Current projects

Resilience

I’ve been reading and thinking a lot about the concept of resilience across different disciplines. I am presenting something related to this at the 2026 online Pacific APA. Abstract below:

In board terms, operationalism is a project aimed at clarifying, concretizing, and arguably reducing the meaning of a concept into its measurement procedures. Despite its long-standing critical reception, a recent wave of defenders of operationalism advocate that aspects of the operational analysis can be recovered to solve specific problems in the philosophy of science surrounding theoretical entities. Both the criticisms and the defences of operationalism share a central assumption, which is that operationalization, when done carefully, is capable of capturing a substantive portion of the intuitive understanding of a concept, though perhaps not all of it. This paper challenges that assumption by arguing that passing the usual criteria for operational success – predictive efficacy, test reliability, cross-context applicability, etc. – does not entail that our sense of the concept’s meaningfulness comes from this operational success. I make this argument by looking at the case study of resilience and argue that operationalizations have consistently failed to provide or clarify meaning for resilience. To the extent that we take resilience to be a meaningful concept, the conceptual content comes from either our extra-operational association with the word “resilience” or its overlap with other, already-established concepts, rather than what we have learned from its successful operationalizations. This does not mean that the concept of resilience is thereby epistemically suspect, but it does mean that its operational success ought not to be taken as corroboratory evidence for its epistemic legitimacy. More broadly, the case study aims to caution philosophers and scientists alike from being too immediately impressed by a concept’s operational success.

Ground Truths Are Theory-Laden: Is That a Problem?

The concept of ground truth is crucial to statistics and machine learning. Ideally, in supervised machine learning, the algorithm is “supervised” by a ground truth training dataset, which presents the algorithm with facts which it is meant to learn. Ideally, after learning, the algorithm is tested against a ground truth benchmark dataset, which assesses its performance against facts. Through an examination of the discussion around ground truth across multiple scientific disciplines, the current paper attempts to accomplish two tasks. The first is to restate, in philosophical terms, a conclusion that has already been voiced by several scientists: ground truth is always theory-laden. The second is to ask whether this should lead to the usual problems associated with theory-laden observations. The answer is going to be that it is a problem if we expect ground truth data to be verificatory, but not if it is only corroboratory.

Data, Data Quality, and the Garbage In Garbage Out Principle

Data quality, or the lack thereof, is often blamed for inferential failures. But what is data quality? This paper pulls together multiple threads in the philosophy of data and scientists’ complaints about data quality under the banner “garbage in, garbage out”. I show how there exist conflicting considerations in theorizing about data quality that pull in different directions. I conclude that we ought to be cautious in our discussions of data quality as what a data gatherer mean by good data may, justifiably, be very different from what a data user expects.

I am presenting this project at the 2026 Central APA.

On bias-in, bias-out (precise title redacted for blind review)

Abstract. Discussions of algorithmic bias often assume, without reflection, that biases in predictive outcomes reflect biases in society. That is, machine learning algorithms are biased because they are bias-preserving processes. This paper challenges this assumption by pointing to a tension between the philosophy of data literature, which largely rejects the idea that data possess essential features that are always preserved during analysis, and the claim that social biases are easily preservable despite attempts at getting rid of them.

What is model interpretability good for?

When scientists fit statistical models to a set of data, it is often the case that there are still multiple options available even after the standard statistical criteria (fit, parsimony, etc.) have been applied. One strategy to break such ties is to interpret the model in non-statistical language by, for example, reading into the content of the variables or connecting the model with existing theory. Some have justified this strategy along the lines that research ought to be theory-driven, while others worry that this is a dressed up version of the old HARKing problem (“hypothesizing after results are known”). In this project, I evaluate the two options and reject them both. On the one hand, I do not think that model interpretability serves the function expected by theorists. On the other hand, I do not think it is quite the same as HARKing or other questionable research practices.

I presented this project at the 2025 EPSA. It’s currently still just ideas in my head. I’m happy to chat about it.

Complimentary science and the person-situation debate

This is a project in collaboration with Mike Schneider. Abstract below.

An understudied component of the social-intellectual dynamics of science is the languish of empirically live options of scientific pursuit. One notable exception is Hasok Chang’s work on “complementary science”, a proposal whereby historical research methods and philosophical scrutiny interact to yield opportunities for superior scientific knowledge concerning past cases. We observe that there is a viable generalization of Chang’s concept of complementary science, which also scopes over contemporary cases. On this conceptualization, one is primed to consider as well the social-intellectual dynamics that are triggered when some scientists abruptly pick up such a live option, breaking with the mainstream. Drawing on examples from both physics and psychology, we argue that complementary science engenders dispute within a scientific community about implicit commitments of the mainstream. We point out that these disputes are distinct from, but nonetheless evocative of, social-intellectual dynamics that Kuhn famously assigns to periods of crisis.

Past projects

What Are Statistical Modeling Assumptions About? An Answer From Perspectival Pluralism (published, 2025)

Abstract: This paper presents a perspectivist framework for understanding and evaluating statistical assumptions. Drawing on the thesis of perspectivism from the philosophy of science, this framework treats statistical assumptions not as empirical hypotheses which are descriptively accurate or inaccurate about the world but as prescribing a particular perspective from which statistical knowledge is generated. What this means is that we ought not judge statistical models solely by how closely they correspond with the world as we independently understand it, but by whether they paint a picture of the world that is epistemically significant.

This paper is published Open Access in Harvard Data Science Review.

Measuring the non-existent: validity before measurement (published, 2023)

Abstract: This paper examines the role existence plays in measurement validity. I argue that existing popular theories of measurement and of validity follow a correspondence framework, which starts by assuming that an entity exists in the real world with certain properties that allow it to be measurable. Drawing on literature from the sociology of measurement, I show that the correspondence framework faces several theoretical and practical challenges. I suggested the validity-first framework of measurement, which starts with a practice-based validation process as the basis for a measurement theory, and only posits objective existence when it is scientifically useful to do so.

This paper is published Open Access in Philosophy of Science.

I made a video abstract for this paper, available here

Sample Representation in the Social Sciences (published 2021)

Abstract: The social sciences face a problem of sample nonrepresentation, where the majority of samples consist of undergraduate students from Euro-American institutions. The problem has been identified for decades with little trend of improvement. In this paper, I trace the history of sampling theory. The dominant framework, called the design-based approach, takes random sampling as the gold standard. The idea is that a sampling procedure that is maximally uninformative prevents samplers from introducing arbitrary bias, thus preserving sample representation. I show how this framework, while good in theory, faces many challenges in application. Instead, I advocate for an alternative framework, called the model-based approach to sampling, where representative samples are those balanced in composition, however they were drawn. I argue that the model-based framework is more appropriate in the social sciences because it allows for systematic assessment of imperfect samples and methodical improvement in resource-limited scientific contexts. I end with practical proposals of improving sample quality in the social sciences.

A post-peer-review, pre-copyedit version of this article, published in Synthese, can be found here. The final authenticated version is available online at: http://dx.doi.org/10.1007/s11229-020-02621-3

A Statistical Learning Approach to a Problem of Induction (unpublished)

Abstract: One “easier” form of the problem of induction questions our ability to pick out true regularities in nature, using limited data, with the assumption that such regularities do exist. Harman and Kulkarni (2012) take this problem to be a challenge on our ability to identify precise conditions under which the method of picking hypotheses based on limited datasets is or is not reliable. They identify an influential result from statistical learning theory, hereafter referred to as the VC theorem (Vapnik and Chervonenkis, 2015), which states that, under the condition that the starting hypotheses set has finite VC dimension, the hypothesis chosen from it converges to the true regularity as the size of the dataset goes to infinity.

This result seems to provide us with a condition (i.e., having finite VC dimension) under which a method (i.e., choosing a hypothesis based on its performance over data), is reliable. Indeed, Harman and Kulkarni take this result to be an answer to the form of the problem of induction they have identified. This paper examines this claim. By discussing the details of how VC theorem may be construed as an answer and the connection between VC theorem in statistical learning theory and the NIP property in model theory, I conclude that the VC theorem cannot give us the kind of general answers needed for Harman and Kulkarni’s response to the problem of induction.

A shorter version of the draft that was presented in the 2018 PSA meeting can be found here.

Intuitionistic Probabilism in Epistemology (unpublished)

Abstract: This paper examines the plausibility of a thesis of probabilism that is based on intuitionistic logic and exposits the difficulties faced by such a program. The paper starts by motivating intuitionistic logic as the logic of investigation along a similar reasoning as Bayesian epistemology. It then considers two existing axiom systems for intuitionistic probability functions — that of Weatherson (2003) and of Roeper and Leblanc (1999) — and discusses the relationship between the two. It will be shown that a natural adaptation of an accuracy argument in the style of Joyce (1998) and de Finetti (1974) to these systems fails. The paper concludes with some philosophical reflections on the results.

The paper has not been published. You can read a draft of it here. It was presented at the 2018 Philosophy of Logic, Mathematics, and Physics Graduate Conference (LMP) at the University of Western Ontario.