* May's job gain of 95,000 well above any sampling error
* Statscan takes measures to eliminate variability
* Steps including personal visits, followed by phone calls
By Randall Palmer
OTTAWA, June 7 (Reuters) - There's a note of caution surrounding the stunning gain of 95,000 Canadian jobs in May, because the method Statistics Canada uses to get its figures brings both volatility and a wide margin of error.
But Statscan experts still see a 95 percent chance that the employment gains were between 37,600 and 152,400 in May - a very substantial monthly increase given the overall size of the Canadian economy.
Unlike the main U.S. payrolls data, Canada's most closely watched labor market data is based on monthly interviews with individual households, a method that is always subject to a substantial sampling error.
The federal agency surveys about 56,000 households, representing about 110,000 individuals aged 15 and older. Every month, five-sixths of those in the sample are people who the agency has contacted before and one out of six is new.
Economists occasionally question the reliability of the monthly results or dismiss dramatic swings like the one in May as anomalies, and even the government looks beyond data from a single month.
"We don't obviously want to pay too much attention to what happens month to month, because we know this is very volatile," Prime Minister Stephen Harper said on Friday.
Statscan, which has a sterling reputation internationally, insists its methodology and sample size are more rigorous than those used in political opinion polls and says it places a lot of emphasis on eliminating improper variability from its Labor Force Survey, one of the most closely watched economic indicators of each month.
"Accuracy is very good compared with the typical political poll," Statscan analyst Jason Gilmore said.
The agency takes pains to cut the problems pollsters face in getting an accurate sample, because of the prevalence of cellphones and because people may simply refuse to answer a phone if it's a number they do not recognize.
A Statscan official visits most of the households in the survey in person, and then contacts them again by telephone over the following five months. During the initial visit, the agency can collect cellphone and other data to make later contact more sure.
A separate Statscan report on payroll employment, earnings and hours is based largely on data from businesses and is more like the U.S. payrolls survey. However, it appears with a delay of nearly two months - March data came out on May 29. Canada does not have any weekly or bi-weekly employment indicators.
Statscan says the standard error in its labor force survey is 28,700 given the size of its sample and the number of employed Canadians at 17.75 million.
That means the number of jobs lost or created in any given month could be up to 28,700 above or below the number provided by Statscan two times out of three.
That standard error increases to 57,400 when one uses the more common "19 times out of 20" measure of accuracy. So both April's gain of 12,500 jobs and March's loss of 54,500 jobs were both within the margin of error.
The data is adjusted for seasonal factors, such as the normal upswing in construction employment in the spring and the layoffs that take place in the autumn.
Hence when May construction employment rose by 42,700, it rose by that much above the normally higher spring levels.
Gilmore said this may partly reflect unseasonably cold weather in April, which may have led to builders postponing some of their hiring into May.
April's seasonally adjusted construction employment figure was little changed, but the average increase in the two months of 21,600 does show a boost in activity in the sector beyond the usual spring jump.