Publication Choice Problem

• Publication Choice Problem is a multifaceted decision challenge where research outputs are matched to publication outlets, time windows, and mechanisms to optimize impact.
• Key methods include causal modeling, inverse-probability weighting, and game-theoretic analyses that balance selection bias, strategic venue impact, and evaluation frequency.
• Empirical and simulation studies reveal trade-offs between citation maximization, ranking reliability, and publication system design in varied academic fields.

Searching arXiv for the cited papers to ground the article in current records. arXiv search: id:([2010.09157](/papers/2010.09157)) OR id:([1810.13375](/papers/1810.13375)) OR id:([2511.13678](/papers/2511.13678)) OR id:([2603.00807](/papers/2603.00807)) OR id:([2504.21156](/papers/2504.21156)) OR id:([0911.0344](/papers/0911.0344)) OR id:([1908.08702](/papers/1908.08702)) The publication choice problem denotes a family of decision problems concerning how research outputs are matched to publication outlets, publication periods, and publication mechanisms. In one line of work, the problem is to choose the venue at which a paper would receive the largest number of citations by estimating counterfactual outcomes across venues (Sato et al., 2020). In another, it is the choice of the publication window used in research assessment, where reliability of rankings must be balanced against frequency of evaluation (Abramo et al., 2018). More recent work formalizes publication choice as a strategic coordination problem in which researchers’ venue selections both respond to and shape venue impact (Wang et al., 17 Nov 2025), and related empirical work studies whether fields converge on common publication preferences or remain fragmented across many outlets (Buskirk et al., 28 Feb 2026). Adjacent literatures analyze how peer-review protocols, journal acceptance rules, and publication-biased incentives alter these choices and their consequences (0911.0344, Jagadeesan et al., 29 Apr 2025, Braganza, 2019).

1. Scope of the term

The expression is used in several distinct, though related, literatures. In the scholarly-publication setting, the core question is usually how a paper, researcher, or field should choose among outlets or evaluation rules. In a separate privacy-preserving data-publication literature, the same phrase is used for a different problem: how to release a single published object with different utility levels for different authorization tiers (Jiang et al., 2021).

Formulation	Decision variable	Criterion
Venue recommendation	$r(X)\in\mathcal T$	Maximize estimated citations
Research assessment window	Publication period $L$	Balance ranking reliability and evaluation frequency
Strategic publication allocation	$a_i=(a_{i,1},\dots,a_{i,k})$	Trade off venue impact and cost
Collective preference coordination	Field-level venue ordering	Consensus versus fragmentation
Data-publication variant	Published object $y_w$	Multi-level utility-privacy trade-offs

The main scholarly uses differ in what is treated as endogenous. Venue-recommendation models take venues as candidate treatments and estimate their effects on citations. Research-assessment models take the publication window itself as the object of choice. Game-theoretic and survey-based analyses instead treat venue prestige or preference orderings as endogenous outcomes of decentralized behavior. This terminological breadth is important because otherwise results about citation-maximizing venue choice, evaluation-window design, and community consensus can be conflated despite solving different optimization problems (Sato et al., 2020, Abramo et al., 2018, Wang et al., 17 Nov 2025, Buskirk et al., 28 Feb 2026).

2. Venue choice as counterfactual impact estimation

A causal formulation of venue choice models the publication venue as a treatment. With $n$ past papers, covariates $X_i\in\mathbb R^d$, observed venue $T_i\in\mathcal T=\{1,\dots,K\}$, and observed citations $Y_i\in\mathbb R_+$, the objective is to learn a recommender

$r:\mathbb R^d \longrightarrow \mathcal T$

that returns the venue at which a query paper is expected to receive the largest number of citations. Under Rubin’s framework, each paper $i$ has potential outcomes

$L$0

but only the factual outcome $L$1 is observed. The target is therefore

$L$2

The central technical difficulty is selection bias: papers are not randomly assigned to venues, so venue and content are not statistically independent. The framework therefore estimates propensity scores

$L$3

and uses inverse-probability weighting. In the implementation described for Poincare, paper features are a binary bag-of-fields-of-study vector in $L$4; outcome estimation uses a T-learner with one linear ridge-regression model per venue,

$L$5

and propensities are estimated with multi-class logistic regression. Bias correction is implemented by minimizing, for venue $L$6,

$L$7

which is equivalent to weighted ridge regression with sample weight $L$8.

The reported experiment uses five top CS conferences,

$L$9

papers accepted in 2015 from DBLP, bag-of-fields-of-study covariates of approximately $a_i=(a_{i,1},\dots,a_{i,k})$0 dimensions, citations counted over the next 5 years, and a stratified 70/30 train/test split per venue. Association-based recommenders—logistic-regression, random forest, SVM, and MLP trained only to predict $a_i=(a_{i,1},\dots,a_{i,k})$1—show near-zero correlation with actual citations, whereas Poincare’s predicted citation counts achieve overall Spearman’s $a_i=(a_{i,1},\dots,a_{i,k})$2 with $a_i=(a_{i,1},\dots,a_{i,k})$3. Covariate distributions differ significantly across venues by MMD two-sample tests with $a_i=(a_{i,1},\dots,a_{i,k})$4, confirming selection bias. On a toy counterfactual simulation, the IPW version outperforms an unweighted variant and all association baselines in both counterfactual ranking-accuracy and average true citations in the chosen venue, with $a_i=(a_{i,1},\dots,a_{i,k})$5. The method also agrees more often with “influential” authors, defined as the top 50% by historical citations (Sato et al., 2020).

The substantive significance of this formulation is that it distinguishes expected acceptance from expected impact. A venue recommender trained only on historical placement estimates where a paper is likely to be published; the treatment-effect formulation instead estimates what would happen to the same paper under alternative venue assignments. Its limitations are explicit: it relies on ignorability, $a_i=(a_{i,1},\dots,a_{i,k})$6, assumes acceptance at any recommended venue, and notes that richer covariates such as abstract text, author features, and acceptance probabilities would make the assumptions more plausible (Sato et al., 2020).

3. Publication-period choice in research assessment

In research-evaluation settings, the publication choice problem concerns the length of the publication period over which productivity should be assessed. The two competing objectives are reliability of performance rankings and frequency of evaluations. Longer windows smooth random year-to-year fluctuations; shorter windows provide more frequent feedback for policy, funding, and strategy. The empirical question is where the trade-off stabilizes.

The study of Italian university scientists in the hard sciences uses a population of 30,611 researchers continuously on faculty between 2003 and 2008 and 196,996 Web of Science publications, with citations counted as of June 30, 2009. The analysis covers nine university disciplinary areas and 181 scientific disciplinary sectors, excluding sectors with fewer than 10 faculty or with under 50% of members publishing at least once during 2003–2008. Researcher productivity is measured by Fractional Scientific Strength,

$a_i=(a_{i,1},\dots,a_{i,k})$7

where citations are standardized by the Italian median for the relevant year and subject category, then adjusted for coauthorship and, in life sciences, author position. Researchers are ranked within their scientific disciplinary sector on a percentile or quartile scale.

Window length is evaluated through contiguous and overlapping comparisons. Ranking variation is measured by average quartile shift and by Spearman rank correlation,

$a_i=(a_{i,1},\dots,a_{i,k})$8

Empirically, average quartile shift declines as $a_i=(a_{i,1},\dots,a_{i,k})$9 increases: for $y_w$0, it ranges from $y_w$1 to $y_w$2 quartiles depending on the disciplinary area; for $y_w$3, from $y_w$4 to $y_w$5; and for $y_w$6, from $y_w$7 to $y_w$8, with several areas below $y_w$9. In physics, Spearman correlation rises from $n$0 for one-year windows to $n$1 for two-year windows and to above $n$2 for three-year windows. Similar trajectories are reported for chemistry, biology, and medicine, while engineering and earth sciences stabilize more slowly.

The main result is a threshold at $n$3 years. Plots of average quartile shift against $n$4 display a clear elbow at three years, especially in physics, chemistry, biology, and medicine. Beyond that point, gains in stability are modest. Annual windows exhibit an average shift of about $n$5 quartiles, biennial windows about $n$6–$n$7, and triennial windows about $n$8, indicating that the random component has substantially diminished by three years. The recommended solution is therefore a triennial publication window, possibly with overlapping triennial windows when more frequent feedback is desired. The study also notes discipline-specific caveats: some engineering and earth-science areas may warrant $n$9, while rapidly evolving fields may tolerate shorter windows if qualitative measures are added (Abramo et al., 2018).

This formulation shows that publication choice need not refer to venue selection at all. It can instead denote the choice of the observation interval that determines how publication records are interpreted by evaluators. A common misconception in this setting is that shorter windows necessarily provide more accurate or more responsive assessment; the reported evidence states that windows under three years are heavily influenced by randomness, publication lag, and citation delay (Abramo et al., 2018).

4. Strategic and collective models of venue preference

A game-theoretic treatment models publication choice as a two-way interaction between researchers and venues. There is a continuum mass $X_i\in\mathbb R^d$0 of researchers and a finite set of venues $X_i\in\mathbb R^d$1. Researcher types lie in $X_i\in\mathbb R^d$2 with population masses $X_i\in\mathbb R^d$3. A type-$X_i\in\mathbb R^d$4 agent chooses a nonnegative vector of publication counts $X_i\in\mathbb R^d$5 subject to the budget constraint

$X_i\in\mathbb R^d$6

Given venue impacts $X_i\in\mathbb R^d$7, utility is

$X_i\in\mathbb R^d$8

with $X_i\in\mathbb R^d$9 and $T_i\in\mathcal T=\{1,\dots,K\}$0. Venue impact is endogenous: $T_i\in\mathcal T=\{1,\dots,K\}$1

Within this model, equilibrium is a pure-strategy Nash equilibrium in the large-population sense. Existence follows by defining a continuous fixed-point map on $T_i\in\mathcal T=\{1,\dots,K\}$2 and applying Brouwer. In the binary-type case, uniqueness is obtained under a Monotone-Cost-Ratio assumption through a characteristic equation $T_i\in\mathcal T=\{1,\dots,K\}$3 in the action ratio $T_i\in\mathcal T=\{1,\dots,K\}$4. The best response has closed form: $T_i\in\mathcal T=\{1,\dots,K\}$5 with proportionality chosen to satisfy the budget. Under monotone cost ratios, higher-impact types invest strictly more on the top venue. The same framework is extended to spotlight labeling, where a fraction $T_i\in\mathcal T=\{1,\dots,K\}$6 of papers receive an impact multiplier $T_i\in\mathcal T=\{1,\dots,K\}$7. In the binary-type case, there is again a unique equilibrium, and a threshold venue index $T_i\in\mathcal T=\{1,\dots,K\}$8 separates two comparative statics: for strongly competitive venues, spotlight labeling lowers the equilibrium impacts of all regular sessions; for less competitive venues, it raises them (Wang et al., 17 Nov 2025).

Empirical work on publication preferences studies whether actual fields resemble coordinated or fragmented equilibria. An adaptive survey of 3,510 tenure-track and tenured faculty at U.S. PhD-granting institutions yields 163,002 pairwise comparisons across 8,044 venues. Respondents self-identified into 19 broad fields, with analysis focusing on the 13 with at least 100 respondents. Field-level consensus is measured by pairwise-prediction accuracy based on field SpringRank scores,

$T_i\in\mathcal T=\{1,\dots,K\}$9

and fragmentation by

$Y_i\in\mathbb R_+$0

Economics, Chemistry, and Physics display strong consensus; Computer Science and Engineering display high fragmentation. Consensus scores range from $Y_i\in\mathbb R_+$1 in Computer Science to $Y_i\in\mathbb R_+$2 in Economics, and mean individual-versus-field top-5 overlap ranges from $Y_i\in\mathbb R_+$3 in Computer Science to $Y_i\in\mathbb R_+$4 in Economics. For the $Y_i\in\mathbb R_+$5 of comparisons in which both venues have a Journal Impact Factor, respondents choose the higher-JIF venue only $Y_i\in\mathbb R_+$6 of the time, while field-consensus SpringRank predicts $Y_i\in\mathbb R_+$7 correctly on the same subset. Preferences correlate positively with institutional prestige, $Y_i\in\mathbb R_+$8 per prestige decile with $Y_i\in\mathbb R_+$9, and with male gender, $r:\mathbb R^d \longrightarrow \mathcal T$0 with $r:\mathbb R^d \longrightarrow \mathcal T$1, while career stage is not statistically different from zero. Respondents have published in $r:\mathbb R^d \longrightarrow \mathcal T$2 of their own top-5 venues but only $r:\mathbb R^d \longrightarrow \mathcal T$3 of their field-consensus top-5, a difference of $r:\mathbb R^d \longrightarrow \mathcal T$4 venues with $r:\mathbb R^d \longrightarrow \mathcal T$5 after BH correction (Buskirk et al., 28 Feb 2026).

Taken together, these works separate two notions often treated as identical: venue impact as an equilibrium object and venue quality as a consensus object. A plausible implication is that a field with low consensus and high fragmentation, such as Computer Science in the survey data, is structurally less amenable to simple prestige ordering than a field with concentrated agreement. Another is that Journal Impact Factor is an incomplete proxy for the publication-choice problem because it explains only part of the choice behavior observed in the survey (Wang et al., 17 Nov 2025, Buskirk et al., 28 Feb 2026).

5. Publication mechanisms, peer review, and journal design

Publication choice is also shaped by the mechanism through which manuscripts are reviewed and accepted. An agent-based model compares a Current System, in which authors choose journals, to an Alternative System, in which journals bid on manuscripts. The model contains 500 authors, 50 journals, and single-authored papers over 120 months. Authors are characterized by Beta distributions over topic preference, technical quality, and novelty; journals are characterized by corresponding Beta distributions and an impact

$r:\mathbb R^d \longrightarrow \mathcal T$6

Editorial acceptance probability is

$r:\mathbb R^d \longrightarrow \mathcal T$7

In the Current System, authors rank journals by $r:\mathbb R^d \longrightarrow \mathcal T$8 and may resubmit up to five times after rejection. In the Alternative System, authors deposit manuscripts, commit to reviewing three manuscripts in the following month, papers are reviewed exactly three times, revised once, and then journals bid. Over ten-year runs, the Current System produces 14,936 manuscripts, of which 9,526 are published ($r:\mathbb R^d \longrightarrow \mathcal T$9); the Alternative System produces 14,683 manuscripts, of which 14,683 are published ($i$0), with 301 abandoned. Mean reviews per manuscript are 10.02 total in the Current System versus 3.00 fixed in the Alternative System. Median publication delay is 21.3 months in the Current System and 8.65 months in the Alternative System. In the Alternative System, 92% of authors publish more, 58% of journals increase output, and author impact rises for all authors. At the same time, mean published manuscript merit $i$1 is $i$2 in the Current System and $i$3 in the Alternative System, while abandoned-manuscript merit is $i$4 and $i$5, respectively. Allowing ten submissions in the Current System raises publication to $i$6 but increases delay to 32.5 months and approximately 14.2 reviews per paper (0911.0344).

A different mechanism-design perspective asks how a journal should choose an acceptance rule when authors select both research design and possible manipulation. The journal chooses $i$7, the probability of publishing a study with design $i$8 and reported result $i$9. Authors choose a design $L$00 at cost $L$01, generating

$L$02

then may manipulate by reporting $L$03 at cost $L$04. Upon publication, the audience forms the posterior mean

$L$05

The journal’s loss is

$L$06

and the author’s payoff is

$L$07

Several comparative-statics results follow. When manipulation is not possible but research costs are substantial, it may be optimal to incentivize cheaper designs even if they are less accurate. When manipulation is possible, the optimal rule publishes some manipulated results and also some results that would not have received attention absent manipulability. Even when pre-registration could deter manipulation, it is suboptimal to do so if experiments entail high research costs. With a linear manipulation cost $L$08, a sharp cutoff induces bunching at the acceptance threshold, while an optimal randomized rule can smooth incentives in a narrow band below a higher cutoff $L$09 (Jagadeesan et al., 29 Apr 2025).

These mechanism studies show that throughput, merit selection, reviewer load, and manipulation deterrence need not move together. Faster publication and lower referee burden can coexist with lower average published merit in simulation, and full suppression of manipulation can be suboptimal when it discourages costly but informative research designs. The publication-choice problem is therefore also a design problem for the institutions that mediate publication (0911.0344, Jagadeesan et al., 29 Apr 2025).

6. Publication-biased incentives and the choice of sample size

A further extension shifts the object of choice from venue or mechanism to study design under publication bias. In a simple economic model, a scientist chooses sample size $L$10 for a study testing $L$11 against $L$12 with standardized effect size $L$13. Each sample costs $L$14, and income $L$15 is earned only from statistically significant positive findings. If $L$16 is the prior probability that the effect exists, $L$17 is Type I error, and $L$18 is power, the probability of publication is

$L$19

and expected profit is

$L$20

The first-order condition for the optimal sample size $L$21 is

$L$22

The model implies that economically rational sample size is small when effects have low base probability, small effect size, or low grant income per publication. Using the normal approximation for a two-sided test,

$L$23

the qualitative result is that $L$24 grows with $L$25 and $L$26, and shrinks with $L$27. With heterogeneous projects drawn from plausible distributions over $L$28, $L$29, and $L$30, the model yields a strongly bimodal distribution of attained power and low overall positive predictive value. The positive predictive value in a niche is

$L$31

The same framework examines conditional equivalence testing, in which a non-significant result may still become publishable if an equivalence test succeeds. Total publishable rate under CET adds publishable negative outcomes to the payoff-relevant outcome set, thereby increasing the marginal revenue of extra samples and shifting $L$32 upward. The model reports that even a modest equivalence bound can markedly increase power and PPV (Braganza, 2019).

This strand generalizes the publication-choice problem beyond venue selection. The publication system determines not only where papers are sent, but which experiments are economically rational to conduct in the first place. A common misconception is that underpowered studies merely reflect methodological error. The model instead states that, under positive publication bias and per-sample costs, small sample sizes can be economically rational. In that sense, publication choice operates upstream of publication itself, through incentive effects on research design (Braganza, 2019).