3.1. Survey Sample

The primary survey framework employed a multi-stage sampling design across the flood-prone agricultural municipalities of Laguna, yielding an initial baseline sample of \({N=349}\) respondents. This original aggregate sample comprised two distinct livelihood cohorts: 214 fisherfolk households and 135 crop-cultivating smallholder farmers.

Because the core objective of this study is to evaluate the impact of crop insurance—a mechanism institutionally irrelevant to fishing operations—a strict inclusion criterion was applied to preserve sectoral homogeneity. Consequently, the 214 fisherfolk observations were excluded from the analysis. The final empirical estimation was conducted exclusively on the specialized sub-sample of 135 smallholder farmers (\({n=135}\)), ensuring a direct alignment between the treatment variable and the targeted production livelihoods.

3.2. PSM Method

To estimate the causal impact of insurance participation on farming household vulnerability and post-disaster coping strategies following the severe flooding caused by Typhoon Ulysses in 2020, PSM was employed. The covariates included the farmer’s age, farming experience of the household head, farmland size, percentage of irrigated farmland, number of flood events experienced, and geographic elevation. These variables were used to achieve balance between insured (treated) and uninsured (control) farmers. A Logit model was estimated to generate propensity scores, after which a nearest-neighbor matching algorithm was applied. The treatment effect was measured using the Average Treatment Effect on the Treated (ATT). Because the coping indicators are binary, the ATT represents the change in probability (percentage points) of an outcome occurring among the treated group (those with insurance) compared to the counterfactual scenario in which they had not received insurance. The three outcome variables—post-flood debt accumulation, reduced household spending, and reduced food consumption or meal skipping—serve as binary indicators of coping behavior. Accordingly, they do not measure the specific magnitude of debt, the extent of expenditure and food reductions, or the quantitative trade-offs among these responses. Furthermore, they do not reflect how household income ultimately changed. Instead, these variables serve as proxies for coping behavior and welfare stress rather than actual fluctuations in income.

According to Milton Friedman’s Permanent Income Hypothesis and standard lifecycle consumption models, households do not alter their day-to-day welfare in reaction to temporary income shocks unless they are completely credit-constrained or lacking a liquid safety net (crop insurance). Hence, when a climate shock occurs, a household’s budget constraint behaves as a rigid accounting identity:

To generate a robust PSM analysis, two assumptions must be satisfied. The first is the common support assumption, which requires that comparisons be made only among individuals with similar ranges of propensity scores. The second is the CIA, which posits that, conditional on observed baseline covariates, insurance participation is effectively random.

a. PSM Model Specification

The estimation strategy follows a three-stage sequential framework to establish a valid counterfactual.

1. Stage 1:

   Propensity Score Estimation (First-Stage Logit)

   To satisfy the CIA, the probability that a smallholder farming household participates in crop insurance, conditional on observable baseline covariates, was estimated using a parametric logistic regression framework:

   \begin{align} P(X) &= \operatorname{Pr}\left(D_{i} = 1 \mid X_{i}\right) \notag \\ & = \dfrac{\exp\left(\beta_{0} + \beta_{1}X_{1i} + \dots + \beta_{k} X_{ki}\right)} {1 + \exp\left(\beta_{0} + \beta_{1}X_{1i} + \dots + \beta_{k}X_{ki}\right)}, \label{eq:2} \tag{2} \end{align}

   where

   1. ・

      \(D_i\) is the binary treatment indicator (\(D_{i}=1\) if the household is insured; \(D_{i}=0\) otherwise),

   2. ・

      \(X_{ki}\) denotes the vector of multi-dimensional vector baseline covariates (e.g., household demographics, land assets, and geophysical variables), and

   3. ・

      \(P(X)\) is the generated univariate propensity score representing assignment probability.

2. Stage 2:

   Distance Matrix Metric (Mahalanobis Extension)

   To address the perfect prediction separation boundaries caused by localized covariates (e.g., irrigated plot clustering), the univariate propensity scores were incorporated into a multi-dimensional covariate balancing distance metric. The geometric distance \(d(i,j)\) between treated household \(i\) and control household \(j\) was minimized using the covariance matrix:

   \begin{equation} d(i,j) = \left(X_{i} - X_{j}\right)' \mathit{\Sigma}^{-1}\left(X_{i} - X_{j}\right), \label{eq:3} \tag{3} \end{equation}

   where \(\mathit{\Sigma}\) is the variance-covariance matrix of the vector of baseline covariates in the pooled matched sample.

3. Stage 3:

   Treatment Effect Estimation

   The ATT was computed using nearest-neighbor vector mapping (\(nn=1,3\)) to measure changes in the behavioral income proxies:

   \begin{align} \textrm{ATT} &= E\left[Y_{1i}-Y_{0i} \mid D_{i}=1\right] \notag \\ & = \dfrac{1}{N_T} \sum_{i \in \left\{D_{i}=1\right\}} \left[Y_{1i} - \sum_{j \in C(i)} w_{ij}Y_{0j}\right], \label{eq:4} \tag{4} \end{align}

   where

   1. ・

      \(Y_{1i}\) and \(Y_{0i}\) denote the observed post-disaster outcomes for household \(i\) under treated and control states, respectively,

   2. ・

      \(N_T\) is the number of successfully matched insured households within the common support boundaries, and

   3. ・

      \(C(i)\) defines the set of control counterparts matched to treated unit \(i\), and \(w_{ij}\) is the weight assigned to each control match.

b. Variable Definitions and Operational Matrix

Table 1 presents the operational parameters of the model, which are categorized into binary treatments, behavioral outcome proxies, and baseline selection controls.

c. Theoretical Justification of Variables Outcome Proxies (Behavioral Income Channels)

Direct self-reported post-disaster income profiles are susceptible to substantial recall bias and high seasonal volatility. Following the permanent income and consumption-smoothing paradigms, households smooth transitory income shocks through distinct behavioral responses.

1. 1.

   food represents severe, non-discretionary constraints where income shocks extend beyond survival margins.

2. 2.

   hhspend captures immediate contractions in discretionary or semi-discretionary non-food expenditures.

3. 3.

   debt tracks informal borrowing dependencies used to cover cash-flow shortfalls.

d. Baseline Matching Covariates (X)

To satisfy the CIA, the matching vectors must influence both the propensity to acquire insurance and post-disaster household income generation:

1. 1.

   Demographics (age, hhyf): Control for risk aversion preferences and farming experience.

2. 2.

   Asset Endowments (landsize, perirrigated_10): Reflect initial physical wealth constraints and production capacity, influencing both formal crop insurance demand and systemic flood vulnerability.

3. 3.

   Geophysical Context (floodevent, elevation): Capture baseline physical exposure of the crop fields to flooding associated with the Laguna Lake basin.

e. Validating the Setup

A PSM model is correctly specified only if the groups are balanced. After matching, insured and uninsured farmers must have similar average values for age, years of experience, landsize, percentage of irrigated land, flood events experienced in the last five years, and farm elevation. The common support condition is imposed by excluding treated observations whose propensity scores fall outside the range of the control group. To ensure the validity of the match, Standardized Mean Differences (SMD) were used, with a threshold of \(|0.1|\) indicating acceptable balance.

The multi-dimensional Mahalanobis distance metric successfully balanced these parameters (Table 2). Prior to matching, substantial selection bias was observed. In particular, insured farmers had a significantly higher percentage of irrigated infrastructure (perirrigated raw SMD \(=\) 0.693) and larger holding distributions (landsize variance ratio \(=\) 2.18), indicating that baseline program participation was skewed toward better-endowed farming assets. The nearest-neighbor matching algorithm (\({nn=3}\)) effectively neutralized these overt selection dynamics. After matching, the absolute SMD for all six parameters were well below the conservative threshold of 0.10, while variance ratios stabilized near the ideal 1.0 target value. This corrected distribution balance confirms that the matched control group closely resembles the treatment group, validating the subsequent causal ATT estimates.

4.1. Worst Flood Experience in 2017–2022

Figures 1(a)–1(c) show the sample flood extent map and flood hazard maps of the municipalities of Bay and Sta. Cruz during Typhoon Ulysses, together with the locations of the survey respondents. Fig. 1(a) shows that Barangay Tagumpay in the municipality of Bay experienced the highest inundation depth of at 3 m, while Barangays San Antonio, Sto. Domingo, and Maitin experienced inundation depths ranging from 0.3 m to 2 m. Fig. 1(c) shows that Santo Angel Norte, Santisima Cruz, San Pablo Norte, and San Pablo Sur were the flood-prone barangays in Sta. Cruz, with inundation depths ranging from 1 m to 4 m.

When asked about the worst flood events they had experienced during the previous five years (2017–2022), the respondents identified the years 2020 (Typhoon Ulysses) and 2022 (Typhoons Paeng and Karding) because these were the strongest typhoons to affect Laguna during the study period (Table 3). Their houses and farms were inundated twice during the five-year period.

The average duration of flooding on rice farms was 42 days, or almost 1.5 months, with an average flood depth of 31 inches (more than 2.5 feet). Flooding inside the house lasted an average of 30 days, with an average depth of 10 inches.

4.2. Flood Damages from Typhoon Ulysses

Table 4 presents the estimated flood damages caused by Typhoon Ulysses among the survey respondents. Of the 135 rice farmers, 46 (34%) reported damage to their standing crops, amounting to an estimated total loss of PHP 1,840,000.

Because crop or flood insurance has been shown in the literature to help farmers reduce losses and recover from disaster-related damages, this study examines why some farmers availed themselves of crop insurance while others did not.

4.3. Crop Insurance Profile of Respondents

The PCIC is the primary agency responsible for providing agricultural insurance covering crops, non-crop assets, and crop losses caused by natural disasters, plant diseases, and pest infestations. Free insurance is provided to farmers registered under the RSBSA.

Table 5 presents the crop insurance profile of rice farmers and the responses of those who did not avail themselves of crop insurance.

Table 6 presents the main reasons why rice farmers did not avail themselves of crop insurance. The most frequently cited reasons were unclear insurance policies, lack of information, and having small farmland that was too small to qualify for coverage.

4.4. Results and Discussion of PSM and Mahalanobis Distance Metric Analysis

To establish the causal impact of insurance on household food vulnerability (defined as food consumption reduction and meal skipping), we estimated the ATT. Households were paired using a 1-to-1 nearest neighbor matching approach based on a first-stage logit model (\({N=135}\)). Selection into the insurance program was modeled using agricultural human capital—measured by the age of the household head (age) and years of farming experience (hhyf)—together with rescaled irrigation increments (perirrigated_10), land size (landsize), and environmental risk factors (floodevent and elevation). A strict caliper of 0.25 standard deviations was imposed to preserve match quality, and standard errors were adjusted using the Abadie–Imbens robust correction method.

The empirical results indicate that insurance plays a highly transformative role in protecting agricultural households from consumption shocks. The estimated ATT is negative and highly statistically significant (\({\textrm{ATT}=-0.316}\), \({z=-3.51}\), \({p<0.001}\)). Controlling for farming experience and baseline geographic vulnerabilities, holding a crop insurance policy reduces the probability of a household reducing food consumption or skipping meals by 31.63 percentage points. These findings strongly support the assertion that formal insurance instruments serve as vital institutional safety nets, preventing localized agricultural shocks from degrading household food security. Using \(nn(3)\) for the primary food outcome variable, the result remains highly robust, confirming that crop insurance has a statistically significant causal impact on reducing the probability of food insecurity at the 5% significance level (\({p=0.021}\)). When a wider multi-neighbor comparison group is used, crop insurance decreases the probability of a household skipping meals or experiencing severe reductions in food consumption by 14.97%. Consistent with other configurations, no observations violated the common support assumption, and all 135 households remained in the analysis. Because both the negative trajectory and statistical significance (\({p<0.05}\)) persist under this adjustment, the food security finding is highly stable and robust to parameter changes.

Figure 2 shows the propensity score overlap (common support). The horizontal axis (\(x\)) represents the estimated probability of a farmer purchasing crop insurance (0 to 1), while the vertical axis (\(y\)) shows the concentration of households at that probability.

Both the blue curve (insured) and red curve (uninsured) span the same general horizontal range. This visually demonstrates that the sample satisfies the overlap assumption. For nearly every insured household, there is an uninsured household with a similar baseline probability profile. However, although the curves overlap, their peaks may shift (e.g., the insured peak might be further to the right). This shift provides the visual evidence of selection bias, confirming that insured and uninsured households are structurally different at baseline. This justifies the use of advanced matching metrics such as the Mahalanobis distance metric to improve covariate balance.

Figure 3 illustrates the distributional balance of the estimated propensity scores across the treated (insured) and control (uninsured) smallholder cohorts, both before and after the matching routine. The left panel (Raw) displays the baseline, unmatched sample data. In this baseline state, a clear divergence is visible: the median propensity score of the treated group is substantially higher than that of the control group, and the entire interquartile range (IQR) for the insured cohort is shifted upward. This structural asymmetry confirms substantial selection bias, demonstrating that households with higher baseline assets or greater flood exposure had a fundamentally higher probability of purchasing agricultural crop insurance.

The right panel (Matched) demonstrates the corrective performance of the matching algorithm. After matching, the median lines within the boxes are brought into near-perfect horizontal alignment, and the interquartile distributions exhibit highly symmetrical distributions. By systematically selecting and weighting control households that closely resemble the multi-dimensional profiles of the insured farmers, the algorithm successfully eliminates baseline selection bias. The appearance of minor outliers at the lower boundary of the matched control distribution represents the mathematical trade-off required to preserve sample size while forcing the overall distributions into parity. This visual alignment confirms that the matched control group constitutes a highly rigorous, statistically valid counterfactual framework for isolating the true treatment effects of the crop insurance program.

To test the localized quality of the counterfactual matches, a strict caliper of 0.25 standard deviations of the propensity score was imposed on the baseline 1-to-1 nearest neighbor specification. The estimation yielded treatment coefficients and standard errors identical to those obtained from the unconstrained nearest-neighbor model. Post-estimation diagnostics confirmed that zero observations were dropped (e(N_dropped) \(=\) 0), indicating strong common support and high baseline match proximity between the insured and uninsured cohorts. Consequently, the final estimates preserve maximum statistical power while guaranteeing that matches do not suffer from distance-driven interpolation bias.

Table 7 summarizes the three outcome variables across matching frameworks, explicitly tracking sample size modifications, dropped units, standard errors, and significance markers. For both Models (1) and (2), zero units were dropped, which proves that the 37 uninsured households provide adequate baseline coverage range for the 98 insured households, eliminating the risk of trimming bias. Model (3) excluded eight treated observations because the non-parametric kernel matching procedure relies on bounded distributions. These observations had extreme propensity scores, likely reflecting exceptionally large landholdings or distinct demographics, and therefore could not be validly matched with any weighted combinations of the control group. Excluding these observations improved the statistical validity of the kernel estimates. The negative signs remain stable for food and hhspend across all specifications. Although the coefficients varied because of the bias-variance trade-off—with \(nn(1)\) identifying the single closest match and multi-matches producing estimates closer to the sample average—the substantive conclusion remained unchanged. Crop insurance effectively protects immediate domestic welfare but has no impact on long-term agricultural reinvestment credit. However, the model exhibits a covariate imbalance issue. Although matching substantially reduced landsize imbalance, several other variables became slightly less balanced in the matched sample than they were in the raw data.

To address the covariate distribution and variance asymmetries observed in the baseline propensity score models, a multivariate Nearest-Neighbor Covariate Matching routine based on the Mahalanobis distance metric with three neighbors (\(nn(3)\)) was implemented. Unlike PSM, which reduces multi-dimensional data to a single probability index, the Mahalanobis metric directly measures proximity across the entire matrix of household and geographic traits, automatically adjusting for the scaling differences and underlying covariances among the variables.

Post-estimation balance diagnostics (Table 8) indicate that this geometric approach achieved superior sample symmetry. With the exception of the highly skewed farm size variable, the matched variance ratios for household head age (1.13), farming experience (1.01), and regional elevation (1.03) fell comfortably within the conservative academic threshold of 0.80 to 1.25. Furthermore, the SMD for localized climate shocks (floodevent) and topography (elevation) were reduced to negligible levels (\({\beta=0.042}\) and \({\beta=-0.025}\), respectively).

Under this highly optimized and robust counterfactual framework, the ATT confirms a powerful, statistically significant protective effect of crop insurance. Holding crop insurance reduces the probability that a smallholder household will be forced to reduce essential food consumption during a flood event such as Typhoon Ulysses in 2020 by 24.49 percentage points (\({z=-2.26}\), \({p=0.024}\)). This finding confirms that formal risk-transfer mechanisms provide an effective consumption-smoothing buffer, protecting vulnerable rural households from immediate nutritional deprivation during environmental shocks.

To evaluate the sensitivity of the food security findings to hidden unobserved selection bias (endogeneity), a Rosenbaum bounds sensitivity analysis was conducted using the matched sample (\({N=135}\) pairs). The analysis evaluates the extent to which an unobserved covariate (e.g., hidden psychological risk aversion or innate management ability) would have to influence the odds of insurance enrollment to invalidate the estimated treatment effect.

As shown in Table 9, the treatment effect on food security shocks is resilient to hidden bias. At the baseline value of \({\gamma=1.0}\) (assuming no unobserved bias), the upper-bound significance level remains highly robust (\({p<0.01}\)). More importantly, even when the simulated hidden bias is incrementally increased up to \({\gamma=2.0}\)—representing a situation in which an unobserved factor doubles a household’s likelihood of acquiring insurance—the upper-bound \(p\)-value remains significant with \({p<0.01}\). Because the critical tipping point lies well beyond \({\gamma=2.0}\), these findings indicate that the estimated consumption-smoothing benefits of crop insurance in Laguna are structurally stable and immune to confounding unobserved selection effects.

5.1. Policy Implications

The empirical findings carry important policy implications for the redesign of agricultural safety nets administered by institutional bodies such as the PCIC. Because the matching estimates confirm that current insurance frameworks effectively protect basic food security (\({\textrm{ATT}=-0.244}\), \({p=0.030}\)) but fail to reduce cuts in domestic non-food expenditures (\({p=0.119}\)) or alleviate post-flood credit dependence (\({p=0.753}\)), policy frameworks must move away from rigid input-cost replacement paradigms. Current indemnity structures tie payouts exclusively to variable farm production inputs (such as seeds and fertilizers), leaving households vulnerable to broader asset destruction and critical living expenses during the recovery window. To address this structural vulnerability, national agencies and regional planners should transition toward bundled parametric or index-based insurance designs that trigger immediate, automated cash transfers based on satellite-derived flood heights and regional rainfall thresholds. Although such schemes may be undergoing pilot testing in a few municipalities, optimizing liquidity velocity and expanding coverage parameters beyond simple field inputs to include broader household asset protection could effectively insulate smallholder farmers from localized informal debt traps and strengthen long-term climate resilience throughout the Laguna Lake basin.

5.2. Conclusion

In conclusion, this study provides a rigorous, robustly matched evaluation of agricultural crop insurance as a post-disaster financial safety net for smallholder farmers in the flood-prone municipalities of Laguna. By shifting from traditional parametric matching models to a multi-dimensional Mahalanobis distance nearest-neighbor matching framework, the study successfully addressed localized perfect-prediction errors and established a credible counterfactual comparison across critical human capital, physical asset, and environmental variables. The empirical results reveal that while current formal crop insurance enrollment achieves its primary objective of protecting household food security by significantly reducing the probability of reduced food consumption or meal skipping following a flood event such as Typhoon Ulysses in 2020, it fails to provide comprehensive economic insulation. The treatment effects show no statistically significant reduction in the likelihood of substantial cuts in domestic non-food expenditures or continued reliance on high-interest informal borrowing. These findings suggest that although insurance prevents immediate survival crises, its current restrictive calibration to focus on compensating production inputs leaves households exposed to broader asset shocks, highlighting an urgent need for multi-risk or index-bundled policy designs to that can support long-term agricultural resilience.