
CME 




Year : 2006  Volume
: 31
 Issue : 1  Page : 46 

Confidence intervals and test of significance
MV Ajay Kumar
Department of P&SM, JIPMER, Pondicherry6, India
Date of Web Publication  8Aug2009 
Correspondence Address: M V Ajay Kumar Department of P&SM, JIPMER, Pondicherry6 India
Source of Support: None, Conflict of Interest: None  Check 
DOI: 10.4103/09700218.54942
How to cite this article: Ajay Kumar M V. Confidence intervals and test of significance. Indian J Community Med 2006;31:46 
In the article entitled "Prevalence and Risk Factors of Hypertension in Adults in an Urban Slum, Tirupati, A.P^{ [}¹^{ ]}", TableVI the various risk factors with their strength of associations are given. All the given factors in the table were shown as statistically significant. Usually if the 95% CI of odds ratio includes 1, it is considered to be 'not significant'. The 95% CI of two of the factors given in the table  'saturated fat intake' and 'non vegetarian diet' included 1 but were shown to be significant (p<0.05). The chi square test as expected, revealed it to be 'not significant' (p values of 0.775 and 0.577 respectively). Confidence interval of 'male sex' also included 1 [Table 1] (p value of 0.25).
Confidence interval of odds ratio is asymmetrical around the point estimate  the upper limb is always longer than the lower limb. For example take age >40 years. The odds ratio (95% CI) is 30.2 (1370) as calculated by Woolf's method [Table 1]. The upper limb of the confidence interval is 39.8(70 minus 30.2) and the lower limb is 17.2(30.2 minus 13). This is because of the fact that the distribution of odds ratio is skewed extending from '0 to 1' on one end and '1 to infinity' on the other end. This skewness is captured in the 95% CI of the odds ratio. Hence the upper limb is always longer. But some of the confidence intervals published in the article showed equal upper and lower limbs (non vegetarian diet, no regular exercise, family history of hypertension) and some showed that upper limb is lesser than lower limb (excess salt, saturated fat intake). The most common method used is 'Woolf's method' and uses the formula
for calculating the standard error of natural log of odds ratio^{ [2]} . Once that is found out 95% CI for log natural odds ratio can be found out. Taking antilog gives the actual CI. The other method is 'Chi Square test based method' and uses the formula OR(1±.1.96/x) where OR is odds ratio and x_{ } is the square root of the chi square value^{ [2]} . The Cornfield's method is used in Epiinfo software^{ [3]} . These three methods were used to calculate CI [Table 1]. Though there are inter method variations in the intervals obtained, they do not violate the two interpretations made above  the asymmetry of the interval and the statistical significance interpretation. This shows that if there is a mention of the method used for the calculation of the confidence interval, it brings much more clarity and helps in verifiability. Such errors can easily be avoided.
Fischer's exact test should have been used instead of chisquare test for 'history of previous events'. That gives a p value of 0.003 and not p<0.001 as mentioned in the article.
References   
1.  Singh US, Choudhary SK. Prevalence and risk factors of hypertension in adults in an urban slum, Tirupati, AP. IJCM, 2005;30:8486. 
2.  Silman AJ, Macfarlane GJ. Epidemiological studies A Practical guide. Second edition; London Cambridge University Press 2002. 
3.  Epi info Version 3.3.2 software available on www.cdc.gov as on September 23, 2005. 
[Table 1]
