Skip to content

Methodological foundations

Why theoryforge

Psychology and the social sciences accumulate findings far faster than they build theories that explain them. Constructs proliferate and blur into one another (the jingle-jangle problem), predictions stay vague and directional where they ought to commit, auxiliary assumptions accrete to shield a theory from the tests it should face, and a revision is rarely judged by whether it adds testable content or merely deflects refutation. Eronen and Bringmann (2021) call this the theory crisis.

theoryforge treats a theory as a structured, versioned object and makes the specification of these properties checkable, where prose left it to the reader's judgement. It scores the theory against twelve structural criteria drawn from the methodology literature, screens its constructs for redundancy, grades how risky each prediction's claim is, distinguishes progressive from degenerating amendments in Lakatos's sense, and renders the structure so it can be read and tested. Every value it reports is produced by a fixed, documented rule, with no model and no randomness, and the R and Python implementations return identical verdicts and byte-identical diagram intermediate representations, so a verdict is reproducible and can be audited against the specification. See Scope and limits below for what this checking does and does not establish.

Each item in the rigour checklist follows from a result in the methodology literature, and the package records the supporting work next to the check (the citation field of the report). The supporting works are listed in APA style below.

Grounding for each rigour check

The twelve checks are set out below with the criterion each one applies and the methodological work that motivates it. The name in brackets is the identifier the check carries in the report, and the supporting works are given in full in the references at the foot of the page. The criteria are quoted from the checklist file the package ships, read when this page is built.

Rigour check Criterion Supporting work
Falsifiability (falsifiability) At least one prediction forbids an observation Popper (1959), Bacharach (1989)
Predictive precision (precision) Predictions are point/interval, not merely directional Meehl (1967), Meehl (1990a)
Prediction severity (risk_severity) Mean prediction severity (declared, else the claim-form rubric) at or above threshold Popper (1959), Meehl (1967), Meehl (1990b)
Parsimony (parsimony) No auxiliary assumption added in response to an anomaly lacks independent corroboration Lakatos (1970), Meehl (1990a)
Construct non-redundancy (non_redundancy) No pair of construct definitions is a near-duplicate or contained in the other (lexical screen) Kelley (1927), Le et al. (2010), Lawson & Robins (2021)
Construct clarity (construct_clarity) Every construct has definition + measurement + boundary conditions Suddaby (2010), Cronbach & Meehl (1955), Flake & Fried (2020)
Scope and boundary conditions (scope) Boundary conditions explicitly stated Whetten (1989), Bacharach (1989)
Mechanism (logical_why) Each proposition states a mechanism, not just a correlation Sutton & Staw (1995), Whetten (1989)
Causal testability (causal_testability) At least one proposition states a causal relation Textor et al. (2016), Eronen & Bringmann (2021)
Diagnosticity (diagnosticity) At least one prediction names a registered alternative it is declared to discriminate from (declared, not verified) Platt (1964), Fiedler (2017)
Formalisation (formalisation) A formal model of a recognised type (ode, abm, network, sem) is declared Robinaugh et al. (2021), Guest & Martin (2021)
Derivation chain (derivation_chain) Every prediction cites at least one proposition, and every cited id is a declared proposition (reference check only) Scheel et al. (2021), Szollosi et al. (2020)

How the rigour score is computed

Each of the twelve items returns a status (pass, warn or fail) and a score between 0 and 1. Present-or-absent items score 1 or 0. Items that measure a proportion (for example, the share of constructs that carry a definition, a measurement and boundary conditions) score that proportion. An item can therefore pass with a score below 1: predictive precision passes from a share of 0.5, and diagnosticity once a single prediction qualifies. An item with nothing to assess has the status n/a and no score. That is the case for construct non-redundancy when the theory has fewer than two constructs, for the derivation chain when it has no predictions and for parsimony when no auxiliary assumption answers an anomaly.

Five default thresholds come from the checklist file. Predictive precision passes at a share of 0.5 and prediction severity at a mean of 0.5. The redundancy screen flags a pair of definitions at a Jaccard similarity of 0.85, or at an overlap coefficient of 0.85 when both hold at least three tokens. The optional embedding screen flags a pair at a cosine of 0.85.

The headline number, the aggregate score, is the weighted mean of the applicable items' scores, scaled to a 0 to 100 range. The weights are fixed in the checklist file. Falsifiability carries the most weight (0.15) and formalisation the least (0.05). The report also gives the coverage, the share of the checklist's weight that was applicable: 0.92 when parsimony alone has nothing to assess. Leaving such an item out keeps a theory from scoring for what it lacks. When the three items scored 1 with nothing to assess, an empty theory scored 26 and adding a prediction lowered its score. The aggregate is compensatory: a high score on one item offsets a low score on another, and a failed blocker counts like any other item. A blocked theory can therefore outscore one whose gate passes, and the score is best read together with the gate and the item table.

The gate is the one-word verdict. Two items are blockers: falsifiability and the derivation chain. If either blocker fails, the gate is blocked. A theory at the draft maturity stage is always advisory, so the blockers inform without gating early work. Otherwise the gate is pass.

Each item is scored as follows.

  • Falsifiability passes if at least one prediction is point, interval or directional, that is, a claim that forbids some observation.
  • Predictive precision is the share of predictions that are point or interval, and passes at 0.5 or above.
  • Prediction severity averages one value per prediction, its declared severity where it has one and otherwise the computed severity of the severity() rubric, and passes at 0.5 or above. With no predictions, it warns with a score of 0. A declared value counts as given, and the dossier prints it beside the rubric's value, with a note when it exceeds that value by more than 0.2. The rubric's value reads the same type label as predictive precision and the same named rival as diagnosticity, so a prediction that declares no severity is credited for those features twice. The item grades claims before any data: the severity of a test in Mayo's (2018) sense depends on the design and the data, which no item reads.
  • Parsimony assesses only the auxiliary assumptions added in response to an anomaly, those whose added_for names the prediction they answer. It fails if any of them is ad hoc, that is, if no prediction it protects other than that anomaly is corroborated (Lakatos, 1970), and passes otherwise. With no such assumption, it is n/a. Core assumptions do not count against a theory, since every derivation relies on auxiliaries (Meehl, 1990a) and declaring them makes the theory's protective belt visible. With no prior version to compare, every prediction an assumption protects counts as support, so the amendment appraisal is the stricter check, since it counts only the content an amended version adds.
  • Construct non-redundancy passes when the redundancy screen flags no pair of construct definitions, and warns with a score of 0 otherwise. The screen flags near-duplicate definitions and a definition contained in another, as described below. With fewer than two constructs, it is n/a.
  • Construct clarity is the share of constructs that carry a definition, a measurement and boundary conditions, and passes only at 1.
  • Scope passes when boundary conditions are stated, either for the theory or on every construct.
  • Mechanism is the share of propositions whose statement names a mechanism, a bare correlation being no help, and passes only at 1.
  • Causal testability passes if at least one proposition states a causal relation, that is, one of the five directed relations: increases, decreases, causes, mediates or moderates. The item checks that such a relation is present and nothing more. The causal_dag export, which takes the five directed relations as arrows and an association between two of their constructs as a bidirected edge, is not checked for acyclicity: a theory whose causal relations form a feedback loop, as the panic-network example does, is emitted as written. dagitty accepts such a graph but reads it by d-separation, which a model with feedback is guaranteed to satisfy only in special cases, a linear model among them (Bongers et al., 2021). Reading the graph for the conditional independencies it implies is the job of implications(), which checks for acyclicity and by default refuses a cyclic graph. With the sigma option it reads a cyclic graph by sigma-separation, whose statements hold when each feedback loop has a unique equilibrium (Bongers et al., 2021), and it then derives one statement for the panic-network example. The shipped modality-switching example is acyclic and gets the same statements either way.
  • Diagnosticity is the share of predictions that name a registered alternative they are declared to discriminate from, and passes once at least one does. The alternative is declared, not verified. The theory records nothing the rival predicts, so no function checks that the prediction would tell the two apart, and the 0.1 bonus that the severity rubric gives such a prediction rests on the same declaration.
  • Formalisation passes if the theory declares a formal model of a recognised type, ode, abm, network or sem. Any other type, none included, does not count.
  • Derivation chain is the share of predictions that cite at least one proposition and cite only declared propositions, and passes only at 1. Otherwise, it fails. It checks references only, so it does not show that a prediction follows from what it cites. With no predictions, it is n/a.

Scope and limits

The checklist is a structural and lexical screen, not a substantive review. Each item asks whether a required property is present in the theory object, such as a boundary-conditions field or a mechanism string, or what share of a collection exhibits it. The non-redundancy screen additionally measures the lexical overlap between construct definitions. Being lexical, it cannot detect empirical redundancy: two constructs that correlate almost perfectly once measurement error is corrected for may be defined in different words (Le et al., 2010), and telling them apart needs data, as in Rönkkö and Cho's (2022) guideline for assessing discriminant validity. No item reads or judges the content of a definition, a mechanism or a boundary condition. A theory can score highly by being completely and precisely specified while resting on a false premise, and can score poorly for being under-specified while resting on a sound one. The aggregate score indexes how completely a theory is specified against the checklist. It says nothing about whether the theory is true or well-reasoned.

The weights that combine the twelve items into the aggregate score, and the five numeric thresholds (the two redundancy thresholds, the embedding threshold, the minimum precision share and the minimum mean severity), are the package's own defaults. They are not derived from an external validation study. They are fixed in the shipped checklist file, and every report records that file's version as checklist_version, so scores stay comparable across theories. The package offers no option to substitute other weights or thresholds. A field with different norms can recombine the per-item scores the report lists, but the result is not comparable with the default aggregate. The gate reflects only two of the twelve items, falsifiability and the derivation chain. A theory can show gate: pass while several other items are warn or fail, so the gate reports a minimal precondition for testing and the item table carries the rest. Read the two together with the aggregate score.

How the other outputs are computed

Every prediction declares one of four types, and both the rigour checklist and the severity rubric read that label. An existence prediction asserts that an effect or relation exists, without a direction. A directional prediction asserts a sign or an order, which covers comparisons, interactions, the invariance of a direction across groups and claims that an effect occurs only when a condition holds. An interval prediction asserts that a quantity lies in a stated range, the range the theory permits. A point prediction asserts one value, with the tolerance that measurement requires. Its width is measurement tolerance, not latitude the theory allows, and that is what separates it from an interval. The labels are self-declared, and no function checks a statement against its type.

The severity rubric ranks the form of each prediction's claim before any data. It gives each prediction a base risk fixed by its type: existence 0.1, directional 0.4, interval 0.7 and point 0.9. The order follows Popper's (1959, sections 31–33) comparison of falsifiability: of two nested claims about one quantity, the one that forbids more is the more falsifiable. The reported risk score is this base. The computed severity then adjusts it. A merely directional prediction is discounted by 25 per cent, following Meehl's (1967, 1990b) argument that a bare sign risks little because almost everything in psychology correlates a little. A prediction that names a registered alternative in its diagnostic_vs field, declaring that it would discriminate between the theories, earns a bonus of 0.1. The declaration is not verified, since nothing in the theory records what the rival predicts. The computed severity is the discounted base plus the bonus, capped at 1. The four base values, the discount and the bonus are the package's conventions, and neither Popper nor Meehl gives them. The .25 and .30 that Meehl (1990b) mentions are sizes of those ambient correlations, a different quantity from the discount. Because the rubric reads only the declared type and the named alternatives, it says nothing about how severely a claim has been tested, which depends on the design and the data.

Post-data severity reads what the rubric leaves out. A prediction may state the region its claim allows in bounds, a lower and an upper value, either of which may be left open. A test outcome may record the estimate the test observed and its standard error se. For each such outcome, post_data_severity() gives the severity with which the claim passed, in Mayo and Spanos's (2006, 2011) sense. It is the probability that the test would have produced a result less in accord with the claim, had the claim been false at its nearer bound. Under a normal approximation to the sampling distribution of the estimate, it is the smaller of Φ((estimate − lower) / se) and Φ((upper − estimate) / se) over the finite bounds. A one-sided claim, such as an effect above a smallest effect of interest, leaves the upper bound open. With two finite bounds, the value equals 1 minus the p value of the two one-sided tests of equivalence. Take a claim that a correlation is at most .10 in absolute value, tested with an estimate of 0. It passes with a severity of 1.0, .841 and .691 at standard errors of .01, .10 and .20, where the rubric gives the claim one value whatever the test. A point claim, whose two bounds coincide, can pass with a severity of at most .5, so it needs a tolerance around its value. The rigour checklist does not read the value, and the dossier gives it in the test record. The modality-switching example records one outcome, with values labelled as illustrative, to show the computation.

A prediction may also record the sampling plan of its test in a test block. The block holds the planned sample size and what it counts, alpha and whether the test is one- or two-sided, and the smallest effect of interest and its scale. It also holds the planned power or decision probability, with an interval and the method that gave it, a reference to the design analysis and a justification. A Stage 1 Registered Report asks for such a plan for each hypothesis. The preregistration lists the plans after the hypotheses, and no function scores them. As precision grows, the probability of a significant difference in the predicted direction approaches one half even when the theory is false (Meehl, 1967). The corroboration a directional prediction earns therefore weakens as power grows, and power must not raise a claim's severity. Power does decide how informative a failed prediction is. A test unlikely to detect the smallest effect of interest says little when it misses that effect (Lakens, 2022), and the reader of the preregistration judges this from the plan. The modality-switching example records a plan for its interval prediction, with values labelled as illustrative.

The construct redundancy screen reduces each construct definition to a set of content tokens. The text is first folded by a fixed table that both twins share. Accented Latin letters go to their ASCII base (é to e, ß to ss), Greek and Cyrillic letters to small unaccented ones, and ASCII letters are lowercased. The text is then split on anything that is not a letter, mark or digit in any script, and tokens shorter than three characters or in a small English stop-word list are dropped. Other scripts are compared as written, so case, accents and the Unicode normalisation form count there, and text written without spaces, such as Chinese, forms one token per run. The similarity of two constructs is the Jaccard index of their token sets, the size of the intersection divided by the size of the union, and their overlap is the overlap coefficient, the size of the intersection divided by the size of the smaller set. The screen reports every construct pair with both values. It flags a pair for review when the similarity reaches 0.85, which marks a near-duplicate, or when both definitions hold at least three tokens and the overlap reaches 0.85, which marks one definition contained in the other. The two motivation constructs of the weak example are such a pair: one definition adds "towards goals" to the other, so their similarity is 0.8 and their overlap 1. This is a deliberately simple lexical screen, and its two thresholds are default working values. An optional embedding-based screen is available for a semantic comparison. It has a cosine threshold of its own, since how high a cosine runs depends on the embedding model.

The amendment appraisal compares an amended theory with a prior version by what the two claim, not by the identifiers of their predictions. Lakatos (1970, p. 118) calls a problemshift progressive when the new theory predicts a fact its predecessor did not and some of that excess content is corroborated, and degenerating when it is not. The appraisal turns this into rules. A prediction that changed only its identifier, with the same statement and type, is reported as renamed and is not new. A new prediction derived only from propositions the prior already held, with the same endpoints and relation, articulates old content and is reported as such. A new auxiliary assumption is new content, so a prediction it protects is not an articulation. A prediction is corroborated when at least one test outcome passes it and none fails it, so a failed replication leaves it mixed. An assumption added in response to an anomaly names the prediction it answers in its added_for. It is ad hoc unless a prediction it protects that is new in this version, other than the anomaly, is corroborated, since accommodating the anomaly is no independent support. The verdict is progressive when a corroborated new prediction adds content, no assumption is ad hoc and no corroborated prediction was dropped. It is degenerating when an assumption is ad hoc and no corroborated prediction adds content. Lakatos's scheme is a dichotomy, and neutral is theoryforge's label for an amendment that meets neither of these two rules. The lists returned tell its cases apart. An amendment may add no predictions, only ones not yet tested (in new_predictions but not corroborated_new) or only articulations of old content (articulated). It may also set a corroborated advance against an ad hoc assumption or a dropped corroboration (ad_hoc_assumptions, dropped_corroborated). Dropped predictions that were never tested (content_lost) and new failures of retained ones (new_anomalies) are reported without changing the verdict. The appraisal reads a test outcome's registered only to report which corroborations were preregistered (corroborated_new_registered). It reads no date, no severity_at_test and no version field, so which version is the prior is the caller's responsibility.

The theory landscape sets the theory against a literature corpus and labels each theme by how many of the theory's accounts address it, counting the focal theory and its registered alternatives. An account addresses a theme when a word of its construct labels, or of an alternative's label and key constructs, appears among the theme's keywords, and the result lists the words that matched. Three kinds of word never match. Words carried by more than a set share of the records, half by default, belong to the whole field. Words of the theory's title name the phenomenon that every account explains, so a construct word that also appears in the title does not match. Words such as theory, model and account name a kind of account. A theme that no account touches is under-theorised, one that a single account touches is covered, and one that two or more touch is crowded. The labels count the registered accounts only, so an under-theorised theme may be addressed by an account the theory does not register. A crowded theme calls for predictions that discriminate between its accounts, and does not show that any of them is redundant. The themes are the connected components of the keyword co-occurrence graph. On a real corpus, a few keywords carried by most records join nearly every keyword into one component, and the package warns when one theme holds more than half the linked keywords. The simple-centres method of co-word analysis (Coulter et al., 1998; Cobo et al., 2011) bounds each theme instead. It weights each pair of keywords by its equivalence index and grows each theme from its strongest link to at most ten keywords. Each theme is then placed in the strategic diagram by its centrality and density (Callon et al., 1991), split at their medians.

The simulation makes each construct a state variable, and each directed proposition contributes a signed coupling between two states, positive for increases, causes and mediates and negative for decreases, scaled by the coupling gain. Moderates and associates couple nothing, and the functional_form field is not read. Every state also decays towards zero at the damping rate, and all states start from a common initial value, so constructs with the same incoming couplings follow the same trajectory. The trajectory is the sequence of state vectors. It is a deliberately transparent linear system, meant to expose the qualitative dynamics a network of propositions implies. Fitting data is not what it is for. A linear system has one equilibrium, or a continuum of them when its matrix is singular. It never has two separate ones, so it cannot show the bistability that formal models of panic disorder rely on (Robinaugh et al., 2024). In a network without directed cycles, every eigenvalue of the system equals minus the damping rate. Such a theory decays at that rate in the long run, but not in its transients: at the default gain and damping, the last construct of a five-construct chain rises to 5.77 times its initial value before it decays. The regime of a theory with feedback loops is set by the coupling gain against the damping. The record lists the propositions that couple nothing and any pair of constructs carrying couplings of both signs, which offset each other.

Two methods propagate the system. The exact method multiplies the state at every step by the matrix exponential of the system matrix times the step size, which it computes by scaling and squaring a Taylor polynomial (Moler & Van Loan, 2003). Each row is then the solution of the linear system at its time, whatever the step size. The Euler method, the default for this release, takes fixed explicit steps. Its step is stable only when |1 + dt·λ| < 1 for every eigenvalue λ of the coupling matrix minus the damping. In a network without directed cycles, once the step size times the damping exceeds 2, the states explode with alternating sign although the system they approximate decays. Raising the damping therefore does not cure a divergence and can cause one. A smaller step size cures it, as does a smaller coupling gain when the network's own feedback outgrows the damping. Euler steps also inflate a sustained oscillation into growth, so a run warns when its trajectory departs from the exact one by more than 5 per cent. The default will change to the exact method in the next minor release. A run whose values grow beyond what can be rounded stops with a message naming the step and the construct.

References

Bacharach, S. B. (1989). Organizational theories: Some criteria for evaluation. The Academy of Management Review, 14(4), 496–515. https://doi.org/10.5465/amr.1989.4308374

Bongers, S., Forré, P., Peters, J., & Mooij, J. M. (2021). Foundations of structural causal models with cycles and latent variables. The Annals of Statistics, 49(5), 2885–2915. https://doi.org/10.1214/21-AOS2064

Callon, M., Courtial, J. P., & Laville, F. (1991). Co-word analysis as a tool for describing the network of interactions between basic and technological research: The case of polymer chemistry. Scientometrics, 22(1), 155–205. https://doi.org/10.1007/BF02019280

Cobo, M. J., López-Herrera, A. G., Herrera-Viedma, E., & Herrera, F. (2011). An approach for detecting, quantifying, and visualizing the evolution of a research field: A practical application to the Fuzzy Sets Theory field. Journal of Informetrics, 5(1), 146–166. https://doi.org/10.1016/j.joi.2010.10.002

Coulter, N., Monarch, I., & Konda, S. (1998). Software engineering as seen through its research literature: A study in co-word analysis. Journal of the American Society for Information Science, 49(13), 1206–1223. https://doi.org/10.1002/(SICI)1097-4571(1998)49:13<1206::AID-ASI7>3.0.CO;2-F

Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957

Eronen, M. I., & Bringmann, L. F. (2021). The theory crisis in psychology: How to move forward. Perspectives on Psychological Science, 16(4), 779–788. https://doi.org/10.1177/1745691620970586

Fiedler, K. (2017). What constitutes strong psychological science? The (neglected) role of diagnosticity and a priori theorizing. Perspectives on Psychological Science, 12(1), 46–61. https://doi.org/10.1177/1745691616654458

Flake, J. K., & Fried, E. I. (2020). Measurement schmeasurement: Questionable measurement practices and how to avoid them. Advances in Methods and Practices in Psychological Science, 3(4), 456–465. https://doi.org/10.1177/2515245920952393

Guest, O., & Martin, A. E. (2021). How computational modeling can force theory building in psychological science. Perspectives on Psychological Science, 16(4), 789–802. https://doi.org/10.1177/1745691620970585

Kelley, T. L. (1927). Interpretation of educational measurements. World Book Company.

Lakatos, I. (1970). Falsification and the methodology of scientific research programmes. In I. Lakatos & A. Musgrave (Eds.), Criticism and the growth of knowledge (pp. 91–196). Cambridge University Press. https://doi.org/10.1017/cbo9781139171434.009

Lakens, D. (2022). Sample size justification. Collabra: Psychology, 8(1), Article 33267. https://doi.org/10.1525/collabra.33267

Lawson, K. M., & Robins, R. W. (2021). Sibling constructs: What are they, why do they matter, and how should you handle them? Personality and Social Psychology Review, 25(4), 344–366. https://doi.org/10.1177/10888683211047101

Le, H., Schmidt, F. L., Harter, J. K., & Lauver, K. J. (2010). The problem of empirical redundancy of constructs in organizational research: An empirical investigation. Organizational Behavior and Human Decision Processes, 112(2), 112–125. https://doi.org/10.1016/j.obhdp.2010.02.003

Mayo, D. G. (2018). Statistical inference as severe testing: How to get beyond the statistics wars. Cambridge University Press. https://doi.org/10.1017/9781107286184

Mayo, D. G., & Spanos, A. (2006). Severe testing as a basic concept in a Neyman–Pearson philosophy of induction. The British Journal for the Philosophy of Science, 57(2), 323–357. https://doi.org/10.1093/bjps/axl003

Mayo, D. G., & Spanos, A. (2011). Error statistics. In P. S. Bandyopadhyay & M. R. Forster (Eds.), Philosophy of statistics (pp. 153–198). Elsevier. https://doi.org/10.1016/B978-0-444-51862-0.50005-8

Meehl, P. E. (1967). Theory-testing in psychology and physics: A methodological paradox. Philosophy of Science, 34(2), 103–115. https://doi.org/10.1086/288135

Meehl, P. E. (1990a). Appraising and amending theories: The strategy of Lakatosian defense and two principles that warrant it. Psychological Inquiry, 1(2), 108–141. https://doi.org/10.1207/s15327965pli0102_1

Meehl, P. E. (1990b). Why summaries of research on psychological theories are often uninterpretable. Psychological Reports, 66(1), 195–244. https://doi.org/10.2466/pr0.1990.66.1.195

Moler, C., & Van Loan, C. (2003). Nineteen dubious ways to compute the exponential of a matrix, twenty-five years later. SIAM Review, 45(1), 3–49. https://doi.org/10.1137/S00361445024180

Platt, J. R. (1964). Strong inference. Science, 146(3642), 347–353. https://doi.org/10.1126/science.146.3642.347

Popper, K. R. (1959). The logic of scientific discovery. Hutchinson.

Robinaugh, D. J., Haslbeck, J. M. B., Ryan, O., Fried, E. I., & Waldorp, L. J. (2021). Invisible hands and fine calipers: A call to use formal theory as a toolkit for theory construction. Perspectives on Psychological Science, 16(4), 725–743. https://doi.org/10.1177/1745691620974697

Robinaugh, D. J., Haslbeck, J. M. B., Waldorp, L. J., Kossakowski, J. J., Fried, E. I., Millner, A. J., McNally, R. J., Ryan, O., de Ron, J., van der Maas, H. L. J., van Nes, E. H., Scheffer, M., Kendler, K. S., & Borsboom, D. (2024). Advancing the network theory of mental disorders: A computational model of panic disorder. Psychological Review, 131(6), 1482–1508. https://doi.org/10.1037/rev0000515

Rönkkö, M., & Cho, E. (2022). An updated guideline for assessing discriminant validity. Organizational Research Methods, 25(1), 6–47. https://doi.org/10.1177/1094428120968614

Scheel, A. M., Tiokhin, L., Isager, P. M., & Lakens, D. (2021). Why hypothesis testers should spend less time testing hypotheses. Perspectives on Psychological Science, 16(4), 744–755. https://doi.org/10.1177/1745691620966795

Suddaby, R. (2010). Editor's comments: Construct clarity in theories of management and organization. Academy of Management Review, 35(3), 346–357. https://doi.org/10.5465/amr.35.3.zok346

Sutton, R. I., & Staw, B. M. (1995). What theory is not. Administrative Science Quarterly, 40(3), 371–384. https://doi.org/10.2307/2393788

Szollosi, A., Kellen, D., Navarro, D. J., Shiffrin, R., van Rooij, I., Van Zandt, T., & Donkin, C. (2020). Is preregistration worthwhile? Trends in Cognitive Sciences, 24(2), 94–95. https://doi.org/10.1016/j.tics.2019.11.009

Textor, J., van der Zander, B., Gilthorpe, M. S., Liśkiewicz, M., & Ellison, G. T. H. (2016). Robust causal inference using directed acyclic graphs: The R package ‘dagitty’. International Journal of Epidemiology, 45(6), 1887–1894. https://doi.org/10.1093/ije/dyw341

Whetten, D. A. (1989). What constitutes a theoretical contribution? The Academy of Management Review, 14(4), 490–495. https://doi.org/10.5465/amr.1989.4308371