The [O iii] Profiles of Infrared-selected Active Galactic Nuclei: More Powerful Outflows in the Obscured Population

We assemble obscured and unobscured AGN samples following DiPompeo et al. (2017a), and refer the reader there for complete details. Briefly, we start by selecting AGN candidates from the WISE survey catalog data, utilizing a simple color-cut of W1 − W2 > 0.8 and magnitude cut of W2 < 15.05 (Vega, or 0.16 mJy; Stern et al. 2012), where W1 and W2 represent the 3.6 and 4.5 μm bands, respectively. Following DiPompeo et al. (2016a, 2017a), we only include sources satisfying these criteria in both the "All-sky" ⁵ and "ALLWISE" ⁶ catalogs, in order to provide the cleanest and most conservative sample. Regions of potentially problematic WISE data are excluded, as described in detail in DiPompeo et al. (2017a). These include regions of high Galactic extinction and Moon contamination, flagged WISE data (such as optical ghosts and other artifacts), and poor photometric quality.

These sources are then limited to the Sloan Digital Sky Survey (SDSS; York et al. 2000) photometric footprint, as complementary optical data are needed to split IR-selected AGNs into obscured and unobscured populations. We match to the SDSS eighth data release⁷ (DR8; Aihara et al. 2011) photometry with a radius of 2''. Since we will be utilizing optical spectroscopic measurements (below), sources without optical counterparts are discarded. Removing optical non-detections does not have a strong impact on other obscured AGN measurements (e.g., DiPompeo et al. 2017a), though we cannot be certain it does not impact this work. At the redshifts of our final sample (<0.4, see below), Hickox et al. (2017) identify u − W3 colors as optimal to distinguish between obscured and unobscured AGNs, with a W2 − W3 dependent split of (u − W3 [AB] > 1.4(W1 − W2 [Vega]) + 3.2 used to define obscured sources. We utilize SDSS-provided Galactic extinction corrections and the transformations provided with the WISE catalogs and by Blanton & Roweis (2007) to convert between Vega and AB magnitude systems. We will discuss the impact of other optical−IR color-cuts (e.g., r − W2) on our results in Section 5. A very small number of sources with spectroscopic measurements (below) are undetected in the u-band, and these are considered obscured.

To probe outflows in the NLR of WISE AGNs, we utilize the publicly available AGN Line Profile And Kinematics Archive (ALPAKA⁸ ; Mullaney et al. 2013). This catalog provides fits to optical SDSS spectra covering the [O iii] λλ4959, 5007 doublet for a sample of 24,264 optically selected AGNs (including both spectroscopically classified Type 1 and Type 2 sources). We refer the reader to Mullaney et al. (2013) for full details of this catalog and the spectral fits, but provide a brief summary here.

The catalog starts with all extragalactic spectroscopic sources in the SDSS DR7 catalog (Abazajian et al. 2009) with [O iii] λ5007, [N ii] λ6584, Hα and Hβ emission detected at >3σ confidence in the automated SDSS fitting routine. Requiring these emission lines limits the sample to z < 0.4. A combination of line widths and emission-line diagnostics are then used to identify potential AGNs that are re-fit in a more robust way than the SDSS pipeline. In brief, the fits use multiple Gaussian (narrow and broad) components to fit the SDSS continuum-subtracted (using a fifth-order polyonomial) profiles of the emission lines. In the case of [O iii], both the λ4959 and λ5007 lines are fit simultaneously with two Gaussians each, with their flux ratios fixed at 3 and the velocity offsets and widths of their respective broad and narrow components tied. Sources initially classified as AGNs based on the SDSS pipeline fits, but later classified as non-AGNs with the improved ALPAKA fits, are included in the final ALPAKA catalog. Finally, the catalog includes Type 1/2 classifications based on these emission line fits—sources with a significant (>50% of the line flux) broad (>600 km s⁻¹) component in the Hα line are considered Type 1, and sources without broad lines but satisfying AGN emission-line diagnostics (e.g., Kauffmann et al. 2003) are considered Type 2.

We find 5314 matches between our WISE AGN sample and the ALPAKA catalog (1392 obscured and 3922 unobscured, or 1.7% and 3.1% of the WISE-selected samples with an optical counterpart in the spectroscopic footprint, respectively). The common area for the two samples is ∼7500 deg². A small subset of these matches (∼3% in each subset) do not have reliable fits to the [O iii] lines or enough reliable photometry to perform SED fits to determine independent AGN luminosities (Section 3), and these are discarded, leaving 1366 obscured and 3772 unobscured objects. This obscured/unobscured ratio of 27%/73% in the sources with optical spectra is similar to the overall ratio in the full WISE-selected population once objects without optical counterparts are removed from the obscured sample.

The dotted red and blue lines in Figure 1 show the redshift distributions of the WISE−ALPAKA-matched obscured and unobscured samples, respectively. While the full populations of WISE-selected unobscured and obscured AGNs are generally well-matched in redshift (based on deep spectroscopic fields and angular cross-correlations; Hickox et al. 2011; DiPompeo et al. 2017a), the additional effects imposed by the SDSS spectroscopic selection lead to a redshift mis-match with obscured sources at preferentially lower redshift. Before comparing the properties of the two samples we randomly sample them to have matched redshift distributions.

Zoom In Zoom Out Reset image size

In absolute numbers (rather than the fractions shown in the figure), there are more obscured than unobscured sources up to z ∼ 0.14. Below this value, we randomly downsample the obscured sample to match the unobscured distribution, and retain all of the unobscured sources. At redshifts greater than 0.14, we downsample the unobscured sources to match the obscured distribution, and keep all obscured sources. To account for the additional variance this downsampling produces, we repeat this process 100 times to provide many redshift-matched samples and average over these when calculating population-level properties (e.g., median [O iii] FWHM), and include the standard deviation among these 100 samples in our error analysis. We also select a single representative obscured and unobscured sample with median [O iii] properties consistent with the median of all random subsamples in order to assemble composite spectra (Section 4.1), visualize distributions (e.g., Section 3), and perform additional error analysis (Section 4.2). These include 1042 obscured and 924 unobscured AGNs, and their redshift distributions are shown with solid lines in Figure 1. The median redshift of each is ∼0.17.

Given the two methods of classification available for this sample, a comparison of differences between photometric (obscured and unobscured, split at (u − W3 [AB] = 1.4(W1 − W2 [Vega]) + 3.2) and spectroscopic (Type 1 and Type 2, defined by their optical emission line properties) classifications is in order as there is overlap between them. In the redshift-matched obscured samples, the average fraction of spectroscopic Type 2, spectroscopic Type 1, and non-AGNs is 57%, 39%, 4%, respectively. Similarly, for the unobscured AGN subsamples, average fractions are 4%, 93%, and 3%, respectively. We include objects not spectroscopically classified as AGNs in the ALPAKA catalog (i.e., they satisfied AGN line-width or emission-line diagnostic tests with SDSS pipeline measurements, but not after more detailed line fitting) in this analysis since we are interested in the properties of the WISE-selected populations, but excluding them does not impact the results.

To aid in this comparison, in Figure 2 we show composite spectra (median combined and normalized between 6100 and 6300 Å rest-frame) of all the spectroscopic Type 1/2 sources in the ALPAKA catalog and the WISE-selected obscured/unobscured sources in this analysis (both samples are limited to z < 0.3 to include the entire Hα region in all sources). It is clear that the unobscured sources have slightly narrower Balmer lines compared to Type 1 sources, and vice-versa for the obscured/Type 2 sources, highlighting the mixing introduced in the photometric classification. However, Hβ is clearly dominated by the narrow component in the obscured sources, while Hα retains a significant broad component, though it is narrower than in the unobscured sample. This indicates that there is a significant fraction of the obscured population with only moderate column densities along the line of sight to the broad line region. Because the ALPAKA catalog used Hα to determine AGN types, these "intermediate" or partially obscured sources (i.e., Type 1.8 or 1.9) are grouped into the Type 1 sample while they are more likely to be photometrically classified as obscured. In both classifications, the continua of the unobscured/Type 1 sources is clearly quite blue and "AGN-like," while the obscured/Type 2 sources are dominated by galaxy starlight. Based on this analysis (see also Section 4.3), it is likely that our obscured sample includes a significant number of reddened Type 1 AGNs, a population of considerable interest in the literature (Richards et al. 2003; Glikman et al. 2007, 2013; Urrutia et al. 2009; Ross et al. 2015).

Zoom In Zoom Out Reset image size

While using optical−IR color to identify obscured AGNs at these redshifts includes a significant fraction of broad-line AGNs (spectroscopic Type 1 sources), it is clear that objects classified in this way have at least moderate amounts of nuclear obscuration, based on the shape of the continuum and width of the Hβ line in the composite spectra (Figure 2).

In order to gain qualitative insight into the average properties of the [O iii] emission lines in the obscured and unobscured samples, we first compare composite spectra. Starting with the continuum-subtracted SDSS spectra, we do a simple median combine with no outlier rejection (including various rejection types makes a negligible impact on the results). Since the spectra are already continuum subtracted, no normalization is applied prior to combination. This does lead to more scatter in the peaks of the lines due to the range of luminosities. To generate errors on the composites, we randomly sample (with replacement, such that individual sources can appear more than once in a given subsample) a new set of spectra of the same size as the data set being considered and create a new composite N = 1000 times, and take the standard deviation of these composites at each wavelength.

First, we show the median spectra of the full Hβ/[O iii] complex in Figure 4 (errors here are omitted for clarity). It is clear from the Hβ line that the unobscured sample has a much stronger broad component than the obscured sample, despite the disagreement in photometric and spectroscopic classification for a significant fraction of the objects. There is also a clear broad, blueshifted wing in both of the [O iii] lines, which is broader in the obscured sample. In spectroscopic samples, the Type 2 AGNs tend to have similar or less prominent broad blue wings (Mullaney et al. 2013). However, we note that the difference seen here is possibly due to the difference in median luminosity of the samples, considering the correlation between the broad component of [O iii] and luminosity (below).

Zoom In Zoom Out Reset image size

In order to account for the difference in average luminosity, we break the samples into three L_bol bins, with boundaries defined to split the obscured sample roughly evenly (∼350 objects per bin; see Table 1): log L_bol < 45.3, 45.3 < log L_bol < 45.7, and log L_bol > 45.7 for bolometric luminosities determined from L_15μm. For comparison, we also split the sample based on de-reddened L_{[O iii]} luminosity, and since these values tend to be higher the divisions are at log L_bol < 45.9, 45.9 < log L_bol < 46.15, and log L_bol > 46.15. The composites of just the [O iii] λ5007 line in these bins are shown in Figure 5; the composites using the [O iii]-based bolometric luminosities are similar and so are not shown. It is clear that the broad, blueshifted component of [O iii] gets stronger and wider with luminosity in both samples, in agreement with previous work (see Section 1) and suggestive of radiatively driven winds in both obscured and unobscured WISE AGNs. However, it is also apparent that in all luminosity bins, regardless of how the bolometric luminosity is determined, the obscured sources have wider broad components, potentially more blueshifted (though the lines are qualitatively more symmetric in the obscured sample), indicative of stronger outflows.

Zoom In Zoom Out Reset image size

Table 1.  [O iii] Properties Binned by L_bol

	Obscured						Unobscured
	N	$\langle \mathrm{log}{L}_{\mathrm{bol}}\rangle$	$\langle \mathrm{FWHM}\rangle$	FWHMc	$\langle v\rangle$	vc	N	$\langle \mathrm{log}{L}_{\mathrm{bol}}\rangle$	$\langle \mathrm{FWHM}\rangle$	FWHMc	$\langle v\rangle$	vc
		(erg s−1)	(km s−1)	(km s−1)	(km s−1)	(km s−1)		(erg s−1)	(km s−1)	(km s−1)	(km s−1)	(km s−1)
15 μm-based Lbol
Lbol < 45.3	351	45.1	643	822	67	128	456	45.0	486	593	43	78
45.3 < Lbol < 45.7	359	45.5	735	907	84	138	287	45.5	617	679	79	119
Lbol > 45.7	332	46.9	889	1058	98	97	181	46.9	654	773	103	133
[O iii]-based Lbol
Lbol < 45.9	333	45.7	688	836	75	106	586	45.6	496	639	44	79
45.9 < Lbol < 46.15	354	46.0	740	829	87	120	191	46.0	647	711	114	149
Lbol > 46.15	355	46.3	805	998	105	81	147	46.4	692	787	109	94

Note. The quantities in brackets represent sample medians, rather than means. The "c" subscript denotes values from fits to the composite spectra, following a similar prescription as Mullaney et al. (2013). The velocity offsets v are given as positive for simplicity, as they are all blueshifted. Typical errors on the composites are ∼50 km s⁻¹ for the FWHM and ∼20 km s⁻¹ for the velocity offset. The median values and their errors are shown in Figure 6.

Download table as:  ASCIITypeset image

To provide some quantitative comparison of the composite spectra, we also perform fits to the [O iii] λ4959 and λ5007 lines in the composites, following a similar methodology as Mullaney et al. (2013) using the mpfitfun IDL routine. The results of these fits are listed in Table 1 (values from composite fits are listed with a "c" subscript). The top half shows results based on SED-derived L_bol, and the bottom half shows the results using dereddened [O iii]-based L_bol. These results verify that the obscured sample has both wider (indicated by the FWHM) and more blueshifted (indicated by the velocity offset v, which we always list as positive since all are in the same direction) broad components. The exception to this is the highest-luminosity bin where the composite obscured [O iii] profile appears more symmetric (while remaining wider), leading to lower blueshift values. However, in the next section we show that this is not true when considering the distributions of sources in each bin, suggesting that the composite is being affected by outliers in the wings of the line.

While the composite spectra are useful for qualitative comparisons, the precise details of creating them can complicate quantitative comparisons. We therefore focus the remainder of this work on the distributions of [O iii] parameters in each sample, utilizing the same bins in luminosity as in Section 4.1. In cases where no broad component was necessary in fitting the lines (Mullaney et al. 2013), we used the properties of the narrow [O iii] lines as limits on the broad components, though simply neglecting these sources does not change the general results or conclusions.

In calculating the median of any property, we first calculate the median of each of the 100 obscured and unobscured redshift-matched subsamples, and then adopt the median of these as the final value. Errors on median values are generated with a conservative bootstrapping method. We randomly resample each quantity under consideration for each object in accordance with its errors in the individual ALPAKA fits, while simultaneously selecting a random subset of objects (with replacement) of the same size as the initial set N = 10,000 times, recalculating the median for each resample. The standard deviation of the resulting distributions of medians is taken as the 1-σ error on the median. To account for variance in median properties introduced by resampling to match in redshift, we calculate the standard deviation of the medians of each of the 100 subsamples and add this in quadrature to the bootstrapped errors.

In Figure 6 we show both the median broad [O iii] FWHM (top panel) and broad [O iii] blueshift (v, bottom panel; note that these values are always given as positive, as they are all in the same direction) in each bin of L_bol. Points are placed along the x-axes at the median L_bol of the sample in each bin, which because of the differences in the overall distributions (Figure 3) are not always identical. We also include points for the full ALPAKA AGN spectroscopically classified Type 1 and 2 samples for comparison. The median values of the FWHM, v, and L_bol are all listed in Table 1. The table also includes results using [O iii]-based L_bol bins, but these are not shown in the figure as they are qualitatively similar.

Zoom In Zoom Out Reset image size

It is immediately clear that both samples have increasing velocity widths and blueshifts with luminosity, and the obscured sample has consistently higher velocities in all bins. We also note that, in all bins, the obscured sample has a higher fraction of objects that require a broad component to improve the fit than the unobscured sample, by ∼10%, further evidence of an enhanced broad component in the obscured population.

In order to quantify the differences between the WISE obscured and unobscured samples, we parameterize the FWHM–L_bol and v–L_bol relationships with simple power laws that generally provide good fits to the data:

$$\begin{eqnarray}&&{\rm{FWHM}}={{AL}}_{{\rm{bol}}}^{\alpha },\end{eqnarray} \tag{ 2 }$$

$$\begin{eqnarray}&&v={{BL}}_{{\rm{bol}}}^{\beta }.\end{eqnarray} \tag{ 3 }$$

The best-fit parameters (using simple χ² minimization) are listed in Table 2. The error ranges of the fits are produced by randomly resampling each point in accordance with its errors and re-performing the fit 1000 times. Because the fit parameters are highly covariant, we do not list errors on the individual parameters in the table. It is clear that the two samples have different normalizations and similar slopes (for the SED-derived luminosities).

Given the low redshifts and quality photometry and spectra for this data set, our SED fitting allows us to explore the [O iii] outflow properties simply as a function of AGN reddening rather than enforcing a hard split between obscured and unobscured populations. We note that this is not possible for example for the full WISE AGN population, which is why the bulk of the analysis thus far has focused on cuts that are more broadly applicable. The optical−IR SED modeling includes the E(B − V) of the AGN component, and as expected the reddening in the obscured sample is, on average, larger than the unobscured sample (with medians of 0.74 and 0.03, respectively). While there is no strict definition, if we define a "red AGN" as having 0.1 < E(B − V) < 3 (e.g., Glikman et al. 2007), we find that 38% of our total sample meets this definition, or 41% of the obscured sample and 37% of the unobscured sample. This fraction is consistent with the number of broad-lined AGNs in the obscured sample.

In Figure 7, we show the broad [O iii] FWHM and blueshift in logarithmically even spaced E(B − V) bins. The top panel shows the median bolometric luminosity in each bin, highlighting that there is no significant trend in luminosity with reddening that could impact any trend with FWHM or blueshift. It is clear that both the broad [O iii] FWHM and blueshift increase from zero to moderate reddening values, mimicking the difference found earlier between obscured and unobscured sources. However, around E(B − V) = 0.5, this trend flattens with FWHM and reverses with blueshift. We note that the reddening values in the range 0.1 to 1 where this transition occurs are typical of what is found in red quasars at higher redshift, which may be in a specific evolutionary phase (e.g., Urrutia et al. 2008; Glikman et al. 2015). We will discuss the implications of this further in Section 5.3.

Zoom In Zoom Out Reset image size

Several studies (including Mullaney et al. 2013, whose spectral fits make up the basis of much of this work) have found that outflow velocity is correlated with radio power, sometimes more so than bolometric luminosity, suggesting that radio jets are the primary driver of large-scale outflows. In our sample, there is a significant fraction of sources with radio counterparts, and these are preferentially in the obscured population—10% of the unobscured sources have counterparts at 1.4 GHz in either Faint Images of the Sky at Twenty-Centimeters (FIRST, ∼0.6 mJy 5σ detection limit; Becker et al. 1995) or the NRAO VLA Sky Survey (NVSS, ∼2.5 mJy 5σ detection limit; Condon et al. 1998), while 40% of the obscured sources have a radio counterpart (Mullaney et al. 2013). The vast majority of detections in both samples are in FIRST. This discrepancy between the two samples is at least partly due to biases imposed by the SDSS spectroscopic selection algorithms, in particular that the obscured sources are more likely to be selected for spectroscopic follow up based on radio emission rather than, e.g., optical properties (e.g., Zakamska et al. 2004; Hickox et al. 2017). We perform two checks on the impacts of radio jets on our results.

The first is to simply remove all radio-detected sources from the samples and repeat the analysis, though we note that this can have other effects on the sample that may impact outflow properties, such as changing the black hole mass and Eddington ratio distributions. In doing so, the results are not significantly changed, though the median FWHM and blueshift in each sample and L_bol bin are reduced slightly. For example, the obscured FWHM from lowest to highest luminosity becomes 614, 695, and 848 km s⁻¹, a reduction of ∼5% in each bin. The change is very similar in the unobscured sample, and the general trend is still preserved.

We also apply the same analysis using bins in radio luminosity instead of L_bol, where we subsample from each bin such that the median bolometric luminosity in each is within 0.25 dex of the others (all at log L_bol ∼ 45.2 erg s⁻¹). In this case, we find that the unobscured FWHM–L_1.4GHz relationship is roughly flat, while for the obscured sample it increases slightly by ∼150 km s⁻¹ from low to high radio luminosity. However, the obscured FWHM is generally smaller than, or consistent with, that of the unobscured sample, rather than consistently larger. We note that controlling for L_bol in each bin does reduce the sample size significantly, particularly for the highest radio luminosity bin, and the errors are increased accordingly. The tests performed here imply that, while radio jets may play a role in the outflow properties of these samples, they are not the dominant factor, and radiation pressure is needed to produce the enhanced velocities in the obscured sample.

While we have matched our samples in luminosity bins, it is possible that other key parameters tied to either the luminosity indicators or the outflows are not matched, such as black hole mass or the Eddington ratio λ_Edd. Recently it was shown that the obscured fraction, and hence the nuclear dust covering factor, is a function of λ_Edd in nearby X-ray AGNs, with higher Eddington ratio sources less likely to be obscured (Ricci et al. 2017). This could imply that we are overestimating the BC for the obscured population. However, in this case applying a smaller BC to the obscured sample implies shifting the relationship in Figure 6 to the left, with higher velocities at lower luminosity, serving to enhance the difference between obscured and unobscured samples further. A similar argument could be used in the case of a luminosity-dependent BC, as might be expected in receding torus models (e.g., Lawrence 1991).

In addition, when considering radiation pressure on dusty gas, the boost factor (or ratio of effective cross-section to the Thompson cross section) depends on both the column density and the ionization parameter. Since the ionizing flux changes with λ_Edd, so too does the boost factor (e.g., Fabian et al. 2008). If the obscured sources are preferentially lower Eddington, this would manifest as a lower boost factor, serving to reduce the outflow velocity (all else being equal).

Finally, it is also possible that we underestimate the intrinsic IR luminosity of the obscured sample relative to the unobscured. This could be due to anisotropic emission (e.g., Hönig et al. 2011), viewing angle differences, or even extinction in the IR in the most heavily obscured sources—though the strong, broad Hα emission in the obscured composite implies that only a minority of sources are deeply buried in this sample. However, DiPompeo et al. (2017a) found that these effects do not seem prevalent in WISE-selected AGNs, and in order to completely explain the results here would require us to underestimate the average obscured luminosity by nearly an order of magnitude, while the effects above would impact the luminosities at most by a factor of a few.

Because we are working with both averages over populations as well as spatially averaged spectra, modeling the differences we find between the two samples in a detailed way is extremely difficult (e.g., Zakamska & Greene 2014). The optical SDSS spectra are from a fiber-fed spectrograph with a fixed 3 arcsec fiber size, which means that the spectra are spatially averaged over different physical scales. At the median redshift of 0.17, the spectroscopic fibers reach radii of ∼4 kpc.

There are many models of radiation-pressure-driven winds in AGNs, and the general components are the AGN luminosity, the wind launch radius, the effective cross section (which increases in dusty gas), and the mass profile of the host galaxy (e.g., Equation  (7) of Fischer et al. 2017). These parameters can be considered broadly in the context of our measurements here.

Since dust enhances the ability of radiation to accelerate a wind due to increased cross section, an angular dependance of the dust content of the wind could lead to orientation affecting the observed velocities. If the wind is conical in shape but more dusty along the outer edge than in the center, for example due to dust ablated off the edge of the torus (similar to, e.g., the two-walled bi-cone of Nevin et al. 2017), we could have a scenario in which the outer edge of the conical outflow is moving at higher velocity. If the obscured sources are then seen from preferentially large viewing angles, nearer to the edge of the cone, this could lead to an enhanced asymmetric velocity component. However, winds launched from the nuclear regions in AGNs can have much higher velocities than what we observe here, for example >10⁴ km s⁻¹ in broad absorption line quasars.

One simple scenario is that the wind launching radius is different for the obscured and unobscured samples, as gas launched from closer to the nucleus will begin with a higher velocity. In general evolutionary scenarios for obscured AGNs that involve a "blowout" phase in which the AGN is clearing its immediate surroundings, the outflowing gas distribution may be more compact and lead to a higher velocity at a given luminosity. Of course, in such a scenario it is expected that the dust content of this cocoon of material is much higher, which can enhance the radiation pressure cross section but also result in more "dead weight" behind the material being accelerated (Fabian et al. 2008).

Guided by Figure 7, we propose that the need to balance these effects results in the enhanced obscured velocities being dominated by sources with intermediate column densities, explaining why this velocity difference has not been seen in spectroscopically classified samples. In this scenario, not all obscured AGNs are in the blowout phase, and the relative numbers of obscured AGNs that are spectroscopically classified as Type 1 and Type 2 would correspond to the relative lifetime of such a phase. Based on these fractions discussed Section 2.4, the blowout phase would correspond to approximately half of the quasar lifetime, roughly consistent with some models and observations (e.g., Hopkins et al. 2008; DiPompeo et al. 2017a). If we repeat our analysis using sources in the obscured sample that are also classified spectroscopically as Type 2, the difference in the relationship between luminosity and outflow velocity compared to the unobscured sample is reduced. Conversely, if we limit the obscured sample to only spectroscopic Type 1 sources, the difference is enhanced. This suggests that at least a subset of the obscured population in this work are linked to red AGNs, which may be a distinct population tied to certain evolutionary phases such as mergers (e.g., Urrutia et al. 2008; Glikman et al. 2015; Hamann et al. 2017), though we note that we are probing different redshift ranges and luminosities.

At low redshifts such as these, other color cuts are less efficient at separating obscured and unobscured AGNs (Hickox et al. 2017). If we instead use an optical−IR color cut of r − W2 = 6 to define our obscured and unobscured sources (Hickox et al. 2007), which is commonly used for the full WISE AGN population which peaks at z ∼ 1, we indeed find more mixing of spectroscopic types within the photometrically defined samples. Roughly 50% of the obscured sources have detected broad lines, and 30% of the unobscured sample are Type 2 sources. This additional mixing of reddened Type 1 sources into the obscured sample does enhance the median outflow velocities, as expected in the scenario described above.