Opinion for the Court filed by Circuit Judge STEPHEN F. WILLIAMS.
STEPHEN F. WILLIAMS, Circuit Judge:Under authority provided by § 6(b) of the Occupational Safety and Health Act of 1970, 29 U.S.C. § 655(b), the Occupational Safety and Health Administration (“OSHA”) in December 1987 adopted new standards for formaldehyde in the workplace. The standards reduce the permissible exposure limit (“PEL”) to one part per million (“ppm”) as an 8-hour time-weighted average and the short term exposure limit (“STEL”) to 2 ppm. 52 Fed.Reg. 46,168-312, codified at 29 C.F.R. § 1910.1048 (1988). They also provide for a variety of ancillary measures, including employee exposure monitoring, medical surveillance, hazard communication and recordkeeping. 29 C.F.R. § 1910.1048(d)-(o).
Several unions representing workers in the affected industries claim the standards are not stringent enough to satisfy § 6(b)(5)’s requirements for the regulation of toxic materials. First, they assert that OSHA erred in finding that formaldehyde would present no significant risk at the 1 ppm level. If they are right, OSHA would be required to tighten the standard, to the point where there remains no significant risk or where further tightening is infeasi*289ble. (They propose 0.5 ppm, or possibly lower.) Second, the unions contend that OSHA was arbitrary and capricious in failing to take stricter measures on medical surveillance, examinations, and medical protection.
Industry petitioners initially attacked the standards governing labeling requirements for formaldehyde, but they now request deferral of their challenge, pending the outcome of OSHA’s ongoing reappraisal of these requirements. See Industry Petitioner’s Reply Brief at 37-39.
We hold that OSHA has insufficiently explained its finding of insignificant risk at an exposure of 1 ppm, and its attendant refusal to impose a lower formaldehyde standard. We also find inadequate OSHA’s explanation of its failure to include a “medical removal protection” requirement. We remand to the agency for reconsideration of these issues, but reject the other union claims.
Background
Formaldehyde is a colorless, pungent gas which, when mixed with stabilizers (primarily water and alcohol) is a major industrial chemical. See 52 Fed.Reg. at 46,171-73. The chemical has four basic commercial uses: as an “intermediate” in both resin and industrial chemical production, as a bactericide or fungicide, and as a component of consumer products. Id. at 46,172. Approximately 65% of formaldehyde sold is used in the production of resins, which are used as wood binders in the construction and furniture industries. OSHA Brief at 6-7. Resins are also used as finishers on 60-85% of all fabric manufactured domestically. Id. Formaldehyde also appears, often in very small quantities, in such diverse products as plastic moldings, housewares, auto parts, sporting goods, disinfectants and cosmetics. 52 Fed.Reg. at 46,-172.
OSHA first regulated workers’ exposure to formaldehyde in 1972, when it imposed a PEL of 3 ppm on the basis of known risk of eye, skin and respiratory irritations. 29 C.F.R. § 1910.1000, Table Z-2. Later research, particularly a rat study by the Chemical Industry Institute of Toxicology (“CUT”) in 1979, raised the specter that formaldehyde might also be a human carcinogen. See 52 Fed.Reg. at 46,169-70. The UAW and 13 other unions petitioned OSHA in 1981 to issue an Emergency Temporary Standard. See Joint Appendix (“J.A.”) at 422. After OSHA resisted, the UAW sued in district court to compel OSHA either to issue an emergency standard or to start rulemaking proceedings to revise the permanent standard. The district court found OSHA’s action arbitrary and ordered the agency to reconsider its decision, and this court, finding that original jurisdiction lay here, adopted a variation of the district court’s order as its own. See International Union, UAW v. Donovan, 590 F.Supp. 747 (D.D.C.1984); 756 F.2d 162 (D.C.Cir.1985). OSHA published an Advance Notice of Proposed Rulemaking and then a Notice of Proposed Rule-making, and, after accumulating the usual massive record (more than 1400 exhibits and 30,000 pages of transcript), issued the standard before us on December 4, 1987.
As noted above, the new standard establishes a 1 ppm PEL and a 2 ppm STEL. Employers who are unable to achieve these levels for the ambient air are required to provide respirators to their employees. 29 C.F.R. § 1910.1048(g). They are also required to supply personal protective equipment such as gloves and suits to protect workers from exposure of their skin to formaldehyde. 29 C.F.R. § 1910.1048(h)-(i). The regulations set an “action level” of 0.5 ppm, which triggers requirements for employers to initiate periodic exposure monitoring, annual medical surveillance and annual training. 29 C.F. R. § 1910.1048(b), (d)(3H4), (k)(1)(i), (n)(2)(ii). All employers must provide medical surveillance for employees manifesting formaldehyde-related symptoms and maintain certain records, and all are subject to housekeeping requirements.
In regulating toxic materials, the Secretary is instructed by § 6(b)(5) of the Act to
set the standard which most adequately assures, to the extent feasible, on the basis of the best available evidence, that *290no employee will suffer material impairment of health or functional capacity even if such employee has regular exposure to the hazard dealt with by such standard for the period of his working life.
29 U.S.C. § 655(b)(5). We review under the substantial evidence standard, 29 U.S.C. § 655(f), which in this setting, as courts have many times noted, does not require the agency to “support its finding ... with anything approaching scientific certainty ... so long as they are supported by a body of reputable scientific thought.” Industrial Union Department, AFL-CIO v. American Petroleum Institute, 448 U.S. 607, 656, 100 S.Ct. 2844, 2871, 65 L.Ed.2d 1010 (1980) (“Benzene ”). To this end, OSHA is expected to
identify relevant factual evidence, to explain the logic and the policies underlying any legislative choice, to state candidly any assumptions on which it relies, and to present its reasons for rejecting significant contrary evidence and argument.
Building and Construction Trades Department, AFL-CIO v. Brock, 838 F.2d 1258, 1264 (D.C.Cir.1988) (quoting United Steelworkers of America v. Marshall, 647 F.2d 1189, 1207 (D.C.Cir.1980)).
We first will address OSHA’s rulings on formaldehyde’s carcinogenic and non-carcinogenic effects at the 1 ppm PEL. Then we will turn to the asserted inadequacy of the ancillary restraints.
I. Carcinogenic Effects
The Supreme Court has construed § 6(b)(5) to authorize regulation only to diminish a “significant risk of harm.” See Benzene, 448 U.S. at 652, 100 S.Ct. at 2869. While observing in Benzene that determination of a level of significant risk must be “based largely on policy considerations,” id. at 655-56 n. 62, 100 S.Ct. at 2870-71 n. 62,1 the Court also gave an example, saying that “if the odds are one in a thousand that regular inhalation of gasoline vapors that are 2% benzene will be fatal, a reasonable person might well consider the risk significant and take appropriate steps to decrease or eliminate it,” id. at 655, 100 S.Ct. at 2870. In this proceeding and in earlier ones, OSHA appears to have incorporated that observation as a policy norm, at least in the sense of believing that it must regulate if it finds a risk at the 1/1000 level (either in a workplace free of regulation or a workplace subject to an actual or hypothesized standard such as the 1 ppm under consideration here). Counsel’s acknowledgement of this standard at oral argument is confirmed by its practice. See 52 Fed.Reg. at 46,233/3 (“OSHA concluded in post-1980 rulemakings that it was the intent of Congress to reduce risks of average magnitude and the Supreme Court indicated that a reasonable person ‘might well consider a risk of 1/1000 significant.’ ...”); see also Public Citizen Health Research Group v.. Tyson, 796 F.2d 1479, 1502-03 (D.C.Cir.1986) (OSHA found that 12-23 excess deaths per 10,000 workers was a significant risk); Building and Construction Trades Department, AFL-CIO v. Brock, 838 F.2d 1258, 1265 (D.C.Cir.1988) (“the Secretary found a risk of 1.64 excess mesothelioma deaths per 1000____ This he expressly found significant.”); but compare 52 Fed.Reg. at 46,234/1 (stating that OSHA “believes that figures of 0.6/100,000 (i.e., 6 in a million) ... may be approaching a level that can be viewed as safe”).
No one could reasonably expect OSHA to adopt some precise estimate of fatalities likely from a given exposure level, and indeed the Supreme Court has said that the agency has “no duty to calculate the exact probability of harm.” Benzene, 448 U.S. at 655, 100 S.Ct. at 2871. Nonetheless, OSHA necessarily seeks to quantify the risk posed by each toxic threat. While our deference to the agency is at a peak for its choices among scientific predictions, we must still look for some articulation of reasons for those choices.
Here the estimates are all over the lot. One, derived from an epidemiological study *291of workers in industries using formaldehyde, predicted 864 cases of fatal lung cancer at 1 ppm per 100,000 workers exposed over a 45-year period, or 8.64 per 1000. See 52 Fed.Reg. at 46,221/2. The agency placed its primary reliance, however, on the 1979 CUT rat study mentioned above. For reasons that will become clear, no one seriously disputes its authority to make that choice. The estimates derived from the rat study itself, however, cover an extraordinary range, depending upon various assumptions. They stretch from .0056 per 1000 at the low end to 2.64 per 1000 at the high. See id. at 46,223/2. While OSHA did not explicitly select among these, see id., its conclusion that the risk was not significant necessarily means, given its 1/1000 standard, that it gave considerably more weight to the low end than the high. (An unweighted average of these two estimates from the rat study would be very slightly over 1.32 per 1000.)
The epidemiological study underlying the estimate of 8.64 per 1000 was conducted by Aaron Blair and others, “Mortality Among Industrial Workers Exposed to Formaldehyde,” 76 Journal of the National Cancer Institute 1071 (June 1986), discussed at 52 Fed.Reg. at 46,185-91 and throughout the rest of the agency’s explanation of the final rule. Ironically the authors of the study, working under the auspices of the National Cancer Institute in conjunction with the Formaldehyde Institute, interpreted its data to exonerate formaldehyde. The study evaluated 26,561 individuals employed in 10 plants, with a minimum followup period of 14 years since their first exposure. 52 Fed.Reg. at 46,186/1-2. The results showed a statistically significant increase in lung cancer risk for employees with a greater than 20-year latency. Id. at 46,187/2. But because Blair and his colleagues did not find a correlation between cumulative exposure and lung cancer risk, they concluded that
[tjhese data provide little evidence that mortality from cancer is associated with formaldehyde exposure at levels experienced by workers in this study.
Id. (quoting Blair report at 1071).
Other scientists subjected the Blair study to a vigorous critique before OSHA. See 52 Fed.Reg. at 46,187-92. Some disputed the proposition that a positive cumulative dose-response relationship was essential to establish carcinogenicity, while others argued that several factors — including unduly subjective exposure estimates and a faulty aggregation of lightly exposed workers with others — might have obscured such a relationship. See id. at 46,187-89. OSHA seems to have conclusively rejected the authors’ reading of their data, 52 Fed. Reg. at 46,187-88, and industry petitioners have not challenged OSHA’s finding of carcinogenicity.
Two scientists not only attacked the Blair study but reworked its data. See Theodor D. Sterling and James J. Weinkam, “Reanalysis of a National Cancer Institute Study on ‘Mortality Among Industrial Workers Exposed to Formaldehyde,’ ” J.A. 1943-64 (November 1986) (unpublished). Among the objections they made were to its overlooking the “Healthy Worker Effect” (the fact that fully employed individuals are generally more healthy than the population at large, so that for accurate identification of a workplace hazard the subjects’ mortality rate should be compared with other workers’); its including as exposed those who worked only a short time in any of the studied plants; and its inadequately separating hourly and salaried employees and thereby failing to account for the differing mortality rates prevailing among those categories in the population and for their differences in probable exposure. See id. at 1946-51. On the basis of their recalculation, Sterling and Weinkam reported a 36% excess in lung cancer risk in employees exposed to formaldehyde concentrations above 0.5 ppm (averaging 1.15 ppm), as opposed to no increase for exposures below 0.5 ppm. Id. at 1956. A simple extrapolation from this figure was the source of the 8.64 per 1000 estimate of the risk of lifetime exposure. See 52 Fed.Reg. at 46,221/2.
OSHA took up a somewhat equivocal position between the original study and its overhaul, saying that it “believes that both the Blair et al. analysis and the Sterling *292and Weinkam analysis are reasonable, and both analyses have been useful to OSHA’s interpretation of the data.” 52 Fed.Reg. at 46,190/1. At least in its finding of carcinogenicity, however, it implicitly gave more credence to the Sterling and Weinkam view. This is not inconsistent, however, with some discounting of its results; after all, it does not appear that Sterling and Weinkam even claim to have fully solved the problem of subjective exposure estimates.
The story of the Blair study and its critics suggests some of the difficulties of epidemiological cancer studies — here, primarily those of getting exact exposure data and holding constant other possible causes of cancer. Accordingly, despite a general preference for epidemiological data, see OSHA Brief at 31, OSHA here turned to and relied somewhat more on animal studies. These too have their limitations. Humans are not rats, and it is far from clear how readily one may generalize from one mammalian species to another. But in light of the epidemiological evidence here, that was not the main problem. Rather it was the absence of data at low levels. About 200 rats were subjected to each of four levels of exposure. The table below shows the number of rats with squamous cell carcinomas:
Dose (in ppm) Rats with Carcinomas
14.3 103
5.6 2
2 0
0 0
J.A. at 546; see also 52 Fed.Reg. at 46,205/2-3.
Obviously a risk of 1/1000 or more could be consistent with these figures, for only 200 rats were exposed at 2 ppm. (Under a 1 ppm ceiling nearly 200,000 humans will be exposed to formaldehyde at levels between 0.5 and 1 ppm. See 52 Fed.Reg. at 46,239.) The question before OSHA was what assumptions it should make in extrapolating to the low-dose range from the positive data at 14.3 and 5.6 ppm.
OSHA selected a so-called “multistage” mathematical model for purposes of this extrapolation, see id. at 46,216/3; this selection is not disputed. It is a “no threshold” model, i.e., it assumes that there is no (non-zero) exposure below which formaldehyde will have no carcinogenic effect at all. The model generates several sets of figures, however, and the parties are in sharp dispute over the choice between the “maximum likelihood estimate” (“MLE”) and the “upper confidence limit” (“UCL”). As we understand it, the MLE is, as its name suggests, an “average” calculation of risk. The UCL is a risk-averse calculation, i.e., a figure that (assuming the validity of the implicit assumptions) has only a 5% probability of understating the risk. The MLE figures, though they vary among themselves, were all quite low, ranging from .006/1000 to 0.23/1000; the UCL figures were far higher, ranging from 1.3/1000 to 2.64/1000.2 Although OSHA found no “compelling argument in support of either” the MLE or the UCL, see id. at 46,223/1, its action reflects an implicit decision to give more weight to the MLE: a simple averaging of the four estimates just quoted would yield a figure slightly over the 1/1000 lifetime risk that OSHA regards as significant.
Were the UCL’s only difference the built-in risk aversion, we should have no trouble affirming OSHA’s choice to place greater reliance on the MLE. Under the system established by § 6(b)(5), as construed in Benzene, OSHA can consider risk aversion in the policy judgment that comes in characterizing a given risk estimate as “significant.” See Benzene, 448 U.S. at 655-56 n. 62, 100 S.Ct. at 2870-71 n. 62. The purpose of the risk estimation itself is to secure facts — or at least the predictions that must pass for facts in these circumstances. The union points to nothing in the statute that prevents the agency from adopting the most probable predictions, reserving for the characterization phase any policy judg*293ment as to the proper degree of risk aversion.3
But there is another distinction between the MLE and the UCL in the multistage model. At low doses the UCL is based on a linear measure of risk, while the MLE is based on a convex measure. See Murray S. Cohn, “Estimated Carcinogenic Risks due to Exposure to Formaldehyde Released from Pressed Wood Products,” (Consumer Product Safety Commission, May 1986), J.A. at 1295, 1297.4 In practical terms for our purposes, a convex dose-response curve for low doses means, relative to a linear one, that the responses rise more gradually from the zero-dose level.5 The core issue is whether OSHA adequately explained its implicit refusal to assume linearity (or, more strictly, its very severe downweighting of the data based on that assumption).
There appears no dispute that the assumption of linearity or convexity depends on whether the “background” cancer risk from other carcinogens increases from interaction with the carcinogen in question (entailing linearity) or is “independent” (entailing convexity). Jean Chesson and Mary Anne Zanetos, “Documentation of High- to Low-Dose Extrapolation Models Used in Quantitative Risk Assessment: Assessment of Models,” EPA Draft Report (February 2, 1984), J.A. 1863, 1879-81. OSHA here explicitly accepted this proposition:
If various exposures to formaldehyde and other carcinogens present in the general environment contribute synergistically to the same chain of events leading to production of a tumor, the maximum likelihood estimate (MLE) for the multistage model based on animal data will not correct for this additional risk. Consequently, some groups prefer to use the upper confidence level (UCL) because it is less likely to underestimate human risk.
52 Fed.Reg. at 46,215/2. Chesson & Zanetos express it in terms of the relevant carcinogenic mechanism: “The additivity [non-independence] assumption is appropriate whenever the test agent shares a mechanism with at least one of the agents producing the background response----” J.A. 1879.
As to the application of these principles, OSHA not only recognized that the carcinogenic mechanism of formaldehyde was similar to that for other relevant carcinogens, 52 Fed.Reg. at 46,217/1, but also explicitly acknowledged strong evidence of synergy between formaldehyde and other carcinogens, id. at 46,224/1, and particularly formaldehyde and wood dust, id. at 46,185/2-3. The latter is a combination affecting a large share of formaldehyde-exposed workers. See id. at 46,239.
OSHA’s reasoning here is impenetrable. It recognized the need for a linear low-dose risk curve when conditions of synergy exist and accepted the evidence of synergy between formaldehyde and other carcinogens, especially the highly relevant wood dust. Under its own view, these seem to entail use of the UCL — presumably adjusted, to the extent possible, to exclude its inherent risk aversion.
We note that OSHA purported to address the problem of synergy in a brief passage in which it “assum[ed] a plausible but arbitrary background rate to substitute into the multistage model.” Id. at 46,217/2. It then proceeded to do so, and came up with the conclusion that the effects were fairly trivial. Id. at 46,217/3. We confess that we are unable to assess the validity of this operation. But given OSHA’s recognition of a quite different standard approach to the issue of synergy, and its failure to give *294much weight to the results that flow from that approach, the passage does not alter our conclusion that OSHA has failed adequately to explain its decision.
That inadequacy is underscored by the gulf between the MLE risk estimates and the epidemiological data — with the Blair study, as adjusted by Sterling and Weinkam, showing a lung cancer risk 1,543 times greater than the MLE estimate. See Memorandum to John Pendergrass, Assistant Secretary, from John Martonik, Acting Director, Office of Risk Assessment, Ex. 255-20 in Appendix of Statutes, Regulations and other Referenced Materials. Granted, the Sterling and Weinkam data were subject to significant uncertainties, as OSHA noted in explaining its decision to downweight the UCL. See 52 Fed.Reg. at 46,223/3. But given its use of those data for the basic finding of carcinogenicity, they at least tended to strengthen the case for relying on the UCL, and thus to leave OSHA’s contrary decision the more obscure.
Finally, the agency sought to justify the 1 ppm PEL with an argument that the other aspects of the regulation, the STEL and the ancillary measures, would “likely decrease the risk of exposure to levels more representative of the lower end of the range of risks [i.e., near to the MLE].” See 52 Fed.Reg. at 46,224/1. The STEL of course provides direct protection only against short bursts of high-dose exposure. We may assume that its existence will force adoption of controls that will in practice tend to reduce average exposures to below the PEL and thus have some beneficent spillover effect on cancer incidence stemming from the eight-hour average. But OSHA offers no data suggesting that this effect is strong enough to justify its implicit rejection (or downweighting) of the linearity assumption and resulting calculations. Compare Building & Construction Trades Dep’t, 838 F.2d at 1265-66 (recognizing potential significance of gap between nominal ceiling and the levels that its adoption would actually bring about). The same objection applies to its reliance on the other ancillary measures, to the extent we may assume they will actually reduce exposures.
OSHA acknowledged the truth of facts that appear to compel it to assume linearity at low doses in its risk assessments. It then proceeded either to discard the resulting estimates or at least give them comparatively scant weight. As there may be some valid explanation for this course, as yet not adequately disclosed, we remand for the agency to reconsider the matter and either express an adequate reason or adjust its implicit risk estimates.
II. Non-Carcinogenic Effects of Formaldehyde
Formaldehyde also induces irritant effects on the skin, eyes, and respiratory organs. See 52 Fed.Reg. 46,173-83. Exposure may occur either by direct contact with formaldehyde solutions or finished products containing the chemical, or through inhalation and other contact with airborne molecules. 52 Fed.Reg. at 46,173/1. The primary dermal effects are redness and thickness of the areas that have come into contact with formaldehyde. Reactions from airborne exposure include burning or itching of the eyes or nose, sore throats and asthma episodes. Id. at 46,-179-80.
Normally, reactions heal within days or a few weeks of the cessation of exposure. Indeed, most persons exposed to formaldehyde over any extended period become acclimated to it and thus come to experience its effects more mildly. Some individuals, however, become “sensitized” and suffer a condition known as allergic contact dermatitis. Id. at 46,174/1-2. These workers suffer itching, redness, swelling, multiple small blisters and scaling whenever subsequent contact occurs, and these reactions tend to become more severe with repeated exposures. Id. at 46,174/1. As a consequence, sensitized individuals find it difficult, if not impossible, to remain in formaldehyde-related jobs. OSHA received evidence that “there are currently no data to determine precisely a safe level of dermal contact to formaldehyde for a sensitized individual.” Id. Preliminary research has *295shown that the level of formaldehyde must be reduced to approximately 0.3 percent in liquid products before the majority of sensitized persons can tolerate them. Id. at 46,174/2. Some sensitized workers react even to concentrations below this amount. Id.
There are obvious difficulties in ascertaining who is prone to sensitization. OSHA did note, however, that people who are atopic — that is, those who suffer from allergic asthma, allergic hay fever, or infantile or childhood eczema — are most likely to experience severe irritant effects from formaldehyde. Id. at 46,174/1. Roughly 20-30% of the population falls into this category. Id.
The unions assert that the non-cancer effects of formaldehyde remain significant at the 1 ppm PEL level. They also claim that further reductions in the PEL are feasible and would produce more than de minimis benefits to affected workers. See Building and Construction Trades, 838 F.2d at 1271 (“[a] party challenging an OSHA standard must bear the burden of demonstrating that the variations it advocates will be feasible to implement and will provide more than a de minimis benefit for worker health.”).
A. Dermal Effects
Most dermal reactions can be traced to direct contact with formaldehyde. Sensitized persons, however, apparently can suffer dermal reactions when their only exposure is to the airborne form. It is uncertain whether non-sensitized individuals can be dermally afflicted when their only exposure is airborne. See 52 Fed.Reg. at 46,177/2 (“[wjhile allergic skin reaction can result from airborne exposure to formaldehyde, at low doses this reaction probably occurs only in persons who have already been sensitized to formaldehyde.”).
OSHA concludes that for non-sensitized workers its requirements of equipment such as gloves and goggles will protect most, if not all, workers from dermal exposure. 29 C.F.R. § 1910.1048(h); 52 Fed. Reg. at 46,269-73. That subsection requires employers to conform to OSHA’s generally applicable regulations governing use of protective garments, 29 C.F.R. § 1910.132-33,6 and also sets out supplemental rules specific to formaldehyde.
The unions argue that the conditions prevailing in textile and apparel plants are such as to undermine OSHA's predictions of employer and employee compliance with these subsections:
[f]or workers who handle formaldehyde ... use of long sleeve shirts and gloves are far from pleasant or convenient. To be effective gloves and sleeves must be impervious to water. Textile and apparel operations are not industries where air conditioning is the norm; to the contrary, it is usually hot, heavy work____ Under these circumstances, it is unreasonable to simply assume — as OSHA does — that workers will be able to wear such protective equipment constantly and effectively perform these jobs.
Unions’ Reply Brief at 17. They point to two locations in the record as support, but we find that neither citation sustains the unions’ assertions. The first, Reply Brief at 16, is to the testimony of Ms. Sandra Simpson, J.A. at 270. After studying the record, we are unable to find the phrase attributed to Ms. Simpson (or anything like it) in her recorded testimony. The second, Reply Brief at 17, is to a Booz-Allen study which the unions claim supports the proposition that “[Tjextile and apparel operations are not industries where air conditioning is the norm.” The Booz-Allen statement actually discusses conditions in “warehousing, shipping and receiving areas,” which it notes “normally [are] not air conditioned.” *296J.A. at 434. Nothing required OSHA to follow the unions in their inferential leap from those specific areas of the workplace to the apparel and textile industries as a whole.
No evidence undermines the agency belief that employers will obey the regulations governing them. That being so, we must sustain OSHA’s reliance on the provisions of § 1910.1048(h) to prevent significant risk of dermal effects.
As to the dermal effects of airborne exposure for sensitized workers, the unions have pointed to no evidence establishing that any feasible lowering of the PEL would result in greater than de minimis benefits, Building and Construction Trades, 838 F.2d at 1271; the apparent absence of any threshold tolerance for formaldehyde indicates its improbability. For such individuals the only effective remedy may be medical removal protection, discussed below at 23-25.
B. Airborne Exposure Effects
OSHA determined that the lingering airborne exposure effects at the 1 ppm level do not rank as significant. The agency credits three studies which show that at the 1 ppm level only 20% of workers suffer “mild discomfort,” while 30% more experience “slight discomfort.” See 52 Fed.Reg. at 46,224/1-3. The first study, supervised by lb Anderson and Lars Molhave, tested student volunteers’ reactions to varying levels of formaldehyde exposure for 5-hour periods. At an exposure level of 1.6 ppm, the subjects’ average discomfort level never rose above 18 on a scale of zero-to-100, where zero meant no effects, 17 meant “mild to slight discomfort,” and 100 meant “intolerable discomfort.” 52 Fed.Reg. at 46,224/1-2. OSHA found the evidence from a Booz-Allen study conducted for the Formaldehyde Institute’s predecessor organization, SOCMA, to be in line with the Anderson-Molhave conclusions. Id. The third study, done by the National Academy of Sciences in 1981 reached the following results:
Percent of population giving Concentration indicated Degree of (in ppm)_response irritation
•1.5-3.0 10-20 7-10
30 5-7
0.5-1.5 10-20 5-7
30 3-5
0.25-0.5 20 3
Less than 0.25 20 1-3
Irritation Index: 10 = strong irritation, great discomfort; 7 = moderate irritation, discomfort; 5 = mild irritation, mild discomfort; [3 = slight irritation, slight discomfort;]7 1 = minimal irritation, minimal discomfort; 0 = no effects.
Id. at 46,224/3. OSHA concluded that at 1 ppm about half of all non-acclimated persons who entered the worksite would suffer “[m]ild irritation of the eyes, nose, and throat,” of whom about half would complain of it. Id.
The unions question OSHA’s reliance on the studies. They object that the youth and health of Anderson and Molhave's student volunteers, and the comparatively brief periods of their exposures, make it improper to generalize from their findings to workers of all ages exposed to formaldehyde over a lifetime. The OSHA preamble recognized several possible flaws in the Anderson-Molhave and SOCMA estimates, but concluded that their general conformity with other irritant studies rendered them probative. 52 Fed.Reg. at 46,224/2-3. Additionally, at oral argument the agency noted that while the students’ brief exposure to formaldehyde was certainly different from the average worker’s, individuals normally become acclimated to the chemical, so that students might experience formaldehyde’s irritant effects more severely than employees who were regularly exposed.
We think that at least for the non-sensitized workers OSHA could legitimately conclude that the residual irritant effects *297at 1 ppm were not “significant.” Under Benzene it is bound to fit the facts into this necessarily somewhat subjective classification. Here all the facts are themselves highly subjective, again necessarily so, and also are framed in terms (“minimal,” “slight,” “mild,” etc.) that do not correspond to a dichotomy between significant and insignificant. We have no basis for upsetting OSHA’s finding.
As to the sensitized workers, we again note the absence of a showing by the unions that any feasible reduction would generate more than de minimis benefits and the resulting likelihood that medical removal protection may be the only relevant potential solution. See below at 399-400.
While upholding OSHA’s determination that the noncancer effects of formaldehyde at 1 ppm are not significant for non-sensitized persons, we do not understand that anything other than analytic convenience causes their segregation from formaldehyde’s carcinogenic effects. Accordingly, one can imagine that the remand we order for the cancer issues might lead to a finding that at 1 ppm the carcinogenic effects alone would be insignificant but that the two types together would be significant.
III. Ancillary Requirements Under the Standard
The unions challenge several aspects of OSHA’s ancillary requirements. The hazard communication provisions, attacked by the unions as too weak and by industry as too strong, have as noted above been taken under reconsideration by the agency and are viewed by all parties as unripe. We treat the remaining objections in turn.
A.Action Level
The unions are dissatisfied with OSHA’s decision to trigger the imposition of some of the ancillary measures — particularly those concerning training and medical programs — only at the action level of 0.5 ppm. Yet, as in Building and Construction Trades, the unions have “failed to point to any evidence that triggering the ... duties at [a lower] level ... would result in a greater than de minimis incremental benefit,” see 838 F.2d at 1274, and as there we do not find that the Secretary has abused his discretion.
B. Medical Examinations
OSHA has not required employers to provide annual medical examinations to workers exposed to formaldehyde. The unions complain about this omission, citing the provisions for such tests in prior toxic material standards, notably those for asbestos and benzene. (See UAW Brief at 42 n. 36 for a full listing.) OSHA responds by observing that, unlike the health hazards from the chemicals identified by the unions, formaldehyde’s dangers are not likely to be isolated during annual checkups. Its short-term effects (such as the dermal and respiratory effects) are usually immediately noticeable, and thus a yearly examination would do nothing for an employee suffering from them. Its longer-term, carcinogenic consequences do not give warning signs. OSHA Brief at 64-65; 52 Fed.Reg. at 46,278/1 (“the [medical surveillance] tests are oriented toward detection of reversible effects and not to chronic effects such as cancer.”). This distinguishes formaldehyde from, for example, asbestos (in which reversible and early signs can appear) or benzene (in which pre-leukemia tumors arise). OSHA Brief at 64-65. Given the apparent inefficacy of annual examinations in the formaldehyde context, we find nothing arbitrary in the agency’s decision to rely on medical surveillance by the employer, coupled with examinations for workers who manifest symptoms.
C. Medical Removal Protection
Medical removal protection (“MRP”) programs, which OSHA has authorized in other standards, provide temporary assistance to workers whose health is affected by exposure to a toxic substance. The protection routinely includes maintenance of the employee’s earnings and seniority at the employer’s expense during a rehabilitative period. See, e.g., United Steelworkers of America v. Marshall, 647 F.2d 1189, 1228-30 (D.C.Cir.1980) (describing MRP provisions under lead and cotton dust standards). OSHA declined to establish a MRP program for formaldehyde-exposed workers. See 52 Fed.Reg. at 46,282/1-2.
*298The agency claimed that MRP would be ineffective for sensitized workers on the theory that they would likely never be able to return to an occupational environment with formaldehyde present. 52 Fed.Reg. at 46,282/1. But the union’s theory was that MRP would benefit even these workers by shielding them from financial insecurity during an employment retraining period. OSHA did not respond to this point in the preamble. In its brief before us, agency counsel attempted to justify OSHA’s position by asserting that MRP was not intended to cover permanent shifts from exposed to unexposed work. See OSHA Brief at 67 (“[wjhen permanent removal from exposure is needed, workman’s compensation is the prescribed remedy”) (citing Steelworkers, 647 F.2d at 1235).
OSHA's explanation is feeble even without regard to its post hoc appearance. OSHA has in fact used MRP in the past to aid workers with permanent disabilities similar to sensitization, requiring it in order to enable workers exposed to coal dust, and bearing signs of Black Lung disease, to transfer to jobs with lower coal dust levels. See Occupational Exposure to Lead, 43 Fed.Reg. at 54,447/2-3. In doing so, OSHA provided for worker retention of the same rate of pay and explicitly recognized that because the disease was “irreversible,” the transfers would be “permanent.” Id. at 54,447/3. Thus, OSHA’s permanency argument represents a thoroughly unexplained “swerve” from prior practice, requiring at the least some explanation. Greater Boston Television Corp. v. FCC, 444 F.2d 841, 852 (D.C.Cir.1970). Moreover, OSHA appears to completely misread Steelworkers, which acknowledges that as a practical matter MRP may overlap with workmen’s compensation. See 647 F.2d at 1235-36.
As to non-sensitized workers suffering acute irritant effects, OSHA reasoned that the “nonspecificity of signs and symptoms, which makes an accurate diagnosis of formaldehyde-induced irritation difficult and complicated,” and the “quick resolution of the effects,” made MRP inappropriate. 52 Fed.Reg. at 46,282/1. On the second aspect, the non-sensitized workers’ supposed quick recovery when exposure stops, the unions aptly point out that the preamble itself states that “[wjhen the damaging exposure ceases, irritant reactions usually heal within days or a few weeks at most.” 52 Fed.Reg. at 46,173/3. If this is accurate, affected workers surely could benefit from receiving MRP during the recovery period.
Finally, although we are unsure of OSHA’s point in its observation about the difficulty of diagnosing formaldehyde-induced irritation, we take it to relate to union claims that without effective MRP workers would not report their symptoms —thereby undermining OSHA’s reliance on worker self-reporting in the medical surveillance scheme. UAW Brief at 45-47. The unions cite undisputed testimony that, in the absence of MRP, workers would underreport symptoms for fear of economic loss. See UAW Brief at 45 nn. 39, 40. This is, indeed, precisely the view that OSHA took in promulgating the lead standard:
the possibility that some workers would fully participate without MRP in no way detracts from the pressing need to provide MRP for the many workers who would otherwise resist meaningful participation in medical surveillance programs. OSHA is determined to protect all inorganic lead workers so far as is feasible, and feels the conclusion is inescapable that MRP is essential to maximize meaningful worker participation in the medical surveillance program provided by the final standard.
43 Fed.Reg. 54,354, 54,449 (November 21, 1978). In light of these past statements, we find the allusions to “non-specificity” of symptoms too vague and obscure either to show consistency with OSHA’s prior stance or to justify a reversal of position.
IV. Conclusion
We remand the case to OSHA for reconsideration of its calculation of the risk of cancer from formaldehyde at 1 ppm (particularly its choice of a non-linear curve for extrapolating the risk from animal studies to low-dose human exposures) and for reconsideration of its failure to require medi*299cal removal protection. Apart from those matters, we affirm.8
So ordered.
. These policy considerations evidently are not to include, at least as a material factor, a weighing of the significance of what must be done to avert or reduce the risk. See American Textile Mfrs. Inst. v. Donovan, 452 U.S. 490, 506-22, 101 S.Ct. 2478, 2489-97, 69 L.Ed.2d 185 (1981).
. No party has sought to explain to us the variation among MLE estimates or among UCL figures, and if the record discloses the reasons, no member of the panel has been able to detect them. See J.A. 644, 650 for figures used above. Curiously, the table showing the high UCL shows the low MLE, and vice versa.
. See generally Albert L. Nichols & Richard J. Zeckhauser, The Perils of Prudence: How Conservative Risk Assessments Distort Regulation, 10 Regulation 13 (Dec. 1986).
. With the figures presented at 11 above, the model cannot be linear throughout, given the no-threshold assumption (i.e., the assumption that risk reaches zero only when exposure reaches zero). Attempting to graph the four points (response at doses of 0, 2, 5.6 and 14.3 ppm) will immediately reveal this point.
.No party has sought to explain why the risk-aversion of the UCL is bundled with linearity, whereas the relative neutrality of the MLE is bundled with convexity. But see Cohn, J.A. 1299, noting the bundling of the two features and the possibility of a "crude" unbundling.
. The general thrust of 29 C.F.R. § 1910.132 is that "[P]rotective equipment, including personal protective equipment for eyes, face, head, and extremities, protective clothing, respiratory devices, and protective shields and barriers, shall be provided, used, and maintained ... wherever it is necessary by reason of hazards ... encountered in a manner capable of causing injury or impairment in the function of any part of the body through absorption, inhalation or physical contact____’’ 29 C.F.R. § 1910.133 is a more specific provision concerning eye and face equipment.
. The reference to the category for slight discomfort was apparently inadvertently omitted by OSHA in the preamble table. The NAS Report discusses the category, however. See J.A. at 532-34.
. The unions moved to strike the various industry trade group petitioners, noting that once OSHA agreed to reconsider the hazard communication provisions, industry had no further objection to the rule as promulgated. The only significance of this motion is that as mere intervenors in support of the rule, industry would have been limited to a brief of 35 typewritten pages, rather than being free to file one of 50 typewritten pages as petitioners. Obviously parties must not manipulate their status to evade page limits. As nothing in the industry brief has affected our decision, however, we dismiss the motion as moot.