Last week we examined the movement of support for all the main UK parties since the start of 2015. Since then we now have a complete set of polling data from YouGov for February so we can explore movement during that month.
In the analysis below we have built a simple OLS regression model on the polling data to see if any trends are significant. From this first model we remove clear outliers (looking at Cook’s D) and then rebuild.
The results are below.
The table below summarises the data.
|Party||Linear Trend at 90% Significance Level||Linear Trend at 95% Significance Level|
|Liberal Democrat||Yes – Upwards||Yes – Upwards|
|Greens||Yes – Downwards||Yes – Downwards|
Last week we saw that since the start of the year UKIP support had declined at a statistically significant level (i.e. we believe that the fall in polling values is not just due to random effects but is an underlying trend across the UK in general). Our analysis of all of the February data shows that that statistically significant trend has now vanished – we are not certain there continues to be a decline in UKIP support.
We tested this hypothesis by building a polynomial model of UKIP support. We found that a polynomial trend gave an r-squared of 0.1519 and Pr > F of 0.2466 compared to the linear trend model with an r-squared of 0.0530 and Pr > F of 0.3429. Neither model is statistically significant, but the polynomial model did not remove the outlier of the 4th of February and produced a shallowing curve, indicative of the decline in UKIP support “flattening out”.
Green Party support has clearly declined in February – a simple runs tests shows a clustering of values around 7% in the first half of the year and 6% in the second. Our linear model indicated a clear decline and polynomial analysis indicates that it is not flattening off.
A runs test on the Conservative polling data in February indicates there is the possibility of a statistically significant growth trend emerging – we will need to repeat this analysis in two weeks time to see if it has clearly developed.
There are a number of improvements that could be implemented on this analysis.
- There were not enough data items in the set of YouGov polling data since the start of 2015 to conduct a Kolmogorov-Smirnov test to assess randomness. It might also be appropriate to conduct an Augmented Dickey-Fuller test on the data (more appropriate for a time series set), but again sample size limitations make this difficult.