brl1Θ′′ l1· (t) Θ′′ l2· (t) Brl2 dt = B T r· Z Θ′′(t)[Θ′′(t)]T dt Br· = B T r·PBr· , (2) where Br· is the r-th row of matrix B, Θl·(t) is the l-th spline basis, and P is the L×L penalty matrix with the element of l1-th row and l2-th column being Pl1l2 = R [Θ′′ l1· (t)]T [Θ′′ l2· (t)]dt. As shown in Jiang, Crainiceanu and Cui (2025), incorporating the penalty R (u ′′ r (t))2dt into the log-likelihood is equivalent to imposing a multivariate normal prior p(Br·) ∝ exp − BT r·PBr· 2σ 2 B on each row of the spline coefficients matrix B, where σ 2 B is the smoothing parameter.

p. 5
Where the paper says it

Under the spline basis expansion UT = BΘ, we have Z (u ′′ r (t))2 dt = Z X L l=1 BrlΘ′′ l· (t) 2 dt = Z X L l1=1 X L l2=1 Brl1Θ′′ l1· (t) Θ′′ l2· (t) Brl2 dt = B T r· Z Θ′′(t)[Θ′′(t)]T dt Br· = B T r·PBr· , (2) where Br· is the r-th row of matrix B, Θl·(t) is the l-th spline basis, and P is the L×L penalty matrix with the element of l1-th row and l2-th column being Pl1l2 = R [Θ′′ l1· (t)]T [Θ′′ l2· (t)]dt. As shown in Jiang, Crainiceanu and Cui (2025), incorporating the penalty R (u ′′ r (t))2dt into the log-likelihood is equivalent to imposing a multivariate normal prior p(Br·) ∝ exp − BT r·PBr· 2σ 2 B on each row of the spline coefficients matrix B, where σ 2 B is the smoothing parameter.

p. 5
Where the paper says it

Under the spline basis expansion UT = BΘ, we have Z (u ′′ r (t))2 dt = Z X L l=1 BrlΘ′′ l· (t) 2 dt = Z X L l1=1 X L l2=1 Brl1Θ′′ l1· (t) Θ′′ l2· (t) Brl2 dt = B T r· Z Θ′′(t)[Θ′′(t)]T dt Br· = B T r·PBr· , (2) where Br· is the r-th row of matrix B, Θl·(t) is the l-th spline basis, and P is the L×L penalty matrix with the element of l1-th row and l2-th column being Pl1l2 = R [Θ′′ l1· (t)]T [Θ′′ l2· (t)]dt. As shown in Jiang, Crainiceanu and Cui (2025), incorporating the penalty R (u ′′ r (t))2dt into the log-likelihood is equivalent to imposing a multivariate normal prior p(Br·) ∝ exp − BT r·PBr· 2σ 2 B on each row of the spline coefficients matrix B, where σ 2 B is the smoothing parameter.

p. 5
Where the paper says it

Under the spline basis expansion UT = BΘ, we have Z (u ′′ r (t))2 dt = Z X L l=1 BrlΘ′′ l· (t) 2 dt = Z X L l1=1 X L l2=1 Brl1Θ′′ l1· (t) Θ′′ l2· (t) Brl2 dt = B T r· Z Θ′′(t)[Θ′′(t)]T dt Br· = B T r·PBr· , (2) where Br· is the r-th row of matrix B, Θl·(t) is the l-th spline basis, and P is the L×L penalty matrix with the element of l1-th row and l2-th column being Pl1l2 = R [Θ′′ l1· (t)]T [Θ′′ l2· (t)]dt. As shown in Jiang, Crainiceanu and Cui (2025), incorporating the penalty R (u ′′ r (t))2dt into the log-likelihood is equivalent to imposing a multivariate normal prior p(Br·) ∝ exp − BT r·PBr· 2σ 2 B on each row of the spline coefficients matrix B, where σ 2 B is the smoothing parameter.

M. and Leroux, A. (2021), ‘Additive functional Cox model’, Journal of Computational and Graphical Statistics 30(3), 780–793

Cui, E., Li, R., Crainiceanu, C. M. and Xiao, L. (2023), ‘Fast multilevel functional principal component analysis’, Journal of Computational and Graphical Statistics 32(2), 366–377

Di, C.-Z., Crainiceanu, C. M., Caffo, B. S. and Punjabi, N. M. (2009), ‘Multilevel functional principal component analysis’, The annals of applied statistics 3(1), 458

Gelman, A. and Rubin, D. p. 1

Where the paper says it

Because single imputation treats the reconstructed trajectories as the truth, the uncertainty measures in downstream analyses can often be overly optimistic, particularly in cases of severe sparsity (Rao and Reimherr, 2021; Petrovich et al. In contrast, the multiple imputation framework (Rubin, 1996) explicitly incorporates the imputation uncertainty by generating multiple plausible completions of missing observations.

B., ‘Inference from iterative simulation using multiple sequences’, Statistical science 7(4), 457–472

Greven, S., Crainiceanu, C., Caffo, B. and Reich, D. (2011), Longitudinal functional principal component analysis, in ‘Recent advances in functional data analysis and related topics’, Springer, pp. p. 0

Where the paper says it

Lock, Erjia Cui∗ Division of Biostatistics, University of Minnesota Abstract Missing data are pervasive in modern functional datasets, where trajectories are often sparsely or irregularly observed. Although Functional Principal Component Analysis (FPCA) is widely used to reconstruct incomplete curves, existing approaches typically employ single imputation, leading to overly optimistic inferences in downstream analyses.

149–154

Han, R., Shi, P. and Zhang, A. R., ‘Guaranteed functional tensor singular value decomposition’, Journal of the American Statistical Association 119(546), 995–1007

Happ, C. and Greven, S. p. 1

Where the paper says it

, 2011; Cui et al. , 2023), longitudinal (Greven et al.

p. 1
Where the paper says it

, 2011; Cui et al. , 2023), longitudinal (Greven et al.

, ‘Multivariate functional principal component analysis for data observed on different (dimensional) domains’, Journal of the American Statistical Association 113(522), 649–659

He, Y., Yucel, R. and Raghunathan, T. E., ‘A functional multiple imputation approach to incomplete longitudinal data’, Statistics in medicine 30(10), 1137–1156

Jang, J. H., Manatunga, A. K., Chang, C. and Long, Q., ‘A bayesian multiple imputation approach to bivariate functional data with missing components’, Statistics in medicine 40(22), 4772–4793

Jiang, Z., Crainiceanu, C. and Cui, E. (2025), ‘Tutorial on bayesian functional regression using stan’, Statistics in Medicine 44(20-22), e70265

Jiang, Z., Li, G. and Lock, E. p. 0

Where the paper says it

BAMIFun: Bayesian Multiple Imputation for Functional Data Ziren Jiang, Lei Xuan, Eric F. Lock, Erjia Cui∗ Division of Biostatistics, University of Minnesota Abstract Missing data are pervasive in modern functional datasets, where trajectories are often sparsely or irregularly observed.

p. 0
Where the paper says it

BAMIFun: Bayesian Multiple Imputation for Functional Data Ziren Jiang, Lei Xuan, Eric F. Lock, Erjia Cui∗ Division of Biostatistics, University of Minnesota Abstract Missing data are pervasive in modern functional datasets, where trajectories are often sparsely or irregularly observed.

F., ‘BAMITA: Bayesian multiple imputation for tensor arrays’, Biostatistics 26(1), kxae047

John, D., Tang, Q., Albinali, F. and Intille, S., ‘An open-source monitor-independent movement summary for accelerometer data processing’, Journal for the measurement of physical behaviour 2(4), 268–281

Kidzi´nski, L. and Hastie, T. p. 1

Where the paper says it

, 2005), Bayesian joint registration and curve estimation (Matuk et al. , 2022), and matrix completion (Kidzi´nski and Hastie, 2024).

, ‘Modeling longitudinal data using matrix completion’, Journal of Computational and Graphical Statistics 33(2), 551–566

Kong, D., Staicu, A.-M. and Maity, A., ‘Classical testing in functional linear models’, Journal of nonparametric statistics 28(4), 813–838

Lang, S. and Brezger, A., ‘Bayesian p-splines’, Journal of computational and graphical statistics 13(1), 183–212

Li, G. and Lock, E. F., ‘Integrative analysis of multimodal omics data’, Annual Review of Statistics and Its Application 13

Matuk, J., Bharath, K., Chkrebtii, O. and Kurtek, S., ‘Bayesian framework for simultaneous registration and estimation of noisy, sparse, and fragmented functional data’, Journal of the American Statistical Association 117(540), 1964–1980

Muirhead, R. J., Aspects of multivariate statistical theory, John Wiley & Sons

Petrovich, J., Reimherr, M. p. 1

Where the paper says it

The PACE approach produces a single imputed value at each missing location, hence belonging to the “single imputation” approaches. Because single imputation treats the reconstructed trajectories as the truth, the uncertainty measures in downstream analyses can often be overly optimistic, particularly in cases of severe sparsity (Rao and Reimherr, 2021; Petrovich et al.

and Daymont, C., ‘Highly irregular functional generalized linear regression with electronic health records’, JRSSC 71(4), 806–833

Rao, A. R. and Reimherr, M. p. 1

Where the paper says it

The PACE approach produces a single imputed value at each missing location, hence belonging to the “single imputation” approaches. Because single imputation treats the reconstructed trajectories as the truth, the uncertainty measures in downstream analyses can often be overly optimistic, particularly in cases of severe sparsity (Rao and Reimherr, 2021; Petrovich et al.

, ‘Modern multiple imputation with functional data’, Stat 10(1), e331

Rubin, D. B., ‘Multiple imputation after 18+ years’, JASA 91(434), 473–489

Ruppert, D., Wand, M. P. and Carroll, R. J., Semiparametric regression, number 12, Cambridge university press

Shou, H., Zipunnikov, V. p. 1

Where the paper says it

This FPCA-based approach has since been extended to multilevel (Di et al. , 2009; Zipunnikov et al.

, Crainiceanu, C. M. and Greven, S., ‘Structured functional principal component analysis’, Biometrics 71(1), 247–257

Speckman, P. L. and Sun, D., ‘Fully bayesian spline smoothing and intrinsic autoregressive priors’, Biometrika pp. p. 5

Where the paper says it

Under the spline basis expansion UT = BΘ, we have Z (u ′′ r (t))2 dt = Z X L l=1 BrlΘ′′ l· (t) 2 dt = Z X L l1=1 X L l2=1 Brl1Θ′′ l1· (t) Θ′′ l2· (t) Brl2 dt = B T r· Z Θ′′(t)[Θ′′(t)]T dt Br· = B T r·PBr· , (2) where Br· is the r-th row of matrix B, Θl·(t) is the l-th spline basis, and P is the L×L penalty matrix with the element of l1-th row and l2-th column being Pl1l2 = R [Θ′′ l1· (t)]T [Θ′′ l2· (t)]dt. As shown in Jiang, Crainiceanu and Cui (2025), incorporating the penalty R (u ′′ r (t))2dt into the log-likelihood is equivalent to imposing a multivariate normal prior p(Br·) ∝ exp − BT r·PBr· 2σ 2 B on each row of the spline coefficients matrix B, where σ 2 B is the smoothing parameter.

289–302

Sun, D., Tsutakawa, R. K. and He, Z., ‘Propriety of posteriors with improper priors in hierarchical linear mixed models’, Statistica Sinica pp. 77–95

Wood, S. and Wood, M. p. 1

Where the paper says it

Major difficulties include the implicit reliance of many methods on dense design assumptions (Kong et al. , 2016), the limited number of tailored software programs (Wood and Wood, 2015), and performance degradation under severe missingness.

S., ‘Package ‘mgcv”, R package version 1(29), 729

Yao, F., M¨uller, H.-G. and Wang, J.-L., ‘Functional data analysis for sparse longitudinal data’, Journal of the American statistical association 100(470), 577–590

Zipunnikov, V., Caffo, B., Yousem, D. M., Davatzikos, C., Schwartz, B. S. and Crainiceanu, C. p. 5

Where the paper says it

Under the spline basis expansion UT = BΘ, we have Z (u ′′ r (t))2 dt = Z X L l=1 BrlΘ′′ l· (t) 2 dt = Z X L l1=1 X L l2=1 Brl1Θ′′ l1· (t) Θ′′ l2· (t) Brl2 dt = B T r· Z Θ′′(t)[Θ′′(t)]T dt Br· = B T r·PBr· , (2) where Br· is the r-th row of matrix B, Θl·(t) is the l-th spline basis, and P is the L×L penalty matrix with the element of l1-th row and l2-th column being Pl1l2 = R [Θ′′ l1· (t)]T [Θ′′ l2· (t)]dt. As shown in Jiang, Crainiceanu and Cui (2025), incorporating the penalty R (u ′′ r (t))2dt into the log-likelihood is equivalent to imposing a multivariate normal prior p(Br·) ∝ exp − BT r·PBr· 2σ 2 B on each row of the spline coefficients matrix B, where σ 2 B is the smoothing parameter.

, ‘Multilevel functional principal component analysis for high-dimensional data’, Journal of Computational and Graphical Statistics 20(4), 852–873

Supplementary Material

S1 Bayesian multiple imputation algorithm for functional data (BAMIFun)

S1.1 The Bayesian imputation model

S3 Cross-validation algorithm for the rank R (with the smoothing parameter sampled through MCMC)

S4 Bayesian multiple imputation algorithm with smoothing parameter selected by cross-validation

S5 Bayesian multiple imputation algorithm for multiway functional data

S6 Additional simulation results

S7 Additional data application results

Data-Generating Mechanism

We generate single-level functional data using the same mechanism as in the first simulation experiment (Section 5.1 of the main manuscript) For each subject i = 1,. p. 3

Where the paper says it

Moreover, BAMIFun achieves substantially higher imputation accuracy compared to BAMITA, especially in cases of severe missingness. The remainder of the manuscript is organized as follows.

.., N, Xi(tk) = X h i(t) + epsilon i(t), k = 1,..., K, where X h are the 12 eigenfunctions obtained from our NHANES data application We vary the total number of time points K ∈ [50, 400] and the sample size N ∈ [100, 300] The resulting missing proportion is therefore 8% to 97.5% for K = 50 to 400 respectively

Methods and Metrics

We compare BAMIFun (using L = 15 B-spline basis functions), the FPCA-based PACE algorithm, and BAMITA, as in the first simulation experiment We report the relative imputation RMSE over the missing cells (median over replications) and the elementwise 95% coverage of the imputed values (mean over replications) For PACE, it is the point estimate ± 1.96 times the estimated standard error from face. p. 8

Where the paper says it

2: Run the frequentist face. sparse algorithm to {X, O = 1} with number of principal component equals R and get the estimated eigenfunctions Uˆ init and scores Vˆ init.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Bayesian multiple imputation for sparse functional data allows AI systems to reconstruct incomplete trajectories using a low-rank structure enforced by penalized splines, leading to substantially improved coverage and more reliable downstream inference. This enables AI to perform more accurate prediction of continuous time-series data, such as physical activity profiles, even when only sparsely observed.

  2. The framework extends Bayesian multiple imputation to multiway functional data using the Functional Tensor Singular Value Decomposition (FTSVD) model, allowing AI systems to handle complex datasets like genomics where observations are indexed over multiple modes (e.g., subject and visit). This capability enables the imputation of high-dimensional, multiway data structures not supported by existing methods.

  3. The method properly accounts for estimation uncertainties in downstream analysis by pooling estimates using Rubin’s rules, resulting in a total variance that includes within-imputation and between-imputation components. This allows AI models to generate confidence intervals for functional regression coefficients that are more valid inference than those produced by single imputation methods.

  4. The system can perform downstream scalar-on-function regression with reliable uncertainty quantification, constructing 95 percent credible intervals for functional coefficients using the total variance formula VˆT (t) = VˆW (t) + (1+ 1/S)VˆB(t). This ensures that inferences drawn from the imputed data are robust and calibrated.

  5. The AI system can remain stable and well-calibrated even as the functional grid becomes arbitrarily fine by incorporating a smoothness constraint via the penalty R(u''r(t)) squared dt, which allows for stable and well-calibrated performance across varying observation densities K.

Abstract

Missing data are pervasive in modern functional datasets, where trajectories are often sparsely or irregularly observed. Although Functional Principal Component Analysis (FPCA) is widely used to reconstruct incomplete curves, existing approaches typically employ single imputation, leading to overly optimistic inferences in downstream analyses. To address these challenges, we develop a novel Bayesian multiple imputation framework for functional data (BAMIFun). For single-level functional data, we impose a Bayesian low-rank model that incorporates penalized spline representations to enforce smoothness of the functional domain and derive an efficient Gibbs sampler algorithm for posterior computation. In addition, we demonstrate and validate how to properly account for estimation uncertainties in downstream analysis. Furthermore, we extend the framework to multiway functional data using Functional Tensor Singular Value Decomposition (FTSVD) model, enabling Bayesian multiple imputation in settings not supported by existing methods. Simulation studies show that BAMIFun achieves substantially improved coverage and more reliable downstream inference compared to existing methods, while maintaining similar imputation accuracy. Case studies using a physical activity dataset and an infant gut microbiome dataset further demonstrate the practical advantages of our proposed methods under severe missingness. The code is available at https://github.com/ZirenJiang/BAMIFun.

Related papers