Survival-Aware and Time-Sensitive Network Analysis of Early Modern Correspondence: Methodological Refinements for Epistolary Data

Abstract

Epistolary networks are now central objects of inquiry in digital humanities and early modern studies, promising new perspectives on scholarly exchange, patronage, and the circulation of knowledge. Yet correspondence data are structurally fragile: letters are lost, calendared imperfectly, preserved unevenly across repositories, and digitized according to contemporary priorities. Conventional correspondence network analysis—often based on the subset of surviving letters treated as a complete graph—can therefore misestimate centrality, distort community structure, and encourage overconfident historical interpretations.

This methodological article develops a set of refinements for network analysis of early modern letter correspondence designed to (1) represent loss and inferability explicitly, (2) incorporate temporal dynamics beyond static aggregation, and (3) account for varying archive survival and collection effects. The paper introduces a “survival-aware correspondence event” framework that treats letter exchange as a temporally ordered event stream filtered by archival preservation and editorial selection. It proposes practical modeling tools—bounded reconstructions, probabilistic “shadow edges,” kernel-based time weighting, and sensitivity analyses for preservation bias—implemented within familiar network workflows. Validation is provided through comparative analyses and worked case studies drawn from well-studied early modern scholarly communication (including the circles of Henry Oldenburg and Marin Mersenne). The goal is not to replace interpretive reading, but to improve methodological rigor and transparency when using network analysis on correspondence corpora.

Keywords: network analysis; correspondence; early modern; epistolary networks; methodology; temporal networks; missing data; archival bias

Introduction

Letters are among the most information-dense artifacts of early modern culture. They document intellectual exchange, institutional formation, the pragmatics of patronage, and the social infrastructures that allowed “learning” to travel. For historians of art and culture, epistolary evidence also illuminates the everyday labor of cultural production: negotiation over images, objects, commissions, books, credit, and reputation. In recent years, these materials have become increasingly accessible through large-scale editorial and digital aggregation projects, which in turn has encouraged the application of network analysis to correspondence as a way of modeling connectivity, brokerage, and structural change in cultural fields. ¹

But epistolary networks are not networks “in the wild.” They are networks after survival. The letters that reach us are a remainder shaped by accidents of custody, deliberate destruction, geopolitical rupture, the vagaries of cataloging, and the priorities of editors and digitizers. Moreover, the letter—unlike many other relational traces—arrives with strong temporal structure: each item is (at minimum) a dated and directed event (sender → recipient), often with places, intermediaries, delivery conditions, and explicit references to prior or missing communications. Treating correspondence as a static graph created by aggregating all known letters into a single adjacency matrix is therefore methodologically convenient but historically risky. It can conflate activity separated by decades, underestimate actors whose archives suffered, and misread silence as disconnection.

Three methodological problems recur across the literature and practice of correspondence network analysis in the digital humanities. First, lost letters and incomplete runs of correspondence generate missing edges and missing events, sometimes in ways that are systematically related to the very variables we want to study (status, geography, institutional embedding). Second, temporal dynamics are often simplified to coarse slices (e.g., by decade) or ignored through total aggregation, obscuring bursts of exchange, periods of dormancy, and the shifting “centers” of communication. Third, archive survival and selection effects vary sharply: the correspondence of a prominent secretary preserved in institutional holdings may be abundant, while equally consequential exchanges survive only as scattered copies—or not at all.

This article argues that correspondence network analysis in early modern studies must become more explicit about its data-generating conditions. That entails both conceptual reframing and practical adjustments. Conceptually, we should treat the observed epistolary network as a sample from an underlying (latent) communication process filtered by survival and curation. Practically, we should (a) encode degrees of inferability rather than forcing every relation into “present/absent,” (b) adopt temporal representations that respect event structure, and (c) incorporate preservation heterogeneity through estimation, bounding, and sensitivity analysis.

The aim is methodological refinement rather than software prescription. The tools proposed here can be implemented with standard network packages in R or Python, complemented by event-modeling and Bayesian workflows when appropriate. The paper focuses on early modern scholarly correspondence because it is both a canonical case (the “Republic of Letters” and its associated historiography) and a domain where survival bias is particularly salient. ² Nevertheless, the principles extend to epistolary networks in artistic, ecclesiastical, mercantile, and colonial contexts.

Scope and Contributions

The article makes five contributions:

• A data model for correspondence that distinguishes observed letters, inferred letters, and uncertain metadata while remaining compatible with common network representations.

• A typology of loss and inferability and a set of operational strategies (bounds, “shadow edges,” probabilistic imputation, and documentation standards) for handling lost letters.

• A temporal methodology for epistolary networks that emphasizes event streams and time-weighted ties rather than static aggregation, with explicit parameterization and interpretive guidance.

• A survival-aware adjustment framework that treats archive survival as heterogeneous and potentially actor-specific, and that encourages sensitivity analysis rather than single-number correction.

• Worked case-study demonstrations on early modern scholarly communication, illustrating how conclusions can shift under refined assumptions.

Background: Correspondence Networks Between Source Criticism and Computation

Network analysis has long offered cultural historians a language for relational structure—centers and peripheries, intermediaries, clusters, and pathways. In the humanities, the “network turn” has been enabled by digitization and by the availability of computational methods that can scale from dozens to tens of thousands of edges. ³ Projects such as Mapping the Republic of Letters demonstrated that even partial correspondence corpora can reveal large-scale spatial patterns and intermediaries, while also underscoring how strongly results depend on editorial scope and survival. ⁴ Parallel efforts such as Early Modern Letters Online (EMLO) emphasize aggregation and metadata standardization, making it easier to perform cross-collection queries while also foregrounding provenance complexity. ⁵

At the same time, early modernists have long insisted on the material and social specificity of letters. Manuscript circulation, scribal copying, epistolary conventions, and the “social life” of letters complicate any simplistic model of direct dyadic exchange. ⁶ A letter may be authored by one party, copied by another, delivered by intermediaries, read aloud, and preserved as part of an archive created for reasons unrelated to the communication itself. For network analysis, this means that “sender” and “recipient” are not always sufficient; nor is it safe to assume that a surviving letter is a random sample of all letters once written.

Methodologically, the key challenge is to combine source criticism with formal modeling . Social network analysis has rich literatures on missing data, temporal networks, and statistical inference, but these are not always aligned with humanities questions or data structures. ⁷ Conversely, historians’ nuanced accounts of archival formation are not always translated into computational pipelines. The methodological refinements proposed below seek to bridge this gap: to treat early modern correspondence with the same critical attention we bring to any other source, but in ways that remain computationally actionable.

Method Description: A Survival-Aware Correspondence Event Framework

Overview: From Letters-as-Edges to Letters-as-Events

The foundational shift is to treat each letter as an event with attributes, rather than as a mere contribution to an aggregated edge count. Formally, we represent a correspondence corpus as a set of events E , where each event e minimally includes (i) a sender i , (ii) a recipient j , and (iii) a time t (possibly uncertain). Additional fields (places, language, genre, medium, repository, editorial status) are not optional ornamentation but essential for diagnosing survival and selection.

This event-first approach aligns with temporal network theory and relational event modeling in the social sciences, which emphasize that ties are produced by sequences of actions rather than existing as static channels. ⁸ It also aligns with early modern source realities: letters appear as dated artifacts within archival series, often with observable gaps and reply chains.

Core Data Schema and Minimal Metadata for Methodological Transparency

Because methodological refinements depend on distinguishing types of evidence, the dataset should include fields that many “edge list” exports omit. Table 1 proposes a minimal schema tailored to early modern correspondence network analysis. It does not require full-text encoding; it requires traceable metadata decisions.

Several fields deserve emphasis. First, repository and collection are essential for diagnosing preservation heterogeneity. Second, representing dates as intervals (start–end) avoids the false precision of forced single-day assignment when catalog descriptions provide only “circa” or month-level dating. Third, evidence_type allows us to include “inferred” events without collapsing them into the same evidentiary status as surviving letters.

Refinement 1: Handling Lost Letters Through Inferability, Bounds, and Shadow Edges

Why “Missingness” in Correspondence Is Not a Single Problem

In network methodology, missing data can arise from node non-response, edge non-observation, sampling designs, or censoring mechanisms. ⁹ In early modern correspondence, “missingness” is more heterogeneous. It includes:

• Physical loss : letters written but not preserved (discarded, destroyed, misfiled).

• Archival fragmentation : partial survival across multiple repositories, often unevenly cataloged.

• Editorial omission : letters excluded from editions by scope, language, perceived importance, or redundancy.

• Inferential visibility : letters known to have existed because they are referenced, answered, summarized, or registered, even if the text is lost.

These modes differ in what can be inferred and in how uncertainty should be represented. The methodological aim is to avoid binary edge logic (“there was a tie” vs. “there was no tie”) when the evidence supports intermediate claims.

A Typology of Evidence for “Lost” Correspondence

We propose encoding at least four evidence categories for each event record:

• Observed letter (O) : a surviving item (autograph, signed copy, scribal copy) with identifiable sender/recipient.

• Registered letter (R) : an entry in a letterbook, dispatch register, or calendar that attests sending/receipt but not necessarily content.

• Referenced letter (F) : an otherwise lost letter attested by explicit mention (e.g., “I received your letter of…”).

• Hypothesized letter (H) : inferred from conversational structure (e.g., an answer implies a prompt) but not explicitly referenced.

Only (O), (R), and (F) provide strong warrant for including an event in a dataset; (H) should be used sparingly and always flagged as model-dependent.

Shadow Edges as Probabilistic Events

To incorporate referenced or registered letters without pretending they are equivalent to surviving manuscripts, we introduce shadow edges : events whose existence is supported but whose metadata (date, direction, addressee, or even sender) may be incomplete. Shadow edges are not a metaphor; they are a data structure with (a) a source pointer (quotation, calendar entry, shelfmark), (b) an uncertainty model (date interval; candidate recipients), and (c) an optional probability weight.

At the tie level, a common need is to estimate communication intensity between i and j while acknowledging survival. Let be observed count of letters from i to j in a time range, and the latent count. A simple survival model treats observation as binomial thinning:

(1)

where is the probability a letter in that dyad/time window survives into the dataset (through preservation and editorial selection). Equation (1) is not intended as a full generative truth; it is a disciplined reminder that counts are filtered . Even this simple expression clarifies why unadjusted comparisons can be misleading: if varies sharply by dyad, actor, repository, or period, then is not proportional to communication effort.

Bounding Rather Than Point-Estimating When Survival Is Unidentifiable

In many humanities datasets, is not identifiable from the data alone. Rather than producing a single corrected number (which risks false precision), we recommend reporting bounds and scenario ranges . For example, if archival assessment suggests that survival for a particular collection likely lies between 10% and 40%, then the plausible latent count is:

(2)

Equation (2) yields an interval estimate that can be propagated into downstream metrics via simulation (generate networks under multiple plausible values and report metric distributions). This approach aligns with sensitivity-analysis norms in missing-data research, where the goal is often to show how conclusions depend on assumptions rather than to claim a single “corrected” network. ¹⁰

Reply-Chain Reconstruction and the “Epistolary Reciprocity Check”

Early modern letters frequently refer to “your last” or answer specific queries. Such reply chains can be operationalized cautiously to detect likely missing events. A practical method is the epistolary reciprocity check :

• For each observed letter j → i at time t , search within a preceding window (e.g., 120 days) for an observed i → j .

• If none exists but the letter explicitly references a prior incoming item (“I received…”), create a shadow event i → j (type F) with a date interval anchored by the reference (e.g., between last known exchange and t ).

This procedure must remain conservative: not all letters are replies, and not all replies imply a missing letter in the same dyad (some respond to intermediaries or oral messages). The value is not “filling gaps” for its own sake; it is to make explicit when the dataset itself contains evidence of its incompleteness.

Refinement 2: Representing Temporal Dynamics Beyond Static Aggregation

Why Time Matters More Than “Before/After”

Static aggregation (combining all letters across a long period) implicitly assumes that ties are stable and commensurable across time. Yet early modern correspondence is often bursty—intense around controversies, travel, patronage campaigns, publication projects—and quiet otherwise. Temporal aggregation can also create anachronistic bridges: an actor who corresponded with two otherwise separate circles in different decades may appear as a structural broker in a static network even if those circles never coexisted.

Temporal network research offers two broad options: (1) discrete time slices (yearly, quinquennial, decadal) and (2) continuous-time models that treat events as time-stamped actions. ¹¹ In correspondence corpora with uneven dating precision, both are useful, but each must be used self-critically.

Time-Weighted Tie Strength via Kernel Smoothing

A practical compromise between event streams and static ties is to compute a time-weighted tie strength that decays with the age of letters. Let be the set of letters from i to j , each at time . Define tie strength at time as:

(3)

where is a decay parameter (e.g., 180 days, 365 days, 3 years) chosen to match the historical tempo of exchange under study. Equation (3) produces a dynamic weighted adjacency that emphasizes recent communication while still retaining memory of past contact. It is interpretable: is the “half-life” scale of epistolary relevance (after rescaling), not a purely technical knob.

Temporal Centrality as a Trajectory, Not a Rank

Once is defined, centrality becomes time-indexed. For example, a simple time-dependent out-strength is:

(4)

Rather than ranking actors once, we track trajectories: who becomes central during controversies, who fades after institutional transitions, who emerges as an intermediary during crises. This shifts interpretation from “great men” permanence to the temporality of cultural work.

Event Models for Correspondence as Social Action (Optional, But Powerful)

For researchers prepared to model correspondence in continuous time, relational event models treat each letter as an event whose probability depends on covariates and network history. ¹² In simplified form, the intensity (hazard) of a letter from i to j at time t can be expressed as:

(5)

where are exogenous covariates (distance, shared institution, confessional alignment where known) and are endogenous history statistics (reciprocity, triadic closure, prior volume), with representing the event history up to time t . While such models can be overkill for small corpora, they offer a principled way to separate “opportunity to write” from “propensity to write,” and to test whether apparent clusters arise from shared attributes or from endogenous network processes.

Refinement 3: Accounting for Varying Archive Survival and Collection Effects

Survival Is Heterogeneous and Often Actor-Specific

Archive survival is not merely a global loss rate. It is shaped by household practices, institutional recordkeeping, political rupture, and later editorial canons. Secretaries and institutional administrators (e.g., society secretaries) may leave unusually complete outgoing letterbooks; elite scholars may have their incoming preserved by heirs; marginal actors may appear only through others’ archives. Treating the observed network as uniformly thinned invites systematic distortions.

Methodologically, we recommend treating survival as a function of actor , repository , and time . Even simple stratification can improve transparency: compute network measures separately by repository or edition; compare structures; report divergence. When estimation is attempted, it should be framed as exploratory and accompanied by sensitivity analysis.

A Simple Hierarchical Survival Model (Exploratory)

When sufficient metadata exist, one can posit a logistic model for survival probability for letters associated with sender i , repository r , and time period p :

(6)

Here, captures actor-specific preservation propensity (e.g., institutional officers), captures repository/collection effects (cataloging completeness, editorial scope), and captures period effects (wars, administrative change). Equation (6) is not a claim that survival is “really” logistic; it is a tractable scaffold for estimating heterogeneous thinning and—more importantly—for forcing the analyst to specify what sources of heterogeneity are being assumed.

Collection Triangulation and “Multi-View” Network Construction

Because survival is difficult to estimate from one collection, we recommend “multi-view” construction:

• View A (edition view): network built from a major published edition (often curated, coherent, but selective).

• View B (repository view): network built from archival catalogs for the same actors (broader, but uneven metadata).

• View C (aggregator view): network built from cross-collection aggregators (broadest scope, but variable authority control).

Comparing measures across views provides an empirical check on robustness. Divergence is not failure; it is evidence of selection effects that should shape interpretation.

End-to-End Workflow and Documentation Standards

Methodological refinement requires not only modeling but documentation. We recommend that published correspondence network analyses include a methodological appendix (or repository README) that specifies:

• Corpus definition and exclusions (languages, genres, time bounds).

• Deduplication rules across editions and repositories.

• Evidence-type coding (O/R/F/H) and how shadow events were created.

• Date handling rules (intervals; imputation; slice assignment).

• Temporal parameter choices (window size; in Eq. (3)).

• Sensitivity analysis design (assumed survival ranges; scenarios).

Pseudocode: Building a Survival-Aware Temporal Network

# Inputs:
# events: table with sender_id, recipient_id, date_start, date_end,
#         evidence_type (O/R/F/H), repository, certainty_score
# params: tau (decay), include_types = {O, R, F}, survival_scenarios = [...]

# 1) Filter by evidentiary policy
events_use = events[events.evidence_type in include_types]

# 2) Resolve dates
# If interval-dated, sample within interval during simulation
def sample_date(row):
    return random_uniform(row.date_start, row.date_end)

# 3) Build event stream
event_stream = []
for row in events_use:
    t = sample_date(row)
    event_stream.append((row.sender_id, row.recipient_id, t, row.certainty_score, row.repository))

# 4) Compute time-weighted adjacency at requested times T
def weight(i, j, t, tau):
    return sum(exp(-(t - tk)/tau) for (si, rj, tk, *_ ) in event_stream if si==i and rj==j and tk <= t)

# 5) Sensitivity analysis loop over survival scenarios
for scenario in survival_scenarios:
    # Optionally thin or reweight events by scenario assumptions
    # Compute centralities as trajectories C_i(t)
    pass

Validation and Comparison

Baseline vs. Refined Approaches: What Changes and Why

To validate methodological refinements, we compare three analytical modes:

1. Naïve static aggregation: build one directed graph from all observed letters; compute degree/centrality once.

2. Temporal aggregation (slices): build decade-by-decade graphs from observed letters; compare ranks and communities.

3. Survival-aware temporal modeling: include shadow events (F/R), represent uncertain dates as intervals, compute time-weighted ties (Eq. (3)), and perform sensitivity analysis over survival scenarios (Eq. (2), Eq. (6)).

The core validation criterion is not whether a refined method produces a single “better” network (often unknowable), but whether it yields more stable interpretations under plausible assumptions and more transparent uncertainty where stability is not achievable.

Simulation as a Methodological Check (Conceptual Validation)

Because ground truth is unavailable for historical correspondence, simulation can provide a disciplined test bed. The idea is to generate a synthetic “true” correspondence event stream, then apply known thinning mechanisms (uniform loss, actor-specific loss, repository-specific selection), and evaluate how well different methods recover known properties.

A minimal simulation design:

• Generate a latent event stream among N actors with known centralization and community structure (e.g., two clusters connected by a broker).

• Apply heterogeneous thinning: high-status actors have higher survival; one “repository” captures outgoing letters for a secretary; peripheral actors are under-preserved.

• Compare naive static centrality, sliced centrality, and survival-aware methods to the known latent structure.

In such simulations (common in missing-network-data research), naïve methods tend to inflate the centrality of actors whose letters survive better and can create spurious brokerage when time is aggregated. ¹³ The value of simulation here is pedagogical and diagnostic: it helps researchers see which claims are fragile under plausible archival conditions.

Case Study Demonstration I: Henry Oldenburg and the Institutional Archive Effect

Henry Oldenburg, as a secretary of the Royal Society, is a paradigmatic example of how institutional roles shape epistolary traces. Major edited volumes of Oldenburg’s correspondence provide a rich backbone for network analysis, but that richness is itself historically structured: secretarial work produces copies, registers, and systematic retention. ¹⁴ A naïve network built from the edited correspondence can therefore mistake institutional recordkeeping for personal sociability.

Applying the refinements proposed here changes the analysis in three ways:

• Evidence stratification: outgoing letters preserved via institutional mechanisms (register/copy) are separated from incoming letters surviving in private archives. This supports analysis of asymmetries (who is visible because they wrote to an institution vs. who is visible because their papers survived).

• Temporal trajectories: time-weighted centrality (Eq. (4)) highlights periods when Oldenburg’s brokerage intensifies (e.g., around coordinated publication exchanges) versus periods where the network’s center shifts to other institutional actors.

• Sensitivity checks: survival scenarios can ask: if outgoing survival for secretarial archives is much higher than incoming survival for peripheral correspondents, how stable are conclusions about “core” members of the network?

Rather than “correcting away” Oldenburg’s importance—which would be a category error—the refined method clarifies what kind of importance is being measured: centrality in the observed archive versus centrality in the broader communicative system. For cultural historians, this distinction matters because it separates infrastructural labor from charisma and because it reveals how institutional forms produce the very evidence we analyze.

Case Study Demonstration II: Marin Mersenne and the Fragmented Hub

Marin Mersenne’s correspondence is frequently described as a hub of early seventeenth-century scholarly communication, connecting mathematicians, natural philosophers, and theologians across Europe. Yet the corpus is distributed and edited across volumes that draw on multiple repositories and copies, with uneven survival across correspondents. ¹⁵ In such a case, “hubness” is both a historical hypothesis and a product of editorial reconstruction.

The survival-aware event framework supports three concrete checks:

• Shadow-event enrichment: referenced letters (F) can be included to reduce reply-chain distortion. If Mersenne’s outgoing letters survive better than incoming (or vice versa), referenced letters can partially rebalance tie visibility—without claiming full recovery.

• Dating uncertainty: many letters have approximate dates (“early spring,” “before Easter”). Representing dates as intervals reduces misleading spikes in time-sliced graphs caused by forced assignment to a single day or month.

• Multi-view comparison: compare network measures computed from the printed edition versus aggregator metadata (where available). If Mersenne’s brokerage persists across views, the claim is more robust; if it collapses, the analysis should foreground selection effects.

Illustrative Metric Comparison (Hypothetical Example)

To make the methodological stakes concrete without claiming new empirical results, Table 3 shows a hypothetical comparison of outcomes that frequently occur when moving from naïve to refined methods. The numbers are illustrative representations only, intended to show directions of change rather than actual measurements.

Discussion

Interpretive Payoffs: What These Refinements Enable

The methodological refinements proposed here do not merely “clean data.” They enable different kinds of cultural-historical claims.

First , survival-aware modeling makes it easier to treat epistolary networks as evidence about infrastructures rather than simply about individuals. Secretaries, scribes, and institutions become visible as preservation engines and as communicative coordinators—roles that are culturally significant even when they complicate hero-centric narratives.

Second , temporal methods shift attention from static hierarchies to episodes : controversies, campaigns, and moments of accelerated exchange. For arts and cultural studies, this helps connect correspondence networks to eventful cultural processes (commissioning cycles, publication debates, diplomatic crises) rather than to generalized “connectedness.”

Third , evidence stratification encourages better integration between quantitative network analysis and qualitative reading. When shadow events are linked to explicit quotations (“your letter of…”), the network visualization becomes an index into textual moments of reference, not a replacement for them.

Methodological Risks and How to Manage Them

Risk 1: Over-modeling and False Precision

Estimating survival probabilities or fitting relational event models can create an aura of scientific certainty. In historical corpora, the danger is that a sophisticated model may outrun the evidentiary base. The remedy is to prefer bounded inference , to publish assumptions prominently, and to report distributions (from scenario simulation) rather than single corrected values.

Risk 2: Conflating Archival Centrality with Social Centrality

Even refined methods cannot fully separate “importance in the surviving record” from “importance in historical life.” But they can prevent unmarked conflation. Researchers should explicitly name the target construct: archival centrality , editorial centrality , or (more ambitiously) latent communicative centrality under stated assumptions.

Risk 3: Neglecting Non-Letter Channels

Early modern communication also flowed through oral exchange, print, manuscript circulation beyond letters, and intermediaries. A correspondence network is therefore at best a partial representation of a broader media ecology. ¹⁶ Temporal network methods can help by aligning letter bursts with known events, but researchers should avoid treating letters as the totality of communication.

Recommendations for Best Practice (Methodological Checklist)

• Model letters as events first ; derive edges second.

• Encode evidence type (O/R/F/H) and expose it in analysis (e.g., layered visualizations).

• Represent date uncertainty explicitly using intervals; avoid forced precision.

• Use time-weighted ties (Eq. (3)) or event models where appropriate; treat centrality as a trajectory.

• Stratify by repository/edition at least once to diagnose collection effects.

• Perform sensitivity analyses over plausible survival rates; report robustness rather than single outcomes.

• Publish decisions (deduplication, exclusions, imputation rules) as part of the scholarly argument.

Conclusion

Network analysis of early modern letter correspondence has matured from novelty to a routine component of digital humanities research. With that maturation comes responsibility: correspondence networks are not neutral graphs but archival artifacts shaped by survival, selection, and time. This article has proposed methodological refinements that make those conditions visible and analytically tractable: evidence typologies and shadow events for lost letters, temporal kernels and trajectories for dynamic interpretation, and survival-aware stratification and sensitivity analysis for uneven archival preservation.

The central claim is modest but consequential. We cannot fully reconstruct the latent epistolary networks of early modern Europe. But we can avoid treating the surviving remainder as if it were complete, and we can discipline our interpretations by showing how they depend on assumptions about loss, dating, and archival formation. In doing so, correspondence network analysis becomes not merely a computational technique but a form of source criticism—one that remains quantitative, explicit, and historically responsible.

Notes

References