CONCEPTUALIZING AND MEASURING AGENCY IDEOLOGY

This article contributes to the growing body of literature that seeks to uncover the latent preferences of bureaucratic agencies within the U.S. executive branch. A central motivation behind these studies is the idea that measuring agency ideology is essential to understanding the scope of the principal-agent problem in bureaucratic politics. If agencies lack stable ideological proclivities of their own, their policymaking may simply mirror that of their political principals, thus dissolving the core tension at the heart of the principal-agent problem.

Previously, researchers have utilized a variety of methods to gauge agency ideology. For instance, survey responses from bureaucrats have been a popular source, as seen in studies by Clinton, Jackman, and Rivers (Reference Clinton, Bertelli, Grose, Lewis and Nixon2012) and Richardson, Clinton, and Lewis (Reference Richardson, Clinton and Lewis2018), who ask federal executives how they perceive the ideology of other federal agencies. Other researchers have aggregated individual-level data on bureaucrats to infer their agency’s ideological leanings, such as campaign contributions (Bonica, Chen, and Johnson Reference Bonica, Chen and Johnson2015; Chen and Johnson Reference Chen and Johnson2015), voter registration files (Spenkuch, Teso, and Xu Reference Spenkuch, Teso and Xu2023), and public statements during congressional hearings (Bertelli and Grose Reference Bertelli and Grose2011). Clinton and Lewis (Reference Clinton and Lewis2008) utilized expert surveys to deduce an agency’s ideological stance, offering a direct but subjective measure of agency ideology. Nixon (Reference Nixon2004) took a novel approach by analyzing cases where an agency head also served in Congress to determine the agency’s ideology. Acs (Reference Acs2020) used a method similar to the current study, albeit with ideal point estimates that are static and limited to two administrations, failing to capture variations across different administrative contexts.

Compared to most existing studies of agency ideology, the approach in this article has several notable distinctions. First, it concentrates on regulation, an important yet arguably understudied domain of federal policymaking. Second, the estimates are grounded in the actual policymaking activities of bureaucrats, not their survey responses, voter registration, campaign contributions, or other peripheral activities. This approach is more similar to the canonical legislative ideal point model, as first developed by Poole and Rosenthal (Reference Poole and Rosenthal2000), which infers the spatial or ideological positions of legislators from their voting records. As such, the method in this article not only provides a more direct measure of agency ideal points, based on policymaking decisions, but also allows for dynamic updating as new administrative data from OIRA become available for future administrations. Finally, this approach estimates ideal points for a comparatively large number of agencies over an extended period of time, reaching back to the Clinton administration.

Despite the growing literature on measuring agency ideal points, important questions remain about what these estimates actually capture. Unlike legislators, agencies are complex organizations whose policymaking is typically shaped by the interaction of career civil servants, political appointees, and the statutes and legal constraints that structure their authority. In any given case, which of these elements contributes most to the observed ideology of the agency? This is a difficult knot to untangle. Nonetheless, the simple premise of this article is that the OIRA review process—specifically, the decision to review a given regulation—reflects a composite of these influences, with their relative weights likely shifting over time and across administrations.

For example, the influence of political appointees may grow stronger in one administration relative to the influence of civil servants or the agency’s foundational statutes. If appointees were to exert near-complete control—such that the agency became a perfect conduit of presidential preferences and policymaking output was fully aligned with those preferences—then the OIRA review process would largely become redundant. In such a case, one would expect patterns of under-reviewing, which would shift the estimated ideal point of the agency closer to the president. In most instances, however, political appointees are unlikely to be perfect agents of the president, and even less likely to fully control career staff or fundamentally redirect the agency’s core policy agenda. The OIRA review process reflects these institutional barriers, as illustrated by the observed patterns of over- and under-review that are explored in this article.Footnote ²

DATA

The agency ideal points are estimated using two publicly available datasets. One is the Unified Agenda, a semi-annual catalog of pending and completed regulatory activities for federal agencies. Agencies are required by the executive order to post their regulatory plans to the Unified Agenda, as articulated in EO 12866: “Each agency shall prepare an agenda of all regulations under development or review, at a time and in a manner specified by the Administrator of OIRA.” OIRA reviews the Unified Agenda to determine which regulations to select for further scrutiny. To aid in this selection, the executive order requires agencies to include specific information for each regulation in the Unified Agenda: “at a minimum, a regulation identifier number, a brief summary of the action, the legal authority for the action, any legal deadline, and the name and telephone number of a knowledgeable agency official.”

The Unified Agenda aims to provide transparency in the regulatory process, allowing OIRA, political overseers in the White House and Congress, and other stakeholders to track forthcoming regulations. Prior to review, however, the details of each regulation remain unknown. It is only when OIRA selects a regulation for review that the agency must submit a draft regulation to OIRA, along with any requested analysis, such as a cost–benefit analysis or a statement on regulatory flexibility for small businesses.

Once the review is underway, OIRA can share these materials with interested parties within the president’s circle, typically beginning with a “heads-up” memo to alert White House stakeholders. This step is crucial because OIRA review is not merely an assessment by neutral bureaucrats; rather, it is a political review conducted within the president’s Office of Management and Budget (OMB), where OIRA resides, as part of the larger Executive Office of the President and the White House’s political staff.

The second dataset is the OIRA docket, which lists the regulations reviewed. For each regulation, the docket indicates when the review started, when it ended, whether it was for a proposed or final rule, and the outcome of the review (e.g., “accepted without change” or “returned”). As mentioned, Clinton’s EO 12866 dramatically altered the OIRA docket. Prior to EO 12866, OIRA reviewed virtually every regulation that agencies promulgated, or at least every regulation was entered into OIRA’s docket. Facing resource constraints, the OIRA offices of the Reagan and George H.W. Bush administrations focused their attention on politically significant regulations, often ignoring rules with little political consequence. To improve the efficiency of the OIRA review process, EO 12866 introduced the practice of selective review, reducing the docket to only those regulatory actions deemed significant by OIRA. As noted, however, few clear criteria exist to define what makes a regulatory action significant.

The final dataset was created by merging the Unified Agenda and the OIRA docket using the eight-digit identifier for each regulation, known as the regulation identifier number (the RIN).

With the merged dataset, a binary review variable was created for each RIN to indicate whether the associated rule was reviewed at least once—that is, whether it appears in the OIRA docket. The analysis focuses on rules with a planned Notice of Proposed Rulemaking (NPRM), the stage in the rulemaking process when agencies publish a proposed regulation in the Federal Register and solicit public comments. Whether an agency planned to publish an NPRM is determined by entries in the Unified Agenda.

When a rule is reviewed, the vast majority of first reviews (over 90%) occur before the NPRM is published. It is primarily during this first review that OIRA’s decision depends most heavily on the agency’s reputation, a point revisited in the next section because it is central to the theoretical claim that the review decision sheds light on the ideological distance between the president and the relevant agency. OIRA also reviews final rules, but in most cases, the rule has already been reviewed prior to the NPRM.Footnote ³ In addition, some agencies issue prerule documents that OIRA may review—such as an ANPRM, or Advance Notice of Proposed Rulemaking—before the NPRM, typically to solicit input on specific elements of the proposal. In a small number of cases, OIRA’s first review is of an ANPRM.

Several types of entries in the Unified Agenda were excluded from the dataset to prevent overcounting the pool of potentially reviewable rules. For each rule, the administration during which the relevant agency first planned to publish its NPRM was identified. Planned NPRM dates often appear in the Unified Agenda six months to a year before the NPRM is published, as EO 12866 requires an entry for “all regulations under development.” Planned NPRMs were excluded if they were withdrawn or abandoned by the agency prior to OIRA review or publication of the NPRM. Rules were also excluded if they were reviewed in one administration but initially planned by a prior administration. Such inherited rules are interesting, but they are likely to be viewed with heightened suspicion by the new administration and would, all else equal, be reviewed at a higher rate.

The dataset includes 69 administrative units—that is, regulatory agencies—that were active during the Clinton, George W. Bush, Obama, and Trump administrations. Each unit is identified by the numeric identifier found in the first four digits of each RIN. In some cases, the unit is an office within a larger agency, such as the EPA’s Office of Air or its Office of Water. In other cases, the agency is an office within a department, such as the Agricultural Marketing Service within the Department of Agriculture. Sometimes, the unit refers to an entire department, as with the Department of Veterans Affairs.Footnote ⁴ For better estimation of the measurement model, regulatory agencies that did not issue a regulation in each administration or issued fewer than 40 regulations during the entire period (averaging less than 10 per administration) were excluded. Supplementary Appendix A lists all 69 regulatory agencies used in the measurement model.Footnote ⁵

For each rule, a binary indicator variable was created to show whether the agency listed the rule as significant in the Unified Agenda. This indicator collapses the five-category “priority” variable for each rule into two categories. The original variable categories are (i) economically significant, (ii) significant, (iii) substantive but not significant, (iv) routine, and (v) informational. Following the EO 12866 requirement that OIRA should review “significant” regulatory actions, the first two categories were coded as significant (significant = 1), while the latter three categories were coded as nonsignificant (significant = 0). Since there is no objective difference between the second and third categories—a “significant” regulatory action versus a “substantive but not significant” action—OIRA still reviews “substantive” rules, particularly when it is suspicious of the agency.

Using the review indicator variable described above, a review rate can be calculated for each agency, yielding the share of an agency’s regulatory agenda that OIRA reviewed. Supplementary Appendix A displays the review rates for each agency and administration. The table also includes columns for the proportion of rules listed as significant and the total number of proposed rules listed in the Unified Agenda. If OIRA reviewed all the rules listed as significant, the review rate and the proportion of rules listed as significant, or significance rate, would be identical. When the review rate exceeds the significance rate, it suggests, all else equal, that the agency is being over-reviewed. Conversely, when the review rate falls below the significance rate, it suggests that the agency is being under-reviewed.

The Unified Agenda is not the only source of rulemaking activity, as all proposed and final rules are published in the Federal Register. However, for the purposes of the measurement model, the Unified Agenda is a better source of rulemaking activity, despite criticisms of its incompleteness.Footnote ⁶ For one, the Federal Register does not include planned NPRMs that were ultimately not published, so it lacks information about the many rules that OIRA reviews at the proposed rule stage and that are subsequently withdrawn or abandoned before publication as proposed rules. Moreover, even if the Unified Agenda does not provide a complete accounting of all rules, it arguably captures the rules that OIRA is aware of, as OIRA oversees the development of the agenda (Copeland Reference Copeland2015). In the data collected for this article, all rules reviewed by OIRA were also present in the Unified Agenda, a fact that was easily confirmed, as the rules are identified in both data sets by their RIN number. This correspondence between reviewed rules and the Unified Agenda is expected, as the Unified Agenda serves as a written record that OIRA uses to identify rules for review before their publication as NPRMs. Thus, if a regulation were not published in the Unified Agenda, it would potentially escape OIRA’s attention. Although such an evasive maneuver is intriguing and worthy of further study, it is not an issue for the measurement model; the focus here is on OIRA’s review decisions based on the planned regulations that OIRA is aware of, not on those that might be concealed.

THEORETICAL FRAMEWORK AND STATISTICAL MODEL

The measurement model is based on a simple theory of the OIRA review process, presented here informally to facilitate interpretation. The theory revolves around a two-player game involving the President and an Agency. Here, the Agency represents the regulator, including its appointees, and the President represents the centralized regulatory oversight process led by OIRA. The President is aware when the Agency is developing a regulation but knows only the Agency’s general reputation as a liberal or conservative regulator, represented by the Agency’s ideal point, rather than the specific content of the regulation. (As mentioned, OIRA may, in practice, have some limited knowledge of the regulation’s content from the summary statement included in each Unified Agenda entry.)

The President can choose to incur a cost to review the regulation and thereby learn its content. If the President reviews the regulation, there are two ways to model the President’s influence over it: (i) the President may revise the regulation to a preferred policy, presumably aligned with the President’s ideal point, or (ii) the President is limited to accepting or rejecting the regulation. Each assumption reflects a different level of presidential power. The second assumption suggests a more constrained presidency, acknowledging that presidents often lack the expertise to develop regulations independently. Consequently, the OIRA review process cannot simply transform a liberal regulation into a conservative one, or vice versa; instead, OIRA serves primarily as a tool to block unwanted regulations. The choice between these two assumptions makes little difference to the theoretical results, so the analysis that follows focuses on the simpler case of OIRA as a veto player.

The players’ strategies consist of the following choices. The Agency decides whether to regulate and, if so, whether to make the regulation acceptable to the President—ensuring that, if reviewed, the President would approve it—or to align the regulation with the Agency’s own preferred policy or ideal point. This decision is private information and not available to the President prior to review. The President, upon observing a planned regulation, decides whether to review it. If the President reviews the regulation, the President must then decide whether to accept or reject it. If the regulation is rejected, a status quo policy prevails, and the President imposes a penalty on the Agency. In practice, this penalty could take the form of personnel or budgetary changes or a reduction in discretion or trust in future regulatory actions. Regardless of its specific form, the penalty creates an incentive for the Agency to ensure the policy is acceptable from the outset.

When the players have different ideological preferences, they find themselves in a zero-sum game that mirrors the strategic dilemma in game theory known as “matching pennies” (Gibbons Reference Gibbons1997). If the Agency believes the President will review the regulation, it has an incentive to propose a regulation that aligns with the President’s preferred policy to avoid the penalty. Conversely, if the Agency believes the President will not review the regulation, it will propose its preferred policy based on its ideal point. The President’s decision to review depends on the Agency’s actions. For example, if the President believes the Agency’s proposal is acceptable, there is no incentive to incur the cost of reviewing it. However, if the President decides not to review, the Agency is incentivized to propose its preferred policy. This cycle, in turn, prompts the President to reconsider and review the regulation. This circular reasoning embodies the essence of the matching pennies game. There is no equilibrium in pure strategies; the cycling between strategies would continue ad infinitum, without reaching a stable conclusion.

To reach an equilibrium, the actors adopt a different approach and choose their strategies probabilistically, incorporating an element of randomness. The President optimizes by reviewing the Agency’s regulation with a probability that makes the Agency indifferent between proposing its own preferred policy or a policy acceptable to the President—that is, a policy the President views more favorably than the status quo. Similarly, the Agency proposes this policy, as opposed to its most preferred policy, with a probability that makes the President indifferent to reviewing the proposal. The greater the difference in policy preferences between the President and the Agency, the higher the President’s review rate must be to make the Agency indifferent. This is because an Agency with divergent preferences has a stronger incentive to propose its own policy rather than one acceptable to the President. Conversely, if the President and the Agency have similar preferences, a lower review probability will suffice to make the Agency indifferent about which policy to pursue.

This game-theoretic reasoning implies that the observed OIRA review rate of an agency reflects the divergence in preferences between OIRA—working at the behest of the president—and the agency. As this divergence grows more pronounced, the agency’s incentive to propose its preferred policies, rather than those of the president, increases. To counteract this, OIRA has an incentive to review the agency’s proposals at a higher rate. The implication is a positive correlation between review rates and ideological distance.Footnote ⁷ This model resembles “police patrol” oversight, as described by McCubbins and Schwartz (Reference McCubbins and Schwartz1984) in the congressional context and by Acs (Reference Acs2018) in the context of OIRA review. That is, given the president’s finite resources, OIRA reviews are targeted where they are most needed, specifically at agencies most inclined to avoid review, similar to how strategic police patrols focus on areas where the incentives for criminality are strongest.Footnote ⁸

Statistical Model

OIRA review rates can be analyzed statistically in a model that estimates unobserved agency ideal points. Based on the theoretical discussion above, the probability that OIRA reviews proposal $ i $ depends on the spatial distance between the president and the agency, $ {X}_p-{X}_a $ , where $ {X}_p $ is the ideal point of the president and $ {X}_a $ is the ideal point of the agency. The OIRA review decision also depends on several additional factors that can be modeled:

1. 1. A binary indicator, $ {S}_i\in \left\{0,1\right\} $ , for whether the agency listed proposal $ i $ as significant in the Unified Agenda;

2. 2. An intercept for each agency, $ {\alpha}_a, $ which captures unobserved, nonspatial factors that result in different average review rates for agency $ a $ . This parameter could be interpreted as a measure of agency “quality” or professionalization, unrelated to the dimension of ideological conflict; an agency trusted more by all presidents to develop sound regulatory plans would have a lower value of $ {\alpha}_a $ , all else equal, reflecting a lower average review rate for the agency.

Each factor is included additively in a model of OIRA’s binary review decision, $ {Y}_i $ ,

(1) $$ \mathit{\Pr}\left({Y}_i=1\right)=\Phi \left({\alpha}_a+{\beta}_{S_i}\right), $$

where the review probability is modeled as a Bernoulli distribution with the probability given by the cumulative distribution function, $ \Phi \left(\cdot \right) $ , of the standard normal distribution. The parameter $ {\alpha}_a $ captures unobserved, nonspatial factors that influence the OIRA review decision, and $ {\beta}_{S_i} $ represents the effect of rule significance and spatial distance as follows:

(2) $$ {\beta}_{S_i}={\gamma}_{S_i}+{\eta}_{S_i}{\left({X}_p-{X}_{a,t}\right)}^2. $$

Spatial distance is determined by $ {X}_p $ , the ideal point of each president, and $ {X}_{a,t} $ , the agency ideal point in administration $ t $ . The effect of rule significance is reflected in the parameters $ {\gamma}_{S_i} $ and $ {\eta}_{S_i} $ , each of length two, where $ {\gamma}_{S_i} $ is the independent effect of rule significance on the review decision and $ {\eta}_{S_i} $ is the effect when interacted with the distance measure. Since significance is a binary variable, the IRT model estimates $ {\gamma}_0 $ , $ {\gamma}_1 $ , $ {\eta}_0 $ , and $ {\eta}_1 $ . Significance is a strong predictor of OIRA review, so it is not surprising that estimates of $ {\gamma}_1 $ and $ {\eta}_1 $ , the effects of rule significance on the review decision, are both positive and substantially larger than $ {\gamma}_0 $ and $ {\eta}_0 $ , their respective counterparts for nonsignificant rules.

The last term in Equation 2, the interaction effect between $ \eta $ and the president-agency spatial distance, requires further elaboration. As mentioned, the model will estimate $ {\eta}_1>{\eta}_0 $ , with $ {\eta}_0 $ being positive but small (i.e., close to zero). This is intuitive, given that rule significance is a strong predictor of review and that the vast majority of nonsignificant rules are not reviewed. The estimates of $ \eta $ are, by definition, average effects and are not specific to any particular agency.

Now consider a hypothetical agency where all significant rules are reviewed, and a sizeable number of nonsignificant rules are also reviewed. This is an agency that is being over-reviewed. To account for this phenomenon, the model will adjust the estimate of $ {X}_a $ further away from $ {X}_p $ , ensuring that nonsignificance is more predictive of review in the formulation $ {\eta}_0{\left({X}_p-{X}_{a,t}\right)}^2 $ .

Conversely, consider the opposite case: a hypothetical agency where nonsignificant rules are never reviewed, and even some significant rules are not reviewed. This agency is being under-reviewed. To account for this, the model will adjust the estimate of $ {X}_a $ closer to $ {X}_p $ , ensuring that significance is less predictive of review in the formulation $ {\eta}_1{\left({X}_p-{X}_{a,t}\right)}^2 $ .

The model incorporates the dynamic nature of regulatory politics by allowing each agency’s ideal point, $ {X}_{a,t} $ , to evolve over time across presidential administrations. These ideal points follow a random walk process where the ideal point in administration $ t $ has the mean of the ideal point from the previous period and a variance of $ {\sigma}^2 $ . Formally, the random walk is specified as follows:

$$ {\displaystyle \begin{array}{l}{X}_{a,1}\sim \mathcal{N}\left(0,1\right),\\ {}{X}_{a,t}\sim \mathcal{N}\left({X}_{a,t-1},{\sigma}^2\right)\mathrm{for}t=2,\dots, T.\end{array}} $$

Allowing agency ideal points to vary across administrations reflects the inherently dynamic nature of the regulatory process, shaped by a shifting political landscape as administrations and appointees cycle in and out of office. The random walk can be viewed as a middle ground between estimating ideal points independently during discrete time periods, such as for pairs of administrations (e.g., Clinton–Bush or Obama–Trump), and using a static model that constrains agency positions to remain fixed across each administration.

To identify the model, the presidential ideal points need to be anchored. The presidential ideal points used in this analysis are drawn from NOMINATE, which estimates ideal points by inferring the president’s ideological position based on pseudo-votes or inferred positions related to congressional roll-call votes.Footnote ⁹ While this approach has limitations—particularly because it estimates presidential ideal points through pseudo-votes rather than actual voting behavior—it provides a consistent method of inferring the location of presidents over time. Notably, it also results in presidential ideal points that are close to the corresponding party medians in the House of Representatives during the relevant Congresses: Clinton: $ -0.438 $ (House Democrats: $ -0.38 $ ); Bush: $ 0.69 $ (House Republicans: $ 0.38 $ ); Obama: $ -0.36 $ (House Democrats: $ -0.39 $ ); Trump: $ 0.40 $ (House Republicans: $ 0.50 $ ).Footnote ¹⁰

In addition to the 69 agencies, a hypothetical benchmark agency was added to the sample. This benchmark exhibits no under- or over-reviewing. Artificial data were generated for the benchmark agency such that it issued 100 proposed rules in total, 25 in each administration—a productivity close to the median agency in the sample. Half of these “rules” were randomly selected to be significant, all of which were indicated as reviewed. As will become clear, this benchmark represents the prototypical moderate agency.

The model was estimated using Hamiltonian Monte Carlo methods in the Stan programming language (Betancourt and Girolami Reference Betancourt and Girolami2015; Stan Development Team 2024). Supplementary Appendix B includes the Stan code used for estimation and details on the distributional assumptions and priors of all model parameters. In total, when including the benchmark agency, the model estimated 355 parameters of interest, including 280 agency ideal points, 70 agency intercepts (each $ {\alpha}_a $ ), 2 independent significance effects ( $ {\gamma}_1 $ and $ {\gamma}_0 $ ), 2 additional interactive significance effects ( $ {\eta}_1 $ and $ {\eta}_0 $ ), and the standard deviation of the random walk ( $ \sigma $ ). The Hamiltonian Monte Carlo sampler was run for 7,000 iterations, of which the first 1,000 were used for warmup and adaptation.Footnote ¹¹ To analyze model convergence, four chains were used and $ \hat{R} $ statistics were calculated for each parameter. The $ \hat{R} $ statistic calculates the ratio of the variance between the multiple chains to the average variance within the chains. When the $ \hat{R} $ statistic equals one, the model has converged, although values as large as 1.1 still suggest convergence. The max $ \hat{R} $ value across all parameters after estimation was 1.00, suggesting that all parameters converged.

EMPIRICAL ESTIMATES OF AGENCY IDEOLOGY

The point estimates for each agency ideal point are presented in Figure 1, grouped by the Clinton, Bush, Obama, and Trump administrations. Each estimate represents the mean of the posterior distribution for that parameter. The uncertainty associated with each estimate is depicted by 90% Bayesian credible intervals, shown as horizontal line segments extending from each point estimate. These intervals correspond to the 5th and 95th percentiles of the posterior distribution.Footnote ¹² For reference, the figure also includes the benchmark agency, indicated by a horizontal dashed line, showing the position of a hypothetical agency that exhibited no over- or under-reviewing.

Figure 1. Dynamic Estimates of Agency Ideal Points with Credible Intervals

Note: Please see Table A1 in the Supplementary Appendix A for full names of all agencies mentioned in this figure.

In each plot, a vertical dashed line at zero marks the ideological center, where many “moderate” agencies are located, such as NASA, the National Park Service, the Patent and Trademark Office, and the Bureau of Economic Analysis. This raises an important question about how to interpret moderate agencies within this framework. Technically, these agencies are reviewed by OIRA at similar rates regardless of the president’s ideal point—possibly due to their relatively noncontroversial regulatory agendas or the limited partisan interest they attract from the White House. Notably, the benchmark agency is located almost exactly at zero, serving as a reference point for ideologically neutral or moderate agencies.

This empirical pattern differs from how moderation is understood in other ideal point models. In the legislative context, ideological moderation typically results from patterns of cross-partisan voting. Moderate lawmakers may include independently minded mavericks as well as conciliatory centrists inclined toward bipartisan compromise. Ideological moderation in agencies, by contrast, requires a different interpretation. Many of these agencies engage in activities that are rarely the subject of partisan debate—such as collecting data, making maps, or approving patents—and are thus reviewed at similar rates by OIRA, including under presidents of varying ideological commitments (Richardson Reference Richardson2024).Footnote ¹³

More controversial regulators are located at the ideological extremes of the distribution. Well-known examples include the EPA’s Office of Air, the central hub for federal climate change policy, and OSHA, an agency so embroiled in partisan politics that the director of legislation for the National Federation of Independent Business (NFIB) has described OSHA as “the worst four-letter swear word in small business’s vocabulary.”Footnote ¹⁴ Not surprisingly, the model estimates OSHA to be more liberal than the most liberal Senate Democrat (see Figure 2). Other agencies on the left side of the ideological spectrum are associated with key liberal interests, such as entitlement programs (e.g., Social Security Administration), racial and ethnic minority issues (e.g., Bureau of Indian Affairs), government healthcare (e.g., Centers for Medicare and Medicaid Services and the Department of Veterans Affairs), foreign aid (e.g., USAID), and immigration court proceedings (e.g., Executive Office for Immigration Review).

Figure 2. Agency Ideal Points Distributed within Departments

Note: Agency ideal points are derived from averages across administrations using the static model. Each department is ordered from top (liberal) to bottom (conservative) based on a weighted average of the agencies’ ideal points in each department, with weights determined by the number of proposals made by each agency. The congressional ideal points were estimated using an IRT voting model on the 112th Congress.

Compared to the liberal agencies, there are fewer conservative agencies—that is, those skewed to the far right of the distribution—and no agency is positioned as far to the right as some liberal agencies are to the left. These general trends reflect the liberal orientation of the federal bureaucracy as found in other studies, such as those reported by Richardson, Clinton, and Lewis (Reference Richardson, Clinton and Lewis2018). The most conservative agency, averaging across presidential administrations, is the Small Business Administration, which works, in part, to reduce regulatory barriers for small firms, followed by the Food and Drug Administration (FDA), a public health agency with close ties to the industry it regulates, according to surveys of FDA staffers.Footnote ¹⁵

The 276 separate agency ideal points can be overwhelming, so simplifying the data through higher levels of aggregation is beneficial for expository reasons. One approach is to average the estimates across administrations, reducing the 276 points to 69 agency ideal points. To compute an average ideal point, the model was modified to fix the agency ideal points during the study period, resulting in a single set of 69 static ideal points for each agency, hereafter referred to as static ideal points. Footnote ¹⁶ Further aggregation is achieved by averaging these static estimates within their parent departments. This level of aggregation is depicted in Figure 2, which plots the distribution of agency ideal points across 16 separate departments. This excludes independent agencies not part of any department structure, except for the EPA.Footnote ¹⁷ In the figure, departments are sorted from the most liberal at the top to the most conservative at the bottom based on a weighted average of their constituent agencies. Each weight is proportional to the number of regulatory proposals issued, so more active regulators contribute more to the department’s ranking. Two health-related departments, Veterans Affairs (VA) and Health and Human Services (HHS), top the list as the most liberal departments, while the Department of the Interior appears at the bottom as the most conservative. The figure also highlights ideological variation among the bureaus within several departments. For example, although the Interior Department is conservative overall, its Bureau of Indian Affairs is one of the most liberal agencies in the sample. In the figure, several departments are represented by a single vertical dash. For the Department of Energy, this is because it has only one constituent agency in the sample—the Office of Energy Efficiency and Renewable Energy, a left-of-center agency. For the VA and State Department, this is because they have no constituent agencies or offices in the Unified Agenda data. Finally, the figure compares the agency estimates to legislative ideal points from the 112th Senate, a time corresponding to roughly the midpoint of the study period. Interestingly, no agency is estimated as more conservative than the chamber’s most conservative member, but several agencies are more liberal than its most liberal member, reflecting the liberal orientation of the federal bureaucracy.Footnote ¹⁸

By aggregating the ideal points at the department level, they can be compared to the results from several other studies, which typically estimate static ideal points at this aggregate level. Interestingly, the department ranking depicted in Figure 2 largely aligns with the results from other studies, despite each employing different methodologies for estimating ideal points. Figure 3 illustrates the extent to which these studies concur on whether a department ranks in the top half or bottom half of the “most liberal” categorization.Footnote ¹⁹ The rankings show several notable similarities. All five studies place the EPA and the departments of Labor and Health in the top liberal half and the departments of Defense and Homeland Security in the bottom conservative half. Complete rankings from each study are presented in Supplementary Appendix D, which also includes a correlation plot showing a positive correlation among all the measures.

Figure 3. Department Ranking Consensus Across Studies (Ranking Liberal to Conservative)

Notes: Each study estimated an ideal point for each of the 16 departments. To construct the figure, the estimates were ranked from liberal to conservative, and the rankings were then compared to evaluate the degree of consensus about which departments ranked in the top, liberal half (top eight), and which ranked in the bottom, conservative half (bottom eight).

For a more granular comparison, without aggregating ideal points at the department level, the static estimates can be directly compared to those in Richardson, Clinton, and Lewis (Reference Richardson, Clinton and Lewis2018) (hereafter referred to as RCL), an impressive study that estimates agency-level ideal points at the subdepartment level. Of the 69 agencies included in the present study, 61 can be matched to the RCL estimates. Although the correlation between these sets is positive, the strength of the correlation is limited by four outlier agencies: Veterans Affairs (VA), Fish and Wildlife Service (FWS), the Defense Department’s Office of Health Affairs (DOD-OASHA), and the Drug Enforcement Administration (DEA). Excluding these outliers, the correlation is 0.42, whereas including them drops the correlation to 0.11.Footnote ²⁰

The presence of such outlier agencies underscores an important distinction about the methodology used in this study. Most other studies fundamentally rely on information about agency employees, as measured by survey responses, voter registration, or campaign contributions. While this approach is generally reasonable, discrepancies can arise between the ideology of agency employees and the actual policy goals of the agency. For instance, conservative employees may advance a liberal policy mission—a paradox evident in several of these outlier agencies. The VA, for example, is generally considered conservative or right-of-center based on its employees’ views.Footnote ²¹ Even the RCL estimates—which are based on survey respondents’ perceptions of an agency’s policy views—rank the VA as the 130th most conservative agency out of 165, potentially reflecting the fact that the agency has close ties to the military, an institution generally more conservative than the broader federal bureaucracy. However, the VA’s actual policy mission—providing government healthcare to veterans—arguably pushes its policies in a liberal direction. This may explain why, during Republican administrations, OIRA tends to review VA regulations more aggressively than during Democratic administrations. A similar case is found with the Defense Department’s Office of Health Affairs (DOD-OASHA). Although engaged in a liberal policy mission—government-run healthcare—the agency’s placement within the DOD may attract more conservative employees and create a perception of conservatism. RCL ranks the Defense Health Agency, which reports to OASHA, as the 134th most conservative agency out of 165, but the estimates here show OASHA to be consistently liberal, likely due to the nature of its actual policy mission, as perceived by OIRA.

Another distinction between this study and others is its reliance on rulemaking activity. While this approach has its advantages, it also introduces several caveats. For instance, although it is rare in practice, some agencies may choose to bypass the rulemaking process, particularly when advancing their more controversial policies, which could make their ideal points appear more moderate than expected. The Department of Education (ED) serves as a notable example. Historically, ED has favored aggressive federal interventions to advance liberal causes (Califano Reference Califano1981; Melnick Reference Melnick2018). Other studies consistently find ED’s employees to be liberal, ranking it as the third most liberal department in two studies and fifth in another (see Supplementary Appendix D). However, the model here estimates ED as relatively moderate (ninth out of 16 departments) likely due, in part, to its history of issuing informal guidance documents to schools and colleges instead of promulgating its controversial policies as formal rules—a tactic particularly evident in the Office of Civil Rights’ regulation of school sexual harassment policies, which were effectively binding in practice (Melnick Reference Melnick2018). More recently, increased political pressure has led ED to move its policy changes through the formal rulemaking process, which could impact ED’s estimated ideal point in future administrations, potentially shifting it to the left.Footnote ²² In addition to the deceptive practice of using guidance documents to avoid scrutiny, agencies may also attempt to advance regulatory policy changes by classifying rules as “direct final” or “interim final”—a strategy that may succeed if interest groups do not subsequently file comments that would trigger the normal notice-and-comment rulemaking process (Nou Reference Nou2012).

Temporal Dynamics of Agency Ideal Points

The study period witnessed significant shifts in the ideal points of many agencies. The first part of this section examines the nature of these shifts, predominantly in the conservative direction, and how they align with contemporaneous factors, such as agency politicization, evolving agency missions, and shifts in the partisan alignment around specific policies. The second part explores an important implication of these shifts, demonstrating that agencies become more productive regulators as they move toward a president ideologically and, conversely, their productivity declines as they shift away.

A prominent trend in the data is the shift toward conservatism. Notably, 40% of agencies were more conservative during the Trump administration than during any previous period, making it the most conservative administration in terms of the average ideal point across all agencies. The primary disruptor of this secular conservative trend is the influence of the president’s party on agency ideology; agencies under Republican administrations tend to be more conservative, on average, than those under Democratic administrations.Footnote ²³ This partisan pattern can largely be explained by politicization, where the president’s appointees incrementally shift agency policies closer to those of the president. The broader conservative trend in agency ideology over time remains an intriguing puzzle.

To explore what might explain this conservative trend, it is helpful to focus on the agencies that shifted the most. Among the top 10 agencies that experienced the largest shifts, seven moved in a conservative direction, as detailed in Table 1.Footnote ²⁴ To better understand these conservative shifts, Figure 4 plots the trajectory of each of the seven agencies over the study period, revealing two core patterns: a gradual shift in the conservative direction from the Clinton to the Trump administrations, evident among both liberal and conservative agencies (the first and third panels), and an intensified conservative shift during the Trump administration (the middle panel).

Table 1. Agencies with Largest Shifts in the Conservative Direction

Figure 4. Conservative Trends in Agency Ideal Points

What explains these patterns? In the cases of the Office of Personnel Management (OPM) and the EPA’s Office of Air, which both shifted significantly toward conservatism during the Trump years, politicization is a likely explanation. Under the Trump administration, EPA Administrator Scott Pruitt, known for his resistance to the agency’s established policy agenda, aggressively blocked new regulatory efforts and rolled back Obama-era climate policies. Pruitt’s reputation for conservative, pro-energy policies during his tenure as Oklahoma Attorney General, coupled with his outspoken skepticism of federal climate change policies, likely diminished OIRA’s scrutiny during the Trump administration, as he was viewed as a reliable ally. This decrease in OIRA scrutiny, as reflected in lower review rates, is precisely what drives the model to estimate the EPA as a more conservative agency, more aligned with the Trump administration’s positions. Interestingly, this pattern contrasts with the EPA during the Bush administration, which also experienced tension with the Office of Air. The estimates suggest that Bush relied more on OIRA reviews to manage the EPA, rather than direct politicization. While Bush appointees were conservative, including former Utah governor Mike Leavitt and Christine Todd Whitman, who identified herself as a Rockefellar Republican, arguably none matched the deregulatory zeal of Scott Pruitt.Footnote ²⁵ A similar phenomenon occurred at OPM, a traditionally liberal agency that was heavily politicized during the Trump administration, to the extent that several high-level appointees advocated for its dissolution. The shifting ideal point estimates indicate OPM moved distinctly in the conservative direction during the Trump years, suggesting that Trump’s OIRA had greater trust in the agency than previous Republican administrations.Footnote ²⁶

Another pattern observed among agencies with significant ideological shifts is a gradual turn toward conservatism. This trend occurred both in agencies initially liberal during the Clinton administration, such as the VA and DOD’s Health Affairs Office, and in moderate agencies, including the Treasury’s Financial Crimes Enforcement Network (FINCEN), and DOD’s Defense Acquisition Regulation Council (DARC). One explanation for these shifts is the changing politics surrounding key policy debates. For instance, FINCEN evolved from a moderate position during the Clinton era to a more conservative agency under subsequent presidents, all of whom served after the attacks of 9/11. During this period, FINCEN’s authority expanded under the Patriot Act of 2001 to investigate financial transactions for terrorism links, a development that alarmed civil liberties groups within the Democratic coalition, such as the ACLU. Procurement regulation also became more politically charged during this period, especially following the wars in Iraq and Afghanistan, which may account for DARC shifting in a conservative direction—that is, generating more scrutiny from OIRA under Democratic administrations than previously. A similar pattern occurred for both the VA and DOD’s Health Affairs Office, both of which were initially quite liberal under Clinton but faced more scrutiny under Obama’s OIRA, shifting their estimated ideal points to the right. The VA’s conservative shift during the Trump administration also coincides with the landmark VA MISSION Act of 2018, a law supported by Republicans and segments of the medical establishment—including the American Medical Association—for expanding veterans’ access to private-sector health care.

However, several agencies bucked the trend of becoming more conservative. Notably, the Minerals Management Service (MMS), discussed earlier, was at its most conservative during the Clinton administration and became more liberal during the Obama and Trump administrations. This reverse trend likely reflects major regulatory reforms in offshore energy oversight following the Deepwater Horizon oil spill. In response to the disaster, the Obama administration implemented a suite of changes aimed at reducing industry influence and increasing environmental safeguards, culminating in the creation of the Bureau of Ocean Energy Management (BOEM).Footnote ²⁷ Before and during the fallout from the spill, MMS was frequently accused of being heavily influenced by the conservative-leaning oil and gas industry. This helps explain the agency’s conservative orientation during the Clinton era and the subsequent push by the Obama administration in 2011 to implement stronger oversight and insulate the agency from industry capture.

One way to validate these over-time trends in agency ideology is to examine whether public comments submitted by stakeholders align with the estimated directional shifts. Stakeholders are often explicit in their feedback on proposed rules, clearly indicating their support or opposition. As such, the content and tone of public comments offer a useful source of external validation. For instance, one would expect reliably liberal stakeholders, such as public sector unions like the American Federation of Government Employees (AFGE), to express greater support for OPM rulemakings during the Obama administration than during the Trump administration, when the agency shifted in a more conservative direction, as shown in Figure 4. This approach serves as a form of convergent validity assessment—validating the estimates against independent, behaviorally observable indicators—and complements the earlier validation strategy based on comparison to existing ideal point estimates (McMann et al. Reference McMann, Pemstein, Seim, Teorell and Lindberg2022).

Three of the agencies mentioned above were selected for closer examination based on the presence of active stakeholder groups with reliably identifiable ideological orientations.Footnote ²⁸ For the EPA Office of Air, comments were analyzed from the U.S. Chamber of Commerce, a consistently pro-business organization that generally favors deregulatory efforts. The Chamber reliably opposed Office of Air regulations during the Obama administration and supported them during the Trump administration. For the Office of Personnel Management (OPM), comments from the American Federation of Government Employees (AFGE) and the Alliance for Federal Employees (AFFE) were examined, both of which advocate for strong protections for federal workers. These unions supported Obama-era rules but expressed consistent opposition to OPM’s rules during the Trump administration, aligning with the rightward shift in the agency’s ideal point. For the Department of Veterans Affairs (VA), comments from the American Medical Association (AMA) were reviewed, a prominent physician organization that has historically supported efforts to privatize core VA services. The AMA was more engaged with, and more supportive of, Trump rulemakings than those under Obama. Across all three agencies, the directional shift in stakeholder support is broadly consistent with the conservative movement estimated in the ideal points, offering external validation of the measurement model.

Regulatory Productivity

Shifts in agency ideology have important policymaking implications: regulatory productivity varies with how closely an agency’s position aligns with the president’s. For example, during the Trump administration, the Office of Personnel Management gained more political leeway to advance its agenda as its ideological position moved closer to the president’s—greater leeway than if it had remained more liberal. More generally, when an agency aligns with the president, regulatory activity tends to increase; when their preferences diverge, productivity diminishes. These findings validate the estimates by showing that the ideal points are correlated with substantively meaningful variation in regulatory activity. The next section considers how these ideal point estimates can be applied more broadly in empirical research.

To empirically test the hypothesis that ideological distance is inversely related to regulatory productivity, the variable President-Agency Distance was constructed. This variable measures the absolute value difference between the estimated ideal points of an agency in administration $ t $ and the president’s ideal point during the same period. It is then employed as a predictor in the following negative binomial model with mean $ \mu $ and overdispersion parameter $ \theta $ :

(3) $$ {\displaystyle \begin{array}{c}{\mathrm{Regulatory}\ \mathrm{Activity}}_{at}\sim \mathrm{NegBin}\left({\mu}_{at},\theta \right)\\ {}\ln \left({\mu}_{at}\right)={\pi}_a+\beta \cdot {\mathrm{PresidentAgencyDistance}}_{at}\end{array}} $$

where Regulatory Activity quantifies the regulatory output promulgated by agency $ a $ in administration $ t $ and $ {\pi}_a $ represents a fixed effect for each agency. Including a fixed effect for each agency allows the model to isolate changes in an agency’s position relative to the president, rather than comparing across different agencies. The data encompass 276 unique president-agency distance measures, derived from 69 agency estimates across 4 administrations ( $ 69\times 4=276 $ ).

There are several ways to measure regulatory activity, the dependent variable, using information about the page length of final rules published in the Federal Register. Footnote ²⁹ Generally, a greater number of pages indicates greater regulatory activity. Although this measure is not without flaws, it arguably provides a more precise estimate of regulatory activity than simply counting the number of final rules, as individual rules can vary significantly in their scope and societal impact. In this analysis, Regulatory Activity for each administration is quantified as the average page length per year of all final rules published during the years of that administration.Footnote ³⁰

It is important to acknowledge that the empirical relationship examined here does not necessarily imply causality. The aim, rather, is to highlight patterns between the distance measure and regulatory activity, in part as a validation exercise to show that the ideal points are correlated in intuitive ways with substantively important patterns of policymaking. Several knotty econometric issues nonetheless persist, and the analysis here cannot thoroughly resolve them. For one, the distance measure is predominately based on OIRA’s decision to review proposed rules. When the distance is large, it generally indicates that the agency has been over-reviewed by OIRA, all else equal. The dependent variable, on the other hand, is the volume of regulatory activity, measured by the page length of final rules, a fraction of which are reviewed by OIRA. Although the final rule length occurs downstream of the review decision, both variables are derived from the same regulatory process. This overlap could be problematic if OIRA’s review of proposed rules systematically reduces the scope of the rule. For example, if OIRA—staffed largely by economists—has a bias toward reducing economic costs, then a greater distance measure could correlate with less regulatory activity, as measured by shorter or less stringent regulations. What appears to be an effect of ideological distance limiting regulatory activity could instead reflect the consequence of the OIRA review process itself. One alternative is to measure the dependent variable in a way that excludes rules reviewed by OIRA. Although not perfect, this alternative approach helps to ensure that the distance measure is not simply capturing the direct effects of OIRA review. Both options are considered in the following analysis.

Five different regression models were estimated, each measuring the dependent variable in a distinct manner. The first three measures are more inclusive. Model 1 (“Formal”) includes final rules provided that they went through the notice-and-comment process and were not temporary. Model 2 (“Formal and Significant”) builds on Model 1 by further requiring the rule to be listed as significant. Model 3 (“Formal, Significant, and Regulatory”) combines the criteria of Models 1 and 2, excluding deregulatory rules.Footnote ³¹ The final two models exclude rules reviewed by OIRA during the proposed rule stage. Model 4 (“Formal”) includes rules that underwent the notice-and-comment process and were not temporary. Model 5 (“Formal and Regulatory”) further excludes rules identified as deregulatory. (Among these last two models, it does not make sense to additionally exclude nonsignificant rules, since these constitute the majority of the sample.)Footnote ³²

The results from estimating Equation 3 are presented in Table 2. Across all models, a negative association is observed between President-Agency Distance and Regulatory Activity. Specifically, a one unit increase in the distance between a president and an agency is associated with roughly a 5%–20% decrease in the agency’s regulatory output, with variations depending on the model.

Table 2. Ideological Distance and Regulatory Productivity (Negative Binomial)

Note: *p $ < $ 0.05. Models 4 and 5 exclude rules reviewed by OIRA at the proposed rule stage. Models 3 and 5 also exclude deregulatory rules. Weights equal to the inverse variance of each agency ideal point are included.

What it means, substantively speaking, to shift the distance measure warrants some explanation. A unit shift in the distance measure corresponds almost exactly to the average change observed across the four administrations—that is, the typical shift expected to be seen in an agency. This average is calculated by taking the mean of the absolute values of the maximum and minimum distance measures for each agency.