Integrating Social Protection Strategies for Improved Impact: A Comparative Evaluation of Cash Transfers and Index Insurance in Kenya

Abstract

Social protection programmes aim to use public funds to reduce poverty and vulnerability. Cash transfers have proven to be an effective social protection strategy in many contexts but are extremely expensive. Researchers have suggested that integrating multiple interventions could improve the efficiency of protection programmes, but there are few evidence-based recommendations on how to best implement such approaches. This study uses household-level panel data to estimate the marginal impacts of observed cash and index insurance transfers (subsidies) on household income in northern Kenya. Those estimates are used to simulate sample-level poverty indices as the outcome of a standard cash transfer programme and of a similar programme that reallocates a small portion of the budget as an insurance subsidy to the vulnerable. We find that the integrated programme reduces poverty to a greater degree than do cash transfers alone, highlighting the importance of protecting the vulnerable in addition to supporting the poorest.

Appendices

Appendix A: Description of key variables and summary statistics

See Table A1.

Table A1 Definition of variables and descriptive statistics in the first season (SRSD 2008)

Appendix B: HSNP and IBLI instrumental variables

This section provides details of the HSNP and IBLI instrumental processes.

HSNP

Programme participation in HSNP I is neither random nor is there perfect targeting.²⁹ As a precaution against endogenous participation, we instrumented for HSNP I participation. In theory, we had three exogenous sources of variation in participation at our disposal. The first was within-individual and captures the periods before and after HSNP receipt as well as changes to the transfer amount. The baseline IBLI survey collected data from before the first HSNP I transfers were made and, for HSNP, target sublocations were randomly designated into early and late categories. In addition, the HSNP rollout was staggered across sublocations in each early and late category.

The second source of exogeneity was between communities—HSNP I randomly targeted communities from a list of communities. Unfortunately, the list excluded a subset of communities deemed too insecure for the programme officers. While we were unable to observe if the surveyed non-HSNP I communities were or were not included in the list used for randomisation, households from the seven IBLI survey communities that were not targeted by HSNP I did have statistically significant differences at baseline from the HSNP I communities, indicating that some of them may not have been on that list or that small sample statistics were an issue here (Table B1).⁴⁴

Table B1 Summary statistics of HSNP and non-HSNP sublocations in SRSD2008

Thus, the between-community variation in participation was, in part, correlated to the variables of interest and could have posed more of a problem than a tool. We argue that such differences are largely due to community-level fixed effects, such as access to markets, remoteness, security and infrastructure. In that case, a household fixed effects model would address those issues. As a precaution against remaining differences, we included interactions between period and index division, which capture differences in average changes across seasons between index divisions. In Table B2, the robustness of our estimates to these assumptions was checked by examining the estimated mean impacts of HSNP on participants, while varying the sublocations included in the analysis. No statistically significant differences were found (Table B2, last row).

Table B2 Testing for robustness of estimated impact of HSNP across subsamples and the entire sample

HSNP I’s targeting criteria offered a third source of exogenous variation in participation. As appropriate targeting was a critical determinant of HSNP’s overall impact, the first phase randomised the beneficiary selection criteria at the sublocation level across three targeting mechanisms: age-based, dependency ratio and community-based targeting (CBT). If compliance had been perfect, we could have controlled for the targeting dimension (assuming that we observed it) and eligibility would have been exogenous. Unfortunately, HSNP targeting compliance was far from perfect.²⁹ However, assuming partial compliance, the intended eligibility criteria would still generate exogenous variation in participation and thus could also be used as an instrument variable.

The community-based targeting (CBT) presents a difficulty with respect to the above strategy. We did not observe the CBT process nor did we have information on the dimensions used by community members for their targeting. HSNP’s own impact evaluation provided some insight into the process though. In it, Hurrell and Sabates-Wheeler⁴⁵ controlled for household characteristics that were likely to have played a role in the CBT. We used a similar set of those characteristics from the period before the first transfers in each CBT sublocation and employed a probit model to predict the likelihood of HSNP participation in each CBT sublocation. We then constructed a household-level intent-to-treat variable within each CBT sublocation that was equal to one if the predicted likelihood of HSNP participation was greater than 50 per cent. By this approach, the generated intent-to-treat variable correctly sorted 69 per cent of households in each CBT sublocation during the season that HSNP launched in that sublocation.⁴⁶

Altogether, we had three potential sources of exogenous variation, one that distinguished between targeted and non-targeted sublocations, one that captured variation in date of initial transfers and variation in transfer size and one that helped distinguish which households within a target community were most likely to have received a transfer.

The first stage used to instrument for HSNP participation is described by Equation B.1.

$${\text{HSNP}}_{lit} = \delta_{0} + \delta_{1} {\text{ITT}}_{lit} + \delta_{2} X_{li1} + \delta_{lt} + \varepsilon_{lit},$$

(B.1)

where

$${\text{ITT}}_{it} = \left[ {\begin{array}{*{20}l} 1 \hfill & {{\text{if }}A_{li} = 1 {\text{and }}L_{lt}^{A} = 1 {\text{and }}S_{lt} = 1 } \hfill \\ 1 \hfill & {{\text{if }}B_{li} = 1 {\text{and }}L_{lt}^{B} = 1 {\text{and }}S_{lt} = 1} \hfill \\ 1 \hfill & {{\text{if }}C_{li} = 1 {\text{and }}L_{lt}^{C} = 1 {\text{and}}\;S_{lt} = 1} \hfill \\ 0 \hfill & {\text{otherwise}} \hfill \\ \end{array} } \right]$$

$$A_{li} = \left[ {\begin{array}{*{20}c} 0 & {\hbox{max}\; {\text{household age}}_{li1} < 64} \\ 1 & {\hbox{max}\; {\text{household age}}_{li1} \ge 64} \\ \end{array} } \right]$$

$$B_{li} = \left[ {\begin{array}{*{20}c} 0 & {{\text{dependency ratio}}_{li1} < 0.54} \\ 1 & {{\text{dependency ratio}}_{li1} \ge 0.54} \\ \end{array} } \right]$$

$$\begin{aligned} C_{li} = \left[ {\begin{array}{*{20}c} 0 & {\hat{p}_{{lit}} , < \,0.5} \\ 1 & {\hat{p}_{{lit}} , \ge \,0.5} \\ \end{array} } \right] , \quad p_{{lit^{\prime } }} = \gamma_{0} + \gamma_{0} X_{{lit^{\prime } }} + \varepsilon_{li1}, \hfill \\ {\text{where }}p_{{lit^{\prime } }} = 1\quad {\text{if }}S_{{l,t^{\prime } }} = 0, S_{{l,t^{\prime } + 1}} = 1\quad {\text{and}}\quad L_{l}^{C} = 1; \hfill \\ \end{aligned}$$

$$L_{l}^{A} = 1\quad {\text{if sublocation }}l \,{\text{is an age-targeted community}};$$

$$L_{l}^{B} = 1 \quad{\text{if sublocation }}l \,{\text{is a dependency ratio-targeted community }};$$

$$L_{l}^{C} = 1 \quad{\text{if sublocation }}l \,{\text{is a CBT-targeted community }};$$

$$S_{lt} = 1 \quad {\text{if HSNP transfers are being distributed in sublocation }}l {\text{in time period }}t;$$

$$X_{li1} = {\text{Baseline household characteristics including maximum age}}, {\text{dependency ratio, and }}\hat{p}_{li1};$$

$$\delta_{{{\text{l}}t}} = {\text{Index division * period fixed effects}}.$$

As mentioned above, this approach has a few potential issues. First, there were statistically significant differences in the observed means of household characteristics in HSNP and non-HSNP sublocations, a possible symptom of fundamental differences that could conflate our estimates. If those differences extended beyond levels to rates of change, this would mean that the non-HSNP communities were not a valid counterfactual and should not have been included in this analysis. The second was that we did not observe the CBT ranking, and our generated ranking may perform poorly as an ITT variable.

To determine the magnitude of these potential issues, we ran our main regression, excluding the IBLI variables, in three ways: (1) using only dependency ratio- and age-targeted HSNP sublocations, (2) using all non-HSNP sublocations and dependency ratio- and age-targeted HSNP sublocations and (3) using the entire sample including CBT sublocations (Table B2). Although there are a few small changes to the parameter estimates between the three models, all were the same sign, none were statistically different and all were within an order of magnitude. We also estimated the average impact among participants. The results are found at the bottom of Table B2. All three estimates are indistinguishable from zero and statistically indistinguishable from each other. We erred on the side of power and used the full sample in the main analysis.

Using the full sample, we regressed HSNP participation onto the ITT variable to provide some indication of the strength of our IV (Table B3). Here we saw that the IV alone was able to account for about 40 per cent of the variation that we observed in participation and the correlation was robust to index region-period controls and household fixed effects.

Table B3 The impact of HNSP’s intent-to-treat on likelihood of HSNP participation (OLS)

IBLI

There was also a risk of endogeneity in IBLI purchases. To identify the causal impacts of IBLI purchases, we used exogenous variation in IBLI purchasing behaviour, which was caused by the random distribution of premium discounts.

Before each sales season, premium discount coupons were randomly distributed to 60 per cent of the sample. The discounts provided by the coupons ranged from 10 to 60 per cent and were only valid for the sales season immediately following their distribution. By LRLD2013, 2,472 coupons had been distributed to the households. Because there had been six distribution events, nearly every household had received at least one discount coupon, and most households had received more than one coupon (Table B4).

Table B4 Number of discount coupons received by each household by LRLD2013

The log-level of discount had a strong and positive impact on the log-level of coverage purchased (Table B5). The discount elasticity of demand was near 0.4.

Table B5 Impact of discounts provided by coupons on IBLI purchases

Appendix C: Demand for insurance

We assume that households make their demand decisions in two steps. First, they choose whether to purchase. Second, they choose how much to purchase provided they have chosen to purchase. The assumption that these are two different decisions allows factors to affect one aspect of demand without affecting the other. For example, in our study, a household’s trust in the vender or the distance from a household to the nearest vender may have been an extremely important factor in one decision but not in the other. The parameter estimates are found in Table C1. We include the full set of covariates included in the main regression. The implied average marginal price elasticities of demand are included at the bottom of the table. Note that the final sales season (LRLD2013) was not used in these demand equations because that season was not used in the main analysis due to lagging.

Table C1 IBLI demand equation estimates

Appendix D: Bootstrapping

We used a bootstrapping process to inform on the precision of the FGT estimates and to test for differences between the programmes. To do so, we used a double bootstrapping approach to transform bootstrapped estimates into t-statistics, which were then used to identify the 95 per cent confidence intervals around the original estimates.

A summary of that process follows. Here we used the subscripts FULL, BOOT and REBOOT to indicate that a statistic was estimated using the full original sample (FULL), a sample developed by resampling with replacement from the full original sample (BOOT) and a sample developed by resampling with replacement from the BOOT sample (REBOOT), respectively.

• 1. Estimate the FGT metrics using the full sample generating the estimates described in Eqs. (1) and (2) in this article: $\widehat{\text{FGT}}_{\text{FULL}}^{\alpha }$.

• 2. Resample with replacement from the full sample to generate the sample BOOT.

• 3. Use the resample to estimate the FGT metrics: $\widehat{\text{FGT}}_{\text{BOOT}}^{\alpha }$.

• 4. Resample with replacement from BOOT, generating the sample REBOOT.

• 5. Estimate the FGT metrics using the REBOOT sample: $\widehat{\text{FGT}}_{\text{BOOT, REBOOT}}^{\alpha }$

• 6. Repeat steps 4–5, A times.

• 7. Repeat steps 2–6, B times.

Steps (1)–(7) result in three sets of estimates. For ease, we use the characters $\tau , \varphi , {\text{and}} \omega$ to reference each set.

$$\tau = \left\{ {\widehat{\text{FGT}}_{\text{FULL}}^{\alpha } } \right\}; \varphi = \left\{ {\widehat{\text{FGT}}_{1}^{\alpha } , \ldots , \widehat{\text{FGT}}_{A}^{\alpha } } \right\} ; \omega = \left\{ {\widehat{\text{FGT}}_{1,1}^{\alpha } , \ldots , \widehat{\text{FGT}}_{1,B}^{\alpha } , \widehat{\text{FGT}}_{2,1}^{\alpha } , \ldots , \ldots , \widehat{\text{FGT}}_{A,B}^{\alpha } } \right\}.$$

The above estimates are used to develop confidence intervals around $\widehat{\text{FGT}}_{\text{FULL}}^{\alpha }$ as follows:

$${\text{CI}}^{\rho } = \left[ {\widehat{\text{FGT}}_{\text{FULL}}^{\alpha } - t^{{{\raise0.7ex\hbox{$\rho $} \!\mathord{\left/ {\vphantom {\rho 2}}\right.\kern-0pt} \!\lower0.7ex\hbox{$2$}}}} *SE_{{\widehat{\text{FGT}}_{\text{FULL}}^{\alpha } }} ; \widehat{\text{FGT}}_{\text{FULL}}^{\alpha } + t^{{(1 - {\raise0.7ex\hbox{$\rho $} \!\mathord{\left/ {\vphantom {\rho 2}}\right.\kern-0pt} \!\lower0.7ex\hbox{$2$}})}} *{\text{SE}}_{{\widehat{\text{FGT}}_{\text{FULL}}^{\alpha } }} } \right],$$

where

$${\text{SE}}_{{\widehat{\text{FGT}}_{\text{FULL}}^{\alpha } }} = \frac{{\sigma_{\varphi } }}{\sqrt A }$$

$$t^{{{\raise0.7ex\hbox{$\rho $} \!\mathord{\left/ {\vphantom {\rho 2}}\right.\kern-0pt} \!\lower0.7ex\hbox{$2$}}}} = {\text{the}}\,{\raise0.7ex\hbox{$\rho $} \!\mathord{\left/ {\vphantom {\rho 2}}\right.\kern-0pt} \!\lower0.7ex\hbox{$2$}}{\text{th ranked}}\, t_{i} ;\, t_{i} = \frac{{\widehat{\text{FGT}}_{\text{FULL}}^{\alpha } - \widehat{\text{FGT}}_{\text{BOOT}}^{\alpha } }}{{{\text{SE}}_{{\widehat{\text{FGT}}_{\text{BOOT}}^{\alpha } }} }}$$

$${\text{SE}}_{{\widehat{\text{FGT}}_{i}^{\alpha } }} = \frac{{\sigma_{i,\omega } }}{\sqrt B }.$$