Institutions ’ effect on households ’ savings in Kenya : A ranked ordered multinomial / conditional probit model approach

Savings is a vital source of investment funds especially for developing economies. However, like in many developing countries, domestic savings in Kenya remain low. Hence, posing a significant development challenge. Household savings contribute a sizeable share of domestic and national savings in both industrial and developing countries. Households should not however, be considered as fully autonomous actors without the influence of institutions. Institutions influence behavior and therefore outcomes. The institutional theory of saving thus indicates that institutional factors significantly affect the ability to save. This study uses a ranked ordered multinomial/conditional probit model to analyze the effect of institutions on households’ savings in Kenya. Data from the Financial Access National 2006, 2009, and 2013 surveys was used in the analysis. The study results show that institutional factors including the travel cost to access a saving option, trust in a saving option, information and saving expectations influence the saving levels in Kenya. It is therefore important to address the travel cost of accessing the saving options through the promotion of non-traditional means of provision of saving services, build trust in saving options, and enhance financial education in the country. Further, enhancing formal education, income levels and reducing gender gaps is important in order to improve saving performance in the country.


INTRODUCTION
Savings, defined as deferred consumption, assists in the accumulation of capital which can be used in producing further output that can possibly be consumed in future.Savings permit increases in income and consumption as well as smoothing consumption when uncertainty arises (Gersovitz, 1988).Savings facilitate investment leading to an increase in economy's productive capacity (Rillo and Miyamoto, 2016).Indeed, the importance of savings in economic development is recognized.For example, Rostow's (1956) stages of growth assert that the preconditions for take-off include an initial ability to mobilize domestic savings.Lewis's growth theory *Corresponding author.E-mail: njengam@kippra.or.ke.Tel: +254721268073.
Author(s) agree that this article remain permanently open access under the terms of the Creative Commons Attribution License 4.0 International License indicates that saving is necessary for growth because investment has to be matched by savings (Lewis, 2013).Further, the neo-classical paradigm asserts that sustained growth of output is possible only when there is an increase in the propensity to save and invest (Sahoo et al., 2001).
The literature however identifies domestic savings as a fundamental source of investment especially in developing economies (Feldstein and Horioka, 1980;Feldstein, 1983;Gersovitz, 1988;Mason, 1988;Wood, 1995).Elbadawi and. Mwega (2000) argue that in sub Saharan Africa, domestic savings are crucial in increasing investment finance required to enhance economic growth.Other benefits of higher national saving rate are a decline in vulnerability to an economy's dependence on foreign capital (Hussein et al., 2017).Indeed, Bresser-Pereira and Nakano (2002) underline the need to increase the internal savings capacity.As Mason (1988) indicates, a higher national saving rate is an important macroeconomic objective for a sizable number of developing countries.
Kenya's vision 2030 is a long term plan aimed at propelling the country to new heights of economic growth with the principal target of making Kenya a middle income country.Under Vision 2030, the economy is expected to achieve a consistent annual growth rate of 10%.To achieve this objective, the financial sector is expected to mobilize additional savings to support higher investment rates of above 30% of gross domestic product (GDP).The country has however, continued to experience low rates of savings.For example, the gross national savings (as a % of GDP) decreased from 15.9% in 2007/2008 to 11.3% in 2011/2012, well below the Medium Term Plan 1 (2008)(2009)(2010)(2011)(2012) set target of 24.4%.In 2013, the savings rate stood at 11.8% against a target of 27.7% (Republic of Kenya, 2007, 2013).In 2014 and 2015, the savings rate marginally increased to 12.2% and 12.7% (Kenya National Bureau of Statistics, 2016).Low savings generates low savings behaviour hence low capital accumulation (Hussein et al., 2017).Low savings therefore, remain a key development challenge to achieving Vision 2030 goals.
Like in other countries, households in Kenya contribute a significant part of national savings.It is therefore important to find ways of encouraging households to save in order to boost the national savings.Using the traditional savings theories, the saving literature attempts to link aspects of household saving behavior to household members' demographics and economic factors.The institutional theory of savings however, argues that institutions play a crucial role in promoting savings (Sherraden et al., 2003).According to the institutional model of saving, savings in households mainly happen through institutional arrangements and there are institutional constructs that lead to savings.
The literature on institutions' and households' saving behavior is scarce in Kenya.This study therefore endeavors to improve the understanding of saving behavior in Kenya through a close investigation of the institutional theory of saving as a crucial framework which can assist in explaining Kenya's saving performance.Hussein et al. (2017), assert that while creating economic policies for investment and growth, it's vital to understand savings behaviour.The study uses a ranked ordered multinomial/conditional probit model to analyze the effect of institutions on households' saving behavior in Kenya.Data from the Financial Access National Surveys (2006,2009,2013) is used in the analysis.

LITERATURE REVIEW
Several traditional savings theories have been identified in the literature.These include the Keynesian saving theory by John Maynard Keynes in 1936; relative-income hypothesis by James S Duessenberry in 1949; the permanent income hypothesis (PIH) by Milton Friedman in the 1950s; and the life-cycle hypothesis proposed by Modigliani andBrumberg in 1954 andAndo andModigliani in 1963.These theories have led to empirical studies which test demographic and economic factors as determinants of saving behavior.
The Keynesian approach assumes that current consumption is a function of disposable income and is a good description of consumption and savings behavior at very low levels of income, where subsistence is a predominant concern and inter-temporal considerations are absent.Keynes reasoned that as income falls relative to recent levels, people will protect consumption standards by not cutting consumption proportionally to the drop in income, and conversely as income rises, consumption will not rise proportionally.This implies that the level of income positively influences saving (Wood, 1995;Branson, 2003) The relative-income hypothesis is based on two principles (Branson, 2003).Firstly, consumers are not concerned about their absolute level of consumption but rather their consumption relative to that of the rest of the population.Thus, the consumption behavior of individuals is interdependent and not independent.The second principle states that present level of absolute relative income and levels of consumption attained in previous periods influence present consumption.It is harder to reduce a level of consumption once reached than to reduce the portion of income saved in any period.Hence, over time, consumption relations are irreversible.This principle implies that the aggregate ratio of saving to income depends on the level of present income relative to previous peak income.As present income rises relative to its previous peak, the aggregate ratio of saving to income increases, and vice versa.Branson (2003) indicates that Duesenberry's theory has however not been successful in terms of acceptance among economists.
The PIH was first proposed by Milton Friedman in the 1950s who used the term permanent income to signify the average income the household should expect over a long time horizon (Sachs and Larrain, 1993).According to PIH, consumption responds to permanent income, defined as an average of present and future incomes.Transitory income is the difference between the current and the permanent income.The prediction that transitory income is entirely saved or more generally that saving and borrowing are used solely for consumption smoothing purposes has formed the basis for a number of empirical tests of the PIH in developing countries (Agenor, 2004).This is despite Snyder (1974) asserting that the PIH was difficult to test because of measurement problem.
The life cycle hypothesis argues that income fluctuates in a systematic manner over the course of a person's life.Therefore, personal saving behavior is significantly established by one's stage in the life cycle.When people are young, their incomes are low, and they often go into debt (dissave) because they know that they will be earning more money later in their lives.During their working years, income rises to reach a peak at around middle age, and they repay the debt incurred earlier and save for their retirement.When retirement comes, income from work goes to zero and people consume their accumulated resources (Sachs and Larrain, 1993).
According to the emerging institutional theory of saving, institutional factors significantly affect the ability to save.Savings and asset accumulation are largely as a consequence of institutional arrangements entailing explicit connections, rules, incentives and subsidies (Sherraden, 1991).Indeed, such institutional factors shape and influence opportunities (Neale, 1987;North, 1990;Weaver and Rockman, 1993;Beverly and Sherraden, 1999;Guy Peters, 1999).The institutional theory of saving advances access to saving services, incentives, information, facilitation and expectations as five key institutional constructs important in influencing saving and asset building behavior especially amongst households with low incomes (Beverly and Sherraden, 1999;Sherraden, 1991Sherraden, , 1999;;Sherraden et al., 2003).
Empirical studies on the traditional saving theories reveal various factors as determinants of saving behavior.These findings show that the dependency ratio negatively affects the saving rates, age of the household head (main income earner) positively influences household saving behavior, household income has a positive effect on households' saving, household head's education is an important determinant of households' savings, household size has a negative effect of financial savings of households, urbanization negatively impacts on saving rate, household saving is affected by gender of household head, and marital status negatively affects savings.Studies on the institutional theory of savings show that institutional factors are associated with saving performance.A summary of selected empirical literature is summarized in Appendix 1.

Theoretical framework
The choice theory forms the theoretical foundation of this study.This is because the study analyzes the effect of institutions on households' choice of saving levels.Random utility model is the main common theoretical basis of choice models and is more in line with the consumer theory.The model assumes as does the consumer theory that the decision maker has a perfect discrimination capability (Ben-Akiva and Lerman, 1985;Ben-Akiva and Bierlaire, 1999).The individual is continually assumed to choose the alternative with utmost utility.However, the utilities are not known with certainty and are hence treated as random variables.In order to reflect uncertainty, the utility is therefore modeled as a random variable.For example, the utility that decision maker n associates with alternative i in the choice set n C is expressed as: where in V is the deterministic (systematic part of utility) and in  is the random term capturing the uncertainty.The alternative with the highest utility is chosen.Hence, the probability that alternative i is chosen by decision maker (2)

Empirical model specification
A multinomial choice model is used in this study, where the probability that decision maker i chooses the jth alternative is: where i x are regressors and  is a vector of regressor coefficients.The functional form for j F should be such that probabilities lie between 0 and 1 and sum over j to 1. Different functional specifications for j F , however, leads to different multinomial choice models.
The expected utilities can be modeled in terms of characteristics of alternatives (herein referred to as institutional factors) rather than the attributes of the decision maker (Mcfadden, 1974): (Choice j for i|z)=U ( ) where ij z represents a vector of institutional factors of the jth alternative for decision maker i and ij  represents the random individual specific terms which are assumed to be independently distributed each with an extreme (Gumbel) distribution.The model in Equation 4 termed a conditional logit (CL) model, however, suffers from the independence of irrelevant alternatives (IIA) assumption or the red bus-blue bus problem.This is because the model assumes that the random terms are independent across alternatives and therefore adding another alternative does not change the choices by the decision maker (Cameron and Trivedi, 2005;Wooldridge, 2002).The IIA assumption therefore leads to unrealistic predictions.
While the multinomial probit (MNP) model solves the IIA problem, it is only able to capture the decision maker attributes.The MNP model (Cameron and Trivedi, 2005;Mwangi and Sichei, 2011) is a m choice multinomial model, with utility of decision maker i from the jth choice expressed as: where ij U represents the utility derived by decision maker i from choosing alternative j , ij x is the observed attributes of decision maker i and alternative j chosen, ( ; ) Vx is the deterministic component of the utility, and j  is the error term denoting the random component of the utility.
In an MNP model, the errors are assumed to follow a multivariate normal distribution and are correlated across the choices: where  is a Kronecker product.The probability that decision maker i will select j is therefore expressed as: Hence, this study uses the multinomial/conditional probit model because it captures both the institutional factors and the decision maker attributes; and fully solves the IIA problem.To get a multinomial/conditional probit model, Equation 5 is reformulated so that the utility derived from choosing alternative j depends on the institutional factors and the decision maker attributes: ( ; ) where ij Z represents a vector of institutional factors that vary across choices.
To analyze the effect of institutions on households' choice of saving levels, we assume that a decision maker i , ranks the choices of saving levels in order of the choices indices, k k ,..., 2 , 1  , such that choice k is the preferred choice and choice 1 is the least preferred.We therefore fit a rank ordered multinomial/conditional probit model to estimate the probability of this ranking of choices as follows:

Definition and measurement of variables
The dependent variable is the choice of saving levels.In a model of individual savings, an individual (decision maker) decides on the level of savings (Gersovitz, 1988).This means that each household has a different saving rate calculated as the ratio of household saving to household disposable income (Schmidt-Hebbel et al., 1992).It is therefore assumed that households choose their level of savings as follows: Low saving levels (<1/2 of monthly income); Moderate saving levels (=1/2 of monthly income); and High saving levels (>1/2 of monthly income).The measurement of independent variables is shown in Table 1.

Data type and sources
This study uses three cross-sectional data sets from the Financial Access 2006, 2009 and 2013 national surveys.Though, the respondents were not the same in the three surveys, they share similarities.These three surveys were conducted by the Financial Access Partnership which is a public-private partnership comprising the government of Kenya and its agencies, financial sector providers, research organizations and development partners.The main goal of these surveys was to measure financial access landscape in Kenya (Financial Sector Deepening Kenya and Central Bank of Kenya, 2006Kenya, , 2009Kenya, , 2013)).The surveys were national representative with the sampling undertaken by the Kenya National Bureau of Statistics.Across the three surveys, similar questions were asked to the respondents.This study therefore uses data from these similar questions to do the analysis across the three periods.Further, the institutions' (whose effect is being analyzed in this study) environment is assumed to have remained fairly the same across the three periods of analysis.The surveys' respondents were individuals within households which were randomly selected throughout the country based on the rural and urban clusters.Every individual aged 16 years and above was eligible to participate in the surveys.In each household, however, only one individual was sampled for the interview.The total respondents were 4,418 (2006), 6,598 (2009) and 6,449 (2013).According to the choice theory which forms the foundation of this study, the decision maker requires a decision rule to arrive at a unique choice from a choice set containing two or more alternatives.Since the individuals with multiple choice of a saving option did not conform to the decision rule, they were eliminated from the study's sample size.Having specific savings expectations increases savings.

Age Age of decision maker in years (+)
The age is positively related to one's ability to hold financial savings during the working period Gender A dummy given as 1 if decision maker is female, 0 otherwise (+) Being female, an individual tends to be more cautious in spending thus increasing savings.

Level of education
Highest level of formal education completed by the decision maker: (1) None, (2) Some primary, (3) Primary completed, (4) Some secondary, (5) Secondary completed, (6) Technical training, and The more the dependants, the lower the savings.

Descriptive statistics
Table 2 present the summary statistics for the dependent and independent variables.In 2006 and 2009, majority of the respondents preferred low saving levels, that is, 76.5% (in 2006) and 65.8% (in 2009).However, in 2013, majority of the respondents (37.4%) preferred high saving levels.On institutional factors, the perceived average travel cost to access the nearest saving option is high.The mean travel cost is 0.81 and 0.94 in 2009 and 2013, respectively.These two figures are above the perceived low travel cost of accessing a saving option.The perceived average interest rate on savings is high.In 2006, the average interest rate on savings is 0.49, while in 2013, the average is 0.25.These mean figures are above the perceived low interest rate on savings.On trust in a saving option, the mean values are 0.45 in 2006, 0.24 in 2009 and 0.2 in 2013.This indicates that the perceived average trust in all the three periods of the study is high.On average, the saving option chosen is the most source of financial advice.This is revealed by the mean values of 0.64 in 2009 and 0.25 in 2013.Lastly, expectations to use savings in the saving option to deal with the highest risk are on average perceived to be high.The mean values of 0.95 in 2009 and 0.6 in 2013 are above the low perceived expectations to use savings in the saving option to deal with the highest risk.
The decision maker attributes included in this study are age, gender, level of education, income, region of residence; marital status and the number of dependants.The average age was 37.20 years in 2006, 40.17 years in 2009 and 37.21 years in 2013.The proportion of females was higher in all the three periods, that is, 61% in 2006, 62% in 2009 and 64% in 2013.The average level of education completed was primary school.This indicates that the education status across the three periods of the study is low.
On income, the average was Ksh.15,784 in 2009 and Ksh.10,922 in 2013.The distribution of income was uneven as indicated by the high standard deviation figures.In terms of region of residence, less than a half of the respondents lived in urban areas, that is, 33, 29 and 39% in 2006, 2009 and 2013, respectively.Most of the decision makers were married, that is, 65% in 2006, 68% in 2009 and 72% in 2013.Finally, the average number of dependants was 2.59 in 2006, 2.03 in 2009 and 4.40 in 2013.In all the three periods, the minimum number of dependants was zero with maximum being 20 dependants.

Estimation results
The regression results from a rank ordered multinomial/ conditional probit model are presented in Appendix 2. A post-estimation diagnostic Wald test is subsequently done to determine the model's fitness.The Wald test results are shown in Table 3.The Wald test results for the 2006 and 2009 periods of study reject the null hypothesis that all coefficients associated with institutional factors and the decision maker attributes are equal to zero.The Chi 2 statistics for the institutional factors and decision maker attributes in 2006 and 2009 are statistical significant at 1% significance level.Hence, the coefficients associated with the institutional factors and the decision maker attributes in these periods are jointly significant.However, the test results for the 2013 period does not reject the null hypothesis that all the coefficients associated with the institutional factors and the decision maker attributes are not jointly significant.Therefore, the Wald test results in Table 3 confirm the model's fitness in 2006 and 2009 data sets.Table 4 presents the institutional factors' marginal effects on choice of saving levels.
The result shows that institutions affect the saving levels as anticipated.A perceived high travel cost to access a saving option is associated with a low probability of having significant saving levels.When the travel cost to get to the nearest bank is perceived to be high instead of low, the probability that an individual will have significant saving levels in a bank will decrease by 2.4% in 2009.Conversely, the probability of having significant levels of saving in MFIs, Ascas/Roscas and Saccos increases by 0.4, 2 and 0.8%, respectively.Also in 2009, the probability of significant saving levels in MFIs will decrease by 1.3% when the travel cost to get to the nearest MFI is perceived to be high instead of low.In the same period, however, the probability of significant saving levels in banks, Ascas/Roscas and Saccos will increase by 0.4, 0.8 and 0.1%, respectively.The results also show that when the travel cost to get to the nearest Sacco is perceived to be high instead of low, the probability of having significant saving levels in a Sacco reduces by 0.06%.Correspondingly, the probability of saving significant levels in banks increases by 0.1%.The probability of having significant saving levels in MFIs also increases by 0.03%.
The trust in a saving option and significant saving levels are positively related.When the trust in banks is perceived to be high instead of low, the probability of having significant saving levels in banks will increase by 0.65% in 2006 and 0.4% in 2009.However, the probability of saving significant levels in MFIs will reduce by 0.6% in 2009.Also, the probability of having significant saving levels in Ascas/Roscas will reduce by 0.7% in 2006 and 0.3% in 2009.Lastly, the probability of having significant saving levels in Saccos will reduce by 0.5% and 0.2% in 2006 and 2009, respectively.
Similarly, when the trust in MFIs is perceived to be high instead of low, the probability of having significant saving levels in MFIs will increase by 0.3% in 2009.In the same period however, the probability of having significant saving levels in banks, Ascas/Roscas and Saccos will decline by 0.07, 0.2 and 0.02%, correspondingly.The trust in Ascas/Roscas and significant saving levels are also positively related.When trust in Ascas/Roscas is perceived to be high instead of low, the probability of significant saving levels in Ascas/Roscas will increase by 9% in 2006.On the other hand, the probability of saving significant levels in banks, MFIs and Saccos declines by 7.47, 0.2 and 1.3%, respectively.Lastly, when the trust in Saccos is perceived to be high instead of low; the probability of saving significant levels in Saccos increases by 4.5% in 2006.Conversely, the probability of saving significant levels in banks, MFIs and Ascas/Roscas reduces by 3, 0.03 and 1.3%, respectively.
The results also indicate a positive relationship between the source of financial advice and significant saving levels.The probability of having significant saving levels in a bank increases by 1.1% in 2009 whenever the bank is perceived to be the most rather than the least source of financial advice.The probability of significant savings levels in Ascas/Roscas and Saccos however reduces by 0.4 and 0.2%, respectively.Also, whenever the MFI is perceived to be the most instead of the least source of financial advice, the probability of significant saving levels in MFIs increases by 0.5%.However, the probability of significant saving levels in banks, Ascas/Roscas and Saccos decline by 0.2, 0.2 and 0.04%, respectively.
Saving expectations affect significant saving levels.The probability of significant savings in banks increases by 11% whenever the expectation to use savings in banks to deal with the highest risk is perceived to be high instead of low.The probability of significant saving levels in Ascas/Roscas and Saccos in contrast, declines by 1.1 and 0.6%, respectively.Also, the probability to have significant savings in MFIs increases by 7.5% when expectation to use savings in MFIs to deal with the highest risk is perceived to be high instead of low.However, the probability to have significant savings in banks, Ascas/Roscas and Saccos reduces by 0.8, 0.09 and 0.4%, correspondingly.Saving expectations also affect significant saving levels in Ascas/Roscas.Whenever the expectation to use savings in Ascas/Roscas to deal with the highest risk is perceived to be high instead of low, the probability of significant savings in Ascas/Roscas increases by 0.1%.However, the probability to have significant savings in MFIs and Saccos reduces by 0.3% each.Finally, the probability of significant savings in Saccos increases by 0.5% whenever the expectation to use savings in Saccos to deal with the highest risk is perceived to be high instead of low.The probability of significant savings in MFIs subsequently decreases by 0.1%.
The results on institutional factors conform to the institutional theory of saving which advances access (travel cost) to saving option, incentives (trust), information (financial advice) and expectations as important constructs in predicting saving behavior.As predicted, the travel cost to access the saving option, trust in the saving option, saving option being the most source of financial advice and saving expectations have a statistically significant relationship with significant saving levels.A perceived high travel cost to access the saving option is associated with reduced probability of having significant saving levels.Also, high trust in the saving option affect significant saving levels positively.The source of financial advice also positively affects significant saving levels.Lastly, the expectations to use savings to deal with the highest risk have a positive association with significant saving levels.
These results are consistent with the findings of other studies.In their study in Kenya, Kibet et al. (2009) found that higher transport costs had a negative effect on saving habits.Chowa et al. (2012) also showed that in Uganda, proximity to the saving option, financial education and financial incentives had positive association with higher saving performance.The implications of these results is that institutional factors including the travel cost to access a saving option, trust in a saving option, information and saving expectations influence the saving levels in Kenya.
Table 5 shows the decision maker attributes' marginal effects on the choice of saving levels.The results show that only income, education level, gender, number of dependants and region affect significant saving levels.On income, a one shilling increases in income beyond Kshs 15,783.80 in 2009 increases the probability of significant saving levels in banks by 0.3%.According to the Keynesian theory of saving, the level of income positively influences saving.Therefore, as income increases, the probability of significant saving levels also increases.
On education, an increase in one level of education beyond the primary level increases the probability of significant savings levels in banks and Saccos by 1.76 and 7.4%, respectively in 2006.Also in 2009, an increase in one level of education beyond the primary level increases the probability of significant savings levels in banks and MFIs by.0.2 and 0.4%, respectively.
However, an increase in one level of education beyond the primary level reduces the probability of significant savings levels in Ascas/Roscas by 26% in 2006.With a higher level of education, one is likely to earn more income due to one's ability to earn more.This leads to higher savings.Similar results have been found by other studies (Kibet et al., 2009).However, as income increase, one's probability to choose formal finance rather than informal finance increases (Carpenter and Jensen 2002;Ouma and Rosner 2003;Mbuthia, 2011).
The results in general indicate that compared to males, females have a lower probability of having significant saving levels in banks and Saccos but have a higher probability of having significant saving levels in Ascas/Roscas.A female's probability of having significant saving levels in banks was lower than that of a male by 0.6% in 2009.Also in 2006, a female's probability of having significant saving levels in Saccos was lower than that of a male by 5.4%.However, in 2006, a female's probability of having significant savings levels in Ascas/Roscas was higher than that of a male by 14.9%.These results are consistent with the literature that women participation in informal finance (e.g.Ascas/Roscas) is higher than that of men (Anderson and Baland, 2002).
On dependants, an increase in one dependant beyond 2 dependants reduces the probability of having significant saving levels in banks and MFIs by 0.1 and 0.2% in 2009, respectively.Also, an increase in one dependant beyond 2.59 dependants reduces the probability of having significant saving levels in Saccos by 3.5% in 2006.An additional dependant can lead to a higher household expenditure leading to reduced significant saving levels.This finding is consistent with the literature on savings including Kibet et al. (2009).
Urban residence is associated with a higher probability of having significant saving levels in banks and MFIs.The results indicate that in 2006 and 2009, urban residents' probability of having significant saving levels in banks was higher than that of rural residents by 1.3 and 0.7%, respectively.Also, the urban residents' probability of having significant saving levels in MFIs was higher than that of rural residents by 0.8% in 2009.However, the urban residents' probability of having significant saving levels in Ascas/Roscas was 15.1% lower than that of rural residents in 2006.According to Atieno (2001), location in urban area positively affects the choice of a formal saving option.This is because formal finance is more predominant in urban than in rural areas.

CONCLUSIONS AND POLICY IMPLICATIONS
This study establishes that institutions influence saving levels in Kenya.A perceived high travel cost to access a saving option is associated with reduced probability of having significant saving levels.Also, a perceived high trust in a saving option increases the probability of having significant saving levels.The source of information (financial advice) also affects significant saving levels.Lastly, the expectations to use savings to deal with the highest risk have a positive association with significant saving levels.On decision maker attributes, only income, education level, gender, number of dependants and region affect saving levels.As income increases, the probability of having significant saving levels in banks increases.An increase in one level of education beyond the primary level increases the probability of having significant savings levels in banks, Saccos and MFIs but reduces the probability of having significant savings levels in Ascas/Roscas.On gender, the results in general indicate that compared to males, females have a higher probability of having significant saving levels in Ascas/Roscas.increase in one dependant reduces the probability of having significant saving levels in banks, Saccos and MFIs.This is because an additional dependant can lead to a higher household expenditure leading to reduced saving levels.Urban residence is associated with a higher probability of significant saving levels in banks and MFIs.However, the urban residents' probability of having significant saving levels in Ascas/Roscas is lower than that of rural residents.
This study draws the following policy implications that can be used in enhancing saving levels in Kenya.There is need to address the travel cost of accessing a saving option by promoting and leveraging on digital innovation in provision of saving services.Thus, developing sustainable digital services that are affordable to low income households is crucial.These digital services must be also secured to promote their wide adoption.Further, savers need to be assured that their savings are safe for them to trust the saving services providers.This in turn will have a positive effect on their savings performance.Promotion of financial education is important because it equips savers with knowledge and information.When people are financially literate, they make better financial decisions and improve their saving actions.The national government should continue with its efforts to boast education in the country since formal education is important in enhancing saving performance.There should therefore be adequate funding of the education sector at all levels including the adult education program.Designing income re-distribution policies is critical in enhancing income levels in the country.The higher the income levels, the higher the probability of saving performance.Lastly, there is need to boost women saving performance in formal saving options by increasing their participation in the labor market in order to enhance their incomes.
The final sample size was 1,503, 2,430 and 1,843 respondents in 2006, 2009 and 2013 data sets, respectively.

Table 2 .
Descriptive statistics for the dependent and independent variables (Author's calculations based on the study data).
-: Means data missing

Table 3 .
Wald test for independent variables: Rank ordered multinomial/conditional probit model estimates for saving levels option (Source: Author's calculations).

Table 4 .
Saving levels rank ordered multinomial choice: Institutional factors' marginal effects for rank ordered multinomial/conditional probit model.