Confidence intervals:
Precision = repeatibility from one sample to the next.
Accuracy = absence of bias.
Point estimate = single point value for parameter.
Interval estimate = range of values around point estimate which have a given probability of including the population parameter.
95% confidence interval (CI) = if the confidence intervals are calculated for one hundred different sample studies, then 95 of them will PROBABLY contain the true value (population parameter). Generally if CI includes 0 then test is not siginificant. For large samples (n>30), the confidence interval is the mean ñ 1.96 standard dedviations. n is the number of values in the sample.
For small samples (n<30), central limit theroem does NOT hold, so use 'Students t distribution'. Degrees of freedom = n-1. Confidence interval = mean ñ t(0.025,n-1) x SD.

Hypothesis testing:
Null hypothesis (H0) = the characteristic we are assuming about the population. Usually says that a population parameter takes a specified value or that two population parameters are equal. eg. congenital hip dislocations occur girls than boys equally often.
Althernative hypothesis (H1) = contracticts the null hyphotesis, and will be accepted if the null hypothesis is rejected. Usually very general in nature. eg. there are more congenital hip dislocations in girls than boys. P-value = the probability of getting a test statistic as extreme as the one determined. If p<0.05 and H0 is true, then only 1 in 20 times should we get such extreme result. So reject H0. If p>0.05 then cannot reject H0, but we have not proved it either. Significance level (à) = probability cut off for the test. The most commonly used value is 0.05 (5%).
Power (1-Beta) = probability of rejecting H0 when H1 is true. Power>0.8 is acceptable. Power is increased by taking larger samples or using less stringent significance level (eg. 0.1).
Type I error = rejecting H0 (as p<0.05) when it is actually true. So can only occur when p<0.05. The probability of a type I error is the significance level, à.
Type II error = not rejecting H0 (as p>0.05) when H1 is true. So can only occur when p>0.05. The probability of a type II error is Beta. So Power is probability of NOT making type II error.
Paired t-test: t=d/SE(d), where d=mean within-pair difference in sample, SE=(standard deviation of within-pair difference)/(ûn) Independant samples t-test: t=(x1-x2)/SE(x1-x2), where X1 and X2 are sample means, etc. Sometimes only non-parametric tests (eg. Mann-Whitney U test, or Wilcoxon Signed-Rank test) using ranked (ie. ordered) values can be used. These have less power and are less flexible. One tailed and two tailed tests use one or two ends of the normal or t distribution.

Contingence tables:
Use 2x2 table and chi-squared (or Fisher's Exact test or McNemar's test) to determinine if there is significant association between usually a disease and a possible cause.
 

(a) Osteoperosis risk factors: Female, early menopause, smoking, elderly, steroid drugs, poor diet, little exercise, family history.
Future changes of disease causes: more elderly in population, lifestyle & diet, effective prevention & treatment programmes.

In case-control studies: decide threshold disease level (eg. eosinophils>x), match cases & control on age, sex, where they live. Exclude people with eg. related allergies or disease, as may bias study. Referals of cases from local authorities may bias study. Note dates of disease occurance, as perhaps due to bad batch of food or drugs.
Cohort study (following people taking drug for a long time) is best study, but takes much time. Body Mass Index (BMI) = Weight/(Heightý). Obese=BMI>30, Underweight=BMI<20.

ETC