The Response of Recreation Demand to Recessionary Forces: Evidence from Local Lake Usage

2. Background and Theoretical Motivation

Two important components that determine recreation behavior are income and the opportunity cost of time (Bockstael, Strand, and Hanemann 1987; Cesario 1976; Englin and Shonkwiler 1995; Feather and Shaw 1999; Larson and Shaikh 2004; Smith, Desvousges, and McGivney 1983). If recreation is a normal good, the impact of a rise in income is positive, and vice versa if it is an inferior good. Time spent for recreation services has two components: travel time and time spent on site. Phaneuf and Requate (2017) provide a simple recreation demand model to motivate a household’s optimization between consumption of nonrecreation necessities and recreation goods and services. The household is naturally endowed with T units of time, out of which the household works for H hours in the market for an hourly wage of w, and allocates the remaining time, T – H, between recreation (R) and leisure (l) so that utility from consumption of R, l, and the numeraire good (z) is maximized. For simplicity, we assume that the hours of work, H, are determined outside the model and are independent of choices for R, l, and z. Formally, the household wants to maximize the utility function U(z, R, l; q), where q represents taste parameters, subject to two separate constraints:

• 1. Money income constraint: wH = cR + z, where c is the out-of-pocket cost of a trip.

• 2. Time resource constraint: T = H + l + tR, where time remaining after work hours, T – H, is used for leisure and recreation, and t is the time cost for consumption of each unit of R.

The household solves the following two-constraint utility maximization problem: Manipulation of the first-order conditions results in At the optimum, the marginal benefit from one recreation trip (U_R/U_z = δz/δR) must equal the marginal cost (c + φt). The recreation price consists of a direct part, c, and an opportunity cost of time, φt. Solving the first-order conditions yields demand equations for each of z*(c,t,w,H,T,q), R^*(c,t,w,H,T,q, l^*(c,t,w,H,T,q), and φ^*(c,t,w,H,T,q).

The model suggests the possible pathways through which recession might influence recreation demand behavior. If the recession affects any recreationist household directly through a reduction in working hours, or job loss, that household experiences a fall in monetary income but has more available time for leisure and recreation. The fall in income raises the value of λ, the Lagrange multiplier on the monetary budget constraint. The unemployment releases more time for time-intensive consumption and leisure, which implies that the shadow value of time (μ) would fall. The combination of these two changes suggests a fall in ϕ, the opportunity cost of time.

Note that an opposing effect takes place through the decrease in working hours and resulting fall in income if outdoor recreation is a normal good. These two effects are comparable to a substitution and income effect resulting from a price change.⁴ Whether the time effect dominates the income effect will determine the overall effect of unemployment on recreation demand under a normal good assumption. If, instead, outdoor recreation is an inferior good, the two effects will reinforce each other. However, empirical evidence suggests that outdoor recreation may be a normal, though income inelastic, good (Loomis and Keske 2009; Phaneuf and Smith 2005).

In modeling local recreation demand, the choice set often includes an element known as the “stay-at-home” option (e.g., Egan et al. 2009). This option captures everything outside the model, including options for other recreation activities such as exotic vacations or an international trip. If recreationist households have planned for such a trip but experience a fall in income due to unemployment during a recession, they are less likely to follow through with those expensive trips. In such cases, the “stay-at-home” option becomes less appealing and might induce an increase in demand for cheap, local recreation activities. In the model specified, this is equivalent to saying that corner solution (i.e., U_R/U_z < c + φt) is less likely to occur. This is equivalent to the case when outdoor recreation is of inferior type. However, since we do not know whether an average household would consider outdoor recreation as an inferior or normal good, the ultimate effect is an empirical question. The net impact on local recreation will ultimately depend on the combination of income effect and substitution effects resulting from unemployment during the recession.

Additional reasons for an ambiguous sign arise in a dynamic setting. Hoque, Kling, and Herriges (2013) introduce a simple two-period model that includes a precautionary motive and consumption smoothing to develop intuition on how dynamic factors affect outdoor recreation demand during a recession. During a recession, the precautionary motive induces risk averse people to increase savings and spend less, thereby reducing spending on all goods, including recreation. This effect would work in the opposite direction for risk loving households. However, a household that becomes unemployed will see a decline in the opportunity cost of leisure, leading to increased consumption of recreation. Unemployment also induces consumption smoothing between time periods. Since money and time are complements for recreation demand, an intratemporal allocation of time from reduced market hours toward nonmarket goods such as outdoor recreation is another consumption smoothing mechanism that households may adopt during recession. Nevo and Wong (2018) documented that U.S. consumers smoothed consumption expenditures during the Great Recession by intratemporal substitution of time across activities. In sum, the combined effect of uncertainty and unemployment on outdoor recreation depends on the relative strength of these effects within and across households. The net effect of a recessionary shock is ultimately an empirical question.

3. Econometric Framework

To provide insight on the empirical question of how recreation use responds in a recession, we study the impact of employment changes during the Great Recession on lake recreation use. The treatment group includes those who experience a change in employment status during a recession, and assignment to this treatment group is nonrandom. This nonrandom treatment assignment, also known as a selection effect, is likely due to both observable and unobservable factors. The selection problem can obscure possible causal effects of a change in employment status during a recession on recreation behavior due to confounding factors that affect both selection into the treatment (change in employment status) and the outcome variable (a recreation visit). The propensity score matching (PSM) method, presented by Rosenbaum and Rubin (1983), is one approach to overcome the selection problem. PSM is widely used in the program evaluation literature (Dehejia and Wahba 2002; Jalan and Ravallion 2003; List et al. 2003; Liu and Lynch 2011; Chabé-Ferret and Subervie 2013). Under certain assumptions, the method solves the problem of missing counterfactuals in nonexperimental settings.

The periods for this analysis are 2005, the prerecession year, and 2009, the recession year. The unit of analysis in this study is the household. We consider three treatment groups, each of which consists of households whose employment status changed during the recession. The control group consists of the households that remained employed full-time in both 2005 and 2009. Each household i may fall in any of the three treatment groups defined as follows: [1] [2] [3] The first treatment group includes those households whose employment status changed to unemployed, part-time employed, or retired. Because the retirement option may be fundamentally different than reduced hours or full unemployment, we form two additional treatment groups to consider these categories separately. The second group contains those households that became unemployed or moved to part-time employment status. The third group includes just those that moved into retirement.

The first outcome variable of interest is a binary variable: Trip_ikt = {0,1}, where 1 indicates that household i in treatment group k took a trip in year t = {Y09,Y05}, and 0 indicates not. The subscript k = {0, 1, 2, 3}, where 0 indicates control group while 1, 2, and 3 indicate the treatment group. The second outcome variable is NTrip_ikt, which denotes the total number of trips for household i in group k in year t.

We estimate separate probit regression models for each treatment group for selection into change in employment status during the 2009 recession, using the following specification: where X_i,k,Y05 is the set of observable covariates for household i in treatment group k in prerecession year Y05. To reduce bias, the program evaluation literature (Heckman, Ichimura, and Todd 1997; Smith and Todd 2005; Caliendo and Kopeinig 2008) suggests choosing a rich set of explanatory variables based on economic theory, previous research, and institutional settings.

To estimate the impact of a change in employment status during a recession on outdoor recreation, we need a counterfactual estimate of what recreationist households in the treatment group would have done were they not affected by an employment shock during the recession. This counterfactual works if the conditional independence assumption holds. This states that after controlling for a set of observable covariates X that are not affected by treatment, potential outcomes are independent of treatment assignment.⁵ We use the recreation behavior of households in the control group in 2009 as the counterfactual recreation behavior for households in the treatment group, at both the intensive and extensive margin. [4] [5] In addition, the common support assumption requires that for a recreationist household that has become unemployed or retired, there must exist comparable recreationist households that are still employed but are otherwise similar with respect to other characteristics.

We estimate the following average treatment effect on the treated (ATT) estimators: [6] and [7] where I₁ is the set of treated observations, I₀ is the set of control observations, S_p is the region of common support, and N_T is the number of observations that belong to the set I₁ ⋂ S_p. Moreover, is the counterfactual for treatment i, which is constructed as the weighted average of all of the matched control outcomes. Note that is the weight assigned to each matched control j corresponding to the treatment i that will depend on the distance between the propensity scores of treatment i and match j, and the number of matches as well.

The estimators in equations [6] and [7] are unbiased if selection into employment change during the recession is due to observable factors. However, there may exist unobservable factors, both time variant and time invariant, that affect both the likelihood of a household’s exposure to employment change during a recession and that household’s recreation behavior. For example, in our context, factors such as distance to lake, residence amenities, local attributes, and labor market conditions might confound the results.

Heckman, Ichimura, and Todd (1997), Smith and Todd (2005), Imbens and Wooldridge (2009), and Arriagada et al. (2012) recommend using a DID approach when geographic and other individual-specific fixed factors play a potentially confounding role. Using the DID matching estimators with our panel structure we are able to difference out all potential time-invariant unobservable factors. The estimation procedure follows equations [6] and [7] except that the impact is measured as the differences of outcomes.⁶ Note that the DID matching estimator requires the parallel trend assumption; that is, any differences in the outcome variable between treatment and control groups must remain constant over prerecession periods (Imbens and Wooldridge 2009).

We apply four different matching algorithms including nearest neighbor matching with replacement, four nearest neighbor matching, radius matching within a radius distance of one-half standard deviation of the estimated propensity score, and kernel matching. Standard errors are estimated following Abadie and Imbens (2008, 2016) for nearest neighbor matching estimators, and a bootstrapped procedure for radius and kernel matching estimators.⁷ The nearest neighbor matching estimators perform better in terms of bias reduction, while the radius and kernel matching estimators are more efficient, as larger shares of controls are utilized in the construction of the counterfactual mean.

4. Iowa Lakes Survey

We utilize data from the Iowa Lakes Survey of households, a random population survey, which collects a rich set of information on Iowan’s lake visitation patterns as well as demographics on gender, age, education, employment status, income, and household composition. The survey was administered in both 2005 and the recession year 2009.⁸ The surveys in 2005 and 2009 together comprise a complete panel with all necessary information for 1,778 households, giving us information on their recreation behavior (both participation and number of trips) and relevant demographics both before and during the recession.⁹ We first identify the households that experienced a change in employment status during the recession to construct the treatment and control group for our study.

The upper panel of Table 1 presents the employment status of the respondents. To study how the movement from full-time employment to unemployment or retirement affects recreation, our initial focus is on the 1,018 households summarized in the first row of the table.¹⁰ Among 1,018 households employed full-time in 2005, 83% (n = 848) reported unaltered employment status in 2009. Approximately 6.5% of the households employed full-time in 2005 reported either unemployment or a decrease in working hours in 2009 (n = 66). In addition, 10% of the previously fulltime employed households retired in 2009 (n = 100).¹¹

The lower panel of Table 1 shows the composition of the control and three treatment groups. Among the three, treatment group 1 is the largest, consisting of 155 observations in total, as it includes retired, unemployed, and part-time-employed households. Comparing with the counterfactual group size, the treatment control ratio is approximately 1:5. The relatively small treatment group compared to the control legitimately raises a concern of small sample bias; however, we are not aware of any test to detect such small sample bias in matching estimators. Nonetheless, by using a large set of control observations we are able to find strong matches, which helps reduce the bias due to incomplete matches. In addition, we note that it is common in the applied literature to use matching estimators with a small sample and relatively large control group (e.g., Bertram-Huemmer and Kraehnert 2018; Rao, Brümmer, and Qaim 2012; Arriagada et al. 2012; Liu and Lynch 2011).

Information on participation, average number of trips, and demographics across treatment and control groups is reported in Table 2. Participation, on average, remains unchanged for the control group across the years 2005 and 2009. However, unlike treatment group 3, treatment groups 1 and 2 increased participation in lake recreation in 2009. For total number of trips, the pattern is a little different. For the control group, treatment group 1, and treatment group 3, the mean number of trips fell in 2009, when compared to 2005. In contrast, treatment group 2 exhibits an increase in the mean number of trips in 2009. This suggests possible differences in recreation behavior across retired and unemployed households.

The covariates we include to estimate the propensity score are age; polynomials of age; education; gender; number of children in the household; interaction terms between education, age, and gender; recreation patterns in previous years; and boat ownership as reported in the 2005 survey. The inclusion of education and age are motivated by the earning function literature in labor economics (Mincer 1974; Heckman, Lochner, and Todd 2006). We assume that factors that determine one’s earnings are strong predictors for labor market status as well. The Iowa lake surveys also contain information on households’ residence county and ZIP code. We match this with the rural-urban commuting area (RUCA) codes maintained by the U.S. Department of Agriculture’s Economic Research Service to classify households by four location types: rural, small town, micropolitan, and metropolitan.¹²

Finally, information on lake recreation activities, such as fishing, boating, and hunting are included to capture preferences for differing activities. Boating, fishing, and hunting are common lake recreation activities (Evans, Herriges, and Kling 2011). Boat activities combine information on household boat ownership as well as participation in any boating activities such as canoeing, boating, jet skiing, and sailing. Habit persistence has been found in the recreation literature (Adamowicz 1994; Moeltner and Englin 2004). This suggests that prerecession recreation behavior might influence recreation choices during the recession. Information on recreation usage (total number of day and overnight trips) taken in the prerecession year is used to group households with similar recreation preferences. The bottom panel of Table 2 presents summary statistics on various covariates observed in the prerecession year.

6. Robustness

We conduct three robustness checks. First, we use a placebo recession year to check if any preexisting differences across treatment and control groups are contaminating our estimates. Second, we change the specification for propensity score estimation to include a subset of covariates previously used; specifically, we exclude recreation preference variables. Third, we match each treatment observation with controls from the same geographic region to control for time-variant, spatial, unobservable factors, given that rural and urban areas may be affected differently during a recession.

The objective of the placebo exercise is to check whether it is unemployment during the recession or some preexisting unobservable factors affecting the treatment and control groups differentially. If the treatment and control groups exhibit differentiated trends in the prerecession years, and the recession truly has no impact on recreation, the DID matching estimator can pick up this difference in trends and make it appear as if it is due to the change in employment during the recession. For the placebo exercise, we use 2003 as a placebo recession year, as the Iowa lakes survey was also undertaken in that year. Tables B1–B3 of Appendix B show that balancing of covariates is well satisfied. We report the estimates from the placebo exercise in Table 4. In all of the matching processes, neither participation nor frequencies of trips in lake recreation are statistically different across the treatment and control groups in 2003. As a second placebo check, a comparison of lake-based recreation of the treatment and control households between a prerecession (2005) and postrecession (2014) period is drawn.¹⁷ Due to a significant gap between the two survey years, approximately 30% of the treatment and control households from the original sample are absent in the 2014 survey.¹⁸ Nevertheless, the results from this exercise, as presented in Table B4 of Appendix B, suggest that the mean recreation behavior is not significantly different across treatment and control groups in a postrecession year. This finding gives us confidence that our analyses based on matching exercises as reported in the previous section are not contaminated due to differential group trends.

As a second robustness check, we estimate the propensity score excluding the recreation preference variables, but including the demographics and region types. Table 5 reports the estimates for participation. Covariates balancing results, as reported in Tables C1–C3of Appendix C, reveal that the quality of the match is good. For treatment groups 1 and 2, the ATT estimates on the level for participation in lake recreation are not robust to the inclusion of a limited set of covariates. However, for these two treatment groups, the DID matching estimates are robust to this different set of observables covariates. Similar to the pattern reported in Table 3, Table 5 reveals that the retired households (treatment group 3) did not alter their participation behavior in lake recreation during the recession.

For the frequency of lake trips, with the new set of covariates, only one out of the 12 matching estimators across the three treatment groups exhibits statistical significance. Our previous finding that the frequency of lake trips does not change due to unemployment or retirement during the recession is robust to the choice of covariates.

Although we account for the effects of time-invariant unobservables through matching on the differences, our estimates may still be confounded by unobservable factors that vary across regions with time. For example, employment statistics in a rural agricultural county may not change during the recession, while employment in a metropolitan area may drop sharply. If there are differences across region types, we would wish to avoid matching a rural treatment with an urban counterfactual. Since DID matching estimators cannot control for such region-specific time-variant unobservable confounding factors, we match each treatment observation with controls from the same RUCA region. The results are reported in Table 6, while Tables D1–D3 of Appendix D report covariate balance for the cell matching. In general, the covariates balance well.

The estimates in Table 6 reveal that when matching is done within the RUCA cell, none of the four matching estimators for participation in treatment group 1 are statistically significant. In contrast, for treatment group 2, all four matching estimators are statistically significant. However, once we apply the DID matching, the results for treatment group 1 are statistically significant in three of the four cases. For treatment group 2 (unemployed and partially unemployed), two out of the four DID matching estimators exhibit statistical significance. The finding of unaltered participation in lake recreation during the recession for the retired group is robust to matching within RUCA cells. In the case of frequencies of lake trips, as can be seen in Table 6, matching within RUCA cells generates similar estimates as observed before (reported in Table 3).

7. Extension: County Unemployment and Recreation

In this section, we adopt an alternative approach to investigate lake recreation during the recession by exploiting cross-county cross-period variation in county unemployment rates. Iowa was not resilient to the adversities of great recession. The average yearly unemployment rate in Iowa rose by 49%, from 4.3% to 6.4%, between 2005 and 2009. Moreover, county-level unemployment rates, extracted from the Local Area Unemployment Statistics by the Bureau of Labor Statistics, exhibit wider variation compared to those reported at the state level. The county-level unemployment rate in our sample varies in the range of 3.9% to 10%, with a median unemployment rate of 4.3% in 2009. In contrast, the corresponding range for the 2005 unemployment rate is 2.9% to 6.6%,with the median at 6%.

We use data from five years of the Iowa Lake Project surveys 2002, 2003, 2004, 2005, and 2009 (the survey was not conducted from 2006 to 2008). Because many households did not report employment information in all years of the survey, a complete panel with the needed data is quite small.¹⁹ Consequently we use a proxy for employment/economic status: the county unemployment rate. This approach has been adopted in a number of economics studies assessing the change in economic behavior during a recession (Ruhm 2000; Dehejia and Lleras-Muney 2004; Aguiar, Hurst, and Karabarbounis 2013).

We utilize a panel spanning the years 2002–2005 and 2009 to estimate a fixed effects model relating county-level unemployment rates to household participation in lake recreation. In the lake surveys, 3,035 observations from 2009 have a matching observation in at least one of the years from 2002 to 2005. Defining 2009 as the recession year, we estimate the following specification: where Trip_ict is a binary variable indicating whether household i in county c takes any lake trip in year t, or not; Un_ct indicates unemployment in county c in year t; RecYr is an indicator variable assuming a value of 1 if year t is a recession year; γ_i are household-specific fixed effects to address time-invariant demographics such as race, location, education, preference for recreation or work, risk attitudes, and so forth; and γ_ct are county-specific time trends.²⁰ In this specification, if β₃ > 0, this provides evidence of a positive effect of unemployment during a recession on outdoor lake recreation. Since the error term, e_ict, might exhibit correlation both across counties within year and within a county across years, standard errors are clustered at the county and year levels (Cameron and Miller 2015).

Table 7 shows the fixed effect estimates of the impact of county unemployment rate on recreation participation. Panels (a) and (b) report results for the unbalanced panel, while panel (c) reports results for the subsample used in the matching exercises in the previous section. The coefficient of recession year is negative and significant. The estimates in columns I–IV in panel (a) reveal that in a re-cession year, participation in outdoor lake recreation decreases, ranging from 14% to 22% relative to a nonrecession year. However, no such pattern is observed for county unemployment rates. Neither the coefficients on county unemployment rate nor those on the recession indicators consistently exhibit statistical significance in panel (c).

In panel (a), the coefficient of the interaction term between county-level unemployment and recession year, after controlling for the level effect of unemployment and recession, is positive and significant in three of the five specifications. Thus, it appears that the average participation rate falls during a recession year, but during a recession a household from a county with a high unemployment rate is more likely to participate in lake recreation compared to one from a low unemployment county. In panel (b), where 2004 is assumed to be a placebo recession year, none of the coefficients on the recession-year indicator or interaction term between recession and county unemployment rates are statistically significant in any of the specifications. Overall, participation in lake recreation responds to unemployment in a different manner, depending on whether there is also a recession at the time.