Do public libraries impact local labour markets? Evidence from Appalachia

ABSTRACT This paper investigates the effect of public library programmes and participation on unemployment and labour force participation in Appalachia. Appalachia is an economically distressed area, mostly rural, and with a sustained lower level of labour force participation and a higher level of unemployment. As public library programmes can be cyclical to business cycles, that is, labour market outcomes, I use public library staff and the amount computers available as instruments. While ordinary least squares (OLS) estimates show no effect of adult or children’s programmes and participation on local labour market outcomes, spatial econometric estimates provide evidence of direct and indirect effects of adults programmes and participation on labour force participation.


INTRODUCTION
The idea that public libraries are only about books amid advances in technology, such as computers and the internet, has reduced the perceived importance of public libraries to local communities. Yet, in 2014, there were 9305 public libraries in the United States, 3.9% more than in 2010. Public libraries in the United States received an overall yearly, non-unique, 4.6 in-person visits per capita and US$12.1 billion in revenue in 2014, which, compared with 2010, represents a 12% increase in yearly in-person visits and 7% increase in revenues.
To remain relevant, public libraries have been adapting their services to match new demand for services and materials (Goulding, 2006;Hunt, 2017;Jerrard, 2009). For example, public libraries have been expanding the number of programmes and resources offered. In 2014, there were a total of 4.5 million programmes offered to adults and children, with 101.9 million non-unique attendees in the country. This represented a 20% increase in the number of programmes and 17% increase in the number of attendees compared with 2010. In terms of collection materials, in 2014 there were over 1.2 billion materials covering books, e-books, video and audio. This corresponded to an increase of 29% of collection materials from 2010. The composition of these materials has also changed. In 2014, books were 66.1% of the materials and e-books were 18.4%, while in 2010, books represented 86.4% of the materials and e-books 2% (The Institute of Museum and Library Services (IMLS), 2017).
Library programmes can be roughly divided into those for children and those for adults. Children's programmes usually focus on book-related activities, educational and entertainment activities. Adult programmes focus on book activities, development of skills and job search services. 1 This paper investigates the impact of these public library programmes for children and adults and programme participation on local labour markets outcomes, in particular, unemployment and labour force participation. To evaluate the impact of the public library programmes on local labour market outcomes, I combine datasets on county demographic characteristics and labour statistics with a novel dataset on public libraries. The Public Library System (PLS) dataset is an annual survey considered to be the census of public libraries in the United States. From the PLS, I collect data on the number of programmes and programme participation, as well as a variety of information about each public library system. 2 By focusing on public library programmes, this paper first contributes to the literature of urban amenities and its effect, in particular, the effect of library use. The most relevant work in this literature is by Bhatt (2010), who finds that an increase in library usage increases time spent reading, decreases time spent watching television and, for school-age children, increases homework completion rates. Betts (1995) and Farber and Gibbons (1996) use the possession of a library card at age 14 as a proxy for innate ability, but they did not evaluate the impact of library programmes directly on wages. Further, Liu (2004) uses a cross-section of countries and finds that public libraries' literacy programmes affect economic productivity measured by gross domestic product per capita.
This paper also contributes to the local labour market literature, in particular, the active labour market programmes (ALMPs) (Ashenfelter, 1978), which has been summarized by Heckman et al. (1999) and Card et al. (2010Card et al. ( , 2018, among others. The ALMP literature mostly focuses on the government programmes created by the Area Redevelopment Act in 1961(LaLonde, 2003, and its most relevant results show the ineffectiveness of public sector employment programmes and the positive impact of job search assistance programmes. In addition, there is some heterogeneity in the results depending on the investigated outcomes, programme type and treatment groups, with larger effects for women and those who were unemployed for longer (Card et al., 2010(Card et al., , 2018Heckman et al., 1999;Kluve, 2010). Also, Card et al. (2018) argue that ALMPs have larger effects during recession times, that is, low growth and high unemployment.
Similar to some of the government programmes, most public library adult programmes have a focus on employment by helping develop new skills and finding jobs (Bertot et al., 2012;Rainie, 2016). Children's programmes, on the other hand, can have an impact on local labour markets since parents may see public libraries as possible substitutes for daycare services (Parrish, 2013;Smith & Rivera, 2004). Thus, I contribute by focusing on overlooked labour market programmes and rural areas, which have lower levels of private and public provided labour market programmes.
I restrict the analysis to the Appalachian region. Appalachia is comprised of 420 counties across 13 states 3 covering remote rural and urban areas. However, the region is mostly rural as 70% of its counties are non-metropolitan areas hosting 42% of its population (Appalachian Regional Commission, 2018;Stephens & Partridge, 2011). Appalachia is, and has been, a systematically lagging region associated with lower levels of labour force participation and higher levels of unemployment (Dorsey, 1991;Isserman & Rephann, 1993;Stephens & Deskins, 2018). For instance, this is the only region in the United States with a dedicated policy-making commission, the Appalachian Regional Commission, which has been in place for over 50 years (Isserman & Rephann, 1993;Sayago-Gomez et al., 2018).
Library programmes and programme participation, however, can be endogenous to local labour markets. If public library programmes are used as counter-cyclical policies, areas with high unemployment and low labour force participation may be more likely to have more adult programmes and fewer children's programmes, for example, rendering ordinary least squares (OLS) estimation biased. Hence, I make use of an instrumental variables (IV) approach, thus contributing to the literature by extending the Halleck Vega and Elhorst (2016) model to account for endogeneity. More specifically, I use the number of librarians without a master's degree and the number of computers for public use as instruments for the number of programmes and the participation in these programmes.
Libraries need both monetary and physical resources to promote programmes and to attract patrons. Because public library funding comes mostly from local government, this is likely to be contemporaneously correlated with local labour markets. In turn, physical resources, such as computers and books, are less likely to be contemporaneously associated with local labour markets outcomes. On the one hand, although the flow of purchases of books and computers may change during recessions and booms, the volume of these resources in the library is less likely to change over time. This should be especially true for rural areas given historical building constraints, and the reduced access to the internet and newer technologies such as e-readers and computers by patrons (Real & Rose, 2017;Swan et al., 2013).
On the other hand, librarians without a master's degree are less likely to take on managerial positions, hence being responsible to offer and run programmes, which should influence the quantity and selection into specific programmes. Since most public librarian positions require a Master of Library Science (MLS) or a Master of Library and Information Science (MLIS) degree, preferably from a school accredited by the American Library Association (ALA), the number of librarians without a master's degree is the result of past public library employment policies that are uncorrelated with current labour market outcomes.
Spatial econometric estimations that account for spatial dependence and possible spillovers find suggestive results that there are some direct and indirect effects from adult programmes and participation on the labour force participation. These results are consistent with those in the ALMP literature in that programmes are largely ineffective. Data limitation in the lack of participants data on public library programmes and on the programmes themselves do not allow for additional analysis closer to those in the ALMP literature that explore both individual and programme heterogeneous effects. Hence the analysis is at an aggregate level. The spatial econometric results are especially important in light of evidence that job search service assistance benefits participants at the expense of those who do not participate in such programmes (Gautier et al., 2017).
With the recent trend of budget cuts to public libraries and the shift of public library focus to programmes to help local communities, it is important to understand the effectiveness of public library programmes. This is the first attempt to explore such questions, even though data limitations prevent the analysis at a more granular level.

PUBLIC LIBRARIES IN THE UNITED STATES
Public libraries are usually taken for granted (Dubner, 2007). In the United States they started as privately financed institutions that offered book-lending services. Public libraries have been, and are still, valued by patrons (Wiegamd, 2015). According to the Pew Research Center, Black and Hispanic populations, as well as students, job seekers and people without internet access at home, value public libraries services the most (Pew Research Center, 2013b. In turn, parents, more educated people and the high-income population are more likely to use public library services (Pew Research Center, 2013c. However, there is an overall lack of knowledge of the services public libraries offer (Bertot et al., 2012;Pew Research Center, 2013a, 2013bRainie, 2016).
Nevertheless, public libraries are little studied by economists and policy scholars. 4 The first economic study on public libraries is by Tiebout and Willis (1965), who discuss the public nature of public libraries. Most studies that followed can be classified into two strands: one focusing on demand, unit costs and cost-benefit analysis (DeBoer, 1992;Feldstein, 1976;Getz, 1980;Goddard, 1970;Hammond, 1999;Pfister & Milliman, 1970;Stratton, 1976); and another focusing on the technical efficiency of public libraries (Ferreira Neto & Hall, 2018;Hemmeter, 2006;Sharma et al., 1999;Vitaliano, 1997Vitaliano, , 1998. Few studies analyse the impact of public libraries on different outcomes. For instance, using an IV approach, Bhatt (2010) finds that an increase in library use increases time spent reading, decreases time spent watching television and, for children at schools, it increases homework completion rates. Fujiwara et al. (2017) use a survey of users and non-users of public libraries in the UK showing a positive association between public library use and self-reported happiness and health status. Conversely, Ferreira Neto (2018) studies the impact of government funding on private donations to public libraries in the United States, finding suggestive results of a crowd-in effect. In terms of the labour market, the research on the impact of public libraries is scarce. Stine (2008) investigates the effect of volunteer workers on public libraries' demand for labour, and finds a complementary relationship between volunteer work and library staff.

DATA
3.1. Labour market outcomes Unemployment and labour force participation data come from the Local Area Unemployment Statistics (LAUS) provided by the Bureau of Labour Statistics (BLS). I follow the BLS and define the labour force participation rate as the ratio between labour force status and the population over 15 years old. Table 1 provides summary statistics for all US counties and those in Appalachia.
Similar to previous studies, Appalachian counties have lower labour force participation and higher unemployment compared with other US counties. Table 1 also splits the counties into those with and without a public library system. In both Appalachia and the United States, counties with a public library system have, on average, lower unemployment. However, while in the United States these have higher labour force participation, in Appalachia they have lower labour force participation. Although these groups are not directly comparable, this shows suggestive evidence on the uniqueness of Appalachia with respect to its labour market. This uniqueness of Appalachia is multifaceted. Durlauf (2012) points to poverty traps to explain Appalachia's persistent poverty and inequality, in particular focusing on educational attainment and migration pattern issues. Betz and Partridge (2012) point out that migration in Appalachia has different effects compared with the rest of the United States in that economic growth attracts lower skilled migrants. As noted by James and James (2015), Appalachia has been dependent of its natural resources; however, it can be a heterogenous region with subregional differences and concentration of self-employment (Stephens & Partridge, 2011). Kahn (2009) adds that Appalachia misses large cities, and its urban centres are far from high-amenity areas. In addition, the region has difficulties in attracting firms and retaining talent. Lastly, Bollinger et al. (2011) concludes that Appalachia suffer from 'missing markets', that is, the lack of high skilled labour and low returns to skill.
Thus, according to the ALMP literature, labour market programmes should be effective in areas such as Appalachia, with low-skilled workers and those who have been unemployed for longer periods. However, rural areas have a lower number of private and public labour market programmes (Dunham et al., 2005;Green et al., 2003;Whitener, 1991). In addition, rural areas have lower levels of internet access, which is an important tool in today's labour market (Hampton, 2018;Stenberg et al., 2009). Thus, in such areas, public libraries could bridge this gap by offering both some labour market programmes and internet access.

The Public Library Survey (PLS)
Information from public libraries come from the PLS. The PLS has been collected annually since 1988 covering approximately 9300 public library systems comprising over 17,000 individual public library outlets (central library, branches and bookmobiles). The survey covers all 50 US states, including the District of Columbia, and outlying territories, and has over a 98% rate of response. As such, it is considered the census of public libraries in the United States (IMLS, 2018).
The Institute of Museum and Library Services (IMLS) reports that no governmental programme is attached to the PLS, and it is not mandatory. Therefore, there are no incentives for over-or underreporting information provided, which covers several features including location, administrative data such as staff information, revenue by source, expenditures, etc.; and service and use, such as circulation, visits, programmes, materials, etc. Until 2005, the PLS was collected by the Institute of Education Sciences and the US Department of Education. Since 2006, the survey has been collected by the IMLS. Since 2009 the PLS has reported the rate of response per state. Appalachian states have a 100% response rate, with the exception of Pennsylvania which had an average response rate of 99.6%.
The variables of interest are the adult and children library programmes. The PLS collects data on the number programmes and participation in these programmes, and reports these data for all (total) programmes, children's programmes and, since 2009, young adult programmes. Ideally the PLS would record not only the total number of programmes and participation, but also the repeated participation in these programmes. Unfortunately, as this is not the case, I am unable to differentiate between extensive and intensive margins of public library use.
For the analysis, I calculate the number of adult programmes as the difference in the total programmes from children's programmes. Similarly, the participation in adult programmes is the difference in total programme attendance minus children's programme attendance. Because the number of programmes and participation are likely endogenous to labour market outcomes, I use other library information as instruments for the number of programmes and participation. More specifically, I use the number of computers with internet access for public use, Do public libraries impact local labour markets? Evidence from Appalachia and the number of librarians without an ALA certified master's degree. The number of librarians without a master's degree is the difference between the number of librarians and those with a master's degree. These variables proxy for quality and capacity of running programmes and attracting patrons. To take into account the heterogeneity due to location and density, library programmes and participation are scaled by county population. The IV, number of computers and librarians without a master's degree are scaled by the unduplicated service population, which is calculated by the IMLS and represents the service area population without overlapping state service areas. 5

Demographic and industry characteristics
Other independent variables used are demographic and industry controls. Demographic control, namely, race, gender and age composition, comes from the Census Bureau through the Area Health Resource Files. Ideally, I would like to incorporate some measure of education. However, there is no dataset that I know of that systematically collects education attainment at the county level on a yearly basis. The average weekly wage for total manufacturing and total services come from the Quarterly Census of Employment and Wages (QCEW) provided by the BLS. Table 2 provides the descriptive statistics for the variables used in the econometric model. In terms of demographics, most of the population is White (91%), female (50.5%), and between 15 and 64 years of age (65.5%). For industry and employment, the weekly wage in manufacturing (US$779) is larger than the weekly wage in services (US$528). With regards to library programmes and participation, there are on average more than two times the number of children's programmes (277) than adult programmes (135). Also, the participation in children's programmes (7269) is on average about three times the participation in adult programmes (2290).

EMPIRICAL STRATEGY
Halleck Vega and Elhorst (2016) note three stylized facts about local unemployment rates: the strong correlation over time (Blanchard & Katz, 1992), the parallel to national trends (Pesaran, 2006) and the correlation across space (Manning & Petrongolo, 2017;Patacchini & Zenou, 2007). Two methods to take the three features of local labour market outcomes into account have been proposed: Bailey et al. (2016) suggest a two-step procedure in which the aggregate shocks are defactored from local labour market outcomes, and the resulting variables modelled using spatial econometrics; and Halleck Vega and Elhorst (2016) argue against this method presenting an alternative that deals with the three issues concomitantly.
Therefore, since the approach described by Halleck Vega and Elhorst (2016) is able to account for all three stylized facts at the same time, I follow their approach more closely and use different local spatial econometric model specifications. This approach has also been followed by Zeilstra and Elhorst (2014) and Rios (2017) analysing unemployment across European regions using a hierarchical model and a spatial Durbin model, respectively.
A general formulation of spatial econometric models is: where W is the spatial weight matrix used to spatially lag the variable of interest. The spatial autoregressive model (SAR) includes only the r parameter, the spatial error model (SEM) includes only the l parameter, and the spatial lag of X model (SLX) includes only the g parameter. LeSage (2014) argues that most applied works such as this should focus on two models only: the spatial Durbin model (SDM), which is the linear combination of the SAR and SEM models, and the spatial Durbin error model (SDEM), which is the nested version of the SEM and SLX models. The former includes both r and g parameters, while the latter includes λ and γ parameters.
The key difference between the SDM and SDEM models is that while the SDM is a global spillover specification, the SDEM is a local spillover one. Global spillover implies an endogenous feedback effect, which are spillovers from higher order neighbours as well (LeSage, 2014). LeSage (2014) argues that global spillover phenomena should be rarer than local spillovers. For the case of local labour markets, local spillovers should be more likely (Halleck Vega & Elhorst, 2016;Patacchini & Zenou, 2007), as these spillover effects would work through commuting patterns, for example, and would not generate further spillovers from neighbouring regions. An extra benefit of local spillover models lies in the fact that the spatially lagged variables can be interpreted as the indirect effect, while the non-spatially lagged are the direct effect. Therefore, I estimate both an SDEM model and an SLX model for robustness.
To investigate the effects of public library programmes on the local labour market outcomes, I estimate the following model: where y is either the unemployment rate (UR) or labour force participation rate (LFPR) in county c and year t. is the spatial weight matrix, Lib is a vector with the library programme variables: average adult and children number of programmes per capita, or average adult and children participation per capita. The vector X contains relevant control variables following the previous literature (Halleck Vega & Elhorst, 2016;Stephens & Deskins, 2018) that explains local labour market outcomes such as demographic controls (race, gender and age composition), industry controls (average weekly wage in manufacturing and services as a whole), the time-lagged unemployment rate or labour force participation rate, and contemporaneous and time-lagged national unemployment rate. m c are county fixed effects; and 1 ct is an error term. All regressors are spatially lagged including the time-lagged dependent variable. As noted by Halleck Vega and Elhorst (2016), the inclusion of the common factors (national unemployment rate) precludes the use of time fixed effects due to perfect collinearity. Appendix A in the supplemental data online shows the results using time fixed effects and the results are similar to those in the main analysis. All variables included in the regression are presented in Table 2 and are described in the table's notes. The inclusion of the time-lagged dependent variable imposes a dynamic setting which is biased under least squares estimation. For the spatial specifications, we use the estimator described by Millo and Piras (2012) which relies on a generalized moments estimator (GM) proposed by Kapoor et al. (2007) and Mutl and Pfaffermayr (2011) and based on the full set of moments conditions to address any issues with the initial condition. The estimation procedure described in Kapoor et al. (2007) is a generalization of the GM estimator, allowing the definition of feasible generalized least squares (GLS) estimator which is identical to an OLS calculated on the 'doubly' transformed model (Millo & Piras, 2012, p. 17). Also, as noted by Millo and Piras (2012), in a local spillover model a within estimator will produce consistent estimates.
The coefficients of interest in this model are b 1 and b 2 , which should be interpreted as the percentage point impact of the additional programme or participation per person on the unemployment rate and labour force participation rate. If b 1 and b 2 are positive, this suggests that public library programmes have a negative impact on the unemployment rate as larger participation and more programmes would be associated with a higher unemployment rate. Conversely, a negative sign would suggest positive impact on the unemployment rate. The opposite is true for the labour force participation rate. That is, if b 1 and b 2 are positive (negative), then public library programmes will have a positive (negative) impact on the labour force participation rate.
However, OLS estimations are likely biased due to an endogeneity problem. As the local labour markets changes, that is, unemployment rate and labour force participation rate increase (decrease), libraries can respond to these changes by offering (cutting) programmes or by incentivizing (discouraging) participation (Hunt, 2017;Jerrard, 2009). If there is a procyclical relationship between public libraries programmes and unemployment rate, the OLS estimates would be biased upward, or vice versa. Therefore, I use an IV approach. Because I have two endogenous variables, adult and children's programmes, at least two instruments are needed for proper identification. The two instruments used are: the average number of computers with internet for public use and the average number of librarians without a master's degree per served person for programme participation. The unbiased effect of public library programmes on the local labour markets is estimated using a two-stage least square framework, in which in the first stage, the instruments are regressed on the endogenous variables, also controlling for other control variables used in the second stage, as well as the regional fixed effects. 6

Instrument validity
There is a possible simultaneity of the unemployment rate and labour force participation rate with the number of public library programmes and their attendance. Public libraries are not randomly assigned to location throughout the country; however, most of them have been in place for over five decades at minimum. According to Stratton (1976), in 1972 there were 7109 public libraries in the country, which corresponds to over 78% of the libraries that existed in 2014 according to the IMLS (2017). Further, public libraries can be used as a policy instrument providing more or fewer programmes in response to changes in the local labour markets.
To properly identify the effect of public library programmes on local labour markets, I need a set of instruments that are uncorrelated with the local labour markets, but highly correlated with the number of programmes and participation in these programmes. I argue that the capacity of offering a programme and attracting patrons meet both criteria.
To offer a programme, public libraries require both an appropriate level of funding and availability of resources for the programmes. According to the IMLS, in 2014, approximately 85.2% of public library funding came from local government, while the remaining part comes from state government (7%), federal government (0.4%) and other sources (7.4%) (IMLS, 2017). Since most revenue stems from local government, this funding is expected to be contemporaneously correlated with regional business cycles and local labour markets. This is corroborated by several reports of public libraries across the country losing part of their revenue due to struggling local governments (Blau, 2011;Cleaver, 2015;Davis, 2015;Kelley, 2015;Smith, 2015;Stepleton, 2015;Warburton, 2013;Woods, 2015). 7 Public libraries also require trained staff, physical space and materials (print and computers), at a minimum, in order to offer such programmes. A priori, because these variables are related to the capacity of the library, they should not be contemporaneously correlated with regional business cycles, but they should be strongly correlated to the programmes offered by the library. Such features make these variables good candidates for instruments.

Exclusion criteria
According to the ALA (2018), there are six occupations in a public library: pages, library assistants or technicians, librarians, library managers, library directors and other professionals. 8 Librarians should be less susceptible to business cycle fluctuations since education (bachelor's and master's degrees) is shown to be a determinant in job security (Hashimoto & Raisian, 1985;Kambayashi & Kato, 2017). Librarians without a master's degree should be the result of past decision-making, thus uncorrelated with current regional business cycles, as most public librarian positions require a master's degree, preferably from a school accredited by the ALA.
The number of computers can be considered a stock variable. Even though the flow 9 of purchases of (upgrade) computers by libraries vary with budgetary allocation, libraries have physical constraint for storage and use of their materials. This physical constraint from the public library building is likely to be historically determined, thus contemporaneously uncorrelated to both unemployment rate and labour force participation rate.

Relevance criteria
The set of variables chosen also proxy for the quantity and quality of programmes and number of programmes. According to the ALA, 10 a master's degree is required by employers for most librarian positions. Given the different occupations in public library, one should expect that the librarians with a master's degree to take on administrative duties as managers and directors, while those without a master's degree to be more likely responsible for library programmes. Additionally, people may select into those programmes in which the librarian is better prepared, more approachable or have a better reputation, making it good predictors for participation as well.
On the other hand, computers and internet access are usually required for adult programmes focused on job seekers. Also, the amount and quality of inputs (books and computers) available should make it easier to provide more and better programmes for both children and adults. Tables 3 and 4 report the first-stage F-statistics and the Wu-Hausman F-statistics. The firststage F-statistics are mostly above 18, suggesting the set of instruments used is good (Stock & Yogo, 2005). These results are corroborated by the Wu-Hausman test, especially for the unemployment rate results. 11 Tables 3 and 4 show the OLS and IV results for two sets of regressions. The first two columns report the results for the unemployment rate, while the last two columns report the results for the labour force participation rate. Table 3 focuses on the number of programmes and Table 4 focuses on the participation. 12 The OLS results show no statistically significant correlation between adult's programmes and participation with the unemployment rate and the labour force participation rate. Children's programmes are also not statistically correlated unemployment rate, but are positively and statistically associated with labour force participation rate. Children's participation has no statistically significant association with neither unemployment rate nor labour force participation rate. As previously discussed, the OLS estimations are likely endogenous to labour market outcomes, hence the IV approach. The IV results show that neither adult's nor children's programmes and participation affect unemployment rate, similar to the OLS results, but also they do not affect the labour force participation. However, the results in Tables 3 and 4 are likely to be biased given the dynamic specification and the omission of spatial dependence. Therefore, I address both of these issues in Table 6, which reports both SLX and SDEM models using an IV approach and a GM estimator. 13 The first empirical step in the estimation of the a-spatial econometric model is the determination of spatial dependence and the spatial weight matrix. From Patacchini and Zenou (2007) and Halleck Vega and Elhorst (2016) we should expect spatial dependence on local labour market outcomes. I test this hypothesis for the case of Appalachia by calculating Moran's I statistic for each year in the sample for both unemployment rate and labour force participation rate, and using a k = 1, . . . , 10 nearest neighbours weight matrix since not all counties with a public library in Appalachia are contiguous. I report the Moran's I statistics in Table 5, which shows evidence of spatial dependence for every year, regardless of the spatial weight matrix and sample. Given these results, I use a spatial weight matrix of k = 1 nearest neighbours. 14 The CD test (Pesaran, 2015(Pesaran, , 2021 for all spatial regressions rejects the hypothesis of weak cross-sectional dependence. Hence, neither the estimation including common factors in the form F-test (first stage) 7.947*** 45.3*** Note: Clustered standard errors are shown in parentheses at the county level. **p,0.1; **p,0.05; ***p,0.01. Observations = 3649 in all regressions (unbalanced panel: T = 10, N = 420). Controls: per cent population Asian, Black, American Indian, other race, Latin, and two or more races, per cent female, per cent population between 15 and 64 years old, time-lagged dependent variable, time-lagged average weekly wage on manufacturing and service, national unemployment rate and time-lagged national unemployment rate, and county fixed effects. Instruments: average librarians without a master's degree per served population and average number of computers for public use per served population. Do public libraries impact local labour markets? Evidence from Appalachia of contemporaneous and time-lagged national unemployment rate, nor the estimation including time fixed-effects, account for all cross-section dependence. While this is undesirable from an empirical standpoint, this is not unexpected given the setting under investigation. The CD test may underperform given small time dimension, suggesting evidence that some strong cross-section dependence remains. Spatial methods to approximate it are the best available option.

RESULTS
The results in Table 6 are similar to those in the non-spatial setting for unemployment rate. Mostly there is no statistically significant association between adults' and children's programme and participation on local unemployment rate. However, for both SLX-IV and SDEM-IV, the results suggest a negative association between adult's programmes and participation on local labour force participation rate in terms of direct and indirect effects. The children's programme and participation effect on labour force participation rate are not different from zero.
The additional adults' programme per 1000 served people decreases the labour force participation by approximately 0.27 percentage points, while the additional neighbouring adults' programme per 1000 served people decreases the labour force participation by approximately 0.10 percentage points. The additional adults' participation per 1000 served people and neighbouring adults' participation per 1000 served people decrease labour force participation by approximately 0.03 and 0.005 percentage points, respectively.
Public library programmes (and participation) may not help people find jobs; however, they should reduce the cost of joining the labour market, especially for adult programmes. Adult programmes focus on job services and skills training (Bertot et al., 2012;Hunt, 2017). These programmes are designed to help adults find and keep their jobs, which should positively impact labour force participation and negatively impact unemployment. This should be particularly true in Appalachia, where people have less access to formal training (Haaga, 2004;Pollard & Jacobsen, 2017) and to the internet at home (Stenberg et al., 2009). However, if adults are selecting into (participating) programmes in their own county and neighbouring counties, they may opt out of the labour market or not accept jobs they would otherwise.
Similar to ALMPs (Card et al., 2018;Heckman et al., 1999;LaLonde, 2003), these programmes may suffer from a selection bias. In other words, these programmes target low-skilled and/or first-time workers (Goulding, 2006;Jerrard, 2009) who may have a higher cost of joining the labour market, especially in rural areas such as Appalachia. Hence, individual-level data would be ideal to disentangle these heterogenous effects, but they are not available for public library programmes. Finally, in this setting I am not able to explicit account for other programmes similar to those from public libraries that can be offered by private agents, government and not-for-profits such as Goodwill, for instance. On the one hand, these programmes can be considered a substitute for library programmes offered to adults. However, more likely, these are programmes are complementary to both adults and children programmes. Although each of these agencies focuses on different issues (skillset, wrap-around, etc.), all aim at lowering the cost of access to labour market. Therefore, while they likely do not influence the number of and participation in library programmes, they can influence individual outcomes; however, the ALMP literature suggests they are ineffective.

CONCLUSIONS AND POLICY IMPLICATIONS
The objective of this paper is to investigate the effect of public library programmes in local labour markets. More specifically, I focus on the impact of the number of children's and adult's programmes and participation on unemployment and labour force participation. I restrict my analysis to the Appalachian region because it is a lagging region, suggesting a high level of unemployment and poverty; mostly rural, which implies fewer private and public labour market programmes and lower levels of internet access; and with unique features in terms of labour market outcomes, in particular, lower levels of labour force participation.
Since the provision of public library programmes can be endogenous to local business cycles, I use an IV approach. Spatial econometric estimates show suggestive result that there are some negative direct and indirect effects of public library adults' programmes and participation on local labour market outcomes. As indirect effects follow the same pattern as direct effects, in settings such as that of Appalachia, public library programmes evaluation may be underestimated.
Therefore, in light of the spatial econometric models, the results suggest that public libraries not only may provide education services but also have other direct and indirect effects on local communities. Some library programmes aim at reducing the cost of (re-)joining the labour market, but other programmes target local community leisure. While I do not expect that library programmes alone will change local labour markets outcomes, they may add incentives and costs that need to be taken into account to better understand observed outcomes.
Although there is no statistically significant effect for children's programmes and participation, they may have both short-and long-term effects on labour market outcomes (not in the time-series sense). In the short-term they may be used by parents to join the labour market, while also being important for educational outcomes, which is an important predictor of longterm employment and income (Bhatt, 2010;Karger, 2021). In addition, if adults are selecting into programmes that allow them to find better job opportunities, these should spill over into other outcomes such as income and health, for example.
These results should be taken with caution because they may vary within the population (gender, race, education level, etc.) and across programmes. From the ALMP literature similar programmes are largely ineffective, but for some subpopulation groups and for some types of programmes (Card et al., 2018). In this paper, however, I am unable to test for these heterogenous effects. The differential effects from types of programmes and across subpopulation groups precludes some generalization of the results, especially in light of the uniqueness of Appalachia, previously discussed. However, this paper adds to the understanding of how these programmes affect labour market outcomes on this unique setting, and how public libraries can affect local economy (e.g., Karger, 2021;Gilpin et al., 2021).
Future studies should focus on acquiring, ideally, data at the individual level, and identifying the patrons who participate in each library programme. Also, focusing on the type of library programmes is important to make the results more comparable with studies on private and publicly provided ALMPs. Lastly, extending a mean group approach, as described by Aquaro et al. (2021), and a partial identification approach in the spirit of Pepper (2000, 2018) to incorporate the spatial spillovers can help improve the understanding and identification of the effect of the library programme given its aggregate nature. on the manuscript. A working paper version of this manuscript is available on the MRPA website.

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the author.

NOTES
1 For a complete list of programmes for school-aged children, see http://www.ala.org/alsc/ kickstart; and for a list of services and programmes for adults, see http://www.ala.org/tools/ atoz/adultservices/adult_lib_svcs/. 2 A public library system is composed of a central library and its branches and bookmobiles. 3 Appalachia is comprised of the following states: Alabama, Georgia, Kentucky, Maryland, Mississippi, New York, North Carolina, Ohio, Pennsylvania, South Carolina, Tennessee, Virginiaand West Virginia. 4 For instance, Knight and Nourse's (1969) commission's report asked for further studies on public libraries instead of providing recommendations per se. Even though there are numerous journals specialized on libraries, a focus on the impact or policy outcome of libraries is scarce. 5 The analysis with library population scaled by unduplicated service population yields similar results in terms of magnitude, sign and statistical significance and is available from the author upon request. 6 Further, conditioning on county fixed effects and lags of characteristics should be enough to mitigate the issue that neighborhood target variables can be endogenous to labour market outcomes and correlated with county-level target variables. 7 In response to budget cuts, several proposals for levies have been introduced into the ballots to specifically fund public libraries, either creating, renewing or increasing existing levies. These proposals have been mostly successful in the ballots (Fallows, 2014;Howard Fleeter & Associates, 2017;Hrin, 2018;Spokane Public Library, 2017). 8 A page is usually part-time and is responsible for keeping items in order. A library assistant can be either part-or full-time and generally performs clerical duties. Librarians are full-time employees who decide the items needed, offer programmes and training, and help people in general. Library managers are middle managers responsible for daily operations. Library directors are the main leadership in the library. For more details, see http://www.ala.org/educationcareers/ careers/librarycareerssite/typesofjobs/. 9 The IMLS (2017, 2019) shows that for the period of study, there is no big change in the kind of collections held by public libraries, even though changes in the collection have occurred over time. 10 See http://www.ala.org/educationcareers/libcareers/become/. 11 One concern may be that adult and children programmes should be considered separately. Appendix B in the supplemental data online presents the spatial econometric analysis considering adults' and children's programmes and participation separately, relying on the number of computers as instrument. The results are in line with those of the main analysis. 12 One possible concern is the heterogenous effects across the conditional distribution of labour force participation and unemployment given differential costs associated with the labour market. Appendix D in the supplemental data online shows the results for quantile regressions without the spatial dependence and reveals no difference across the conditional distribution of labour market outcomes. 13 Because the estimation of spatial panel models relies on balanced panels only, the results presented in the main text consider only the 360 counties with public libraries during the whole period of analysis. Appendix C in the supplemental data online provides estimates considering counties with no public libraries as zero programmes and participation. The results are consistent in terms of sign, magnitude and statistical significance. 14 LeSage and Pace (2014) argue that the specification of the weight matrices should not have a large impact on estimates and inferences.