STRONG [O iii] OBJECTS AMONG SDSS BROAD-LINE ACTIVE GALAXIES

We selected a large sample of broad-line AGN spectra from the SDSS Data Release 5 (DR5; Adelman-McCarthy et al. 2007; York et al. 2000; Stoughton et al. 2002). The AGNs in SDSS are selected by two primary methods. First, quasar candidates were targeted for spectroscopy by a photometric color and morphological selection above a limiting magnitude of i < 19.1 (Richards et al. 2002). Second, AGNs were spectroscopically discovered in spectra that were targeted for other reasons, for instance as probable galaxies (Strauss et al. 2002; Eisenstein et al. 2001). When the latter spectra were cross-correlated with the SDSS quasar template spectrum, they were identified as quasars due to their prominent broad lines and blue continua.

We obtained our sample and spectral measurements from Salviander et al. (2007), which they have since extended to DR5. We selected all the DR5 objects with spectral coverage of the rest-frame region 4000–6000 Å on which we performed the SPCA. This selection includes Fe ii and [O iii] in the observed wavelength range of the SDSS. We chose not to extend this wavelength region further to the red because it would result in a more severe redshift restriction, nor to the blue because we wanted the ability to do a more straightforward comparison with BG92 and they did not use data blueward of 4000 Å for their sample. This limited our sample to the redshift range z ≲ 0.56. We also chose to exclude objects with z < 0.1 to reduce the need of fitting the host galaxy continuua in low-luminosity AGNs. There are 12040 quasars in SDSS DR5 with 0.1 < z ≲ 0.56.⁴ For consistency, all redshifts used in this paper are the SDSS DR5 spectroscopic redshifts derived from template matching rather than the Salviander et al. redshifts derived from line fitting.

Salviander et al. fitted the spectra in the wavelength region 4000–5500 Å with a power law, Fe ii pseudocontinuum derived from the Marziani et al. (2003a) template, the narrow-line [O iii] λλ4960, 5008, and narrow and broad Hβ. Note that there was no fit to the host galaxy continuum. The [O iii] line ratios were fixed to their laboratory value of 1:3 for the [O iii] λ4960 and λ5008 lines, respectively, and all wavelengths were fixed to their vacuum values as well. The narrow component of Hβ was fixed to the same redshift and line profile as the [O iii] line, with 10% of the flux. The broad component of Hβ was allowed to vary independently. The emission lines were all fitted using Gauss–Hermite polynomials, which allow straightforward parameterization of line shapes and asymmetries. The measurements include L_5100 Å, which is λL_λ at 5100 Å, rest-frame EW_Fe ii, and the line shape, rest-frame EW, integrated line flux, and FWHM for each of the emission lines. The fitting procedure is described in more detail in Salviander et al.

We decided to be more conservative in our sample selection than Salviander et al. by excluding potential Type II AGNs from our sample, since we wished to investigate relationships between the BLR and NLR. To do this, we conservatively eliminated objects with FWHM of broad Hβ < 2000 km s⁻¹. We compared the results of performing an SPCA on this sample with a sample including objects with FWHM Hβ as low as 1050 km s⁻¹ and found that the principal components were virtually identical. The linewidth cut reduced the sample to 9362 objects.

Finally, like Salviander et al., we eliminated the objects with a failed fit of the Hβ broad line, since we wished to use Hβ parameters in our analyses. In general, the fits failed when the spectrum had a low S/N, the absence of the broad line, or cosmetic defects such as cosmic rays. However, unlike Salviander et al., we chose to include objects with failed fits of [O iii] since they were likely weak [O iii] objects. The Hβ failed fits removed only 316 objects, or 3% of the sample. Our final data set consisted of 9046 objects.

The objects in our sample were targeted for spectroscopic follow-up by the SDSS pipeline for a variety of reasons. We found that 88% of the sample were primarily targeted as quasars of some sort. The very low luminosity objects were generally targeted as galaxies, but this represents only 0.4% of the total sample. Only 4% of our sample were "serendipitous," but these do not have unusual principal component properties (see below). The other 7.6% were targeted primarily for an assortment of other reasons, including ROSAT and Faint Images of the Radio Sky (FIRST) detections, blue stars, etc.

PCA is a powerful tool because it provides a means of classifying and potentially distilling important relationships among the data (Francis et al. 1992, and references therein). The PCA extracts orthogonal eigenvectors, each consisting of a linear combination of input variables. They are conventionally ordered according to the fraction of sample variance represented by each eigenvector. Although there is no a priori reason, the hope is that the sample variance can be represented by only a small number of eigenvectors, so the PCA results in a drastic simplification of the entire data set. For example, BG92 found that 51% of the variance in their sample was represented by the first two eigenvectors. Additionally, one hopes that these eigenvectors illuminate physical relationships. Indeed, BG92 (updated by Boroson 2002) found that EV1 is related to the M_BH or Eddington ratio, and EV2 is related to L. However, in nature the relationships among various parameters of an object may not be well described by a linear analysis. For instance, a certain characteristic may apply to only a subset of objects. This may make physical interpretations difficult and confuse relationships among the input parameters, but the PCA still appears to provide a useful description of the data.

While many authors have used direct line and continuum measurements as input data for the PCA, there is another approach. Spectral PCA or SPCA has been proven to reduce various samples (Brotherton et al. 1994; Shang et al. 2003; Yip et al. 2004) down to only a handful of components that reproduce the vast majority of quasar-to-quasar spectral differences. The SPCA uses the flux densities binned in wavelength as input parameters to solve for orthogonal eigenvectors, or principal components (PCs; Francis et al. 1992). Since the input variables are not the same, the features responsible for most of the variance within a sample can be different than in a PCA. For example, many continuum bins are likely to result in a greater contribution from the continuum to the PCs in an SPCA. The resulting PCs are not typical spectra, with flux versus wavelength, but are representations of how wavelength regions correlate with one another (C_λ). Positive correlations between wavelength regions appear as features pointing the same direction, while anticorrelations appear with opposite directions. The SPCA retains information for each input spectrum in the form of the coefficients (scores) for each PC. The coefficients can then be compared with various measured properties of the AGNs (e.g., black hole masses) in the hopes of gaining additional insights. One benefit of the SPCA is the ability to easily identify extreme objects as outliers in a PC–PC plane using the PC coefficients for each object. Another benefit of the SPCA is that there is no need to extract measured quantities from the spectra, so there are no added uncertainties from continuum estimation or fitting errors. In our case, the spectrum-to-spectrum noise in the results is reduced by the large numbers of SDSS spectra. We can include spectra with low signal-to-noise ratios whose features may be difficult to measure accurately, and these spectra will not bias the results.

While SPCA is a different approach to a PCA like BG92's because it uses very different input parameters, in the case of optical spectra of luminous quasars, they still yield meaningful and comparable results. Shang et al. (2003) carried out an SPCA on the same limited optical region that we used for a sample made up of a subset of the quasars in BG92. By comparing their components with measured spectral parameters, they were able to justify their physically meaningful interpretations of the PCs. They also showed that, in the same optical region as in BG92, as few as two PCs dominate the sample variance. Moreover these, PC1 and PC3, appeared to represent the same physical quantities as BG92's EV1 and EV2. They showed this by a direct correlation of the coefficients of their PCs for each spectrum with the physical quantities apparently underlying the BG92 eigenvectors. These earlier studies were for relatively small samples of luminous quasars (M_V ≲ −23). Can we extend these studies of the optical region to the much larger SDSS sample, covering a wider range of L?

The mean spectrum and first two PCs for our entire data set are shown in Figure 1. We use ALL:PC1 to denote the first PC from the SPCA run on the entire data set, ALL:PC2 for the second, and so on. ALL:PC1, which accounts for 35% of the variance, appears to show a strong correlation among EWs of lines from the NLR, including [O iii], the narrow Balmer lines, and He ii λ4686. There is a less prominent anticorrelation between the EW of the narrow lines and the EW of the Fe ii blends that originate in the BLR. If this qualitative interpretation of the PCs is correct, one would expect to see a correlation between the ALL:PC1 coefficients for individual spectra when compared with the measured values of EW_[O iii] and an inverse correlation with EW_Fe ii. We plot these values in Figure 2 and find that, in fact, the EW_[O iii] is indeed strongly correlated with ALL:PC1 coefficients, with a Spearman correlation coefficient of 0.95. Figure 2(b) shows a less obvious anticorrelation between EW_Fe ii and ALL:PC1, which has a Spearman correlation coefficient of −0.297. Given these relationships, ALL:PC1 is the component that most closely reproduces the correlations in the traditional EV1. However, in at least one respect ALL:PC1 deviates significantly from the BG92 EV1. ALL:PC1 does not appear to include any correlation with the line width of broad Hβ, which would be present in the PC as a "M"-shaped pattern centered at Hβ (Shang et al. 2003, hereafter S03). There is no strong correlation between ALL:PC1 and the FWHM of Hβ (Spearman correlation coefficient of 0.076; Figure 2(c)). At face value, our result disagrees with the classic BG92 finding that the strength of the Fe ii emission strongly anticorrelates with the FWHM of Hβ. By plotting the EW_Fe ii directly against the FWHM of Hβ (Figure 2(d)), we note that the entire sample appears to have an anticorrelation between the two, but there is a large amount of scatter (Spearman correlation coefficient of −0.198). Several authors (S03, Yip et al. 2004) have performed similar SPCAs using the rest-frame optical region of the spectrum. While the small, high-luminosity sample used by S03 seems to recover EV1 correlations, the large samples used here and in Yip et al. (2004) do not include the strong linewidth dependence of broad Hβ.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The dominant correlations present in ALL:PC2, which accounts for 20% of the variance, are positive correlations between all broad lines and continuum shape (objects with a strongly sloped continuum will have more contribution from PC2 compared with objects with a flat continuum). We also find that ALL:PC2 is correlated with L₅₁₀₀, with a Spearman correlation probability of <0.001 that they are not correlated. Further components of the entire data set each account for <12% of the variance among the spectra and we do not offer interpretations of them.

Since our large sample and others do not seem to entirely recover the traditional EV1 correlations, we wonder if our choice of such a large, diverse sample is potentially masking our ability to notice relationships among the data. Over the range of properties that the objects in our sample exhibit, the distribution of a given property may saturate for a subset of objects. Alternatively, the relationship between properties need not be linear. For instance, the EW_[O iii] spans such a large range relative to other properties that it dominates ALL:PC1 and therefore the SPCA results of the entire data set. Thus, an SPCA that uses the entire sample is possibly less likely to result in PCs that are directly interpretable in terms of physical properties.

One way to investigate relationships that might be diluted by the large range of properties in the entire sample is to divide the sample into subsets. The SDSS, with the large sample it provides, gives us an opportunity to investigate subsets with particular properties. Because ALL:PC1 encompasses the relationships representing the largest spectrum-to-spectrum variation among the objects in the entire data set, we chose to divide our sample into subsets based on the ALL:PC1 coefficient of each object. We settled on the following three subsets, which isolate the extremes of the EW_[O iii] range: S1, which has ALL:PC1 <0 and includes 6317 objects, S2 with 0< ALL:PC1 <7 includes 2307 objects, and S3 with ALL:PC1 >7 includes only 422 objects (see Figure 3). These subset divisions are roughly equivalent to EW_[O iii] < 15 Å, 15–50 Å, and >50 Å for S1, S2, and S3, respectively. However, by restricting the range of EW_[O iii], we realize we might be reducing the importance of any relationships that do include EW_[O iii].

Zoom In Zoom Out Reset image size

3.2.1. Subset Results

The resulting PCs for S1, which includes the low EW_[O iii] objects where ALL:PC1 <0, are in the left column of Figures 4 and 5. It includes ∼70% of the original data set and therefore illuminates the relationships among the bulk of broad-line AGNs. This subset consists of objects with a negative value of ALL:PC1, which means that reconstructing their spectra requires subtracting NLR emission from the original mean spectrum, so this subset includes the weakest EW_[O iii]. As objects representing the strongest EW_[O iii] have been removed, ALL:PC2, representing the relationship between BLR emission and the continuum slope, becomes the first principal component (S1:PC1). S1:PC2 includes a correlation among EWs of all emission (Balmer lines and He ii) and [O iii]. This subset also more closely recovers EV1 relationships in S1:PC3. Note that there is a "W" linewidth signature for Hβ in S1:PC3, so this component links FWHM_Hβ, Fe ii, and [O iii] in a way similar to EV1. To cross-check our qualitative interpretation of the PCs, we compare these three quantities with the coefficient of S1:PC3 (Figure 6). The EW_[O iii] and FWHM_Hβ anticorrelate with the S1:PC3 coefficient, with Spearman correlation coefficients of −0.447 and −0.549, respectively. The EW_Fe ii shows a correlation with S1:PC3 and a Spearman correlation coefficient of 0.310. These are the sense of relationships we would expect to see within BG92's EV1, and when we examined EW_Fe ii versus FWHM_Hβ (Figure 6(d)), we found an inverse correlation for S1 (Spearman correlation coefficient of −0.221), although there is an enormous amount of scatter in the data. Additional PCs account for only a few percent of the object-to-object variation within the subset, and so we do not interpret them.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The third subset, S3, which includes less than 5% of the entire data set, is responsible for the most prominent relationships in ALL:PC1. S3:PC1 (right column of Figures 4) shares the relationships seen in ALL:PC1. Effectively, the extreme variation of EW_[O iii] embodied in S3 compared to the rest of the entire sample was dominating the relationships that account for the most object-to-object variance among the 9046 object data set.

As is evident after examining the S3:PCs, they are entirely dominated by variations in NLR emission lines, except for S3:PC4, which shows the same continuum slope and broad Balmer lines as ALL:PC2. The S3:PCs represent correlations among NLR lines in terms of EW, asymmetry or shift, line width, and even stronger asymmetry in S3:PC5. However, note that even though these 422 objects are clearly dominated by their NLR emission, they are selected to be broad-line AGNs, not Type II AGNs (see Section 4). The mean spectrum for this subset shows a distinct broad Hβ contribution to the spectrum. We will address in Section 5 possible reasons for these objects' extraordinarily strong NLR emission lines.

In the middle column of Figures 4 and 5, we present results for the second subset, S2, which includes objects with ALL:PC1 coefficients between 0 and 7. This subset, which is comprised of 2307 objects or about 1/4 of the entire data set, has properties intermediate between S1 and S3. In S2:PC1, we basically recover ALL:PC2, which is a component dominated by broad Balmer lines and continuum shape. S2:PC2 seems similar to ALL:PC1, dominated by NLR emission, but with additional continuum information as well. S2:PC3 represents asymmetry in the NLR lines that could be caused by errors in the redshifts or real shifts in the emission lines.

Past authors (e.g., Ho & Peng 2001, and references therein) investigated the degree to which radio power correlates with L_[O iii], and found that for radio-quiet sources, the correlation is quite strong. We investigate the radio properties of the objects in the strong [O iii] subset to see if they have correspondingly strong radio emission. We cross-correlated our list of 9046 SDSS objects with the Faint Images of the Radio Sky at Twenty-Centimeters (FIRST) catalog of radio sources. The FIRST survey was conducted using the Very Large Array (VLA) and covers over 9000 square degrees of sky. The detection threshold is 1mJy at 1.4 GHz in the observed frame. We find 735 radio detections within 2'' of the optical source for our sample. We chose a 2'' search radius, as other authors show (Ivezić et al. 2002; Rafter et al. 2009; Husemann et al. 2008) that the number of meaningful correlations with optical AGN detections drops off rapidly outside 2''.

Of the 9046 objects in our entire data set, 8460 lie within the FIRST footprint, rejecting less than 10%. Therefore, the 735 detections represent 8% of the data set within the observing area of FIRST. However, of the 422 objects in S3, 404 were within the footprint, with 92 detections, yielding a fraction of 22% detected. Thus, a higher proportion of the strong [O iii] objects have detectable radio emission.

The S3 objects have the same redshift distribution as the entire data set (Figure 10), so the higher detection percentage for S3 is possibly meaningful but susceptible to various selection biases. As an extra precaution, we follow the prescription in Ivezić et al. (2002) for defining a "stringent radio sample" that is not subject to a Malmquist bias for detections from either FIRST or SDSS. The Malmquist bias results in extra detections near the flux limit of a given survey due to statistical errors in detection.

Zoom In Zoom Out Reset image size

We selected a complete sample above, well within the flux limits of both the SDSS and FIRST. In Figure 11, we have plotted the observed radio magnitude, t = −2.5log(F_rad/3631Jy), where we use the integrated radio flux density of the detected objects versus their i-band magnitude following Ivezić et al. (2002). We calculate the radio to optical flux density ratio as log R = 0.4(i − t) (Ivezić et al. 2002). The flux limits of both SDSS, at i < 19.1, and FIRST, at t < 16.5, are shown as the dotted vertical and horizontal lines. To simplify the analysis, we did not include a K-correction, but chose to use the observed magnitudes for the radio and i band.

Zoom In Zoom Out Reset image size

To select a complete sample, we wanted to span the full range observed in log R, where the short-dashed lines are lines of constant log R. We selected the stringent radio sample by taking objects that lie above the long-dashed line perpendicular to the lines of constant log R and within 0< log R < 3, which then includes only objects safely within the flux limits. The solid line denotes log R = 1, which separates radio-loud objects (log R > 1) from radio-quiet objects (log R < 1). Within this stringent sample, we have plotted objects included in S1 as red triangles, objects in S2 as blue squares, and objects in S3 as green circles.

If radio properties were really physically related to extreme NLR emission, then we would expect the S3 objects within the stringent radio sample to have significantly higher values of log R than those in S1 and S2. However, within the stringent sample, the distribution of S3 objects is not noticeably biased toward the radio-loud side of the sample. Figure 10 includes the histogram of the distribution of S3 objects in t, compared with the rest of the stringent sample. While the K–S test results in only a 0.6% chance that the whole stringent radio sample and S3 objects within it are from the same parent distribution, there does not appear to be a striking correlation of radio loudness with [O iii] emission. One caveat to this conclusion is the small number statistics within the stringent radio sample, as there are only nine objects from S3. The radio properties might be somewhat different between the S3 objects and the rest of the stringent sample, and S3 does appear to have a higher percentage of objects that are detected in the radio compared with the entire data set, but there is not a simple increase of radio loudness with EW_[O iii].

In the orientation Unified Scheme, AGNs are surrounded by an optically thick dusty torus (e.g., Antonucci 1993). If the observer's line of sight to the nucleus misses the torus, the observer sees both the BLR and the NLR, but if the line of sight passes through the torus, the BLR is obscured—the AGN is "buried"—and the observer sees only a narrow-line spectrum from unobscured gas far from nucleus. Indeed, given the proper orientation, an optically thick cloud of dust anywhere in the host galaxy can obscure the AGNs. If, for instance, the line of sight to the AGNs just grazes the torus, the AGN is partially obscured. The broad-line emission and the AGN continuum appear weaker; the latter leading to higher equivalent widths for the narrow lines. Examples of these objects include Seyfert 1.5–1.8 galaxies (Antonucci 1993; Tran 1995; Leighly et al. 1997; Cohen et al. 1999) that have strong narrow-line emission and weak broad-line emission.

Our composite S3 spectrum closely resembles the spectra of Seyfert 1.5–1.8 galaxies. Does obscuration of objects in S3 cause them to have high EW_[O iii]? We first note that the S3 sample definitely includes buried AGNs, possibly many buried AGNs. We will see that the broad-line Hα/Hβ flux ratio is large in some of the S3 AGNs, a clear signal of reddening; a visual inspection of individual spectra in the S3 sample quickly finds classic Seyfert 1.5–1.8 objects; and component S3:PC4 is a broad-line, blue-continuum component devoid of NLR features that is introduced by variable obscuration and reddening. The question, though, is whether the obscuration in the S3 sample is causing higher EW_[O iii]. We first investigate whether the obscuration in the S3 sample as a whole is different from the obscuration in the S1 and S2 samples. There are several tests that might answer this question.

First, if the AGNs in the S3 sample are more obscured than the AGNs in the S1 and S2 samples, they might be fainter than the AGNs in S1 and S2. The mean L_5100 Å of each subset is 1.39 × 10⁴⁴, 1.06 × 10⁴⁴, and 9.68 × 10⁴³ erg s⁻¹ for S1, S2, and S3, respectively. Thus, the mean L_5100 Å of the S3 AGN is, indeed, lower than the mean luminosity of the AGNs in the other sets, but the maximum difference is a factor of just 1.4, much less than the difference in mean [O iii] equivalent widths. This test is not a strong one, given the steep luminosity function of AGNs. We also compared the distribution of L_5100 Å luminosities within the S3 sample to the distribution for the entire sample (see Figure 10). (The S3 objects comprise a small fraction of the entire sample, so a comparison of the S3 sample with the entire sample is essentially a comparison with S3 to S1 plus S2.) A K–S test yields a P = 0.36 probability that the S3 sample comes from the same parent distribution as the entire sample. Thus, there is no evidence that the luminosities of the S3 AGNs are lower than the luminosities of the entire AGN sample by enough to account for the higher narrow-line EWs.

Second, if the AGNs in the S3 sample are more obscured than the entire sample, they should be more reddened and their colors should be different. We defined the color to be the log of the ratio of the median fluxes in 20 Å bins at 3610 Å and 5100 Å that are relatively free of emission lines. We compared the color distribution for S3 to that of the entire data set (see Figure 10). A K–S test finds a P = 0.78 probability that the two distributions are drawn from the same parent sample. Thus, this test finds no evidence for more reddening in the S3 sample than in the entire sample.

Third, because the continuum and the BLR emission lines do not necessarily arise in the same spatial regions, they could be reddened differently. Therefore, we tested whether the behavior of the Hα/Hβ flux ratio, which is a good reddening indicator for the BLR region, is different from the behavior of the continuum colors. We measured the Balmer decrement for the 51 objects in the S3 sample with z < 0.35 so that both Hα and Hβ are visible in their SDSS spectra. We fitted their spectra using the IRAF package specfit between 6400 and 6800 Å, using eight components for the fits: a fixed power law for the continuum, Gaussians for the [S ii] λλ6717, 6730 doublet, the [N ii] λλ6547, 6583 doublet, the narrow component of Hα, and two Gaussians for the broad component of Hα, with the narrow lines all forced to have the same width. The Balmer decrement for the BLR is the ratio of broad Hα flux to broad Hβ flux. We measured the mean Balmer decrement for the 51 objects as 8.311, and it ranged from 3.37 to 78.2. We found a Spearman correlation probability of P < 0.001 between the Balmer decrements and the measured continuum flux ratios, indicating a low probability of no correlation. We conclude that we are justified in using the continuum color as a reddening indicator.

In summary, the tests we applied to the data yielded no evidence for more obscuration in the S3 sample than in the entire sample. The interpretation of the test for a color difference is somewhat muddied by the presence of host galaxy contamination in some low-luminosity objects. For example, one might imagine that the S1 and S2 samples are redder because they have more galaxy contamination, while the S3 objects are redder because they are more obscured, leaving the two groups with the same colors. Thus, colors would not necessarily reveal extra obscuration in the S3 sample. To see whether host-galaxy contamination might be affecting the color test, we looked at the colors of only the most luminous objects (where z > 0.45 and log L_5100 Å > 43.5). Even in this high-luminosity bin, where galaxy contamination is minimal, the mean color of the S3 sample is similar to the mean color of the entire sample (log F₃₆₁₀/F_5100 Å = 0.34 for the S3 sample and 0.35 for the entire sample). Balmer decrements are difficult to measure if Hβ is weak. Since we have restricted our discussion of the Balmer decrement to those systems for which the Balmer decrement is reliably measured, we have excluded exactly those systems with large decrements that might have cast doubt on color as a measure of reddening. On the other hand, the number excluded with unreliable measurements of Hβ is only 3% of the entire sample, so this is not likely to introduce a large bias.

The previous paragraph shows that one must be cautious not to overinterpret the tests for excess obscuration in the S3 sample. We are, nevertheless, left with no convincing evidence for more obscuration in the S3 sample than in the entire sample, and must conclude that the greater EW of the [O iii] lines in the S3 sample cannot be attributed entirely to extra obscuration. Some additional factor is affecting the equivalent widths. Our comparisons of luminosity and color suggest that obscuration and galaxy contamination are similar in S1 and S3; AGNs in S3 may be those in which the strength of [O iii] is sufficient to be further enhanced by obscuration. Nevertheless, we ask what produces this strong [O iii].

Since neither radio loudness nor orientation can alone account for the larger NLR EW in the S3 sample, we now turn to a global comparison of the spectral properties of each subset. One difference is that the EW_Fe ii appears much larger in the S1 and S2 samples than the S3 sample. This is visible in the mean spectra for each sample and also in ALL:PC1, in which Fe ii is anticorrelated with [O iii]. The Fe ii features are extremely weak or non-existent in the mean S3 spectrum and all the S3 principal components. The broad components of the Balmer lines are still present in the S3 sample, so the BLR is still visible on average even though the Fe ii lines are not. For a crude test of the ionization state among the subsets, we used the IRAF routine splot to measure the flux ratio of He ii to narrow Hβ from the mean spectra (Figure 4). We measured the ratio to be 0.432 for S1, 0.309 for S2, and 0.237 for S3. Thus, it appears that there may be a decrease in ionization state with increasing NLR strength.

To investigate other parameters that may shed light on the physics in the objects in S3, we used K–S tests to compare the distributions of the entire data set with those of S3. We compared distributions of Gauss–Hermite coefficients h3 and h4, EW, and FWHM for Hβ and [O iii] (see Figure 10). With the exception of Hβ h4 and the [O iii] FWHM the K–S tests yielded probabilities between 0.35 and 0.99 that the samples are drawn from the same parent distribution (0.04 and 0.10 for Hβ h4 and [O iii] FWHM, respectively). We conclude that with the possible exception of Hβ h4 and [O iii] FWHM the objects in the S3 sample do not differ significantly from the entire sample in any of these properties. A narrower [O iii] FWHM in S3 could imply that the NLR gas is further from the center of the AGN, which would be expected if orientation or covering factor played into the explanation for strong EW_[O iii]. An investigation of the X-ray properties of these objects would be interesting, as X-ray luminosity has been shown to correlate with L_[O iii] (Heckman et al. 2005; Panessa et al. 2006) and α_ox indicates weaker X-ray flux, compared with the optical, for higher EW_[O iii] (BG92), but that is beyond the scope of this work.

Variations in EW_[O iii] can also result from differences in the covering factor, density, or a relatively high ionization parameter. Baskin & Laor (2005), hereafter BL05, investigated the [O iii] line strengths in the PG quasars. Their photoionization models suggested that it is the covering factor that causes differences in [O iii]. If the covering factor is large, the NLR gas intercepts a larger amount of continuum and produces relatively stronger lines.

To determine whether our strong [O iii] objects show any evidence of special physical conditions, such as unusual electron density or ionization parameter, we examined the spectra of the 16 objects in S3 with largest EW_[O iii]. Using Figures 1 and 2 in BL05, and the [O iii] λ4363/λ5007 and [O iii] λ5007/broad Hβ line flux ratios (exhibiting a mean log [O iii] λ4363/λ5007 = −1.62 ±  0.10, and a range of log [O iii] λ5007/broad Hβ from 1.0–1.18), we conclude that physical conditions are typical of AGN NLRs (Peterson 1997). The covering factor then becomes the most likely reason for the large [O iii] EW in the S3 AGN. Assuming that we see the true continuum in the bluest AGN (the presence of Ca ii K absorption and the 4000 Å break in the continuum slope show that host galaxy starlight contributes in the redder AGN, although dust reddening is also probably involved), we deduce that large but physical NLR covering factors of ≳ 0.35 are required. However, we cannot make a convincing choice among the various reasons why the covering factor might be large. There might be more illuminated gas in the NLR of the S3 objects, perhaps as a result of a larger filling factor in the ionization cone, or a larger opening angle. These two scenarios would result in different infrared properties. If S3 objects have a larger opening angle, we would expect these objects to have less radiation absorbed and reprocessed into the infrared. However, if the larger covering factor were due to the larger filling factor of the NLR, the infrared properties of S3 objects should be the same as the rest of the data set.