Panel graphical VAR: temporal, contemporaneous, and between-person networks

Written by Mohammed Saqr and Sonsoles López-Pernas

Panel studies measure the same people on a few occasions, for instance weekly during a course. Two different questions can be asked of such data. The between-person question is whether students who are generally more efficacious also plan more. The within-person question is whether a student who feels more efficacious than usual in one week plans more than usual, in the same week or in the next. The answers can differ, and a cross-sectional network of the pooled data mixes them. The panel graphical vector-autoregressive model (panel-GVAR; Epskamp, 2020) separates each person’s scores into a stable personal level and weekly deviations from it, and estimates three networks: a temporal network of lagged effects between the deviations, a contemporaneous network of deviations that move together within a week, and a between-person network of the stable levels.

This article fits the model to a four-wave panel of five self-regulated learning indicators, reads the three networks with their uncertainty, selects a sparser model by two methods, and evaluates the fit, the numerical solution and the division of variance between and within persons.

library(psychnets)

Data

srl_panel is a synthetic panel of 100 students measured in four consecutive weeks. It was simulated from a panel-GVAR whose structure was taken from real experience-sampling data (see ?srl_panel), so it is realistic and can be shared without restriction. The data are in long format, one row per student per week.

head(srl_panel)
#>   student week efficacy planning effort enjoyment anxiety
#> 1    S001    1     38.7     54.1   52.1      53.1    44.2
#> 2    S001    2     32.9     55.4   50.7      62.2    48.8
#> 3    S001    3     34.5     40.9   52.9      50.3    36.9
#> 4    S001    4     28.6     44.5   41.0      42.4    39.6
#> 5    S002    1     45.0     54.7   48.6      48.1    38.4
#> 6    S002    2     62.8     67.2   63.7      47.5    44.8
vars <- c("efficacy", "planning", "effort", "enjoyment", "anxiety")

The model

For person \(p\) at wave \(t\), the vector of scores is decomposed as

\[\mathbf{y}_{pt} = \boldsymbol\mu + \boldsymbol\eta_p + \mathbf{w}_{pt}, \qquad \mathbf{w}_{pt} = \mathbf{B}\,\mathbf{w}_{p,t-1} + \boldsymbol\zeta_{pt},\]

where \(\boldsymbol\mu\) is the grand mean, \(\boldsymbol\eta_p\) the person’s stable deviation from it, and \(\mathbf{w}_{pt}\) the person’s deviation from their own level at wave \(t\). The three networks are:

The parameters are estimated by maximum likelihood, and each edge has a standard error and a Wald test. The model assumes that the means do not change across waves.

What do the three networks show?

panel_gvar() takes the data with the names of the variables, the person identifier and the wave. By default every edge is estimated and kept, and an edge is flagged as significant at \(p < 0.01\).

fit <- panel_gvar(srl_panel, vars = vars, id = "student", wave = "week")
fit
#> <psychnet_panelgvar> panel graphical VAR (temporal + contemporaneous + between)
#>   persons: 100   waves: 4   nodes: 5
#>   edges (significant/estimated, p < 0.01): temporal 1/20 (+4/5 autoregressive)   contemporaneous 5/10   between 2/10
#>   selection: none (full model)
#>   fit: chisq(170) = 207.19, p = 0.0273   CFI = 0.957   TLI = 0.946   RMSEA = 0.047
#>   optimality (ML score): 9.99e-13
#>   starts: 3 of 3 reached the best saturated optimum

The model has 100 students in 4 waves. Of the 20 lagged effects between variables, 1 is significant, and 4 of the 5 autoregressive effects are; 5 of the 10 contemporaneous and 2 of the 10 between-person edges are significant. The edge table has one row per edge, with its standard error, p-value and 95% Wald interval:

head(as.data.frame(fit), 8)
#>    network     from        to       weight         se           p   ci_lower
#> 1 temporal  anxiety   anxiety  0.296628278 0.10027278 0.003094278  0.1000972
#> 2 temporal  anxiety  efficacy  0.026009263 0.08031053 0.746045017 -0.1313965
#> 3 temporal  anxiety    effort -0.014936851 0.08549642 0.861309611 -0.1825068
#> 4 temporal  anxiety enjoyment  0.036480078 0.08218780 0.657142045 -0.1246051
#> 5 temporal  anxiety  planning  0.003375992 0.07857850 0.965730789 -0.1506350
#> 6 temporal efficacy   anxiety -0.122338773 0.13119129 0.351066802 -0.3794690
#> 7 temporal efficacy  efficacy -0.096672635 0.13467200 0.472857381 -0.3606249
#> 8 temporal efficacy    effort -0.185033790 0.13554948 0.172232751 -0.4507059
#>    ci_upper significant
#> 1 0.4931593        TRUE
#> 2 0.1834150       FALSE
#> 3 0.1526331       FALSE
#> 4 0.1975652       FALSE
#> 5 0.1573870       FALSE
#> 6 0.1347914       FALSE
#> 7 0.1672796       FALSE
#> 8 0.0806383       FALSE
summary(fit)
#>           network directed nodes edges significant mean_abs_weight
#> 1        temporal     TRUE     5    20           1       0.1046718
#> 2 contemporaneous    FALSE     5    10           5       0.1877681
#> 3         between    FALSE     5    10           2       0.1831649

summary() gives one row per network. The contemporaneous and between-person edges are on average larger in absolute value (0.188 and 0.183) than the temporal ones (0.105). plot() draws the three networks on one layout: solid edges are significant, dashed edges are not, blue is positive and vermillion negative, and the weight is printed on each edge.

plot(fit)

The significant edges answer the two questions separately:

Which edges should a sparser model keep?

Flagging significance leaves the model unchanged. Two methods select a sparser model.

select = "prune" removes every edge with \(p > 0.01\), refits, and then adds edges back one at a time while each addition significantly improves the fit and lowers the BIC (step-up search). The step-up stage matters because edges can substitute for each other: an autoregressive effect and a correlated lagged effect may each look unnecessary while the other is present, and a single pruning pass would remove both.

pruned <- panel_gvar(srl_panel, vars = vars, id = "student", wave = "week",
                     select = "prune")
pruned
#> <psychnet_panelgvar> panel graphical VAR (temporal + contemporaneous + between)
#>   persons: 100   waves: 4   nodes: 5
#>   edges (significant/estimated, p < 0.01): temporal 2/2 (+4/4 autoregressive)   contemporaneous 5/5   between 4/4
#>   selection: prune (alpha = 0.01, adjust = none; stepup (bic) re-added 3)
#>   fit: chisq(200) = 227.13, p = 0.0913   CFI = 0.969   TLI = 0.966   RMSEA = 0.037
#>   optimality (ML score): 2.58e-13
#>   starts: 3 of 3 reached the best saturated optimum

The pruned model keeps 4 autoregressive effects, 2 lagged effects between variables, and 5 contemporaneous and 4 between-person edges. The step-up search re-added 3 edges that pruning had removed. The sparser model still fits well (CFI = 0.969, RMSEA = 0.037), so the removed edges were not needed to reproduce the data. Its edge table:

as.data.frame(pruned)
#>            network      from        to    weight         se            p
#> 1         temporal   anxiety   anxiety 0.3097634 0.10142690 2.257702e-03
#> 2         temporal    effort  efficacy 0.2745331 0.05772377 1.974862e-06
#> 3         temporal    effort    effort 0.3050205 0.06630006 4.212623e-06
#> 4         temporal enjoyment  efficacy 0.1771524 0.04998380 3.938117e-04
#> 5         temporal enjoyment enjoyment 0.3811826 0.07083908 7.408767e-08
#> 6         temporal  planning  planning 0.2937572 0.06026909 1.093089e-06
#> 7  contemporaneous  efficacy    effort 0.3736857 0.05047766 1.331492e-13
#> 8  contemporaneous  efficacy  planning 0.5750641 0.03916613 8.322431e-49
#> 9  contemporaneous    effort enjoyment 0.2875605 0.04897720 4.323548e-09
#> 10 contemporaneous  planning    effort 0.1747501 0.05512077 1.522752e-03
#> 11 contemporaneous  planning enjoyment 0.2427458 0.04454979 5.069247e-08
#> 12         between  efficacy  planning 0.3994936 0.08012929 6.176604e-07
#> 13         between    effort enjoyment 0.3784409 0.11858021 1.415685e-03
#> 14         between  planning    effort 0.3085720 0.10517305 3.346829e-03
#> 15         between  planning enjoyment 0.3560514 0.10934260 1.128777e-03
#>      ci_lower  ci_upper significant
#> 1  0.11097033 0.5085565        TRUE
#> 2  0.16139658 0.3876696        TRUE
#> 3  0.17507478 0.4349663        TRUE
#> 4  0.07918596 0.2751189        TRUE
#> 5  0.24234055 0.5200246        TRUE
#> 6  0.17563197 0.4118825        TRUE
#> 7  0.27475132 0.4726201        TRUE
#> 8  0.49829988 0.6518283        TRUE
#> 9  0.19156696 0.3835541        TRUE
#> 10 0.06671533 0.2827848        TRUE
#> 11 0.15542977 0.3300617        TRUE
#> 12 0.24244302 0.5565441        TRUE
#> 13 0.14602795 0.6108538        TRUE
#> 14 0.10243661 0.5147074        TRUE
#> 15 0.14174384 0.5703590        TRUE

Both lagged effects lead into efficacy: a week of more effort than usual (0.27) or more enjoyment than usual (0.18) is followed by higher efficacy than usual. The contemporaneous edges are those of the full model, and the between-person network gains two edges of planning, with effort (0.31) and enjoyment (0.36), which were not significant at \(p < 0.01\) in the full model but are significant once the other between-person edges are removed.

plot(pruned)

select = "ebic" instead shrinks the edges with a lasso penalty chosen by the extended BIC, the panel counterpart of the graphical lasso. Weak edges become exactly zero in one estimation. The estimates are penalized and therefore have no p-values; refit = TRUE refits the selected network without the penalty. A path of 20 penalty values keeps the example short.

lasso <- panel_gvar(srl_panel, vars = vars, id = "student", wave = "week",
                    select = "ebic", nlambda = 20)
summary(lasso)
#>           network directed nodes edges significant mean_abs_weight
#> 1        temporal     TRUE     5     2          NA       0.1436679
#> 2 contemporaneous    FALSE     5     5          NA       0.3385725
#> 3         between    FALSE     5     0          NA       0.0000000
as.data.frame(lasso)
#>            network      from        to     weight se  p ci_lower ci_upper
#> 1         temporal   anxiety   anxiety 0.01126022 NA NA       NA       NA
#> 2         temporal  efficacy  efficacy 0.01940377 NA NA       NA       NA
#> 3         temporal    effort  efficacy 0.16267521 NA NA       NA       NA
#> 4         temporal    effort    effort 0.16069469 NA NA       NA       NA
#> 5         temporal enjoyment  efficacy 0.12466058 NA NA       NA       NA
#> 6         temporal enjoyment enjoyment 0.25094809 NA NA       NA       NA
#> 7         temporal  planning  planning 0.17284327 NA NA       NA       NA
#> 8  contemporaneous  efficacy    effort 0.35072292 NA NA       NA       NA
#> 9  contemporaneous  efficacy  planning 0.58378554 NA NA       NA       NA
#> 10 contemporaneous    effort enjoyment 0.29072556 NA NA       NA       NA
#> 11 contemporaneous  planning    effort 0.21147396 NA NA       NA       NA
#> 12 contemporaneous  planning enjoyment 0.25615455 NA NA       NA       NA
#>    significant
#> 1           NA
#> 2           NA
#> 3           NA
#> 4           NA
#> 5           NA
#> 6           NA
#> 7           NA
#> 8           NA
#> 9           NA
#> 10          NA
#> 11          NA
#> 12          NA

The lasso keeps the same two lagged effects into efficacy and the same five contemporaneous edges as pruning, and no between-person edge. It also keeps all five autoregressive effects, two of them close to zero (anxiety 0.011, efficacy 0.019). The two methods therefore agree on the within-person structure. With 100 students, the penalty is more conservative than significance-based selection for the between-person network.

How well does the model fit?

The fit indices compare the model with the saturated model, which reproduces the observed covariances exactly, and with a baseline without temporal or network structure.

as.data.frame(fit, what = "fit")
#>        logl npar  df    chisq     pvalue baseline_chisq baseline_df       cfi
#> 1 -6802.528   60 170 207.1872 0.02733013       1084.025         215 0.9572082
#>         tli      rmsea      aic      bic
#> 1 0.9458809 0.04677052 13725.06 13881.37

The chi-square test rejects exact fit (\(\chi^2(170) = 207.19\), \(p = 0.027\)), as is common for a sample of this size, but the approximate indices show good fit: CFI = 0.957, TLI = 0.946 and RMSEA = 0.047. The model reproduces the covariances across the four waves well.

Is the estimate the maximum-likelihood solution?

The likelihood of the panel-GVAR is not convex, so the optimizer is started from three different points. At a maximum the gradient of the log-likelihood (the score) is zero; the largest remaining score is the certificate, and a value near zero shows that the estimate is a maximum. The diagnostics also report whether any parameter lies on the boundary of its admissible range, whether standard errors could be computed, and how many starts reached the same optimum.

as.data.frame(fit, what = "diagnostics")
#>     n waves        score boundary boundary_part se_available starts_agree
#> 1 100     4 9.987888e-13    FALSE          <NA>         TRUE            3
#>   starts
#> 1      3

The largest score is below \(10^{-11}\), no parameter is on a boundary, standard errors are available, and all three starts reached the same optimum. The estimate is therefore the maximum-likelihood solution, and its standard errors can be used.

How much of each variable is stable between persons?

The variance of each variable is divided into the variance of the personal levels (between) and of the weekly deviations (within).

as.data.frame(fit, what = "variance")
#>    variable  between   within share_between
#> 1  efficacy 65.39817 66.42409     0.4961087
#> 2  planning 72.92447 57.68286     0.5583490
#> 3    effort 82.63922 69.22678     0.5441588
#> 4 enjoyment 53.83746 65.44361     0.4513496
#> 5   anxiety 95.00513 59.25030     0.6158949

Between 45% (enjoyment) and 62% (anxiety) of the variance lies between students. About half of the variation in each indicator is a stable difference between students and half a fluctuation from week to week, so both the between-person and the within-person networks describe a substantial part of the data.

What if the wave means drift?

The model assumes that the mean of each variable is the same in every wave. If the means drift, for instance a steady decline over a term, standardize = "z_per_wave" standardizes every variable within every wave, which removes the drift; the edges are then on a standardized scale.

z_fit <- panel_gvar(srl_panel, vars = vars, id = "student", wave = "week",
                    standardize = "z_per_wave")
subset(as.data.frame(z_fit, what = "fit"), select = c(chisq, df, cfi, rmsea))
#>     chisq  df       cfi      rmsea
#> 1 182.507 170 0.9851576 0.02712385

On these data, which were simulated with stable means, standardizing within waves gives a similar and slightly better fit (CFI = 0.985, RMSEA = 0.027). On data with drifting means the difference would be larger, and the standardized model would be the one to report.

Assumptions and when to use the panel-GVAR

Reference

Epskamp, S. (2020). Psychometric network models from time-series and panel data. Psychometrika, 85(1), 206-231. doi:10.1007/s11336-020-09697-3