The primary outcome in this study is the weekly number of patients visiting the International Centre for Diarrhoeal Disease Research, Bangladesh (ICDDR,B) Dhaka Hospital due to non-cholera diarrhoea. The hospital serves an urban population of approximately 10 million individuals6 and provides free treatment to more than 100 000 cases of diarrhoea each year. Every 50th patient visiting the hospital has been enrolled in a surveillance system since 1996. For all patients enrolled in the surveillance, microbiological examination of stool or a rectal swab sample was conducted to identify enteric pathogens. The patient or family members were interviewed by health workers who collect socio-demographic data at the time of the hospital visits. We abstracted individual information on the date of the hospital visit, age, sex, socio-economic status (educational level and roof structure of the house), hygiene and sanitation practices (drinking water source, distance to the water source and type of toilet) and pathogens identified from stool specimen during a 7-year period (January 1996–December 2002). A patient was classified as non-cholera diarrhoea when Vibrio cholerae was not identified from the stool specimen. The cause of non-cholera diarrhoea was categorized as rotavirus, Shigella, Salmonella, Campylobacter, Escherichia coli, Aeromonas or other diarrhoea to show the components of non-cholera diarrhoea. Salmonella includes all Salmonella species except for S. typhi. E. coli consists of enterotoxigenic and enteropathogenic E. coli. Other diarrhoea includes diarrhoea with none of pathogens identified. When two or more pathogens amongst these non-cholera pathogens were identified from the same stool specimen, the patient was classified in each category of pathogen-specific diarrhoea for pathogen-specific descriptive analysis. Parasites were also routinely examined during the study period including Cryptosporidium parvum, Entamoeba histolytica and Giardia lamblia of which the number of cases was small.
Meteorological and river-level data
We obtained daily rainfall and maximum and minimum temperature in Dhaka from the Bangladesh Meteorological Department. The daily river-level of the Brigonga River at Mill Barrack in Dhaka was recorded by the Bangladesh Water Development Board. The weekly means for maximum temperature and maximum river-level and the total weekly rainfall were calculated from the daily records.
Statistical methods are summarized here and described in detail in the supplementary online appendix. We examined the relationship of the number of weekly non-cholera cases with rainfall and temperature using generalized linear Poisson regression models allowing for overdispersion.7 To account for the seasonality of non-cholera counts not directly due to the weather, Fourier terms up to the sixth harmonic were introduced into the model. Indicator variables for the years of the study were incorporated into the model to allow for long-term trends and other variations between years. An indicator variable for public holidays was incorporated into the model to control bias in the event that holidays affected access to the hospital, as was suggested in a previous time-series study in the UK.8 To allow for the autocorrelations an autoregressive term at order one was incorporated into the models.9
Models for rainfall
From exploratory analyses, existing literature and considerations of interpretational difficulty with very long lags, we considered lags (delays in effect) of up to 16 weeks for rainfall. In the initial analyses designed to identify the broad shape of any association, we fitted natural cubic splines (3 df)10 to (i) the average rainfall over lags 0–16 and (ii) the average over 0–8 weeks and 9–16 weeks, as separate splines simultaneously included in the model. We also included temperature as a natural cubic spline (3 df) in all models to control confounding, with lag 0–4 weeks, following expectation from published work.3,8,11 Because initial analyses suggested a broad ‘U’ shape, we then fitted a double-thresholds model, comprising linear terms for rainfall above and below ‘high’ and ‘low’ thresholds, respectively, with no association (i.e. flat) in between.12 Guided by the spline analysis the low and high rainfall terms were based on the 0–16 and 0–8 week average, respectively. The thresholds were estimated by maximum likelihood. An increase or decrease in the number of cases that were associated with a 10 mm increase or decrease in a given measure of rainfall, estimated as coefficients from the regression model, was reported as percentage change.
With the simple thresholds model we then examined lag effects in more detail, by fitting linear unconstrained distributed lag models, comprising terms for low and high rainfall at each lag up to the previous 16 weeks.12 The simple linear-thresholds model also allowed investigation of the modification of rainfall effects by patient characteristics (e.g. socio-economic status), by fitting the models separately to incidence series according to their characteristics.
Models for temperature
Delayed effects of temperature on diarrhoea due to some pathogens, have been known to be approximately 1 month,3,8,11 so we considered lags of up to 4 weeks for temperature. Rainfall terms with natural cubic spline (3 df) were included in all models to control confounding of rainfall over lags 0–8 and 9–16 weeks. Because a smooth relationship using natural cubic splines (3 df) suggested a log-linear association through the whole range of temperature, we fitted simple linear models. Detailed lag effects were examined by fitting linear unconstrained distributed lag models comprising terms for temperature at each lag up to previous 4 weeks.12 We also investigated the modification of temperature effects by patient characteristics, by the same methods with models for rainfall.
The same analyses were conducted for non-cholera diarrhoea without the inclusion of rotavirus because its seasonality differs from the rest of the pathogens. Sensitivity of estimates to the degree of seasonal control (3 and 12 harmonics) was also examined. All statistical analyses were carried out using Stata 9.0 (Stata Corporation, College Station, TX, USA).