The Macroeconomic Effects of Government Asset Purchases: Evidence from Postwar Us Housing Credit Policy

We document the portfolio activity of federal housing agencies and provide evidence on its impact on mortgage markets and the economy. Through a narrative analysis, we identify historical policy changes leading to expansions or contractions in agency mortgage holdings. Based on those regulatory events that we classify as unrelated to short-run cyclical or credit market shocks, we find that an increase in mortgage purchases by the agencies boosts mortgage lending, in particular refinancing, and lowers mortgage rates. Agency purchases influence prices in other asset markets, stimulate residential investment and expand homeownership. Using information in GSE stock prices to construct an alternative instrument for agency purchasing activity yields very similar results as our benchmark narrative identification approach.


Introduction
The residential mortgage market in the United States is one of the largest capital markets in the world and by far the dominant source of credit for American households. The mortgage market finances housing, which is a key component of both household wealth and aggregate spending, see e.g. Leamer (2007). Many accounts of the causes and propagating factors of the 2007/08 financial crisis assign an important role to a boom and bust in the availability of mortgage credit. 1 The US mortgage market is also subject to heavy government involvement through various federal agencies, including the housing government-sponsored enterprises (GSEs). In the decades preceding the 2007/08 crisis, the various agencies collectively accumulated a large share of the total outstanding US mortgage debt on their balance sheets. In this paper, we investigate whether agency portfolio purchases of mortgage assets influence the availability and cost of housing credit, and whether there are spillovers to other debt markets and economic activity more broadly.
While the history of agency activity offers a rich source of variation to study the effects of government asset purchases, it also presents a number of challenges. The largest agencies, Fannie Mae and Freddie Mac, have been privately owned for much of their existence and therefore carry responsibilities to stock owners as well as to their public missions of providing "stability" and "ongoing assistance" in mortgage markets. Both profit and public objectives cause these agencies to systematically and rapidly respond to market conditions, such that changes in their mortgage purchasing activity reflect changes in housing credit demand and many other influences. Some of the correlation between agency balance sheets on the one hand and credit growth or mortgage rates on the other is therefore likely to reflect reverse causality.
We propose two different strategies to isolate changes in agency purchasing activity free of confounding influences. Our first and principal strategy is to focus on historical credit policy interventions affecting agency mortgage holdings, in the spirit of the approaches in Romer (1989, 2010) and Ramey (2011) to studying monetary and fiscal policy. Based on a narrative analysis of the regulatory history of the housing agencies, we identify and quantify significant policy events affecting agency purchases. These include adjustments to capital requirements, portfolio caps, or statutory borrowing authority, direct appropriations and 1 See e.g. Mian and Sufi (2009), Justiniano, Primiceri, and Tambalotti (2014), or Di . capital injections by the Treasury, or changes to the pool of mortgages eligible for agency purchase, such as changes in conforming loan limits or authorizations to enter new mortgage market segments.
Credit policy changes are often reactions to cyclical conditions in mortgage and housing markets, the recent crisis being a prime example. However, many interventions are motivated by other longer run objectives such as increasing homeownership. Based on an extensive analysis of historical sources, we classify each significant credit policy change as motivated by either cyclical considerations or by other non-cyclical objectives. 2 This results in an indicator summarizing the non-cyclically motivated policy events, which we use as an instrumental variable in regressions of a variety of outcome variables on measures of agency purchasing activity. Similar to the approach in Ramey and Zubairy (2016) to estimating government spending multipliers, we estimate the cumulative effects of an increase in agency purchases on mortgage credit and originations, as well as impulse responses to news shocks about future agency purchasing activity.
Our second and complementary identification approach is based on instrumenting measures of agency purchasing activity with orthogonalized innovations in Fannie and Freddie excess stock returns. This alternative strategy is analogous to Fisher and Peters (2010), who use excess return innovations in major US defense stocks as a measure of news shocks to military spending. Passmore (2005) estimates that the advantages granted by federal housing credit policy account for much of the market value and portfolio size of Fannie and Freddie. We show that news about policy interventions affecting GSE balance sheets is reflected in their stock market valuation. Positioned last in a causal ordering behind credit aggregates, interest rates, and other macro variables, we find that residual variation in Fannie and Freddie excess stock returns predicts agency mortgage purchases. This motivates us to use this residual variation as an alternative instrumental variable to estimate the response of credit aggregates to shocks to agency mortgage purchases.
It is not clear ex ante that government purchases of mortgage assets have meaningful effects on the cost and availability of housing credit. If financial market frictions are relatively unimportant, an increase in agency purchases may have little impact on the volume of mortgage credit, and simply lead to crowding out of private holdings. If such frictions are instead pervasive, mortgage market policies may on the other hand be very important for the provision of credit to residential borrowers. Based on our methodology, we find that agency purchases indeed lead to statistically significant expansions in mortgage credit. Our estimates indicate that each additional dollar in agency mortgage purchases leads to a 3 to 4 dollar cumulative increase in mortgage originations over the course of three to four years, and a net expansion in the stock of mortgage debt of around one dollar. The rise in originations is largely driven by an increase in refinancing activity, but is also followed by a greater volume of originations financing home purchases. The expansionary effects on housing credit are accompanied by temporary reductions in mortgage interest rates, which fall by 10 to 15 basis points for more than a year following an increase in agency purchases of one percent of trend originations.
Agency purchases also affect prices in other asset markets. We estimate that the 10-year Treasury rate and the 3-month T-bill rate both decline when the agencies increase their purchases of mortgages. A key policy objective behind President Hoover's introduction of housing credit policies in the 1930s was to stimulate the construction sector, while another recurrent motivation has been to promote homeownership. We find evidence that supports these roles of the agencies in that new housing starts and homeownership rates rise following an increase in agency mortgage purchases. We also find some evidence that agency mortgage purchases increase house prices and stimulate private sector consumption. There is no clear evidence of any significant impact on the unemployment rate or personal income.
Perhaps our most surprising finding concerns the relationship between housing credit and monetary policies. We show that the narratively identified housing credit policy shocks have forecasting power for the residual shock component of the  decomposition of funds rate target changes, while the reverse is not true. Instead, we find that cyclically motivated housing credit policy changes lean against the wind of contractionary monetary disturbances. Housing credit policy shocks have larger effects on refinancing originations than interest rate shocks, and influence homeownership independent of shortterm interest rates. The quantitative effects of housing credit policy and conventional monetary shocks are very similar along many other dimensions. These findings suggest that both may share similar transmission channels, and that the interplay between monetary and credit policy deserves more attention.

Mortgage Purchases as Credit Policy in the United States
The US government intervenes in the mortgage market in many ways. We focus attention on the federal involvement in purchasing residential mortgages. The first significant use of this type of policy dates back to the Great Depression. The sharp and sustained downturn in credit markets motivated Congress to create the Home Owners' Loan Corporation in 1933. Financed by bonds, the Corporation purchased delinquent mortgages from lenders and refinanced these mortgages into fully amortizing fixed-rate loans with long maturities to lower monthly payments for distressed mortgagors. In 1938, Congress created Fannie Mae to support a secondary market for government-guaranteed mortgages. Fannie's authority to acquire mortgage debt was increased greatly after WWII to support the construction sector and promote homeownership among veterans. The late 1960s saw the creation of Ginnie Mae to provide continued support the market for government-guaranteed mortgages. In 1970, Fannie Mae obtained permission to enter the conventional market, i.e. the market for loans not directly guaranteed or insured by the government, and the newly created Freddie Mac joined Fannie Mae in developing a nationwide secondary market for conventional mortgages.
Over time, the agencies have played an increasingly active role. The two largest GSEs, Fannie and Freddie, acquire mortgages through advance commitments to buy loans from mortgage lenders, which are delivered once the loans are originated in the primary market. 3 Until the late 1960s, the purchases by Fannie were financed predominantly by borrowing from the Treasury. Afterwards, as quasi-private entities, Fannie and Freddie have financed these purchases with a mix of private capital and debt issued in capital markets. A third financing option is the issuance of mortgage pools, i.e. mortgage-backed securities (MBS). Securitization was brought to the conventional market by Freddie Mac in the early 1970s, and took off in the 1980s when it was also adopted by Fannie Mae. Mortgage securitization has consistently been GSE-dominated, perhaps with the brief exception of the 2004-2006 private-label securitization boom. In the process of packaging whole mortgages into securities, the agencies also assume the credit risk in return for guarantee fees.
From the early 1990s onwards, the agencies increasingly retained their own and acquired each other's MBS, as opposed to selling them to private investors. a quarter of all residential mortgage debt resided on the balance sheet of a federal agency, with roughly 20 percent owned by Fannie and Freddie alone. In early September 2008, Fannie and Freddie were taken into conservatorship and were required to gradually wind-down their balance sheets by two-thirds. The Federal 4 Other agencies include the Home Owners' Loan Corporation, Treasury, Veterans Administration, Federal Housing Administration, Federal Farmers Home Administration, Resolution Trust Corporation, Federal Deposit Insurance Corporation, and Public Housing Administration. We do not include mortgages in government pension funds. See the data appendix for sources. 5 Because purchases may include loans originated in prior periods, the market shares may occasionally exceed 100 percent. 6 Net additions to the stock of mortgage debt are considerably smaller than originations since both existing home sales as well as refinancing transactions typically lead to minor net changes in mortgage debt.
Reserve subsequently pursued several rounds of large-scale purchases of agency MBS under its quantitative easing (QE) programs, and its current holdings amount to roughly 15 percent of total mortgage debt outstanding. For readers wishing more information about the institutional history of the housing agencies, the appendix provides more background.
The focus of this paper is on the portfolio purchases of the housing agencies, shown in blue in the lower right panel in Figure 1. Prior to the Fed's QE programs, Fannie and Freddie accounted for the bulk of agency mortgage acquisitions. Even as privately owned corporations, Fannie and Freddie have been key agents of federal housing policy and differ from traditional financial intermediaries in a number of important ways. First, they have always maintained authorization to borrow from the Treasury. While this authorization was limited and never formally exercised, it sufficed to create the widely held belief that the US government would never allow a GSE to default. This perception, eventually justified by the government takeover of Fannie and Freddie in 2008, meant that interest rates on agency bonds have typically been close to Treasury rates. Second, agency debt is eligible for open market operations by the Fed. In the 1960s and 1970s the Fed made significant purchases of agency debt, see Haltum and Sharp (2014), and again so under the QE programs. Third, the prudential supervision of the GSEs is separate from private banks and, prior to 2008, resided within the Department of Housing and Urban Development (HUD). 7 Regulatory oversight of the GSEs was traditionally light compared to that of private banks, and the GSEs generally enjoyed much less stringent capital and reporting requirements. For instance, despite being publicly listed companies, Fannie and Freddie were exempt from filing with the Security and Exchange Commission until the early 2000s.
Finally, for much of their existence, the GSEs have also benefitted from various preferential tax treatments.
In exchange for the privileges granted by federal law, the GSEs face a number of restrictions and obligations. Fannie and Freddie cannot originate loans in the primary market and are not allowed to diversify portfolio holdings much beyond mortgage assets. Their purchases are limited to conforming mortgages that must meet certain underwriting standards, and the principal on the loans cannot exceed a maximum amount, known as the conforming loan limit. The authority for adjusting the limit and other loan characteristics that 7 Since 2008, the regulatory authority lies with the Federal Housing Finance Agency, an independent federal agency. determine what mortgages are conforming has generally lied with Congress and the HUD Secretary. In 1980 the conforming loan limit became indexed to a house price index maintained by Freddie Mac. Since then typically around 80 percent of mortgages have been conforming. 8 Finally, the GSEs are expected to balance stock owner interests with certain public policy objectives, including the stabilization and enhancement of mortgage markets, as well as assistance with the provision of credit to lower-income households.

Related Literature
There are relatively few attempts at identifying the dynamic effects of agency purchases on mortgage credit, residential investment or homeownership. An early literature estimates reduced form models of credit and housing markets to assess the impact of GSE activity in the 1970s, e.g. Arcelus andMeltzer (1973), Meltzer (1974), Villani (1977, 1980), Jaffee and Rosen (1978), and Kaufman (1985). Although no clear consensus emerges from this early work, Smith, Rosen, and Fallis (1988) conclude that an additional dollar in government lending increases mortgage debt by 25 to 35 cents after three to four quarters. Arcelus and Meltzer (1973) and Meltzer (1974), however, argue there is no effect on residential investment or home purchases, while Jaffee and Rosen (1978) and Villani (1977, 1980) find a positive impact of agency activity on home construction.
Starting with Hendershott and Shilling (1989), a number of studies document significant interest rate spreads between conforming and jumbo loans, which suggests that the GSEs affect the cost of mortgage credit. Hendershott and Shilling (1989) attribute this result to a credit supply channel operating through agency securitization. A number of studies investigate the time series relationship between GSE activity and credit costs. Naranjo and Toevs (2002), for instance, find a negative long-run relationship between GSE purchases and mortgage rates, while González-Rivera (2001) finds only a negative short-run relationship. 9 Lehnert, 8 In response to the financial crisis, the limit was increased substantially for the financing of homes in urban areas, which further expanded the pool of mortgage debt eligible for GSE purchase. 9 Naranjo and Toevs (2002), who use vector error-correction (VEC) and GARCH (generalized autoregressive conditional heteroskedastic) models and monthly time series data from 1986 to 1998, find that both GSE purchases and securitization reduce conforming mortgage spreads and volatility, while documenting some spill over to reductions in non-conforming loans, which they attribute to investor substitution effects. González-Rivera (2001), who uses VEC models and monthly data from 1994 to 1999, finds a negative short-run relationship of GSE purchases responding to widening secondary mortgage market spreads, and some evidence of a pass through from secondary to primary mortgage rates from agency purchases. Passmore, and Sherlund (2008) study the impact of GSE activities on primary and secondary market mortgage spreads using both generalized impulse response analysis and causal orderings in VAR models. Based on monthly data from 1993 to 2005, these authors find little evidence that higher GSE purchases impact mortgage spreads, which is consistent with the Meltzer view that credit market interventions are neutral. In a May 2005 speech, Federal Reserve Chairman Alan Greenspan conveys a similar view of the role of the GSEs' portfolio activities, stating that "Fannie's and Freddie's purchases... with their market-subsidized debt do not contribute usefully to mortgage market liquidity, to the enhancement of capital markets in the United States, or to the lowering of mortgage rates for homeowners" (Greenspan, 2005).
In this paper, we contribute new evidence against the Greenspan-Meltzer view that agency mortgage purchase have little effect on the cost and availability of mortgage credit. Our approach is similar in spirit to Lehnert et al. (2008), but adopts novel and arguably better identification strategies to control for the endogeneity of agency purchases. We also study a much longer time frame than any of the earlier papers, and we estimate the effects on both credit aggregates and mortgage rates. Moreover, our dynamic regressions allow us to study many other variables of interest, including housing starts, home prices, homeownership rates, cyclical indicators, and various other interest rates and credit spreads.
Our paper is related to the many analyses of the large-scale MBS purchases by the Federal Reserve under the QE programs. To isolate the effects of these purchases, the literature typically restricts attention to high frequency financial data, and most studies conclude that the MBS purchases lowered secondary market mortgage yields on impact, see e.g. Gagnon et al. (2011), Krishnamurthy and Vissing-Jørgensen (2011), Patrabansh, Doerner, and Asin (2014, and Passmore (2011, 2015). 10 Exploiting cross-sectional variation, a few recent studies also uncover evidence that is suggestive of a positive impact on mortgage lending. Di Maggio, Kermani, and Palmer (2016), for instance, find that, after the first QE intervention, originations of mortgages qualifying for inclusion in securities eligible for purchase by the Fed increased substantially more than those of non-qualifying mortgages. No such differential effects are evident after the second QE intervention, which did not include MBS purchases. Darmouni and Rodnyansky (2016) 10 Stroebel and Taylor (2012) instead find no effects of the MBS purchases under QE1. find that banks with larger mortgage positions increased lending relative to banks with smaller positions, and Chakraborty, Goldstein, and MacKinlay (2016) show that banks with MBS exposure increased their mortgage origination share relative to other banks. By studying a longer history of housing credit policy interventions, we are able to circumvent some key limitations of the event studies of the Fed's large-scale MBS purchases. Our approach permits an analysis beyond the very short-run response of financial variables, and unlike the cross-sectional studies, provides direct evidence on aggregate rather than relative effects.
Our study also fits in a broader empirical literature that aims to identify credit supply shocks and estimate their aggregate effects. Peek, Rosengren, and Tootell (2003), for instance, use bank health indicators as proxies for loan supply shocks and find substantial effects on inventory investment and other aggregates. Gilchrist and Zakrajsek (2012) look at innovations in corporate bond spreads after removing cyclical default premia, and demonstrate their strong predictive content for macroeconomic fluctuations. Bassett, Chosak, Driscoll and Zakrajsek (2014) study residual variation in survey measures of bank lending standards and find an impact on economic activity. Mian, Sufi, and Verner (2017) use variation in the timing of bank branching deregulation in the 1980s to construct differential state-level credit supply shocks, and find that these shocks impact household borrowing and employment. Both our narrative policy indicator and the GSE excess return shocks can similarly be viewed as proxies for credit supply shocks in the mortgage market.
Many existing theories of financial frictions can explain the non-neutrality of agency mortgage purchases. Krishnamurthy andVissing-Jørgensen (2011) and, among others, discuss a variety of potential transmission channels associated with the MBS purchases under the QE programs. Many of these channels have similar implications for mortgage purchases by the GSEs. Through the portfolio rebalancing channel, for instance, private investors bid up the price of mortgages when rebalancing assets towards some desired composition of mortgages and agency liabilities. For the GSEs, the latter are not reserves, but debt instruments that closely substitute for Treasuries in terms of liquidity and (perceived) safety. 11 Depending on the level of segmentation in financial markets, rebalancing effects may spill over to other asset markets, in which case yields on mortgage substitutes-particularly other types of long-term debt-may fall as well. 11 This difference may be less important if the Federal Reserve simultaneously acquires agency debt. Agency mortgage purchases also matter when private mortgage lenders face capital constraints because of regulations or binding incentive constraints, for instance as in the theoretical models of Gertler and Kiyotaki (2010) or Cúrdia and Woodford (2011). Because the GSEs are more highly leveraged than private lenders, aggregate lending capacity increases with agency market share. Agency purchases that drive up the price of mortgages may additionally improve the net worth position of private mortgage lenders, while the exchange of mortgages for agency debt lowers their risk-weighted leverage ratios. Increased agency activity in the secondary mortgage market may also reduce liquidity premia. Our findings support a role for credit supply channels in determining household debt, homeownership, and residential investment, but it is beyond the scope of this paper to isolate precisely which of these channels may be more important.

Endogeneity Problems
To assess the impact of agency portfolio purchases, one might be tempted to simply correlate measures of agency activity, such as those in Figure 1, with credit and other macroeconomic aggregates. This would, however, ignore various endogeneity problems. For one, the agencies respond to changes in market conditions. To maintain market share, for instance, the GSEs vary purchases with the supply of mortgages into the secondary market, which in turn depends on fluctuations in the housing market and the economy. A different endogeneity concern is that agency purchases typically expand relative to the mortgage market when credit is tight and/or conditions in the housing market are deteriorating. This was evidently the case during the latest financial crisis through the actions of the Fed and Treasury, but is also true of earlier episodes. Figure 2 shows the average real levels of agency and private holdings of mortgage debt over the course of business and credit cycles since the mid-1950s. The left panel of Figure 2 shows the average real levels of agency and privately held mortgage debt centered around NBER business cycle peaks. On average, growth in agency holdings is high relative to growth in private holdings prior to a business cycle peak. The growth in private mortgage holdings slows down just prior to the peak and remains low for a prolonged period after the start of a recession. The pace of growth of agency holdings, in contrast, remains roughly unchanged for at least two years after the beginning of an economic downturn.
The right panel of Figure 2 shows the average real levels of mortgage holdings centered around the peak of credit cycles, defined as the quarter preceding the start of credit crisis episodes based on the datings in Eckstein and Sinai (1986) and subsequent updates. 12 Agency and private holdings grow at roughly similar rates prior to a credit crunch. Growth in private holdings of mortgage debt slows markedly following the start of a credit crisis. In contrast, growth in agency holdings accelerates at the onset of a credit crunch and remains elevated for about ten quarters, before flattening toward the pre-crunch trend.
The evidence thus indicates that agencies tend to increase their share of the market in cyclical downturns and credit crunches. These countercyclical purchase dynamics are robust to omitting the 2007/08 crisis and the Federal Reserve's interventions. There are a number of reasons why the agencies jointly maintain or expand purchases during cyclical downturns. A public mission to provide stability to mortgage markets is mandated in the GSEs' statutory charters. Credit crises also offer particularly profitable opportunities for the GSEs because their lending spreads widen relative to private intermediaries, due to countercyclical mortgage spreads and the implicit guarantee provided by the US government. Finally, the federal government often undertakes deliberate regulatory or legislative actions to further enable agency expansions during downturns. The fact that agency purchases tend to accelerate when mortgage spreads are elevated and/or credit is tight induces a negative relationship with mortgage credit aggregates. This negative association needs to be accounted for in order to determine the causal effects of agency mortgage purchases.

Narrative Analysis of Policy Changes Affecting Agency Mortgage Holdings
Our principal strategy to control for reverse causality in the relationship between agency mortgage purchases and credit conditions is to use a narrative identification approach involving major regulatory events impacting agency mortgage holdings. By focusing on policy interventions by the federal government, we exclude variation in purchase activity resulting from the agencies' regular response to market developments. Be-cause policymakers themselves often respond to conditions in mortgage and housing markets, we exclude interventions with short-run stabilization motives as the primary objective. The end result of our narrative analysis is a record of housing credit policy events that we use as an instrumental variable for agency purchase activity. Here, we summarize the methodology of the narrative analysis, and describe the resulting policy indicators. A companion background paper, Fieldhouse and Mertens (2017), provides the full narrative analysis of credit policy events, including explanations of relevant findings for each policy event and extensive documentation that allows verification of our analysis.

Overview of Methodology
The development of the narrative instrumental variable follows five steps: Identifying significant policy changes affecting agency portfolios; quantifying their ex ante projected impact on agency holdings; pinpointing the timing of when the policies became publicly known; classifying each policy change as either cyclically or non-cyclically motivated; and restricting the sample for consistent use as an instrument for agency purchasing activity. Next, we describe the procedures used in each of these steps. Table 1 provides an overview of the historical primary sources used in the narrative analysis.
I. Identifying Significant Policy Changes Policy changes affecting agency purchases and mortgages holdings have historically been directed by a range of policymakers, notably Congress, the President and the Cabinet, particularly the Secretaries of the Treasury and HUD, various regulatory agencies in the executive branch, and the Federal Reserve. The relevant regulatory institutions were disbanded and reinvented several times over the decades, and as a result there is no single consistent source tracking the history of housing credit policy. Instead, a wide range of sources is required for identifying and analyzing policy changes.
Policy actions generally originate from one of three sources: enacted legislative changes, regulatory policy changes published in the Federal Register or as other binding agreements with regulators, and macroeconomic stabilization policies managed by the Federal Reserve or Treasury. We restrict attention to significant policy actions, meaning actions that would either be expected to directly impact agencies' permissible volume of net purchases and retained portfolio holdings, or else considerably expand the pool of eligible mort-gages an agency was authorized or required to purchase. Interventions determined at the legislative level include adjusting statutory leverage ratios, capital requirements, and conforming loan limits, provision of working capital, mandatory retirements of public stock, and direct appropriations or borrowing authority for purchases, among others. Regulatory policy actions include setting permissible debt-to-capital ratios, imposing capital surcharges in excess of statutory capital requirements, capping portfolio size or growth, setting bill. 13 We then analyze relevant sections of these primary sources to confirm these laws' material impact on mortgage holdings and better understand the nature of the policy changes.
Legislative actions often set in motion the drafting of new regulatory rules. Identified significant legislative events are the starting point for a directed search of the related regulatory changes in HeinOnline's Federal Register Library. We also obtain information from the GSEs' annual reports about significant regulatory changes, as well as from 10-K filings in more recent years. We additionally used sections of the Economic 13 The ProQuest Congressional Publications Database provides a comprehensive compilation of all public laws, committee reports, and hearings. Public laws and related legislative actions since 1973 are available from Congress.gov, a project of the Library of Congress, along with committee reports since 1995. Most older public laws are available through LegisWorks Statutes at Large Project. Most hearing transcripts are digitally available since 1985 from the US Government Publishing Office.
Report of the President and Annual Report of the Board of Governors of the Federal Reserve, as well as the various reports by regulators to collect information about regulatory rulings. We use newspapers, financial newswires, and mortgage industry newsletters to help direct the search for information about the rulings in the Federal Register, particularly the Wall Street Journal, American Banker, and National Mortgage News. 14 Final rules published in the Federal Register almost always include a detailed background and overview of the initial proposed rule, public comments received, and subsequent modifications.
Using these procedures, we are confident that we have identified the overwhelming majority of significant policy events. The main concern is developing a policy indicator that is correlated with underlying regulatory shocks to agency purchasing activity. The larger the number of significant policy events identified, the higher the relevance of the instrument.

II. Quantification
To be included, we require that primary sources either explicitly cite projections of the policy change's impact, or contain information that can be used to quantify the impact. We describe here our general approach, and show later that the resulting projections align closely with the ex post estimated balance sheet impact.
For each policy change, we use contemporaneous sources to obtain an ex ante estimate of the projected impact on the agencies' capacity to purchase mortgages, measured in annualized billions of dollars. If a baseline is needed for quantifying a policy change, say for Fannie's regulatory capital when its debt-tocapital ratio is increased, we use the most recent data publicly available prior to the policy change. We use ex ante balance sheet data on regulatory capital, liabilities, and/or assets in conjunction with standing leverage or capitalization requirements to estimate the impact of related changes, such as increases in permissible leverage ratios. Similarly, public capital injections are quantified as a multiple of one more than the prevailing leverage ratio, to capture the potential increase in assets supported by related debt issues plus the working capital itself. Direct appropriations are straightforward to quantify, at most requiring a pro-rata annualization adjustment based on relevant implementation lags. To quantify potential impacts of discretionary conforming loan limit changes, we rely on estimates from Congressional committee reports accompanying legislation. Such reports typically cite the extent to which a large conforming loan limit increase would restore a GSE's real purchase activity. We quantify the impact of such adjustments as the difference between annualized purchase volumes immediately preceding the policy change and the home price index-adjusted purchase volume of the benchmark year being restored. For relatively large, open-ended changes, such as leverage ratio increases, potential effects on mortgage holdings are annualized using a two-year rule, which assumes half of the full potential impact would be realized within the first year of taking effect.
For other policies that are inherently harder to quantify, such as authorizations for program expansions into new mortgage market segments, we search for ex ante estimates of projected impacts on purchasing activity from committee reports, market analysts, regulators, or agency executives. We do not include policies that would not have been expected to impose or alleviate binding constraints on agency activity. For instance, when adjustments to leverage ratios or affordable housing goals are viewed as non-binding by most accounts and this appears consistent with the agencies' balance sheet and purchase behavior, we do not consider the policy change significant. We also exclude any laws or regulations that merely extend prior authorizations, and for certain authorizations affecting Ginnie Mae, we use a current policy baseline as opposed to a current law baseline for scoring annual funding changes.
When estimating the quantitative aspects of the policies, we rely on information released by the Congressional Budget Office, Government Accountability Office, Treasury Department, and Congressional Research Service that contain detailed analyses of policy changes, background information, and/or balance sheet data for the agencies in question, see Table 1. We also use information from the annual or periodic reports of the agencies and regulators, particularly regarding balance sheet data, and from appropriations bills and budget appendices for certain policies affecting Ginnie Mae. Committee report language occasionally cites projected effects of a pending policy change, and we also use the financial press and industry newsletters to search for projections of the impact of policies that are difficult to quantify.

III. Timing
At the operational level, the agencies sell commitments to purchase conforming mortgages from primary market lenders, which may then be exercised by the mortgagee up to an expiration date. Consequently, actual agency purchases tend to lag behind the issuance of commitments to purchase mortgages from primary market originators. Together with the usual policy implementation lags, the policy events are therefore best thought of as news shocks about agency mortgage purchases. We date each policy intervention to the month in which we estimate that it became publicly anticipated, rather than the month in which it was formally announced or took effect. We show in the next section that this timing approach is roughly consistent with the observed movements in consolidated agency mortgage holdings.
The ProQuest Congressional Publications Database, HeinOnline's Federal Register Library, the CQ Almanac, and financial press are the primary sources used for documenting pertinent news surrounding policy changes and the implementation dates. For regulatory changes, we use the month in which proposed rules were first published in the Federal Register or reported in the press. We date legislative changes when the provision including the policy change was agreed upon in the House, Senate, or conference version of a bill, rather than upon subsequent enactment. For Fannie and Freddie, we additionally check the timing by cross-referencing policy announcements with GSE stock price movements and the financial press, as often policy news is priced into GSE shares.

IV. Classification by Motivation
The classification of the policy events distinguishes between interventions that are guided by prevailing business cycle and financial conditions, and those that are plausibly free of such contemporaneous influences. Our instrument for agency mortgage purchases only includes the latter to avoid bias due to the systematic relaxation of policies during periods of stress in mortgage or housing markets. The classification is based on identifying the primary motivations underlying each of the policy interventions. To make this classification, we parse historical documents, paying particular attention to the rationales invoked by policymakers and the press, the nature of the legislative vehicles or regulatory processes, the relation to known periods of economic and financial stress, and the time horizon of policy objectives.
The principal data sources for identifying policy motives include Congressional committee reports and hearings, Presidential speeches and signing statements, the Budget of the US Government, Economic Report of the President, Federal Reserve Bulletin, Annual Report of the Board of Governors of the Federal Reserve, CQ Almanac, and the financial press (see Table 1). For legislated policies, the accompanying reports of the Senate Committee on Banking, Housing and Urban Affairs and the House Financial Services Committee typically detail congressional intent and any pertinent economic context. Major housing policy laws are also usually accompanied by a Presidential signing statement explaining the bill's motivation, context, and intended impact. Budget appendices and/or committee reports accompanying appropriations bills usually explain the impetus for certain policy changes affecting Ginnie Mae. Final rules published in the Federal Register also almost always include a detailed background and history, shedding light on regulators' motives.
Based on these sources, we classify the policy changes as either cyclically motivated or non-cyclically motivated. Interventions classified as cyclically motivated tend to emphasize short-term outcomes, such as boosting housing starts in a recession. Legislative vehicles for such policy actions tend to be quickly drafted and enacted, with a relatively concise legislative history and narrow focus. Policymakers are typically quite explicit about cyclical concerns and objectives, overwhelmingly so when policies are implemented in close proximity to recessions or credit crunches. Language we search for in committee reports and signing statements as strong evidence of cyclical motivations include "emergency, crisis, recession, credit shortage, credit crunch, housing starts, employment, construction, downturn, depressed, stimulus, boost", etc. Policies enacted during or near a recession or credit crunch are held to a particularly high bar for being classified as non-cyclical, but are not automatically classified as cyclically motivated.
Interventions motivated by social policy, budgetary, or other more ideological objectives are classified as unrelated to the business or financial cycle, provided the various historical sources do not at the same time indicate significant short-term economic or financial market concerns. Political rather than economic context shapes the development of these interventions, such as an administration's emphasis on expanding affordable homeownership opportunities to lower-income households, concerns regarding the structural budget deficit, or ideological hostility toward the GSEs. It is often hard to establish a single rationale for the non-cyclical ac-tions, which can be motivated by a mix of objectives. For our purposes, however, a more precise distinction between these objectives is not essential. Language we search for as indicative of non-cyclical motivations include "long-term, farsighted, comprehensive, low-income, affordable housing, American Dream, homeownership, budget deficit, reduce borrowing, off-budget, privatize," etc. Legislative actions classified as non-cyclical emphasize longer-term outcomes, such as increasing homeownership rates. Legislative vehicles for such interventions tend to be slower-moving bills, particularly deliberate overhauls of housing policy with a lengthy legislative history; the National Housing Acts, Housing and Urban Development Acts, and Housing and Community Development Acts of various years tend to meet this description, being slowly crafted and negotiated between the House, Senate, and White House, and focusing on broad, long-term objectives for housing policy, such as urban revitalization or access to affordable housing for various constituencies. New regulatory rules set in motion by such bills also tend to be classified as non-cyclical, such as HUD setting new affordable housing goals for the GSEs. Occasionally, interventions are prompted by specific events that we view as unrelated to the cycle, such as the regulatory actions taken in the aftermath of accounting scandals at Fannie and Freddie in 2003-2004. V. Sample Restrictions Occasionally a law or public rule sets in place changes in purchase authorizations or balance sheet restrictions to take effect only multiple years after announcement. To obtain a good indicator for news about pending purchase behavior, we exclude changes with very long implementation delays and focus on interventions taking effect within nine months of their news being made public. We also restrict attention to policy events after January 1967. This choice is made to select a period of relative institutional stability, as it roughly coincides with the creation of Ginnie and Freddie, the emergence of a nationwide secondary market for conventional mortgages, and the beginning of the ascendancy of the privatized GSE era. This starting point is also in part determined by the availability of time series used in the empirical analysis. We focus exclusively on the mortgage portfolio activity of Fannie, Freddie, and Ginnie, ignoring less significant government entities for which monthly data is not easily available. We also include purchases by the Federal Reserve and Treasury in the recent financial crisis, but in most of the analysis in Sections 5 and 6 the sample is truncated at December 2006 to deliberately exclude the financial crisis and the Fannie and Freddie conservatorship period. As shown in Figure 1, the three housing agencies that we analyze account for the large majority of government agency mortgage holdings between 1967 and 2006. Table 2 lists the policy events resulting from the narrative analysis. Each intervention is described by the agencies affected, by its annualized projected impact (in billions of US dollars), timing, and motivation. The  Do the policy changes that we have narratively identified lead to actual changes in agency purchases and retained mortgage portfolios? To investigate this, it is important to take into account the potentially significant delays between the policy events and their impact on the agency portfolios. Recall that agencies initially make advance commitments to buy loans from mortgage providers and subsequently effectuate these as loans are originated in the primary market. We regress three activity indicators, net mortgage purchase commitments made by the agencies, the actual net purchases of mortgages, and the stock of agency mortgage holdings, on the indicator for non-cyclical policy events m t :

The Narrative Measures of Policy Changes
We use monthly observations from January 1967 to December 2014 in log first differences of current dollars. Because monthly commitment and purchase flows are relatively volatile, we run the regressions for a 36 month backward moving average of these two variables. The event indicator m t is the non-cyclically motivated narrative measure scaled by the average level of agency mortgage holdings over the prior 12 months.
All regressions include the current value of m t as well as three years of lags. For each activity measure, we estimate two versions of (1), one in which we set l = 0 and one in which l = 12. The second version includes a full year of leads of m t , which allows us to verify the plausibility of our timing of the interventions. Figure   4 shows the estimated dynamics of the agency activity measures around the policy events, obtained as the sum of the slope coefficients β j over various horizons, together with 95 percent Newey and West (1987) confidence bands.
The policy change indicator strongly predicts significant changes in agency purchase commitments (left panel) which pick up at date 0 and peak at a 3 to 4 percent higher level roughly two years out. Actual purchases (middle panel) follow a very similar trajectory as commitments, but with a lag of a few months, demonstrating that the policy changes act as news shocks for actual purchases. The right panel shows that the higher pace of mortgage purchases is followed by a persistent and statistically significant increase in the agencies' retained portfolio.
The specifications allowing a lead response to the events suggest that our indicator accurately captures the timing of the change in purchasing activity. It is also encouraging that the size of the estimated response of agency mortgage holdings converges to around one percent within 24 months of the policy event, which is consistent with our scoring of the projected impact of the policy changes.

The Cumulative Effects of Agency Mortgage Purchases on Mortgage Credit
We estimate the cumulative impact of agency mortgage purchases on credit aggregates by using the noncyclically motivated policy changes as an instrument for agency purchasing activity. As shown in Figure 1, the growth in mortgage debt over our sample period has on average exceeded growth in GDP, while agency holdings have increased at an even faster pace. Because of these differential trends, expressing the impact on credit aggregates and other variables in terms of elasticities can be misleading. 16 To address these issues, we run local projections-IV regressions similar to those used by Ramey and Zubairy (2016) to estimate cumulative government spending multipliers.
Our results are based on local projections for horizon h and outcome variable y t of the form where p t is either the volume of commitments or actual purchases by the agencies in month t. Both y t and p t are expressed in constant dollars using the core PCE price index. For every horizon h, the change in y t as well as the cumulative change in commitments or purchases p t is expressed as a ratio of X t , a deterministic trend in real personal income obtained by fitting a third degree polynomial of time to the log of personal income deflated by the core PCE price index. 17 For stock variables, the dependent variable is the change in the stock between t − 1 and period t + h, scaled by X t . For credit flow measures, we construct y t by cumulating the flows f t such that y t+h − y t−1 = ∑ h j=0 f t+ j .
Each regression includes a full year of monthly lags of a number of control variables Z t , such that ϕ h (L) is a lag polynomial of order 11. The controls include variables with predictive content for the dependent variables, and always include lagged values of y t /X t (or f t /X t for flow variables), as well as lags of agency 16 Both the news aspect and the scaling issues also occur in other contexts, in particular in the measurement of the effects of fiscal policy interventions, see for instance Ramey (2011), Mertens and Ravn (2012), and Ramey and Zubairy (2016). 17 The results do not differ meaningfully when we use polynomials of different order. In online appendix A.1, we also show that the results are robust to using a trend in mortgage originations instead of personal income. net purchases and commitments as a ratio of X t . In addition Z t contains lagged growth rates of the core PCE price index, a nominal house price index and total mortgage debt, the log level of real mortgage originations, housing starts, and lags of several interest rate variables: the 3-month T-bill rate, the 10-year Treasury rate, the conventional mortgage interest rate, and the BAA-AAA corporate bond spread. Finally, we add lags of two cyclical indicators: the unemployment rate and the growth rate of real personal income. All growth rates are quarter-over-quarter. The data appendix provides full details on the sources and construction of the time series. In online appendix A.1, we discuss results for a number of alternative control (sub)sets.
The coefficient γ h in (2) measures the multiplier associated with an additional dollar in commitments or purchases made between period t and t + h. This multiplier is the total cumulative dollar change in y t between period t and t + h. We estimate γ h by two-stage least squares (2SLS) using the dollar impact estimates of the non-cyclically motivated policy events reported in Table 2, deflated by the core PCE price index and expressed as a ratio of X t , as the instrument. Our baseline estimates use an effective sample of 480 monthly observations, starting in January 1967. 18 In online appendix A.1, we look at different sample periods.
The central identifying restriction is exogeneity of the instrument, which requires that the residuals in (2) and the narrative measure are uncorrelated. To the extent that the lagged controls are informationally equivalent to all relevant impulses to the dependent variables occurring prior to time t, the regression residuals correspond to their horizon h forecast errors. The latter depend only on unpredictable shocks occurring between period t and t + h. Our instrument is based on the projected impact of policy events constructed from ex ante information. These estimates should therefore be uncorrelated with shocks occurring after time t. The identifying restriction then boils down to the assumption of orthogonality between the instrument and all shocks in month t other than the one associated with the policy event itself. If the control set does not fully capture all impulses prior to date t − 1, then the exogeneity requirement is stricter and the instrument must be uncorrelated with the history of relevant impulses to the left hand side variables. The omission of the cyclically motivated events aims at dropping policy actions that may be correlated with all other time t shocks.
Our narrative classification retains the non-cyclically motivated events for which correlation with contemporaneous shocks is unlikely, while the lagged controls provide additional insurance that the confounding effects of any remaining correlations with prior shocks are eliminated, see also Ramey (2016).

First-Stage Diagnostics
We first assess the strength of our narrative instrument. The left panel of Figure 5 shows the Newey and West (1987) robust F-statistics on the excluded instrument in the first-stage regressions for horizons h = 0 to h = 60, which are of the form wherem t is the non-cyclical narrative policy indicator expressed in real dollars. The figure shows the Fstatistics when using either cumulative commitments or purchases as the measure of agency activity p t .
The results indicate that the narrative measure is a reasonably strong instrument for agency purchasing activity for horizons between 4 to 48 months after the policy events, with robust F-test statistics exceeding or close to 10. The F-statistics are low for very short horizons, which is natural given the implementation lags and our timing according to the arrival of news about impending regulatory changes. Beyond horizons of 48 months, the F-statistics fall to lower levels, which is also not surprising as other influences on agency purchases accumulate with the forecast horizon. Given these results we restrict attention to the 4-to 48month horizon. The F-statistics are very similar when we instrument for either purchases or commitments.
The right panel of Figure 5 depicts IV estimates of the dollar change in agency purchases for every dollar of commitments issued over the various time horizons based on the regressions in (2) using cumulative agency purchases as the outcome variable and cumulative commitments as the independent variable. The fine lines denote 95 percent Newey and West (1987) confidence intervals. Because of the time delays associated with secondary market transactions, the pass-through from commitments to purchases is high but smaller than unity for shorter horizons. After about one year the relationship becomes one-for-one with very narrow confidence intervals. The interpretation of the credit multiplier estimates presented next therefore depends somewhat on the denominator used, but only for horizons of less than one year. At longer horizons, there is essentially no difference between using commitments or purchases as the agency action measure.

Cumulative Credit Multipliers
According to the Meltzer-Greenspan view, the portfolio activities of the agencies have no meaningful impact on housing or household debt. Without credit market imperfections, the ownership of mortgage debt is irrelevant. Any change in agency mortgage holdings has no impact on total mortgage debt, but leads instead to perfect crowding out of private holdings. If, on the other hand, there are frictions impeding on the private flow of credit to residential borrowers, agency activity may not be neutral for the volume of mortgage lending. We now examine whether agency mortgage purchases indeed impact housing credit, and test the neutrality hypothesis using the local projections in (2) and the narrative policy instrument. Figure 6 shows the impact of an increase in either agency commitments or purchases, together with the 95 percent Newey and West (1987) confidence bands. There is a marked difference between the short and long-run effects. In the short run, the results are consistent with neutrality: The upper left panel shows that a dollar purchased increases agency mortgage holdings initially by almost a dollar. The short-run effect of a dollar increase in commitments on agency holdings is lower at around 60 cents, which is expected given the time delay between commitments and purchases. The right panel shows that private holdings decline initially by roughly the same amount as the increase in agency holdings, although the confidence bands are wide. 19 The middle panels in Figure 6 show that as the dollar in mortgage debt changes from private to agency ownership, there are initially no significant changes in originations or mortgage debt.
Over longer horizons, however, there is clear evidence against the notion that agency purchases are neutral for mortgage credit. The cumulative impact on total mortgage originations increases with the horizon and becomes statistically significant after 6 months. Over the course of 3 years and beyond, there is a cumulative increase in originations of 3 dollars or more for every dollar purchased by the agencies. The estimated long-run multipliers for total originations are highly statistically significant and nearly identical for commitments and purchases. The point estimates for the impact on the stock of mortgage debt at shorter horizons are roughly in line with the range reported in Smith, Rosen, and Fallis (1988). The increase in mortgage debt becomes statistically significant after three to four years and in the longer run reaches a level of around one dollar. As the time horizon grows, the increase in agency holdings slowly dissipates toward levels expected before the expansion. Similarly, the negative impact on the level of private mortgage holdings vanishes over time and eventually turns into an increase, although not one that is statistically significant.
The results in the middle row of Figure 6 imply that agency portfolio expansions lead to a substantial rise in mortgage lending activity. Originations take place when borrowers refinance, purchase an existing home, or and by a substantially larger amount than purchase originations. Refinancing originations see a statistically significant increase beyond 6 months, and within 3 years are higher by roughly 3 dollars per dollar of agency purchases. Home purchase originations rise more slowly and are statistically significantly higher after 18 months, increasing by nearly one dollar within 4 years. The rise in purchase originations occurs somewhat faster than the rise in total mortgage debt, suggesting that existing home sales respond before new home sales. The longer-run cumulative change in purchase originations is comparable to the increase in mortgage debt, which suggests a positive impact on residential construction. In the impulse response analysis below, we indeed find evidence for an increase in housing starts. We also document positive effects on homeownership rates and, although less clearly, on home prices, both of which also contribute to the rise in mortgage debt. The bulk of the effect on originations is nevertheless due to refinancing. 21 Given the procyclicality of mortgage originations, the large positive effect on originations makes it unlikely that the estimates are severely contaminated by the countercyclical actions of the agencies over the sample.
The decrease in private holdings and the delayed effect on originations also suggest that the estimates are not merely picking up increased supply of mortgages to the secondary market, as would occur if private lenders simply passed on newly originated loans to the agencies; if this were the case, originations would rise before or roughly simultaneously with agency purchases and without a decline in private holdings. It is precisely in this direction that the results change when we instead estimate γ h in (2) by ordinary least-squares (OLS).
To illustrate this, Figure 7 compares the OLS and 2SLS responses of mortgage originations. Regardless of whether the baseline or full set of controls is included, the OLS estimates suggest that the bulk of the increase in originations occurs within a few months. Given the decision lags and time delays associated with making new mortgage loans, the delayed and gradual rise in originations that appears after instrumentation is much more consistent with a causal interpretation. The comparison in Figure 7 is generally informative about the endogeneity concerns in practice. Systematic GSE expansions during times of high primary mortgage demand or high private sector credit supply would lead to an upward bias using OLS, while the stabilizing actions of the agencies lead to a downward bias. Since the OLS estimates exceed the 2SLS estimates for horizons up to 2 years, the former appears the dominant practical concern in our sample. 22

Impulse Response Analysis of News Shocks to Agency Purchases
To evaluate the effects of agency purchases on residential investment and homeownership, as well as analyze the response of interest rates and other macro aggregates, in this section we conduct an impulse response analysis of shocks to agency mortgage purchases. Given the gradual and anticipated nature of agency balance sheet expansions, our goal is to identify the response to shocks to expectations of future agency purchasing 21 This is consistent with Di , who document an increase in refinancing activity by 170% during the Fed's first QE program. 22 In online appendix A.1, we elaborate on the role of instrumenting, and we discuss additional results on agency securitization.
We also verify robustness in several dimensions, such as the choice of scaling variable X t , the sample choice, the set of controls, as well as the exclusion of specific policy events in the narrative instrument. The expansionary effects of agency purchases on mortgage credit are shown to be robust to many details of the analysis. activity. We adopt a local projections approach and use the narrative instrument for identification. We also present results for an alternative instrumentation strategy that exploits information in GSE stock prices.

Empirical Specification with the Narrative Instrument
For a given monthly outcome variable y t , we estimate the response at horizon h based on The right hand side variable of interest measures annualized agency commitments made over an 8 month period, expressed as a ratio ofX t , a long-run trend in annualized originations. The latter is obtained by fitting a third degree polynomial of time to the log of real originations obtained using the core PCE price index as the deflator. The control variables Z t−1 are the same as in equation (2) estimating dollar cumulative effects.
When an outcome variable is not included in the benchmark control set, we always add 12 monthly lags of that variable as additional controls (in growth rates for trending variables and in levels for other variables).
The regression in (4) estimates the month h ≥ 0 response to a time 0 news shock to agency purchases.
Expected agency purchases are proxied by agency commitments made over the next 8 months. We choose an 8 month horizon to measure expected future commitments because at this horizon the robust F-statistic associated with the narrative instrument in the first-stage regression is the largest, and equals 11.68. The results are very similar for somewhat shorter or longer horizons. To address endogeneity, we use the indicator of non-cyclical policy events, deflated by the core PCE price index and scaled by trend originations X t , as the instrument. The IV estimates of δ h in (4)

Empirical Specification with the GSE Excess Returns Instrument
Although our narrative instrument is a good predictor of agency purchasing activity, it is based on relatively few policy events. It is therefore always possible that our findings are driven by the small sample size. To gain confidence that this is not the case, as well as address other potential concerns with the narrative identification method, we also pursue an alternative approach inspired by Fisher and Peters (2010). These authors interpret innovations in excess stock returns of major defense contractors as news shocks about future military spending. Fisher and Peters (2010) obtain these innovations by ordering the excess returns last in a recursively identified structural vector autoregressive system. The recursive scheme assumes that none of the macro aggregates included in the analysis are affected on impact by the news shock, while excess stock returns react contemporaneously to all macroeconomic shocks. We follow a similar strategy to identify the response to news shocks to agency purchases. 23 Passmore (2005) estimates that 44 percent to 89 percent of Fannie and Freddie's stock market value is derived from their special GSE status, and that the GSEs would hold far fewer mortgages in portfolio and have higher capital ratios if they were purely private. Any news about changes in the policies guiding the GSEs' portfolios business and leverage is thus likely to affect their market value. To construct an alternative measure of news of agency purchasing activity, we identify innovations in the excess return of GSE stock based on similar recursivity assumptions as Fisher and Peters (2010). To implement these assumptions, we estimate the response at horizon h through the following regression: Relative to (4), this specification adds the contemporaneous controls W t , which includes the interest rate variables (3-month T-bill, 10-year Treasury, the conventional rate, BAA spread), the log of real originations, the log changes in mortgage debt, real house prices, the core PCE price index and personal income, the log of housing starts, and the unemployment rate. When we rotate in another variable, we also include it in W t .
One other modification relative to (4) is that the log of the GSE stock price-to-S&P 500 ratio is added to Z t .
The response coefficient δ h in (5) is estimated using 2SLS, where the first stage is a regression of the log change in the GSE stock price-to-S&P 500 ratio on W t and Z t−1 . This specification imposes that shocks to expected agency purchases have no contemporaneous impact on the variables in W t . The instrument used in the 2SLS estimation of (5) is therefore the monthly innovation in the GSE excess stock return, orthogonalized to the monthly innovations in the variables in W t . By assumption, other endogenous influences on GSE excess returns, such as shocks to household demand for mortgages, are eliminated by allowing excess returns to respond contemporaneously to mortgage credit, interest rates, prices, and the cyclical indicators. Figure 8 plots the standardized cumulative sum of these innovations, together with the non-cyclical narrative indicator for reference. The GSE excess returns shocks are followed by statistically significant increases in agency purchasing activity. The F-statistic for the GSE excess returns instrument is the highest for agency commitments at a horizon of 5 months, and equals 8.55. 24 To provide further evidence that GSE stock prices reflect changes in agency purchasing activity, Figure 9 plots the response of the GSE stock price-to-S&P 500 ratio to a one pp. increase in the expected future agency market share, measured by agency commitments as a ratio of trend originations. This response is estimated by (4) using the narrative policy indicator as the instrument, and reveals a clear relative increase in GSE stock prices that is statistically significant at horizons of 4 to 12 months. Fieldhouse and Mertens (2017) also provide narrative evidence that announcements of policy changes affecting Fannie and Freddie are generally associated with adjustments in their stock prices.
Because the GSE excess returns instrument contains monthly observations, it potentially contains more information about variation in agency purchases than the narrative policy indicator. 25 Our narrative indicator contains, for instance, little information for the 1990s because of the scarcity of quantifiable and binding regulatory changes. However, this period witnessed a rapid expansion of GSE balance sheets and may be particularly important for learning the effects of agency purchases. As is well known, however, equity prices 24 For simplicity, we keep the horizon for cumulating commitments in equation (5) at 8 months, the same as in equation (4). The value of the first-stage robust F-statistic for this horizon is 7.68. Changing the horizon for cumulating commitments in specification (5) to 5 months does not lead to any meaningful changes in the results. 25 Assuming the GSE excess returns shocks contain all of the information about agency purchase shocks, it becomes possible to estimate the variance contribution of these shocks to any endogenous variables of interest. In online appendix A.5, we do this in the context of an SVAR model. are volatile, and the GSE excess return shocks are, on the other hand, also relatively noisy. While the GSE excess returns shocks clearly have predictive power for agency commitment activity, the first-stage F-statistics are somewhat lower than for the narrative instrument. Another caveat is that the GSE excess return shocks may also pick up unanticipated variation in the scale of the GSEs' securitization business. Nevertheless, we view this identification strategy as a useful alternative to the narrative approach. The next sections show that both instruments generally lead to very similar impulse response estimates. The first row in Figure 10 displays the responses of real originations and mortgage debt to the agency purchase shock. Based on the narrative instrument, mortgage originations start rising after a few months and reach peak increases of 4 percent to 5 percent between 12 and 18 months after the shock. With a slightly longer delay, the stock of mortgage debt also gradually rises to levels that are about 0.3 percent higher after two years. Using the GSE excess returns instrument, the rise in originations occurs slightly more rapidly but is more transitory. The peak increase in originations is around 4 percent and occurs between 10 and 14 months. The size and shape of the rise in the stock of mortgage debt is also very similar across instruments.

Effects on Mortgage Credit and Interest Rates
The expansions in both the stock and gross flow of mortgage credit following a positive shock to agency purchases are statistically significant for multiple periods for both instruments. The results again indicate that agency purchases stimulate mortgage lending significantly. Online appendix A.3 shows that the impulse response analysis also confirms that refinancing originations account for a large share of the increase.
The second row in Figure 10 shows the impact on interest rates on 30-year fixed rate mortgages in the primary market. The left panel illustrates the interest rate effect on newly originated conventional conforming mortgages, whereas the right panel contains the impact on interest rates of mortgages guaranteed by the Federal Housing Administration. According to the narrative instrument, the mortgage rates in the primary market are largely unaffected in the initial months after the increase in agency mortgage purchase commitments. As the agencies' purchasing activity picks up, however, both mortgage rates gradually decline and are lower by around 10 basis points after 6 months. The declines in mortgages rates appear quite persistent, and help explain the increase in refinancing activity. When using the GSE excess returns instrument, by construction there is no impact on interest rates in the first month. After 6 to 18 months, both mortgage rates are lower by 10 to 15 basis points. For both instruments, the declines in mortgage rates are statistically significant for multiple periods. This decrease in mortgage cost is consistent with agency purchases affecting the aggregate supply of housing credit, for instance because of portfolio rebalancing effects or because private mortgage lenders are capital constrained. Agency purchases also alleviate any constraints faced by private intermediaries, for instance because the higher prices of mortgage assets improve their net worth, or because the sale of mortgages in exchange for agency debt lowers their risk-weighted leverage.
The issuance of agency debt to finance the mortgage purchases potentially puts upward pressure on interest rates on other debt instruments. Such pressure may be limited if significant amounts of agency debt are purchased by foreign investors, as has been the case since the mid-1980s, or by the Federal Reserve, as was the case in the early years of our sample, see Haltum and Sharp (2014). Depending on the level of segmentation in financial markets, the rebalancing and other effects may also spill over to other asset markets and cause the yields on substitutes to mortgages to fall. These include other high duration instruments such as long-term Treasuries and corporate bonds. In addition, lower mortgage rates lead to more prepayments, which do not carry any penalty in the United States. There is considerable evidence that lower effective durations cause mortgage investors to bid up the price of higher duration instruments, see for instance Boudoukh et al. (1997), Perli and Sack (2003), Hanson (2014), and Malkhozov et al. (2016). The broader impact on long-term yields is therefore ex ante not clear.
The left panel of the bottom row in Figure 10 shows the estimated response of the 10-year Treasury rate.
The results are very similar to those for the long-term mortgage rates just discussed: The 10-year Treasury rate responds little the first couple of months, but as the agency mortgage purchases commence, it declines in a gradual and persistent manner by up to 5 to 10 basis points. The drop is significant between 3 and 6 months after the shock. The results are similar for the two instruments but confidence intervals are wider using excess returns. In online appendix A.2, we discuss additional results on the effects on other interest rates and credit spreads. 26 Overall, the results indicate that there are significant spill-overs from agency actions in mortgage markets to other longer-term asset markets.
The right panel in the bottom row of Figure 10 reports the impact on the 3-month T-bill rate. The results are qualitatively similar to those for the long-term rates discussed above. Quantitatively, we find some indication of a larger drop in short-term rates than in the longer term rates. Based on the narrative instrument, and with a delay of a few months, the T-bill rate drops persistently by 15 to 20 basis points with a partial reversion taking place at longer forecast horizons. The results using the GSE excess return instrument yields again similar results (although quantitatively smaller). The negative response of short-term interest rates indicates that a potentially important explanation for the expansion in mortgage lending and the decline in mortgage rates is a more accommodative stance of monetary policy. In Section 7 below, we investigate the role of monetary policy and its interactions with housing credit policy in greater detail.
The finding that increases in agency mortgage purchases produces a boom in mortgage lending and declining interest rates is robust. In online appendix A.4, we report very similar results for samples that omit the Volcker years, or for the subsample starting in October 1982, the end of the period of non-borrowed reserve targeting by the Federal Reserve. Thus, the results are not driven by differences in Federal Reserve operating procedures in the 1970s or by the inclusion of the Volcker period. There is narrative evidence that political pressure to support the GSEs was exerted with some success in the late 1960s and 1970s, leading for instance the Federal Reserve to purchase significant amounts of agency debt, see Haltum and Sharp (2014).
In the post-1982 sample, however, it seems less likely that political pressure to support government housing policies can explain an accommodative monetary policy response. Finally, the results also do not appear to be particularly sensitive to the inclusion of any individual policy event, see online appendix A.4.

Effects on Housing and Other Macro Aggregates
Next, we assess the evidence for the broader macroeconomic effects of government asset purchases. Figure   11 shows the responses of a range of monthly macro aggregates to an agency purchase shock, together with 95 percent Newey and West (1987) confidence bands. As in Figure 10, the responses are to an anticipated increase in purchases by one percentage point of trend originations, estimated using either the narrative instrument (blue) or the GSE excess returns instrument (red). We consider the following additional outcome variables at the monthly frequency: housing starts, real house prices, the homeownership rate, real personal consumption expenditures, real personal income, and the unemployment rate. 27 The first panel in Figure 11 shows the effects on residential investment, as measured by monthly housing starts. Based on the narrative instrument, the number of new housing starts rises to levels that are roughly 1 to 2 percent higher after about 6 months. Housing starts remain elevated for about a year and drop off to prior levels afterwards. Using the GSE excess returns instrument, housing starts pick up more quickly and remain around 2 percent higher between 4 and 12 months after the shock. For both instruments, the size and shape of the response of housing starts is similar and statistically significant for multiple periods. We thus find evidence that the expansion in the stock of mortgage debt following a shock to agency purchases is associated with higher levels of residential investment. 28 The top middle panel in Figure 11 plots the impact on real house prices, as measured by the Freddie Mac house price index deflated by the core PCE price index. Using the narrative instrument, we find that real house prices rise gradually but very persistently over time, with a point estimate that becomes significantly positive at longer forecast horizons only. The GSE excess returns instrument, in contrast, does not reveal a significant change in house prices at any horizon. Thus, we have no clearcut evidence of an impact of agency mortgage purchases on house prices. Moreover, the point estimates imply that very little of the increase in 27 All these variables, except the unemployment and the homeownership rate, are included logs and all nominal variables are deflated by the core PCE price index. The homeownership rate is only available at quarterly frequency, and the monthly series in this case simply consists of the quarter values. See the data appendix for precise definitions and sources. 28 The more immediate effects on residential construction are consistent with the more delayed impact on mortgage debt in Figure 10. This is because financing in the building phase is typically through a short-term construction loan that is converted into a residential mortgage loans only after the borrower takes up occupancy of the house. the dollar volume of mortgage credit is explained by increases in house prices.
The top right panel in Figure 11 shows the response of the homeownership rate, as measured by the Census Bureau, which is often cited as one of the primary motivations for housing credit policy. Using the narrative instrument, there is a sustained increase in homeownerhip by around 5 basis points beyond a horizon of 10 months. Based on the GSE excess returns instrument, homeownerhip also rises by a similar magnitude, but the increase occurs somewhat more rapidly. While there is considerable uncertainty in the estimates, the responses are in both cases statistically significant at the 95 percent level for multiple months, indicating that agency activity indeed has an effect on homeownership. This also implies that the expansion in the stock of mortgage debt is in part driven by an increase in homeownership.
The remaining panels in Figure 11 show the responses of consumption expenditures, personal income, and the unemployment rate. Using the narrative instrument, we find that an increase in agency mortgage purchases stimulates consumption very modestly and with a delay of more than a year. Personal sector income and the unemployment rate are approximately unchanged over the entire forecast horizon. The increase in consumption is imprecisely estimated and none of the impulse responses are significantly different from zero at the 95 percent level. The point estimates are somewhat different for the GSE excess returns instrument and show an increase in personal consumption expenditures similar to that estimated using the narrative instrument, but with a much shorter delay of 3-4 months. There is a decline in the unemployment rate around a year after the shock. Standard errors are, however, large for this instrument as well.
In sum, we find evidence that agency mortgage purchases stimulate residential investment and homeownership, and some indication of a positive effect on personal consumption expenditures. The confidence bands in Figure 11 are, however, relatively wide, and the power of our instruments to detect a macroeconomic impact of agency mortgage purchases beyond the housing sector is rather limited.

Housing Credit Policy vs. Conventional Monetary Policy
In the previous section, we found that increases in agency mortgage purchases lead to an expansion in mortgage credit and to declines in short-and long-term interest rates. A natural question to ask is to what extent these effects reflect conventional monetary policy actions, and how monetary and credit policies interact more broadly. The left panel in Figure 12 reports the estimated response of the federal funds rate to an agency purchase shock using the methods of the previous section. According to the narrative instrument, there is a delayed and transitory decline in the funds rate by up to 30 basis points after 6 months. This decrease is statistically significant after 4 to 12 months. The GSE excess returns instrument also yields a temporary decline in the funds rate, but it is smaller in size and not statistically significant.
We obtain similar results for the post 1982 subsample, after excluding the non-borrowed reserves targeting period, or after omitting larger policy events from the narrative instrument (see online appendix A.4).
There is therefore evidence that agency mortgage purchases are accompanied by accommodative monetary policy. A possible alternative interpretation is that both identification schemes erroneously pick up the influence of recessionary shocks causing downward adjustments in the Federal Reserve's interest rate target.
However, if this were the case, we would not expect to find increases in strongly procyclical variables such as mortgage originations or housing starts. To gain more insight into the nature of the funds rate response, we make use of the decomposition by   The middle panel in Figure 12 depicts the estimated response of the cumulative  shocks to an agency purchase shock using the regressions in (4) and (5). With a few months delay, the narrative instrument yields a significant and persistent decline by up to 10 basis points. The funds rate decline is therefore not explained by inflation and output considerations alone, and possibly reflects also an inde- 29 We use the updates by Wieland and Yang (2016) to extend the sample of the original series. pendent reaction to credit market conditions and/or credit policies. We note, however, that the GSE excess returns instrument does not yield a similar significant decline in the  residual.
To investigate whether monetary policy affects housing credit policy, the right panel in Figure 12 reports the response of the cumulated narrative measures of credit policy changes in Table 2, deflated by the core PCE price index and as a percentage of trend originations, to a monetary shock. There is no evidence for monetary policy shocks impacting the non-cyclical measure of agency purchase shocks, as our narratively identified housing credit policy instrument is not itself predictable by the  residuals.
This provides assurance that our narrative instrument does not erroneously pick up the effects of monetary policy shocks. Similarly, adding the current and lagged values of the  shocks as additional control variables in (4) also has very little effect on the results, see online appendix A.4. The cyclical housing policy measure (in red), on the other hand, does show a statistically significant decline following an expansionary monetary policy shock, which illustrates the importance of accounting for the endogeneity of credit policies. Consistent with an objective of stabilizing credit flows, we thus find that housing credit policies on average act to offset the effects of monetary policy disturbances.
To further judge the extent to which agency purchase shocks operate through more conventional monetary transmission channels, Figure 13 compares the impact of a traditional monetary policy shock (in red) with the response to the agency purchase shock identified using the narrative instrument (in blue). The response to a monetary shock is obtained by similar regressions as in equation (4), but replacing the agency market share on the right hand side with the contemporaneous level of the 3-month T-bill rate, and using the  shock measure as an instrument. 30 In the figure, the impact of the interest rate shock is scaled such that the maximum decline in the 3-month T-bill rate is the same as for the agency purchase shock identified with the narrative instrument. For easier comparison, the responses to the monetary policy shock in Figure 13 are shifted forward by 4 months such that the maximum interest declines for each of the policy shocks coincide. As before, the bands are 95 percent Newey and West (1987) intervals. 30 Conditional on including an informationally sufficient set of lagged variables as controls, valid identification under this approach requires only contemporaneous exogeneity of the  shocks. The predictability of the  shocks by agency purchase shock therefore does not necessarily invalidate the identification of the response to monetary shocks. There are, however, also some notable differences between the responses in Figure 13. The first is that agency purchases lead to a rise in originations that is roughly twice as large as that of the interest rate shock. There is little indication that a conventional monetary policy shock causes a significant rise in real house prices, while the decline in long-term interest rates is slightly more pronounced and persistent after an agency purchase shock. Both the rise in housing starts and mortgage debt, on the other hand, are very similar for both policy shocks. Taken together, the results indicate that agency purchases have a larger effect on mortgage repayments than conventional interest rate policy. In appendix A.3, we compare the responses of refinance and purchase originations. Whereas purchase originations respond very similarly to both shocks, refinancing originations react more strongly to the agency purchase shock, and account for the entire difference in the effect on total originations. Another notable difference between credit policy and traditional interest rate shocks is the effect on the homeownership rate (right panel, third row in Figure 13). Unlike the response to an agency purchase shock, there is no indication that a conventional interest rate shock has any positive effect on homeownership. In most months, the estimated effect on homeownership instead is negative, though small and generally not sta-tistically significant. Apart from the different response of originations and homeownership, however, it does appear as if credit policy operates through similar transmission channels as conventional monetary policy.
In online appendix A.5, we compare agency activity and conventional monetary shocks in terms of their contribution to fluctuations in credit aggregates and interest rates. Because our local projections approach is not well suited for this purpose, we assess the variance contributions in an SVAR setting using the GSE excess returns identification strategy and the  residuals as a proxy for monetary shocks. The main finding is that GSE excess returns shocks explain up to 15% and 10% of the medium-run forecast error variance of mortgage originations and housing starts, respectively, which is roughly comparable to the contribution of monetary policy shocks. In addition, while shocks to monetary policy are substantially more important for the variance of interest rates in the short run, the role of GSE excess returns shocks for long-term interest rates exceeds the role of monetary policy shocks at longer horizons. The SVAR-based analysis overall indicates that the contribution of credit policy shocks to fluctuations in housing and credit markets is non-negligible.
To explore the potential effects of agency mortgage purchases when conventional interest rate policy does not respond, for instance because it is constrained by the zero lower bound, Figure 14 reports the results from a counterfactual experiment in which the short-term interest rate is assumed to remain constant. As before, the responses are to an increase in anticipated agency purchases by one percentage point of trend originations, as in (4) and (5). However, we now additionally assume the realization of a sequence of monetary shocks such that the 3-month T-bill rate remains unchanged at every horizon. 32 An important caveat with this experiment is that the short-term rate remains constant because of successive monetary surprises rather than an anticipated policy response. As such, the results are clearly subject to the Lucas critique. Figure 14 shows the counterfactual responses in red and the earlier baseline estimates in blue, in both cases with 95 percent Newey and West (1987) bands. Panel A shows the results when using the narrative instrument to identify the effects of agency purchase shocks, and panel B when we use the GSE excess returns instrument.
The results from the counterfactual experiment in Figure 14 indicate that conventional monetary policy plays an important role in explaining the effects of agency purchase shocks. Under both identification schemes, the rise in originations is only about half as large when short-term interest rates remain constant, and there is no longer any sign of an increase in the stock of mortgage debt. The drop in long-term interest rates is much reduced with the narrative instrument, but remains more clearly present with the excess returns instrument.
The positive effect on housing starts is smaller using the GSE excess returns instrument, and disappears entirely with the narrative instrument. The combination of expansionary monetary and credit policy therefore seems particularly important for stimulating residential investment. Even with constant interest rates, however, purchases of mortgage assets continue to have statistically significant effects on mortgage lending, regardless of the instrument. In addition, the path of short-term interest rates appears largely irrelevant for the increase in the homeownership rate following a shock to agency purchases.

Concluding Remarks
The postwar period witnessed a remarkable expansion in residential mortgage debt. During the same period, an increasing share has come to reside on what is ultimately the balance sheet of the federal government.
In this paper, we provide evidence that government mortgage purchases influence the volume and cost of mortgage lending. In order to tackle reverse causality, we make use of a number of policy changes that have impacted the ability of government agencies to acquire mortgage debt. Using policy interventions that we classify as non-cyclically motivated to construct an instrumental variable for (news about) agency mortgage purchases, we find that an increase in these purchases stimulates mortgage originations and debt, and temporarily lowers mortgage rates. Consistent with the evidence in Di  regarding the effects of the QE interventions, we find that agency purchases have particularly large effects on refinancing activity.
We also find a positive impact on housing starts and homeownership, and some indications of positive effects on house prices and consumption expenditures. An alternative identification strategy based on instrumenting with orthogonalized innovations in GSE excess returns yields very similar results overall.
One important aspect of our findings is the apparent similarity and interaction between housing credit policies and conventional interest rate policy. We find that greater agency mortgage purchases lead to broad declines in short and long-term interest rates. Our measure of non-cyclically motivated credit policy changes predicts the  monetary policy shock measure, and expansionary credit policy appears to be accommodated by monetary policy. In contrast, we find that credit policy adjusts in order to offset the effects of monetary disturbances. It may therefore be necessary to account for credit policy to understand the effects of monetary policy. Agency purchase shocks have relatively larger effects on mortgage originations and refinancing activity than interest rate shocks, and influence homeownership regardless of the path of short-term interest rates. The quantitative effects of credit and monetary policy shocks on many other variables, including residential investment, are otherwise remarkably similar.
There are several interesting avenues for future research: Unlike theoretical or multivariate statistical models, our local projections/IV-approach does not allow an assessment of the historical contribution of structural shocks without further assumptions. In online appendix A.5, we apply our GSE excess returns identification strategy in an SVAR setting, and we find that the contribution to the short-run variability of mortgage credit and housing starts is substantial and similar to that of monetary policy shocks. Future work can verify in more detail whether credit policy shocks are important causal factors in past housing or credit cycles, in particular during the most recent housing boom and bust. 33 Our results can be used to help evaluate the credit policy interventions in the recent financial crisis, the possible impact of unwinding the Fed's current mortgage holdings, or the various proposals for GSE reform. We have made no attempt at understanding more precisely the nature or implications of the credit frictions and transmission channels through which housing credit policies operate. Future work may apply similar cross-sectional identification strategies as Di , Darmouni andRodnyansky (2016), or Chakraborty, Goldstein, andMacKinlay (2016) to other housing credit policy events documented in our narrative analysis. Finally, it is possible to apply a similar analysis to government mortgage guarantees and securitization. 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 1955 1960 1965 1970 1975 1980 1985 1990 1995 1955 1960 1965 1970 1975 1980 1985 1990 1995  Notes: Residential mortgage debt and originations include home as well as multifamily mortgages. Agency holdings include holdings of both whole loans and pools. Agency purchases are net purchases for portfolio investment, whereas pool issues approximate purchases backing new issuance of mortgage pools (mortgage-backed securities). The grey bars are NBER-dated recessions. Sources: see data appendix.  1960Q1, 1969Q4, 1973Q4, 1980Q1, 1981Q3, 1990Q3, 2001Q1, 2007Q4. Right panel: average of credit crunches beginning one quarter after the following dates: 1955Q3, 1959Q1, 1965Q4, 1968Q4, 1972Q4, 1978Q1, 1980Q4, 1990Q1, 1998Q2, 2007Q2. See data appendix for sources.

Figure 5 Preliminary Diagnostics
Notes: The left panel shows Newey and West (1987) robust F-statistics of the first-stage regressions of cumulative agency commitments and purchases, respectively, on the narrative instrument. The right panel shows the estimated dollar increase in agency purchases per dollar increase in commitments. Finer lines in the right panel are 95% Newey and West (1987) confidence bands. Sample: Jan 1967 to Dec 2006.   (2), comparing OLS and 2SLS estimates. The specification with baseline controls excludes the interest rate and cyclical controls.

Historical Background
This appendix provides some more historical background to the evolution of agency market shares depicted in Figure 1. 34 During the Depression, the Home Owners' Loan Corporation took ownership of nearly 15% of mortgage debt. Housing and homeownership reemerged as a priority at the end of WWII, which is reflected in the strong growth of Fannie holdings in the late 1940s until the Korean War again shifted priority away from housing. A struggling Fannie was rechartered as a mixed private-public ownership corporation in 1954. In 1968, Fannie Mae was split into a publicly listed private corporation and a government-owned Ginnie Mae. In the 1970s, Fannie expanded almost without interruption and the agencies mortgage holdings reached close to 10 percent of total mortgage debt. However, Fannie's large debt-financed balance sheet incurred heavy losses after interest rates rose sharply in 1979. Profitability was only restored through a strategy of aggressive portfolio expansion and by entering the securitization business. At its creation in 1970, ownership of Freddie Mac was restricted to the savings and loans, which had no interest in creating a competitor. As a result, Freddie focused on the securitization of conventional loans, maintaining only a relatively modest mortgage portfolio for warehousing until the late 1980s. In the second half of the 1980s, rising delinquencies and a more hostile attitude of the Reagan administration towards the GSEs led to a reduction in the agencies' market share.
34 Table 1 contains references to various books an articles that contain more comprehensive overviews.
Various reforms in the aftermath of the 1980s S&L crisis set the stage for a prolonged rise in agency activity in the 1990s and early 2000s, and by 2002, the agencies held close to one quarter of the total outstanding mortgage debt on their portfolios. Before, in 1989 Freddie was turned into a publicly traded company with the same profit incentives for balance sheet growth as Fannie, while the Federal Home Loan Banks were granted permission to invest in MBS. Prudential regulations were tightened for private banks, but remained light for the GSEs despite a 1992 reform. The agencies increasingly retained their own and acquired each other's MBS, as opposed to selling them to private investors. As part of an ambitious homeownership strategy, the Clinton administration was supportive of the efforts by Fannie and Freddie to develop automated underwriting systems and ramped up affordable housing goals for their purchases.
The rapid rise in agency ownership of mortgage debt increasingly became a cause of concern with public officials, and in the wake of the Enron scandal Fannie and Freddie were required to start filing reports with the Security and Exchange Commission. The 2008 conservatorship agreement allowed for continued GSE balance sheet growth in the short run, but also mandated a long-run wind-down of their portfolios at an annual rate of 10%, increased to 15% in 2012, until they reach a $250 billion each. The day after the agreement, the Treasury announced its own MBS purchase program, while the Federal Reserve's MBS program was launched a few weeks later. As a result of the Fed and Treasury programs, the combined agency ownership share regained levels similar to the early 2000s despite a gradual decline in holdership by the traditional housing agencies. In contrast, Fannie and Freddie have been allowed to grow their MBS guarantee book essentially without limits. Since the financial crisis, the vast majority of conforming loans originated have been acquired, guaranteed, and sold off in MBS by the agencies.

Data Sources and Construction
Data underlying Figure  The upper left panel of Figure 1 shows annual data up to 1952 and quarterly data afterwards. Missing quarterly data on FHLB holdings is obtained by linear interpolation of annual data. Residential mortgage originations shown in the lower left panel of Figure 1 is the quarterly aggregate of the monthly series described below.

Agency Net Portfolio Purchases and Pool
Issues is the sum of net portfolios purchases of both whole loans as well as mortgage pools, and of issues of mortgage pools respectively, by Fannie Mae, Freddie Mac, Ginnie Mae, the FHLBanks, the Treasury Department, the Federal Reserve, and a number of other government agencies: Fannie Mae: Monthly data on Fannie's net portfolio purchases from 1953 to 1998 is from various issues of the Federal Reserve Bulletin (portfolio purchases less sales). More recent data is from Fannie's monthly volume summary cross-checked with the annual reports from OFHEO/FHFA for consistency. While data on purchases is available over the entire sample, data on portfolio sales is missing for 1986 and 1988-1997. We impute the missing observations using data on Fannie's commitments to purchase and sell, actual purchases, and the net change in the retained portfolio. The imputation is done by Kalman smoothing in a state space model estimated by maximum likelihood as in Shumway and Stoffer (1982) using monthly data from 1980 to 2014. The model used is a vector autoregressive process for the net portfolio purchase rate, retained mortgage portfolio growth, and the ratio of purchases and net commitments to the retained portfolio. Data on Fannie pool issues from 1993 is from Lehnert, Passmore, and Sherlund (2008), extended to 2014 using Fannie's monthly volume summaries. Pre-1993 monthly data is obtained by subtracting Freddie and Ginnie pool issues from total net purchases by agency mortgage pools from monthly releases by the Department of Housing and Urban Development from the Survey of Mortgage Lending Activity (obtained through the National Archives and Records Administration).
Freddie Mac: Monthly data on Freddie's net portfolio purchases from 1993 onwards is from Lehnert, Passmore, and Sherlund (2008) and Freddie's monthly volume summaries. Data before 1984 is obtained by subtracting Freddie pool issues from total wholesale loan purchases available from the Federal Reserve Bulletin. Data between 1984 and 1993 is imputed using data on Freddie holdings and repayment rates in Fannie's portfolio. The imputation is done by Kalman smoothing in a state space model estimated by maximum likelihood as in Shumway and Stoffer (1982) using monthly data from 1980 to 2014. The model used is a vector autoregressive process for Freddie's net portfolio purchase rate, retained mortgage portfolio growth and repayment rates in Fannie's retained portfolio. Monthly data on Freddie pool issuance is from the journal of the Federal Home Loan Bank Board (various issues, 1971-1980), the Federal Reserve Bulletin (1980)(1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998), and the monthly volume summaries (1998 onwards).
Ginnie Mae: Monthly data on Ginnie's net portfolio purchases from 1968 to 1971 is from various issues of the Federal Reserve Bulletin. Subsequent data is imputed by assuming that repayment rates for mortgages packaged in pools backed by Ginnie are the same as for mortgages held in portfolio.
Monthly data on Ginnie pool issues since 1968 was provided to us directly by the Department of Housing and Urban Development.
FHLBanks: Data on net purchases by the FHLBanks is imputed using net changes in holdings and assuming that the combined repayment rate on mortgages debt in Fannie, Freddie and Ginnie pools is identical to the repayment rate on mortgages in mortgage-backed securities held by the FHLBanks.
Federal Reserve: Data on MBS purchases using the date of settlement, available from the Board of Governors https://www.federalreserve.gov/newsevents/reform_mbs.htm and the Federal Reserve Bank of New York https://www.newyorkfed.org/markets/ambs/ambs_schedule.html.
Other Agencies: Data on combined net purchases by the other agencies is imputed using net changes in holdings and by assuming that the combined repayment rate on mortgages debt in Ginnie pools is identical to the repayment rate on mortgages in mortgage-backed securities held in portfolio.
The lower right panel of Figure 1 shows quarterly data from 1952 onwards.
Data underlying Figure 2: Agency mortgage holdings is the quarterly series from Figure 1. Private mortgage holdings is total residential mortgage debt from Figure 1 less agency holdings. Both series are deflated by the price index for personal consumption expenditures excluding food and energy from NIPA (series PCEPILFE from the FRED database at the Federal Reserve Bank of St. Louis). The chronology for pre-1986 credit crunches is from Eckstein and Sinai (1986). The dating of post-1986 crunches is based on Owens and Schreft (1993) for the 1990 commercial real estate crunch), Lehnert, Passmore, and Sherlund (2008) for the 1998 Russian default/LTCM crisis, and Bordo and Haubrich (2010) for the 2007 Financial Crisis.
Monthly agency data: The monthly series for consolidated agency mortgage holdings and net portfolio purchases sums the monthly series for Fannie, Freddie, Ginnie, the Federal Reserve, and the Treasury described above (see data underlying Figure 1). All series are seasonally adjusted using the X-13 program from the Census Bureau. Agency purchase commitments are the sum of the following series: No data for net commitments is available for Ginnie Mae and the Treasury, and we simply use the series for net portfolio purchases.
Monthly mortgage market data: The conventional mortgage rate is the 30-year fixed-rate conventional conforming mortgage rate. From 1971 onwards, the conventional rate is the monthly average commitment rate from the Freddie Mac primary mortgage market survey. Pre-1971 data is from the Federal Housing Adminstration (FHA)/Departement of Housing and Urban Development (HUD) series for the primary conventional market rate, available from the Federal Reserve Bulletin (various issues).
The FHA mortgage rate is the 30-year fixed-rate FHA-guaranteed mortgage rate. Rate data for FHAmortgages offered in the secondary market from 1963 is provided by FHA/HUD and is available from various issues of the Federal Reserve Bulletin. Earlier data is from the NBER's macrohistory database (series m13045). The series has a handful of missing observations and was discontinued in 2000. We impute data by Kalman smoothing in a VAR/state space model estimated by maximum likelihood as in Shumway and Stoffer (1982) using several closely related interest rate series over the 1976-2014 period: the conventional 30-year rate (FHA/HUD as well as the Freddie Mac series), the 3-month and 10-year Treasury rates, and yields on Ginnie Mae securities (from the Federal Reserve Bulletin as well as the MTGEGNSF Index from Bloomberg). A couple of missing observations prior to 1976 were imputed in a similar fashion using data on the 3-month and 10-year Treasury rates, on interest rate data provided by Saul B. Klaman's 1961 NBER publication "The Postwar Residential Mortgage Market", and on interest rate ceilings on FHA loans applicable at the time. The 10-year and 3-month Treasury rates are from the FRED database (GS10 and TB3MS).
The primary source of monthly data on residential mortgage originations are monthly news releases from the Survey of Mortgage Lending Activity (SMLA) conducted by HUD from 1970 to 1997, accessed through the National Archives and Records Administration (Tables 2 and 3: total originations of long term mortgage loans for 1-to-4 nonfarm homes and multifamily residential properties). The monthly series is interpolated after 1997 using quarterly data on originations (series USMORTORA in Datastream) and weekly data on mortgage applications (series MBAVBASC on Bloomberg), both from the Mortgage Bankers' Association (MBA). The interpolation is done through Kalman smoothing of an estimated VAR/state space model as in Shumway and Stoffer (1982). Observations before 1965 are based on data of total new non-farm mortgages of $20,000 or less recorded from the Federal Home Loan Bank Board and available from the NBER's macrohistory database (series m02173). To obtain an estimate of total originations, we assume that the share of originations of $20,000 or less in all originations is the same as the share in originations by Savings & Loans associations. Data on S&L originations (total and $20,000 or less) is available from various issues of the Savings and Home Financing Sourcebooks, a publication by the Federal Home Loan Bank Board up prior to 1990. Data between 1965and 1970 is imputed using total originations by S&L associations based on Kalman smoothing in a VAR/state space model estimated as in Shumway and Stoffer (1982) using monthly data from 1954 to 1985. The series is seasonally adjusted using the X-13 program from the Census Bureau. Unfortunately, the monthly SMLA releases do not contain information on the purpose of the mortgage loans. However, the Savings and Home Financing Sourcebooks published prior to 1990 contain monthly data on refinancing originations by S&L banks (although observations from May 1985 to December 1986 are missing). After 1990, quarterly totals of refinancing originations are available from the MBA (series USMORRVLA in Datastream). As an estimate of the share of refinancing loans, we use the monthly shares at S&L banks before 1990, and the quarterly shares from the MBA afterwards. Our monthly series on refinance and purchase originations are obtained by applying the estimated share of refinancing to our series for total residential mortgage originations.
The monthly series for mortgage debt is based on interpolation of the quarterly mortgage debt series from the Financial Accounts of the United States (see Figure 1) using the series on monthly originations. The series is constructed by linear interpolation of the implied quarterly repayment rates. The final series is seasonally adjusted using the X-13 program from the Census Bureau.  (2015), and seasonally adjusted using the X-13 program from the Census Bureau. The series is deflated by the nominal price level, measured by the core PCE price index to obtain a real house price index (series PCEPILFE from FRED). To the best of our knowledge, no monthly data on the homeownership rate is available. We therefore simply use quarterly values of series RHORUSQ156N from FRED. Monthly personal consumption expenditures is from NIPA (series PCE from FRED). Monthly personal income is from NIPA (series PI from FRED). The unemployment rate is series UNR from FRED. The short and longterm nominal interest rates 3-month and 10-year Treasury rates are series TB3MS and GS10 from FRED. This section discusses a number of robustness checks of the results presented in Section 5 regarding the cumulative effects of agency purchases.

Scaling by Trend Originations
The baseline specification in (2) uses a trend in personal income as the scaling variable. Figure  in originations, implying that the agency portfolio purchases are instead of whole loans. MBS issuance gradually rises, but the total cumulative increase is a smaller share of the total increase in originations. This pattern is more similar to agency behavior before the growth of mortgage securitization in the mid-1980s.  We highlight the following patterns from the results in Table A.1. First, the point estimates across the IV regressions are all quite similar. Controlling for interest rates is the most consequential. When leaving out interest rates in [2], we find somewhat smaller increases in mortgage originations and debt. The results are essentially unchanged by including the cyclical controls (unemployment and income growth). Interestingly, and conditional on including the richest control set as in our benchmark specification, the results remain similar when we also include the cyclically motivated policy events in the instrument, see column [8]. This is in our view not too surprising. Based on our reading of the various historical policy actions, see Fieldhouse and Mertens (2017), recognition and decision lags likely exceed one month in practice. With a sufficiently rich set of lagged controls, including the cyclical actions may therefore not lead to any meaningful violation of the exogeneity requirements. Instrumentation with policy events, however, is important for the results. The in originations that appears after instrumentation is more consistent with a causal interpretation. Figure A.3 also reveals that the total agency mortgage holdings increase by a substantially smaller amount than the dollar purchased or committed, even at relative short horizons. This indicates that agency purchases tend to coincide with higher sales of mortgage assets to private investors and/or with higher repayment rates, both of which are likely to reflect other influences on the private demand for mortgage credit in the primary or secondary market. All variations of the baseline specification reported in Table A.2 yield cumulative origination multipliers in the range of 2.5 to 4.5 after 3 to 4 years. Moreover, the impact on originations is consistently highly statistically significant. The estimated cumulative change in mortgage debt also remains in the range of the benchmark specification. The credit multipliers are the lowest when we extend the sample to include the recent financial crisis (column [2]) and when we add GSE excess stock returns to the control set (column [8]).
In these cases, the impact on mortgage debt is no longer significant at conventional levels. We also highlight that the inclusion of the September 2004 policy event is important for the precision of the estimates. The

A.2 Additional Results on Interest Rates and Credit Spreads
This section discusses a number of additional results regarding the effects of news shocks to agency purchases on interest rates and credit spreads. Figure  The first two panels in Figure A.4 show the responses of the AAA-rated and BAA-rated long term corporate bond yields. Taken together, the results suggest that agency purchases exert a downward pressure on corporate yields with a timing that coincides with the actual purchasing of mortgage assets by the agencies.
The response of the corporate yields is qualitatively similar to those of mortgage and Treasury rates, showing initially no effect, and subsequently a gradual decline. The 95% confidence bands around the responses are relatively wide, and the responses are only marginally significant in the case of the narrative instrument.
The declines in corporate bond yields are also quantitatively smaller than mortgage and Treasury rates, in particular for the narrative instrument. The third panel in Figure A.4 shows statistically significant shortrun increases in the spread between AAA-rated corporate bonds and 10-year Treasuries. Agency purchases appear therefore to induce the greatest spill-overs on the demand for the relative liquidity and safety of Treasuries, which do not have prepayment risk. The increases are, however, relatively short-lived, with the effects dissappearing after 7 or 8 months in the case of the narrative instrument, and only after a few months in the case of the GSE excess returns instrument. The next panel shows evidence for a drop in the spread between BAA and AAA-rated corporate bonds after 7 or 8 months, suggesting also some positive spill-over effects on the demand for riskier long-term bonds.
The middle and right panels in the second row of Figure  spread increases by 5 basis points or more when using the narrative instrument for identification. While this may indicate some upward pressure on short-term interest rates relative to the funds rate target, the increase is not statistically significant at conventional confidence levels. The GSE excess returns instrument does not yield any indication of an impact. Finally, the right panel on the bottom row shows the response of the slope of the Treasury yield curve, measured by the difference between the 10-year and 3-month yields. In the case of the narrative instrument, there is initially no impact on the slope of the yield curve. However, as the agency purchases induce a broad-based decline in both long and short-term interest rates, the slope of the yield curve gradually steepens for the first 8 months. Subsequently, the yield curve flattens relatively quickly and returns to prior levels one year after the new shock. In contrast, the GSE excess returns instrument does not produce any significant impact on the slope of the Treasury yield curve.

Figure A.4 Impulses Responses to A Shock to Anticipated Agency Purchases
Notes: The figure shows responses to a one pp. increase in the expected future agency market share measured by agency commitments as a ratio of trend originations. Estimates are from local projections-IV regressions instrumented with the (1) narrative policy indicator, see equation (4), or orthogonalized GSE excess stock returns innovations, see equation (5). Finer lines and shaded areas are 95% Newey and West (1987)   nancing the purchase of a home. We note that for the estimates in these two panels, the sample excludes May 1985 to December 1986 because of missing data on refinance shares, see data appendix.
The results in Figure A.5 are consistent with those for the dollar credit multipliers reported in Section 5.
Refinancing originations show a gradual increase following the agency purchase shock, regardless of the instrument used for identification. Purchase originations also rise, but with a longer delay relative to refinance originations. Using the narrative instrument, purchase originations are initially lower for the first six months or so, before rising by similar levels between 12 and 24 months as for the GSE excess returns instrument. Using the average share of refinancing originations of 25% over the benchmark sample, the estimates in Figure   A.5 imply that refinancing originations account for the larger share of the increase in total originations.

Figure A.5 Impulse Responses to a Shock to Anticipated Agency Purchases
Notes: The figure shows responses to a one pp. increase in the expected future agency market share measured by agency commitments as a ratio of trend originations. Estimates are from local projections-IV regressions instrumented with the narrative policy indicator, see equation (4), or orthogonalized GSE excess stock returns innovations, see equation (5). Finer lines and shaded areas are 95% Newey and West (1987)   response to the agency purchase shock identified using the narrative instrument (in blue). Responses to monetary shocks are identified using the  monetary policy shock measure as an instrument, as explained in Section 7. As in Figure 13, the impact of the interest rate shock is scaled such that the maximum decline in the 3-month T-bill rate is identical as for the agency purchase shock identified with the narrative instrument. The left panel repeats the responses of total mortgage originations shown in Figure 13 and shows that the agency purchase shock generates a larger increase in total originations. The middle panel shows the responses of refinancing originations, while the right panel shows the estimated responses of originations financing the purchase of a home. The results in Figure A.6 indicate that the differential impact on total originations is due to the different impact on refinancing activity. The response of purchase originations (right panel) is very similar in timing and size across both shocks. The response of refinancing originations to monetary policy shock, on the other hand, is much more muted than the response to the agency purchase shocks.

Figure A.6 Responses to A Shock to Anticipated Agency Purchases Versus a Monetary Policy Shock
Notes: The figure shows responses to a one pp. increase in the expected future agency market share as well as the response to a monetary policy shock. Estimates are from local projections-IV regressions instrumenting agency commitments with the narrative policy indicator, see equation (4), and instrumenting the 3 month T-Bill rate with the Romer and Romer (2004)

A.4 Impulse Response Analysis: Sensitivity Checks
Omitting the 1977-1982 and NBR-targeting Periods Figure A.7 shows the response to a shock to anticipated agency purchases by one percentage point of trend originations, together with 95% Newey and West (1987)  Omitting Policy Events from the Narrative Instrument Figure A.9 shows the response to a shock to anticipated agency purchases for the benchmark specification together with those when we omit in turn each of the three largest policy interventions from the narrative instrument: the October 1977 conforming loan limit increase and expansion of the Brooke-Cranston Tandem program, the December 1982 increase in Fannie Mae's debt-to-capital limit, and the September 2004 tightening of capital requirements. In each case we add the omitted event as a separate dummy variable, including both the contemporaneous value and twelve lags to the control variables. While there is some variation in the size of the responses, the results remain qualitatively similar to the benchmark narrative estimates. In all cases, there are increases in originations and mortgage debt, and declines in short-and long-term interest rates.
Including  Shocks as Controls Figure A.10 compares the benchmark narrative impulse response estimates of Figure 10 with those from a specification that includes both the contemporaneous value as well as 12 lags of the  monetary shock measure as additional controls. Figure A.10 shows that controlling for the  shocks has little effect on the estimation results.

A.5 Forecast Error Contributions from an SVAR Model
The local projections-IV specifications do not allow an assessment of the historical role of structural shocks to housing credit policy, which requires knowledge of the variance contribution of these shocks to the cumulative purchase measures in equations (2) or (4). In order to gain some insight into the importance of GSE activity for the dynamics of credit aggregates and interest rates, this section estimates the variance contribution of the orthogonalized GSE excess returns innovations in a structural vector autoregressive (SVAR) model. The main finding is that the contribution of GSE excess returns shocks to the short-run variability of mortgage credit and housing starts is roughly as important as monetary policy shocks. In addition, shocks to monetary policy are substantially more important for the forecast error variance of interest rates in the short run. The role of GSE excess returns shocks for long-term interest rates exceeds the one of monetary policy shocks at horizons beyond 18 months.
In order to estimate forecast error variance contribution of shocks to GSE activity, we adopt a VAR model for the joint dynamics of the ratio of agency purchases and commitments to trend originations, as well as all of the variables included as controls in the LPIV regressions: the log levels of core PCE and house price indices, the log difference of total mortgage debt, the log levels of real mortgage originations and housing starts, the 3-month T-bill rate, the 10-year Treasury rate, the conventional mortgage interest rate, the BAA-AAA corporate bond spread, the unemployment rate, and the log of real personal income. In addition, the VAR system also includes the log of the S&P 500 index and the log ratio of the GSE stock index to the S&P 500. We estimate the VAR by OLS using 12 lags of all the endogenous variables and using monthly data from August 1971 to December 2006. The start of the sample reflects the month in which Fannie Mae stock was traded for the first time on the New York Stock Exchange, which was August 1970.
The impact of a shock to orthogonalized GSE excess returns is the response to an innovation to the GSE stock index variable, which is obtained by taking the lower triangular Choleski decomposition of the estimated covariance matrix of the VAR residuals, ordering all of the variables except agency purchases/commitments above the GSE stock index variable. This approach imposes the same exclusion restrictions as the LPIV model in (4) within the SVAR context, which amounts to assuming that none of the variables ordered before the GSE stock index variable responds within the same month to orthogonalized GSE excess returns innovations. Figure A.11 shows the resulting impulse responses, which for ease of comparison are scaled to imply a similar 6 month impact on originations as the LPIV estimates in Figure 10. The GSE excess returns shocks lead to statistically significant increases in agency net commitments and net purchases (not shown). Consistent with our main findings, Figure 10 shows that originations, mortgage debt and housing starts all rise significantly following a positive innovations in GSE excess returns, while interest rates decline in the short run. One difference between the VAR and the LPIV estimates in Figure 10 is that interest rates rise in the longer run. The SVAR estimates are otherwise generally similar to those obtained using LPIV regressions using the GSE excess returns as an instrument for agency mortgage purchases. An advantage of the SVAR model is that it is straightforward to evaluate of the relative importance of shocks in driving fluctuations in the endogenous variables. Figure A.12 depicts the share of the forecast error variance at various horizons that is due to the identified GSE excess returns innovations. For comparison, Figure A.12 also shows the variance contribution of monetary policy shocks identified using the  measure as a proxy using the methodology in Stock and Watson (2012) and Mertens and Ravn (2013). We find that the GSE excess returns shocks explain up to 9% of the agency net purchases and commitments forecast variance (not shown). The contribution of monetary policy shocks remains below 2% at all horizons considered. Figure A.12 reveals that both shocks account for a substantial fraction of the forecast variance of originations and housing starts at horizons beyond 6 months. GSE excess returns shocks explain up to 14% of the forecast variance of originations at horizons between 12 and 18 months, and around 7% to 8% of housing starts between 8 and 14 months. In comparison, monetary shocks explain between 8% to 9% of originations, and around 14% of housing starts at similar horizons. Neither of the shocks accounts for much of the forecast variance of the stock of mortgage debt at horizons up to 36 months.
Monetary shocks account for a substantial share of the short-run forecast variance of the 3-month T-bill rates, and up to 11%, respectively 7%, of the variance in mortgage and 10-year Treasury rates at horizons around 6 months. GSE excess returns shocks are relative less important for the variability in interest rates at shorter horizons, but become relatively more important than monetary policy shocks in accounting for the uncertainty in long-term interest rates at horizons exceeding 18 months.