• Open Access

The potential trade-off between artisanal fisheries production and hydroelectricity generation on the Kafue River, Zambia


Andrew M. Deines, 177 Galvin Life Sciences, University of Notre Dame, Notre Dame, IN 46556, U.S.A.
E-mail: adeines@nd.edu


1. Freshwater resource managers are increasingly obligated to consider the impacts of large river engineering projects on ecosystem services. We evaluated the effect of altered water regime from the operation of a large dam on the production of the downstream tropical floodplain fishery of the Kafue River, Zambia. We compared the benefits of increased hydropower relative to potentially lost fishery production.

2. We compiled a long-term data set consisting of experimental gillnet catches, artisanal harvesting effort and monthly river flows for 25 years prior to and 29 years after the 1977 completion of the upstream Itezhi-Tezhi Dam. As a metric of the flood regime, we calculated a canonical correlation score for each hydrological year before and after dam closure. For the period following dam construction, we used the Muskingum method of flood routing to estimate ‘no-dam’ flows through the fishery area and downstream hydroelectric turbines at the Kafue Gorge Dam.

3. We compared 16 alternative models of catch per unit effort (CPUE) with and without an effect of water regime on fish population growth rate. Using the two best fitting models, we estimated the total observed fishery harvest and simulated ‘no-dam’ fisheries harvest and found no significant effect of altered water regime on fishery production.

4. We estimate that the large upstream dam increases downstream hydropower production by about $18 million USD per annum. The reduction in fishery production caused by the altered water regime is not significantly different than zero, although the average reduction amounts to about $2.3 million annually. The total estimated value of harvest ranges from $1.3 million to $56 million annually.

5. Large observed declines in fish abundance over the 54-year study period are attributed primarily with similarly large increases in total fishing effort in this mostly open-access artisanal fishery.

6. These results contrast with other examples of the effects of flow alteration on fish, probably because levels of fisheries exploitation on the Kafue River are very high relative to better studied regions on other continents; our focus on the whole fish community; and the unprecedented length of the time series we considered. If the goal is to sustain fishery production, investments in altering flow regime are likely to be less effective than investments to decrease fishing effort.


Maximising the production of ecosystem services is a desirable outcome for resource management, particularly when increasing the provision of one service decreases the provision of another. In these instances, it is important to estimate the value of ecosystem services – and the trade-offs between potentially competing services – to efficiently use resources. We estimate two major ecosystem services provided by the Kafue River, Zambia, hydroelectricity generation and fisheries production, and discuss management implications of the potential trade-off between them.

Globally, there is increasing scientific and policy interest in ‘environmental flows’, that is, flows that more closely mimic natural hydrological patterns, as a tool to sustain ecosystem services and human well-being and better balance potential trade-offs in flow alteration and other ecosystem services. This interest derives from the fact that large, impounded rivers around the world are facing competing uses of water resources as seasonal flood regimes are altered for hydropower or irrigation by storing water from high-flow seasons for use in dryer periods. Changes in water flow can impact river biodiversity by altering the physical channel structure, disrupting organisms’ life history patterns, severing connectivity and by encouraging species invasions (Dudgeon et al., 2005). These water regime changes also impact fisheries production and related ecological services (Welcomme, 2008; Poff & Zimmerman, 2010).

The task of assessing trade-offs in ecosystem services as a result of flow alteration is especially relevant in Africa where there are 20 large dams now under construction or advanced planning, 42 undergoing expansions or rehabilitation and 83 proposed new dams (International Rivers Africa Program, 2010). Little knowledge exists for similar rivers about the effect of changes in flow regime on ecosystem services. Most work on environmental flows has occurred in the northern hemisphere in mostly small streams and rivers, has focused on particular species (e.g. Chinnok salmon, Oncorhynchus tshawytscha; Service, 2011) or has been conducted in the absence of significant fisheries. The responses of ecosystems, particularly fisheries, to changes in hydrological regime, thus remain largely unstudied in Africa (Poff & Zimmerman, 2010). This general paucity of data regarding ecological responses to hydrological change has led to the development of the comprehensive Integrated Basin Flow Assessment (IBFA; King & Brown, 2010) approach over the last 15 years, which integrates expert ecological knowledge and socio-economic factors to inform river management decisions and observe hydrological effects on ecosystems as they are implemented. The IBFA process may yield data and effective management outcomes in the future, but it is still too soon to see the long-term effectiveness. A retrospective analysis of previous hydrological changes and fisheries response could provide immediate guidance for river resource managers. The Kafue River, Zambia, presents an opportunity to examine the importance of water regime in a highly relevant social, economic and environmental context: heavily exploited artisanal fisheries.

The Kafue River provides many ecosystem services including irrigation, cattle grazing and wildlife and bird habitat in national parks and Ramsar wetland sites (Lagler, Kapetsky & Stewart, 1971; Mumba & Thompson, 2005). Hydropower generation and capture fisheries are, however, the most direct and obvious uses of this water resource. They are important to the public and policy makers, the most likely to be in direct conflict for water, and have the most complete records available.

Pre-dam fisheries studies conducted on the Kafue in the mid-1970s found annual fisheries harvest to be significantly correlated to flood regime (Chapman et al., 1971; Lagler et al., 1971; Dudley, 1974; Kapetsky & Illies, 1974; Welcomme, 1975; Muncy, 1977). These relationships between hydrological regime and fishery yield are complicated, however, because these studies focused on total harvest which overlooks the reciprocal relationships between fish abundance and fishery effort. Moreover, no studies have evaluated the impact of change in water regime on Kafue fisheries since dam construction was completed, although negative impacts on fisheries are often assumed (Chipungu, 1981; Schelle & Pittock, 2005). Changes in other ecosystem properties such as decreased extent of wetland habitat (Munyati, 2000; Mumba & Thompson, 2005) and water chemistry (Obrdlik, Mumeka & Kasonde, 1989) have also been reported. In response to these observations, environmental flows have been advocated by some stakeholders (Schelle & Pittock, 2005).

We hypothesised that the upstream construction of Itezhi-Tezhi (ITT) Dam altered the Kafue River water regime and reduced downstream fish abundance. To determine the relationship between water regime and fishery production, we compiled catch per unit effort (CPUE) from scientific surveys, artisanal fishing effort, total harvest and monthly mean discharge hydrographs from the Kafue River for the years 1954–2010 and developed state-space population growth models to test the effects of flood regime on the multi-species fish community population growth rate. Using flow data from above ITT reservoir, we simulated the water regime on the Kafue River for the post-ITT Dam period as if the dam had not been constructed. Using this ‘no-dam’ simulated flow, we used the best fitting population growth models to estimate fisheries production and hydroelectric generating capacity downstream of ITT. Finally, we compared the revenue derived from these ecosystem services with and without water regulation by ITT Dam.


Site description

The Kafue Flats are a large, flat floodplain of the Kafue River in Zambia (Fig. 1). Historically, after the onset of the rainy season in November, flood waters began to rise from a dry season low of about 30 m3 s−1 and peaked in April or May at more than 1500 m3 s−1. More than 6000 square kilometres were under water during typical flood stage; for comparison, this is an area more than 10 times that of Lake Constanz (Germany, Austria and Switzerland) and roughly one-third the size of Lake Ontario of the Laurentian Great Lakes. The fishery has remained primarily artisanal since its inception in the 1950s. Fishers typically use dugout canoes or fibreglass ‘banana boats’ and multifilament gillnets. Although illegal, large (>100 m) hand-drawn seines of <1 mm mesh are also common, as are monofilament gillnets and gillnets of mesh less than 50 mm mesh. Fish-driving is also practiced, by beating the water to drive fish into gillnets. The main large-bodied fish species harvested are cichlids, principally species of the genera Oreochromis, Tilapia, Sargochromis and Serranochromis, as well as Clarias catfish and the Kafue pike, Hepsetus odeo. Most species of all sizes are also harvested and marketed, including many Cyprinid species and to a lesser extent, Mormyrids and Synodontids.

Figure 1.

 Map of the Kafue River, Zambia, in southern Africa (insert) focusing on the locations used to describe the fishery and hydrological regime from the upstream Hook Bridge and Itezhi-Tezhi Dam, through the Kafue floodplain fishery area (grey hatched) and to the downstream Kafue Gorge Dam and power station. Fishery sampling locations (▲) included in this analysis are from west to east: Namwala, Maala, Chunga Lagoon, Nyimba, Mazabuka and Chinyanya.

The main fishery area is bracketed by two dams (Fig. 1). The downstream Kafue Gorge Dam (KG), completed in 1972, was originally installed with 600 megawatt (MW) of generating capacity, but was later expanded to 900 MW (Smardon, 2009). The estimated maximum capacity of the reservoir at KG is 800 million m3 (van der Knaap, 1994), and without additional water regulation during the dry season, the dam would only have enough water available for 207 MW of power output. To increase hydropower capacity, the ITT Dam was built at the upstream end of the Kafue Flats in 1977. With a much larger reservoir holding 4950 million m3, ITT provides steady flow downstream during the dry season, reducing the risk of insufficient water for maximum power generation at KG (Smardon, 2009).

Data compilation

Experimental gillnet fisheries data from scientific surveys (Fig. 2a) were compiled from published literature and unpublished data from the Zambian Department of Fisheries (DoF) for the years 1954–2010 and were assumed to be an index of total fish abundance as CPUE calculated in mass per metre of net per night (Williams, 1960, CSO, 1970, 1978, 1984; Everett, 1971, 1974; Kapetsky, 1974). In this analysis, we included only those mesh sizes and sampling locations (Table S1) as those reported in the pre-dam data to produce a data string that is comparable across all years. Misidentification and changing nomenclature of taxa complicated comparisons of catches across time. We unified species to current nomenclature to the genus or species level (Mortimer, 1965; Skelton, 2001; Froese & Pauly, 2012) and made sure representatives of each taxa at genus level were represented whenever species-specific data were available for specific collections (Table S2). Fishing effort data on the Kafue were available as the number of gillnets and/or the fleet size as the number of boats in use in each year (Fig. 2b) (Mortimer, 1965; Kapetsky & Illies, 1974; CSO, 1978, 1984; Muyanga & Chipungu, 1978; Chipungu, 1981; Lupikisha, 1992, 1993; DoF, 1993; DoF & CSO, 2007). Total harvest (Fig. 2c) estimates were as reported for 1954–1996 by Nyimbili (2006), but comparable total harvest data after 1996 were not available.

Figure 2.

 Fishery and hydrological data for the Kafue River fishery, Zambia. (a) Box plots of annual experimental gillnet catch per unit effort showing the mean, quartile range, approximate 95% confidence interval and outliers outside this interval. (b) Annual artisanal effort in metres of gillnet (○) with the last data point slightly offset on the x-axis for visibility and number of boats (•). (c) Annual total reported harvest and (d) mean monthly discharge at Itezhi-Tezhi. The first vertical dashed line indicates the date of closure of Kafue Gorge Dam (downstream), while the second indicates the closure of Itezhi-Tezhi Dam (upstream).

Monthly mean discharge (m3 s−1) at ITT was obtained from the Food and Agriculture Organization (FAO, 1968) for the years 1953 through 1963, the Zambian Ministry of Energy and Water Development for 1960 through 1991 (unpublished data) and Nyimbili (2006) from 1980 through 2005 (Fig. 2d). Using the before and after damming mean monthly hydrographs, we preformed canonical correlation analysis (CCA) with the CCA package (González et al., 2008) in R (vers 2.14.0; R Development Core Team, 2011), to simplify the hydrograph into one variable for each year for use in population modelling. CCA is similar to principal component analysis (PCA) in that it assists description of multivariate data by describing new axes that are linear combinations of the original data that better describe patterns of sample variance. Whereas PCA simplifies patterns of correlations of multivariate data measured among individual samples, CCA identifies patterns between mutually exclusive groups of samples. Specifically, PCA finds the eigenvalues and eigenvectors of the sample variance–covariance matrix, while CCA is constrained to maximise the ratio of within-group and between-group variance–covariance matrixes. In this way, the first canonical variate represents a new orthogonal axis that most effectively discriminates the sample groups under consideration, in this case, hydrographs before and after damming (Zelditch et al., 2004). Thus, the correlation score associated with this axis in each year was used as a flood regime metric in the fisheries model described below.

Modelling the impact of flood regime on fishery production

We used a multivariate autoregressive state-space (MARSS) model to fit time series of experimental CPUE, fisheries effort and water regime to population growth models using maximum likelihood estimation. This state-space approach allowed the simultaneous estimation of the unobserved state process of fish abundance (CPUE) and fisheries effort (metres of gillnet) with observation error and including the effect of water level as a covariate to the CPUE process.

The MARSS model assumes Gompertz population growth expressed as a linear model by taking the natural log of CPUE and effort (See Supporting information). Gompertz growth is similar to logistic growth (commonly known as the Ricker model in fisheries) but implies growth is density dependent on the natural log of the population, meaning the growth rate varies exponentially with population size rather than linearly as with logistic. This formulation has considerable computational and statistical advantages and performs well in density-dependent populations (Dennis & Taper, 1994). In multivariate state-space, state and observation processes are arranged into a system of equations expressing population growth in matrix form including covariates (Holmes & Ward, 2012).


where xi are state vectors at time t defined by a state process equation (eqn 1a) with i superscripts representing the estimated variates (v) CPUE and effort, or the covariates (cv), the CCA score representing water regime. Bi are parameter matrices to be estimated. Process error, wt, is modelled as a multivariate normal distribution with mean zero and variance–covariance matrix of process (environmental) stochasticity Q. Vectors of observed data yi are related to the process states through the observation process equation (eqn 1b). Z are identity matrices that associate one or more observations to unobserved state processes, with a parameters that linearly scale multiple observations of the same state and multivariate normal observation error vt with R variance–covariance matrixes.

In total we specified three state processes: CPUE, effort and water. Following standard practice, we demeaned and standardised all data and used the resulting z-scores for estimation. We make the simplifying assumption that processes do not covary, that is for example, the variance of CPUE and effort do not scale with water level. The alternative, including all possible covarying effects, greatly increases computation and decreases power (Bolker, 2008). We also fixed CCA score variance at unity to give the process model the flexibility to exactly equal the true covariate values; thus, the covariate processes are not modelled but exactly specified (Holmes & Ward, 2012). Initial results indicated process errors <1 e−15 in all cases, leading to instability in the estimation algorithm. We therefore fixed process error for CPUE and effort at a trivially small value, 1 e−5. Fixing process error in this way assumes that most of the variation present in the data stems from how the data were observed and reported, not environmental variation in the populations themselves. This assumption is consistent with the inherent difficulties of measuring fish stocks and compiling data from multiple studies.

We specified two observation vectors for the effort process, metres of gillnet and boat counts. Gillnets are the more accurate measure of actual fishing effort, and therefore, boat counts are linearly scaled to gillnets by estimating the number of gillnet metres per boat in vector a(v). Assuming this scaling is constant for all years with boat data is reasonable, given that the common-type canoe has changed little from early accounts and even pictures of the fishery (Lagler et al., 1971) and that these boats likely always carried their full capacity of nets, which are limited to only a few by the size of the boats and yet are cheap in terms of investment relative to the boat itself.

We created 16 base models that variously included or excluded all combinations of the effect of density dependence for CPUE and effort, of CPUEt-1 on effort and watert − 1 on CPUE. Models were estimated using the MARSS package (vers 2.8; Holmes & Ward, 2012) in R and ranked by Akaike’s information criteria (corrected, AICc). For the two best models, we estimated the 95% bootstrap confidence intervals (CI) for the parameters and the CPUE and effort states. To perform the bootstrap, we resampled the mean CPUE data for each year and parametrically simulated 500 bootstrap replicates of gillnets and boat counts based on the error estimated in each model. Using the original model estimate for initial conditions, these bootstrap data were used to re-estimate parameters and states for each model and estimate the 95% CI for the observations of CPUE, gillnets and boats.

Sensitivity analysis

The treatment of observation error in state-space models is critical to estimating environmental effects on population dynamics (Linden & Knape, 2009) and the key motivation for sensitivity analysis. Ives et al. (2003) recommended against directly fitting observation error from data without independent estimates and suggested instead fitting the model with rough estimates of observation variance and then test the sensitivity of the model to these estimates. We therefore fit models while estimating observation error, fixing observation error with rough estimates and fixing observation error at unity. We fixed rough estimates of observation error as follows. For CPUE, we used the mean variance from 10000 bootstrap samples of each year with ≥4 sampling periods (=0.069). Gillnet observation variance derives from the variance observed in gillnets used in several areas over a 12-month period in 1972 (=0.321, CSO, 1978). To roughly estimate observation error in boat counts, we assumed that fishermen and boats were censused with the same error (counts of fishermen roughly coincide with boats counts; data not shown), and we therefore conservatively used the maximum variance in any year between multiple counts of boats and/or counts of fishermen, which occurred in 2006 (=0.085). This yielded a total of 48 models for comparison.

We also tested the ability of the MARSS and AICc model selection procedure to select the correct model by re-running the selection for each model replacing the data with the estimated state vectors. That is, for each of the 16 base models, the state vectors represented data simulated from a known model with known parameters, and we tested whether running model selection on that simulated data would recapture the model from which it was generated.

Simulating flood regime and hydroelectric generating capacity

The ITT Dam is intended to provide a more consistent supply of water during the dry season and thereby keep the turbines at KG Dam capable of running at full capacity year-round. We applied the widely used Muskingum method of flood routing, or predicting downstream flows based on known upstream flows (McCarthy, 1938) to predict actual and ‘no-dam’ flows at KG based on flows at ITT. Detailed methods are provided in the supporting information and summarised below.

First, we estimated simulated ‘no-dam’ discharge for ITT using daily discharge at Hook Bridge monitoring station upstream from ITT reservoir (Figs 1 & S2) to estimate the inflows into the Muskingum model and discharges at ITT to represent outflows. Similarly, to model hydropower production at the downstream KG Dam under both actual and simulated ‘no-dam’ flows, we used the actual, and this estimated ‘no-dam’ simulated daily discharge at the ITT Dam to represent inflows to the Muskingum model and flows at KG Dam to represent outflows. To calculate the ‘no-dam’ scenario water regime CCA metric, we multiplied the CCA loadings calculated from the observed ITT hydrograph by the ‘no-dam’ simulated mean monthly flow at ITT.

To estimate hydropower production under alternate observed and ‘no-dam’ flow regimes, we took the annual average for each simulated KG hydrograph as an expected difference in flow attributed to the ITT Dam’s influence on the Kafue River’s flood regime. Reduced discharges into the reservoir at KG do not necessarily imply reduced generating capacity as the KG generating station can still operate until the reservoir is emptied, and even then it can use the reduced inflow directly. Therefore, we conservatively compute the minimum reduction of power output in the ‘no-dam’ scenario relative to that which uses the full capacity of the reservoir. To estimate the value of this reduction in generation, we use the replacement cost of importing the foregone electricity, which is estimated at about $31 per MWh (PB Power, 2006). Using the forgone difference makes the reasonable assumption that all the electricity that can be generated is readily consumed and that imported electricity is available and would be consumed at this price.

Harvest revenue

Using the actual and ‘no-dam’ CCA water regime metrics and the parameter estimates of the best models, we calculated harvest for each year up to 2006 for each of 500 bootstrap CPUE replicates by rescaling the state process estimates and taking the product of the CPUE (kg m−1 night−1), effort (m) and activity rate of fishers (proportion of nights per year spent fishing), for which a range of estimates exist. We multiplied the upper 95% CI of CPUE × effort by an activity rate = 1 (365 days of fishing per year) to retrieve maximum total harvest per year assuming that all nets were deployed every night of the year. For an intermediate estimate of annual harvest, we multiplied the median of the product by the ‘standard’ activity rate reported by the DoF = 0.65 (237.25 days) (DoF, 1993). For a minimum harvest estimate, we used activity rate = 0.4 (146 days), the minimum reported in the compiled data (Lupikisha, 1993).

There is no known record of the price of fish in Zambia for any years in the time series analysed. We can provide only a point estimate of the value of harvest in 2006, the last year of the time series analysed, based on a survey at six markets in Lusaka of the retail price of fresh tilapia (species of the genus Oreochromis) in 2008: about 15 000 Kwacha per kg (B. Klco and A.M. Deines, unpublished data). We assume that the real price of fish was relatively constant between 2006 and 2008, adjust for inflation and convert the nominal 2008 price in kwacha into real 2006 dollars and calculate the harvest revenue, D, in US dollars as


where H is the harvest described above and P is the market price of tilapia in 2008, CPI is the consumer price index (CSO 2008) and z is the 2006 exchange rate for Zambian Kwacha to USD, equal to 2.9 e−4 (http://www.xe.com, accessed 3-18-12).


The simulated ‘no-dam’ flow at ITT estimated from Hook Bridge flows is shown in Fig. 3a. The CCA flood regime metric (Fig. 3b) captured 95% of the variance between the pre- and post-dam mean monthly hydrographs, with 67% of the CCA loading assigned to the historical low water months of September and October. These CCA loadings and the ‘no-dam’ flows were used to simulate what the CCA metric would have been had the ITT Dam not existed (Fig. 3b) and these appear consistent with the pre-dam regime.

Figure 3.

 Hydrograph modelling results. (a) Simulated ‘no-dam’ mean monthly discharge in cubic metres per second at Itezhi-Tezhi and (b) the canonical correlation analysis (CCA) scores that represent a flow index which best discriminates hydrographs before and after dam construction (•) and the simulated ‘no-dam’ CCA scores (▪). The first vertical dashed line indicates the date of closure of Kafue Gorge Dam (downstream), while the second indicates the closure of Itezhi-Tezhi Dam (upstream).

The impact of flood regime on fishery production

We excluded from consideration models that were non-stationary and models that estimated negative intrinsic population growth (uCPUE). Non-stationary models have a variance distribution that increases over time, which can cause the model estimates to diverge and the estimation algorithm to become unstable, indicating the particular model is structurally a poor representation of the system (Holmes & Ward, 2012). Similarly, models that estimate negative intrinsic population growth are not biologically feasible representations of the fishery, which suggests the particular parameterization of these models does not adequately represent the system. We also excluded models with a difference in AICc (dAICc) greater than 10 from the best model. Full modelling results for all models are presented in Table S3.

The difference in dAICc between the three best models (dAIC = 0.06) provides very little support for choosing from among these models, although there is some support from dAICc that these three models are better fits to the data than the other models (dAIC ≥ 2.25) (Table 1). All three models included density dependence (b1-1) for CPUE and negative effects of effort on CPUE (b3), consistent with biological intuition. Contrary to our expectation, two of the top three models, including the best model, did not include an effect of water regime on CPUE. The best model included a small positive effect of CPUE on effort (b2). The second best model included a small negative impact of water regime on CPUE and implies a strictly linear increase in effort over time. The third model also did not include a water regime effect or an effect of CPUE on effort, but did include weak density dependence on effort.

Table 1. Results of model selection and parameter estimates. Estimates in bold indicate parameters that were fixed a priori as part of the model specification and 95% confidence intervals are indicated in italics
 −LL dAICc UCPUEe Ueffort CPUE density dependent (b1-1)Effect of CPUE on effort (b2)Effect of effort on CPUE (b3)Effort density dependent (b4-1)Effect of water regime on CPUE (b5) RCPUE Rgillnet Rboats
  1. ind = independent (not included), *significant, only evaluated for models 1 & 2.

  2. − LL, negative log likelihood; dAICc, delta Akaike’s information criterion, corrected; UI, growth parameters for catch per unit effort (CPUE) and effort, respectively; Ri, observation variance parameters for CPUE, gillnets and boats, respectively. See text for details.

1−93.2107.4 e−2
1.5 e4 to 1.7
8.0 e−2*
6.2 e2 to 9.6 e2*
 0.30 to −0.13
5.5 e−3
2.4 e2 to 2.7 e2
− 0.1*
0.2 to −7.4 e2
ind ind 1.4 e−1
1.3 e1 to 2.5 e1
2.2 e−1
8.4 e2 to 3.5 e1
2.4 e−1
8.8 e1 to 4.3 e1
2− e−2*
2.7 e2 to 1.4 e1
7.9 e−2
6.4 e2* to 9.3 e2
− 0.2*
0.4 to −0.1
ind − 0.1*
0.2 to −8.8 e2
ind − 2.0 e−2
0.1 to 4.2 e2
1.4 e−1
1.3 e1 to 2.4 e1
2.2 e−1
9.8 e2 to 3.5 e1
2.4 e−1
8.8 e2 to 4.1 e1
3− e−28.31 e−2− 0.22 ind − 0.14− 3.21 e−3 ind 1.4 e−12.2 e−12.4 e−1
4− e−28.04 e−2− 0.267.18 e−3− 0.14 ind − 2.9 e−21.4 e−12.2 e−12.4 e−1
5− 93.172.357.58 e−28.41 e−2− 0.25 ind − 0.14− 4.01 e−3− 2.5 e−21.4 e−12.2 e−12.4 e−1
6− 99.142.536.18 e−27.89 e−2− 0.21 ind − 0.13 ind ind 1.9 e−1 3.2 e−1 8.6 e−1
7− 99.104.736.32 e−27.89 e−2− 0.23 ind − 0.13 ind − 2.0 e−2 1.9 e−1 3.2 e−1 8.6 e−1
8− 99.124.776.93 e−27.94 e−2− 0.223.46 e−3− 0.14 ind ind 1.9 e−1 3.2 e−1 8.6 e−1
9− 99.134.806.58 e−28.04 e−2− 0.21 ind − 0.13− 1.3 e−3 ind 1.9 e−1 3.2 e−1 8.6 e−1
10− 96.975.131.387.90 e−2− 1.89 ind − 1.42 ind ind 1.7 e−12.2 e−12.4 e−1
11− 99.076.977.46 e−27.95 e−2− 0.255.00 e−3− 0.14 ind − 2.5 e−2 1.9 e−1 3.2 e−1 8.6 e−1
12− e−28.15 e−2− 0.24 ind − 0.14− 2.2 e−3− 2.2 e−2 1.9 e−1 3.2 e−1 8.6 e−1
13− 100.810.352.25 e−11.13− 0.37− 1.28− 0.27− 9.98 e−1 ind 1.9 e−1 3.2 e−1 8.6 e−1

The sensitivity analysis showed that all but one model was successfully recaptured, the exception being a non-stationary model that was removed from the analysis. In all remaining models except one, the estimated observation error was smaller than our conservative fixed estimates. This excepted model was, however, removed from consideration because it also estimated negative population growth. Whether observation error was fixed using rough estimates, fixed at large variance (unity) or estimated within the model, the results of model selection were very similar. In all cases, three of the top four models corresponded to the top three models in Table 1, demonstrating reasonable model selection over a wide range of observation variance. The exception to this pattern was model 6 (Table 1), which was estimated as the best model when observation error was fixed at rough estimates or at unity. Model 6 with estimated observation error (as in Table 1), however, still fits better than either the roughly fixed model or the fixed at unity model.

We used model 1 and model 2 to estimate the current – or status quo – fishery harvest and model 2 to estimate the fishery using simulated ‘no-dam’ discharge from ITT (Fig. 4). Similarity in model likelihoods suggests there was no significant difference in CPUE estimated between these best two models. The 95% CI around parameters for models 1 and 2 (Table 1) demonstrated no significant effect of CPUE on effort levels or water regime on CPUE. Moreover, the CI on the estimate of CPUE (Fig. 4a) demonstrated no significant difference between the observed CPUE estimated from model 1 and model 2, or between those models and the ‘no-dam’ water regime simulation.

Figure 4.

 The first and second best model estimates of the Kafue River fisheries modelling results shown in log scale to highlight low abundances and mean observed data. In each panel, black symbols are the mean observations, and solid and dashed lines are median model estimates and 95% confidence intervals, respectively. Black and red lines are model 1 and model 2 estimates under the observed water regime, respectively, and blue lines are model 2 estimates under the simulated ‘no-dam’ water regime. (a) Catch per unit effort with additional available data shown for the year 2008 and 2010 that were not included in the model. (b) Effort in metres of gillnet (○) and boats (•) transformed into gillnets units. (c) The reported and estimated total harvest from the Kafue River. The substantial over-plotting of the fitted lines highlights the general result that hydrological regime has little effect on the fishery.

Effort increased exponentially in both model 1 and model 2 (Fig. 4b). In model 1, the effect of CPUE on effort was non-significant, while model 2 did not include any covariant or density-dependent effects of effort (Table 1). In both models, the intrinsic growth rate of effort was greater than that for the fish population indexed by CPUE (Table 1). This difference in growth rate was only significant in model 2 where the 95% CI of CPUE and effort do not overlap, whereas the CPUE growth rate in model 1 was itself not significant as the 95% CI includes zero and overlaps with the effort CI (Table 1).

The median total harvest (Fig. 4c) corresponded well, within about an order of magnitude, to the reported DoF harvest data, although the total harvest was not directly included in the model. There was no significant difference in harvest between the observed status quo models and the no-dam simulation, although the no-dam scenario suggests a slightly larger harvest. We estimated that revenue derived from total harvest in 2006 was approximately USD $7 million, while under the ‘no-dam’ scenario, it was approximately $9 million, but could range from approximately $1.3 million up to about $56 million per year as result of the large range of the harvest confidence intervals and fishing activity rates. While not statistically different, the difference in median harvest between the status quo and no-dam scenarios is about 900 metric tons, equivalent to about $2.3 million.

The effect of flood regime on hydroelectric generating capacity

The analysis of the impact of ITT Dam on the value of hydropower generated at KG Dam suggests that with ITT Dam in place, the KG turbines can keep 254 m3 s−1 of constant flow during the dry season, which corresponds to 888 MW, since each cubic metre per second generates 3.501 MW (estimated by OLS regression of daily power output on discharge through turbines; R2 = 0.961, = 2161, SE = 0.015) (Fig. 5). The installed generating capacity is 900 MW (corresponding to 256 m3 s−1); therefore, the KG power generation is on average unconstrained. Without ITT Dam, the KG turbines could keep only 203 m3 s−1 of constant flow during the dry season, corresponding to 713 MW. This implies a total power deficit of (888 − 713) MW × 136 days × 24 h × day−1 = 571 200 MWh. The cost of importing electricity to Zambia is about $31 per MWh (PB Power, 2006); thus, we estimate the total annual replacement cost as about $17.7 million if the ITT Dam were not in place.

Figure 5.

 Averaged simulated seasonal hydrographs at Kafue Gorge with (light grey) and without (black) Itezhi-Tezhi Dam. Dashed lines represent the dry-season generating capacity in each scenario, corresponding to 254 m3 s−1 (888 MW) and 203 m3 s−1 (770 MW) with and without Itezhi-Tezhi, respectively. Hashed areas represent the differences in turbine flow during the low water season.


We showed that the construction of the Itezhi-Tezhi Dam had substantial impacts on the water regime in the Kafue Flats. This hydrological manipulation has allowed gains in hydroelectric generating capacity of about $18 million per year at Kafue Gorge Dam, estimated by the replacement cost method (i.e. the cost of purchasing the same amount of electricity from another source at the current price). A more accurate measure of gains would be lost total surplus, which accounts for losses to consumers who must pay higher prices for electricity and to taxpayers who must make up lowered revenues. There is considerable evidence that the Zambian power authority does not price its electricity according to market conditions, however; so it is impossible to estimate lost total surplus by using observed price data (IPA Energy Consulting, 2007). Thus, our estimate is likely an underestimate of the true value of the hydroelectric production benefit of ITT.

Our fishery modelling sensitivity analysis indicated that our modelling methods were able to select appropriate models. The best models, however, did not indicate a significant impact of the dam-altered water regime on the fisheries production of the Kafue River. Our estimates are the first published for the monetary value of this fishery, which is as large as $56 million annually: potentially more than three times as great as the replacement value of hydropower generated as a result of the construction of ITT Dam. Our model selection and simulations suggest, however, that under current fishery practices, no trade-off or at most a small trade-off of about $2.3 million exists between hydropower and fisheries production. The price of tilapia, a preferred species used here to represent prices for all fish species, may overestimate the total value of the fishery; however, the large uncertainty around the total value of the fishery due to unknown fishing activity rates probably overwhelms those for fish price. Moreover, if our price estimate using tilapia overestimates fishery values, and our hydropower analysis undervalues electricity, this further supports a general conclusion of only small, if any, trade-offs between fisheries and hydropower.

Total harvests calculated from our model were consistent with the independently reported DoF statistics. Inland fisheries harvests are notoriously underreported in general (Welcomme, 2011), but considering that harvest was not included in the model and that the CPUE data were apparently not used in constructing the DoF harvest estimates, the similarity of these independent estimates of the Kafue River fishery lends confidence both to the models presented here and to the long-term harvest records reported by the DoF. Differences in these harvest estimates are, however, apparent particularly in the early years where the models predict dramatic declines in harvest, while the DoF data report increasing harvests. It is likely that during the early years of development, the activity rate was lower than even the low estimate (146 days per year) considered here due to large changes in social structures as populations transformed from largely agro-pastoralist with seasonal fishing to largely fishing based (Haller & Merten, 2008). Low overall activity rates implied by seasonal fishing would significantly reduce the estimated total harvest during that time to be more in line with DoF records. The generally close agreement of the model and DoF estimates suggests that in the future, reported harvest should be explicitly incorporated into the model estimation.

In apparent contrast to the results reported here, previous studies have found significant relationships between total harvest and various aspects of pre-dam water regimes on the Kafue River (Chapman et al., 1971; Lagler et al., 1971; Dudley, 1974; Kapetsky & Illies, 1974; Welcomme, 1975; Muncy, 1977). Two studies examined the effect of changed flood regime on fisheries, and these both occurred after the construction of KG Dam downstream, but before ITT Dam was built upstream. These studies are largely consistent with our results, finding no detectable influence on the growth rate of two important tilapia species (Dudley, 1979) or on CPUE for most species (Dudley & Scully, 1980).

It would be incorrect to interpret our results as indicating that there is no relationship between hydrology and fishery production on the Kafue River. Rather, our results are more specific, indicating only that there is no relationship between total fish abundance and the hydrological changes most influenced by the dam, as indexed by the CCA. Selection of appropriate biologically relevant water metrics for comparing average yearly CPUE to the monthly mean flows is a multivariate reduction problem seeking a complex balance between accounting for as much hydrological variation as possible while minimising covariance between indices (Olden & Poff, 2003). Hundreds of metrics have been published for this purpose (Poff et al., 2010). The task of selecting metrics was simplified in this case because we only need to find the differences in the hydrograph along one dimension, before and after the ITT Dam, rather than search for metrics with biological relevance. The ability of the CCA to clearly discriminate the multivariate differences in hydrological regime before and after dam construction is a novel and critical strength of our analysis. This approach could also be applied predictively to the many regions where dams are being constructed or are proposed, given expectations about future dam-induced alteration of flows.

The results of this study also apparently contrast with previous studies of the effects of damming on fish in many streams and rivers around the world (Poff & Zimmerman, 2010). We suggest four reasons that these conflicting results may arise. First, we considered the whole fish community in both experimental gillnets and total harvest, while previous work has dealt with particular specialist or sensitive taxa (Poff & Zimmerman, 2010). It is possible that by considering the total fish community, negative impacts on sensitive or target taxa may be obscured by positive effects on other species, and vice versa, due to CPUE aggregating effects (Kleiber & Maunder, 2008) and/or the portfolio effect (Hooper et al., 2005). The fishery also appears to have few strong species preferences; all species of all sizes are harvested and successfully marketed. Species-specific contributions to the fishery markets may therefore also be obscured by the portfolio effect. A species-specific analysis to hydrological change in the Kafue River would be a logical next step.

Species-specific data are relatively sparse in the early data set (Table S2), and the selectivity of gillnets data further obscure such analysis, but some examples are informative. For instance, the spiny eel, Mastacembelidae, and eastern bottlenose, Mormyrus longirostris, have not been reported for almost 30 years, since just after dam construction. Conversely, the Nile tilapia Oreochromis niloticus was introduced to the Kafue River in the 1980s (Schwanck, 1995), and our gillnetting efforts in 2008 and 2010 reveal that this species is now as common as the native O. andersonii and is caught throughout the Kafue River between ITT and KG. It seems unlikely that species losses or additions would, however, obscure the general result reported. Species losses would only strengthen the signal of negative hydrological effects. Nile tilapia, while increasingly common, make up only a small portion of the total fish caught in experimental gillnets: 0.2% of all fish and 0.4% of the biomass in the whole post-dam era and a maximum of 0.7% of all fish and 5% of the biomass in 2010, placing their particular contribution well within the CPUE 95% confidence intervals. It also seems unlikely that changes in species-specific fisher preferences were significant, and lists of important fisheries from the early fishery (e.g. Williams, 1960) are largely similar to more recent observations. These factors together suggest that the whole fishery production, by biomass and value, is the most appropriate metrics for our purposes of comparing fishing and hydropower production.

A second reason our results seemingly conflict with previous studies may be due to the exceptionally long and highly resolved data set from both before and after flow alteration. The relevant fish-related papers reviewed in Poff & Zimmerman (2010) contain a maximum of 10 years of pre-dam data and 45 years of post-dam data, although the average and median post-dam data set is only 6 and 2 years, respectively (n = 33). We provide 25 years of pre-dam observations and 29 years of post-dam observations. It seems plausible that long-term compensatory changes in the fishery could swamp short-term effects of damming detected in other studies.

Third, many problems that are well known to exist with the quality of long-term inland fisheries data may apply less strongly to the Kafue River data. Globally, underreporting of inland fisheries statistics is a serious problem, particularly in small subsistence fisheries with diffuse landing sites and local consumption which hampers collection of harvest records and ultimately the total production reported (Welcomme, 2011). The Kafue CPUE data considered here avoids these common complications because data are derived directly from research sampling and are not dependent on accurate accrual of harvest data from fishers. In many systems, effort in terms of fish collection gears and protocols has been inconsistent and difficult to estimate accurately (De Graaf et al., 2012), but this also appears less so in the Kafue fishery where boats and gillnets have changed little over time. On the other hand, we know catches with seine nets and other prohibited methods in the Kafue were underreported and therefore poorly represented in the model. We also did not include the possibility of changes in fisher behaviour as a direct result of hydrological regime, although it is possible that dam construction drew new participants to the fishery. Overall, however, the consistency of the independent CPUE, effort and total harvest observations over the full length of the data and compiled across multiple studies leads us to believe that the Kafue has fared very well in comparison with other systems in terms of the quality of data collection. Indeed, the Kafue was considered the best-documented African floodplain fishery before damming (Welcomme, 1975).

Finally, we reconcile the differences in the Kafue River and other studies of the relationship between hydrology and fish abundance as the result of harvesting effort in the Kafue River over the period of study that is likely more intense (and increasing) than in most if not all other studied ecosystems. The impact of harvest on fish abundance probably overwhelms the effect from hydrological manipulation. The z-scored data used in the population modelling allows direct comparison of the relative magnitudes of the effects (and uncertainty) of effort and CCA scores (Table 1). In models 1 and 2, the effects of fishing were not significantly different, but in model 2, the best model that included an effect of water regime, the effect of water regime on fish abundance was an order of magnitude lower than the effect of effort. Moreover, the growth rate of fishing effort is significantly larger than the growth rate of the fish. Meanwhile, the observation error variance for both gillnets and boats in models 1 and 2 is about twice that for CPUE and an order of magnitude larger than the effect of water regime, indicating that the relatively small effects of water regime are easily lost in the noise that surrounds effort in the system.


Dam construction does not seem to have had significant impacts on the Kafue Flats fishery. The overall trends in CPUE, effort and harvest in the Kafue fishery are largely consistent with overfishing, particularly the concept of fishing-down in open-access fisheries (Allan et al., 2005). Our data and analysis of the effort on the Kafue River demonstrate that fishing effort has been continually and exponentially increasing over time mostly independent from the abundance of the fish and fishers already present, suggesting little internal control over effort and the primacy of increased fishing effort on the observed decline in fish abundance. These results do not, however, rule out an interaction between fishing effort and hydrological regime such that a reduction in fishing effort could prompt a response to changes in hydrology or magnify a currently unobserved response into the realm of detectability.

The implication of these results is that effort reduction will be the most effective way to increase fisheries output and thereby improve the livelihoods of fishermen. Management tactics that aim to directly increase fish abundance without addressing effort, such as habitat improvements via environmental flows, direct stocking or species introductions, are likely to be ineffectual or short-term solutions. These issues closely mirror general observations of global fisheries, where a wide assortment of context-specific management strategies are seeing some success at rebuilding important fish stocks by decreasing exploitation rates (Worm et al., 2009). In Zambia, national policy safeguards the right to fish for all Zambians, limiting the power of both the state and traditional institutions to regulate access in the Kafue fishery (Haller & Merten, 2008). Yet, indirectly limiting effort through closed fishing seasons, closed areas and gear restrictions are the main regulatory instruments available to the DoF. Devolution of management to local levels with participatory management schemes has had varying levels of success in Lake Kariba and the Mweru–Luapula complex in northern Zambia, at least in terms of stakeholder acceptance (Kapasa, 2004, AMD Pers. Obs.). These and future management strategies will need to explore and accommodate broader socio-economic forces which we show as external drivers of the Kafue fishery in terms of effort and ultimately fish abundance and harvests. We have provided a foundation for future, more comprehensive analyses of alternative management scenarios by estimating linkages to the wider economy in terms of the monetary value of the Kafue fishery and hydropower production. We have not considered other river-related ecosystem services that are potentially modified by damming, such as the provisioning of pasturing for livestock and habitat for wildlife, not because we think these values are less significant, only less tenable. Given the foundation provided here, changes in management practices in the future could probably increase total ecosystem services from the Kafue River.


P. Ngalande, L. Njobvu Chilufya, R. Nkhata, I. Bbole and many others from the Zambian Department of Fisheries contributed data and other assistance. H.G. Mudenda, B. Klco, J. Deines and the World Fish Center assisted with much on the ground support. The Zambian Department of Water Affairs and J. Kolding contributed data and J. McLachlan made helpful suggestions on the fisheries model. M. Wittmann and the Lodge Lab and two anonymous reviewers provided comments on earlier versions of the manuscript. Funding was provided by NSF awards #0504495 and #1046682, and The Kellogg Institute at the University of Notre Dame.