Philosophers on Philosophy: The 2020 PhilPapers Survey

2.1 Target population

In the 2009 survey, we were restricted to a relatively small group of departments for which we had faculty lists (mainly drawn from the Philosophical Gourmet Report’s faculty lists for ranked departments). In 2020, the PhilPeople database included information on philosophers and philosophy departments around the world (with strongest coverage on English-speaking and English-publishing philosophers), so we could survey a broader and more representative group.

After a period of consultation, we decided on a target group including:

(1) in Australia, Canada, Ireland, New Zealand, the UK, and the US: all regular faculty members (tenure-track or permanent) in BA-granting philosophy departments with four or more members (according to the PhilPeople database); and

(2) in all other countries: English-publishing philosophers in BA-granting philosophy departments with four or more English-publishing faculty members.

For the purposes of this study, we defined an English-publishing philosopher as someone with one or more publications (according to the PhilPapers database) in one of a wide range of English-language venues. We limited our coverage to English-publishing philosophers because we do not have adequate information on philosophers who do not publish in English, and because the majority of our questions are drawn from English-language traditions.

For meaningful longitudinal comparisons, we also designated a “2009-comparable departments” target group of 100 departments in the same regions as the 2009 survey, selected by similar criteria (all Ph.D.-granting departments with a 2017-2018 Philosophical Gourmet Report score of 1.9 or above, plus two leading departments with MA programs and a selected group of European departments based on expert recommendations). This group was used only for longitudinal comparisons.

We used data entry from departmental websites to supplement existing PhilPeople records and make our information as complete as possible. After data entry and cleanup, our target population included 7685 philosophers, including 6112 in group (1) and 1573 in group (2). The 2009-comparable target group included 2407 philosophers. The online data is imperfect, so our group of 7685 philosophers almost certainly excludes some philosophers who meet criterion (1) or (2) and includes some philosophers who do not.

Every member of the target group was sent an initial email invitation to take the survey, followed by additional email requests after approximately 10 days and 20 days if they had not yet responded.

2.2 Philosophical questions

In the 2009 survey, we asked 30 questions each with 2-4 answer options: for example, “God: theism or atheism?” and “Mind: physicalism or nonphysicalism?”.

In the 2020 survey, we used the 30 questions from the 2009 survey unaltered (although we made some answer options slightly more fine-grained, as discussed below), to allow meaningful longitudinal comparisons. We expanded the list of 30 questions to a list of 40 main questions that would be asked of all participants. We also added a further group of 60 additional (often more specialized) questions, each of which would be asked to one-sixth of participants (selected randomly). As a result, each participant was asked to answer a minimum of 50 questions (40 main questions and 10 additional questions). Participants were also given the option of answering some or all of the other 50 additional questions if they chose to, with a maximum of 100 philosophical questions per participant.

We determined the 70 new questions through an extended period of consultation, including consultation with PhilPapers editors and extended discussion on social media including PhilPeople, Facebook, and philosophy blogs. We also had a lengthy period of beta testing the survey questions and the survey interface with PhilPapers editors using the interface. We aimed for questions that covered many areas of philosophy, that worked in the multiple choice format, and that would be familiar to at least half of our target population.

The 100 resulting questions included approximately 50 questions drawn from metaphysics and epistemology (broadly construed), 30 questions drawn from value theory, 9 from the philosophy of science, logic, and mathematics, 6 from the history of philosophy, and 5 from metaphilosophy.

As in 2009, we did not include any questions drawn from non-Western and non-analytic traditions, as it proved too difficult to find questions from these traditions that met the familiarity and multiple-choice constraints. We attempted to include some new questions reflecting philosophy as it stands in 2020 (adding two questions each about gender and race, for example), but we acknowledge an overall bias toward certain relatively traditional issues in the analytic and English-speaking canons. In retrospect, we could have done more to reflect the diversity of contemporary philosophy. In future surveys, we will try to do so.

As in the earlier survey, we allowed respondents to indicate that they “accept” or “lean toward” a view, and we allowed a range of other options. The options are shown in Figure 1. We changed the 2009 answer options slightly to allow respondents more fine-grained options when endorsing multiple answers. Where the 2009 survey just had an option for “Accept both” (binary questions) or “Accept more than one” (ternary questions), the 2020 survey allowed respondents to accept, reject, or lean toward or against each answer separately if they chose to (as shown on Figure 2). We also allowed respondents to write in alternative answers if they chose to. Two questions, one about other minds and one about philosophical methods, were given special treatment because we didn’t expect a majority of respondents to choose a single answer to these questions. For these questions, respondents had to say whether they accept or reject each option individually as if they had selected “Evaluate multiple options”.

Figure 1:

Figure 1: Example survey form for one philosophical question.

Figure 2:

Figure 2: Selecting combinations of answers.

2.3 Philosophical orientation

Respondents were asked the following questions about their philosophical orientation:

Areas of specialization: Respondents had to choose from the following list of areas (the primary areas in the PhilPapers category system): 17th/18th Century Philosophy; 19th Century Philosophy; 20th Century Philosophy; Aesthetics; African/Africana Philosophy; Ancient Greek Philosophy; Applied Ethics; Asian Philosophy, Continental Philosophy; Decision Theory; Epistemology; European Philosophy; Feminist Philosophy; General Philosophy of Science; Logic and Philosophy of Logic; Medieval and Renaissance Philosophy; Meta-ethics; Metaphilosophy; Metaphysics; Normative Ethics; Philosophy of Action; Philosophy of Biology; Philosophy of Cognitive Science; Philosophy of Computing and Information; Philosophy of Gender, Race, and Sexuality; Philos ophy of Language; Philosophy of Law; Philosophy of Mathematics; Philosophy of Mind; Philosophy of Physical Science; Philosophy of Religion; Philosophy of Social Science; Philosophy of the Americas; and Social and Political Philosophy.

Philosophical tradition: As in 2009, respondents could choose either “analytic”, “continental”, or “other tradition”. When selecting “other tradition” they could enter a tradition as free text.

Identification with philosophers: Respondents were asked “For which nonliving philosophers X would you describe yourself or your work as X-ian, or the equivalent?” Respondents could choose from a list of well-known philosophers or select “other” to specify philosophers manually. The 2009 list was based on online surveys of the greatest philosophers of the last 200 years and of all time. It included: Anscombe, Aquinas, Aristotle, Augustine, Berkeley, Carnap, Davidson, Descartes, Frege, Hegel, Heidegger, Hobbes, Hume, Husserl, Kant, Kierkegaard, Leibniz, Lewis, Locke, Marx, Mill, Moore, Nietzsche, Plato, Quine, Rawls, Rousseau, Russell, Socrates, Spinoza, and Wittgenstein. For 2020 we added Dewey, Foucault, James, Merleau-Ponty, Peirce, Popper, Reid, Rorty, Sellars, and Whitehead (the ten most popular write-in choices in 2009); Parfit and Putnam (the leading candidates per previous criteria who died since the previous survey); and Arendt, Avicenna, Beauvoir, Buddha, Confucius, Deleuze, Derrida, Du Bois, Laozi, Nagarjuna, Rand, Sartre, and Wollstonecraft (to expand coverage of other traditions).

2.4 Background questions

Respondents were also asked the following background questions: year of birth, nationality, gender, doctorate in philosophy (specifying the granting institution and year), and current affiliation (including role).

2.5 Consent

Under consent guidelines approved by The Western University Non-Medical Research Ethics Board, respondents were told how their answers would be used, and at the end of the survey were asked to consent to the use of their answers. The Survey was anonymous by default, although respondents were given the option to make their answers public eventually. Respondents were also told that their answers would be retained for use in possible follow-up surveys, and that any question could be skipped if they were uncomfortable in answering.

3.1 Main questions

The results for the 40 main questions (those asked of all respondents) are shown in Table 1. For each question and each option, Table 1 presents the total number of respondents and the percentage who either “accept” or “lean toward” that option. This figure can be calculated either as an “inclusive” figure, where respondents who endorse multiple options are included in the totals for each options, or as an “exclusive” figure, which counts only respondents who endorse that option and no other option. We present inclusive figures in all cases. To simplify the table, we present exclusive figures only when at least one of them differs by 3% or more from inclusive figures. This gives some indication of questions for which choosing multiple options is popular. The last question shows respondents who reject options rather than exclusive numbers. The figures include all respondents who completed the survey except those who indicated “insufficiently familiar with the issue” or who skipped the question. The survey website offers a detailed breakdown of “other” answers.

The figures in Table 1 should not be longitudinally compared to the main results presented in the 2009 PhilPapers Survey paper, for three main reasons. First, the 2020 population is much broader (not restricted to 99 departments). Second, the 2020 survey made it easier to endorse multiple answers than the 2009 survey. Third, the main results presented in the 2009 paper included respondents who skipped the question or checked “insufficiently familiar”, whereas the results below exclude those respondents. For meaningful longitudinal comparisons, see Section 5, where we present 2020 results that are more directly comparable to the 2009 results (restricting to 2009-comparable departments, exclusive answers, and including skip/unfamiliar answers under “other”).

Among the 40 main questions, views mentioned most often as alternative answers (according to a semi-automated analysis) included: abstract objects: Aristotelian realism (24 respondents, or 1.5%); aesthetic value: intersubjective (25); knowledge: pragmatism (30), knowledge: Kantian (25), God: agnosticism (23); logic: pluralism (35); normative ethics: pluralism (31), normative ethics: particularism (23), perceptual experience: direct realism (23), perceptual experience: phenomenological (20), philosophical methods: phenomenology (30), truth: pragmatism (26). More information on combined and alternative answers can be found on the survey website on the pages presenting detailed results for each question.

Main questions for which combined answers were the most popular include: aim of philosophy (27%), gender (20%), normative ethics (15%), race (10%), knowledge (8%), political philosophy (8%), and vagueness (8%).

The “pluralism” answers in the cases of logic and normative ethics (as well as numerous cases discussed in the next section) bring out that pluralist views were often expressed as alternative answers (choosing “Alternative view” and then endorsing pluralism) rather than as combined answers (choosing “Evaluate multiple options”, and then endorsing multiple views). On the logic question, for example, 76 respondents endorsed a combined answer (accepting or leaning toward both classical and nonclassical logic) while 35 endorsed pluralism as an alternative answer. Insofar as pluralism can be considered a combined view, a consequence is that combined answer numbers alone may somewhat understate the popularity of combined views, and information on both alternative and combined answers is required for a full analysis.²

3.2 Additional questions

The results for the additional questions are found in Table 2. Of the 60 additional questions, one-sixth of the 1785 respondents, or about 300 respondents, were presented with the question as part of their mandatory 50 questions. Typically, another 800 respondents (45%) were presented with the question by answering additional questions, for a total of around 1100 respondents (62%) presented with the question. The figures below include these respondents, excluding those who chose to skip the question or who indicated “insufficiently familiar”. As before, the results are subject to selection bias, which is discussed in Section 8.

4. Demographics

4.2 Gender and age.

Just over 20% of respondents who indicated gender indicated “female”, while about .5% indicated “other”, with the rest indicating “male”. The number of female respondents who completed the survey is slightly lower than numbers in previously reported gender distributions (see Leslie et al. (2015), Schwitzgebel & Cushman (2012), Schwitzgebel & Cushman (2015)). As seen in Table 13 in Section 8, there appears to be a small gender effect in response bias.

The most common decade of birth was the 1970s (500), followed by the 1980s (424) and the 1960s (395). Figure 4 shows the breakdown of years of birth in slices of five years, with genders color-coded. The gender imbalance appears to be somewhat smaller among the youngest respondents.

Figure 4:

Figure 4: Breakdown of years of birth and gender.

5. Longitudinal analysis

This section compares the 2020 and 2009 results with respect to their 30 shared questions. The longitudinal results discussed in this section can be found in Table 18 in Appendix A.

For the purpose of longitudinal comparison, we restricted the 2020 results to the target group of 100 2009-comparable departments in Australasia, Europe, and North America. As explained in Section 2, the 2009-comparable departments were selected using ranking criteria analogous to those used in the 2009 survey.

It should be noted that the 2009-comparable department group differs systematically from the broader target population in a number of respects. Demographically, it includes a higher proportion of UK-based philosophers and analytic-tradition philosophers than the target population. Philosophically, it includes a lower proportion of theists, along with many other differences evident in comparing the 2020 results in Table 1 (all departments) to the results under “Comparable departments” in Table 18.

Figure 5:

Figure 5: Areas of specialization.

For longitudinal purposes, we use “exclusive” rather than “inclusive” answer figures: that is, we exclude respondents who endorse multiple options. Exclusive answers were used in presenting our 2009 results, so using them here maximizes continuity with existing results. Furthermore, inclusive answers were handled somewhat differently in 2009 and 2020, so using exclusive answers maximizes comparability of the results. Using them should make no difference to relative results on binary questions (though absolute results are lower in some cases, as multiple options could be chosen more easily in 2020 than in 2009). On non-binary questions, this method sets aside fine-grained information about respondents who endorse some but not all of the options; but since this fine-grained information was not available in the 2009 survey, it would be hard to use it for longitudinal purposes. We have also included “skip” and “insufficiently familiar” answers as “other” answers in this context, to maximize continuity with how results were presented in 2009.

Our main longitudinal measure is the swing toward or against a position on the survey, defined as its relative strength in 2020 minus its relative strength in 2009. The relative strength of a position on a survey is defined as the percentage of respondents who endorse it exclusively, minus the average percentage of respondents who endorse some option exclusively (averaged across all options). For example, in a binary question, if 50% of respondents endorse X exclusively, and 30% of respondents endorse Y exclusively, the relative strength of X is 10% (50% - 40%) and the relative strength of Y is −10% (30% - 40%). The swing is then the change in these relative strengths over time. There is no perfect measure of swing when more than two options are present, but our definition has the nice feature that all swings on a given question sum to zero. Furthermore, in cases where all positions increase or decrease by the same amount (5%, say), as discussed above, the swing toward each position will be zero. The biggest swings toward any position from 2009 to 2020 are shown in Table 8.

The biggest swings in the 2020 survey are away from invariantism and toward contextualism about knowledge claims, and toward non-classical logic, externalism about moral motivation, a priori knowledge, and Humeanism about laws of nature. Changes can be divided into swings toward a majority view (contextualism, a priori knowledge, free will compatibilism) and swings toward a minority view (moral externalism, Humeanism, subjective aesthetic value, trolley non-switching). For binary questions, these swings are accompanied by a corresponding swing away from the alternative minority views (a priori knowledge: no) or majority view (moral internalism, non-Humeanism, objective aesthetic value, trolley switching). Among non-binary questions, the largest swings away include swings away from a minority view (invariantism) and a majority view (correspondence theory of truth). In two cases the majority (or at least plurality) view changed: from an objective to a subjective view of aesthetic value, and from internalism to externalism about moral motivation.

Our data also allows longitudinal comparisons over the same people in 2009 and 2020: that is, over respondents in the 2009 target group who also responded in 2020 (regardless of whether they were in the target group in 2020). We have included these “same people” results with a corresponding swing, to shed light on the issue of how individual views may change over time. The biggest swings for this longitudinal comparison can be found in Table 9. Most swings are under 2%, which tends to confirm the oft-reported impression that philosophers do not commonly change their views.

In order to better understand relationships between institutional and personal changes in views, we computed swing numbers for departments excluding the individuals who answered both surveys (institutional swings) and plotted them against swing numbers for individuals who answered both surveys (personal swings). Figure 6 shows a scatter plot of these swings, with the correlations of positions with year of birth color-coded. As might be expected, we can see that institutional swings are sometimes importantly different from personal swings: for example, there is an institutional swing toward a priori knowledge but a personal swing against, while the reverse is true for metaphilosophical naturalism. In these cases and many other (but not all) cases, views with higher institutional swings relative to personal swings are also views that are more popular with younger voters. This suggests that, as one might expect, much but not all of the difference between institutional and personal swings is due to popularity among younger faculty members who did not take the 2009 survey.

Figure 6:

Figure 6: Institutional-ex-personal-swing vs personal swing.

We can also measure longitudinal changes in demographic and background questions in the 2009-comparable target group. It should be noted that these increases and decreases may reflect changes from 2009 to 2020 in respondent bias (see Section 8) rather than changes in the profession. The number of respondents specifying “female” rose from 17.4% to 22%. The most common decade of birth shifted from the 1960s to the 1970s. Respondents were more often Canadian-born, more often Australian-affiliated, and more often had a UK Ph.D. Areas of specialization with the largest relative increases included (in order) applied ethics, philosophy of cognitive science, social and political philosophy, normative ethics, and epistemology. Areas with the largest relative decreases included philosophy of physical science, ancient Greek philosophy, philosophy of mathematics, philosophy of religion, 19th century philosophy, and 17th/18th century philosophy. Identification with the continental tradition was down from 3.8% to 1.8%. Philosophers with the largest relative increases in identification include Descartes, Lewis, Leibniz, and Marx, while those with largest relative decreases include Davidson, Wittgenstein, Locke, and Hegel.

The survey website includes some further longitudinal comparisons, including comparisons across “same departments” (the 99 target departments from 2009, compared across 2009 and 2020). It also includes comparisons using “weighted scores” (0 for rejecting a view, .25 for leaning against it, .5 for other, .75 for leaning toward it, 1 for accepting it). These scores are more fine-grained than the percentages used elsewhere, which in effect assign 0 for the first three options and 1 for the last two.

6. Correlations

One of the aims of this survey was to get a clearer sense of relationships between philosophical views within the target population. The most highly significant correlations between survey answers can be found in Tables 19-25 in Appendix B. More correlations are available on the survey’s website.

These correlations were calculated as follows. We first assigned a numerical variable for each main answer option (e.g., mind: physicalism or mind: non-physicalism) to a philosophical question. For each respondent, this variable was assigned a value ranging from −2 to 2 depending on whether the respondent rejects, leans toward rejecting, is neutral on (including “agnostic” answers), leans toward accepting, or accepts the position corresponding to the option. When a respondent selects “accept: X” or “lean toward: X” in the main interface without evaluating multiple options, we consider that they also reject or lean against the other options respectively. When they evaluate multiple options, we use those evaluations separately. Respondents who gave answers falling under “other” but that are not considered neutral were left out of correlation calculations for all relevant pairs of variables.

For binary questions, we have included correlation results for only one main answer option (the first option listed in Table 1, e.g. mind: physicalism). The second option (e.g. mind: non-physicalism) is usually strongly anti-correlated with the first, so correlations here will have a similar magnitude in the opposite direction.

For most non-binary questions, we consider correlation results for all main answer options, but a few questions were given a special treatment. The “other minds” question, which asked respondents to say whether they accept that entities of various levels of complexity have consciousness (AI systems and living things from plants to humans), was split into two questions: consciousness in living things and consciousness in AI systems. We assigned respondents a single numerical answer to the first question based on the most complex living things to which they assigned consciousness. We assigned respondents a numerical answer to the second question based on the most advanced AI systems to which they assigned consciousness. Answers to the three-option questions “philosophical knowledge” and “philosophical progress” were similarly converted to a linear scale.

For demographic and orientation questions, each possible answer (e.g., “AOS: Feminist philosophy”) was assigned a variable whose value is 2 for a respondent who selected that answer, 0 for respondents who specified another answer, and N/A for respondents who skipped the question.

7. Dimensionality reduction and clustering

We performed component and cluster analyses to get a clearer sense of the extent to which answers to the main questions can be distilled to a smaller number of underlying views. These results should be taken with a grain of salt, as they are heavily dependent on which questions and topics are included in the survey.

Our principal component analysis used only one of the numerical variables described in the preceding section for each question, so a total of 101 variables (for non-binary questions, we selected the variable corresponding to the most popular option).³ As shown in Figure 7, we found that a few principal components can explain a modest amount of variance. Six components explain 2% or more of the variance each. The two first components explain considerable variance at 12.3% and 8.8% respectively. The variables most correlated with these two first components are represented in Figure 8. The first component appears to correlate with a combination of rationalism (or non-empiricism) and realism, especially in the moral domain. The second component is harder to label but might correlate with a science-friendly outlook.

Figure 7:

Figure 7: Variance explained by 10 first principal components.

Figure 8:

Figure 8: First and second principal components. The axes represent correlation coefficients between variables and components.

To try to shed more light on the nature of views that might unify answers to our questions, we performed linear regressions on the same variables (normalized to a mean of 0 and variance of 1). Table 10 shows the answers that explain 2% or more of the variance among other answers, with the percentage of variance (of all other 100 answers).

We also built multivariate models, employing an iterative procedure to determine which variables should be treated as independent and dependent. We first selected as independent the variable that explains the most variance among other variables (“God: theism”). We then iteratively selected the independent variables that can explain the most additional variance when added to previously selected independent variables.⁴ We found that 10 variables could explain 11.9% of variance. Table 11 shows the variables that were selected by our procedure and the cumulative variance explained.

Clustering methods can also be used to assess relatedness of answers. We used 1 − |ρ_A,B| (one minus the absolute value of the Pearson correla tion coefficient between A and B) as our distance metric. Figure 9 shows the result of clustering the 101 answer options described above using hierarchical clustering with the “average” method, which minimizes the mean distance between members of joined clusters. Other hierarchical clustering methods yielded less readable and balanced trees. Figure 9 allows us to see that while there are some strong correlations between answers and some form small clusters, the correlations do not amount to large clusters of closely related views. Note that because the distance metric used is based on the absolute values of correlation coefficients, anti-correlated positions are often clustered together.

Figure 9:

Figure 9: Hierarchical clustering (“average” method) of answer options using 1 − |ρ_A,B| as distance metric. The position of a junction corresponds to the mean distance between the components of the two branches it joins together. The dotted line marks a mean distance of .8, or a mean correlation between cluster members of .2. Clusters with a mean distance below this threshold are uniformly colored (colors are reused elsewhere within the graph).

8. Selection bias

Selection bias arises when the group who responds to a survey question is not a random sample of the target population. In our survey, selection bias takes two forms. First, the philosophers who respond to the survey at all (completing at least 50 questions and consenting) are not a random sample of the target population. We call this respondent bias. Second, respondents have the option to complete more than 50 (up to 100) questions, and the group who do so are not a random sample of the overall group of respondents. We call this enthusiast bias.

The results presented in Section 3 are subject to respondent bias and enthusiast bias, so they cannot be considered accurate guides to the distribution of views in our target population as a whole. To use survey responses to assess the distribution of views in our target population as a whole, we need to correct for respondent bias and enthusiast bias.

Enthusiast bias: Enthusiast bias does not affect the 40 main questions, which all respondents answered as part of their 50 mandatory questions. It affects only the 60 additional questions.

To correct for enthusiast bias on these questions, we can simply restrict our analysis to those First-50 respondents, who answered these questions as part of their 50 mandatory questions. This group should be a random sample of respondents as a whole. This information is shown in Table 14 under the “F50” column. Enthusiast bias is reflected in the difference between the “All” column and the “F50” column.

We can also assess enthusiast bias by calculating correlations between the number of questions answered by a respondent and their various answers to questions. The highest correlations are shown in Table 12.

Respondent bias: It is less straightforward to assess and correct for respondent bias, as we have less information on philosophers in the target population who did not participate in the survey. However, we were able to rely on the PhilPeople database, which was updated with extensive data entry for relevant departments prior to the launch of the survey. This database contains entries for almost all members of the target population, including their affiliations (hence their regions of affiliation). We estimated AOS using PhilPeople’s publication attributions (taking the AOS of an individual to be the PhilPapers cluster-level topic in which they have the most publications). We estimated gender using PhilPeople’s gender guessing algorithm, which, roughly, assigns gender based on first name statistics from US census records. There are many cases in which the algorithm cannot confidently assign a gender. We treat these cases as a third category. There are limitations on using census data for guessing gender, but these limitations should not affect our results so long as all gender identities are guessed equally well based on census data.

We used this information to assess and correct for respondent bias with respect to AOS, region of affiliation, and gender.⁵ The (modest) biases that we have identified are summarized in Table 13. Tables 14 and 15 summarize the additional and main results corrected for AOS and gender. Relatively few results are corrected by more than 3 percentage points, and only two are corrected by more than 5% compared to First-50 answers: propositional attitudes: representational (-6.15%) and true contradictions: impossible (-9.72%). We did not make regional corrections because regions of affiliation are not very strongly correlated with philosophical views (see Table 20) and the number of data points at our disposal did not allow a three-way stratification of respondents.

It remains possible and likely that there are respondent biases that go beyond AOS, gender, and region of affiliation, but our ability to measure them is limited by the limited information that we have about nonrespondents in the target population.

9. Specialist effects

As in 2009, we were interested to see whether and when specialists in the area of a question tend to give different answers than non-specialists. We associated every question with at least one (sometimes two) area of specialization, which we call the associated AOS. We use AOS as a useful if imperfect proxy for research specialization on a given question. For every question, we then compared specialist answers (percentages of non-“other” responses to each answer over respondents in the associated AOS) to non-specialist answers (percentages of non-“other” responses over respondents who are not in the associated AOS). Effects that are significant at the .05 level or better (based on a chi-squared test) are shown in Table 26 in Appendix C.

By far the biggest specialist effects are in the philosophy of religion, where 78% of specialists endorsed theism compared to 17% of nonspecialists, and 74% endorsed design as an account of fine-tuning compared to 13% of nonspecialists. The next biggest effects are on metaontology (metaphysicians favor heavyweight realism more than nonspecialists), gender and race (philosophers of gender, race, and sexuality favor social views), Newcomb’s problem (decision theorists favor two-boxing), and aesthetic value (aestheticians favor objective views).

Of course there can be many different sources of specialist effects. In some cases, prior philosophical views may influence one’s specialization. In other cases, specialization may influence philosophical views. In still other cases, there may be a complex interaction between specialization and views. If one looked to surveys like this as a guide to truth, there is perhaps a case for giving special weight to specialist opinion, at least in cases where one thinks that specialization influences view rather than vice versa. However, our data do not speak directly to the direction of influence, and do not tell us anything about whether specialist opinion correlates with philosophical truth.

10. Order effects

There is an order effect between two questions Q1 and Q2 if responses to Q1 when it is presented after Q2 differ significantly from responses to Q1 when it is presented before Q2. There exist a number of studies of order effects on philosophical questions, with many focusing especially on order effects between questions about the trolley case and questions about the footbridge case. Prior studies have demonstrated order effects on these questions among both professional philosophers (Schwitzgebel and Cushman 2015; 2012) and non-philosophers (Petrinovich and O’Neill 1996; Lanteri et al 2008).

It is natural to suppose that some of our questions might produce order effects. We assessed order effects for pairs of questions using exclusive normalized percentages that do not include any “other” answers and applying a chi-squared test to assess the significance of differences found. For the most part, in this project we did not engage in hypothesis testing, but given prior work on these issues, we formulated and tested a few hypotheses about order effects.

First, we formulated the primary hypothesis that there may be an order effect between the footbridge and the trolley questions. Consistently with prior work, we found a highly significant effect with Q1 = trolley and Q2 = footbridge (p = 2 x 10⁻⁸). We found only a weakly significant effect with Q1 = footbridge and Q2 = trolley (p = .03). These results are summarized in Table 16.

To investigate the footbridge/trolley effect further, we computed the order effect for each answer to each of these questions (four answers in total). We found that there is a statistically significant difference in the distribution of answers between the before and after conditions only among respondents who answer “don’t push” to the footbridge question (p = 6 x 10⁻⁸). Respondents who are presented the footbridge question first and answer “don’t push” are more likely to answer “don’t switch” on the trolley question than respondents who are presented with the footbridge second and answer “don’t push”. We don’t find such a clearly significant effect in the other cases. These results are summarized in Table 17.

Second, we formulated secondary hypotheses regarding possible order effects involving nine other pairs of questions on related topics (chosen somewhat arbitrarily from many pairs of related questions): footbridge vs. normative ethics, aim of philosophy vs. philosophical progress, personal identity vs. teletransporter, fine-tuning vs. theism, hard problem of consciousness vs. mind, material composition vs. metaontology, semantic content vs. knowledge claims, consciousness vs. other minds, Chinese room vs. other minds. These nine pairs of questions give rise to eighteen possible comparisons. After correcting for multiple comparisons, none of these comparisons was significant at p < .05.

Finally, we assessed order effects for every pair of questions (9900 comparisons), whether or not we had formulated hypotheses regarding those questions. By chance alone, we would expect one result to be significant at p < 10⁻⁴. In fact, two results were significant at that level: the pair with Q1 = trolley and Q2 = footbridge (p = 2 x 10⁻⁸ as above) and the pair with Q1 = moral judgment and Q2 = concepts (p = 4 x 10⁻⁶), with 80% and 61% of respondents respectively endorsing cognitivism about moral judgment when presented before and after the question about concepts (nativism or empiricism). The second result is significant (at p < .05) even when correcting for multiple comparisons. Unlike the first result, the second result does not correspond to a prior hypothesis (and there is no obvious relation between the questions). In the absence of further hypothesis-testing, it remains unsettled whether the second result is a genuine effect or a random fluctuation.

Overall, the results regarding the footbridge/trolley order effects are congruent with results from earlier studies. The result concerning concepts and cognitivism has not yet survived a similar process of robust testing. In any case, it is apparent that the footbridge/trolley order effect is unusually strong, and that order effects of this strength may be more the exception than the rule.

11. Conclusion

As we noted in the introduction, the 2020 PhilPapers Survey was intended to make at least three contributions over and above the 2009 PhilPapers Survey. It was intended to give information about a broader group of academic philosophers’ views about a broader range of philosophical questions, and it was intended to give some information about longitudinal changes in philosophers’ views over time. Our results suggest that it succeeded in these aims.

That said, the survey’s success in these aims is relative and far from complete. There are clear limitations on our survey population and on our survey questions, including (among other limitations) a strong analytic and English-language bias in both. There is considerable room for future work surveying a broader range of philosophers on a broader range of topics, giving more extensive information about philosophers’ philosophical views and how they change over time.