The Rise and Fall of Poetry in Nineteenth-Century Russian Literature: Age-Period-Cohort Analysis

Introduction

What drives population-level cultural and social change? Cultural sociology offers two competing theoretical accounts that aim to answer this question (Vaisey and Lizardo; Kiley and Vaisey). One of them, the stable disposition model (SDM), suggests that, in their formative years, individuals acquire a set of beliefs, preferences, and patterns of action that tend to persist throughout life. The other, the active updating model (AUM), posits that changes in the social environment make individuals adopt new behavior and beliefs that are maintained over time. While both are theories of social learning at the individual level, the SDM and AUM make different predictions about aggregate cultural change.

If the SDM holds, cultural change should be a cohort-led process with generational replacement as its main driver. Conversely, if the AUM holds, cultural change results from aggregated within-individual change over time. As Vaisey and Lizardo argue, these predictions can be formalized from the perspective of age-period-cohort (APC) analysis (Vaisey and Lizardo). APC analysis is based on the idea that changes in various outcomes are the result of the joint work of three temporal effects. Period effects reflect time-specific social and historical changes. Cohort effects are changes due to intergenerational differences. And age effects refer to an individual-level change that unfolds over the life course and biological aging. The SDM implies that cohort effects should be most important in explaining cultural change, whereas the AUM posits that individual change—the composition of period and age effects—plays the primary role. Research shows that generational turnover explains a wide range of cultural change, from religiosity to shifts in public opinion on sensitive issues to the erosion of social capital (Molteni and Biolcati; Restrepo Ochoa and Vaisey; Schwadel and Stout).

When it comes to literature, there are two main systemic explanations for literary change. The first one relies on a loosely understood notion of literary generations. Russian formalists argue that literary “children” tend to fight their “fathers” and align more closely with “grandfathers” with whom the “fathers” had struggled (Tynianov 195; Yarkho 312). However, no Russian formalist elaborated the process of generational turnover.¹ In his theory of “anxiety of influence,” Harold Bloom asserts that writers are driven to overcome the shadow of their precursors and thereby strive for innovation (Bloom). Although Bloom’s psychoanalytic account of intergenerational struggle offers one rationale for cohort replacement, the latter can arise from more prosaic mechanisms. As classical sociological theorists and the most recent researchers point out, cohort differentiation stems from formative experience at critical ages under different social contexts or, in other words, from “generational imprinting” (Ryder; Fosse, “Theorizing in the Cohort Mode”; Keskintürk).

An alternative explanation relies on the notion of literary periods. Consider the way literature is taught in schools and universities. Modern literary history is typically presented as a series of major aesthetic shifts: classicism giving way to romanticism, romanticism to realism, realism to modernism, and so on. This view closely resembles Thomas Kuhn’s model of scientific revolutions: Most writers operate within the prevailing literary “movements” (analogous to Kuhn’s “normal science,” or research that extends and refines, not challenges, prevailing scientific paradigms), following the first-movers, until a literary “revolution” transforms writing practices (Kuhn). Such movements are often viewed as responses to the zeitgeist, that is, to broader historical and social changes. This account applies not only to major aesthetic shifts but also to the rise and fall of specific literary forms, such as genres. For example, Luc Boltanski explains the emergence of detective and spy novels by the development of the modern democratic state in France and England (Boltanski). From this perspective, literary history is a period-led process.

To date, only two closely related studies have directly addressed the role of APC effects in literary change. The first, led by Ted Underwood, analyzes anglophone fiction published from 1880 to 1999 and finds that cohort effects explain more than half of the temporal variance in topics, exceeding period effects (Underwood et al.). Oleg Sobchuk and Bret Beheim reach similar conclusions: Cohort replacement plays a slightly stronger role than individual change in topic dynamics (Sobchuk and Beheim). Analyzing related corpora with different methods, both studies converge on their main finding. Although both research teams are confident that topic modeling provides an adequate representation of literary texts for their research question, their results and, more importantly, their theoretical claims should nevertheless be tested using alternative units of analysis.

In Sobchuk and Beheim’s study, the canonical example of a topic driven by cohort effects was one strongly associated with adventure novels about the American Wild West. The authors refer to this topic as the Western genre and demonstrate that the emergence and disappearance of a single generation of writers drove its ascent and subsequent decline in popularity. This finding provides the starting point for this study. Literary scholars rarely think of long-term literary change as a transformation of computationally inferred topics over time. Not only do the latter still lack conceptual transparency, but literary history is traditionally framed as the evolution of more or less “stable forms”: formulas, narratives, meters, and genres (Underwood; Šeļa et al.).

Russian literary history, in particular, is conventionally described through both aesthetic shifts and the alternation of dominant types of literary genres, especially prose and poetry. The first third of the nineteenth-century Russian literature is primarily considered poetic, whereas the second third is characterized by the dominance of prose, which is followed by the poetic “Silver Age.” Furthermore, in Russian literature, prevailing literary styles or movements are associated with specific genres and with the divide between prose and poetry. Romanticism and modernism literature are primarily expressed through poetic genres, whereas realist fiction is largely expressed in the form of the novel and other prosaic genres, such as sketches and short stories. However, being descriptive, this stylized fact lacks causal explanation.

In this article, I evaluate which temporal driver contributed most to the observed shift from poetry to prose in nineteenth-century Russian literature. Using prosopographical data from the Dictionary of Russian Writers, 1800–1917, I adopt the bounding analysis of APC effects proposed by sociologists Ethan Fosse and Christopher Winship (“Bounding Analyses of Age-Period-Cohort Effects”). I find that the rise and fall of poetic debuts cannot be explained by generational replacement and, instead, was mostly driven by period effects.

Studying APC effects has far-reaching implications for understanding literary and, more broadly, cultural change. If literature is driven by cohort succession, or intergenerational differences, then its history may not resemble a sequence of aesthetic revolutions, as is often assumed, but instead a gradual process of what some might call “reforms.” This view aligns with sociological research on why some cultural artifacts or producers become popular and prestigious, while others do not. In particular, analyzing which musical tracks are more likely to enter and stay in top charts, Noah Askin and Michael Mauskapf find that the most successful songs deviated from their predecessors and contemporaries, but only moderately, an effect called “optimal differentiation” or “optimal distinctiveness” (Askin and Mauskapf). Buhr and colleagues report a similar finding in a study of high-tech firms (Buhr et al.). If the most successful cultural products are those that strike a balance between conformity and innovation, then cultural evolution should be gradual and not abrupt. Because the generational replacement hypothesis similarly suggests a pattern of incremental change, the study of APC effects in literature is relevant for both literary scholars and, more broadly, researchers of cultural evolution (Lambert et al.).

In the next sections, I introduce my data source and outline the identification strategy. Then, I contextualize the historical setting and show that, under different sets of assumptions, the bounding analysis produces qualitatively similar results, supporting period effects as the primary drivers of the poetry and prose alternation in nineteenth-century Russian literature.

Data

I analyze APC data on nineteenth-century Russian writers born between 1760 and 1839. The data were manually compiled from the Dictionary of Russian Writers, 1800–1917 by a research team who was supported by a grant from the Russian Scientific Foundation. This source is the most comprehensive prosopographical and bibliographical dictionary available, covering a long span of Russian literary history. No other resource provides comparable completeness for the population of Russian writers. However, because the authors of the Dictionary did not list all published texts by each writer, I am unable to examine period and cohort effects on genre choices throughout their careers. For this reason, I focus on writers’ debuts for which the genre and publication year were consistently recorded.

The research team collected a wide range of data on the social and literary characteristics of all the authors included in the Dictionary. Three of these were the authors’ birth year, debut year, and debut genre. But the Dictionary contains data on individuals who never published fiction, for example, Dmitri Mendeleev. After excluding from the analysis all debut genres that cannot be considered either poetry or fiction in the nineteenth-century context (e.g., sermons, textbooks, or articles), these individuals were automatically ruled out. Nevertheless, some individuals made their debut with nonfictional publications but eventually produced literary texts. As the research team did not specifically aim to record literary debuts, such cases were not captured. As a member of the team that manually coded data from the biographical articles, I believe the number of authors who transitioned from nonfiction to fiction is relatively small.

There are other data limitations. Because dramatic debut texts could have been written in either prose or verse, it was often difficult to determine the form, even when consulting the corresponding biographical entries. As a result, most of the dramatic debuts (N = 37) were excluded from the analysis. The same situation applies to translations that cannot be easily recoded into prose or poetry (N = 152). Note that translations were frequently published anonymously, and most remain unattributed. Thus, 189 observations had to be left out, and the focus was shifted to non-translated literary debuts. I analyze a sample of writers born in the 1760s through the 1830s. After applying all cleaning procedures, the population under examination consists of 719 writers (see the distribution of authors in Figure 4).

Methods

Bounding Analysis

All scholars examining temporal effects need to face the classical APC identification problem. The APC identification problem arises from the perfect linear dependence between age, period, and cohort (cohort = period – age), making the APC effects not identifiable from the data alone. All solutions require some assumptions, whether explicitly or not (Rohrer).

When measuring the topical variance explained by cohorts or periods, Underwood and his coauthors assume that “age strongly constrains some behavior (e.g., driving) but is not very useful in predicting literature” and represented age as a quadratic polynomial (Underwood et al. 187). Sobchuk and Beheim contrast generational replacement with individual change. By doing so, they sidestep the APC identification problem at the cost of not distinguishing age from period effects, because it is their composition that constitutes individual change. However, in literary studies, much effort has focused on the importance of period effects (Underwood). Since isolating both period and cohort effects is of interest when examining literary change, I apply a recently developed framework called bounding analysis that offers partial identification solution to the APC problem (Fosse and Winship, “Bounding Analyses”).

Although linear APC effects cannot be recovered from data alone, nonlinear effects and combinations of linear effects are fully identifiable (Fosse and Winship, “Moore–Penrose Estimators”). With the help of theoretical assumptions, these estimable quantities can be leveraged to partially identify the linear components and, by extension, total APC effects. This is the core idea behind the bounding analysis that Fosse and Winship propose.

Formally, the bounding analysis relies on the linearized APC model where APC effects are represented as the sum of linear and nonlinear components. Let i = 1, …, I represent age groups, j = 1, …, J be period groups, and k = 1, …, K stand for cohort groups. Then, the outcome of interest Y_ijk will be

$Y_{\mathit{\text{ijk}}}=\mu +\alpha \left(\mathit{\text{ag}}{e}_i\right)+\pi \left(\mathit{\text{perio}}{d}_j\right)+\gamma \left(\mathit{\text{cohor}}{t}_k\right)+{\tilde{\alpha }}_i+{\tilde{\pi }}_j+{\tilde{\gamma }}_k{}_{}+{\epsilon }_{\mathit{\text{ijk}}}$     (1)

where α, π, and γ are linear effects (slopes) for APC, respectively; age_i, period_j, and cohort_k represent APC indices; and ̃α_i, ̃π_j, and ̃γ_k are nonlinear APC effects. Due to the linear dependency of the variables, the linear slopes are not identifiable. However, knowing that period_j = cohort_k + age_i, equation (1) can be transformed to

$Y_{\mathit{\text{ijk}}}=\mu +\underset{{\theta }_1}{\underbrace{\left(\alpha +\pi \right)}}\left(\mathit{\text{ag}}{e}_i\right)+\underset{{\theta }_2}{\underbrace{\left(\pi +\gamma \right)}}\left(\mathit{\text{cohor}}{t}_k\right)+\underset{nonlinear components}{\underbrace{{\tilde{\alpha }}_i+{\tilde{\pi }}_j+{\tilde{\gamma }}_k}}+{\epsilon }_{ijk}$     (2)

Replacing period with the sum of age and cohort makes equation (2) identifiable, but instead of the individual APC slopes, it identifies two linear combinations, θ₁ and θ₂, and the nonlinear components as deviations from the unknown linear parts. In the following empirical analysis (discussed in the next section), these quantities will be obtained via a classical linear regression model.

After estimating θ₁ and θ₂, we can narrow the range of possible APC effects by solving the following system of equations:

$\left\{\begin{array}{c}{\theta }_1 = \alpha + \pi ,\\ {\theta }_2 = \pi + \gamma .\end{array}\right.$     (3)

At this stage of the analysis, it is helpful to adopt a geometric interpretation of the APC identification problem. As combinations of APC linear slopes, θ₁ and θ₂ define two planes in the three-dimensional APC parameter space. Their intersection is the line that contains all possible solutions compatible with the data. Note that this canonical solution line is a solution to the system of equations (3). Panel A of Figure 1 illustrates these ideas by plotting two planes defined by θ₁ = 1 and θ₂ = 0.2, respectively. Panel B presents their intersection in two-dimensional space.

Figure 1: Geometric interpretation of the canonical solution line (simulated data). Panel A shows the three-dimensional APC parameter space of linear slopes. The black line is the canonical solution line. Panel B shows its two-dimensional projection.

Although an exact set of linear slopes can be computed for each point on the canonical solution line, the number of such combinations remains infinite. To move forward, the canonical solution line should be constrained. Fosse and Winship propose strategies to impose bounds by introducing theoretical assumptions about the size, shape, or direction of APC effects (“Bounding Analyses”). Here I discuss shape constraints that will be used in the empirical analysis below.

Shape constraints encode expectations about the overall trajectory of one or more APC effects. Recall that, in equation (2), nonlinear components are identified along with θ₁ and θ₂. If a temporal effect is expected to be monotonic (increasing or decreasing) over a certain range, we can derive upper and lower bounds for the corresponding linear slope to account for nonlinear deviations that distort the expected trend. In Panel A of Figure 2, I present a toy example illustrating nonlinear age effects.

Figure 2: Nonlinear age effects and the bounded canonical solution line (simulated data). Panel A depicts nonlinear age effects. The shaded area denotes the interval subject to the monotonicity constraint, and the red arrow indicates the minimum linear slope needed to meet it. Panel B depicts the canonical solution line after imposing the derived minimum linear age slope. The shaded area represents the space of remaining possible slope combinations.

Suppose we have a theoretically motivated assumption that overall age effects should increase monotonically between the 10 and 40 age groups. If the linear age slope is zero or negative, this assumption is violated because, between the 10 and 20 age groups, the nonlinear deviation reduces the linear effect by 0.3 points. To compensate for this downward deviation, the linear slope must therefore be at least 0.3 (minimal linear slope). By imposing this adjustment, we obtain the lower bound for the canonical solution line (see Figure 2, Panel B). The same procedure can be used to derive the upper bound. Once both bounds are established, the canonical solution line is defined by a finite set of solutions, and the range of possible APC trajectories can be computed.²

To summarize, the bounding analysis achieves partial identification of APC effects in two main steps. First, it defines the canonical solution line by statistically estimating θ₁ and θ₂, which represent combinations of linear slopes. Second, it constrains this line by analytically inferring the minimum and maximum linear slopes required to satisfy the theoretical assumptions.³

In this study, the bounding analysis was performed using the apcR package developed by Fosse (“Bounding Analyses”) and custom functions for visualizing bounded APC effects by Rohrer.⁴ The apcR package is not yet available on CRAN, but it can be downloaded from the Harvard website.⁵

Assessing Contribution of Total Effects

After applying bounding analysis to the data, I obtain a range of estimates for the decomposed APC effects, representing the probability that a debut is poetic (1) instead of prosaic (0) for each APC group. In theory, this could be sufficient to assess whether period- or cohort-driven processes played a greater role in shaping the alternation between prose and poetry. Figure 3 illustrates two idealized scenarios. Under a pure period-based explanation (Panels A–C), the age and cohort curves would be flat at 0.5 across all ages and cohorts, while the period effect varies. In this case, in any given period the total effect of age–cohort composition is the same, and differences arise across periods. The same is true for the pure cohort explanation (Panels D–F). If period and age effects are fixed at 0.5 for each group, changes in the proportion of poetic debuts over time could be explained only by changes in the probability of debuts across cohort groups (Panel F).

Figure 3: Theoretical expectations on total APC effects (simulated data). Panels A–C represent the period-based explanation of poetry/prose shift. Panels D–F stand for the cohort-based explanation.

However, more realistically, it may be unclear which of the APC contributions is primarily responsible for the change in the overall probability of a poetic debut between t + 1 and t. The contribution of the period effect can be measured as the difference between corresponding probabilities displayed on the graph of period-implied trajectories (e.g., at Figure 1, Panel B, it is the difference between 0.32 and 0.6 between the 1840s and 1830s). To determine the contribution of cohorts to the overall change between t + 1 and t, I estimate the difference between the average probability of poetry debut for the cohort composition at t + 1 and the average probability of poetry debut for the cohort composition at t:

${\delta }_{P_{t+1}-P_t}^{cohort}=\frac{{{\displaystyle \sum }}_{i=1}^K\hspace{0.17em}{n}_{C_i}^{P_{t+1}}·p_{C_i}}{{{\displaystyle \sum }}_{i=1}^K\hspace{0.17em}{n}_{C_i}^{P_{t+1}}}-\frac{{{\displaystyle \sum }}_{i=1}^{K^{\prime }}\hspace{0.17em}{n}_{C_i}^{P_t}·p_{C_i}}{{{\displaystyle \sum }}_{i=1}^{K^{\prime }}\hspace{0.17em}{n}_{C_i}^{P_t}}$     (4)

where C_i – cohort category (cohort 1, cohort 2, …, cohort K or K′); K and K′ are the numbers of cohort categories at P_t+1 and P_t; $n_{C_i}^{P_{t+1}}$ and $n_{C_i}^{P_t}$ are the numbers of individuals from cohort C_i at periods P_t+1 and P_t, respectively; and $p_{C_i}$ – probability of a poetic debut by cohort C_i.

Suppose that in period t there are 10 debutants from cohort C₁ and 5 debutants from cohort C₂, with cohort-specific probabilities of a poetic debut equal to 0.6 and 0.3, respectively. In period t + 1, the composition reverses: 5 from C₁ and 10 from C₂.

Then ${\delta }_{P_{t+1}-P_t}^{cohort}=\frac{5·0.6+10·0.3}{10+5}-\frac{10·0.6+5·0.3}{10+5}=-0.1$ . In other words, the change in cohort composition reduces the probability of a poetic debut by 10 percentage points.

The same logic applies to the age contribution:

${\delta }_{P_{t+1}-P_t}^{\mathit{\text{age}}}=\frac{{{\displaystyle \sum }}_{i=1}^I\hspace{0.17em}{n}_{A_i}^{P_{t+1}}·p_{A_i}}{{{\displaystyle \sum }}_{i=1}^I\hspace{0.17em}{n}_{A_i}^{P_{t+1}}}-\frac{{{\displaystyle \sum }}_{i=1}^{I^{\prime }}\hspace{0.17em}{n}_{A_i}^{P_t}·p_{A_i}}{{{\displaystyle \sum }}_{i=1}^{I^{\prime }}\hspace{0.17em}{n}_{A_i}^{P_t}}$     (5)

where A_i is the age category (age 1, age 2, …, age I or I′); I and I′ are the numbers of cohort categories at P_t+1 and P_t; $n_{A_i}^{P_{t+1}}$ and $n_{A_i}^{P_t}$ are the numbers of individuals in age category A_i at periods P_t+1 and P_t, respectively; and p_Ai – probability of a poetic debut by age A_i.

The Rise and Fall of Poetry

Before moving to the bounding analysis of APC of Russian writers’ poetic and prosaic debuts, I first inspect the structure of the data. Figure 4 presents three panels showing the number of debuts, stacked by each APC group with equal 10-year intervals.⁶ Panel B supports the widespread belief of the alternation of prose and poetry in nineteenth-century Russian literature. From the 1790s to the 1820s, poetic debuts were markedly more frequent than prosaic ones. However, beginning in the 1830s, this trend reversed, with the proportion of poetic texts declining significantly. In Russian literary history, this shift is commonly attributed to the transition from Romanticism to Realism.

Figure 4: Poetic and prosaic debuts by age (Panel A), period (Panel B), and cohort (Panel C) groups.

The question, however, is why this alternation occurred in terms of APC effects. Was it, say, fashionable for writers across all age and cohort groups to debut with poetry in the first third of the nineteenth century, and with prose in its second half? Or were there age or cohort effects behind that? Panel A and Panel C suggest that it is possible that older cohorts were inclined to write poetry and, in combination with early debuts that are mostly poetic, they could collectively shape the trend visible in Panel B.

A similar identification exercise could be done to explain the pattern in Panel C. Were older cohorts inclined to choose verses in their debuts, compared to younger cohorts? One possible reason for this could be the different reading practices: The Pushkin generation read French poetry, while the Tolstoy generation also read French novelists at a young age. Alternatively, the probability of making a poetic debut may not depend much on cohort group but instead on the composition of period and age effects. In other words, perhaps the Pushkin and Tolstoy generations did not differ in their attitudes towards poetry and prose, but the former lived in an era of poetry superiority, while the latter lived in an age of prose, which is why their choices are so different. Both explanations can usually be found implicitly in various academic works on the history of Russian literature.

Simply measuring the frequency of poetic and prosaic debuts in different combinations of APC groups might only provide highly suggestive hypotheses about the direction, shape, and size of temporal effects. In Figure 3, I have presented the two competing explanations in terms of the probability of a poetic debut (1) versus a prosaic one (0). In reality, these two extreme models (pure period-based and pure cohort-based) are highly unlikely. However, they provide a sense of what to expect when either period or cohort effects primarily drive the poetry/prose change.

From this point, I proceed to a bounding analysis in two parts corresponding to two sets of assumptions I impose on the data. The first, conducted under period-related assumptions, supports the period-based explanation. The second, conducted under cohort-related assumptions, demonstrates nearly identical results that support the period-based explanation and reject the cohort-based one.

Period Constraints

The bounding analysis requires formalizing field-specific knowledge in order to narrow the set of possible APC effects. Given that the first third of the nineteenth century is commonly associated with the rise of sentimentalist and then romantic poetry, while the second third of the century is associated with the ascent of realist prose (Vaysman et al.), I propose two assumptions about the shape of period effects:

P1. Probability of a poetic debut increases monotonically from the 1800s to the 1820s period intervals.

P2. Probability of a poetic debut decreases monotonically from the 1830s to the 1860s period intervals.⁷

The combination of P1 and P2 assumptions provides bounds for the upper and lower linear slopes that result in qualitatively unambiguous possible APC trajectories. Figure 5 presents both the nonlinear and total APC effects that should be interpreted in percentage points as the probability of debuting in poetry (1) versus prose (0).⁸

Consider the nonlinear period effects. The two shaded areas indicate the intervals over which different assumptions are imposed. The blue-shaded interval corresponds to the P1 assumption, where the nonlinear effects are monotonically increasing. If I require the total period effect to be monotonically increasing over this interval, I can derive a lower bound for the linear period slope. To see this, suppose the linear slope is positive. Then the P1 assumption is automatically satisfied. However, if the linear slope is strongly negative, the nonlinear component cannot offset it to produce an overall increasing trajectory. This implies that there exists a minimum (in absolute value) negative slope that still allows the nonlinear effects to preserve the desired monotonic increase.

Turning to assumption P2, the nonlinear effects are monotonically decreasing over the corresponding interval (orange). Since the total effects are required to decrease as well, any negative linear slope is consistent with this assumption. A strongly positive slope, however, would dominate over nonlinear effects and produce an increasing total trajectory. The upper bound is therefore the least positive slope that still yields a monotonically decreasing total effect.

Together, these two assumptions yield a lower and an upper bound for the linear-period slope. As shown in the canonical solution line graph (see Figure 2), constraining a single slope automatically constrains all remaining slopes. The second row of Figure 5 displays the resulting effects under the P1 and P2 assumptions. First, note that the assumptions are satisfied. All period trajectories increase monotonically between the 1800s and the 1820s and decrease monotonically between the 1830s and the 1860s. If we look at the overall trend, the probability of a poetic debut rises in the 1780s through 1820s and falls within the 1820s–1860s range, with upward and downward fluctuations until the end of the century.

Considering age trajectories, the probability of debuting in poetry generally declines with age, with an exception in the 30s and 40s age groups. The cohort effects also display an overall downward trajectory: Writers born later were less likely to debut with poetry. Had the cohort trend been upward, we could conclude that poetry’s decline occurred in spite of, rather than because of, generational replacement. As it stands, Figure 5 leaves open the possibility that the decline of poetry and the rise of prose shown in Figure 4 were driven more by intergenerational differences than by period effects.

Figure 5: Nonlinear APC effects (top row) and total APC effects under the P1 and P2 assumptions (bottom row). The y-axis shows the probability of a poetic (1) versus prosaic (0) debut, and the x-axis marks 10-year APC intervals. In the top row, the blue-shaded region on the period panel indicates the interval associated with the P1 assumption, and the orange-shaded region indicates the interval for P2. In the bottom row, blue lines represent solutions with maximum linear APC effects, and orange lines represent solutions with minimum linear APC effects. Dashed lines in the middle show the average trajectories between the two extreme bounds for age, period, and cohort. Shaded bands indicate the range of estimated effects.

To adjudicate which temporal effect played the most important role, I compute the total contribution of APC effects across periods. The contribution of period effects to changes in the probability of debuting in poetry across decades is given by the difference between the corresponding probabilities (Figure 5, “Period effects”). The contributions of cohort and age effects are calculated as weighted averages (see Equations (4) and (5) in the Methods section). As the bounding analysis does not identify precise APC effects but rather a range of possible solutions, I use the APC effects obtained with the maximum and minimum linear slopes. The resulting estimates are presented in Figure 6.

Figure 6: APC contributions to the total change (in percentage points) under P1 and P2 assumptions. Panel A shows solutions with the smallest linear APC slopes. Panel B shows solutions with the largest linear slopes.

The total contributions of APC effects, computed under the minimum and maximum linear slopes, display nearly identical patterns. Overall, the total change in the probability of a poetic debut over decades was driven primarily by period effects. The 1820s to 1830s, a turning point in nineteenth-century Russian poetry/prose dynamics when poetic debuts began to decline rapidly, was a joint product of all APC effects. However, period effects clearly led this process. From the 1780s through the 1810s, period-implied changes were the only force driving the rise in poetic debuts. The impact of intergenerational differences appears to be weak. Interestingly, changes in the age composition of debutants played a role comparable to that of cohorts and were even the major driver of the decline in poetic debuts between the 1850s and 1860s. In later decades, the strong apparent contribution of age effects was largely due to data skew: because the analysis focuses on debutants whose cohort groups end in the 1830s, older debutants are overrepresented in the second half of the century. As a result, age effects become more prominent during this period. Consequently, we should be cautious when drawing conclusions about periods after the 1850s and 1860s.

In sum, the bounding analysis under the assumptions P1 and P2 rejects the cohort-based explanation of the rise and fall of nineteenth-century Russian poetry and instead points to period effects as the primary driving force behind the observed trend.

Cohort Constraints

In nineteenth-century Russian literature, the Pushkin generation is regarded as the most poetic generation. However, Pushkin himself, born in 1799, occupies an intermediate position between the cohorts born in the 1790s and those born in the 1800s. The formative influences on this generation are usually attributed to the previous generation, most notably Vasily Zhukovsky. Subsequent generations, from Gogol to Turgenev to Dostoevsky and Tolstoy, are generally characterized as becoming increasingly prosaic. Accordingly, I adopt the following two assumptions:

C1. Probability of a poetic debut increases monotonically from the 1780s to the 1790s cohort groups.⁹

C2. Probability of a poetic debut decreases monotonically from the 1800s to the 1830s cohort groups.

The top row of Figure 7 shows nonlinear cohort effects with two shaded areas corresponding to the C1 and C2 intervals.¹⁰ For the 1780s–1790s cohorts, the C1 assumption holds automatically if the linear cohort slope is zero or positive. Therefore, the minimum slope sufficient to produce monotonically increasing total cohort effects must be determined. For the 1800s–1820s cohorts (C2), any positive slope can be ruled out because it would amplify the increase in total cohort effects already visible in the nonlinear component during the 1810s–1820s. The upper bound is thus the least negative slope that counteracts the 1810–1820 rise (any more negative slope will satisfy the C2 assumption). These constraints together result in the upper and lower bounds that produce total effects shown in the second row of Figure 7.

Figure 7: Nonlinear APC effects (top row) and total APC effects under the C1 and C2 assumptions (bottom row). The y-axis shows the probability of a poetic (1) versus prosaic (0) debut, and the x-axis marks 10-year APC intervals. In the top row, the orange-shaded region on the period panel indicates the interval associated with the C1 assumption, and the blue-shaded region indicates the interval for C2. In the bottom row, blue lines represent solutions with maximum linear APC effects, and orange lines represent solutions with minimum linear APC effects. Dashed lines in the middle show the average trajectories between the two extreme bounds for age, period, and cohort. Shaded bands indicate the range of estimated effects.

Applying cohort assumptions C1 and C2 returns APC probabilities almost the same as those under P1 and P2. Figure 7 mirrors the trajectories seen in Figure 5 but with a slightly narrower range of estimates. Thus, the overall pattern remains unchanged: it is period effects that shape the prosaic and poetic debut dynamics.

Discussion

Departing from recent computational research of literary change, this study tested whether, in Russian literature, generational replacement might explain the fall and rise of poetry and prose, much like it does in the case of topical variation over time. Contrary to the previous studies (e.g., Underwood et al.; Sobchuk and Beheim), the bounding analysis of APC effects does not support this idea. While cohort turnover contributes to the dynamics of poetic and prosaic debuts, its influence is low compared to period effects. In addition to this, the research reports the importance of age effects. Younger authors have a higher probability of writing poetry, whereas older authors are more likely to produce prose, just as poems tend to skew younger and novels skew older.

However, age effects should not be interpreted as a universal life-course trajectory of preference change. The obtained estimates are valid only within the proposed identification assumptions, which are themselves period-specific. This means that age effects are by construction bounded by a particular historical setting. One could further argue that even the period under analysis is internally heterogeneous and that age effects may differ between, say, the era of Romantic dominance and that of Realist fiction. This is indeed one of the limitations of “accounting” APC models, to which the bounding analysis belongs. Technically, this reflects the assumption of non-interaction among APC terms. While interactions between temporal variables are substantively important, they exacerbate the APC identification problem, and appropriate methods for handling them remain underdeveloped (Fosse and Winship, “Analyzing Age-Period-Cohort Data”).

The implications of the main finding depend on the level of generalizability we are ready to accept. If we reject that the overall prose and poetry dynamics mirror those observed for prose and poetry debuts, then the apparent dominance of period effects may be an artifact of the specificity of debut choice. In that case, early-career writers are guided by the contemporary literary context (e.g., editorial cues, authoritative peers, economic constraints), but by mid-career they revert to the literary forms “imprinted” during formative years, along with others in their cohort. This view aligns with the stable disposition model in which individuals may deviate from a baseline but tend to return to it (Kiley and Vaisey). However, one would then expect the same pattern in the topics of debut publications.

If we assume that the period effects shape both debutants’ and non-debutants’ choices of poetry and prose, the reason my results are inconsistent with those of Underwood’s and Sobchuk’s studies may reflect differences in observability and salience of literary features. Whether the text is written as poetry or prose is straightforward to classify, whereas topic assignment is noisier and more time-consuming. It is much easier to determine whether the text is prosaic or poetic than to determine the topics to which it is dedicated. As Restrepo Ochoa and Vaisey argue, beliefs about salient, publicly discussed issues are driven by intra-individual change, whereas more sensitive issues are shaped by cohort replacement (Restrepo Ochoa and Vaisey). Similarly, writers can readily adapt their choices when the dominant type of literary discourse is easy to observe (prose against poetry in this case, but also specific genres or vers libre against syllabo-tonic versification), but they tend to persist in the topics to which they were extensively exposed at formative ages. This salience argument finds indirect support in studies of literary topics: topics associated with external shocks (e.g., wars, technological change) tend to be driven by period effects (Underwood et al.) and individual-level shifts (Sobchuk and Beheim).

Previous researchers have emphasized that different cultural domains are shaped by distinct temporal dynamics, with the literature appearing particularly susceptible to cohort effects (Underwood et al. 185; Sobchuk and Beheim 10). This study suggests that within a single cultural domain there are at least two levels that evolve at different paces and under different primary drivers. Consider the case of science. It is plausible that paradigms tend to shift from period to period, whereas less ossified elements, such as specific research topics, may emerge and fade through cohort replacement. If this is true, literature can be described in a similar manner, but instead of paradigms, we encounter major stylistic movements and more or less fixed forms, both of which seem to be prone to period-implied mechanisms.

Beyond its theoretical implications, the present study raises the question of the period-related mechanisms that led writers to favor poetry or prose. Despite literary scholarship on the subject, the specific drivers behind period effects remain suggestive and generally unclear. To move beyond the often essayistic, historical causal claims, it is desirable to develop mechanism-based approaches to the APC identification problem (Winship and Harding). This strategy underscores that APC effects work through specific mediators and may help to address the directionality of cultural change. For example, taste is known to be formed via familial and schooling socialization (Childress et al.). Cohort effects could therefore operate through exposure to different reading material during the formative years of writers. Economic development, in turn, can serve as a period-implied mechanism, as it was shown to be a causal driver of romantic love in literary fiction (Baumard et al.). Identifying these mechanisms is an important step toward understanding why individuals orient toward particular cultural objects at particular times and why, as in the case of poetry and prose, those orientations shift and reverse over the course of literary history.

Acknowledgements

This study was supported by a grant from the Russian Science Foundation, project No. 24-18-00761 (Institute of Russian Literature, Russian Academy of Sciences), https://www.rscf.ru/project/24-18-00761/. I thank Kirill Maslinsky, Alexey Vdovin, Antonina Martynenko, Alina Bodrova, and Sergei Guskov for their guidance, interest, and support at various stages of this work. Last but not least, I would like to acknowledge the anonymous JCA reviewers whose contributions made the article genuinely better.

Competing Interests

The author has no competing interests to declare.