This vignette describes the
statistical model implemented by fit_baygmst(), connecting
the code to the underlying statistics and to the literature it descends
from. For a hands-on walkthrough of the software, see
vignette("baygmst-intro", package = "BayGMST").
Global mean surface temperature (GMST) is a fundamental state variable of Earth’s climate system and a central target of paleoclimate reconstruction efforts. Because direct instrumental measurements extend back only about 150 years, estimates of GMST over longer periods must be inferred from indirect observations preserved in paleoclimate archives: tree rings, glacier ice, corals and sclerosponges, lake and marine sediments, and historical documents (Mann et al., 1999; Moberg et al., 2005). A variety of statistical strategies have been developed to extract the climate signal contained in these “proxies” (Tingley et al., 2012; Smerdon and Pollack, 2016), whether the target is a spatial climate field (Smerdon et al., 2023) or an aggregated index such as GMST.
BayGMST’s inference framework builds directly on the work of Barboza et al. (2014, 2019) and Wang (2020), who utilized a reduced proxy Bayesian approach that leverages not only the climate proxy network, but also estimates of past forcings (e.g., greenhouse gas concentrations, solar, and volcanic activity), which are known to influence GMST. The model below is their hierarchical formulation; BayGMST packages it, with the implementation choices noted in this vignette.
The approach is based on computing a “reduced proxy” \(\boldsymbol{P} = (P_1, \ldots, P_t, \ldots,
P_n)^T\), an aggregate of the available paleoclimate proxy
observations. Several methods exist to do so, like principal component
regression, partial least squares, sliced inverse regression, or a
simple composite (Barboza et al., 2019); see
reduce_proxies() for the methods implemented here, and
as_baygmst_proxy() for supplying a reduced or composited
series built elsewhere. This quantity is then modeled as autoregressive
and linearly dependent on the GMST, denoted \(\boldsymbol{T} = (T_1, \ldots, T_t, \ldots,
T_n)^T\) (subscript \(t\) refers
to time in years CE).
The model’s data level:
\[ P_t = \phi_P P_{t-1} + \alpha_0 + \alpha_1 T_t + \sigma_P \epsilon_t, \qquad \epsilon_t \overset{\text{iid}}{\sim} N(0, 1) \]
where \(\phi_P\) denotes the lag-1 autocorrelation of \(P\), the \(\alpha\)’s are coefficients, and \(\sigma_P\) controls the amplitude of the unit-variance noise term \(\epsilon\).
The process level models \(T_t\) as autoregressive and linearly
dependent on the climate forcings \(G_t\) (greenhouse gas concentrations),
\(V_t\) (related to volcanic aerosol
optical depth), and \(S_t\) (solar
irradiance), already transformed as described below (see
transform_forcings()):
\[ T_t = \phi_T T_{t-1} + \beta_0 + \beta_G G_t + \beta_V V_t + \beta_S S_t + \sigma_T \eta_t, \qquad \eta_t \overset{\text{iid}}{\sim} N(0, 1) \]
where \(\phi_T\) denotes the lag-1 autocorrelation of \(\boldsymbol{T}\), the \(\beta\)’s are constant coefficients, and \(\sigma_T\) governs the amplitude of \(\eta_t\). Gaussian white noise for both \(\boldsymbol{\eta}\) and \(\boldsymbol{\epsilon}\) provides good results (Wang, 2020). Here \(|\phi_P|, |\phi_T| < 1\) and \(\sigma_P, \sigma_T > 0\).
Missing (pre-instrumental) temperatures are treated as parameters and estimated jointly with the model parameters, so the posterior yields both the reconstruction and its uncertainty.
| Notation | Stan parameter | R argument / role |
|---|---|---|
| \(T_t\) | y (transformed parameter) |
instrumental_T (observed part), y_mis
(reconstructed part) |
| \(P_t\) | z (data) |
proxy, see reduce_proxies() /
as_baygmst_proxy() |
| \(\phi_P\) | phi_R |
|
| \(\sigma_P\) | sigma_z |
|
| \(\alpha_0, \alpha_1\) | alpha0, alpha1 |
|
| \(\phi_T\) | phi_T |
|
| \(\sigma_T\) | sigma_y |
|
| \(\beta_0, \beta_G, \beta_V, \beta_S\) | beta0, betaG, betaV,
betaS |
|
| \(G_t, V_t, S_t\) | G, V, S (data) |
forcing_G, forcing_V,
forcing_S |
Note on naming: the Stan code calls the proxy-equation parameters
phi_R and sigma_z (“R” for “reduced proxy”,
matching this implementation’s single-series design); this vignette
writes them \(\phi_P\) and \(\sigma_P\) for readability. They are the
same quantities.
As specified in the Stan model’s model block (weakly
informative):
\[ \begin{aligned} \alpha_0, \alpha_1, \beta_0, \beta_G, \beta_V, \beta_S &\sim N(0, 0.5) \\ \phi_P, \phi_T &\sim N(0, 0.3) \\ \sigma_P, \sigma_T &\sim \text{Exponential}(1) \end{aligned} \]
Because the data level estimates its own intercept \(\alpha_0\) and slope \(\alpha_1\), the proxy series does not need
to be pre-calibrated to temperature units; these priors do, however,
assume \(P_t\) is on roughly unit scale
(a standardized composite or a temperature-anomaly-scale series). See
as_baygmst_proxy().
Per Barboza et al. (2019, section 2.3), the three external forcings used in this family of models are:
past1000 forcing set (Jungclaus et al.,
2017).All these series measure perturbation to the top of the atmosphere
energy budget, and are thus expressed in \(\mathrm{W\,m^{-2}}\), following the climate
science convention. To effect the conversion from the original units
(aerosol optical depth for \(V_t\);
mole fraction for \(G_t\)),
transform_forcings() implements
\[ G'_t = \texttt{co2\_coef} \times \log(G_t / \texttt{co2\_c0}), \qquad V'_t = -\lvert\texttt{vol\_coef}\rvert \,(1 - e^{-V_t}), \]
where co2_coef = 5.35 is the simplified CO2 radiative
forcing coefficient of Myhre et al. (1998), adopted throughout the IPCC
assessment reports; co2_c0 = 280 ppm is the standard
pre-industrial atmospheric CO2 reference concentration; and
vol_coef = 25 is the stratospheric aerosol optical
depth-to-forcing scaling factor (in \(\mathrm{W\,m^{-2}}\) per unit optical
depth) used in IPCC AR5, traced to the GISS ModelE lineage of Hansen et
al. (2005). Solar irradiance is simply centered on its own mean, \(S'_t = S_t - \bar{S}\). These
physically motivated formulations are a slight departure from the
transformations used by Barboza et al. (2014, 2019).
baygmst.stan, the latent temperature state is set
exactly equal to the raw instrumental values for observed years; there
is no separate instrumental measurement-noise parameter. Keep this in
mind when interpreting instrumental-period fit statistics from
summary() on a fitted model.cv_baygmst() reports the posterior mean of the
Stan-computed r2_cv/mse_cv generated
quantities, which is close to, but not identical to, a plug-in
point-estimate \(R^2\) computed from
the posterior-mean reconstruction.as_baygmst_proxy()), the ensemble is collapsed
to a single series before fitting in this release.